Revert "[nit]: Fix import for the lmcache integration (#27600)"

This reverts commit 05e034f085.

.buildkite/lm-eval-harness/configs/Qwen2-1.5B-Instruct-W8A16-compressed-tensors.yaml

@@ -0,0 +1,12 @@
+# For vllm script, with -t option (tensor parallel size).
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Qwen2-1.5B-Instruct-W8A16-Channelwise -b "auto" -l 1000 -f 5 -t 1
+model_name: "nm-testing/Qwen2-1.5B-Instruct-W8A16-Channelwise"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.595
+  - name: "exact_match,flexible-extract"
+    value: 0.582
+limit: 1000
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Qwen3-235B-A22B-Instruct-2507-FP8.yaml

@@ -1,14 +0,0 @@
-model_name: "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8"
-tasks:
-  - name: "mmlu_pro"
-    metrics:
-      - name: "exact_match,custom-extract"
-        value: 0.82
-limit: 250 # will run on 250 * 14 subjects = 3500 samples
-num_fewshot: 5
-enforce_eager: false # we use false to speed up the eval process
-kv_cache_dtype: fp8 # we use fp8 to speed up the eval process
-max_model_len: 40960
-apply_chat_template: true
-fewshot_as_multiturn: true
-gen_kwargs: "temperature=0,top_p=1,top_k=0,max_gen_toks=5632,until=<|ENDANSWER|>"

.buildkite/lm-eval-harness/configs/models-large-h100.txt

@@ -0,0 +1 @@
+Meta-Llama-4-Maverick-17B-128E-Instruct-FP8.yaml

.buildkite/lm-eval-harness/configs/models-large-hopper.txt

@@ -1 +0,0 @@
-Qwen3-235B-A22B-Instruct-2507-FP8.yaml

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -21,13 +21,10 @@ def launch_lm_eval(eval_config, tp_size):
     max_model_len = eval_config.get("max_model_len", 4096)
     batch_size = eval_config.get("batch_size", "auto")
     backend = eval_config.get("backend", "vllm")
-    enforce_eager = eval_config.get("enforce_eager", "true")
-    kv_cache_dtype = eval_config.get("kv_cache_dtype", "auto")
     model_args = (
         f"pretrained={eval_config['model_name']},"
         f"tensor_parallel_size={tp_size},"
-        f"enforce_eager={enforce_eager},"
-        f"kv_cache_dtype={kv_cache_dtype},"
+        f"enforce_eager=true,"
         f"add_bos_token=true,"
         f"trust_remote_code={trust_remote_code},"
         f"max_model_len={max_model_len},"
@@ -40,13 +37,8 @@ def launch_lm_eval(eval_config, tp_size):
         limit=eval_config["limit"],
         # TODO(yeq): using chat template w/ fewshot_as_multiturn is supposed help
         # text models. however, this is regressing measured strict-match for
-        # existing text models in CI, so only apply it for mm, or explicitly set
-        apply_chat_template=eval_config.get(
-            "apply_chat_template", backend == "vllm-vlm"
-        ),
-        fewshot_as_multiturn=eval_config.get("fewshot_as_multiturn", False),
-        # Forward decoding and early-stop controls (e.g., max_gen_toks, until=...)
-        gen_kwargs=eval_config.get("gen_kwargs"),
+        # existing text models in CI, so only apply it for mm.
+        apply_chat_template=backend == "vllm-vlm",
         batch_size=batch_size,
     )
     return results

.buildkite/nightly-benchmarks/README.md

@@ -2,23 +2,40 @@
 
 ## Introduction
 
-This directory contains a benchmarking suite for **developers** to run locally and gain clarity on whether their PR improves/degrades vllm's performance.
-vLLM also maintains a continuous performance benchmark under [perf.vllm.ai](https://perf.vllm.ai/), hosted under PyTorch CI HUD.
+This directory contains two sets of benchmark for vllm.
+
+- Performance benchmark: benchmark vllm's performance under various workload, for **developers** to gain clarity on whether their PR improves/degrades vllm's performance
+- Nightly benchmark: compare vllm's performance against alternatives (tgi, trt-llm and lmdeploy), for **the public** to know when to choose vllm.
+
+See [vLLM performance dashboard](https://hud.pytorch.org/benchmark/llms?repoName=vllm-project%2Fvllm) for the latest performance benchmark results and [vLLM GitHub README](https://github.com/vllm-project/vllm/blob/main/README.md) for latest nightly benchmark results.
 
 ## Performance benchmark quick overview
 
-**Benchmarking Coverage**: latency, throughput and fix-qps serving on B200, A100, H100, Intel® Xeon® Processors and Intel® Gaudi® 3 Accelerators with different models.
+**Benchmarking Coverage**: latency, throughput and fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!) and Intel® Xeon® Processors, with different models.
 
 **Benchmarking Duration**: about 1hr.
 
 **For benchmarking developers**: please try your best to constraint the duration of benchmarking to about 1 hr so that it won't take forever to run.
 
+## Nightly benchmark quick overview
+
+**Benchmarking Coverage**: Fix-qps serving on A100 (the support for FP8 benchmark on H100 is coming!) on Llama-3 8B, 70B and Mixtral 8x7B.
+
+**Benchmarking engines**: vllm, TGI, trt-llm and lmdeploy.
+
+**Benchmarking Duration**: about 3.5hrs.
+
 ## Trigger the benchmark
 
-The benchmark needs to be triggered manually:
+Performance benchmark will be triggered when:
+
+- A PR being merged into vllm.
+- Every commit for those PRs with `perf-benchmarks` label AND `ready` label.
+
+Manually Trigger the benchmark
 
 ```bash
-bash .buildkite/performance-benchmarks/scripts/run-performance-benchmarks.sh
+bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
 ```
 
 Runtime environment variables:
@@ -30,11 +47,14 @@ Runtime environment variables:
 - `REMOTE_HOST`: IP for the remote vLLM service to benchmark. Default value is empty string.
 - `REMOTE_PORT`: Port for the remote vLLM service to benchmark. Default value is empty string.
 
+Nightly benchmark will be triggered when:
+
+- Every commit for those PRs with `perf-benchmarks` label and `nightly-benchmarks` label.
+
 ## Performance benchmark details
 
 See [performance-benchmarks-descriptions.md](performance-benchmarks-descriptions.md) for detailed descriptions, and use `tests/latency-tests.json`, `tests/throughput-tests.json`, `tests/serving-tests.json` to configure the test cases.
 > NOTE: For Intel® Xeon® Processors, use `tests/latency-tests-cpu.json`, `tests/throughput-tests-cpu.json`, `tests/serving-tests-cpu.json` instead.
-For Intel® Gaudi® 3 Accelerators, use `tests/latency-tests-hpu.json`, `tests/throughput-tests-hpu.json`, `tests/serving-tests-hpu.json` instead.
 >
 ### Latency test
 
@@ -132,3 +152,26 @@ Here is an example using the script to compare result_a and result_b with Model,
 A comparison diagram will be generated below the table.
 Here is an example to compare between 96c/results_gnr_96c_091_tp2pp3 and 128c/results_gnr_128c_091_tp2pp3
 <img width="1886" height="828" alt="image" src="https://github.com/user-attachments/assets/c02a43ef-25d0-4fd6-90e5-2169a28682dd" />
+
+## Nightly test details
+
+See [nightly-descriptions.md](nightly-descriptions.md) for the detailed description on test workload, models and docker containers of benchmarking other llm engines.
+
+### Workflow
+
+- The [nightly-pipeline.yaml](nightly-pipeline.yaml) specifies the docker containers for different LLM serving engines.
+- Inside each container, we run [scripts/run-nightly-benchmarks.sh](scripts/run-nightly-benchmarks.sh), which will probe the serving engine of the current container.
+- The `scripts/run-nightly-benchmarks.sh` will parse the workload described in [nightly-tests.json](tests/nightly-tests.json) and launch the right benchmark for the specified serving engine via `scripts/launch-server.sh`.
+- At last, we run [scripts/summary-nightly-results.py](scripts/summary-nightly-results.py) to collect and plot the final benchmarking results, and update the results to buildkite.
+
+### Nightly tests
+
+In [nightly-tests.json](tests/nightly-tests.json), we include the command line arguments for benchmarking commands, together with the benchmarking test cases. The format is highly similar to performance benchmark.
+
+### Docker containers
+
+The docker containers for benchmarking are specified in `nightly-pipeline.yaml`.
+
+WARNING: the docker versions are HARD-CODED and SHOULD BE ALIGNED WITH `nightly-descriptions.md`. The docker versions need to be hard-coded as there are several version-specific bug fixes inside `scripts/run-nightly-benchmarks.sh` and `scripts/launch-server.sh`.
+
+WARNING: populating `trt-llm` to latest version is not easy, as it requires updating several protobuf files in [tensorrt-demo](https://github.com/neuralmagic/tensorrt-demo.git).

.buildkite/nightly-benchmarks/benchmark-pipeline.yaml

@@ -0,0 +1,184 @@
+steps:
+  - label: "Wait for container to be ready"
+    key: wait-for-container-image
+    agents:
+      queue: A100
+    plugins:
+    - kubernetes:
+        podSpec:
+          containers:
+          - image: badouralix/curl-jq
+            command:
+            - sh .buildkite/nightly-benchmarks/scripts/wait-for-image.sh
+  - label: "Cleanup H100"
+    agents:
+      queue: H100
+    depends_on: ~
+    command: docker system prune -a --volumes --force
+  
+  - label: "A100"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: A100
+    depends_on: wait-for-container-image
+    if: build.branch == "main"
+    plugins:
+    - kubernetes:
+        podSpec:
+          priorityClassName: perf-benchmark
+          containers:
+          - image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT
+            command:
+            - bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+            resources:
+              limits:
+                nvidia.com/gpu: 8
+            volumeMounts:
+            - name: devshm
+              mountPath: /dev/shm
+            env:
+            - name: VLLM_USAGE_SOURCE
+              value: ci-test
+            - name: HF_TOKEN
+              valueFrom:
+                secretKeyRef:
+                  name: hf-token-secret
+                  key: token
+          nodeSelector:
+            nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
+          volumes:
+          - name: devshm
+            emptyDir:
+              medium: Memory
+
+  - label: "H200"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: H200
+    depends_on: wait-for-container-image
+    if: build.branch == "main"
+    plugins:
+    - docker#v5.12.0:
+        image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT
+        command:
+        - bash
+        - .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+        mount-buildkite-agent: true
+        propagate-environment: true
+        ipc: host
+        gpus: 4,5,6,7
+        volumes:
+          - /data/benchmark-hf-cache:/root/.cache/huggingface
+        environment:
+        - VLLM_USAGE_SOURCE
+        - HF_TOKEN
+
+  #- block: "Run H100 Benchmark"
+    #key: block-h100
+    #depends_on: ~
+
+  - label: "H100"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: H100
+    depends_on: wait-for-container-image
+    if: build.branch == "main"
+    plugins:
+    - docker#v5.12.0:
+        image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:$BUILDKITE_COMMIT
+        command:
+        - bash
+        - .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+        mount-buildkite-agent: true
+        propagate-environment: true
+        ipc: host
+        gpus: all # see CUDA_VISIBLE_DEVICES for actual GPUs used
+        volumes:
+          - /data/benchmark-hf-cache:/root/.cache/huggingface
+        environment:
+        - VLLM_USAGE_SOURCE
+        - HF_TOKEN
+
+  # Premerge benchmark
+  - label: "A100"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: A100
+    depends_on: wait-for-container-image
+    if: build.branch != "main"
+    plugins:
+    - kubernetes:
+        podSpec:
+          priorityClassName: perf-benchmark
+          containers:
+          - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+            command:
+            - bash .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+            resources:
+              limits:
+                nvidia.com/gpu: 8
+            volumeMounts:
+            - name: devshm
+              mountPath: /dev/shm
+            env:
+            - name: VLLM_USAGE_SOURCE
+              value: ci-test
+            - name: HF_TOKEN
+              valueFrom:
+                secretKeyRef:
+                  name: hf-token-secret
+                  key: token
+          nodeSelector:
+            nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
+          volumes:
+          - name: devshm
+            emptyDir:
+              medium: Memory
+
+  - label: "H200"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: H200
+    depends_on: wait-for-container-image
+    if: build.branch != "main"
+    plugins:
+    - docker#v5.12.0:
+        image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+        command:
+        - bash
+        - .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+        mount-buildkite-agent: true
+        propagate-environment: true
+        ipc: host
+        gpus: 4,5,6,7
+        volumes:
+          - /data/benchmark-hf-cache:/root/.cache/huggingface
+        environment:
+        - VLLM_USAGE_SOURCE
+        - HF_TOKEN
+
+  #- block: "Run H100 Benchmark"
+    #key: block-h100
+    #depends_on: ~
+
+  - label: "H100"
+    # skip: "use this flag to conditionally skip the benchmark step, useful for PR testing"
+    agents:
+      queue: H100
+    depends_on: wait-for-container-image
+    if: build.branch != "main"
+    plugins:
+    - docker#v5.12.0:
+        image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+        command:
+        - bash
+        - .buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh
+        mount-buildkite-agent: true
+        propagate-environment: true
+        ipc: host
+        gpus: all # see CUDA_VISIBLE_DEVICES for actual GPUs used
+        volumes:
+          - /data/benchmark-hf-cache:/root/.cache/huggingface
+        environment:
+        - VLLM_USAGE_SOURCE
+        - HF_TOKEN

.buildkite/nightly-benchmarks/nightly-annotation.md

@@ -0,0 +1,28 @@
+# Nightly benchmark annotation
+
+## Description
+
+This file contains the downloading link for benchmarking results.
+
+- [benchmarking pipeline](artifact://nightly-pipeline.yaml)
+- [benchmarking results](artifact://results.zip)
+- [benchmarking code](artifact://nightly-benchmarks.zip)
+
+Please download the visualization scripts in the post
+
+## Results reproduction
+
+- Find the docker we use in `benchmarking pipeline`
+- Deploy the docker, and inside the docker:
+    - Download `nightly-benchmarks.zip`.
+    - In the same folder, run the following code:
+
+    ```bash
+    export HF_TOKEN=<your HF token>
+    apt update
+    apt install -y git
+    unzip nightly-benchmarks.zip
+    VLLM_SOURCE_CODE_LOC=./ bash .buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
+    ```
+
+And the results will be inside `./benchmarks/results`.

.buildkite/nightly-benchmarks/nightly-descriptions.md

@@ -0,0 +1,39 @@
+
+# Nightly benchmark
+
+This benchmark aims to:
+
+- Provide performance clarity: Provide clarity on which one (vllm, tensorrt-llm, lmdeploy and SGLang) leads in performance in what workload.
+- Be reproducible: one can run the exact same set of benchmarking commands inside the exact same docker by following reproducing instructions.
+
+Latest results: [results link](https://blog.vllm.ai/2024/09/05/perf-update.html), scroll to the end.
+
+Latest reproduction guide: [github issue link](https://github.com/vllm-project/vllm/issues/8176)
+
+## Setup
+
+- Docker images:
+    - vLLM: `vllm/vllm-openai:v0.6.2`
+    - SGLang: `lmsysorg/sglang:v0.3.2-cu121`
+    - LMDeploy: `openmmlab/lmdeploy:v0.6.1-cu12`
+    - TensorRT-LLM: `nvcr.io/nvidia/tritonserver:24.07-trtllm-python-py3`
+        - *NOTE: we use r24.07 as the current implementation only works for this version. We are going to bump this up.*
+    - Check [nightly-pipeline.yaml](nightly-pipeline.yaml) for the concrete docker images, specs and commands we use for the benchmark.
+- Hardware
+    - 8x Nvidia A100 GPUs
+- Workload:
+    - Dataset
+        - ShareGPT dataset
+        - Prefill-heavy dataset (in average 462 input tokens, 16 tokens as output)
+        - Decode-heavy dataset (in average 462 input tokens, 256 output tokens)
+        - Check [nightly-tests.json](tests/nightly-tests.json) for the concrete configuration of datasets we use.
+    - Models: llama-3 8B, llama-3 70B.
+        - We do not use llama 3.1 as it is incompatible with trt-llm r24.07. ([issue](https://github.com/NVIDIA/TensorRT-LLM/issues/2105)).
+    - Average QPS (query per second): 2, 4, 8, 16, 32 and inf.
+        - Queries are randomly sampled, and arrival patterns are determined via Poisson process, but all with fixed random seed.
+    - Evaluation metrics: Throughput (higher the better), TTFT (time to the first token, lower the better), ITL (inter-token latency, lower the better).
+
+## Known issues
+
+- TRT-LLM crashes with Llama 3.1 8B [issue](https://github.com/NVIDIA/TensorRT-LLM/issues/2105).
+- TGI does not support `ignore-eos` flag.

.buildkite/nightly-benchmarks/nightly-pipeline.yaml

@@ -0,0 +1,196 @@
+common_pod_spec: &common_pod_spec
+  priorityClassName: perf-benchmark
+  nodeSelector:
+    nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
+  volumes:
+    - name: devshm
+      emptyDir:
+        medium: Memory
+    - name: hf-cache
+      hostPath:
+        path: /root/.cache/huggingface
+        type: Directory
+
+common_container_settings: &common_container_settings
+  command:
+    - bash .buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh
+  resources:
+    limits:
+      nvidia.com/gpu: 8
+  volumeMounts:
+    - name: devshm
+      mountPath: /dev/shm
+    - name: hf-cache
+      mountPath: /root/.cache/huggingface
+  env:
+    - name: VLLM_USAGE_SOURCE
+      value: ci-test
+    - name: HF_HOME
+      value: /root/.cache/huggingface
+    - name: VLLM_SOURCE_CODE_LOC
+      value: /workspace/build/buildkite/vllm/performance-benchmark
+    - name: HF_TOKEN
+      valueFrom:
+        secretKeyRef:
+          name: hf-token-secret
+          key: token
+
+steps:
+  - block: ":rocket: Ready for comparing vllm against alternatives? This will take 4 hours."
+
+
+
+  - label: "A100 vllm step 10"
+    priority: 100
+    agents:
+      queue: A100
+    plugins:
+      - kubernetes:
+          podSpec:
+            <<: *common_pod_spec
+            containers:
+              - image: vllm/vllm-openai:v0.6.2
+                <<: *common_container_settings
+
+
+
+  - label: "A100 sglang benchmark"
+    priority: 100
+    agents:
+      queue: A100
+    plugins:
+      - kubernetes:
+          podSpec:
+            <<: *common_pod_spec
+            containers:
+              - image: lmsysorg/sglang:v0.3.2-cu121
+                <<: *common_container_settings
+
+  - label: "A100 lmdeploy benchmark"
+    priority: 100
+    agents:
+      queue: A100
+    plugins:
+      - kubernetes:
+          podSpec:
+            <<: *common_pod_spec
+            containers:
+              - image: openmmlab/lmdeploy:v0.6.1-cu12
+                <<: *common_container_settings
+
+
+
+
+  - label: "A100 trt llama-8B"
+    priority: 100
+    agents:
+      queue: A100
+    plugins:
+      - kubernetes:
+          podSpec:
+            <<: *common_pod_spec
+            containers:
+              - image: nvcr.io/nvidia/tritonserver:24.07-trtllm-python-py3
+                <<: *common_container_settings
+                env:
+                  - name: VLLM_USAGE_SOURCE
+                    value: ci-test
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+                  - name: VLLM_SOURCE_CODE_LOC
+                    value: /workspace/build/buildkite/vllm/performance-benchmark
+                  - name: HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+                  - name: TEST_SELECTOR
+                    value: "llama8B"
+
+
+  - label: "A100 trt llama-70B"
+    priority: 100
+    agents:
+      queue: A100
+    plugins:
+      - kubernetes:
+          podSpec:
+            <<: *common_pod_spec
+            containers:
+              - image: nvcr.io/nvidia/tritonserver:24.07-trtllm-python-py3
+                <<: *common_container_settings
+                env:
+                  - name: VLLM_USAGE_SOURCE
+                    value: ci-test
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+                  - name: VLLM_SOURCE_CODE_LOC
+                    value: /workspace/build/buildkite/vllm/performance-benchmark
+                  - name: HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+                  - name: TEST_SELECTOR
+                    value: "llama70B"
+
+
+  # FIXME(Kuntai): uncomment this after NVIDIA gives us their test docker image 
+  # - label: "A100 trt benchmark"
+  #   priority: 100
+  #   agents:
+  #     queue: A100
+  #   plugins:
+  #     - kubernetes:
+  #         podSpec:
+  #           <<: *common_pod_spec
+  #           containers:
+  #             - image: nvcr.io/nvidia/tritonserver:24.07-trtllm-python-py3
+  #               <<: *common_container_settings
+
+
+  # FIXME(Kuntai): uncomment this after TGI supports `--ignore-eos`.
+  # - label: "A100 tgi benchmark"
+  #   priority: 100
+  #   agents:
+  #     queue: A100
+  #   plugins:
+  #     - kubernetes:
+  #         podSpec:
+  #           <<: *common_pod_spec
+  #           containers:
+  #             - image: ghcr.io/huggingface/text-generation-inference:2.2.0
+  #               <<: *common_container_settings
+        
+  - wait
+
+  - label: "Collect the results"
+    priority: 100
+    agents:
+      queue: A100
+    plugins:
+      - kubernetes:
+          podSpec:
+            <<: *common_pod_spec
+            containers:
+            - image: vllm/vllm-openai:v0.5.0.post1
+              command:
+              - bash .buildkite/nightly-benchmarks/scripts/nightly-annotate.sh
+              resources:
+                limits:
+                  nvidia.com/gpu: 8
+              volumeMounts:
+              - name: devshm
+                mountPath: /dev/shm
+              env:
+              - name: VLLM_USAGE_SOURCE
+                value: ci-test
+              - name: VLLM_SOURCE_CODE_LOC
+                value: /workspace/build/buildkite/vllm/performance-benchmark
+              - name: HF_TOKEN
+                valueFrom:
+                  secretKeyRef:
+                    name: hf-token-secret
+                    key: token
+
+  - block: ":rocket: check the results!"

.buildkite/nightly-benchmarks/performance-benchmarks-descriptions.md

@@ -5,7 +5,7 @@
 - Input length: 32 tokens.
 - Output length: 128 tokens.
 - Batch size: fixed (8).
-- GPU/HPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- GPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
 - CPU Models: llama-3.1 8B.
 - Evaluation metrics: end-to-end latency (mean, median, p99).
 
@@ -16,7 +16,7 @@
 - Input length: randomly sample 200 prompts from ShareGPT dataset (with fixed random seed).
 - Output length: the corresponding output length of these 200 prompts.
 - Batch size: dynamically determined by vllm to achieve maximum throughput.
-- GPU/HPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- GPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
 - CPU Models: llama-3.1 8B.
 - Evaluation metrics: throughput.
 
@@ -28,7 +28,7 @@
 - Output length: the corresponding output length of these 200 prompts.
 - Batch size: dynamically determined by vllm and the arrival pattern of the requests.
 - **Average QPS (query per second)**: 1, 4, 16 and inf. QPS = inf means all requests come at once. For other QPS values, the arrival time of each query is determined using a random Poisson process (with fixed random seed).
-- GPU/HPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
+- GPU Models: llama-3.1 8B, llama-3 70B, mixtral 8x7B.
 - We also added a speculative decoding test for llama-3 70B on GPU, under QPS 2
 - CPU Models: llama-3.1 8B.
 - Evaluation metrics: throughput, TTFT (time to the first token, with mean, median and p99), ITL (inter-token latency, with mean, median and p99).

.buildkite/nightly-benchmarks/scripts/convert-results-json-to-markdown.py

@@ -392,7 +392,7 @@ if __name__ == "__main__":
     json_file = "benchmark_results.json"
     with open(results_folder / md_file, "w") as f:
         results = read_markdown(
-            "../.buildkite/performance-benchmarks/"
+            "../.buildkite/nightly-benchmarks/"
             + "performance-benchmarks-descriptions.md"
         )
         results = results.format(

.buildkite/nightly-benchmarks/scripts/download-tokenizer.py

@@ -0,0 +1,26 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+
+from transformers import AutoTokenizer
+
+
+def main(model, cachedir):
+    # Load the tokenizer and save it to the specified directory
+    tokenizer = AutoTokenizer.from_pretrained(model)
+    tokenizer.save_pretrained(cachedir)
+    print(f"Tokenizer saved to {cachedir}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Download and save Hugging Face tokenizer"
+    )
+    parser.add_argument("--model", type=str, required=True, help="Name of the model")
+    parser.add_argument(
+        "--cachedir", type=str, required=True, help="Directory to save the tokenizer"
+    )
+
+    args = parser.parse_args()
+    main(args.model, args.cachedir)

.buildkite/nightly-benchmarks/scripts/generate-nightly-markdown.py

@@ -0,0 +1,97 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+import json
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from tabulate import tabulate
+
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(
+        description="Parse command line arguments for summary-nightly-results script."
+    )
+    parser.add_argument(
+        "--results-folder",
+        type=str,
+        required=True,
+        help="The folder where the results are stored.",
+    )
+    parser.add_argument(
+        "--description", type=str, required=True, help="Description of the results."
+    )
+
+    args = parser.parse_args()
+    return args
+
+
+def get_perf(df, method, model, metric):
+    means = []
+
+    for qps in [2, 4, 8, 16, "inf"]:
+        target = df["Test name"].str.contains(model)
+        target = target & df["Engine"].str.contains(method)
+        target = target & df["Test name"].str.contains("qps_" + str(qps))
+        filtered_df = df[target]
+
+        if filtered_df.empty:
+            means.append(0.0)
+        else:
+            means.append(filtered_df[metric].values[0])
+
+    return np.array(means)
+
+
+def get_perf_w_std(df, method, model, metric):
+    if metric in ["TTFT", "ITL"]:
+        mean = get_perf(df, method, model, "Mean " + metric + " (ms)")
+        mean = mean.tolist()
+        std = get_perf(df, method, model, "Std " + metric + " (ms)")
+        if std.mean() == 0:
+            std = None
+        success = get_perf(df, method, model, "Successful req.")
+        if std is not None:
+            std = std / np.sqrt(success)
+            std = std.tolist()
+
+    else:
+        assert metric == "Tput"
+        mean = get_perf(df, method, model, "Input Tput (tok/s)") + get_perf(
+            df, method, model, "Output Tput (tok/s)"
+        )
+        mean = mean.tolist()
+        std = None
+
+    return mean, std
+
+
+def main(args):
+    results_folder = Path(args.results_folder)
+
+    results = []
+
+    # collect results
+    for test_file in results_folder.glob("*_nightly_results.json"):
+        with open(test_file) as f:
+            results = results + json.loads(f.read())
+
+    # generate markdown table
+    df = pd.DataFrame.from_dict(results)
+
+    md_table = tabulate(df, headers="keys", tablefmt="pipe", showindex=False)
+
+    with open(args.description) as f:
+        description = f.read()
+
+    description = description.format(nightly_results_benchmarking_table=md_table)
+
+    with open("nightly_results.md", "w") as f:
+        f.write(description)
+
+
+if __name__ == "__main__":
+    args = parse_arguments()
+    main(args)

.buildkite/nightly-benchmarks/scripts/get-lmdeploy-modelname.py

@@ -0,0 +1,9 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from lmdeploy.serve.openai.api_client import APIClient
+
+api_client = APIClient("http://localhost:8000")
+model_name = api_client.available_models[0]
+
+print(model_name)

.buildkite/nightly-benchmarks/scripts/nightly-annotate.sh

@@ -0,0 +1,78 @@
+#!/bin/bash
+
+set -ex
+set -o pipefail
+
+
+main() {
+
+    (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
+    (which jq) || (apt-get update && apt-get -y install jq)
+    (which zip) || (apt-get install -y zip)
+
+    if [ ! -f /workspace/buildkite-agent ]; then
+        echo "buildkite-agent binary not found. Skip plotting the results."
+        exit 0
+    fi
+
+    # initial annotation
+    #description="$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/nightly-descriptions.md"
+
+    # download results
+    cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
+    mkdir -p results/
+    /workspace/buildkite-agent artifact download 'results/*nightly_results.json' results/
+    ls
+    ls results/
+
+    # upload benchmark results
+    zip -r results.zip results/
+    /workspace/buildkite-agent artifact upload "results.zip"
+
+    # upload benchmarking scripts
+    cd "$VLLM_SOURCE_CODE_LOC/"
+    zip -r nightly-benchmarks.zip .buildkite/ benchmarks/
+    /workspace/buildkite-agent artifact upload "nightly-benchmarks.zip"
+
+    cd "$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/"
+    # upload benchmarking pipeline
+    /workspace/buildkite-agent artifact upload "nightly-pipeline.yaml"
+
+    cd "$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/"
+    /workspace/buildkite-agent annotate --style "success" --context "nightly-benchmarks-results" --append < nightly-annotation.md
+    
+
+
+    # The figures should be generated by a separate process outside the CI/CD pipeline
+
+    # # generate figures
+    # python3 -m pip install tabulate pandas matplotlib
+
+    # python3 $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/generate-nightly-markdown.py \
+    #     --description $description \
+    #     --results-folder results/ 
+
+
+    # python3 $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/plot-nightly-results.py \
+    #     --description $description \
+    #     --results-folder results/ \
+    #     --dataset sharegpt
+
+    # python3 $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/plot-nightly-results.py \
+    #     --description $description \
+    #     --results-folder results/ \
+    #     --dataset sonnet_2048_128
+
+    # python3 $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/plot-nightly-results.py \
+    #     --description $description \
+    #     --results-folder results/ \
+    #     --dataset sonnet_128_2048
+    
+    # # upload results and figures
+    # /workspace/buildkite-agent artifact upload "nightly_results*.png"
+    # /workspace/buildkite-agent artifact upload $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/nightly-pipeline.yaml
+    # /workspace/buildkite-agent artifact upload $VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/tests/nightly-tests.json
+    # /workspace/buildkite-agent annotate --style "success" --context "nightly-benchmarks-results" --append < nightly_results.md
+}
+
+main "$@"

.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh

@@ -0,0 +1,464 @@
+#!/bin/bash
+
+set -o pipefail
+set -x
+
+check_gpus() {
+  # check the number of GPUs and GPU type.
+  declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
+  if [[ $gpu_count -gt 0 ]]; then
+    echo "GPU found."
+  else
+    echo "Need at least 1 GPU to run benchmarking."
+    exit 1
+  fi
+  declare -g gpu_type="$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')"
+  echo "GPU type is $gpu_type"
+}
+
+check_hf_token() {
+  # check if HF_TOKEN is available and valid
+  if [[ -z "$HF_TOKEN" ]]; then
+    echo "Error: HF_TOKEN is not set."
+    exit 1
+  elif [[ ! "$HF_TOKEN" =~ ^hf_ ]]; then
+    echo "Error: HF_TOKEN does not start with 'hf_'."
+    exit 1
+  else
+    echo "HF_TOKEN is set and valid."
+  fi
+}
+
+
+upload_to_buildkite() {
+  # upload the benchmarking results to buildkite
+
+  # if the agent binary is not found, skip uploading the results, exit 0
+  if [ ! -f /workspace/buildkite-agent ]; then
+    echo "buildkite-agent binary not found. Skip uploading the results."
+    return 0
+  fi
+  # /workspace/buildkite-agent annotate --style "success" --context "benchmark-results" --append < $RESULTS_FOLDER/${CURRENT_LLM_SERVING_ENGINE}_nightly_results.md
+  /workspace/buildkite-agent artifact upload "$RESULTS_FOLDER/*"
+}
+
+
+get_current_llm_serving_engine() {
+
+  if which lmdeploy >/dev/null; then
+    echo "Container: lmdeploy"
+    export CURRENT_LLM_SERVING_ENGINE=lmdeploy
+    return
+  fi
+
+  if [ -e /tgi-entrypoint.sh ]; then
+    echo "Container: tgi"
+    export CURRENT_LLM_SERVING_ENGINE=tgi
+    return
+  fi
+
+  if which trtllm-build >/dev/null; then
+    echo "Container: tensorrt-llm"
+    export CURRENT_LLM_SERVING_ENGINE=trt
+    return
+  fi
+
+  if [ -e /sgl-workspace ]; then
+    echo "Container: sglang"
+    export CURRENT_LLM_SERVING_ENGINE=sglang
+    return
+  fi
+
+  if [ -e /vllm-workspace ]; then
+    echo "Container: vllm"
+    # move to a completely irrelevant directory, to avoid import vllm from current folder
+    export CURRENT_LLM_SERVING_ENGINE=vllm
+
+    return
+  fi
+}
+
+json2args() {
+  # transforms the JSON string to command line args, and '_' is replaced to '-'
+  # example:
+  # input: { "model": "meta-llama/Llama-2-7b-chat-hf", "tensor_parallel_size": 1 }
+  # output: --model meta-llama/Llama-2-7b-chat-hf --tensor-parallel-size 1
+  local json_string=$1
+  local args=$(
+    echo "$json_string" | jq -r '
+      to_entries |
+      map("--" + (.key | gsub("_"; "-")) + " " + (.value | tostring)) |
+      join(" ")
+    '
+  )
+  echo "$args"
+}
+
+kill_gpu_processes() {
+  pkill -f '[p]ython'
+  pkill -f '[p]ython3'
+  pkill -f '[t]ritonserver'
+  pkill -f '[p]t_main_thread'
+  pkill -f '[t]ext-generation'
+  pkill -f '[l]mdeploy'
+  # vLLM now names the process with VLLM prefix after https://github.com/vllm-project/vllm/pull/21445
+  pkill -f '[V]LLM'
+
+  while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
+    sleep 1
+  done
+}
+
+wait_for_server() {
+  # wait for vllm server to start
+  # return 1 if vllm server crashes
+  timeout 1200 bash -c '
+    until curl -s localhost:8000/v1/completions > /dev/null; do
+      sleep 1
+    done' && return 0 || return 1
+}
+
+ensure_installed() {
+  # Ensure that the given command is installed by apt-get
+  local cmd=$1
+  if ! which "$cmd" >/dev/null; then
+    apt-get update && apt-get install -y "$cmd"
+  fi
+}
+
+run_serving_tests() {
+  # run serving tests using `vllm bench serve` command
+  # $1: a json file specifying serving test cases
+
+  local serving_test_file
+  serving_test_file=$1
+
+  # Iterate over serving tests
+  jq -c '.[]' "$serving_test_file" | while read -r params; do
+    # get the test name, and append the GPU type back to it.
+    test_name=$(echo "$params" | jq -r '.test_name')
+
+    # if TEST_SELECTOR is set, only run the test cases that match the selector
+    if [[ -n "$TEST_SELECTOR" ]] && [[ ! "$test_name" =~ $TEST_SELECTOR ]]; then
+      echo "Skip test case $test_name."
+      continue
+    fi
+
+    # prepend the current serving engine to the test name
+    test_name=${CURRENT_LLM_SERVING_ENGINE}_${test_name}
+
+    # get common parameters
+    common_params=$(echo "$params" | jq -r '.common_parameters')
+    model=$(echo "$common_params" | jq -r '.model')
+    tp=$(echo "$common_params" | jq -r '.tp')
+    dataset_name=$(echo "$common_params" | jq -r '.dataset_name')
+    dataset_path=$(echo "$common_params" | jq -r '.dataset_path')
+    port=$(echo "$common_params" | jq -r '.port')
+    num_prompts=$(echo "$common_params" | jq -r '.num_prompts')
+    reuse_server=$(echo "$common_params" | jq -r '.reuse_server')
+
+    # get client and server arguments
+    server_params=$(echo "$params" | jq -r ".${CURRENT_LLM_SERVING_ENGINE}_server_parameters")
+    client_params=$(echo "$params" | jq -r ".${CURRENT_LLM_SERVING_ENGINE}_client_parameters")
+    client_args=$(json2args "$client_params")
+    qps_list=$(echo "$params" | jq -r '.qps_list')
+    qps_list=$(echo "$qps_list" | jq -r '.[] | @sh')
+    echo "Running over qps list $qps_list"
+
+    # check if there is enough GPU to run the test
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required num-shard $tp but only $gpu_count GPU found. Skip testcase $test_name."
+      continue
+    fi
+
+    if [[ $reuse_server == "true" ]]; then
+      echo "Reuse previous server for test case $test_name"
+    else
+      kill_gpu_processes
+      bash "$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/launch-server.sh" \
+        "$server_params" "$common_params"
+    fi
+
+    if wait_for_server; then
+      echo ""
+      echo "$CURRENT_LLM_SERVING_ENGINE server is up and running."
+    else
+      echo ""
+      echo "$CURRENT_LLM_SERVING_ENGINE failed to start within the timeout period."
+      break
+    fi
+
+    # prepare tokenizer
+    # this is required for lmdeploy.
+    cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
+    rm -rf /tokenizer_cache
+    mkdir /tokenizer_cache
+    python3 ../.buildkite/nightly-benchmarks/scripts/download-tokenizer.py \
+      --model "$model" \
+      --cachedir /tokenizer_cache
+    cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
+
+
+    # change model name for lmdeploy (it will not follow standard hf name)
+    if [[ "$CURRENT_LLM_SERVING_ENGINE" == "lmdeploy" ]]; then
+      model=$(python ../.buildkite/nightly-benchmarks/scripts/get-lmdeploy-modelname.py)
+    fi
+
+    # iterate over different QPS
+    for qps in $qps_list; do
+      # remove the surrounding single quote from qps
+      if [[ "$qps" == *"inf"* ]]; then
+        echo "qps was $qps"
+        qps="inf"
+        echo "now qps is $qps"
+      fi
+
+      new_test_name=$test_name"_qps_"$qps
+
+      backend=$CURRENT_LLM_SERVING_ENGINE
+
+      if [[ $backend = "trt" ]]; then
+        backend="tensorrt-llm"
+      fi
+
+      if [[ "$backend" == *"vllm"* ]]; then
+        backend="vllm"
+      fi
+
+      if [[ "$dataset_name" = "sharegpt" ]]; then
+
+        client_command="vllm bench serve \
+          --backend $backend \
+          --tokenizer /tokenizer_cache \
+          --model $model \
+          --dataset-name $dataset_name \
+          --dataset-path $dataset_path \
+          --num-prompts $num_prompts \
+          --port $port \
+          --save-result \
+          --result-dir $RESULTS_FOLDER \
+          --result-filename ${new_test_name}.json \
+          --request-rate $qps \
+          --ignore-eos \
+          $client_args"
+
+      elif [[ "$dataset_name" = "sonnet" ]]; then
+
+        sonnet_input_len=$(echo "$common_params" | jq -r '.sonnet_input_len')
+        sonnet_output_len=$(echo "$common_params" | jq -r '.sonnet_output_len')
+        sonnet_prefix_len=$(echo "$common_params" | jq -r '.sonnet_prefix_len')
+
+        client_command="vllm bench serve \
+          --backend $backend \
+          --tokenizer /tokenizer_cache \
+          --model $model \
+          --dataset-name $dataset_name \
+          --dataset-path $dataset_path \
+          --num-prompts $num_prompts \
+          --sonnet-input-len $sonnet_input_len \
+          --sonnet-output-len $sonnet_output_len \
+          --sonnet-prefix-len $sonnet_prefix_len \
+          --port $port \
+          --save-result \
+          --result-dir $RESULTS_FOLDER \
+          --result-filename ${new_test_name}.json \
+          --request-rate $qps \
+          --ignore-eos \
+          $client_args"
+
+      else
+
+        echo "The dataset name must be either 'sharegpt' or 'sonnet'. Got $dataset_name."
+        exit 1
+
+      fi
+
+
+
+      echo "Running test case $test_name with qps $qps"
+      echo "Client command: $client_command"
+
+      eval "$client_command"
+
+      server_command="None"
+
+      # record the benchmarking commands
+      jq_output=$(jq -n \
+        --arg server "$server_command" \
+        --arg client "$client_command" \
+        --arg gpu "$gpu_type" \
+        --arg engine "$CURRENT_LLM_SERVING_ENGINE" \
+        '{
+          server_command: $server,
+          client_command: $client,
+          gpu_type: $gpu,
+          engine: $engine
+        }')
+      echo "$jq_output" >"$RESULTS_FOLDER/${new_test_name}.commands"
+
+    done
+
+  done
+
+  kill_gpu_processes
+}
+
+run_genai_perf_tests() {
+  # run genai-perf tests
+
+  # $1: a json file specifying genai-perf test cases
+  local genai_perf_test_file
+  genai_perf_test_file=$1
+
+  # Iterate over genai-perf tests
+  jq -c '.[]' "$genai_perf_test_file" | while read -r params; do
+    # get the test name, and append the GPU type back to it.
+    test_name=$(echo "$params" | jq -r '.test_name')
+
+    # if TEST_SELECTOR is set, only run the test cases that match the selector
+    if [[ -n "$TEST_SELECTOR" ]] && [[ ! "$test_name" =~ $TEST_SELECTOR ]]; then
+      echo "Skip test case $test_name."
+      continue
+    fi
+
+    # prepend the current serving engine to the test name
+    test_name=${CURRENT_LLM_SERVING_ENGINE}_${test_name}
+
+    # get common parameters
+    common_params=$(echo "$params" | jq -r '.common_parameters')
+    model=$(echo "$common_params" | jq -r '.model')
+    tp=$(echo "$common_params" | jq -r '.tp')
+    dataset_name=$(echo "$common_params" | jq -r '.dataset_name')
+    dataset_path=$(echo "$common_params" | jq -r '.dataset_path')
+    port=$(echo "$common_params" | jq -r '.port')
+    num_prompts=$(echo "$common_params" | jq -r '.num_prompts')
+    reuse_server=$(echo "$common_params" | jq -r '.reuse_server')
+
+    # get client and server arguments
+    server_params=$(echo "$params" | jq -r ".${CURRENT_LLM_SERVING_ENGINE}_server_parameters")
+    qps_list=$(echo "$params" | jq -r '.qps_list')
+    qps_list=$(echo "$qps_list" | jq -r '.[] | @sh')
+    echo "Running over qps list $qps_list"
+
+    # check if there is enough GPU to run the test
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required num-shard $tp but only $gpu_count GPU found. Skip testcase $test_name."
+      continue
+    fi
+
+    if [[ $reuse_server == "true" ]]; then
+      echo "Reuse previous server for test case $test_name"
+    else
+      kill_gpu_processes
+      bash "$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/launch-server.sh" \
+        "$server_params" "$common_params"
+    fi
+
+    if wait_for_server; then
+      echo ""
+      echo "$CURRENT_LLM_SERVING_ENGINE server is up and running."
+    else
+      echo ""
+      echo "$CURRENT_LLM_SERVING_ENGINE failed to start within the timeout period."
+      break
+    fi
+
+    # iterate over different QPS
+    for qps in $qps_list; do
+      # remove the surrounding single quote from qps
+      if [[ "$qps" == *"inf"* ]]; then
+        echo "qps was $qps"
+        qps=$num_prompts
+        echo "now qps is $qps"
+      fi
+
+      new_test_name=$test_name"_qps_"$qps
+      backend=$CURRENT_LLM_SERVING_ENGINE
+
+      if [[ "$backend" == *"vllm"* ]]; then
+        backend="vllm"
+      fi
+      #TODO: add output dir.
+      client_command="genai-perf profile \
+        -m $model \
+        --service-kind openai \
+        --backend "$backend" \
+        --endpoint-type chat \
+        --streaming \
+        --url localhost:$port \
+        --request-rate $qps \
+        --num-prompts $num_prompts \
+      "
+
+    echo "Client command: $client_command"
+
+    eval "$client_command"
+
+    #TODO: process/record outputs
+    done
+  done
+
+  kill_gpu_processes
+
+}
+
+prepare_dataset() {
+
+  # download sharegpt dataset
+  cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
+  wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+
+  # duplicate sonnet by 4x, to allow benchmarking with input length 2048
+  cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
+  echo "" > sonnet_4x.txt
+  for _ in {1..4}
+  do
+    cat sonnet.txt >> sonnet_4x.txt
+  done
+
+}
+
+main() {
+
+  # check if the environment variable is successfully injected from yaml
+
+  check_gpus
+  check_hf_token
+  get_current_llm_serving_engine
+
+  pip install -U transformers
+
+  pip install -r requirements/dev.txt
+  which genai-perf
+
+  # check storage
+  df -h
+
+  ensure_installed wget
+  ensure_installed curl
+  ensure_installed jq
+  # genai-perf dependency
+  ensure_installed libb64-0d
+
+  prepare_dataset
+
+  cd "$VLLM_SOURCE_CODE_LOC/benchmarks"
+  declare -g RESULTS_FOLDER=results/
+  mkdir -p $RESULTS_FOLDER
+  BENCHMARK_ROOT="$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/"
+
+  # run the test
+  run_serving_tests "$BENCHMARK_ROOT/tests/nightly-tests.json"
+
+  # run genai-perf tests
+  run_genai_perf_tests "$BENCHMARK_ROOT/tests/genai-perf-tests.json"
+  mv artifacts/ $RESULTS_FOLDER/
+
+  # upload benchmark results to buildkite
+  python3 -m pip install tabulate pandas
+  python3 "$BENCHMARK_ROOT/scripts/summary-nightly-results.py"
+  upload_to_buildkite
+
+}
+
+main "$@"

.buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh

@@ -15,8 +15,6 @@ check_gpus() {
     declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
   elif command -v amd-smi; then
     declare -g gpu_count=$(amd-smi list | grep 'GPU' | wc -l)
-  elif command -v hl-smi; then
-    declare -g gpu_count=$(hl-smi --list | grep -i "Module ID" | wc -l)
   fi
 
   if [[ $gpu_count -gt 0 ]]; then
@@ -25,16 +23,10 @@ check_gpus() {
     echo "Need at least 1 GPU to run benchmarking."
     exit 1
   fi
-  
-  declare -g arch_suffix=''
-  
   if command -v nvidia-smi; then
     declare -g gpu_type=$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')
   elif command -v amd-smi; then
     declare -g gpu_type=$(amd-smi static -g 0 -a | grep 'MARKET_NAME' | awk '{print $2}')
-  elif command -v hl-smi; then
-    declare -g gpu_type=$(hl-smi -q | grep "Product Name" | head -n 1 | awk -F ':' '{print $2}' | sed 's/^ *//')
-    arch_suffix='-hpu'
   fi
   echo "GPU type is $gpu_type"
 }
@@ -146,10 +138,6 @@ kill_gpu_processes() {
     while [ "$(amd-smi metric -g 0 | grep 'USED_VRAM' | awk '{print $2}')" -ge 1000 ]; do
       sleep 1
     done
-  elif command -v hl-smi; then
-    while [ "$(hl-smi -q | grep "Used" | head -n 1 | awk '{print $3}')" -ge 1000 ]; do
-      sleep 1
-    done
   fi
 
   # remove vllm config file
@@ -463,7 +451,6 @@ main() {
      ARCH='-cpu'
   else
      check_gpus
-     ARCH="$arch_suffix"
   fi
   check_hf_token
 
@@ -482,7 +469,7 @@ main() {
   ensure_sharegpt_downloaded
   declare -g RESULTS_FOLDER=results/
   mkdir -p $RESULTS_FOLDER
-  QUICK_BENCHMARK_ROOT=../.buildkite/performance-benchmarks/
+  QUICK_BENCHMARK_ROOT=../.buildkite/nightly-benchmarks/
 
   # dump vllm info via vllm collect-env
   env_output=$(vllm collect-env)

.buildkite/nightly-benchmarks/scripts/summary-nightly-results.py

@@ -0,0 +1,82 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import datetime
+import json
+import os
+from pathlib import Path
+
+import pandas as pd
+from tabulate import tabulate
+
+results_folder = Path("results/")
+
+# serving results and the keys that will be printed into markdown
+serving_results = []
+serving_column_mapping = {
+    "test_name": "Test name",
+    "gpu_type": "GPU",
+    "completed": "Successful req.",
+    "request_throughput": "Tput (req/s)",
+    "mean_ttft_ms": "Mean TTFT (ms)",
+    "std_ttft_ms": "Std TTFT (ms)",
+    "median_ttft_ms": "Median TTFT (ms)",
+    "mean_itl_ms": "Mean ITL (ms)",
+    "std_itl_ms": "Std ITL (ms)",
+    "median_itl_ms": "Median ITL (ms)",
+    "mean_tpot_ms": "Mean TPOT (ms)",
+    "std_tpot_ms": "Std TPOT (ms)",
+    "median_tpot_ms": "Median TPOT (ms)",
+    "total_token_throughput": "Total Token Tput (tok/s)",
+    "output_throughput": "Output Tput (tok/s)",
+    "total_input_tokens": "Total input tokens",
+    "total_output_tokens": "Total output tokens",
+    "engine": "Engine",
+}
+
+if __name__ == "__main__":
+    # collect results
+    for test_file in results_folder.glob("*.json"):
+        with open(test_file) as f:
+            raw_result = json.loads(f.read())
+
+        # attach the benchmarking command to raw_result
+        with open(test_file.with_suffix(".commands")) as f:
+            command = json.loads(f.read())
+        raw_result.update(command)
+
+        # update the test name of this result
+        raw_result.update({"test_name": test_file.stem})
+
+        # add the result to raw_result
+        serving_results.append(raw_result)
+        continue
+
+    serving_results = pd.DataFrame.from_dict(serving_results)
+
+    if not serving_results.empty:
+        serving_results = serving_results[list(serving_column_mapping.keys())].rename(
+            columns=serving_column_mapping
+        )
+
+    serving_md_table_with_headers = tabulate(
+        serving_results, headers="keys", tablefmt="pipe", showindex=False
+    )
+    # remove the first line of header
+    serving_md_table_lines = serving_md_table_with_headers.split("\n")
+    serving_md_table_without_header = "\n".join(serving_md_table_lines[2:])
+
+    prefix = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
+    prefix = prefix + "_" + os.environ.get("CURRENT_LLM_SERVING_ENGINE")
+
+    # document benchmarking results in markdown
+    with open(results_folder / f"{prefix}_nightly_results.md", "w") as f:
+        # document results with header.
+        # for those who wants to reproduce our benchmark.
+        f.write(serving_md_table_with_headers)
+        f.write("\n")
+
+    # document benchmarking results in json
+    with open(results_folder / f"{prefix}_nightly_results.json", "w") as f:
+        results = serving_results.to_dict(orient="records")
+        f.write(json.dumps(results))

.buildkite/nightly-benchmarks/scripts/wait-for-image.sh

@@ -0,0 +1,23 @@
+#!/bin/sh
+TOKEN=$(curl -s -L "https://public.ecr.aws/token?service=public.ecr.aws&scope=repository:q9t5s3a7/vllm-ci-postmerge-repo:pull" | jq -r .token)
+if [[ "$BUILDKITE_BRANCH" == "main" ]]; then
+    URL="https://public.ecr.aws/v2/q9t5s3a7/vllm-ci-postmerge-repo/manifests/$BUILDKITE_COMMIT"
+else
+    URL="https://public.ecr.aws/v2/q9t5s3a7/vllm-ci-test-repo/manifests/$BUILDKITE_COMMIT"
+fi
+
+TIMEOUT_SECONDS=10
+
+retries=0
+while [ $retries -lt 1000 ]; do
+    if [ "$(curl -s --max-time "$TIMEOUT_SECONDS" -L -H "Authorization: Bearer $TOKEN" -o /dev/null -w "%{http_code}" "$URL")" -eq 200 ]; then
+        exit 0
+    fi
+
+    echo "Waiting for image to be available..."
+
+    retries=$((retries + 1))
+    sleep 5
+done
+
+exit 1

.buildkite/performance-benchmarks/tests/latency-tests-hpu.json

@@ -1,55 +0,0 @@
-[
-    {
-        "test_name": "latency_llama8B_tp1",
-        "environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 1,
-            "load_format": "dummy",
-            "num-iters-warmup": 5,
-            "num-iters": 15,
-            "max-model-len": 256,
-            "async-scheduling": ""
-        }
-    },
-    {
-        "test_name": "latency_llama70B_tp4",
-        "environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
-            "tensor_parallel_size": 4,
-            "load_format": "dummy",
-            "num-iters-warmup": 5,
-            "num-iters": 15,
-            "max-model-len": 256,
-            "async-scheduling": ""
-        }
-    },
-    {
-        "test_name": "latency_mixtral8x7B_tp2",
-        "environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "parameters": {
-            "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
-            "tensor_parallel_size": 2,
-            "load_format": "dummy",
-            "num-iters-warmup": 5,
-            "num-iters": 15,
-            "max-model-len": 256,
-            "async-scheduling": ""
-        }
-    }
-]

.buildkite/performance-benchmarks/tests/serving-tests-hpu.json

@@ -1,82 +0,0 @@
-[
-    {
-        "test_name": "serving_llama8B_tp1_sharegpt",
-        "qps_list": [1, 4, 16, "inf"],
-        "server_environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "server_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 1,
-            "swap_space": 16,
-            "disable_log_stats": "",
-            "load_format": "dummy",
-            "max-model-len": 2048,
-            "max-num-seqs": 256,
-            "async-scheduling": ""
-        },
-        "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama70B_tp4_sharegpt",
-        "qps_list": [1, 4, 16, "inf"],
-        "server_environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "server_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
-            "tensor_parallel_size": 4,
-            "swap_space": 16,
-            "disable_log_stats": "",
-            "load_format": "dummy",
-            "max-model-len": 2048,
-            "max-num-seqs": 256,
-            "async-scheduling": ""
-        },
-        "client_parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_mixtral8x7B_tp2_sharegpt",
-        "qps_list": [1, 4, 16, "inf"],
-        "server_environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "server_parameters": {
-            "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
-            "tensor_parallel_size": 2,
-            "swap_space": 16,
-            "disable_log_stats": "",
-            "load_format": "dummy",
-            "max-model-len": 2048,
-            "max-num-seqs": 256,
-            "async-scheduling": ""
-        },
-        "client_parameters": {
-            "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    }
-]

.buildkite/performance-benchmarks/tests/throughput-tests-hpu.json

@@ -1,61 +0,0 @@
-[
-    {
-        "test_name": "throughput_llama8B_tp1",
-        "environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
-            "tensor_parallel_size": 1,
-            "load_format": "dummy",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 1000,
-            "backend": "vllm",
-            "max-model-len": 2048,
-            "max-num-seqs": 512,
-            "async-scheduling": ""
-        }
-    },
-    {
-        "test_name": "throughput_llama70B_tp4",
-        "environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "parameters": {
-            "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
-            "tensor_parallel_size": 4,
-            "load_format": "dummy",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 1000,
-            "backend": "vllm",
-            "max-model-len": 2048,
-            "max-num-seqs": 512,
-            "async-scheduling": ""
-        }
-    },
-    {
-        "test_name": "throughput_mixtral8x7B_tp2",
-        "environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "parameters": {
-            "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
-            "tensor_parallel_size": 2,
-            "load_format": "dummy",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 1000,
-            "backend": "vllm",
-            "max-model-len": 2048,
-            "max-num-seqs": 512,
-            "async-scheduling": ""
-        }
-    }
-]

.buildkite/release-pipeline.yaml

@@ -116,6 +116,24 @@ steps:
     commands:
       - "bash .buildkite/scripts/annotate-release.sh"
 
+  - label: "Build and publish TPU release image"
+    depends_on: ~
+    if: build.env("NIGHTLY") == "1"
+    agents:
+      queue: tpu_queue_postmerge
+    commands:
+      - "yes | docker system prune -a"
+      - "git fetch --all"
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f docker/Dockerfile.tpu ."
+      - "docker push vllm/vllm-tpu:nightly"
+      - "docker push vllm/vllm-tpu:$BUILDKITE_COMMIT"
+    plugins:
+      - docker-login#v3.0.0:
+          username: vllmbot
+          password-env: DOCKERHUB_TOKEN
+    env:
+      DOCKER_BUILDKIT: "1"
+
   - input: "Provide Release version here"
     id: input-release-version
     fields:
@@ -132,7 +150,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_AMXBF16=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest"
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:

.buildkite/scripts/annotate-release.sh

@@ -2,23 +2,16 @@
 
 set -ex
 
-# Get release version, default to 1.0.0.dev for nightly/per-commit builds
-RELEASE_VERSION=$(buildkite-agent meta-data get release-version 2>/dev/null | sed 's/^v//')
-if [ -z "${RELEASE_VERSION}" ]; then
-  RELEASE_VERSION="1.0.0.dev"
+# Get release version and strip leading 'v' if present
+RELEASE_VERSION=$(buildkite-agent meta-data get release-version | sed 's/^v//')
+
+if [ -z "$RELEASE_VERSION" ]; then
+  echo "Error: RELEASE_VERSION is empty. 'release-version' metadata might not be set or is invalid."
+  exit 1
 fi
 
 buildkite-agent annotate --style 'info' --context 'release-workflow' << EOF
-To download the wheel (by commit):
-\`\`\`
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux2014_aarch64.whl .
-
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux1_x86_64.whl .
-\`\`\`
-
-To download the wheel (by version):
+To download the wheel:
 \`\`\`
 aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux1_x86_64.whl .
 aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux2014_aarch64.whl .

.buildkite/scripts/hardware_ci/run-amd-test.sh

@@ -59,7 +59,7 @@ while true; do
         fi
 done
 
-echo "--- Pulling container"
+echo "--- Pulling container" 
 image_name="rocm/vllm-ci:${BUILDKITE_COMMIT}"
 container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 docker pull "${image_name}"
@@ -78,13 +78,17 @@ HF_MOUNT="/root/.cache/huggingface"
 commands=$@
 echo "Commands:$commands"
 
-commands=${commands//"pytest -v -s basic_correctness/test_basic_correctness.py"/"pytest -v -s basic_correctness/test_basic_correctness.py"}
+if [[ $commands == *"pytest -v -s basic_correctness/test_basic_correctness.py"* ]]; then
+  commands=${commands//"pytest -v -s basic_correctness/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s basic_correctness/test_basic_correctness.py"}
+fi
 
 if [[ $commands == *"pytest -v -s models/test_registry.py"* ]]; then
   commands=${commands//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
 fi
 
-commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"pytest -v -s compile/test_basic_correctness.py"}
+if [[ $commands == *"pytest -v -s compile/test_basic_correctness.py"* ]]; then
+  commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s compile/test_basic_correctness.py"}
+fi
 
 if [[ $commands == *"pytest -v -s lora"* ]]; then
   commands=${commands//"pytest -v -s lora"/"VLLM_ROCM_CUSTOM_PAGED_ATTN=0 pytest -v -s lora"}
@@ -169,28 +173,19 @@ fi
 PARALLEL_JOB_COUNT=8
 MYPYTHONPATH=".."
 
-# Test that we're launching on the machine that has
-# proper access to GPUs
-render_gid=$(getent group render | cut -d: -f3)
-if [[ -z "$render_gid" ]]; then
-  echo "Error: 'render' group not found. This is required for GPU access." >&2
-  exit 1
-fi
-
-# check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs.
+# check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs. 
 if [[ $commands == *"--shard-id="* ]]; then
-  # assign job count as the number of shards used
-  commands=$(echo "$commands" | sed -E "s/--num-shards[[:blank:]]*=[[:blank:]]*[0-9]*/--num-shards=${PARALLEL_JOB_COUNT} /g" | sed 's/ \\ / /g')
+  # assign job count as the number of shards used   
+  commands=${commands//"--num-shards= "/"--num-shards=${PARALLEL_JOB_COUNT} "}
   for GPU in $(seq 0 $(($PARALLEL_JOB_COUNT-1))); do
     # assign shard-id for each shard
-    commands_gpu=$(echo "$commands" | sed -E "s/--shard-id[[:blank:]]*=[[:blank:]]*[0-9]*/--shard-id=${GPU} /g" | sed 's/ \\ / /g')
+    commands_gpu=${commands//"--shard-id= "/"--shard-id=${GPU} "}
     echo "Shard ${GPU} commands:$commands_gpu"
     echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
     docker run \
         --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
         --network=host \
         --shm-size=16gb \
-        --group-add "$render_gid" \
         --rm \
         -e HIP_VISIBLE_DEVICES="${GPU}" \
         -e HF_TOKEN \
@@ -222,8 +217,8 @@ else
           --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
           --network=host \
           --shm-size=16gb \
-          --group-add "$render_gid" \
           --rm \
+          -e HIP_VISIBLE_DEVICES=0 \
           -e HF_TOKEN \
           -e AWS_ACCESS_KEY_ID \
           -e AWS_SECRET_ACCESS_KEY \

.buildkite/scripts/hardware_ci/run-cpu-test.sh

@@ -49,7 +49,6 @@ function cpu_tests() {
   # Run kernel tests
   docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
-    pytest -x -v -s tests/kernels/attention/test_cpu_attn.py
     pytest -x -v -s tests/kernels/test_onednn.py"
 
   # Run basic model test
@@ -77,7 +76,7 @@ function cpu_tests() {
   # Run AWQ test
   # docker exec cpu-test-"$NUMA_NODE" bash -c "
   #   set -e
-  #   pytest -x -s -v \
+  #   VLLM_USE_V1=0 pytest -x -s -v \
   #   tests/quantization/test_ipex_quant.py"
 
   # Run multi-lora tests
@@ -117,4 +116,4 @@ function cpu_tests() {
 
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 2.5h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 2h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"

.buildkite/scripts/hardware_ci/run-xpu-test.sh

@@ -20,10 +20,7 @@ trap remove_docker_container EXIT
 
 # Run the image and test offline inference/tensor parallel
 docker run \
-    --device /dev/dri:/dev/dri \
-    --net=host \
-    --ipc=host \
-    --privileged \
+    --device /dev/dri \
     -v /dev/dri/by-path:/dev/dri/by-path \
     --entrypoint="" \
     -e "HF_TOKEN=${HF_TOKEN}" \
@@ -45,7 +42,7 @@ docker run \
     pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
     pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
     pytest -v -s v1/structured_output
-    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.py --ignore=v1/spec_decode/test_speculators_eagle3.py
-    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_shared_storage_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py
+    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.py
+    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_shared_storage_connector.py
     pytest -v -s v1/test_serial_utils.py
 '

.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh

@@ -1,62 +0,0 @@
-#!/usr/bin/env bash
-set -euxo pipefail
-
-# args: [THRESHOLD] [NUM_QUESTIONS] [START_PORT]
-THRESHOLD=${1:-0.25}
-NUM_Q=${2:-1319}
-PORT=${3:-8010}
-OUT_DIR=${OUT_DIR:-/tmp/vllm-scheduled}
-mkdir -p "${OUT_DIR}"
-
-wait_for_server() {
-  local port=$1
-  timeout 600 bash -c '
-    until curl -sf "http://127.0.0.1:'"$port"'/health" > /dev/null; do
-      sleep 1
-    done'
-}
-
-MODEL="deepseek-ai/DeepSeek-V2-lite"
-BACKENDS=("deepep_high_throughput" "deepep_low_latency")
-
-cleanup() {
-  if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
-    kill "${SERVER_PID}" 2>/dev/null || true
-    for _ in {1..20}; do
-      kill -0 "${SERVER_PID}" 2>/dev/null || break
-      sleep 0.5
-    done
-    kill -9 "${SERVER_PID}" 2>/dev/null || true
-  fi
-}
-trap cleanup EXIT
-
-for BACK in "${BACKENDS[@]}"; do
-  VLLM_DEEP_GEMM_WARMUP=skip \
-  VLLM_ALL2ALL_BACKEND=$BACK \
-  vllm serve "$MODEL" \
-    --enforce-eager \
-    --tensor-parallel-size 2 \
-    --data-parallel-size 2 \
-    --enable-expert-parallel \
-    --enable-eplb \
-    --trust-remote-code \
-    --max-model-len 2048 \
-    --port $PORT &
-  SERVER_PID=$!
-  wait_for_server $PORT
-
-  TAG=$(echo "$MODEL" | tr '/: \\n' '_____')
-  OUT="${OUT_DIR}/${TAG}_${BACK}.json"
-  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port $PORT --num-questions ${NUM_Q} --save-results ${OUT}
-  python3 - <<PY
-import json; acc=json.load(open('${OUT}'))['accuracy']
-print(f"${MODEL} ${BACK}: accuracy {acc:.3f}")
-assert acc >= ${THRESHOLD}, f"${MODEL} ${BACK} accuracy {acc}"
-PY
-
-  cleanup
-  SERVER_PID=
-  sleep 1
-  PORT=$((PORT+1))
-done

.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh

@@ -1,61 +0,0 @@
-#!/usr/bin/env bash
-set -euxo pipefail
-
-# args: [THRESHOLD] [NUM_QUESTIONS] [START_PORT]
-THRESHOLD=${1:-0.8}
-NUM_Q=${2:-1319}
-PORT=${3:-8020}
-OUT_DIR=${OUT_DIR:-/tmp/vllm-scheduled}
-mkdir -p "${OUT_DIR}"
-
-wait_for_server() {
-  local port=$1
-  timeout 600 bash -c '
-    until curl -sf "http://127.0.0.1:'"$port"'/health" > /dev/null; do
-      sleep 1
-    done'
-}
-
-MODEL="QWen/Qwen3-30B-A3B-FP8"
-BACKENDS=("deepep_high_throughput" "deepep_low_latency")
-
-cleanup() {
-  if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
-    kill "${SERVER_PID}" 2>/dev/null || true
-    for _ in {1..20}; do
-      kill -0 "${SERVER_PID}" 2>/dev/null || break
-      sleep 0.5
-    done
-    kill -9 "${SERVER_PID}" 2>/dev/null || true
-  fi
-}
-trap cleanup EXIT
-
-for BACK in "${BACKENDS[@]}"; do
-  VLLM_DEEP_GEMM_WARMUP=skip \
-  VLLM_ALL2ALL_BACKEND=$BACK \
-  vllm serve "$MODEL" \
-    --enforce-eager \
-    --tensor-parallel-size 2 \
-    --data-parallel-size 2 \
-    --enable-expert-parallel \
-    --trust-remote-code \
-    --max-model-len 2048 \
-    --port $PORT &
-  SERVER_PID=$!
-  wait_for_server $PORT
-
-  TAG=$(echo "$MODEL" | tr '/: \\n' '_____')
-  OUT="${OUT_DIR}/${TAG}_${BACK}.json"
-  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port $PORT --num-questions ${NUM_Q} --save-results ${OUT}
-  python3 - <<PY
-import json; acc=json.load(open('${OUT}'))['accuracy']
-print(f"${MODEL} ${BACK}: accuracy {acc:.3f}")
-assert acc >= ${THRESHOLD}, f"${MODEL} ${BACK} accuracy {acc}"
-PY
-
-  cleanup
-  SERVER_PID=
-  sleep 1
-  PORT=$((PORT+1))
-done

.buildkite/test-amd.yaml

@@ -38,7 +38,7 @@ steps:
 - label: Pytorch Nightly Dependency Override Check # 2min
   # if this test fails, it means the nightly torch version is not compatible with some
   # of the dependencies. Please check the error message and add the package to whitelist
-  # in /vllm/tools/pre_commit/generate_nightly_torch_test.py
+  # in /vllm/tools/generate_nightly_torch_test.py
   mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
   # grade: Blocking
@@ -48,8 +48,8 @@ steps:
   commands:
   - bash standalone_tests/pytorch_nightly_dependency.sh
 
-- label: Async Engine, Inputs, Utils, Worker Test # 10min
-  timeout_in_minutes: 15
+- label: Async Engine, Inputs, Utils, Worker Test # 36min
+  timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
@@ -226,27 +226,6 @@ steps:
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
   - popd
 
-- label: Distributed Tests (8 GPUs) # 4min
-  timeout_in_minutes: 10
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_8
-  # grade: Blocking
-  gpu: h100
-  num_gpus: 8
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - examples/offline_inference/torchrun_dp_example.py
-  - vllm/config/parallel.py
-  - vllm/distributed/
-  - vllm/v1/engine/llm_engine.py
-  - vllm/v1/executor/uniproc_executor.py
-  - vllm/v1/worker/gpu_worker.py
-  commands:
-  # https://github.com/NVIDIA/nccl/issues/1838
-  #- export NCCL_CUMEM_HOST_ENABLE=0
-  # test with torchrun tp=2 and dp=4 with ep
-  - torchrun --nproc-per-node=8 ../examples/offline_inference/torchrun_dp_example.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
-
 - label: EPLB Algorithm Test # 5min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
@@ -259,11 +238,11 @@ steps:
   commands:
   - pytest -v -s distributed/test_eplb_algo.py
 
-- label: EPLB Execution Test # 10min
+- label: EPLB Execution Test # 5min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
   # grade: Blocking
-  timeout_in_minutes: 20
+  timeout_in_minutes: 15
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -271,7 +250,6 @@ steps:
   - tests/distributed/test_eplb_execute.py
   commands:
   - pytest -v -s distributed/test_eplb_execute.py
-  - pytest -v -s distributed/test_eplb_spec_decode.py
 
 - label: Metrics, Tracing Test # 12min
   timeout_in_minutes: 20
@@ -295,7 +273,7 @@ steps:
 
 - label: Regression Test # 7min
   timeout_in_minutes: 20
-  mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   grade: Blocking
   source_file_dependencies:
@@ -308,9 +286,9 @@ steps:
 
 - label: Engine Test # 25min
   timeout_in_minutes: 40
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
-  # grade: Blocking
+  #grade: Blocking
   source_file_dependencies:
   - vllm/
   - tests/engine
@@ -340,7 +318,7 @@ steps:
 
 - label: V1 Test entrypoints # 35min
   timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
@@ -359,7 +337,6 @@ steps:
     - tests/v1
   commands:
     # split the test to avoid interference
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
     - pytest -v -s -m 'not cpu_test' v1/core
     - pytest -v -s v1/executor
     - pytest -v -s v1/kv_offload
@@ -371,25 +348,10 @@ steps:
     - pytest -v -s -m 'not cpu_test' v1/metrics
     - pytest -v -s v1/test_oracle.py
     - pytest -v -s v1/test_request.py
-    - pytest -v -s v1/test_outputs.py
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 
-# TODO: Add the "V1 Test attetion (MI300)" test group
-
-- label: V1 Test attention (H100) # 10min
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_1
-  # grade: Blocking
-  timeout_in_minutes: 30
-  gpu: h100
-  source_file_dependencies:
-    - vllm/v1/attention
-    - tests/v1/attention
-  commands:
-    - pytest -v -s v1/attention
-
 - label: V1 Test others (CPU) # 5 mins
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
@@ -479,7 +441,7 @@ steps:
       --ignore=lora/test_llm_with_multi_loras.py \
       --ignore=lora/test_olmoe_tp.py \
       --ignore=lora/test_deepseekv2_tp.py \
-      --ignore=lora/test_gptoss_tp.py \
+      --ignore=lora/test_gptoss.py \
       --ignore=lora/test_qwen3moe_tp.py
   parallelism: 4
 
@@ -516,11 +478,10 @@ steps:
   - tests/compile
   commands:
   - pytest -v -s compile/test_basic_correctness.py
-  - pytest -v -s compile/test_multimodal_compile.py
   - pytest -v -s compile/piecewise/
 
-- label: PyTorch Fullgraph Test # 27min
-  timeout_in_minutes: 40
+- label: PyTorch Fullgraph Test # 22min
+  timeout_in_minutes: 35
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
@@ -529,23 +490,8 @@ steps:
   - vllm/
   - tests/compile
   commands:
-  - pytest -v -s compile/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
-    # Limit to no custom ops to reduce running time
-    # Wrap with quotes to escape yaml and avoid starting -k string with a -
-  - "pytest -v -s compile/test_fusions_e2e.py -k 'TRITON and -quant_fp8'"
-
-- label: Cudagraph test
-  timeout_in_minutes: 20
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_1
-  source_file_dependencies:
-  - tests/v1/cudagraph
-  - vllm/v1/cudagraph_dispatcher.py
-  - vllm/config/compilation.py
-  - vllm/compilation
-  commands:
-    - pytest -v -s v1/cudagraph/test_cudagraph_dispatch.py
-    - pytest -v -s v1/cudagraph/test_cudagraph_mode.py
+  - pytest -v -s compile/test_full_graph.py
+  - pytest -v -s compile/test_fusions_e2e.py
 
 - label: Kernels Core Operation Test # 48min
   timeout_in_minutes: 75
@@ -597,8 +543,6 @@ steps:
   - tests/kernels/moe
   - vllm/model_executor/layers/fused_moe/
   - vllm/distributed/device_communicators/
-  - vllm/envs.py
-  - vllm/config
   commands:
     - pytest -v -s kernels/moe --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 2
@@ -617,13 +561,10 @@ steps:
 
 - label: Model Executor Test # 23min
   timeout_in_minutes: 35
-  torch_nightly: true
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
-  - vllm/engine/arg_utils.py
-  - vllm/config/model.py
   - vllm/model_executor
   - tests/model_executor
   - tests/entrypoints/openai/test_tensorizer_entrypoint.py
@@ -675,9 +616,9 @@ steps:
   - uv pip install --system torchao==0.13.0
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
-- label: LM Eval Small Models # 15min
-  timeout_in_minutes: 20
-  mirror_hardwares: [amdexperimental, amdproduction]
+- label: LM Eval Small Models # 53min
+  timeout_in_minutes: 75
+  mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
@@ -686,18 +627,17 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
 
-- label: OpenAI API correctness # 10min
-  timeout_in_minutes: 15
-  mirror_hardwares: [amdexperimental, amdproduction]
+- label: OpenAI API correctness # 22min
+  timeout_in_minutes: 30
+  mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
   - csrc/
   - vllm/entrypoints/openai/
   - vllm/model_executor/models/whisper.py
-  commands: # LMEval
-  # Transcription WER check is skipped because encoder-decoder models are not supported on ROCm, see https://github.com/vllm-project/vllm/issues/27442
-  - pytest -s entrypoints/openai/correctness/  --ignore entrypoints/openai/correctness/test_transcription_api_correctness.py
+  commands: # LMEval+Transcription WER check
+  - pytest -s entrypoints/openai/correctness/
 
 - label: OpenAI-Compatible Tool Use # 23 min
   timeout_in_minutes: 35
@@ -918,11 +858,10 @@ steps:
     - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
     - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
 
-- label: Multi-Modal Accuracy Eval (Small Models) # 10min
-  timeout_in_minutes: 70
-  mirror_hardwares: [amdexperimental, amdproduction]
+- label: Multi-Modal Accuracy Eval (Small Models) # 50min
+  mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
-  # grade: Blocking
+  timeout_in_minutes: 70
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
   - vllm/multimodal/
@@ -969,7 +908,7 @@ steps:
 
 - label: Quantized Models Test # 45 min
   timeout_in_minutes: 60
-  mirror_hardwares: [amdexperimental, amdproduction]
+  mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
@@ -993,7 +932,6 @@ steps:
 - label: Transformers Nightly Models Test
   mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
-  # grade: Blocking
   working_dir: "/vllm-workspace/"
   optional: true
   commands:
@@ -1021,16 +959,11 @@ steps:
   - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
-  - vllm/v1/attention/backends/mla/cutlass_mla.py
-  - vllm/v1/attention/backends/mla/flashinfer_mla.py
-  - vllm/platforms/cuda.py
-  - vllm/attention/selector.py
   commands:
     - nvidia-smi
     - python3 examples/offline_inference/basic/chat.py
     # Attention
     # num_heads2 broken by https://github.com/flashinfer-ai/flashinfer/issues/1353
-    - pytest -v -s tests/kernels/attention/test_attention_selector.py
     - pytest -v -s tests/kernels/attention/test_flashinfer.py -k 'not num_heads2'
     - pytest -v -s tests/kernels/attention/test_flashinfer_trtllm_attention.py
     - pytest -v -s tests/kernels/attention/test_cutlass_mla_decode.py
@@ -1067,39 +1000,12 @@ steps:
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
     # this runner has 2 GPUs available even though num_gpus=2 is not set
     - pytest -v -s tests/compile/test_fusion_all_reduce.py
-    # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
-    # Wrap with quotes to escape yaml
-    - "pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and Llama-3.1 and -quant_fp8 and -rms_norm'"
-
-- label: Blackwell Fusion E2E Tests # 30 min
-  timeout_in_minutes: 40
-  working_dir: "/vllm-workspace/"
-  gpu: b200
-  optional: true
-  num_gpus: 2
-  source_file_dependencies:
-  - csrc/quantization/fp4/
-  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-  - vllm/v1/attention/backends/flashinfer.py
-  - vllm/compilation/
-  # can affect pattern matching
-  - vllm/model_executor/layers/layernorm.py
-  - vllm/model_executor/layers/activation.py
-  - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/test_fusions_e2e.py
-  - tests/compile/test_full_graph.py
-  commands:
-    - nvidia-smi
-    # Run all e2e fusion tests
     - pytest -v -s tests/compile/test_fusions_e2e.py
-    # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
-    - pytest -v -s tests/compile/test_full_graph.py::test_fp8_kv_scale_compile
 
-- label: ROCm GPT-OSS Eval
+- label: Blackwell GPT-OSS Eval
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/"
-  agent_pool: mi325_1
-  mirror_hardwares: [amdproduction]
+  gpu: b200
   optional: true # run on nightlies
   source_file_dependencies:
   - tests/evals/gpt_oss
@@ -1108,7 +1014,7 @@ steps:
   - vllm/v1/attention/backends/flashinfer.py
   commands:
     - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - VLLM_ROCM_USE_AITER_MHA=0 VLLM_ROCM_USE_AITER=1 VLLM_USE_AITER_UNIFIED_ATTENTION=1 pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
+    - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
 
 - label: Blackwell Quantized MoE Test
   timeout_in_minutes: 60
@@ -1311,8 +1217,6 @@ steps:
     - pytest -v -s -x lora/test_llama_tp.py
     - pytest -v -s -x lora/test_llm_with_multi_loras.py
     - pytest -v -s -x lora/test_olmoe_tp.py
-    - pytest -v -s -x lora/test_gptoss_tp.py
-
 
 - label: Weight Loading Multiple GPU Test  # 33min
   timeout_in_minutes: 45
@@ -1345,7 +1249,6 @@ steps:
 - label: NixlConnector PD accuracy tests (Distributed) # 30min
   mirror_hardwares: [amdexperimental]
   agent_pool: mi325_4
-  # grade: Blocking
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -1360,9 +1263,6 @@ steps:
 ##### A100 test #####
 
 - label: Distributed Tests (A100) # optional
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_4
-  # grade: Blocking
   gpu: a100
   optional: true
   num_gpus: 4
@@ -1377,9 +1277,6 @@ steps:
   - pytest -v -s -x lora/test_mixtral.py
 
 - label: LM Eval Large Models # optional
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_4
-  # grade: Blocking
   gpu: a100
   optional: true
   num_gpus: 4
@@ -1391,27 +1288,8 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
-##### H100 test #####
-- label: LM Eval Large Models (H100) # optional
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_4
-  # grade: Blocking
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  commands:
-    - export VLLM_USE_DEEP_GEMM=0  # We found Triton is faster than DeepGEMM for H100
-    - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4
-
 ##### H200 test #####
 - label: Distributed Tests (H200) # optional
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_2
-  # grade: Blocking
   gpu: h200
   optional: true
   working_dir: "/vllm-workspace/"
@@ -1423,7 +1301,6 @@ steps:
     - pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
     - pytest -v -s tests/distributed/test_context_parallel.py
     - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048
-    - pytest -v -s tests/v1/distributed/test_dbo.py
 
 ##### B200 test #####
 - label: Distributed Tests (B200) # optional
@@ -1434,7 +1311,6 @@ steps:
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
     - pytest -v -s tests/distributed/test_nccl_symm_mem_allreduce.py
-    - pytest -v -s tests/v1/distributed/test_dbo.py
 
 ##### RL Integration Tests #####
 - label: Prime-RL Integration Test # 15min
@@ -1450,27 +1326,3 @@ steps:
   - .buildkite/scripts/run-prime-rl-test.sh
   commands:
     - bash .buildkite/scripts/run-prime-rl-test.sh
-
-- label: DeepSeek V2-Lite Accuracy
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_4
-  # grade: Blocking
-  timeout_in_minutes: 60
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace"
-  commands:
-  - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
-
-- label: Qwen3-30B-A3B-FP8-block Accuracy
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_4
-  # grade: Blocking
-  timeout_in_minutes: 60
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace"
-  commands:
-  - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh 0.8 200 8020

.buildkite/test-pipeline.yaml

@@ -25,7 +25,6 @@
 #     and $$BUILDKITE_PARALLEL_JOB_COUNT environment variables.
 # working_dir(str): specify the place where the command should execute, default to /vllm-workspace/tests
 # source_file_dependencies(list): the list of prefixes to opt-in the test for, if empty, the test will always run.
-# autorun_on_main (bool): default to false, if true, the test will run automatically when commit is pushed to main branch.
 
 # When adding a test
 # - If the test belongs to an existing group, add it there
@@ -39,7 +38,7 @@ steps:
 - label: Pytorch Nightly Dependency Override Check # 2min
   # if this test fails, it means the nightly torch version is not compatible with some
   # of the dependencies. Please check the error message and add the package to whitelist
-  # in /vllm/tools/pre_commit/generate_nightly_torch_test.py
+  # in /vllm/tools/generate_nightly_torch_test.py
   soft_fail: true
   source_file_dependencies:
   - requirements/nightly_torch_test.txt
@@ -57,7 +56,7 @@ steps:
   - pytest -v -s -m 'not cpu_test' multimodal
   - pytest -v -s utils_
 
-- label: Async Engine, Inputs, Utils, Worker, Config Test (CPU) # 4 mins
+- label: Async Engine, Inputs, Utils, Worker Test (CPU) # 4 mins
   timeout_in_minutes: 10
   source_file_dependencies:
   - vllm/
@@ -66,7 +65,6 @@ steps:
   - tests/multimodal
   - tests/standalone_tests/lazy_imports.py
   - tests/transformers_utils
-  - tests/config
   no_gpu: true
   commands:
   - python3 standalone_tests/lazy_imports.py
@@ -74,7 +72,6 @@ steps:
   - pytest -v -s test_outputs.py
   - pytest -v -s -m 'cpu_test' multimodal
   - pytest -v -s transformers_utils
-  - pytest -v -s config
 
 - label: Python-only Installation Test # 10min
   timeout_in_minutes: 20
@@ -208,24 +205,6 @@ steps:
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
   - popd
 
-- label: Distributed Tests (8 GPUs) # 4min
-  timeout_in_minutes: 10
-  gpu: h100
-  num_gpus: 8
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - examples/offline_inference/torchrun_dp_example.py
-  - vllm/config/parallel.py
-  - vllm/distributed/
-  - vllm/v1/engine/llm_engine.py
-  - vllm/v1/executor/uniproc_executor.py
-  - vllm/v1/worker/gpu_worker.py
-  commands:
-  # https://github.com/NVIDIA/nccl/issues/1838
-  - export NCCL_CUMEM_HOST_ENABLE=0
-  # test with torchrun tp=2 and dp=4 with ep
-  - torchrun --nproc-per-node=8 ../examples/offline_inference/torchrun_dp_example.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
-
 - label: EPLB Algorithm Test # 5min
   timeout_in_minutes: 15
   working_dir: "/vllm-workspace/tests"
@@ -235,8 +214,8 @@ steps:
   commands:
   - pytest -v -s distributed/test_eplb_algo.py
 
-- label: EPLB Execution Test # 10min
-  timeout_in_minutes: 20
+- label: EPLB Execution Test # 5min
+  timeout_in_minutes: 15
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -244,7 +223,6 @@ steps:
   - tests/distributed/test_eplb_execute.py
   commands:
   - pytest -v -s distributed/test_eplb_execute.py
-  - pytest -v -s distributed/test_eplb_spec_decode.py
 
 - label: Metrics, Tracing Test # 12min
   timeout_in_minutes: 20
@@ -319,7 +297,6 @@ steps:
     - vllm/
     - tests/v1
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
     # split the test to avoid interference
     - pytest -v -s -m 'not cpu_test' v1/core
     - pytest -v -s v1/executor
@@ -332,7 +309,6 @@ steps:
     - pytest -v -s -m 'not cpu_test' v1/metrics
     - pytest -v -s v1/test_oracle.py
     - pytest -v -s v1/test_request.py
-    - pytest -v -s v1/test_outputs.py
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
@@ -346,15 +322,6 @@ steps:
   commands:
     - pytest -v -s v1/attention
 
-- label: V1 Test attention (B200) # 10min
-  timeout_in_minutes: 30
-  gpu: b200
-  source_file_dependencies:
-    - vllm/v1/attention
-    - tests/v1/attention
-  commands:
-    - VLLM_DISABLE_FLASHINFER_PREFILL=1 pytest -v -s v1/attention # TODO: FI prefill is bugged and causes incorrectness, fix this
-
 - label: V1 Test others (CPU) # 5 mins
   source_file_dependencies:
     - vllm/
@@ -432,9 +399,9 @@ steps:
       --ignore=lora/test_llm_with_multi_loras.py \
       --ignore=lora/test_olmoe_tp.py \
       --ignore=lora/test_deepseekv2_tp.py \
-      --ignore=lora/test_gptoss_tp.py \
+      --ignore=lora/test_gptoss.py \
       --ignore=lora/test_qwen3moe_tp.py
-
+      
   parallelism: 4
 
 - label: PyTorch Compilation Unit Tests # 15min
@@ -445,8 +412,6 @@ steps:
     - vllm/
     - tests/compile
   commands:
-    - pytest -v -s compile/test_graph_partition.py
-    - pytest -v -s compile/test_config.py
     - pytest -v -s compile/test_pass_manager.py
     - pytest -v -s compile/test_fusion.py
     - pytest -v -s compile/test_fusion_attn.py
@@ -456,7 +421,6 @@ steps:
     - pytest -v -s compile/test_decorator.py
     - pytest -v -s compile/test_noop_elimination.py
     - pytest -v -s compile/test_aot_compile.py
-    - pytest -v -s compile/test_qk_norm_rope_fusion.py
 
 - label: PyTorch Fullgraph Smoke Test # 15min
   timeout_in_minutes: 30
@@ -467,22 +431,18 @@ steps:
   - tests/compile
   commands:
   - pytest -v -s compile/test_basic_correctness.py
-  - pytest -v -s compile/test_multimodal_compile.py
   - pytest -v -s compile/piecewise/
 
-- label: PyTorch Fullgraph Test # 27min
-  timeout_in_minutes: 40
+- label: PyTorch Fullgraph Test # 22min
+  timeout_in_minutes: 35
   mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
   - tests/compile
   commands:
-    # fp8 kv scales not supported on sm89, tested on Blackwell instead
-  - pytest -v -s compile/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
-    # Limit to no custom ops to reduce running time
-    # Wrap with quotes to escape yaml and avoid starting -k string with a -
-  - "pytest -v -s compile/test_fusions_e2e.py -k 'TRITON and not +quant_fp8 and not Llama-4'"
+  - pytest -v -s compile/test_full_graph.py
+  - pytest -v -s compile/test_fusions_e2e.py
 
 - label: Cudagraph test
   timeout_in_minutes: 20
@@ -538,8 +498,6 @@ steps:
   - tests/kernels/moe
   - vllm/model_executor/layers/fused_moe/
   - vllm/distributed/device_communicators/
-  - vllm/envs.py
-  - vllm/config
   commands:
     - pytest -v -s kernels/moe --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 2
@@ -556,11 +514,8 @@ steps:
 
 - label: Model Executor Test # 23min
   timeout_in_minutes: 35
-  torch_nightly: true
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
-  - vllm/engine/arg_utils.py
-  - vllm/config/model.py
   - vllm/model_executor
   - tests/model_executor
   - tests/entrypoints/openai/test_tensorizer_entrypoint.py
@@ -612,7 +567,6 @@ steps:
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
-  autorun_on_main: true
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
 
@@ -876,12 +830,12 @@ steps:
   optional: true
   commands:
     - pip install --upgrade git+https://github.com/huggingface/transformers
-    - pytest -v -s tests/models/test_initialization.py -k 'not (Gemma3 or ModernBert or Qwen2_5_VL or Qwen2_5vl or Qwen2VL or TransformersMultiModalEmbeddingModel or TransformersMultiModalForSequenceClassification or Ultravox or Phi4Multimodal or LlavaNextVideo or MiniCPMO or Lfm2Moe or PaliGemma or RobertaForSequenceClassification or Ovis2_5 or Fuyu or DeepseekOCR or KimiVL)'
+    - pytest -v -s tests/models/test_initialization.py
     - pytest -v -s tests/models/test_transformers.py
-    # - pytest -v -s tests/models/multimodal/processing/
-    - pytest -v -s tests/models/multimodal/test_mapping.py -k 'not (Gemma3 or Qwen2VL or Qwen2_5_VL)'
+    - pytest -v -s tests/models/multimodal/processing/
+    - pytest -v -s tests/models/multimodal/test_mapping.py
     - python3 examples/offline_inference/basic/chat.py
-    # - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
+    - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
     # Whisper needs spawn method to avoid deadlock
     - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
 
@@ -899,16 +853,11 @@ steps:
   - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
-  - vllm/v1/attention/backends/mla/cutlass_mla.py
-  - vllm/v1/attention/backends/mla/flashinfer_mla.py
-  - vllm/platforms/cuda.py
-  - vllm/attention/selector.py
   commands:
     - nvidia-smi
     - python3 examples/offline_inference/basic/chat.py
     # Attention
     # num_heads2 broken by https://github.com/flashinfer-ai/flashinfer/issues/1353
-    - pytest -v -s tests/kernels/attention/test_attention_selector.py
     - pytest -v -s tests/kernels/attention/test_flashinfer.py -k 'not num_heads2'
     - pytest -v -s tests/kernels/attention/test_flashinfer_trtllm_attention.py
     - pytest -v -s tests/kernels/attention/test_cutlass_mla_decode.py
@@ -926,7 +875,7 @@ steps:
     - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
 
-- label: Blackwell Fusion and Compile Tests # 30 min
+- label: Blackwell Fusion Tests # 30 min
   timeout_in_minutes: 40
   working_dir: "/vllm-workspace/"
   gpu: b200
@@ -945,32 +894,6 @@ steps:
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
     # this runner has 2 GPUs available even though num_gpus=2 is not set
     - pytest -v -s tests/compile/test_fusion_all_reduce.py
-    # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
-    # Wrap with quotes to escape yaml
-    - "pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
-    # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
-    - pytest -v -s tests/compile/test_full_graph.py::test_fp8_kv_scale_compile
-
-- label: Blackwell Fusion E2E Tests # 30 min
-  timeout_in_minutes: 40
-  working_dir: "/vllm-workspace/"
-  gpu: b200
-  optional: true
-  num_gpus: 2
-  source_file_dependencies:
-  - csrc/quantization/fp4/
-  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-  - vllm/v1/attention/backends/flashinfer.py
-  - vllm/compilation/
-  # can affect pattern matching
-  - vllm/model_executor/layers/layernorm.py
-  - vllm/model_executor/layers/activation.py
-  - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/test_fusions_e2e.py
-  - tests/compile/test_full_graph.py
-  commands:
-    - nvidia-smi
-    # Run all e2e fusion tests
     - pytest -v -s tests/compile/test_fusions_e2e.py
 
 - label: Blackwell GPT-OSS Eval
@@ -1176,7 +1099,6 @@ steps:
     - pytest -v -s -x lora/test_llama_tp.py
     - pytest -v -s -x lora/test_llm_with_multi_loras.py
     - pytest -v -s -x lora/test_olmoe_tp.py
-    - pytest -v -s -x lora/test_gptoss_tp.py
 
 
 - label: Weight Loading Multiple GPU Test  # 33min
@@ -1202,7 +1124,7 @@ steps:
   - tests/weight_loading
   commands:
     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt
-
+  
 - label: NixlConnector PD accuracy tests (Distributed) # 30min
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
@@ -1244,19 +1166,6 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
-##### H100 test #####
-- label: LM Eval Large Models (H100) # optional
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  commands:
-    - export VLLM_USE_DEEP_GEMM=0  # We found Triton is faster than DeepGEMM for H100
-    - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4
-
 ##### H200 test #####
 - label: Distributed Tests (H200) # optional
   gpu: h200
@@ -1267,11 +1176,9 @@ steps:
     - pytest -v -s tests/compile/test_async_tp.py
     - pytest -v -s tests/compile/test_sequence_parallelism.py
     - pytest -v -s tests/compile/test_fusion_all_reduce.py
-    - "pytest -v -s tests/compile/test_fusions_e2e.py -k 'not Llama-4'"
-    - pytest -v -s tests/distributed/test_sequence_parallel.py
+    - pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
     - pytest -v -s tests/distributed/test_context_parallel.py
     - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048
-    - pytest -v -s tests/v1/distributed/test_dbo.py
 
 ##### B200 test #####
 - label: Distributed Tests (B200) # optional
@@ -1282,7 +1189,6 @@ steps:
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
     - pytest -v -s tests/distributed/test_nccl_symm_mem_allreduce.py
-    - pytest -v -s tests/v1/distributed/test_dbo.py
 
 ##### RL Integration Tests #####
 - label: Prime-RL Integration Test # 15min
@@ -1295,21 +1201,3 @@ steps:
   - .buildkite/scripts/run-prime-rl-test.sh
   commands:
     - bash .buildkite/scripts/run-prime-rl-test.sh
-
-- label: DeepSeek V2-Lite Accuracy
-  timeout_in_minutes: 60
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace"
-  commands:
-  - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
-
-- label: Qwen3-30B-A3B-FP8-block Accuracy
-  timeout_in_minutes: 60
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace"
-  commands:
-  - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh 0.8 200 8020

.github/CODEOWNERS

@@ -3,13 +3,13 @@
 
 # This lists cover the "core" components of vLLM that require careful review
 /vllm/attention @LucasWilkinson
-/vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @njhill
-/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @njhill @22quinn
+/vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @22quinn
 /vllm/model_executor/layers/fused_moe @mgoin @pavanimajety
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety
 /vllm/model_executor/layers/mamba @tdoublep
 /vllm/model_executor/model_loader @22quinn
-/vllm/multimodal @DarkLight1337 @ywang96 @NickLucche @tjtanaa
+/vllm/multimodal @DarkLight1337 @ywang96 @NickLucche
 /vllm/vllm_flash_attn @LucasWilkinson
 /vllm/lora @jeejeelee
 /vllm/reasoning @aarnphm @chaunceyjiang
@@ -20,15 +20,15 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
 # so spam a lot of people
-/vllm/config @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
-/vllm/config/cache.py @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg @heheda12345
+/vllm/config @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
+/vllm/config/cache.py @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg @heheda12345
 
 # vLLM V1
 /vllm/v1/attention @LucasWilkinson
 /vllm/v1/attention/backends/mla @pavanimajety
 /vllm/v1/attention/backends/flashinfer.py @mgoin @pavanimajety
 /vllm/v1/attention/backends/triton_attn.py @tdoublep
-/vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC
+/vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat @heheda12345 @ApostaC
 /vllm/v1/sample @22quinn @houseroad @njhill
 /vllm/v1/spec_decode @benchislett @luccafong
 /vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
@@ -36,11 +36,11 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /vllm/v1/offloading @ApostaC
 
 # Test ownership
-/.buildkite/lm-eval-harness @mgoin 
+/.buildkite/lm-eval-harness @mgoin @simon-mo
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
-/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @aarnphm @NickLucche
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm @NickLucche
 /tests/evals @mgoin
 /tests/kernels @mgoin @tlrmchlsmth @WoosukKwon @yewentao256
 /tests/models @DarkLight1337 @ywang96
@@ -49,7 +49,7 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/test_inputs.py @DarkLight1337 @ywang96
 /tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm
-/tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC
+/tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat @heheda12345 @ApostaC
 /tests/weight_loading @mgoin @youkaichao @yewentao256
 /tests/lora @jeejeelee
 /tests/models/language/generation/test_hybrid.py @tdoublep
@@ -57,20 +57,10 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/v1/kv_connector @ApostaC
 /tests/v1/offloading @ApostaC
 
-# Transformers modeling backend
+# Transformers backend
 /vllm/model_executor/models/transformers @hmellor
 /tests/models/test_transformers.py @hmellor
 
-# Observability
-/vllm/config/observability.py @markmc
-/vllm/v1/metrics @markmc
-/tests/v1/metrics @markmc
-/vllm/tracing.py @markmc
-/tests/v1/tracing/test_tracing.py @markmc
-/vllm/config/kv_events.py @markmc
-/vllm/distributed/kv_events.py @markmc
-/tests/distributed/test_events.py @markmc
-
 # Docs
 /docs/mkdocs @hmellor
 /docs/**/*.yml @hmellor
@@ -115,21 +105,11 @@ mkdocs.yaml @hmellor
 /vllm/attention/ops/triton_unified_attention.py @tdoublep
 
 # ROCm related: specify owner with write access to notify AMD folks for careful code review
-/vllm/**/*rocm* @tjtanaa
-/docker/Dockerfile.rocm* @gshtras @tjtanaa
-/vllm/v1/attention/backends/rocm*.py @gshtras @tjtanaa
-/vllm/v1/attention/backends/mla/rocm*.py @gshtras @tjtanaa
-/vllm/attention/ops/rocm*.py @gshtras @tjtanaa
-/vllm/model_executor/layers/fused_moe/rocm*.py @gshtras @tjtanaa
-/csrc/rocm @gshtras @tjtanaa
-/requirements/*rocm* @tjtanaa
-/tests/**/*rocm* @tjtanaa
-/docs/**/*rocm* @tjtanaa
-/vllm/**/*quark* @tjtanaa
-/tests/**/*quark* @tjtanaa
-/docs/**/*quark* @tjtanaa
-/vllm/**/*aiter* @tjtanaa
-/tests/**/*aiter* @tjtanaa
+/docker/Dockerfile.rocm* @gshtras
+/vllm/v1/attention/backends/rocm*.py @gshtras
+/vllm/v1/attention/backends/mla/rocm*.py @gshtras
+/vllm/attention/ops/rocm*.py @gshtras
+/vllm/model_executor/layers/fused_moe/rocm*.py @gshtras
 
 # TPU
 /vllm/v1/worker/tpu* @NickLucche
@@ -147,8 +127,3 @@ mkdocs.yaml @hmellor
 /vllm/config/pooler.py @noooop
 /vllm/pooling_params.py @noooop
 /vllm/model_executor/layers/pooler.py @noooop
-
-# Security guide and policies
-/docs/usage/security.md @russellb
-/SECURITY.md @russellb
-/docs/contributing/vulnerability_management.md @russellb

.github/mergify.yml

@@ -108,7 +108,7 @@ pull_request_rules:
       - files~=^benchmarks/
       - files~=^vllm/benchmarks/
       - files~=^tests/benchmarks/
-      - files~=^\.buildkite/performance-benchmarks/
+      - files~=^\.buildkite/nightly-benchmarks/
   actions:
     label:
       add:
@@ -151,23 +151,6 @@ pull_request_rules:
       add:
         - gpt-oss
 
-- name: label-nvidia
-  description: Automatically apply nvidia label
-  conditions:
-    - label != stale
-    - or:
-      - files~=cuda
-      - files~=cutlass
-      - files~=flashinfer
-      - files~=trtllm
-      - title~=(?i)NVIDIA
-      - title~=(?i)CUDA
-      - title~=(?i)CUTLASS
-  actions:
-    label:
-      add:
-        - nvidia
-
 - name: label-rocm
   description: Automatically apply rocm label
   conditions:

.github/workflows/macos-smoke-test.yml

@@ -1,81 +0,0 @@
-name: macOS Apple Silicon Smoke Test
-
-on:
-  push:
-    branches:
-      - main
-  workflow_dispatch:  # Manual trigger
-
-jobs:
-  macos-m1-smoke-test:
-    runs-on: macos-latest
-    timeout-minutes: 20
-
-    steps:
-      - uses: actions/checkout@v4
-
-      - uses: astral-sh/setup-uv@v7
-        with:
-          enable-cache: true
-          cache-dependency-glob: |
-            requirements/**/*.txt
-            pyproject.toml
-          python-version: '3.12'
-
-      - name: Create virtual environment
-        run: |
-          uv venv
-          echo "$GITHUB_WORKSPACE/.venv/bin" >> "$GITHUB_PATH"
-
-      - name: Install dependencies and build vLLM
-        run: |
-          uv pip install -r requirements/cpu.txt --index-strategy unsafe-best-match
-          uv pip install -e .
-        env:
-          CMAKE_BUILD_PARALLEL_LEVEL: 4
-
-      - name: Verify installation
-        run: |
-          python -c "import vllm; print(f'vLLM version: {vllm.__version__}')"
-          python -c "import torch; print(f'PyTorch: {torch.__version__}')"
-
-      - name: Smoke test vllm serve
-        timeout-minutes: 10
-        run: |
-          # Start server in background
-          vllm serve Qwen/Qwen3-0.6B \
-            --max-model-len=2048 \
-            --load-format=dummy \
-            --enforce-eager \
-            --port 8000 &
-
-          SERVER_PID=$!
-
-          # Wait for server to start
-          for i in {1..30}; do
-            if curl -s http://localhost:8000/health > /dev/null; then
-              echo "Server started successfully"
-              break
-            fi
-            if [ "$i" -eq 30 ]; then
-              echo "Server failed to start"
-              kill "$SERVER_PID"
-              exit 1
-            fi
-            sleep 2
-          done
-
-          # Test health endpoint
-          curl -f http://localhost:8000/health
-
-          # Test completion
-          curl -f http://localhost:8000/v1/completions \
-            -H "Content-Type: application/json" \
-            -d '{
-              "model": "Qwen/Qwen3-0.6B",
-              "prompt": "Hello",
-              "max_tokens": 5
-            }'
-
-          # Cleanup
-          kill "$SERVER_PID"

.gitignore

@@ -221,6 +221,3 @@ csrc/moe/marlin_moe_wna16/kernel_*
 
 # Ignore ep_kernels_workspace folder
 ep_kernels_workspace/
-
-# Allow tracked library source folders under submodules (e.g., benchmarks/lib)
-!vllm/benchmarks/lib/

.markdownlint.yaml

@@ -3,9 +3,10 @@ MD007:
 MD013: false
 MD024:
   siblings_only: true
-MD031:
-  list_items: false
 MD033: false
+MD045: false
 MD046: false
+MD051: false
 MD052: false
+MD053: false
 MD059: false

.pre-commit-config.yaml

@@ -38,14 +38,14 @@ repos:
   rev: 0.9.1
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu129, --python-platform, x86_64-manylinux_2_28, --python-version, "3.12"]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu129, --python-platform, x86_64-manylinux_2_28]
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
   - id: format-torch-nightly-test
     name: reformat nightly_torch_test.txt to be in sync with test.in
     language: python
-    entry: python tools/pre_commit/generate_nightly_torch_test.py
+    entry: python tools/generate_nightly_torch_test.py
     files: ^requirements/test\.(in|txt)$
   - id: mypy-local
     name: Run mypy locally for lowest supported Python version
@@ -78,12 +78,12 @@ repos:
     stages: [manual] # Only run in CI
   - id: shellcheck
     name: Lint shell scripts
-    entry: tools/pre_commit/shellcheck.sh
+    entry: tools/shellcheck.sh
     language: script
     types: [shell]
   - id: png-lint
     name: Lint PNG exports from excalidraw
-    entry: tools/pre_commit/png-lint.sh
+    entry: tools/png-lint.sh
     language: script
     types: [png]
   - id: signoff-commit
@@ -100,12 +100,12 @@ repos:
     stages: [commit-msg]
   - id: check-spdx-header
     name: Check SPDX headers
-    entry: python tools/pre_commit/check_spdx_header.py
+    entry: python tools/check_spdx_header.py
     language: python
     types: [python]
   - id: check-root-lazy-imports
     name: Check root lazy imports
-    entry: python tools/pre_commit/check_init_lazy_imports.py
+    entry: python tools/check_init_lazy_imports.py
     language: python
     types: [python]
   - id: check-filenames
@@ -119,11 +119,11 @@ repos:
     pass_filenames: false
   - id: update-dockerfile-graph
     name: Update Dockerfile dependency graph
-    entry: tools/pre_commit/update-dockerfile-graph.sh
+    entry: tools/update-dockerfile-graph.sh
     language: script
   - id: enforce-import-regex-instead-of-re
     name: Enforce import regex as re
-    entry: python tools/pre_commit/enforce_regex_import.py
+    entry: python tools/enforce_regex_import.py
     language: python
     types: [python]
     pass_filenames: false
@@ -131,7 +131,7 @@ repos:
   # forbid directly import triton
   - id: forbid-direct-triton-import
     name: "Forbid direct 'import triton'"
-    entry: python tools/pre_commit/check_triton_import.py
+    entry: python tools/check_triton_import.py
     language: python
     types: [python]
     pass_filenames: false
@@ -144,7 +144,7 @@ repos:
     additional_dependencies: [regex]
   - id: validate-config
     name: Validate configuration has default values and that each field has a docstring
-    entry: python tools/pre_commit/validate_config.py
+    entry: python tools/validate_config.py
     language: python
     additional_dependencies: [regex]
   # Keep `suggestion` last

CMakeLists.txt

@@ -39,13 +39,6 @@ set(PYTHON_SUPPORTED_VERSIONS "3.10" "3.11" "3.12" "3.13")
 # Supported AMD GPU architectures.
 set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151")
 
-# ROCm installation prefix. Default to /opt/rocm but allow override via
-# -DROCM_PATH=/your/rocm/path when invoking cmake.
-if(NOT DEFINED ROCM_PATH)
-  set(ROCM_PATH "/opt/rocm" CACHE PATH "ROCm installation prefix")
-else()
-  set(ROCM_PATH ${ROCM_PATH} CACHE PATH "ROCm installation prefix" FORCE)
-endif()
 #
 # Supported/expected torch versions for CUDA/ROCm.
 #
@@ -244,28 +237,11 @@ set_gencode_flags_for_srcs(
   SRCS "${VLLM_CUMEM_EXT_SRC}"
   CUDA_ARCHS "${CUDA_ARCHS}")
 
-if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
+if(VLLM_GPU_LANG STREQUAL "CUDA")
   message(STATUS "Enabling cumem allocator extension.")
-  if(VLLM_GPU_LANG STREQUAL "CUDA")
-    # link against cuda driver library
-    list(APPEND CUMEM_LIBS CUDA::cuda_driver)
-  else()
-    # link against rocm driver library. Prefer an absolute path to
-    # libamdhip64.so inside ${ROCM_PATH}/lib if available, otherwise fall
-    # back to linking by name "amdhip64".
-    find_library(AMDHIP64_LIB
-      NAMES amdhip64 libamdhip64.so
-      PATHS ${ROCM_PATH}/lib
-      NO_DEFAULT_PATH)
-    if(AMDHIP64_LIB)
-      message(STATUS "Found libamdhip64 at ${AMDHIP64_LIB}")
-      list(APPEND CUMEM_LIBS ${AMDHIP64_LIB})
-    else()
-      message(WARNING "libamdhip64 not found in ${ROCM_PATH}/lib; falling back to linking 'amdhip64' by name")
-      list(APPEND CUMEM_LIBS amdhip64)
-    endif()
-  endif()
-  define_extension_target(
+  # link against cuda driver library
+  list(APPEND CUMEM_LIBS CUDA::cuda_driver)
+  define_gpu_extension_target(
     cumem_allocator
     DESTINATION vllm
     LANGUAGE CXX
@@ -289,7 +265,6 @@ set(VLLM_EXT_SRC
   "csrc/pos_encoding_kernels.cu"
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"
-  "csrc/fused_qknorm_rope_kernel.cu"
   "csrc/layernorm_quant_kernels.cu"
   "csrc/sampler.cu"
   "csrc/cuda_view.cu"
@@ -355,7 +330,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # Keep building Marlin for 9.0 as there are some group sizes and shapes that
   # are not supported by Machete yet.
   # 9.0 for latest bf16 atomicAdd PTX
-  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_ARCHS)
 
     #
@@ -512,9 +487,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
   # require CUDA 12.8 or later
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
   endif()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
@@ -619,9 +594,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # FP4 Archs and flags
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(FP4_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(FP4_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(FP4_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(FP4_ARCHS "10.0a;10.1a;12.0a;12.1a" "${CUDA_ARCHS}")
   endif()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
     set(SRCS
@@ -695,7 +670,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
     cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
   endif()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
@@ -741,9 +716,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
   endif()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS "csrc/quantization/w8a8/cutlass/moe/blockwise_scaled_group_mm_sm100.cu")
@@ -861,7 +836,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # Hadacore kernels
-  cuda_archs_loose_intersection(HADACORE_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(HADACORE_ARCHS "8.0;8.9;9.0" "${CUDA_ARCHS}")
   if(HADACORE_ARCHS)
     set(SRCS "csrc/quantization/hadamard/hadacore/hadamard_transform_cuda.cu")
     set_gencode_flags_for_srcs(
@@ -883,7 +858,7 @@ if (VLLM_GPU_LANG STREQUAL "HIP")
 endif()
 
 message(STATUS "Enabling C extension.")
-define_extension_target(
+define_gpu_extension_target(
   _C
   DESTINATION vllm
   LANGUAGE ${VLLM_GPU_LANG}
@@ -939,7 +914,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
   # 9.0 for latest bf16 atomicAdd PTX
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_MOE_ARCHS)
 
     #
@@ -998,7 +973,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 endif()
 
 message(STATUS "Enabling moe extension.")
-define_extension_target(
+define_gpu_extension_target(
   _moe_C
   DESTINATION vllm
   LANGUAGE ${VLLM_GPU_LANG}
@@ -1019,7 +994,7 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
     "csrc/rocm/skinny_gemms.cu"
     "csrc/rocm/attention.cu")
 
-  define_extension_target(
+  define_gpu_extension_target(
     _rocm_C
     DESTINATION vllm
     LANGUAGE ${VLLM_GPU_LANG}

README.md

@@ -21,9 +21,6 @@ Join us at the [PyTorch Conference, October 22-23](https://events.linuxfoundatio
 
 *Latest News* 🔥
 
-- [2025/11] We hosted [the first vLLM Europe Meetup in Zurich](https://luma.com/0gls27kb) focused on quantization, distributed inference, and reinforcement learning at scale with speakers from Mistral, IBM, and Red Hat. Please find the meetup slides [here](https://docs.google.com/presentation/d/1UC9PTLCHYXQpOmJDSFg6Sljra3iVXzc09DeEI7dnxMc/edit?usp=sharing) and recording [here](https://www.youtube.com/watch?v=6m6ZE6yVEDI)
-- [2025/11] We hosted [vLLM Beijing Meetup](https://mp.weixin.qq.com/s/xSrYXjNgr1HbCP4ExYNG1w) focusing on distributed inference and diverse accelerator support with vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1nQJ8ZkLSjKxvu36sSHaceVXtttbLvvu-?usp=drive_link).
-- [2025/10] We hosted [vLLM Shanghai Meetup](https://mp.weixin.qq.com/s/__xb4OyOsImz-9eAVrdlcg) focused on hands-on vLLM inference optimization! Please find the meetup slides [here](https://drive.google.com/drive/folders/1KqwjsFJLfEsC8wlDugnrR61zsWHt94Q6).
 - [2025/09] We hosted [vLLM Toronto Meetup](https://luma.com/e80e0ymm) focused on tackling inference at scale and speculative decoding with speakers from NVIDIA and Red Hat! Please find the meetup slides [here](https://docs.google.com/presentation/d/1IYJYmJcu9fLpID5N5RbW_vO0XLo0CGOR14IXOjB61V8/edit?usp=sharing).
 - [2025/08] We hosted [vLLM Shenzhen Meetup](https://mp.weixin.qq.com/s/k8ZBO1u2_2odgiKWH_GVTQ) focusing on the ecosystem around vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Ua2SVKVSu-wp5vou_6ElraDt2bnKhiEA).
 - [2025/08] We hosted [vLLM Singapore Meetup](https://www.sginnovate.com/event/vllm-sg-meet). We shared V1 updates, disaggregated serving and MLLM speedups with speakers from Embedded LLM, AMD, WekaIO, and A*STAR. Please find the meetup slides [here](https://drive.google.com/drive/folders/1ncf3GyqLdqFaB6IeB834E5TZJPLAOiXZ?usp=sharing).
@@ -85,7 +82,7 @@ vLLM is flexible and easy to use with:
 - Tensor, pipeline, data and expert parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Support for NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, Arm CPUs, and TPU. Additionally, support for diverse hardware plugins such as Intel Gaudi, IBM Spyre and Huawei Ascend.
+- Support for NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, and TPU. Additionally, support for diverse hardware plugins such as Intel Gaudi, IBM Spyre and Huawei Ascend.
 - Prefix caching support
 - Multi-LoRA support
 

benchmarks/benchmark_batch_invariance.py

@@ -1,380 +0,0 @@
-#!/usr/bin/env python3
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""
-Benchmark to measure the performance overhead of VLLM_BATCH_INVARIANT mode.
-
-This benchmark runs the same workload twice:
-1. With VLLM_BATCH_INVARIANT=0 (baseline)
-2. With VLLM_BATCH_INVARIANT=1 (batch invariant mode)
-
-And reports the timing and throughput metrics for comparison.
-
-Environment variables:
-    VLLM_BENCH_MODEL: Model to benchmark (default: "Qwen/Qwen3-1.7B")
-    VLLM_BENCH_TP_SIZE: Tensor parallel size (default: 1, use 8 for deepseek)
-    VLLM_BENCH_BATCH_SIZE: Max batch size (default: 128)
-    VLLM_BENCH_NUM_TRIALS: Number of trials to run (default: 5)
-    VLLM_BENCH_MIN_PROMPT: Min prompt length in words (default: 1024)
-    VLLM_BENCH_MAX_PROMPT: Max prompt length in words (default: 2048)
-    VLLM_BENCH_MAX_TOKENS: Max tokens to generate (default: 128)
-    VLLM_BENCH_TEMPERATURE: Temperature for sampling (default: 0.0)
-    VLLM_BENCH_GPU_MEMORY_UTILIZATION: GPU memory utilization (default: 0.4)
-    VLLM_BENCH_MAX_MODEL_LEN: Max model length (default: 5120)
-    VLLM_BENCH_BACKEND: Attention backend (default: FLASH_ATTN)
-
-Example usage:
-    # Benchmark qwen3 (default)
-    python benchmarks/benchmark_batch_invariance.py
-
-    # Benchmark deepseek with 8 GPUs
-    VLLM_BENCH_MODEL="deepseek-ai/DeepSeek-V3" VLLM_BENCH_TP_SIZE=8 \\
-        python benchmarks/benchmark_batch_invariance.py
-
-    # Quick test with fewer trials
-    VLLM_BENCH_NUM_TRIALS=2 VLLM_BENCH_BATCH_SIZE=32 \\
-        python benchmarks/benchmark_batch_invariance.py
-"""
-
-import contextlib
-import os
-import random
-import time
-
-from vllm import LLM, SamplingParams
-from vllm.platforms import current_platform
-
-
-def _random_prompt(min_words: int = 1024, max_words: int = 1024 * 2) -> str:
-    """Generate a random prompt for benchmarking."""
-    prompt_templates = [
-        "Question: What is the capital of France?\nAnswer: The capital of France is",
-        "Q: How does photosynthesis work?\nA: Photosynthesis is the process by which",
-        "User: Can you explain quantum mechanics?\nAssistant: Quantum mechanics is",
-        "Once upon a time in a distant galaxy, there lived",
-        "The old man walked slowly down the street, remembering",
-        "In the year 2157, humanity finally discovered",
-        "To implement a binary search tree in Python, first we need to",
-        "The algorithm works by iterating through the array and",
-        "Here's how to optimize database queries using indexing:",
-        "The Renaissance was a period in European history that",
-        "Climate change is caused by several factors including",
-        "The human brain contains approximately 86 billion neurons which",
-        "I've been thinking about getting a new laptop because",
-        "Yesterday I went to the store and bought",
-        "My favorite thing about summer is definitely",
-    ]
-
-    base_prompt = random.choice(prompt_templates)
-
-    if max_words < min_words:
-        max_words = min_words
-    target_words = random.randint(min_words, max_words)
-
-    if target_words > 50:
-        padding_text = (
-            " This is an interesting topic that deserves more explanation. "
-            * (target_words // 50)
-        )
-        base_prompt = base_prompt + padding_text
-
-    return base_prompt
-
-
-def run_benchmark_with_batch_invariant(
-    model: str,
-    tp_size: int,
-    max_batch_size: int,
-    num_trials: int,
-    min_prompt: int,
-    max_prompt: int,
-    max_tokens: int,
-    temperature: float,
-    gpu_mem_util: float,
-    max_model_len: int,
-    backend: str,
-    batch_invariant: bool,
-    seed: int = 12345,
-) -> dict:
-    """
-    Run the benchmark with the specified configuration.
-
-    Returns a dict with timing and throughput metrics.
-    """
-    random.seed(seed)
-
-    # Set environment variables
-    os.environ["VLLM_ATTENTION_BACKEND"] = backend
-    if batch_invariant:
-        os.environ["VLLM_BATCH_INVARIANT"] = "1"
-    else:
-        os.environ["VLLM_BATCH_INVARIANT"] = "0"
-
-    print(f"\n{'=' * 80}")
-    print(f"BENCHMARK: VLLM_BATCH_INVARIANT={int(batch_invariant)}")
-    print(f"  Model: {model}")
-    print(f"  TP Size: {tp_size}")
-    print(f"  Backend: {backend}")
-    print(f"  Max Batch Size: {max_batch_size}")
-    print(f"  Trials: {num_trials}")
-    print(f"  Max Tokens: {max_tokens}")
-    print(f"{'=' * 80}\n")
-
-    sampling = SamplingParams(
-        temperature=temperature,
-        top_p=0.95,
-        max_tokens=max_tokens,
-        seed=20240919,
-    )
-
-    needle_prompt = "There once was a "
-
-    llm = None
-    try:
-        # Create LLM engine
-        start_init = time.perf_counter()
-        llm = LLM(
-            model=model,
-            max_num_seqs=max_batch_size,
-            gpu_memory_utilization=gpu_mem_util,
-            max_model_len=max_model_len,
-            dtype="bfloat16",
-            tensor_parallel_size=tp_size,
-            enable_prefix_caching=False,
-        )
-        init_time = time.perf_counter() - start_init
-        print(f"Engine initialization time: {init_time:.2f}s\n")
-
-        # Generate baseline
-        print("Generating baseline (warmup)...")
-        baseline_out = llm.generate([needle_prompt], sampling)
-        assert len(baseline_out) == 1
-        baseline_text = baseline_out[0].outputs[0].text
-        print(f"Baseline output: '{baseline_text[:50]}...'\n")
-
-        # Run trials and measure timing
-        trial_times: list[float] = []
-        total_tokens = 0
-        total_prompts = 0
-
-        for trial in range(num_trials):
-            # Create a batch
-            prompts: list[str] = []
-            batch_size = random.randint(max_batch_size // 2, max_batch_size)
-            needle_pos = random.randint(0, batch_size - 1)
-            for i in range(batch_size):
-                if i == needle_pos:
-                    prompts.append(needle_prompt)
-                else:
-                    prompts.append(_random_prompt(min_prompt, max_prompt))
-
-            # Measure time for this trial
-            start_time = time.perf_counter()
-            outputs = llm.generate(prompts, sampling)
-            trial_time = time.perf_counter() - start_time
-
-            trial_times.append(trial_time)
-            total_prompts += len(prompts)
-
-            # Count tokens
-            for output in outputs:
-                if output.outputs:
-                    total_tokens += len(output.outputs[0].token_ids)
-
-            print(
-                f"Trial {trial + 1}/{num_trials}: "
-                f"batch_size={batch_size}, "
-                f"time={trial_time:.2f}s"
-            )
-
-            # Verify needle output still matches
-            needle_output = outputs[needle_pos]
-            assert needle_output.prompt == needle_prompt
-
-        # Compute statistics
-        avg_time = sum(trial_times) / len(trial_times)
-        min_time = min(trial_times)
-        max_time = max(trial_times)
-        throughput = total_tokens / sum(trial_times)
-        prompts_per_sec = total_prompts / sum(trial_times)
-
-        print(f"\n{'=' * 80}")
-        print("RESULTS:")
-        print(f"  Average time per trial: {avg_time:.2f}s")
-        print(f"  Min time: {min_time:.2f}s")
-        print(f"  Max time: {max_time:.2f}s")
-        print(f"  Total tokens generated: {total_tokens}")
-        print(f"  Total prompts processed: {total_prompts}")
-        print(f"  Throughput: {throughput:.2f} tokens/s")
-        print(f"  Prompts/s: {prompts_per_sec:.2f}")
-        print(f"{'=' * 80}\n")
-
-        return {
-            "init_time": init_time,
-            "avg_time": avg_time,
-            "min_time": min_time,
-            "max_time": max_time,
-            "total_tokens": total_tokens,
-            "total_prompts": total_prompts,
-            "throughput": throughput,
-            "prompts_per_sec": prompts_per_sec,
-            "trial_times": trial_times,
-        }
-
-    finally:
-        # Cleanup
-        if llm is not None:
-            with contextlib.suppress(Exception):
-                llm.shutdown()
-
-
-def main():
-    # Check platform support
-    if not (current_platform.is_cuda() and current_platform.has_device_capability(90)):
-        print("ERROR: Requires CUDA and >= Hopper (SM90)")
-        print(f"Current platform: {current_platform.device_type}")
-        if current_platform.is_cuda():
-            print(f"Device capability: {current_platform.get_device_capability()}")
-        return 1
-
-    # Read configuration from environment
-    model = os.getenv("VLLM_BENCH_MODEL", "Qwen/Qwen3-1.7B")
-    tp_size = int(os.getenv("VLLM_BENCH_TP_SIZE", "1"))
-    max_batch_size = int(os.getenv("VLLM_BENCH_BATCH_SIZE", "128"))
-    num_trials = int(os.getenv("VLLM_BENCH_NUM_TRIALS", "5"))
-    min_prompt = int(os.getenv("VLLM_BENCH_MIN_PROMPT", "1024"))
-    max_prompt = int(os.getenv("VLLM_BENCH_MAX_PROMPT", "2048"))
-    max_tokens = int(os.getenv("VLLM_BENCH_MAX_TOKENS", "128"))
-    temperature = float(os.getenv("VLLM_BENCH_TEMPERATURE", "0.0"))
-    gpu_mem_util = float(os.getenv("VLLM_BENCH_GPU_MEMORY_UTILIZATION", "0.4"))
-    max_model_len = int(os.getenv("VLLM_BENCH_MAX_MODEL_LEN", "5120"))
-    backend = os.getenv("VLLM_BENCH_BACKEND", "FLASH_ATTN")
-
-    print("\n" + "=" * 80)
-    print("VLLM BATCH INVARIANCE BENCHMARK")
-    print("=" * 80)
-    print("\nConfiguration:")
-    print(f"  Model: {model}")
-    print(f"  Tensor Parallel Size: {tp_size}")
-    print(f"  Attention Backend: {backend}")
-    print(f"  Max Batch Size: {max_batch_size}")
-    print(f"  Number of Trials: {num_trials}")
-    print(f"  Prompt Length Range: {min_prompt}-{max_prompt} words")
-    print(f"  Max Tokens to Generate: {max_tokens}")
-    print(f"  Temperature: {temperature}")
-    print(f"  GPU Memory Utilization: {gpu_mem_util}")
-    print(f"  Max Model Length: {max_model_len}")
-    print("=" * 80)
-
-    # Run benchmark WITHOUT batch invariance (baseline)
-    print("\n" + "=" * 80)
-    print("PHASE 1: Running WITHOUT batch invariance (baseline)")
-    print("=" * 80)
-    baseline_results = run_benchmark_with_batch_invariant(
-        model=model,
-        tp_size=tp_size,
-        max_batch_size=max_batch_size,
-        num_trials=num_trials,
-        min_prompt=min_prompt,
-        max_prompt=max_prompt,
-        max_tokens=max_tokens,
-        temperature=temperature,
-        gpu_mem_util=gpu_mem_util,
-        max_model_len=max_model_len,
-        backend=backend,
-        batch_invariant=False,
-    )
-
-    # Run benchmark WITH batch invariance
-    print("\n" + "=" * 80)
-    print("PHASE 2: Running WITH batch invariance")
-    print("=" * 80)
-    batch_inv_results = run_benchmark_with_batch_invariant(
-        model=model,
-        tp_size=tp_size,
-        max_batch_size=max_batch_size,
-        num_trials=num_trials,
-        min_prompt=min_prompt,
-        max_prompt=max_prompt,
-        max_tokens=max_tokens,
-        temperature=temperature,
-        gpu_mem_util=gpu_mem_util,
-        max_model_len=max_model_len,
-        backend=backend,
-        batch_invariant=True,
-    )
-
-    # Compare results
-    print("\n" + "=" * 80)
-    print("COMPARISON: Batch Invariance vs Baseline")
-    print("=" * 80)
-
-    init_overhead_pct = (
-        (batch_inv_results["init_time"] - baseline_results["init_time"])
-        / baseline_results["init_time"]
-        * 100
-    )
-    time_overhead_pct = (
-        (batch_inv_results["avg_time"] - baseline_results["avg_time"])
-        / baseline_results["avg_time"]
-        * 100
-    )
-    throughput_change_pct = (
-        (batch_inv_results["throughput"] - baseline_results["throughput"])
-        / baseline_results["throughput"]
-        * 100
-    )
-
-    print("\nInitialization Time:")
-    print(f"  Baseline:         {baseline_results['init_time']:.2f}s")
-    print(f"  Batch Invariant:  {batch_inv_results['init_time']:.2f}s")
-    print(f"  Overhead:         {init_overhead_pct:+.2f}%")
-
-    print("\nAverage Trial Time:")
-    print(f"  Baseline:         {baseline_results['avg_time']:.2f}s")
-    print(f"  Batch Invariant:  {batch_inv_results['avg_time']:.2f}s")
-    print(f"  Overhead:         {time_overhead_pct:+.2f}%")
-
-    print("\nThroughput (tokens/s):")
-    print(f"  Baseline:         {baseline_results['throughput']:.2f}")
-    print(f"  Batch Invariant:  {batch_inv_results['throughput']:.2f}")
-    print(f"  Change:           {throughput_change_pct:+.2f}%")
-
-    print("\nPrompts/s:")
-    print(f"  Baseline:         {baseline_results['prompts_per_sec']:.2f}")
-    print(f"  Batch Invariant:  {batch_inv_results['prompts_per_sec']:.2f}")
-
-    print("\n" + "=" * 80)
-    print("SUMMARY")
-    print("=" * 80)
-    if time_overhead_pct > 0:
-        print(
-            f"Batch invariance mode adds approximately {time_overhead_pct:.1f}% "
-            "overhead"
-        )
-    else:
-        print(
-            f"Batch invariance mode is approximately {-time_overhead_pct:.1f}% "
-            "faster (unexpected!)"
-        )
-
-    if abs(throughput_change_pct) < 1.0:
-        print("Throughput difference is negligible (< 1%)")
-    elif throughput_change_pct < 0:
-        print(
-            f"Throughput decreased by {-throughput_change_pct:.1f}% "
-            "with batch invariance"
-        )
-    else:
-        print(
-            f"Throughput increased by {throughput_change_pct:.1f}% "
-            "with batch invariance (unexpected!)"
-        )
-
-    print("=" * 80 + "\n")
-
-    return 0
-
-
-if __name__ == "__main__":
-    exit(main())

benchmarks/benchmark_prefix_caching.py

@@ -69,7 +69,7 @@ def sample_tokens(tokenizer: PreTrainedTokenizerBase, length: int) -> list[int]:
 
     # Remove the special tokens.
     return random.choices(
-        [v for v in vocab.values() if v not in all_special_ids],
+        [v for k, v in vocab.items() if k not in all_special_ids],
         k=length,
     )
 

benchmarks/kernels/bench_block_fp8_gemm.py

@@ -1,18 +1,10 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
-import os
-
-# Disable DeepGEMM for this benchmark to use CUTLASS
-os.environ["VLLM_USE_DEEP_GEMM"] = "0"
-
 import torch
 
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    W8A8BlockFp8LinearOp,
-)
-from vllm.model_executor.layers.quantization.utils.quant_utils import (
-    GroupShape,
+    apply_w8a8_block_fp8_linear,
 )
 from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
     CUTLASS_BLOCK_FP8_SUPPORTED,
@@ -47,14 +39,13 @@ def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
     fp8_info = torch.finfo(torch.float8_e4m3fn)
     fp8_max, fp8_min = fp8_info.max, fp8_info.min
 
-    # Create random input tensor (bfloat16, will be quantized by W8A8BlockFp8LinearOp)
+    # Create random FP8 tensors
     A_ref = (torch.rand(M, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
 
-    # Create quantized weight tensor
     B_ref = (torch.rand(N, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
     B = B_ref.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
 
-    # Create weight scales
+    # Create scales
     block_n, block_k = block_size[0], block_size[1]
     n_tiles = (N + block_n - 1) // block_n
     k_tiles = (K + block_k - 1) // block_k
@@ -64,25 +55,19 @@ def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
         * factor_for_scale
     )
 
-    # Create W8A8BlockFp8LinearOp instance
-    weight_group_shape = GroupShape(block_n, block_k)
-    act_quant_group_shape = GroupShape(1, block_k)  # Per-token, per-group quantization
-
-    linear_op = W8A8BlockFp8LinearOp(
-        weight_group_shape=weight_group_shape,
-        act_quant_group_shape=act_quant_group_shape,
-        cutlass_block_fp8_supported=use_cutlass,
-        use_aiter_and_is_supported=False,
-    )
+    # SM90 CUTLASS requires row-major format for scales
+    if use_cutlass and current_platform.is_device_capability(90):
+        Bs = Bs.T.contiguous()
 
     def run():
-        return linear_op.apply(
-            input=A_ref,
-            weight=B,
-            weight_scale=Bs,
-            input_scale=None,
-            bias=None,
-        )
+        if use_cutlass:
+            return apply_w8a8_block_fp8_linear(
+                A_ref, B, block_size, Bs, cutlass_block_fp8_supported=True
+            )
+        else:
+            return apply_w8a8_block_fp8_linear(
+                A_ref, B, block_size, Bs, cutlass_block_fp8_supported=False
+            )
 
     return run
 

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -16,8 +16,8 @@ from vllm.model_executor.layers.fused_moe.fused_moe import (
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = [
-    "mistralai/Mixtral-8x7B-Instruct-v0.1",
-    "deepseek-ai/DeepSeek-V2-Lite",
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1",
+    "nm-testing/deepseekv2-lite",
     "ibm-granite/granite-3.0-1b-a400m",
     "ibm-granite/granite-3.0-3b-a800m",
 ]

benchmarks/kernels/benchmark_lora.py

@@ -19,24 +19,13 @@ from torch.utils.benchmark import Measurement as TMeasurement
 from utils import ArgPool, Bench, CudaGraphBenchParams
 from weight_shapes import WEIGHT_SHAPES
 
-from vllm.lora.ops.triton_ops.utils import get_lora_op_configs
-from vllm.triton_utils import HAS_TRITON, triton
+from vllm.triton_utils import HAS_TRITON
 
 if HAS_TRITON:
-    from vllm.lora.ops.triton_ops import (  ## added fused_moe_lora
-        LoRAKernelMeta,
-        fused_moe_lora_expand,
-        fused_moe_lora_shrink,
-        lora_expand,
-        lora_shrink,
-    )
-    from vllm.lora.ops.triton_ops.fused_moe_lora_op import (
-        _LORA_PTR_DICT,  ## added _LORA_PTR_DICT for fused_moe_lora
-    )
+    from vllm.lora.ops.triton_ops import LoRAKernelMeta, lora_expand, lora_shrink
     from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT
-from vllm import _custom_ops as ops
+
 from vllm.utils.argparse_utils import FlexibleArgumentParser
-from vllm.utils.math_utils import round_up
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_TP_SIZES = [1]
@@ -70,8 +59,6 @@ DEFAULT_NUM_LORAS = [1, 2, 3, 4]
 DEFAULT_SORT_BY_LORA_IDS = [False, True]
 DEFAULT_SEQ_LENGTHS = [1]
 DEFAULT_EXPAND_FN_ADD_INPUTS = [True, False]
-DEFAULT_TOP_K_NUMS = [1]  # Added for MoE LoRA top_k
-DEFAULT_NUM_EXPERTS = [8]  # Added for MoE LoRA num_experts
 
 
 # Utilities
@@ -204,11 +191,6 @@ class OpType(Enum):
 
     LORA_SHRINK = auto()
     LORA_EXPAND = auto()
-    ## Adding support for fused moe lora
-    FUSED_MOE_LORA_GATE_UP_SHRINK = auto()  ## Gate/Up projection variant with shrink
-    FUSED_MOE_LORA_GATE_UP_EXPAND = auto()  ## Gate/Up projection variant with expand
-    FUSED_MOE_LORA_DOWN_SHRINK = auto()  ## Down projection variant with shrink
-    FUSED_MOE_LORA_DOWN_EXPAND = auto()  ## Down projection variant with expand
 
     @staticmethod
     def from_str(s: str) -> "OpType":
@@ -216,15 +198,6 @@ class OpType(Enum):
             return OpType.LORA_SHRINK
         if s.lower() == "lora_expand":
             return OpType.LORA_EXPAND
-        # Adding support for fused moe lora, both in gate_up and down
-        if s.lower() == "fused_moe_lora_gate_up_shrink":  ## Gate/Up variant with shrink
-            return OpType.FUSED_MOE_LORA_GATE_UP_SHRINK
-        if s.lower() == "fused_moe_lora_gate_up_expand":  ## Gate/Up variant with expand
-            return OpType.FUSED_MOE_LORA_GATE_UP_EXPAND
-        if s.lower() == "fused_moe_lora_down_shrink":  ## Down variant with shrink
-            return OpType.FUSED_MOE_LORA_DOWN_SHRINK
-        if s.lower() == "fused_moe_lora_down_expand":  ## Down variant with expand
-            return OpType.FUSED_MOE_LORA_DOWN_EXPAND
         raise ValueError(f"Unrecognized str {s} to convert to OpType")
 
     def is_shrink_fn(self) -> bool:
@@ -233,56 +206,19 @@ class OpType(Enum):
     def is_expand_fn(self) -> bool:
         return self in [OpType.LORA_EXPAND]
 
-    def is_fused_moe_lora_fn(self) -> bool:  ## adding for fused MoE LoRA
-        return self in [
-            OpType.FUSED_MOE_LORA_GATE_UP_SHRINK,
-            OpType.FUSED_MOE_LORA_DOWN_SHRINK,
-            OpType.FUSED_MOE_LORA_GATE_UP_EXPAND,
-            OpType.FUSED_MOE_LORA_DOWN_EXPAND,
-        ]
-
-    def is_fused_moe_lora_gate_up_fn(
-        self,
-    ) -> bool:  ## adding for fused MoE LoRA Gate/Up
-        return self in [
-            OpType.FUSED_MOE_LORA_GATE_UP_SHRINK,
-            OpType.FUSED_MOE_LORA_GATE_UP_EXPAND,
-        ]
-
-    def is_fused_moe_lora_down_fn(self) -> bool:  ## adding for fused MoE LoRA Down
-        return self in [
-            OpType.FUSED_MOE_LORA_DOWN_SHRINK,
-            OpType.FUSED_MOE_LORA_DOWN_EXPAND,
-        ]
-
-    def is_fused_moe_lora_shrink_fn(self) -> bool:
-        return self in [
-            OpType.FUSED_MOE_LORA_GATE_UP_SHRINK,
-            OpType.FUSED_MOE_LORA_DOWN_SHRINK,
-        ]
-
-    def is_fused_moe_lora_expand_fn(self) -> bool:
-        return self in [
-            OpType.FUSED_MOE_LORA_GATE_UP_EXPAND,
-            OpType.FUSED_MOE_LORA_DOWN_EXPAND,
-        ]
-
     def num_slices(self) -> list[int]:
-        if self.is_fused_moe_lora_gate_up_fn():
-            return [2]
-        elif self.is_fused_moe_lora_down_fn():
-            return [1]
         return [1, 2, 3]
 
     def mkn(
         self, batch_size: int, seq_length: int, hidden_size: int, lora_rank: int
     ) -> tuple[int, int, int]:
         num_tokens = batch_size * seq_length
-        if self.is_shrink_fn() or self.is_fused_moe_lora_fn():
+        if self.is_shrink_fn():
             m = num_tokens
             k = hidden_size
             n = lora_rank
-        elif self.is_expand_fn():
+        else:
+            assert self.is_expand_fn()
             m = num_tokens
             k = lora_rank
             n = hidden_size
@@ -296,36 +232,9 @@ class OpType(Enum):
         """
         if self.is_shrink_fn():
             return op_dtype, op_dtype, torch.float32
-        elif self.is_expand_fn():
+        else:
+            assert self.is_expand_fn()
             return torch.float32, op_dtype, op_dtype
-        else:
-            assert self.is_fused_moe_lora_fn()
-            return op_dtype, op_dtype, op_dtype
-
-    def matmul_shapes_fused_moe_lora(
-        self,
-        m: int,
-        n: int,
-        k: int,
-        num_loras: int,
-        num_slices: int,
-        top_k_num: int,
-        num_experts: int,
-    ) -> tuple[tuple[int], tuple[int], tuple[int], tuple[int]]:
-        if self.is_fused_moe_lora_shrink_fn():
-            input_shape = (
-                (m * top_k_num, n)
-                if self in [OpType.FUSED_MOE_LORA_DOWN_SHRINK]
-                else (m, n)
-            )
-            output_shape = (num_slices, m, top_k_num, k)
-            weight_shape = (num_loras, num_experts, k, n)
-        else:
-            assert self.is_fused_moe_lora_expand_fn()
-            input_shape = (num_slices, m, top_k_num, k)
-            output_shape = (m, top_k_num, n * num_slices)
-            weight_shape = (num_loras, num_experts, n, k)
-        return (input_shape, weight_shape, output_shape)
 
     def matmul_shapes(
         self,
@@ -335,8 +244,6 @@ class OpType(Enum):
         lora_rank: int,
         num_loras: int,
         num_slices: int,
-        top_k_num: int | None = None,
-        num_experts: int | None = None,
     ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
         """
         Given num_slices, return the shapes of the A, B, and C matrices
@@ -351,16 +258,6 @@ class OpType(Enum):
         if self in [OpType.LORA_EXPAND]:
             # LoRA expand kernels support num_slices inherently in the kernel
             return ((num_slices, m, k), b_shape, (m, n * num_slices))
-        if self.is_fused_moe_lora_fn():
-            return self.matmul_shapes_fused_moe_lora(
-                m,
-                k,
-                n,
-                num_loras,
-                num_slices,
-                top_k_num,
-                num_experts,
-            )
         raise ValueError(f"Unrecognized op_type {self}")
 
     def bench_fn(self) -> Callable:
@@ -368,16 +265,6 @@ class OpType(Enum):
             return lora_shrink
         if self == OpType.LORA_EXPAND:
             return lora_expand
-        if self in [
-            OpType.FUSED_MOE_LORA_GATE_UP_SHRINK,
-            OpType.FUSED_MOE_LORA_DOWN_SHRINK,
-        ]:
-            return fused_moe_lora_shrink
-        if self in [
-            OpType.FUSED_MOE_LORA_GATE_UP_EXPAND,
-            OpType.FUSED_MOE_LORA_DOWN_EXPAND,
-        ]:
-            return fused_moe_lora_expand
 
         raise ValueError(f"Unrecognized optype {self}")
 
@@ -431,8 +318,6 @@ class BenchmarkContext:
     sort_by_lora_id: bool
     dtype: torch.dtype
     seq_length: int | None = None
-    num_experts: int | None = None  # num_experts for MoE based ops
-    top_k_num: int | None = None  # top_k for MoE based ops
     num_slices: int | None = None  # num_slices for slice based ops
 
     def with_seq_length(self, seq_length: int) -> "BenchmarkContext":
@@ -488,11 +373,6 @@ class BenchmarkTensors:
             f"{dtype_to_str(self.output.dtype)}"
         )
 
-    def get_num_tokens(self, size: int, top_k_num: int, op_type: OpType):
-        return (
-            size * top_k_num if op_type in [OpType.FUSED_MOE_LORA_DOWN_SHRINK] else size
-        )
-
     @staticmethod
     def make(
         ctx: BenchmarkContext, op_type: OpType, device: str = "cuda"
@@ -505,8 +385,6 @@ class BenchmarkTensors:
             ctx.lora_rank,
             ctx.num_loras,
             ctx.num_slices,
-            ctx.top_k_num,
-            ctx.num_experts,
         )
         a_type, b_type, c_type = op_type.matmul_dtypes(ctx.dtype)
         input_tensor, lora_weights, output_tensor = make_rand_tensors(
@@ -554,27 +432,17 @@ class BenchmarkTensors:
             prompt_lora_indices_tensor,
         )
 
-    def sanity_check(self, ctx: BenchmarkContext, op_type: OpType) -> None:
+    def sanity_check(self) -> None:
         """
         Fails asserts when non-conformality is detected.
         """
-        num_tokens = (
-            self.input.shape[1]
-            if op_type.is_fused_moe_lora_expand_fn()
-            else self.input.shape[-2]
-        )
+        num_tokens = self.input.shape[-2]
         # check metadata tensors
-        ## In down shrink case, each token is repeated top_k_num times
-        assert num_tokens == self.get_num_tokens(
-            torch.sum(self.seq_lens), ctx.top_k_num, op_type
-        ), f"Expected {num_tokens} tokens, but got {torch.sum(self.seq_lens)}"
+        assert torch.sum(self.seq_lens) == num_tokens
         num_seqs = self.seq_lens.shape[0]
         # assert self.seq_start_loc.shape[0] == num_seqs
-        ## In down shrink case, each prompt corresponds to top_k_num sequences
         assert self.prompt_lora_mapping.shape[0] == num_seqs
-        assert self.get_num_tokens(
-            self.lora_kernel_meta.token_lora_mapping.shape[0], ctx.top_k_num, op_type
-        )
+        assert self.lora_kernel_meta.token_lora_mapping.shape[0] == num_tokens
 
     def to_device(self, device: str):
         """
@@ -603,111 +471,21 @@ class BenchmarkTensors:
                 to_device(field) if field_name != "no_lora_flag_cpu" else field,
             )
 
-    def metadata(self, ctx: BenchmarkContext, op_type: OpType) -> tuple[int, int, int]:
+    def metadata(self) -> tuple[int, int, int]:
         """
         Return num_seqs, num_tokens and max_seq_len
         """
         num_seqs = self.seq_lens.shape[0]
-        num_tokens = self.get_num_tokens(
-            self.lora_kernel_meta.token_lora_mapping.shape[0], ctx.top_k_num, op_type
-        )
+        num_tokens = self.lora_kernel_meta.token_lora_mapping.shape[0]
         max_seq_len = torch.max(self.seq_lens).item()
         num_slices = len(self.lora_weights_lst)
         return num_seqs, num_tokens, max_seq_len, num_slices
 
-    def fused_moe_lora_data_prepare(
-        self,
-        block_size: int,
-        token_lora_mapping: torch.Tensor,
-        ctx: BenchmarkContext,
-    ):
-        def moe_lora_align_block_size(
-            topk_ids: torch.Tensor,
-            token_lora_mapping: torch.Tensor,
-            block_size: int,
-            num_experts: int,
-            max_loras: int,
-            expert_map: torch.Tensor | None = None,
-            pad_sorted_ids: bool = False,
-        ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-            """
-            Aligns tokens and experts into block-sized chunks for LoRA-based
-            mixture-of-experts (MoE) execution.
-            """
-            max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
-            if pad_sorted_ids:
-                max_num_tokens_padded = round_up(max_num_tokens_padded, block_size)
-            sorted_ids = torch.empty(
-                (max_loras * max_num_tokens_padded,),
-                dtype=torch.int32,
-                device=topk_ids.device,
-            )
-            max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size)
-            # Expert ids must be set default to -1 to prevent a blank block
-            expert_ids = torch.empty(
-                (max_loras * max_num_m_blocks,),
-                dtype=torch.int32,
-                device=topk_ids.device,
-            )
-            num_tokens_post_pad = torch.empty(
-                (max_loras), dtype=torch.int32, device=topk_ids.device
-            )
-
-            ops.moe_lora_align_block_size(
-                topk_ids,
-                token_lora_mapping,
-                num_experts,
-                block_size,
-                max_loras,
-                max_num_tokens_padded,
-                max_num_m_blocks,
-                sorted_ids,
-                expert_ids,
-                num_tokens_post_pad,
-            )
-            if expert_map is not None:
-                expert_ids = expert_map[expert_ids]
-
-            return sorted_ids, expert_ids, num_tokens_post_pad
-
-        num_tokens = ctx.batch_size
-        curr_topk_ids = torch.randint(
-            0,
-            ctx.num_experts,
-            (num_tokens, ctx.top_k_num),
-            device="cuda",
-            dtype=torch.int32,
-        )
-        topk_weights = torch.randint(
-            0,
-            ctx.num_experts,
-            (num_tokens, ctx.top_k_num),
-            device="cuda",
-            dtype=torch.int32,
-        )
-
-        (sorted_token_ids_lora, expert_ids_lora, num_tokens_post_padded_lora) = (
-            moe_lora_align_block_size(
-                topk_ids=curr_topk_ids,
-                token_lora_mapping=token_lora_mapping,
-                block_size=block_size,
-                num_experts=ctx.num_experts,
-                max_loras=ctx.num_loras,
-            )
-        )
-
-        sorted_token_ids = sorted_token_ids_lora.view(ctx.num_loras, -1)
-        expert_ids = expert_ids_lora.view(ctx.num_loras, -1)
-        num_tokens_post_padded = num_tokens_post_padded_lora
-        return (topk_weights, sorted_token_ids, expert_ids, num_tokens_post_padded)
-
-    def as_lora_shrink_kwargs(
-        self, ctx: BenchmarkContext, op_type: OpType
-    ) -> dict[str, Any]:
-        self.sanity_check(ctx, op_type)
+    def as_lora_shrink_kwargs(self) -> dict[str, Any]:
+        self.sanity_check()
         self.to_device(self.input.device)
 
-        _, num_tokens, _, num_slices = self.metadata(ctx, op_type)
+        _, num_tokens, _, num_slices = self.metadata()
 
         # Sanity check matrix shapes.
         i_shape, lw_shape, o_shape = (
@@ -742,13 +520,11 @@ class BenchmarkTensors:
             "no_lora_flag_cpu": self.lora_kernel_meta.no_lora_flag_cpu,
         }
 
-    def as_lora_expand_kwargs(
-        self, ctx: BenchmarkContext, op_type: OpType, add_inputs: bool
-    ) -> dict[str, Any]:
-        self.sanity_check(ctx, op_type)
+    def as_lora_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]:
+        self.sanity_check()
         self.to_device(self.input.device)
 
-        _, num_tokens, _, num_slices = self.metadata(ctx, op_type)
+        _, num_tokens, _, num_slices = self.metadata()
 
         # Sanity check matrix shapes.
         i_shape, lw_shape, o_shape = (
@@ -785,173 +561,18 @@ class BenchmarkTensors:
             "no_lora_flag_cpu": self.lora_kernel_meta.no_lora_flag_cpu,
         }
 
-    def as_fused_moe_lora_shrink_kwargs(
-        self, ctx: BenchmarkContext, op_type: OpType
-    ) -> dict[str, Any]:
-        self.sanity_check(ctx, op_type)
-        self.to_device(self.input.device)
-
-        _, num_tokens, _, num_slices = self.metadata(ctx, op_type)
-
-        # Sanity check matrix shapes.
-        i_shape, lw_shape, o_shape = (
-            self.input.shape,
-            self.lora_weights_lst[0].shape,
-            self.output.shape,
-        )
-        # Expected input shape : [num_tokens, hidden_size] for gate_up
-        # Expected input shape : [top_k_num * num_tokens, hidden_size] for down
-        assert len(i_shape) == 2
-        assert i_shape[0] == num_tokens
-        hidden_size = i_shape[1]
-        # Expected lora weight shape [max_lora, num_experts, lora_rank, hidden_size]
-        assert len(lw_shape) == 4
-        assert lw_shape[-1] == hidden_size
-        lora_rank = lw_shape[-2]
-        # Expected output shape : [num_slices, num_tokens, top_k_num, lora_rank]
-        assert len(o_shape) == 4
-        assert (
-            o_shape
-            == (num_slices, num_tokens // ctx.top_k_num, ctx.top_k_num, lora_rank)
-            if op_type in [OpType.FUSED_MOE_LORA_DOWN_SHRINK]
-            else o_shape == (num_slices, num_tokens, ctx.top_k_num, lora_rank)
-        )
-        kernel_config = get_lora_op_configs(
-            op_type.name.lower(),
-            max_loras=lw_shape[0],
-            batch=num_tokens,
-            hidden_size=hidden_size,
-            rank=lora_rank,
-            num_slices=num_slices,
-            add_inputs=False,
-        )
-
-        (topk_weights, sorted_token_ids, expert_ids, num_tokens_post_padded) = (
-            self.fused_moe_lora_data_prepare(
-                block_size=kernel_config["BLOCK_SIZE_M"],
-                token_lora_mapping=self.lora_kernel_meta.token_lora_mapping,
-                ctx=ctx,
-            )
-        )
-
-        return {
-            "qcurr_hidden_states": self.input,
-            "lora_a_stacked": self.lora_weights_lst,
-            "a_intermediate_cache1": self.output,
-            "topk_weights": topk_weights,
-            "sorted_token_ids": sorted_token_ids,
-            "expert_ids": expert_ids,
-            "num_tokens_post_padded": num_tokens_post_padded,
-            "top_k_num": ctx.top_k_num,
-            "device": self.input.device,
-            "N": lora_rank,
-            "M": topk_weights.shape[0],
-            "EM": sorted_token_ids.shape[1],
-            "K": self.input.shape[1],
-            "num_tokens": num_tokens,
-            "num_experts": ctx.num_experts,
-            "num_slices": num_slices,
-            "shrink_block_size_m": kernel_config["BLOCK_SIZE_M"],
-            "shrink_block_size_n": kernel_config["BLOCK_SIZE_N"],
-            "shrink_block_size_k": kernel_config["BLOCK_SIZE_K"],
-            "shrink_group_size_m": kernel_config["GROUP_SIZE_M"],
-            "shrink_num_warps": kernel_config["NUM_WARPS"],
-            "shrink_num_stages": kernel_config["NUM_STAGES"],
-            "shrink_split_k": kernel_config.get("SPLIT_K", 1),
-            "mul_routed_weight": op_type.is_fused_moe_lora_down_fn(),
-        }
-
-    def as_fused_moe_lora_expand_kwargs(
-        self, ctx: BenchmarkContext, op_type: OpType
-    ) -> dict[str, Any]:
-        self.sanity_check(ctx, op_type)
-        self.to_device(self.input.device)
-
-        _, num_tokens, _, num_slices = self.metadata(ctx, op_type)
-
-        # Sanity check matrix shapes.
-        i_shape, lw_shape, o_shape = (
-            self.input.shape,
-            self.lora_weights_lst[0].shape,
-            self.output.shape,
-        )
-
-        # Expected input shape : [num_slices, num_tokens, top_k_num, lora_rank]
-        assert len(i_shape) == 4
-        assert i_shape[0] == num_slices
-        assert i_shape[1] == num_tokens
-        lora_rank = i_shape[-1]
-        # Expected lora weight shape : [num_loras, num_experts, hidden_size, lora_rank]
-        assert len(lw_shape) == 4
-        assert lw_shape[-1] == lora_rank
-        hidden_size = lw_shape[-2]
-        # Expected output shape : [num_tokens, top_k_num, hidden_size * num_slices]
-        assert len(o_shape) == 3
-        assert o_shape == (num_tokens, ctx.top_k_num, hidden_size * num_slices)
-
-        kernel_config = get_lora_op_configs(
-            op_type.name.lower(),
-            max_loras=lw_shape[0],
-            batch=num_tokens,
-            hidden_size=hidden_size,
-            rank=lora_rank,
-            num_slices=num_slices,
-            add_inputs=False,
-        )
-
-        (topk_weights, sorted_token_ids, expert_ids, num_tokens_post_padded) = (
-            self.fused_moe_lora_data_prepare(
-                block_size=kernel_config["BLOCK_SIZE_M"],
-                token_lora_mapping=self.lora_kernel_meta.token_lora_mapping,
-                ctx=ctx,
-            )
-        )
-
-        return {
-            "a_intermediate_cache1": self.input,
-            "lora_b_stacked": self.lora_weights_lst,
-            "output": self.output,
-            "topk_weights": topk_weights,
-            "sorted_token_ids": sorted_token_ids,
-            "expert_ids": expert_ids,
-            "num_tokens_post_padded": num_tokens_post_padded,
-            "top_k_num": ctx.top_k_num,
-            "device": self.input.device,
-            "N": lora_rank,
-            "M": topk_weights.shape[0],
-            "EM": sorted_token_ids.shape[1],
-            "K": self.input.shape[1],
-            "num_tokens": num_tokens,
-            "num_experts": ctx.num_experts,
-            "num_slices": num_slices,
-            "max_lora_rank": lora_rank,
-            "w1_output_dim_size": lw_shape[2],
-            "expand_block_size_m": kernel_config["BLOCK_SIZE_M"],
-            "expand_block_size_n": kernel_config["BLOCK_SIZE_N"],
-            "expand_block_size_k": kernel_config["BLOCK_SIZE_K"],
-            "expand_group_size_m": kernel_config["GROUP_SIZE_M"],
-            "expand_num_warps": kernel_config["NUM_WARPS"],
-            "expand_num_stages": kernel_config["NUM_STAGES"],
-            "expand_split_k": kernel_config.get("SPLIT_K", 1),
-            "mul_routed_weight": op_type.is_fused_moe_lora_down_fn(),
-        }
-
     def bench_fn_kwargs(
-        self, ctx: BenchmarkContext, op_type: OpType, add_inputs: bool | None = None
+        self, op_type: OpType, add_inputs: bool | None = None
     ) -> dict[str, Any]:
-        if op_type.is_shrink_fn() or op_type.is_fused_moe_lora_fn():
+        if op_type.is_shrink_fn():
             assert add_inputs is None
         else:
             assert add_inputs is not None
 
         if op_type == OpType.LORA_SHRINK:
-            return self.as_lora_shrink_kwargs(ctx, op_type)
+            return self.as_lora_shrink_kwargs()
         if op_type == OpType.LORA_EXPAND:
-            return self.as_lora_expand_kwargs(ctx, op_type, add_inputs)
-        if op_type.is_fused_moe_lora_shrink_fn():
-            return self.as_fused_moe_lora_shrink_kwargs(ctx, op_type)
-        if op_type.is_fused_moe_lora_expand_fn():
-            return self.as_fused_moe_lora_expand_kwargs(ctx, op_type)
+            return self.as_lora_expand_kwargs(add_inputs)
         raise ValueError(f"Unrecognized optype {self}")
 
     def test_correctness(
@@ -996,7 +617,7 @@ def bench_optype(
     test_correctness: bool = False,
 ) -> TMeasurement:
     assert arg_pool_size >= 1
-    if op_type.is_shrink_fn() or op_type.is_fused_moe_lora_fn():
+    if op_type.is_shrink_fn():
         assert expand_fn_add_inputs is None
     else:
         assert expand_fn_add_inputs is not None
@@ -1006,30 +627,23 @@ def bench_optype(
         BenchmarkTensors.make(ctx, op_type) for _ in range(arg_pool_size)
     ]
     for bt in bench_tensors:
-        bt.sanity_check(ctx, op_type)
+        bt.sanity_check()
 
     # Test correctness of our implementation.
     if test_correctness:
-        assert op_type in [OpType.LORA_SHRINK, OpType.LORA_EXPAND], (
-            f"Correctness testing is not supported for {op_type.name}."
-        )
         assert all(
-            [
-                bt.test_correctness(ctx, op_type, expand_fn_add_inputs)
-                for bt in bench_tensors
-            ]
+            [bt.test_correctness(op_type, expand_fn_add_inputs) for bt in bench_tensors]
         )
 
     # BenchmarkTensors -> dict (kwargs)
     kwargs_list = [
-        bt.bench_fn_kwargs(ctx, op_type, add_inputs=expand_fn_add_inputs)
+        bt.bench_fn_kwargs(op_type, add_inputs=expand_fn_add_inputs)
         for bt in bench_tensors
     ]
 
     # Clear LoRA optimization hash-maps.
     _LORA_A_PTR_DICT.clear()
     _LORA_B_PTR_DICT.clear()
-    _LORA_PTR_DICT.clear()
     # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are set up
     for kwargs in kwargs_list:
         op_type.bench_fn()(**kwargs)
@@ -1179,9 +793,7 @@ def run(args: argparse.Namespace, bench_ctxs: list[BenchmarkContext]):
 
                     # Benchmark bench_op
                     expand_fn_add_inputs = (
-                        [None]
-                        if bench_op.is_shrink_fn() or bench_op.is_fused_moe_lora_fn()
-                        else args.expand_fn_add_inputs
+                        [None] if bench_op.is_shrink_fn() else args.expand_fn_add_inputs
                     )
                     for add_input_arg in expand_fn_add_inputs:
                         seq_len_timers.append(
@@ -1219,22 +831,12 @@ def as_benchmark_contexts(
     hidden_sizes: list[int], lora_ranks: list[int], args: argparse.Namespace
 ) -> list[BenchmarkContext]:
     ctxs: list[BenchmarkContext] = []
-    for (
-        batch_size,
-        hidden_size,
-        lora_rank,
-        num_loras,
-        sort_by_lora_id,
-        top_k_num,
-        num_experts,
-    ) in product(  # noqa
+    for batch_size, hidden_size, lora_rank, num_loras, sort_by_lora_id in product(  # noqa
         args.batch_sizes,
         list(hidden_sizes),
         lora_ranks,
         args.num_loras,
         args.sort_by_lora_id,
-        args.top_k_nums,
-        args.num_experts,
     ):
         ctxs.append(
             BenchmarkContext(
@@ -1249,8 +851,6 @@ def as_benchmark_contexts(
                 seq_length=None,
                 sort_by_lora_id=sort_by_lora_id,
                 dtype=args.dtype,
-                top_k_num=top_k_num,
-                num_experts=num_experts,
                 # To be filled based on the OpType to benchmark
                 num_slices=None,
             )
@@ -1412,22 +1012,6 @@ if __name__ == "__main__":
             ),
         )
 
-        p.add_argument(
-            "--top-k-nums",
-            nargs="+",
-            type=int,
-            default=DEFAULT_TOP_K_NUMS,
-            help="Top-K values for MoE LoRA operations",
-        )
-
-        p.add_argument(
-            "--num-experts",
-            nargs="+",
-            type=int,
-            default=DEFAULT_NUM_EXPERTS,
-            help="Number of experts for MoE LoRA operations",
-        )
-
     parser = FlexibleArgumentParser(
         description=f"""
 Benchmark LoRA kernels:

benchmarks/kernels/benchmark_moe.py

@@ -211,7 +211,7 @@ def get_rocm_tuning_space(use_fp16):
     num_warps_range = [1, 2, 4, 8]
     group_m_range = [1, 4, 8, 16, 32]
     num_stage_range = [2]
-    waves_per_eu_range = [0, 1, 2, 4]
+    waves_per_eu_range = [0]
     matrix_instr_nonkdim_range = [16, 32] if use_fp16 else []
     kpack_range = [1, 2] if use_fp16 else []
 
@@ -590,7 +590,6 @@ def main(args: argparse.Namespace):
         "DeepseekV3ForCausalLM",
         "DeepseekV32ForCausalLM",
         "Glm4MoeForCausalLM",
-        "NemotronHForCausalLM",
     ):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
@@ -616,11 +615,6 @@ def main(args: argparse.Namespace):
         topk = config.moe_topk[0]
         intermediate_size = config.moe_intermediate_size[0]
         hidden_size = config.hidden_size
-    elif config.architectures[0] in ["Qwen3OmniMoeForConditionalGeneration"]:
-        E = config.thinker_config.text_config.num_experts
-        topk = config.thinker_config.text_config.num_experts_per_tok
-        intermediate_size = config.thinker_config.text_config.moe_intermediate_size
-        hidden_size = config.thinker_config.text_config.hidden_size
     else:
         # Support for llama4
         config = config.get_text_config()

benchmarks/kernels/benchmark_rope.py

@@ -1,76 +1,97 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
-import itertools
+from itertools import accumulate
 
+import nvtx
 import torch
 
-from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.triton_utils import triton
+from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding, get_rope
+from vllm.platforms import current_platform
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 
-batch_size_range = [2**i for i in range(0, 8, 2)]
-seq_len_range = [2**i for i in range(6, 10, 1)]
-num_heads_range = [32, 48]
-configs = list(itertools.product(batch_size_range, seq_len_range, num_heads_range))
 
+def benchmark_rope_kernels_multi_lora(
+    is_neox_style: bool,
+    batch_size: int,
+    seq_len: int,
+    num_heads: int,
+    head_size: int,
+    rotary_dim: int | None,
+    dtype: torch.dtype,
+    seed: int,
+    device: str,
+    max_position: int = 8192,
+    base: float = 10000,
+) -> None:
+    current_platform.seed_everything(seed)
+    torch.set_default_device(device)
+    if rotary_dim is None:
+        rotary_dim = head_size
+    # silulating serving 4 LoRAs
+    scaling_factors = [1, 2, 4, 8]
+    # batched RoPE can take multiple scaling factors
+    batched_rope = get_rope(
+        head_size,
+        rotary_dim,
+        max_position,
+        base,
+        is_neox_style,
+        {"rope_type": "linear", "factor": tuple(scaling_factors)},
+    )
+    # non-batched RoPE takes only one scaling factor, we create multiple
+    # instances to simulate the same behavior
+    non_batched_ropes: list[RotaryEmbedding] = []
+    for scaling_factor in scaling_factors:
+        non_batched_ropes.append(
+            get_rope(
+                head_size,
+                rotary_dim,
+                max_position,
+                base,
+                is_neox_style,
+                {"rope_type": "linear", "factor": (scaling_factor,)},
+            )
+        )
 
-def get_benchmark(head_size, rotary_dim, is_neox_style, device):
-    @triton.testing.perf_report(
-        triton.testing.Benchmark(
-            x_names=["batch_size", "seq_len", "num_heads"],
-            x_vals=[list(_) for _ in configs],
-            line_arg="provider",
-            line_vals=["torch", "flashinfer", "vllm"],
-            line_names=["PyTorch", "FlashInfer", "vLLM"],
-            styles=[("blue", "-"), ("green", "-"), ("red", "-")],
-            ylabel="us",
-            plot_name=f"rope-perf{'-neox-style' if is_neox_style else ''}",
-            args={},
+    positions = torch.randint(0, max_position, (batch_size, seq_len))
+    query = torch.randn(batch_size, seq_len, num_heads * head_size, dtype=dtype)
+    key = torch.randn_like(query)
+
+    # create query offsets for batched RoPE, we concat multiple kv cache
+    # together and each query needs to find the right kv cache of its type
+    offset_map = torch.tensor(
+        list(
+            accumulate(
+                [0]
+                + [
+                    max_position * scaling_factor * 2
+                    for scaling_factor in scaling_factors[:-1]
+                ]
+            )
         )
     )
-    def benchmark(batch_size, seq_len, num_heads, provider):
-        dtype = torch.bfloat16
-        max_position = 8192
-        base = 10000
-        rope = get_rope(head_size, rotary_dim, max_position, base, is_neox_style)
-        rope = rope.to(dtype=dtype, device=device)
-        cos_sin_cache = rope.cos_sin_cache.to(dtype=torch.float, device=device)
+    query_types = torch.randint(
+        0, len(scaling_factors), (batch_size, seq_len), device=device
+    )
+    # map query types to offsets
+    query_offsets = offset_map[query_types]
+    # the kernel takes flattened offsets
+    flatten_offsets = query_offsets.flatten()
 
-        positions = torch.randint(0, max_position, (batch_size, seq_len), device=device)
-        query = torch.randn(
-            (batch_size, seq_len, num_heads * head_size), dtype=dtype, device=device
-        )
-        key = torch.randn_like(query)
-
-        quantiles = [0.5, 0.2, 0.8]
-
-        if provider == "torch":
-            ms, min_ms, max_ms = triton.testing.do_bench(
-                lambda: rope.forward_native(positions, query.clone(), key.clone()),
-                quantiles=quantiles,
-            )
-        elif provider == "flashinfer":
-            ms, min_ms, max_ms = triton.testing.do_bench(
-                lambda: torch.ops.vllm.flashinfer_rotary_embedding(
-                    positions,
-                    query.clone(),
-                    key.clone(),
-                    head_size,
-                    cos_sin_cache,
-                    is_neox_style,
-                ),
-                quantiles=quantiles,
-            )
-        else:
-            ms, min_ms, max_ms = triton.testing.do_bench(
-                lambda: rope.forward_cuda(positions, query.clone(), key.clone()),
-                quantiles=quantiles,
-            )
-
-        return 1000 * ms, 1000 * max_ms, 1000 * min_ms
-
-    return benchmark
+    # batched queries of the same type together for non-batched RoPE
+    queries = [query[query_types == i] for i in range(len(scaling_factors))]
+    keys = [key[query_types == i] for i in range(len(scaling_factors))]
+    packed_qkr = zip(queries, keys, non_batched_ropes)
+    # synchronize before start timing
+    torch.cuda.synchronize()
+    with nvtx.annotate("non-batched", color="yellow"):
+        for q, k, r in packed_qkr:
+            r.forward(positions, q, k)
+    torch.cuda.synchronize()
+    with nvtx.annotate("batched", color="green"):
+        batched_rope.forward(positions, query, key, flatten_offsets)
+    torch.cuda.synchronize()
 
 
 if __name__ == "__main__":
@@ -95,12 +116,17 @@ if __name__ == "__main__":
     parser.add_argument(
         "--device", type=str, choices=["cuda:0", "cuda:1"], default="cuda:0"
     )
-    parser.add_argument("--save-path", type=str, default="./configs/rope/")
     args = parser.parse_args()
+    print(args)
 
-    # Get the benchmark function
-    benchmark = get_benchmark(
-        args.head_size, args.rotary_dim, args.is_neox_style, args.device
+    benchmark_rope_kernels_multi_lora(
+        is_neox_style=args.is_neox_style,
+        batch_size=args.batch_size,
+        seq_len=args.seq_len,
+        num_heads=args.num_heads,
+        head_size=args.head_size,
+        rotary_dim=args.rotary_dim,
+        dtype=getattr(torch, args.dtype),
+        seed=args.seed,
+        device=args.device,
     )
-    # Run performance benchmark
-    benchmark.run(print_data=True, save_path=args.save_path)

benchmarks/kernels/benchmark_shapes.py

@@ -78,11 +78,11 @@ WEIGHT_SHAPES = {
 }
 
 WEIGHT_SHAPES_MOE = {
-    "mistralai/Mixtral-8x7B-Instruct-v0.1": [
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1": [
         [8, 2, 4096, 28672],
         [8, 2, 14336, 4096],
     ],
-    "deepseek-ai/DeepSeek-V2-Lite": [
+    "nm-testing/deepseekv2-lite": [
         [64, 6, 2048, 1408],
     ],
     "ibm-granite/granite-3.0-1b-a400m": [

benchmarks/multi_turn/bench_dataset.py

@@ -11,7 +11,6 @@ from bench_utils import (
     Color,
     logger,
 )
-from tqdm import tqdm
 from transformers import AutoTokenizer  # type: ignore
 
 # Conversation ID is a string (e.g: "UzTK34D")
@@ -418,10 +417,6 @@ def generate_conversations(
             data = file.read()
             tokens_in_file = tokenizer.encode(data, add_special_tokens=False)
             list_of_tokens.extend(tokens_in_file)
-        logger.info(
-            f"Loaded {len(tokens_in_file)} tokens from file {filename}, "
-            f"total tokens so far: {len(list_of_tokens)}"
-        )
 
     conversations: ConversationsMap = {}
     conv_id = 0
@@ -454,25 +449,18 @@ def generate_conversations(
         )
         base_offset += common_prefix_tokens
 
-    for conv_id in tqdm(
-        range(args.num_conversations),
-        total=args.num_conversations,
-        desc="Generating conversations",
-        unit="conv",
-    ):
+    for conv_id in range(args.num_conversations):
         # Generate a single conversation
         messages: MessagesList = []
 
         nturns = turn_count[conv_id]
 
         # User prompt token count per turn (with lower limit)
-        input_token_count: np.ndarray = args.input_num_tokens.sample(nturns).astype(int)
+        input_token_count: np.ndarray = args.input_num_tokens.sample(nturns)
         input_token_count = np.maximum(input_token_count, base_prompt_token_count)
 
         # Assistant answer token count per turn (with lower limit)
-        output_token_count: np.ndarray = args.output_num_tokens.sample(nturns).astype(
-            int
-        )
+        output_token_count: np.ndarray = args.output_num_tokens.sample(nturns)
         output_token_count = np.maximum(output_token_count, 1)
 
         user_turn = True

benchmarks/multi_turn/benchmark_serving_multi_turn.py

@@ -55,7 +55,6 @@ class ClientArgs(NamedTuple):
     verify_output: bool
     conversation_sampling: ConversationSampling
     request_rate: float
-    max_retries: int
 
 
 class RequestArgs(NamedTuple):
@@ -64,7 +63,6 @@ class RequestArgs(NamedTuple):
     stream: bool
     limit_min_tokens: int  # Use negative value for no limit
     limit_max_tokens: int  # Use negative value for no limit
-    timeout_sec: int
 
 
 class BenchmarkArgs(NamedTuple):
@@ -216,7 +214,6 @@ async def send_request(
     stream: bool = True,
     min_tokens: int | None = None,
     max_tokens: int | None = None,
-    timeout_sec: int = 120,
 ) -> ServerResponse:
     payload = {
         "model": model,
@@ -238,16 +235,10 @@ async def send_request(
     headers = {"Content-Type": "application/json"}
 
     # Calculate the timeout for the request
+    timeout_sec = 120
     if max_tokens is not None:
         # Assume TPOT of 200ms and use max_tokens to determine timeout
-        token_based_timeout = int(max_tokens * 0.2)
-        if token_based_timeout > timeout_sec:
-            timeout_sec = token_based_timeout
-            logger.info(
-                "Using timeout of %ds based on max_tokens %d",
-                timeout_sec,
-                max_tokens,
-            )
+        timeout_sec = max(timeout_sec, int(max_tokens * 0.2))
     timeout = aiohttp.ClientTimeout(total=timeout_sec)
 
     valid_response = True
@@ -418,7 +409,6 @@ async def send_turn(
         req_args.stream,
         min_tokens,
         max_tokens,
-        req_args.timeout_sec,
     )
 
     if response.valid is False:
@@ -528,25 +518,6 @@ async def poisson_sleep(request_rate: float, verbose: bool = False) -> None:
     await asyncio.sleep(interval)
 
 
-async def exponential_backoff_sleep(
-    attempt_cnt: int,
-    base_rate: float = 1.0,
-    backoff_factor: float = 2.0,
-    jitter_fraction: float = 0.10,
-    verbose: bool = False,
-) -> None:
-    # Sleep with exponential backoff and jitter after a failed request.
-    backoff_delay = base_rate * (backoff_factor**attempt_cnt)
-    jittered_delay = backoff_delay * (
-        1 + np.random.uniform(-jitter_fraction, jitter_fraction)
-    )
-
-    if verbose:
-        logger.info(f"Backoff for {jittered_delay:.3f} seconds...")
-
-    await asyncio.sleep(jittered_delay)
-
-
 async def client_main(
     args: ClientArgs,
     req_args: RequestArgs,
@@ -561,11 +532,8 @@ async def client_main(
         f"{Color.CYAN}Started client {client_id}: max_num_requests={args.max_num_requests}, max_active_conversations={args.max_active_conversations}{Color.RESET}"  # noqa: E501
     )
 
-    # Set unique seed per client (each client runs in its own process)
-    # Add 1 to ensure no client uses the same seed as the main process
-    client_seed = args.seed + client_id + 1
-    random.seed(client_seed)
-    np.random.seed(client_seed)
+    random.seed(args.seed)
+    np.random.seed(args.seed)
 
     # Active conversations
     active_convs: ConversationsMap = {}
@@ -678,62 +646,49 @@ async def client_main(
                 )
                 time_of_last_turn[conv_id] = curr_time_sec
 
-            success = False
-            for attempt_cnt in range(args.max_retries + 1):
-                try:
-                    exception = False
-                    result = await send_turn(
-                        session,
-                        client_id,
-                        conv_id,
-                        messages,
-                        current_turn,
-                        tokenizer,
-                        req_args,
-                        args.print_content,
-                        args.verify_output,
-                    )
-                    if result is not None:
-                        result_queue.put(result)
-                        success = True
-                        break
-                    else:
-                        logger.warning(
-                            f"{Color.YELLOW}Client {client_id} - Request rejected during conversation ID {conv_id} (turn: {current_turn}){Color.RESET}"  # noqa: E501
-                        )
-                except asyncio.exceptions.TimeoutError:
-                    exception = True
-                    logger.error(
-                        "%sClient %d - Timeout during conversation ID %s (turn: %d). "
-                        "Base timeout is %ss (set with --request-timeout-sec), but the "
-                        "effective timeout may be longer based on max_tokens. If this "
-                        "is unexpected, consider increasing the timeout or checking "
-                        "model performance.%s",
-                        Color.RED,
-                        client_id,
-                        conv_id,
-                        current_turn,
-                        req_args.timeout_sec,
-                        Color.RESET,
-                    )
-                except Exception:
-                    exception = True
-                    logger.exception(
-                        f"{Color.RED}Client {client_id} - Exception during conversation ID {conv_id} (turn: {current_turn}){Color.RESET}"  # noqa: E501
+            success = True
+            try:
+                result = await send_turn(
+                    session,
+                    client_id,
+                    conv_id,
+                    messages,
+                    current_turn,
+                    tokenizer,
+                    req_args,
+                    args.print_content,
+                    args.verify_output,
+                )
+                if result is not None:
+                    result_queue.put(result)
+                else:
+                    # None means that the request failed,
+                    # and should not be added to the statistics.
+                    success = False
+                    num_failures += 1
+
+                    logger.warning(
+                        f"{Color.YELLOW}Client {client_id} - Request rejected during conversation ID {conv_id} (turn: {current_turn}){Color.RESET}"  # noqa: E501
                     )
 
-                # Sleep before retry if not last attempt
-                if not success and attempt_cnt < args.max_retries:
-                    await exponential_backoff_sleep(attempt_cnt, verbose=args.verbose)
+                    # Remove the conversation (should not be used again)
+                    active_convs.pop(conv_id)
 
-            if not success:
+            except asyncio.exceptions.TimeoutError:
                 num_failures += 1
-                # Remove the conversation (should not be used again)
-                active_convs.pop(conv_id)
-                if exception:
-                    break  # Exit gracefully instead of raising an error
+                logger.exception(
+                    f"{Color.RED}Client {client_id} - Timeout during conversation ID {conv_id} (turn: {current_turn}){Color.RESET}"  # noqa: E501
+                )
+                break  # Exit gracefully instead of raising an error
 
-            else:
+            except Exception:
+                num_failures += 1
+                logger.exception(
+                    f"{Color.RED}Client {client_id} - Exception during conversation ID {conv_id} (turn: {current_turn}){Color.RESET}"  # noqa: E501
+                )
+                break  # Exit gracefully instead of raising an error
+
+            if success:
                 num_successes += 1
 
                 # Update the turns counter to include the LLM response
@@ -848,7 +803,6 @@ def get_client_config(
         verify_output=args.verify_output,
         conversation_sampling=args.conversation_sampling,
         request_rate=args.request_rate,
-        max_retries=args.max_retries,
     )
 
     if args.limit_min_tokens > 0 or args.limit_max_tokens > 0:
@@ -861,9 +815,6 @@ def get_client_config(
                 "Invalid min/max tokens limits (min should not be larger than max)"
             )
 
-    if args.request_timeout_sec <= 0:
-        raise ValueError("Request timeout must be a positive number")
-
     # Arguments for API requests
     chat_url = f"{args.url}/v1/chat/completions"
     model_name = args.served_model_name if args.served_model_name else args.model
@@ -874,7 +825,6 @@ def get_client_config(
         stream=not args.no_stream,
         limit_min_tokens=args.limit_min_tokens,
         limit_max_tokens=args.limit_max_tokens,
-        timeout_sec=args.request_timeout_sec,
     )
 
     return client_args, req_args
@@ -1018,7 +968,7 @@ async def main_mp(
             f"(is alive: {client.is_alive()}){Color.RESET}"
         )
 
-        client.join(timeout=req_args.timeout_sec + 1)
+        client.join(timeout=120)
 
         if client.is_alive():
             logger.warning(
@@ -1384,16 +1334,6 @@ async def main() -> None:
         help="Expected request rate (Poisson process) per client in requests/sec."
         "Set to 0 for no delay between requests.",
     )
-    parser.add_argument(
-        "--max-retries",
-        type=int,
-        default=int(os.environ.get("MULTITURN_BENCH_MAX_RETRIES", "0")),
-        help="Maximum number of retry attempts for timed-out requests. "
-        "Default is 0 (no retries). "
-        "Set to higher values to retry failed requests and maintain "
-        "fair workload distribution. "
-        "Can also be set via MULTITURN_BENCH_MAX_RETRIES environment variable.",
-    )
     parser.add_argument(
         "--conversation-sampling",
         type=ConversationSampling,
@@ -1411,13 +1351,6 @@ async def main() -> None:
         action="store_true",
         help="Verify the LLM output (compare to the answers in the input JSON file)",
     )
-    parser.add_argument(
-        "--request-timeout-sec",
-        type=int,
-        default=120,
-        help="Timeout in seconds for each API request (default: 120). "
-        "Automatically increased if max tokens imply longer decoding.",
-    )
 
     parser.add_argument(
         "--no-stream",
@@ -1493,10 +1426,11 @@ async def main() -> None:
             f"Invalid --warmup-percentage={args.warmup_percentage}"
         ) from None
 
-    # Set global seeds for main process
     random.seed(args.seed)
     np.random.seed(args.seed)
 
+    if not os.path.exists(args.model):
+        raise OSError(f"Path does not exist: {args.model}")
     logger.info("Loading tokenizer")
     tokenizer = AutoTokenizer.from_pretrained(args.model)
 

benchmarks/multi_turn/requirements.txt

@@ -2,5 +2,4 @@ numpy>=1.24
 pandas>=2.0.0
 aiohttp>=3.10
 transformers>=4.46
-xlsxwriter>=3.2.1
-tqdm>=4.66
+xlsxwriter>=3.2.1

cmake/cpu_extension.cmake

@@ -15,7 +15,6 @@ endif()
 #
 set(ENABLE_AVX512BF16 $ENV{VLLM_CPU_AVX512BF16})
 set(ENABLE_AVX512VNNI $ENV{VLLM_CPU_AVX512VNNI})
-set(ENABLE_AMXBF16 $ENV{VLLM_CPU_AMXBF16})
 
 include_directories("${CMAKE_SOURCE_DIR}/csrc")
 
@@ -141,22 +140,6 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
         set(ENABLE_AVX512VNNI OFF)
         message(WARNING "Disable AVX512-VNNI ISA support, no avx512_vnni found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512VNNI=1.")
     endif()
-
-    find_isa(${CPUINFO} "amx_bf16" AMXBF16_FOUND)
-    if (AMXBF16_FOUND OR ENABLE_AMXBF16)
-        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
-            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
-            list(APPEND CXX_COMPILE_FLAGS "-mamx-bf16" "-mamx-tile")
-            set(ENABLE_AMXBF16 ON)
-            add_compile_definitions(-DCPU_CAPABILITY_AMXBF16)
-        else()
-            set(ENABLE_AMXBF16 OFF)
-            message(WARNING "Disable AMX_BF16 ISA support, requires gcc/g++ >= 12.3")
-        endif()
-    else()
-        set(ENABLE_AMXBF16 OFF)
-        message(WARNING "Disable AMX_BF16 ISA support, no amx_bf16 found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AMXBF16=1.")
-    endif()
     
 elseif (AVX2_FOUND)
     list(APPEND CXX_COMPILE_FLAGS "-mavx2")
@@ -210,30 +193,7 @@ endif()
 if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
     # Fetch and build Arm Compute Library (ACL) as oneDNN's backend for AArch64
     # TODO [fadara01]: remove this once ACL can be fetched and built automatically as a dependency of oneDNN
-    set(ONEDNN_AARCH64_USE_ACL OFF CACHE BOOL "")
     if(ASIMD_FOUND)
-        # Set number of parallel build processes
-        include(ProcessorCount)
-        ProcessorCount(NPROC)
-        if(NOT NPROC)
-            set(NPROC 4)
-        endif()
-        # locate PyTorch's libgomp (e.g. site-packages/torch.libs/libgomp-947d5fa1.so.1.0.0)
-        # and create a local shim dir with it
-        vllm_prepare_torch_gomp_shim(VLLM_TORCH_GOMP_SHIM_DIR)
-
-        find_library(OPEN_MP
-            NAMES gomp
-            PATHS ${VLLM_TORCH_GOMP_SHIM_DIR}
-            NO_DEFAULT_PATH
-            REQUIRED
-        )
-        # Set LD_LIBRARY_PATH to include the shim dir at build time to use the same libgomp as PyTorch
-        if (OPEN_MP)
-            set(ENV{LD_LIBRARY_PATH} "${VLLM_TORCH_GOMP_SHIM_DIR}:$ENV{LD_LIBRARY_PATH}")
-        endif()
-
-        # Fetch and populate ACL
         if(DEFINED ENV{ACL_ROOT_DIR} AND IS_DIRECTORY "$ENV{ACL_ROOT_DIR}")
             message(STATUS "Using ACL from specified source directory: $ENV{ACL_ROOT_DIR}")
         else()
@@ -242,58 +202,43 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
                 SUBBUILD_DIR "${FETCHCONTENT_BASE_DIR}/arm_compute-subbuild"
                 SOURCE_DIR   "${FETCHCONTENT_BASE_DIR}/arm_compute-src"
                 GIT_REPOSITORY https://github.com/ARM-software/ComputeLibrary.git
-                GIT_TAG        v52.6.0
+                GIT_TAG        v52.2.0
                 GIT_SHALLOW    TRUE
                 GIT_PROGRESS   TRUE
             )
             set(ENV{ACL_ROOT_DIR} "${arm_compute_SOURCE_DIR}")
-            set(ACL_LIB_DIR "$ENV{ACL_ROOT_DIR}/build")
         endif()
 
-        # Build ACL with CMake
-        set(ARM_COMPUTE_BUILD_SHARED_LIB "OFF")
-        set(CMAKE_BUILD_TYPE "Release")
-        set(ARM_COMPUTE_ARCH "armv8.2-a")
-        set(ARM_COMPUTE_ENABLE_ASSERTS "OFF")
-        set(ARM_COMPUTE_ENABLE_CPPTHREADS "OFF")
-        set(ONEDNN_ENABLE_PRIMITIVE "MATMUL;REORDER")
-        set(ARM_COMPUTE_ENABLE_OPENMP "ON")
-        set(ARM_COMPUTE_ENABLE_WERROR "OFF")
-        set(ARM_COMPUTE_BUILD_EXAMPLES "OFF")
-        set(ARM_COMPUTE_BUILD_TESTING "OFF")
-
-        set(_cmake_config_cmd
-             ${CMAKE_COMMAND} -G Ninja -B build 
-            -DARM_COMPUTE_BUILD_SHARED_LIB=OFF 
-            -DCMAKE_BUILD_TYPE=Release 
-            -DARM_COMPUTE_ARCH=armv8.2-a 
-            -DARM_COMPUTE_ENABLE_ASSERTS=OFF 
-            -DARM_COMPUTE_ENABLE_CPPTHREADS=OFF 
-            -DARM_COMPUTE_ENABLE_OPENMP=ON 
-            -DARM_COMPUTE_ENABLE_WERROR=OFF 
-            -DARM_COMPUTE_BUILD_EXAMPLES=OFF 
-            -DARM_COMPUTE_BUILD_TESTING=OFF)
-        set(_cmake_build_cmd
-            ${CMAKE_COMMAND} --build build -- -j${NPROC}
+        # Build ACL with scons
+        include(ProcessorCount)
+        ProcessorCount(_NPROC)
+        set(_scons_cmd
+        scons -j${_NPROC}
+            Werror=0 debug=0 neon=1 examples=0 embed_kernels=0 os=linux
+            arch=armv8.2-a build=native benchmark_examples=0 fixed_format_kernels=1
+            multi_isa=1 openmp=1 cppthreads=0
         )
 
+        # locate PyTorch's libgomp (e.g. site-packages/torch.libs/libgomp-947d5fa1.so.1.0.0)
+        # and create a local shim dir with it
+        include("${CMAKE_CURRENT_LIST_DIR}/utils.cmake")
+        vllm_prepare_torch_gomp_shim(VLLM_TORCH_GOMP_SHIM_DIR)
+
+        if(NOT VLLM_TORCH_GOMP_SHIM_DIR STREQUAL "")
+            list(APPEND _scons_cmd extra_link_flags=-L${VLLM_TORCH_GOMP_SHIM_DIR})
+        endif()
+
         execute_process(
-            COMMAND ${_cmake_config_cmd}
-            WORKING_DIRECTORY "$ENV{ACL_ROOT_DIR}"
-        )
-        execute_process(
-            COMMAND ${_cmake_build_cmd}
+            COMMAND ${_scons_cmd}
             WORKING_DIRECTORY "$ENV{ACL_ROOT_DIR}"
             RESULT_VARIABLE _acl_rc
         )
-
         if(NOT _acl_rc EQUAL 0)
             message(FATAL_ERROR "ACL SCons build failed (exit ${_acl_rc}).")
         endif()
-        message(STATUS "Arm Compute Library (ACL) built successfully.")
 
-        # VLLM/oneDNN settings for ACL
-        set(ONEDNN_AARCH64_USE_ACL ON CACHE BOOL "" FORCE)
+        set(ONEDNN_AARCH64_USE_ACL "ON")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wl,-rpath,$ENV{ACL_ROOT_DIR}/build/")
         add_compile_definitions(VLLM_USE_ACL)
     endif()
 
@@ -310,7 +255,7 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
         FetchContent_Declare(
             oneDNN
             GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-            GIT_TAG v3.10
+            GIT_TAG v3.9
             GIT_PROGRESS TRUE
             GIT_SHALLOW TRUE
         )
@@ -330,10 +275,7 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
     set(ONEDNN_VERBOSE "OFF")
     set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
 
-    set(VLLM_BUILD_TYPE ${CMAKE_BUILD_TYPE})
-    set(CMAKE_BUILD_TYPE "Release") # remove oneDNN debug symbols to reduce size
     FetchContent_MakeAvailable(oneDNN)
-    set(CMAKE_BUILD_TYPE ${VLLM_BUILD_TYPE})
     add_library(dnnl_ext OBJECT "csrc/cpu/dnnl_helper.cpp")
     target_include_directories(
         dnnl_ext
@@ -363,14 +305,14 @@ endif()
 #
 set(VLLM_EXT_SRC
     "csrc/cpu/activation.cpp"
+    "csrc/cpu/attention.cpp"
+    "csrc/cpu/cache.cpp"
     "csrc/cpu/utils.cpp"
     "csrc/cpu/layernorm.cpp"
     "csrc/cpu/mla_decode.cpp"
     "csrc/cpu/pos_encoding.cpp"
-    "csrc/moe/dynamic_4bit_int_moe_cpu.cpp"
-    "csrc/cpu/cpu_attn.cpp"
-    "csrc/cpu/scratchpad_manager.cpp"
-    "csrc/cpu/torch_bindings.cpp")
+    "csrc/cpu/torch_bindings.cpp"
+    "csrc/moe/dynamic_4bit_int_moe_cpu.cpp")
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     set(VLLM_EXT_SRC
@@ -401,7 +343,7 @@ message(STATUS "CPU extension source files: ${VLLM_EXT_SRC}")
 # Define extension targets
 #
 
-define_extension_target(
+define_gpu_extension_target(
     _C
     DESTINATION vllm
     LANGUAGE CXX
@@ -412,4 +354,4 @@ define_extension_target(
     WITH_SOABI
 )
 
-message(STATUS "Enabling C extension.")
+message(STATUS "Enabling C extension.")

cmake/external_projects/flashmla.cmake

@@ -92,7 +92,7 @@ if(FLASH_MLA_ARCHS)
         SRCS "${FlashMLA_Extension_SOURCES}"
         CUDA_ARCHS "${FLASH_MLA_ARCHS}")
 
-    define_extension_target(
+    define_gpu_extension_target(
         _flashmla_C
         DESTINATION vllm
         LANGUAGE ${VLLM_GPU_LANG}
@@ -109,7 +109,7 @@ if(FLASH_MLA_ARCHS)
         $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
         $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
 
-    define_extension_target(
+    define_gpu_extension_target(
         _flashmla_extension_C
         DESTINATION vllm
         LANGUAGE ${VLLM_GPU_LANG}

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 58e0626a692f09241182582659e3bf8f16472659
+          GIT_TAG a893712401d70362fbb299cd9c4b3476e8e9ed54
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/utils.cmake

@@ -453,20 +453,21 @@ macro(override_gpu_arches GPU_ARCHES GPU_LANG GPU_SUPPORTED_ARCHES)
 endmacro()
 
 #
-# Define a target named `MOD_NAME` for a single extension. The
+# Define a target named `GPU_MOD_NAME` for a single extension. The
 # arguments are:
 #
 # DESTINATION <dest>         - Module destination directory.
-# LANGUAGE <lang>            - The language for this module, e.g. CUDA, HIP,
-#                              CXX, etc.
+# LANGUAGE <lang>            - The GPU language for this module, e.g CUDA, HIP,
+#                              etc.
 # SOURCES <sources>          - List of source files relative to CMakeLists.txt
 #                              directory.
 #
 # Optional arguments:
 #
-# ARCHITECTURES <arches>     - A list of target architectures in cmake format.
-#                              For GPU, refer to CMAKE_CUDA_ARCHITECTURES and
-#                              CMAKE_HIP_ARCHITECTURES for more info.
+# ARCHITECTURES <arches>     - A list of target GPU architectures in cmake
+#                              format.
+#                              Refer `CMAKE_CUDA_ARCHITECTURES` documentation
+#                              and `CMAKE_HIP_ARCHITECTURES` for more info.
 #                              ARCHITECTURES will use cmake's defaults if
 #                              not provided.
 # COMPILE_FLAGS <flags>      - Extra compiler flags passed to NVCC/hip.
@@ -477,61 +478,63 @@ endmacro()
 #
 # Note: optimization level/debug info is set via cmake build type.
 #
-function (define_extension_target MOD_NAME)
+function (define_gpu_extension_target GPU_MOD_NAME)
   cmake_parse_arguments(PARSE_ARGV 1
-    ARG
+    GPU
     "WITH_SOABI"
     "DESTINATION;LANGUAGE;USE_SABI"
     "SOURCES;ARCHITECTURES;COMPILE_FLAGS;INCLUDE_DIRECTORIES;LIBRARIES")
 
   # Add hipify preprocessing step when building with HIP/ROCm.
-  if (ARG_LANGUAGE STREQUAL "HIP")
-    hipify_sources_target(ARG_SOURCES ${MOD_NAME} "${ARG_SOURCES}")
+  if (GPU_LANGUAGE STREQUAL "HIP")
+    hipify_sources_target(GPU_SOURCES ${GPU_MOD_NAME} "${GPU_SOURCES}")
   endif()
 
-  if (ARG_WITH_SOABI)
-    set(SOABI_KEYWORD WITH_SOABI)
+  if (GPU_WITH_SOABI)
+    set(GPU_WITH_SOABI WITH_SOABI)
   else()
-    set(SOABI_KEYWORD "")
+    set(GPU_WITH_SOABI)
   endif()
 
-  if (ARG_USE_SABI)
-    Python_add_library(${MOD_NAME} MODULE USE_SABI ${ARG_USE_SABI} ${SOABI_KEYWORD} "${ARG_SOURCES}")
+  if (GPU_USE_SABI)
+    Python_add_library(${GPU_MOD_NAME} MODULE USE_SABI ${GPU_USE_SABI} ${GPU_WITH_SOABI} "${GPU_SOURCES}")
   else()
-    Python_add_library(${MOD_NAME} MODULE ${SOABI_KEYWORD} "${ARG_SOURCES}")
+    Python_add_library(${GPU_MOD_NAME} MODULE ${GPU_WITH_SOABI} "${GPU_SOURCES}")
   endif()
 
-  if (ARG_LANGUAGE STREQUAL "HIP")
+  if (GPU_LANGUAGE STREQUAL "HIP")
     # Make this target dependent on the hipify preprocessor step.
-    add_dependencies(${MOD_NAME} hipify${MOD_NAME})
+    add_dependencies(${GPU_MOD_NAME} hipify${GPU_MOD_NAME})
     # Make sure we include the hipified versions of the headers, and avoid conflicts with the ones in the original source folder
-    target_include_directories(${MOD_NAME} PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/csrc
-      ${ARG_INCLUDE_DIRECTORIES})
+    target_include_directories(${GPU_MOD_NAME} PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/csrc
+      ${GPU_INCLUDE_DIRECTORIES})
   else()
-    target_include_directories(${MOD_NAME} PRIVATE csrc
-      ${ARG_INCLUDE_DIRECTORIES})
+    target_include_directories(${GPU_MOD_NAME} PRIVATE csrc
+      ${GPU_INCLUDE_DIRECTORIES})
   endif()
 
-  if (ARG_ARCHITECTURES)
-    set_target_properties(${MOD_NAME} PROPERTIES
-      ${ARG_LANGUAGE}_ARCHITECTURES "${ARG_ARCHITECTURES}")
+  if (GPU_ARCHITECTURES)
+    set_target_properties(${GPU_MOD_NAME} PROPERTIES
+      ${GPU_LANGUAGE}_ARCHITECTURES "${GPU_ARCHITECTURES}")
   endif()
 
-  target_compile_options(${MOD_NAME} PRIVATE
-    $<$<COMPILE_LANGUAGE:${ARG_LANGUAGE}>:${ARG_COMPILE_FLAGS}>)
 
-  target_compile_definitions(${MOD_NAME} PRIVATE
-    "-DTORCH_EXTENSION_NAME=${MOD_NAME}")
+  target_compile_options(${GPU_MOD_NAME} PRIVATE
+    $<$<COMPILE_LANGUAGE:${GPU_LANGUAGE}>:${GPU_COMPILE_FLAGS}>)
 
-  target_link_libraries(${MOD_NAME} PRIVATE torch ${ARG_LIBRARIES})
+  target_compile_definitions(${GPU_MOD_NAME} PRIVATE
+    "-DTORCH_EXTENSION_NAME=${GPU_MOD_NAME}")
+
+
+  target_link_libraries(${GPU_MOD_NAME} PRIVATE torch ${GPU_LIBRARIES})
 
   # Don't use `TORCH_LIBRARIES` for CUDA since it pulls in a bunch of
   # dependencies that are not necessary and may not be installed.
-  if (ARG_LANGUAGE STREQUAL "CUDA")
-    target_link_libraries(${MOD_NAME} PRIVATE torch CUDA::cudart CUDA::cuda_driver ${ARG_LIBRARIES})
+  if (GPU_LANGUAGE STREQUAL "CUDA")
+    target_link_libraries(${GPU_MOD_NAME} PRIVATE CUDA::cudart CUDA::cuda_driver)
   else()
-    target_link_libraries(${MOD_NAME} PRIVATE torch ${TORCH_LIBRARIES} ${ARG_LIBRARIES})
+    target_link_libraries(${GPU_MOD_NAME} PRIVATE ${TORCH_LIBRARIES})
   endif()
 
-  install(TARGETS ${MOD_NAME} LIBRARY DESTINATION ${ARG_DESTINATION} COMPONENT ${MOD_NAME})
+  install(TARGETS ${GPU_MOD_NAME} LIBRARY DESTINATION ${GPU_DESTINATION} COMPONENT ${GPU_MOD_NAME})
 endfunction()

csrc/attention/merge_attn_states.cu

@@ -46,32 +46,6 @@ __global__ void merge_attn_states_kernel(
   s_lse = std::isinf(s_lse) ? -std::numeric_limits<float>::infinity() : s_lse;
 
   const float max_lse = fmaxf(p_lse, s_lse);
-
-  /* In certain edge cases, MLA can produce p_lse = s_lse = -inf;
-     continuing the pipeline then yields NaN. Root cause: with chunked prefill
-     a batch may be split into two chunks; if a request in that batch has no
-     prefix hit, every LSE entry for that request’s position is -inf, and at
-     this moment we merge cross-attention at first. For now we simply emit
-     prefix_output (expected to be all zeros) and prefix_lse (-inf) to fix
-     this problem.
-  */
-  if (std::isinf(max_lse)) {
-    if (pack_offset < head_size) {
-      // Pack 128b load
-      pack_128b_t p_out_pack = reinterpret_cast<const pack_128b_t*>(
-          prefix_head_ptr)[pack_offset / pack_size];
-
-      // Pack 128b storage
-      reinterpret_cast<pack_128b_t*>(output_head_ptr)[pack_offset / pack_size] =
-          p_out_pack;
-    }
-    // We only need to write to output_lse once per head.
-    if (output_lse != nullptr && pack_idx == 0) {
-      output_lse[head_idx * num_tokens + token_idx] = max_lse;
-    }
-    return;
-  }
-
   p_lse = p_lse - max_lse;
   s_lse = s_lse - max_lse;
   const float p_se = expf(p_lse);

csrc/cpu/attention.cpp

@@ -0,0 +1,798 @@
+#include "cpu_types.hpp"
+
+namespace {
+
+template <typename scalar_t>
+struct KernelVecType {
+  using q_load_vec_type = void;
+  using q_vec_type = void;
+  using k_load_vec_type = void;
+  using k_vec_type = void;
+  using qk_acc_vec_type = void;
+  using v_load_vec_type = void;
+};
+
+template <>
+struct KernelVecType<float> {
+  using q_load_vec_type = vec_op::FP32Vec4;
+  using q_vec_type = vec_op::FP32Vec16;
+  using k_load_vec_type = vec_op::FP32Vec16;
+  using k_vec_type = vec_op::FP32Vec16;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP32Vec16;
+};
+
+template <>
+struct KernelVecType<c10::Half> {
+#if defined(__powerpc64__) || defined(__s390x__)
+  // Power and s390x architecture-specific vector types
+  using q_load_vec_type = vec_op::FP32Vec8;
+  using k_load_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP32Vec16;
+#else
+  // Fallback for other architectures, including x86
+  using q_load_vec_type = vec_op::FP16Vec8;
+  using k_load_vec_type = vec_op::FP16Vec16;
+  using v_load_vec_type = vec_op::FP16Vec16;
+#endif
+  using q_vec_type = vec_op::FP32Vec16;
+  using k_vec_type = vec_op::FP32Vec16;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+};
+
+#ifdef __AVX512BF16__
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using q_load_vec_type = vec_op::BF16Vec8;
+  using q_vec_type = vec_op::BF16Vec32;
+  using k_load_vec_type = vec_op::BF16Vec32;
+  using k_vec_type = vec_op::BF16Vec32;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+#else
+  #ifdef __aarch64__
+    #ifndef ARM_BF16_SUPPORT
+    // pass
+    #else
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using q_load_vec_type = vec_op::BF16Vec8;
+  using q_vec_type = vec_op::FP32Vec16;
+  using k_load_vec_type = vec_op::BF16Vec16;
+  using k_vec_type = vec_op::FP32Vec16;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+    #endif
+  #else
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using q_load_vec_type = vec_op::BF16Vec8;
+  using q_vec_type = vec_op::FP32Vec16;
+  using k_load_vec_type = vec_op::BF16Vec16;
+  using k_vec_type = vec_op::FP32Vec16;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+  #endif
+#endif
+
+template <typename T>
+FORCE_INLINE std::pair<T, T> reduceSoftmax(T* data, const int size,
+                                           const int capacity) {
+  T max = data[0];
+  for (int i = 1; i < size; ++i) {
+    max = max >= data[i] ? max : data[i];
+  }
+
+  T sum = 0;
+  for (int i = 0; i < size; ++i) {
+    data[i] = std::exp(data[i] - max);
+    sum += data[i];
+  }
+
+  int i = 0;
+  for (; i < size; ++i) {
+    data[i] /= sum;
+  }
+
+  for (; i < capacity; ++i) {
+    data[i] = 0;
+  }
+
+  return {max, sum};
+}
+
+template <typename T>
+FORCE_INLINE std::pair<T, T> reduceSoftmaxAlibi(T* data, const int size,
+                                                const int capacity,
+                                                const float alibi_slope,
+                                                const int start_index,
+                                                const int seq_len) {
+  data[0] += alibi_slope * (start_index - seq_len + 1);
+  T max = data[0];
+  for (int i = 1; i < size; ++i) {
+    T qk = data[i] + alibi_slope * (start_index + i - seq_len + 1);
+    data[i] = qk;
+    max = max >= qk ? max : qk;
+  }
+
+  T sum = 0;
+  for (int i = 0; i < size; ++i) {
+    data[i] = std::exp(data[i] - max);
+    sum += data[i];
+  }
+
+  int i = 0;
+  for (; i < size; ++i) {
+    data[i] /= sum;
+  }
+
+  for (; i < capacity; ++i) {
+    data[i] = 0;
+  }
+
+  return {max, sum};
+}
+
+template <typename T>
+FORCE_INLINE void reducePartitionSoftmax(const T* max_data, T* sum_data,
+                                         const int size) {
+  T max = max_data[0];
+  for (int i = 1; i < size; ++i) {
+    max = max >= max_data[i] ? max : max_data[i];
+  }
+
+  T rescaled_sum = 0;
+  for (int i = 0; i < size; ++i) {
+    T rescale_factor = std::exp(max_data[i] - max);
+    rescaled_sum += rescale_factor * sum_data[i];
+    sum_data[i] *= rescale_factor;
+  }
+  for (int i = 0; i < size; ++i) {
+    sum_data[i] /= rescaled_sum + 1e-8;
+  }
+}
+
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE, int x>
+struct reduceQKBlockKernel {
+  using q_load_vec_type = typename KernelVecType<scalar_t>::q_load_vec_type;
+  using q_vec_type = typename KernelVecType<scalar_t>::q_vec_type;
+  using k_load_vec_type = typename KernelVecType<scalar_t>::k_load_vec_type;
+  using k_vec_type = typename KernelVecType<scalar_t>::k_vec_type;
+  using qk_acc_vec_type = typename KernelVecType<scalar_t>::qk_acc_vec_type;
+
+  constexpr static int TOKEN_PER_GROUP = k_load_vec_type::get_elem_num() / x;
+  constexpr static int MAX_GROUP_NUM = 16 / TOKEN_PER_GROUP;
+  constexpr static int UNROLL_GROUP_NUM = MAX_GROUP_NUM / 4;
+
+  static_assert(MAX_GROUP_NUM == 8 || MAX_GROUP_NUM == 4);
+  static_assert(k_load_vec_type::get_elem_num() % x == 0);
+  static_assert(q_load_vec_type::get_elem_num() * sizeof(scalar_t) == 16);
+
+  FORCE_INLINE static void call(const scalar_t* __restrict__ q,
+                                const scalar_t* __restrict__ k_block,
+                                float* __restrict__ logits, float scale,
+                                const int token_num) {
+    const int group_num = (token_num + TOKEN_PER_GROUP - 1) / TOKEN_PER_GROUP;
+
+    qk_acc_vec_type group_accums[MAX_GROUP_NUM];
+    if (token_num == BLOCK_SIZE) {
+      for (int q_offset = 0; q_offset < HEAD_SIZE;
+           q_offset += x, k_block += x * BLOCK_SIZE) {
+        q_load_vec_type q_load_group_vec(q + q_offset);
+        q_vec_type q_group_vec(q_load_group_vec);
+
+        vec_op::unroll_loop<int, MAX_GROUP_NUM>(
+            [k_block, &q_group_vec, &group_accums](int token_group_idx) {
+              k_load_vec_type k_load_group_vec(k_block + token_group_idx * x *
+                                                             TOKEN_PER_GROUP);
+              k_vec_type k_group_vec(k_load_group_vec);
+              vec_op::fma(group_accums[token_group_idx], q_group_vec,
+                          k_group_vec);
+              vec_op::prefetch(k_block + x * BLOCK_SIZE +
+                               token_group_idx * x * TOKEN_PER_GROUP);
+            });
+      }
+    } else {
+      for (int q_offset = 0; q_offset < HEAD_SIZE;
+           q_offset += x, k_block += x * BLOCK_SIZE) {
+        q_load_vec_type q_load_group_vec(q + q_offset);
+        q_vec_type q_group_vec(q_load_group_vec);
+        for (int token_group_start = 0; token_group_start < group_num;
+             token_group_start += UNROLL_GROUP_NUM) {
+          vec_op::unroll_loop<int, UNROLL_GROUP_NUM>(
+              [token_group_start, k_block, &q_group_vec,
+               &group_accums](int token_group_idx) {
+                token_group_idx += token_group_start;
+                k_load_vec_type k_load_group_vec(k_block + token_group_idx * x *
+                                                               TOKEN_PER_GROUP);
+                k_vec_type k_group_vec(k_load_group_vec);
+                vec_op::fma(group_accums[token_group_idx], q_group_vec,
+                            k_group_vec);
+                vec_op::prefetch(k_block + x * BLOCK_SIZE +
+                                 token_group_idx * x * TOKEN_PER_GROUP);
+              });
+        }
+      }
+    }
+
+    for (int token_group_idx = 0; token_group_idx < group_num;
+         ++token_group_idx) {
+      vec_op::unroll_loop<int, TOKEN_PER_GROUP>(
+          [&group_accums, logits, scale, token_group_idx](int token_idx) {
+            float dot_v =
+                group_accums[token_group_idx]
+                    .template reduce_sub_sum<qk_acc_vec_type::get_elem_num() /
+                                             TOKEN_PER_GROUP>(token_idx);
+            logits[token_group_idx * TOKEN_PER_GROUP + token_idx] =
+                dot_v * scale;
+          });
+    }
+  }
+};
+
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE,
+          int HEAD_PARTITION_SIZE, typename acc_t>
+FORCE_INLINE void reduceValueBlock(const float* prob, const scalar_t* v_block,
+                                   acc_t&& acc) {
+  using v_load_vec_type = typename KernelVecType<scalar_t>::v_load_vec_type;
+  constexpr int ELEM_NUM = v_load_vec_type::get_elem_num();
+  static_assert(BLOCK_SIZE == ELEM_NUM);
+  vec_op::FP32Vec16 prob_vec(prob);
+
+  vec_op::unroll_loop<int, HEAD_PARTITION_SIZE>([&](int head_elem_idx) {
+    v_load_vec_type v_vec(v_block + BLOCK_SIZE * head_elem_idx);
+    vec_op::FP32Vec16 fp32_v_vec(v_vec);
+    acc[head_elem_idx] = acc[head_elem_idx] + prob_vec * fp32_v_vec;
+  });
+}
+};  // namespace
+
+// Paged attention v1
+namespace {
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE>
+struct paged_attention_v1_impl {
+  static void call(
+      scalar_t* __restrict__ out,            // [num_seqs, num_heads, head_size]
+      const scalar_t* __restrict__ q,        // [num_seqs, num_heads, head_size]
+      const scalar_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size/x, block_size, x]
+      const scalar_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size, block_size]
+      const int num_kv_heads, const float scale,
+      const int* __restrict__ block_tables,  // [num_seqs,
+                                             // max_num_blocks_per_seq]
+      const int* __restrict__ seq_lens,      // [num_seqs]
+      const int max_num_blocks_per_seq,
+      const float* __restrict__ alibi_slopes,  // [num_heads]
+      const int q_stride, const int kv_block_stride, const int kv_head_stride,
+      const int num_seqs, const int num_heads) {
+    constexpr int x = 16 / sizeof(scalar_t);
+    const int num_queries_per_kv = num_heads / num_kv_heads;
+
+    static_assert(BLOCK_SIZE == 16);
+
+    int max_seq_len = max_num_blocks_per_seq * BLOCK_SIZE;
+    int max_seq_len_padded = (max_seq_len + 15) & 0xFFFFFFF0;
+    TORCH_CHECK((max_seq_len_padded * sizeof(float)) % 64 == 0);
+
+    const int parallel_work_item_num = omp_get_max_threads();
+
+    size_t logits_bytes =
+        parallel_work_item_num * max_seq_len_padded * sizeof(float);
+    float* logits = (float*)std::aligned_alloc(
+        64, logits_bytes);  // Cacheline alignment for each context token.
+                            // [parallel_work_item_num, max_seq_len_padded]
+
+#pragma omp parallel for collapse(2) schedule(dynamic, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        int seq_len = seq_lens[seq_idx];
+        const int* seq_block_table =
+            block_tables + max_num_blocks_per_seq * seq_idx;
+        const int block_num = (seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
+        const int64_t kv_head_idx = head_idx / num_queries_per_kv;
+        const scalar_t* __restrict__ q_vec_ptr =
+            q + seq_idx * q_stride + head_idx * HEAD_SIZE;
+        const int last_block_token_num = seq_len - (block_num - 1) * BLOCK_SIZE;
+        float* __restrict__ thread_block_logits =
+            logits + omp_get_thread_num() * max_seq_len_padded;
+
+        // Compute logits
+        for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+          const int64_t physical_block_idx = seq_block_table[block_idx];
+          const scalar_t* __restrict__ k_block_cache_ptr =
+              k_cache + physical_block_idx * kv_block_stride +
+              kv_head_idx * kv_head_stride;
+          float* __restrict__ head_block_logits =
+              thread_block_logits + block_idx * BLOCK_SIZE;
+
+          reduceQKBlockKernel<scalar_t, HEAD_SIZE, BLOCK_SIZE, x>::call(
+              q_vec_ptr, k_block_cache_ptr, head_block_logits, scale,
+              block_idx == block_num - 1 ? last_block_token_num : BLOCK_SIZE);
+        }
+
+        // Compute softmax
+        if (alibi_slopes) {
+          reduceSoftmaxAlibi(thread_block_logits, seq_len,
+                             block_num * BLOCK_SIZE, alibi_slopes[head_idx], 0,
+                             seq_len);
+        } else {
+          reduceSoftmax(thread_block_logits, seq_len, block_num * BLOCK_SIZE);
+        }
+
+        // Compute value
+        constexpr int head_elem_num_per_partition = 16;
+        constexpr int head_partition_num =
+            HEAD_SIZE / head_elem_num_per_partition;
+        for (int head_part_idx = 0; head_part_idx < head_partition_num;
+             ++head_part_idx) {
+          vec_op::FP32Vec16 accums[head_elem_num_per_partition];
+          scalar_t* __restrict__ out_ptr =
+              out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE +
+              head_part_idx * head_elem_num_per_partition;
+          for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+            const int64_t physical_block_idx = seq_block_table[block_idx];
+            const float* __restrict__ prob_vec_ptr =
+                thread_block_logits + block_idx * BLOCK_SIZE;
+            const scalar_t* __restrict__ v_block_cache_ptr =
+                v_cache + physical_block_idx * kv_block_stride +
+                kv_head_idx * kv_head_stride +
+                BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+            reduceValueBlock<scalar_t, HEAD_SIZE, BLOCK_SIZE,
+                             head_elem_num_per_partition>(
+                prob_vec_ptr, v_block_cache_ptr, accums);
+
+            if (block_idx != block_num - 1) {
+              const int64_t next_physical_block_idx =
+                  seq_block_table[block_idx + 1];
+              const scalar_t* __restrict__ next_v_block_cache_ptr =
+                  v_cache + next_physical_block_idx * kv_block_stride +
+                  kv_head_idx * kv_head_stride +
+                  BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+              vec_op::unroll_loop<int, head_elem_num_per_partition>(
+                  [&](int head_elem_idx) {
+                    if (head_elem_idx % 2 == 0) {
+                      vec_op::prefetch(next_v_block_cache_ptr +
+                                       BLOCK_SIZE * head_elem_idx);
+                    }
+                  });
+            }
+          }
+
+          vec_op::unroll_loop<int, head_elem_num_per_partition>(
+              [&](int head_elem_idx) {
+                float value = accums[head_elem_idx].reduce_sum();
+                vec_op::storeFP32(value, out_ptr + head_elem_idx);
+              });
+        }
+      }
+    }
+    std::free(logits);
+  }
+};
+
+#define LAUNCH_V1_ATTENTION_KERNEL(T, HEAD_SIZE, BLOCK_SIZE)                   \
+  paged_attention_v1_impl<T, HEAD_SIZE, BLOCK_SIZE>::call(                     \
+      out_ptr, query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale, \
+      block_tables_ptr, seq_lens_ptr, max_num_blocks_per_seq,                  \
+      alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride, num_seqs,   \
+      num_heads);
+
+template <typename T, int BLOCK_SIZE>
+void paged_attention_v1_impl_launcher(
+    torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
+    const std::optional<torch::Tensor>& alibi_slopes) {
+  int num_seqs = query.size(0);
+  int num_heads = query.size(1);
+  int head_size = query.size(2);
+  int max_num_blocks_per_seq = block_tables.size(1);
+  int q_stride = query.stride(0);
+  int kv_block_stride = key_cache.stride(0);
+  int kv_head_stride = key_cache.stride(1);
+
+  // NOTE: alibi_slopes is optional.
+  const float* alibi_slopes_ptr =
+      alibi_slopes
+          ? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
+          : nullptr;
+
+  T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
+  T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
+  T* key_cache_ptr = reinterpret_cast<T*>(key_cache.data_ptr());
+  T* value_cache_ptr = reinterpret_cast<T*>(value_cache.data_ptr());
+  int* block_tables_ptr = block_tables.data_ptr<int>();
+  int* seq_lens_ptr = seq_lens.data_ptr<int>();
+
+  switch (head_size) {
+    case 32:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 32, BLOCK_SIZE);
+      break;
+    case 64:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 64, BLOCK_SIZE);
+      break;
+    case 80:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 80, BLOCK_SIZE);
+      break;
+    case 96:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 96, BLOCK_SIZE);
+      break;
+    case 112:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 112, BLOCK_SIZE);
+      break;
+    case 128:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 128, BLOCK_SIZE);
+      break;
+    case 192:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 192, BLOCK_SIZE);
+      break;
+    case 256:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 256, BLOCK_SIZE);
+      break;
+    default:
+      TORCH_CHECK(false, "Unsupported head size: ", head_size);
+      break;
+  }
+}
+
+#define CALL_V1_KERNEL_LAUNCHER(T, BLOCK_SIZE)                               \
+  paged_attention_v1_impl_launcher<T, BLOCK_SIZE>(                           \
+      out, query, key_cache, value_cache, num_kv_heads, scale, block_tables, \
+      seq_lens, max_seq_len, alibi_slopes);
+
+#define CALL_V1_KERNEL_LAUNCHER_BLOCK_SIZE(T)                     \
+  switch (block_size) {                                           \
+    case 16:                                                      \
+      CALL_V1_KERNEL_LAUNCHER(T, 16);                             \
+      break;                                                      \
+    default:                                                      \
+      TORCH_CHECK(false, "Unsupported block size: ", block_size); \
+      break;                                                      \
+  }
+}  // namespace
+
+void paged_attention_v1(
+    torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
+    int64_t max_seq_len, const std::optional<torch::Tensor>& alibi_slopes,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
+    const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
+    const int64_t blocksparse_head_sliding_step) {
+  TORCH_CHECK(blocksparse_vert_stride <= 1,
+              "CPU backend does not support blocksparse attention yet.");
+  VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v1_impl",
+                               [&] {
+                                 CPU_KERNEL_GUARD_IN(paged_attention_v1_impl)
+                                 CALL_V1_KERNEL_LAUNCHER_BLOCK_SIZE(scalar_t);
+                                 CPU_KERNEL_GUARD_OUT(paged_attention_v1_impl)
+                               });
+}
+
+// Paged attention v2
+namespace {
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE, int PARTITION_SIZE>
+struct paged_attention_v2_impl {
+  static void call(
+      scalar_t* __restrict__ out,            // [num_seqs, num_heads, head_size]
+      float* __restrict__ exp_sums,          // [num_seqs, num_heads,
+                                             // max_num_partitions]
+      float* __restrict__ max_logits,        // [num_seqs, num_heads,
+                                             // max_num_partitions]
+      scalar_t* __restrict__ tmp_out,        // [num_seqs, num_heads,
+                                             // max_num_partitions, head_size]
+      const scalar_t* __restrict__ q,        // [num_seqs, num_heads, head_size]
+      const scalar_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size/x, block_size, x]
+      const scalar_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size, block_size]
+      const int num_kv_heads, const float scale,
+      const int* __restrict__ block_tables,  // [num_seqs,
+                                             // max_num_blocks_per_seq]
+      const int* __restrict__ seq_lens,      // [num_seqs]
+      const int max_num_blocks_per_seq,
+      const float* __restrict__ alibi_slopes,  // [num_heads]
+      const int q_stride, const int kv_block_stride, const int kv_head_stride,
+      const int num_seqs, const int num_heads, const int max_num_partitions) {
+    constexpr int x = 16 / sizeof(scalar_t);
+    const int num_queries_per_kv = num_heads / num_kv_heads;
+
+    static_assert(BLOCK_SIZE == 16);
+    static_assert(PARTITION_SIZE * sizeof(float) % 64 == 0);
+    static_assert(PARTITION_SIZE % BLOCK_SIZE == 0);
+
+#pragma omp parallel for collapse(3) schedule(static, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int partition_idx = 0; partition_idx < max_num_partitions;
+           ++partition_idx) {
+        for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+          const int seq_len = seq_lens[seq_idx];
+          const int start_token_idx = partition_idx * PARTITION_SIZE;
+
+          if (start_token_idx >= seq_len) continue;
+
+          const int partition_num =
+              (seq_len + PARTITION_SIZE - 1) / PARTITION_SIZE;
+          const bool no_reduce = (partition_num == 1);
+          const int token_num =
+              (std::min(seq_len, start_token_idx + PARTITION_SIZE) -
+               start_token_idx);
+          const int block_num = (token_num + BLOCK_SIZE - 1) / BLOCK_SIZE;
+          const int last_block_token_num =
+              token_num - (block_num - 1) * BLOCK_SIZE;
+          const int* seq_block_table = block_tables +
+                                       max_num_blocks_per_seq * seq_idx +
+                                       start_token_idx / BLOCK_SIZE;
+          const int64_t kv_head_idx = head_idx / num_queries_per_kv;
+          const scalar_t* __restrict__ q_vec_ptr =
+              q + seq_idx * q_stride + head_idx * HEAD_SIZE;
+
+          float logits[PARTITION_SIZE] __attribute__((aligned(64))) = {0};
+
+          // Compute logits
+          for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+            const int64_t physical_block_idx = seq_block_table[block_idx];
+            const scalar_t* __restrict__ k_block_cache_ptr =
+                k_cache + physical_block_idx * kv_block_stride +
+                kv_head_idx * kv_head_stride;
+            float* __restrict__ head_block_logits =
+                logits + block_idx * BLOCK_SIZE;
+
+            reduceQKBlockKernel<scalar_t, HEAD_SIZE, BLOCK_SIZE, x>::call(
+                q_vec_ptr, k_block_cache_ptr, head_block_logits, scale,
+                block_idx == block_num - 1 ? last_block_token_num : BLOCK_SIZE);
+          }
+
+          std::pair<float, float> max_and_sum;
+          if (alibi_slopes) {
+            max_and_sum = reduceSoftmaxAlibi(
+                logits, token_num, block_num * BLOCK_SIZE,
+                alibi_slopes[head_idx], start_token_idx, seq_len);
+          } else {
+            max_and_sum =
+                reduceSoftmax(logits, token_num, block_num * BLOCK_SIZE);
+          }
+
+          auto&& [max_logit, exp_sum] = max_and_sum;
+
+          scalar_t* __restrict__ output_buffer = nullptr;
+          if (!no_reduce) {
+            auto idx = seq_idx * num_heads * max_num_partitions +
+                       head_idx * max_num_partitions + partition_idx;
+            max_logits[idx] = max_logit;
+            exp_sums[idx] = exp_sum;
+            output_buffer =
+                tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
+                head_idx * max_num_partitions * HEAD_SIZE +
+                partition_idx * HEAD_SIZE;
+          } else {
+            output_buffer =
+                out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
+          }
+
+          // Compute value
+          constexpr int head_elem_num_per_partition = 16;
+          constexpr int head_partition_num =
+              HEAD_SIZE / head_elem_num_per_partition;
+          for (int head_part_idx = 0; head_part_idx < head_partition_num;
+               ++head_part_idx) {
+            vec_op::FP32Vec16 accums[head_elem_num_per_partition];
+            scalar_t* __restrict__ out_ptr =
+                output_buffer + head_part_idx * head_elem_num_per_partition;
+            for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+              const int64_t physical_block_idx = seq_block_table[block_idx];
+              const float* __restrict__ prob_vec_ptr =
+                  logits + block_idx * BLOCK_SIZE;
+              const scalar_t* __restrict__ v_block_cache_ptr =
+                  v_cache + physical_block_idx * kv_block_stride +
+                  kv_head_idx * kv_head_stride +
+                  BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+              reduceValueBlock<scalar_t, HEAD_SIZE, BLOCK_SIZE,
+                               head_elem_num_per_partition>(
+                  prob_vec_ptr, v_block_cache_ptr, accums);
+
+              if (block_idx != block_num - 1) {
+                const int64_t next_physical_block_idx =
+                    seq_block_table[block_idx + 1];
+                const scalar_t* __restrict__ next_v_block_cache_ptr =
+                    v_cache + next_physical_block_idx * kv_block_stride +
+                    kv_head_idx * kv_head_stride +
+                    BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+                vec_op::unroll_loop<int, head_elem_num_per_partition>(
+                    [&](int head_elem_idx) {
+                      if (head_elem_idx % 2 == 0) {
+                        vec_op::prefetch(next_v_block_cache_ptr +
+                                         BLOCK_SIZE * head_elem_idx);
+                      }
+                    });
+              }
+            }
+
+            vec_op::unroll_loop<int, head_elem_num_per_partition>(
+                [&](int head_elem_idx) {
+                  float value = accums[head_elem_idx].reduce_sum();
+                  vec_op::storeFP32(value, out_ptr + head_elem_idx);
+                });
+          }
+        }
+      }
+    }
+
+    // Rescale partition softmax and store the factors to exp_sums
+#pragma omp parallel for collapse(2) schedule(static, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        const int seq_len = seq_lens[seq_idx];
+        const int partition_num =
+            (seq_len + PARTITION_SIZE - 1) / PARTITION_SIZE;
+
+        if (partition_num == 1) continue;
+
+        reducePartitionSoftmax(
+            max_logits + seq_idx * num_heads * max_num_partitions +
+                head_idx * max_num_partitions,
+            exp_sums + seq_idx * num_heads * max_num_partitions +
+                head_idx * max_num_partitions,
+            partition_num);
+      }
+    }
+
+    // Reduce values
+    using v_load_vec_type = typename KernelVecType<scalar_t>::v_load_vec_type;
+    static_assert(v_load_vec_type::get_elem_num() == BLOCK_SIZE);
+    constexpr int head_elem_num_per_group =
+        16;  // Note: didn't align with the cacheline size, due to some
+             // HEAD_SIZE didn't align with 64 bytes
+    static_assert(HEAD_SIZE % head_elem_num_per_group == 0);
+    constexpr int head_group_num = HEAD_SIZE / head_elem_num_per_group;
+    const float* __restrict__ rescale_factors = exp_sums;
+#pragma omp parallel for collapse(3) schedule(static, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        for (int group_idx = 0; group_idx < head_group_num; ++group_idx) {
+          const int seq_len = seq_lens[seq_idx];
+          const int partition_num =
+              (seq_len + PARTITION_SIZE - 1) / PARTITION_SIZE;
+
+          if (partition_num == 1) continue;
+
+          const float* __restrict__ seq_head_rescale_factors =
+              rescale_factors + seq_idx * num_heads * max_num_partitions +
+              head_idx * max_num_partitions;
+          const scalar_t* __restrict__ seq_head_tmp_out =
+              tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
+              head_idx * max_num_partitions * HEAD_SIZE +
+              group_idx * head_elem_num_per_group;
+          scalar_t* __restrict__ seq_head_output =
+              out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE +
+              group_idx * head_elem_num_per_group;
+
+          vec_op::FP32Vec16 acc;
+          for (int i = 0; i < partition_num; ++i) {
+            vec_op::FP32Vec16 rescale_factor(seq_head_rescale_factors[i]);
+            v_load_vec_type value(seq_head_tmp_out + i * HEAD_SIZE);
+            vec_op::FP32Vec16 fp32_value(value);
+            acc = acc + fp32_value * rescale_factor;
+          }
+          v_load_vec_type cast_acc(acc);
+          cast_acc.save(seq_head_output);
+        }
+      }
+    }
+  }
+};
+
+#define LAUNCH_V2_ATTENTION_KERNEL(T, HEAD_SIZE, BLOCK_SIZE)                 \
+  paged_attention_v2_impl<T, HEAD_SIZE, BLOCK_SIZE, PARTITION_SIZE>::call(   \
+      out_ptr, exp_sums_ptr, max_logits_ptr, tmp_out_ptr, query_ptr,         \
+      key_cache_ptr, value_cache_ptr, num_kv_heads, scale, block_tables_ptr, \
+      seq_lens_ptr, max_num_blocks_per_seq, alibi_slopes_ptr, q_stride,      \
+      kv_block_stride, kv_head_stride, num_seqs, num_heads,                  \
+      max_num_partitions);
+
+template <typename T, int BLOCK_SIZE, int PARTITION_SIZE = 512>
+void paged_attention_v2_impl_launcher(
+    torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
+    torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int block_size,
+    int max_seq_len, const std::optional<torch::Tensor>& alibi_slopes) {
+  int num_seqs = query.size(0);
+  int num_heads = query.size(1);
+  int head_size = query.size(2);
+  int max_num_blocks_per_seq = block_tables.size(1);
+  int q_stride = query.stride(0);
+  int kv_block_stride = key_cache.stride(0);
+  int kv_head_stride = key_cache.stride(1);
+  int max_num_partitions = exp_sums.size(-1);
+
+  // NOTE: alibi_slopes is optional.
+  const float* alibi_slopes_ptr =
+      alibi_slopes
+          ? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
+          : nullptr;
+
+  T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
+  float* exp_sums_ptr = reinterpret_cast<float*>(exp_sums.data_ptr());
+  float* max_logits_ptr = reinterpret_cast<float*>(max_logits.data_ptr());
+  T* tmp_out_ptr = reinterpret_cast<T*>(tmp_out.data_ptr());
+  T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
+  T* key_cache_ptr = reinterpret_cast<T*>(key_cache.data_ptr());
+  T* value_cache_ptr = reinterpret_cast<T*>(value_cache.data_ptr());
+  int* block_tables_ptr = block_tables.data_ptr<int>();
+  int* seq_lens_ptr = seq_lens.data_ptr<int>();
+
+  switch (head_size) {
+    case 32:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 32, BLOCK_SIZE);
+      break;
+    case 64:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 64, BLOCK_SIZE);
+      break;
+    case 80:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 80, BLOCK_SIZE);
+      break;
+    case 96:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 96, BLOCK_SIZE);
+      break;
+    case 112:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 112, BLOCK_SIZE);
+      break;
+    case 128:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 128, BLOCK_SIZE);
+      break;
+    case 192:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 192, BLOCK_SIZE);
+      break;
+    case 256:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 256, BLOCK_SIZE);
+      break;
+    default:
+      TORCH_CHECK(false, "Unsupported head size: ", head_size);
+      break;
+  }
+}
+
+#define CALL_V2_KERNEL_LAUNCHER(T, BLOCK_SIZE)                              \
+  paged_attention_v2_impl_launcher<T, BLOCK_SIZE>(                          \
+      out, exp_sums, max_logits, tmp_out, query, key_cache, value_cache,    \
+      num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len, \
+      alibi_slopes);
+
+#define CALL_V2_KERNEL_LAUNCHER_BLOCK_SIZE(T)                     \
+  switch (block_size) {                                           \
+    case 16:                                                      \
+      CALL_V2_KERNEL_LAUNCHER(T, 16);                             \
+      break;                                                      \
+    default:                                                      \
+      TORCH_CHECK(false, "Unsupported block size: ", block_size); \
+      break;                                                      \
+  }
+}  // namespace
+
+void paged_attention_v2(
+    torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
+    torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
+    int64_t max_seq_len, const std::optional<torch::Tensor>& alibi_slopes,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
+    const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
+    const int64_t blocksparse_head_sliding_step) {
+  TORCH_CHECK(blocksparse_vert_stride <= 1,
+              "CPU backend does not support blocksparse attention yet.");
+  VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v2_impl",
+                               [&] {
+                                 CPU_KERNEL_GUARD_IN(paged_attention_v2_impl)
+                                 CALL_V2_KERNEL_LAUNCHER_BLOCK_SIZE(scalar_t);
+                                 CPU_KERNEL_GUARD_OUT(paged_attention_v2_impl)
+                               });
+}

csrc/cpu/cache.cpp

@@ -0,0 +1,214 @@
+#include <map>
+#include <vector>
+
+#include "cpu_types.hpp"
+
+#if defined(__x86_64__)
+  #define DISPATCH_MACRO VLLM_DISPATCH_FLOATING_TYPES_WITH_E5M2
+#else
+  #define DISPATCH_MACRO VLLM_DISPATCH_FLOATING_TYPES
+#endif
+
+namespace {
+template <typename scalar_t>
+void copy_blocks_cpu_impl(std::vector<torch::Tensor> const& key_caches,
+                          std::vector<torch::Tensor> const& value_caches,
+                          const torch::Tensor& mapping_pairs,
+                          const int element_num_per_block,
+                          const int layer_num) {
+  const size_t pair_num = mapping_pairs.size(0);
+  const size_t block_bytes = sizeof(scalar_t) * element_num_per_block;
+#pragma omp parallel for collapse(2)
+  for (int layer = 0; layer < layer_num; ++layer) {
+    for (size_t pair = 0; pair < pair_num; ++pair) {
+      int64_t source_offset =
+          element_num_per_block * mapping_pairs[pair][0].item<int64_t>();
+      int64_t target_offset =
+          element_num_per_block * mapping_pairs[pair][1].item<int64_t>();
+      scalar_t* key_cache_ptr = key_caches[layer].data_ptr<scalar_t>();
+      scalar_t* source_ptr = key_cache_ptr + source_offset;
+      scalar_t* target_ptr = key_cache_ptr + target_offset;
+      std::memcpy(target_ptr, source_ptr, block_bytes);
+
+      scalar_t* value_cache_ptr = value_caches[layer].data_ptr<scalar_t>();
+      source_ptr = value_cache_ptr + source_offset;
+      target_ptr = value_cache_ptr + target_offset;
+      std::memcpy(target_ptr, source_ptr, block_bytes);
+    }
+  }
+}
+
+template <typename scalar_t>
+void reshape_and_cache_cpu_impl(
+    const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
+    scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
+    const int64_t* __restrict__ slot_mapping, const int num_tokens,
+    const int key_stride, const int value_stride, const int num_heads,
+    const int head_size, const int block_size, const int x) {
+  const int block_elem_num = num_heads * head_size * block_size;
+
+#pragma omp parallel for collapse(2)
+  for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
+    for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+      const int64_t slot_idx = slot_mapping[token_idx];
+      if (slot_idx >= 0) {
+        int src_key_head_idx = token_idx * key_stride + head_idx * head_size;
+        int src_value_head_idx =
+            token_idx * value_stride + head_idx * head_size;
+        const scalar_t* src_key_head_ptr = key + src_key_head_idx;
+        const scalar_t* src_value_head_ptr = value + src_value_head_idx;
+        const int64_t block_index = slot_idx / block_size;
+        const int64_t block_offset = slot_idx % block_size;
+        scalar_t* target_key_head_ptr = key_cache +
+                                        block_elem_num * block_index +
+                                        head_idx * block_size * head_size;
+        scalar_t* target_value_head_ptr = value_cache +
+                                          block_elem_num * block_index +
+                                          head_idx * block_size * head_size;
+
+        for (int src_key_idx = 0; src_key_idx < head_size; src_key_idx += x) {
+          const int64_t target_offset =
+              src_key_idx * block_size + block_offset * x;
+          for (int i = 0; i < x; ++i) {
+            target_key_head_ptr[target_offset + i] =
+                src_key_head_ptr[src_key_idx + i];
+          }
+        }
+
+        for (int src_value_idx = 0; src_value_idx < head_size;
+             ++src_value_idx) {
+          const int64_t target_offset =
+              src_value_idx * block_size + block_offset;
+          target_value_head_ptr[target_offset] =
+              src_value_head_ptr[src_value_idx];
+        }
+      }
+    }
+  }
+}
+};  // namespace
+
+template <typename scalar_t>
+void concat_and_cache_mla_cpu_impl(
+    const scalar_t* __restrict__ kv_c,  // [num_tokens, kv_lora_rank]
+    const scalar_t* __restrict__ k_pe,  // [num_tokens, pe_dim]
+    scalar_t* __restrict__ kv_cache,  // [num_blocks, block_size, (kv_lora_rank
+                                      // + pe_dim)]
+    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
+    const int num_tokens,                      //
+    const int block_stride,                    //
+    const int entry_stride,                    //
+    const int kv_c_stride,                     //
+    const int k_pe_stride,                     //
+    const int kv_lora_rank,                    //
+    const int pe_dim,                          //
+    const int block_size                       //
+) {
+#pragma omp parallel for
+  for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
+    const int64_t slot_idx = slot_mapping[token_idx];
+    // NOTE: slot_idx can be -1 if the token is padded
+    if (slot_idx < 0) {
+      continue;
+    }
+    const int64_t block_idx = slot_idx / block_size;
+    const int64_t block_offset = slot_idx % block_size;
+
+    auto copy = [&](const scalar_t* __restrict__ src,
+                    scalar_t* __restrict__ dst, int src_stride, int dst_stride,
+                    int size, int offset) {
+      for (int i = 0; i < size; i++) {
+        const int64_t src_idx = token_idx * src_stride + i;
+        const int64_t dst_idx =
+            block_idx * block_stride + block_offset * entry_stride + i + offset;
+        dst[dst_idx] = src[src_idx];
+      }
+    };
+
+    copy(kv_c, kv_cache, kv_c_stride, block_stride, kv_lora_rank, 0);
+    copy(k_pe, kv_cache, k_pe_stride, block_stride, pe_dim, kv_lora_rank);
+  }
+}
+
+// Note: the key_caches and value_caches vectors are constant but
+// not the Tensors they contain. The vectors need to be const refs
+// in order to satisfy pytorch's C++ operator registration code.
+void copy_blocks(std::vector<torch::Tensor> const& key_caches,
+                 std::vector<torch::Tensor> const& value_caches,
+                 const torch::Tensor& block_mapping) {
+  unsigned num_layers = key_caches.size();
+  TORCH_CHECK(num_layers == value_caches.size());
+  if (num_layers == 0) {
+    return;
+  }
+
+  const int element_num_per_block = key_caches[0][0].numel();
+  DISPATCH_MACRO(key_caches[0].scalar_type(), "copy_blocks_cpu_impl", [&] {
+    CPU_KERNEL_GUARD_IN(copy_blocks_cpu_impl)
+    copy_blocks_cpu_impl<scalar_t>(key_caches, value_caches, block_mapping,
+                                   element_num_per_block, num_layers);
+    CPU_KERNEL_GUARD_OUT(copy_blocks_cpu_impl)
+  });
+}
+
+void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
+                       torch::Tensor& key_cache, torch::Tensor& value_cache,
+                       torch::Tensor& slot_mapping,
+                       const std::string& kv_cache_dtype,
+                       torch::Tensor& k_scale, torch::Tensor& v_scale) {
+  int num_tokens = key.size(0);
+  int num_heads = key.size(1);
+  int head_size = key.size(2);
+  int block_size = key_cache.size(3);
+  int x = key_cache.size(4);
+
+  int key_stride = key.stride(0);
+  int value_stride = value.stride(0);
+
+  DISPATCH_MACRO(key.scalar_type(), "reshape_and_cache_cpu_impl", [&] {
+    CPU_KERNEL_GUARD_IN(reshape_and_cache_cpu_impl)
+    reshape_and_cache_cpu_impl<scalar_t>(
+        key.data_ptr<scalar_t>(), value.data_ptr<scalar_t>(),
+        key_cache.data_ptr<scalar_t>(), value_cache.data_ptr<scalar_t>(),
+        slot_mapping.data_ptr<int64_t>(), num_tokens, key_stride, value_stride,
+        num_heads, head_size, block_size, x);
+    CPU_KERNEL_GUARD_OUT(reshape_and_cache_cpu_impl)
+  });
+}
+
+void concat_and_cache_mla(
+    torch::Tensor& kv_c,          // [num_tokens, kv_lora_rank]
+    torch::Tensor& k_pe,          // [num_tokens, pe_dim]
+    torch::Tensor& kv_cache,      // [num_blocks, block_size, (kv_lora_rank +
+                                  // pe_dim)]
+    torch::Tensor& slot_mapping,  // [num_tokens] or [num_actual_tokens]
+    const std::string& kv_cache_dtype, torch::Tensor& scale) {
+  int num_tokens = slot_mapping.size(0);
+  int kv_lora_rank = kv_c.size(1);
+  int pe_dim = k_pe.size(1);
+  int block_size = kv_cache.size(1);
+
+  TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
+  TORCH_CHECK(kv_cache_dtype != "fp8");
+
+  int kv_c_stride = kv_c.stride(0);
+  int k_pe_stride = k_pe.stride(0);
+  int block_stride = kv_cache.stride(0);
+  int entry_stride = kv_cache.stride(1);
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      kv_c.scalar_type(), "concat_and_cache_mla_cpu_impl", [&] {
+        CPU_KERNEL_GUARD_IN(concat_and_cache_mla_cpu_impl)
+        concat_and_cache_mla_cpu_impl<scalar_t>(
+            kv_c.data_ptr<scalar_t>(), k_pe.data_ptr<scalar_t>(),
+            kv_cache.data_ptr<scalar_t>(), slot_mapping.data_ptr<int64_t>(),
+            num_tokens, block_stride, entry_stride, kv_c_stride, k_pe_stride,
+            kv_lora_rank, pe_dim, block_size);
+        CPU_KERNEL_GUARD_OUT(concat_and_cache_mla_cpu_impl)
+      });
+}
+
+void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
+                 const torch::Tensor& block_mapping) {
+  TORCH_CHECK(false, "swap_blocks is unsupported on CPU.")
+}

csrc/cpu/cpu_attn.cpp

@@ -1,249 +0,0 @@
-#include "cpu_attn_vec.hpp"
-#include "cpu_attn_vec16.hpp"
-
-#ifdef CPU_CAPABILITY_AMXBF16
-  #include "cpu_attn_amx.hpp"
-  #define AMX_DISPATCH(...)                                                   \
-    case cpu_attention::ISA::AMX: {                                           \
-      using attn_impl = cpu_attention::AttentionImpl<cpu_attention::ISA::AMX, \
-                                                     scalar_t, head_dim>;     \
-      return __VA_ARGS__();                                                   \
-    }
-#else
-  #define AMX_DISPATCH(...) case cpu_attention::ISA::AMX:
-#endif
-
-#define CPU_ATTN_DISPATCH_CASE(HEAD_DIM, ...) \
-  case HEAD_DIM: {                            \
-    constexpr size_t head_dim = HEAD_DIM;     \
-    return __VA_ARGS__();                     \
-  }
-
-#define CPU_ATTN_DISPATCH_CASE_HEADDIM(HEAD_DIM, ...)           \
-  [&] {                                                         \
-    switch (HEAD_DIM) {                                         \
-      CPU_ATTN_DISPATCH_CASE(32, __VA_ARGS__)                   \
-      CPU_ATTN_DISPATCH_CASE(64, __VA_ARGS__)                   \
-      CPU_ATTN_DISPATCH_CASE(96, __VA_ARGS__)                   \
-      CPU_ATTN_DISPATCH_CASE(128, __VA_ARGS__)                  \
-      CPU_ATTN_DISPATCH_CASE(160, __VA_ARGS__)                  \
-      CPU_ATTN_DISPATCH_CASE(192, __VA_ARGS__)                  \
-      CPU_ATTN_DISPATCH_CASE(224, __VA_ARGS__)                  \
-      CPU_ATTN_DISPATCH_CASE(256, __VA_ARGS__)                  \
-      default: {                                                \
-        TORCH_CHECK(false, "Invalid CPU attention head_dim: " + \
-                               std::to_string(HEAD_DIM));       \
-      }                                                         \
-    }                                                           \
-  }()
-
-#define CPU_ATTN_DISPATCH_IMPL(ISA_TYPE, ...)                                 \
-  [&] {                                                                       \
-    switch (ISA_TYPE) {                                                       \
-      AMX_DISPATCH(__VA_ARGS__)                                               \
-      case cpu_attention::ISA::VEC: {                                         \
-        using attn_impl =                                                     \
-            cpu_attention::AttentionImpl<cpu_attention::ISA::VEC, scalar_t,   \
-                                         head_dim>;                           \
-        return __VA_ARGS__();                                                 \
-      }                                                                       \
-      case cpu_attention::ISA::VEC16: {                                       \
-        using attn_impl =                                                     \
-            cpu_attention::AttentionImpl<cpu_attention::ISA::VEC16, scalar_t, \
-                                         head_dim>;                           \
-        return __VA_ARGS__();                                                 \
-      }                                                                       \
-      default: {                                                              \
-        TORCH_CHECK(false, "Invalid CPU attention ISA type.");                \
-      }                                                                       \
-    }                                                                         \
-  }()
-
-torch::Tensor get_scheduler_metadata(
-    const int64_t num_req, const int64_t num_heads_q,
-    const int64_t num_heads_kv, const int64_t head_dim,
-    const torch::Tensor& seq_lens, at::ScalarType dtype,
-    const torch::Tensor& query_start_loc, const bool casual,
-    const int64_t window_size, const std::string& isa_hint,
-    const bool enable_kv_split) {
-  cpu_attention::ISA isa;
-  if (isa_hint == "amx") {
-    isa = cpu_attention::ISA::AMX;
-  } else if (isa_hint == "vec") {
-    isa = cpu_attention::ISA::VEC;
-  } else if (isa_hint == "vec16") {
-    isa = cpu_attention::ISA::VEC16;
-  } else {
-    TORCH_CHECK(false, "Unsupported CPU attention ISA hint: " + isa_hint);
-  }
-
-  cpu_attention::AttentionScheduler::ScheduleInput input;
-  input.num_reqs = num_req;
-  input.num_heads_q = num_heads_q;
-  input.num_heads_kv = num_heads_kv;
-  input.head_dim = head_dim;
-  input.query_start_loc = query_start_loc.data_ptr<int32_t>();
-  input.seq_lens = seq_lens.data_ptr<int32_t>();
-  if (window_size != -1) {
-    input.left_sliding_window_size = window_size - 1;
-    if (casual) {
-      input.right_sliding_window_size = 0;
-    } else {
-      input.right_sliding_window_size = window_size - 1;
-    }
-  } else {
-    input.left_sliding_window_size = -1;
-    if (casual) {
-      input.right_sliding_window_size = 0;
-    } else {
-      input.right_sliding_window_size = -1;
-    }
-  }
-  input.casual = casual;
-  input.isa = isa;
-  input.enable_kv_split = enable_kv_split;
-  TORCH_CHECK(casual, "Only supports casual mask for now.");
-
-  VLLM_DISPATCH_FLOATING_TYPES(dtype, "get_scheduler_metadata", [&]() {
-    CPU_ATTN_DISPATCH_CASE_HEADDIM(head_dim, [&] {
-      CPU_ATTN_DISPATCH_IMPL(isa, [&]() {
-        input.elem_size = sizeof(scalar_t);
-        input.q_buffer_elem_size = sizeof(attn_impl::q_buffer_t);
-        input.logits_buffer_elem_size = sizeof(attn_impl::logits_buffer_t);
-        input.output_buffer_elem_size =
-            sizeof(attn_impl::partial_output_buffer_t);
-        input.max_num_q_per_iter = attn_impl::MaxQHeadNumPerIteration;
-        input.kv_block_alignment = attn_impl::BlockSizeAlignment;
-      });
-    });
-  });
-
-  cpu_attention::AttentionScheduler scheduler;
-  torch::Tensor metadata = scheduler.schedule(input);
-  return metadata;
-}
-
-void cpu_attn_reshape_and_cache(
-    const torch::Tensor& key,    // [token_num, head_num, head_size]
-    const torch::Tensor& value,  // [token_num, head_num, head_size]
-    torch::Tensor&
-        key_cache,  // [num_blocks, num_kv_heads, block_size, head_size]
-    torch::Tensor&
-        value_cache,  // [num_blocks, num_kv_heads, block_size, head_size]
-    const torch::Tensor& slot_mapping, const std::string& isa) {
-  TORCH_CHECK_EQ(key.dim(), 3);
-  TORCH_CHECK_EQ(value.dim(), 3);
-  TORCH_CHECK_EQ(key_cache.dim(), 4);
-  TORCH_CHECK_EQ(value_cache.dim(), 4);
-  TORCH_CHECK_EQ(key.stride(2), 1);
-  TORCH_CHECK_EQ(value.stride(2), 1);
-
-  const int64_t token_num = key.size(0);
-  const int64_t key_token_num_stride = key.stride(0);
-  const int64_t value_token_num_stride = value.stride(0);
-  const int64_t head_num = value.size(1);
-  const int64_t key_head_num_stride = key.stride(1);
-  const int64_t value_head_num_stride = value.stride(1);
-  const int64_t num_blocks = key_cache.size(0);
-  const int64_t num_blocks_stride = key_cache.stride(0);
-  const int64_t cache_head_num_stride = key_cache.stride(1);
-  const int64_t block_size = key_cache.size(2);
-  const int64_t block_size_stride = key_cache.stride(2);
-  const int64_t head_dim = key.size(-1);
-
-  cpu_attention::ISA isa_tag = [&]() {
-    if (isa == "amx") {
-      return cpu_attention::ISA::AMX;
-    } else if (isa == "vec") {
-      return cpu_attention::ISA::VEC;
-    } else if (isa == "vec16") {
-      return cpu_attention::ISA::VEC16;
-    } else {
-      TORCH_CHECK(false, "Invalid ISA type: " + isa);
-    }
-  }();
-
-  VLLM_DISPATCH_FLOATING_TYPES(
-      key.scalar_type(), "cpu_attn_reshape_and_cache", [&]() {
-        CPU_ATTN_DISPATCH_CASE_HEADDIM(head_dim, [&] {
-          CPU_ATTN_DISPATCH_IMPL(isa_tag, [&]() {
-            attn_impl::reshape_and_cache(
-                key.data_ptr<scalar_t>(), value.data_ptr<scalar_t>(),
-                key_cache.data_ptr<scalar_t>(),
-                value_cache.data_ptr<scalar_t>(),
-                slot_mapping.data_ptr<int64_t>(), token_num,
-                key_token_num_stride, value_token_num_stride, head_num,
-                key_head_num_stride, value_head_num_stride, num_blocks,
-                num_blocks_stride, cache_head_num_stride, block_size,
-                block_size_stride);
-          });
-        });
-      });
-}
-
-void cpu_attention_with_kv_cache(
-    const torch::Tensor& query,  // [num_tokens, num_heads, head_size]
-    const torch::Tensor&
-        key_cache,  // [num_blocks, num_kv_heads, block_size, head_size]
-    const torch::Tensor&
-        value_cache,        // [num_blocks, num_kv_heads, block_size, head_size]
-    torch::Tensor& output,  // [num_tokens, num_heads, head_size]
-    const torch::Tensor& query_start_loc,  // [num_tokens + 1]
-    const torch::Tensor& seq_lens,         // [num_tokens]
-    const double scale, const bool causal,
-    const std::optional<torch::Tensor>& alibi_slopes,  // [num_heads]
-    const int64_t sliding_window_left, const int64_t sliding_window_right,
-    const torch::Tensor& block_table,  // [num_tokens, max_block_num]
-    const double softcap, const torch::Tensor& scheduler_metadata,
-    const std::optional<torch::Tensor>& s_aux  // [num_heads]
-) {
-  TORCH_CHECK_EQ(query.dim(), 3);
-  TORCH_CHECK_EQ(query.stride(2), 1);
-  TORCH_CHECK_EQ(key_cache.dim(), 4);
-  TORCH_CHECK_EQ(value_cache.dim(), 4);
-
-  cpu_attention::AttentionInput input;
-  input.metadata = reinterpret_cast<cpu_attention::AttentionMetadata*>(
-      scheduler_metadata.data_ptr());
-  input.num_tokens = query.size(0);
-  input.num_heads = query.size(1);
-  input.num_kv_heads = key_cache.size(1);
-  input.block_size = key_cache.size(2);
-  input.query = query.data_ptr();
-  input.query_num_tokens_stride = query.stride(0);
-  input.query_num_heads_stride = query.stride(1);
-  input.cache_num_blocks_stride = key_cache.stride(0);
-  input.cache_num_kv_heads_stride = key_cache.stride(1);
-  input.blt_num_tokens_stride = block_table.stride(0);
-  input.key_cache = key_cache.data_ptr();
-  input.value_cache = value_cache.data_ptr();
-  input.output = output.data_ptr();
-  input.query_start_loc = query_start_loc.data_ptr<int32_t>();
-  input.seq_lens = seq_lens.data_ptr<int32_t>();
-  input.block_table = block_table.data_ptr<int32_t>();
-  input.alibi_slopes =
-      alibi_slopes.has_value() ? alibi_slopes->data_ptr<float>() : nullptr;
-  // For now sink must be bf16
-  input.s_aux = s_aux.has_value() ? s_aux->data_ptr<c10::BFloat16>() : nullptr;
-  input.scale = scale;
-  input.causal = causal;
-  input.sliding_window_left = sliding_window_left;
-  input.sliding_window_right = sliding_window_right;
-  if (input.causal) {
-    // to make boundary calculation easier
-    input.sliding_window_right = 0;
-  }
-  float softcap_fp32 = softcap;
-  input.softcap = softcap_fp32;
-
-  VLLM_DISPATCH_FLOATING_TYPES(
-      query.scalar_type(), "cpu_attention_with_kv_cache", [&]() {
-        CPU_ATTN_DISPATCH_CASE_HEADDIM(query.size(2), [&] {
-          CPU_ATTN_DISPATCH_IMPL(input.metadata->isa, [&]() {
-            TORCH_CHECK_EQ(input.block_size % attn_impl::BlockSizeAlignment, 0);
-            cpu_attention::AttentionMainLoop<attn_impl> mainloop;
-            mainloop(&input);
-          });
-        });
-      });
-}

csrc/cpu/cpu_attn_amx.hpp

@@ -1,511 +0,0 @@
-#ifndef CPU_ATTN_AMX_HPP
-#define CPU_ATTN_AMX_HPP
-
-#include "cpu_attn_impl.hpp"
-
-namespace cpu_attention {
-namespace {
-// AMX specific
-constexpr static int64_t AMX_TILE_ROW_BYTES = 64;
-constexpr static int64_t AMX_TILE_ROW_NUM = 16;
-constexpr static int64_t AMX_TILE_BYTES = AMX_TILE_ROW_BYTES * AMX_TILE_ROW_NUM;
-
-typedef struct __tile_config {
-  uint8_t palette_id = 1;
-  uint8_t start_row = 0;
-  uint8_t reserved_0[14] = {0};
-  uint16_t colsb[16] = {0};
-  uint8_t rows[16] = {0};
-} __tilecfg;
-
-// 2-2-4 pattern, for 16 < m <= 32
-// TILE 0, 1: load A matrix, row num should be 16, m - 16
-// TILE 2, 3: load B matrix, row num should be 16
-// TILE 4, 5, 6, 7: store results C matrix, row num should be 16, 16, m - 16, m
-// - 16
-template <typename kv_cache_t>
-class TileGemm224 {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size, void* __restrict__ a_tile,
-                                void* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    TORCH_CHECK(false, "Unsupported kv cache type for TileGemm224");
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    TORCH_CHECK(false, "Unsupported kv cache type for TileGemm224");
-  }
-};
-
-template <>
-class TileGemm224<c10::BFloat16> {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size,
-                                c10::BFloat16* __restrict__ a_tile,
-                                c10::BFloat16* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    const int32_t k_times =
-        dynamic_k_size / (AMX_TILE_ROW_NUM * 4 / sizeof(c10::BFloat16));
-    c10::BFloat16* __restrict__ a_tile_0 = a_tile;
-    c10::BFloat16* __restrict__ a_tile_1 = a_tile + lda * AMX_TILE_ROW_NUM;
-    const int64_t a_tile_stride = [&]() {
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // q_buffer is prepacked
-        return AMX_TILE_ROW_BYTES;
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // logits_buffer is row-major
-        return lda * sizeof(c10::BFloat16);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-    }();
-
-    c10::BFloat16* __restrict__ b_tile_2 = b_tile;
-    c10::BFloat16* __restrict__ b_tile_3 = [&]() {
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // k_cache is prepacked
-        return b_tile + (k_size * AMX_TILE_ROW_BYTES / 4);
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // v_cache is prepacked
-        return b_tile + (block_size * AMX_TILE_ROW_BYTES / 4);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-    }();
-    // k_cache, v_cache are prepacked
-    const int32_t b_tile_stride = AMX_TILE_ROW_BYTES;
-
-    // logits_buffer, output_buffer are not prepacked
-    float* __restrict__ c_tile_4 = c_tile;
-    float* __restrict__ c_tile_5 =
-        c_tile_4 + AMX_TILE_ROW_BYTES / sizeof(float);
-    float* __restrict__ c_tile_6 = c_tile + AMX_TILE_ROW_NUM * ldc;
-    float* __restrict__ c_tile_7 =
-        c_tile_6 + AMX_TILE_ROW_BYTES / sizeof(float);
-    const int32_t c_tile_stride = ldc * sizeof(float);
-
-    if (accum_c) {
-      _tile_loadd(4, c_tile_4, c_tile_stride);
-      _tile_loadd(5, c_tile_5, c_tile_stride);
-      _tile_loadd(6, c_tile_6, c_tile_stride);
-      _tile_loadd(7, c_tile_7, c_tile_stride);
-    } else {
-      _tile_zero(4);
-      _tile_zero(5);
-      _tile_zero(6);
-      _tile_zero(7);
-    }
-
-    for (int32_t k = 0; k < k_times; ++k) {
-      _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_tile_stride);
-      _tile_dpbf16ps(4, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_tile_stride);
-      _tile_dpbf16ps(5, 0, 3);
-      _tile_loadd(1, a_tile_1, a_tile_stride);
-      _tile_dpbf16ps(6, 1, 2);
-      _tile_dpbf16ps(7, 1, 3);
-
-      // update ptrs
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // Q buffer is prepacked
-        a_tile_0 += AMX_TILE_BYTES / sizeof(c10::BFloat16);
-        a_tile_1 += AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // P buffer is not prepacked
-        a_tile_0 += AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-        a_tile_1 += AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-      b_tile_2 += AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_3 += AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    }
-
-    _tile_stored(4, c_tile_4, c_tile_stride);
-    _tile_stored(5, c_tile_5, c_tile_stride);
-    _tile_stored(6, c_tile_6, c_tile_stride);
-    _tile_stored(7, c_tile_7, c_tile_stride);
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    const int32_t m_0 = AMX_TILE_ROW_NUM;
-    const int32_t m_1 = m - AMX_TILE_ROW_NUM;
-    config.rows[0] = m_0;
-    config.rows[1] = m_1;
-    config.rows[2] = AMX_TILE_ROW_NUM;
-    config.rows[3] = AMX_TILE_ROW_NUM;
-    config.rows[4] = m_0;
-    config.rows[5] = m_0;
-    config.rows[6] = m_1;
-    config.rows[7] = m_1;
-    _tile_loadconfig(&config);
-  }
-};
-
-// 1-2-2 pattern, for 0 < m <= 16
-// TILE 0, (1): load A matrix, use extra 1 tile for prefetch, row num should be
-// m, m
-// TILE 2, 3, (4, 5): load B matrix, use extra 2 tiles for prefetch, row
-// num should be 16
-// TILE 6, 7, (6, 7): store results C matrix, row num should be
-// m
-template <typename kv_cache_t>
-class TileGemm122 {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size, void* __restrict__ a_tile,
-                                void* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    TORCH_CHECK(false, "Unsupported kv cache type for TileGemm122");
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    TORCH_CHECK(false, "Unsupported kv cache type for TileGemm122");
-  }
-};
-
-template <>
-class TileGemm122<c10::BFloat16> {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size,
-                                c10::BFloat16* __restrict__ a_tile,
-                                c10::BFloat16* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    c10::BFloat16* __restrict__ a_tile_0 = a_tile;
-    c10::BFloat16* __restrict__ a_tile_1 = [&]() {
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // q_buffer is prepacked
-        return a_tile + AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // logits_buffer is row-major
-        return a_tile + AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-    }();
-    const int64_t a_tile_stride = [&]() {
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // q_buffer is prepacked
-        return AMX_TILE_ROW_BYTES;
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // logits_buffer is row-major
-        return lda * sizeof(c10::BFloat16);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-    }();
-
-    c10::BFloat16* __restrict__ b_tile_2 = b_tile;
-    c10::BFloat16* __restrict__ b_tile_3 = [&]() {
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // k_cache is prepacked
-        return b_tile + (k_size * AMX_TILE_ROW_BYTES / 4);
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // v_cache is prepacked
-        return b_tile + (block_size * AMX_TILE_ROW_BYTES / 4);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-    }();
-    c10::BFloat16* __restrict__ b_tile_4 =
-        b_tile_2 + AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    c10::BFloat16* __restrict__ b_tile_5 =
-        b_tile_3 + AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    int64_t b_stride = AMX_TILE_ROW_BYTES;
-
-    float* __restrict__ c_tile_6 = c_tile;
-    float* __restrict__ c_tile_7 = c_tile + AMX_TILE_ROW_BYTES / sizeof(float);
-    int64_t c_stride = ldc * sizeof(float);
-
-    const int32_t k_times =
-        dynamic_k_size / (AMX_TILE_ROW_NUM * 4 / sizeof(c10::BFloat16));
-    const int32_t k_group_times = k_times / 2;
-    const bool has_tail = (k_times % 2 == 1);
-
-    if (accum_c) {
-      _tile_loadd(6, c_tile_6, c_stride);
-      _tile_loadd(7, c_tile_7, c_stride);
-    } else {
-      _tile_zero(6);
-      _tile_zero(7);
-    }
-
-    for (int32_t k = 0; k < k_group_times; ++k) {
-      _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_stride);
-      _tile_dpbf16ps(6, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_stride);
-      _tile_dpbf16ps(7, 0, 3);
-      _tile_loadd(1, a_tile_1, a_tile_stride);
-      _tile_stream_loadd(4, b_tile_4, b_stride);
-      _tile_dpbf16ps(6, 1, 4);
-      _tile_stream_loadd(5, b_tile_5, b_stride);
-      _tile_dpbf16ps(7, 1, 5);
-
-      // update ptrs
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // Q buffer is prepacked
-        a_tile_0 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-        a_tile_1 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // P buffer is not prepacked
-        a_tile_0 += 2 * AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-        a_tile_1 += 2 * AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-      }
-      b_tile_2 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_3 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_4 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_5 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    }
-
-    if (has_tail) {
-      _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_stride);
-      _tile_dpbf16ps(6, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_stride);
-      _tile_dpbf16ps(7, 0, 3);
-    }
-
-    _tile_stored(6, c_tile_6, c_stride);
-    _tile_stored(7, c_tile_7, c_stride);
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    config.rows[0] = m;
-    config.rows[1] = m;
-    config.rows[2] = AMX_TILE_ROW_NUM;
-    config.rows[3] = AMX_TILE_ROW_NUM;
-    config.rows[4] = AMX_TILE_ROW_NUM;
-    config.rows[5] = AMX_TILE_ROW_NUM;
-    config.rows[6] = m;
-    config.rows[7] = m;
-    _tile_loadconfig(&config);
-  }
-};
-}  // namespace
-
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::AMX, scalar_t, head_dim> {
- public:
-  using query_t = scalar_t;
-  using q_buffer_t = scalar_t;
-  using kv_cache_t = scalar_t;
-  using logits_buffer_t = float;
-  using partial_output_buffer_t = float;
-  using prob_buffer_t = scalar_t;
-
-  constexpr static int64_t BlockSizeAlignment =
-      AMX_TILE_ROW_BYTES /
-      sizeof(kv_cache_t);  // KV token num unit of QK and PV phases
-  constexpr static int64_t HeadDimAlignment =
-      2 * (AMX_TILE_ROW_BYTES / 4);  // headdim num unit of PV phase
-  constexpr static int64_t MaxQHeadNumPerIteration = 32;
-  constexpr static int64_t HeadDim = head_dim;
-  constexpr static ISA ISAType = ISA::AMX;
-  constexpr static bool scale_on_logits = true;
-
- public:
-  AttentionImpl() : current_q_head_num_(0) {
-    // Use all columns in AMX tiles
-    vec_op::unroll_loop<int, 8>([&](int i) { amx_tile_config_.colsb[i] = 64; });
-  }
-
-  ~AttentionImpl() { _tile_release(); }
-
-  template <template <typename tile_gemm_t> typename attention>
-  FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
-    if (q_head_num > AMX_TILE_ROW_NUM) {
-      if (q_head_num != current_q_head_num_) {
-        current_q_head_num_ = q_head_num;
-        TileGemm224<kv_cache_t>::init_tile_config(q_head_num, amx_tile_config_);
-      }
-      attention<TileGemm224<kv_cache_t>> attention_iteration;
-      attention_iteration(CPU_ATTENTION_PARAMS);
-    } else {
-      if (q_head_num != current_q_head_num_) {
-        current_q_head_num_ = q_head_num;
-        TileGemm122<kv_cache_t>::init_tile_config(q_head_num, amx_tile_config_);
-      }
-      attention<TileGemm122<kv_cache_t>> attention_iteration;
-      attention_iteration(CPU_ATTENTION_PARAMS);
-    }
-  }
-
-  // k_cache_token_group_stride: stride of K cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t k_cache_token_group_stride(
-      const int32_t block_size) {
-    return BlockSizeAlignment * head_dim;
-  }
-
-  // v_cache_token_group_stride: stride of V cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t v_cache_token_group_stride(
-      const int32_t block_size) {
-    return BlockSizeAlignment * (AMX_TILE_ROW_BYTES / 4);
-  }
-
-  // v_cache_head_group_stride: stride of V cache when move to next
-  // HeadDimAlignment head dims in a block
-  constexpr static int64_t v_cache_head_group_stride(const int32_t block_size) {
-    return block_size * HeadDimAlignment;
-  }
-
-  static void copy_q_heads_tile(
-      scalar_t* __restrict__ src,  // [q_num, q_heads_per_kv, head_size]
-      scalar_t* __restrict__ q_buffer, const int32_t q_num,
-      const int32_t q_heads_per_kv, const int64_t q_num_stride,
-      const int64_t q_head_stride, const float scale) {
-    constexpr int64_t bytes_per_head = head_dim * sizeof(scalar_t);
-    static_assert(bytes_per_head % AMX_TILE_ROW_BYTES == 0);
-    constexpr int64_t head_size_block_num = bytes_per_head / AMX_TILE_ROW_BYTES;
-    constexpr int64_t head_elem_num_pre_block =
-        AMX_TILE_ROW_BYTES / sizeof(scalar_t);
-
-    int32_t idx = 0;
-    int8_t* __restrict__ q_buffer_iter = reinterpret_cast<int8_t*>(q_buffer);
-    for (int32_t q_num_idx = 0; q_num_idx < q_num;
-         ++q_num_idx, src += q_num_stride) {
-      scalar_t* __restrict__ src_iter = src;
-      for (int32_t q_head_idx = 0; q_head_idx < q_heads_per_kv;
-           ++q_head_idx, src_iter += q_head_stride) {
-        vec_op::unroll_loop<int32_t, head_size_block_num>(
-            [&](int32_t head_size_block_idx) {
-              // Use INT8Vec64 for 64 bytes block
-              vec_op::INT8Vec64 vec(src_iter + head_size_block_idx *
-                                                   head_elem_num_pre_block);
-              vec.save(q_buffer_iter + head_size_block_idx * AMX_TILE_BYTES);
-            });
-
-        ++idx;
-        q_buffer_iter += AMX_TILE_ROW_BYTES;
-        if ((idx & (AMX_TILE_ROW_NUM - 1)) == 0) {
-          // head is in another amx tile
-          q_buffer_iter -= AMX_TILE_ROW_NUM * AMX_TILE_ROW_BYTES;
-          q_buffer_iter += head_size_block_num * AMX_TILE_BYTES;
-        }
-      }
-    }
-  }
-
-  // reshape KV to AMX friendly layout
-  static void reshape_and_cache(
-      const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
-      scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
-      const int64_t* __restrict__ slot_mapping, const int64_t token_num,
-      const int64_t key_token_num_stride, const int64_t value_token_num_stride,
-      const int64_t head_num, const int64_t key_head_num_stride,
-      const int64_t value_head_num_stride, const int64_t num_blocks,
-      const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
-      const int64_t block_size, const int64_t block_size_stride) {
-    // For AMX 2D tiles, size of each line is 64 bytes
-    constexpr int64_t amx_tile_row_size = AMX_TILE_ROW_BYTES;
-    // For AMX B martix, N always is 16
-    constexpr int64_t amx_b_tile_n_size = AMX_TILE_ROW_BYTES / 4;
-    constexpr int64_t amx_b_tile_k_size = amx_tile_row_size / sizeof(scalar_t);
-    // For now suppose block_size is divisible by amx_tile_column_num
-    TORCH_CHECK_EQ(block_size % amx_b_tile_k_size, 0);
-
-#pragma omp parallel for collapse(2)
-    for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
-      for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
-        const int64_t pos = slot_mapping[token_idx];
-        if (pos < 0) {
-          // skip
-          continue;
-        }
-
-        const int64_t block_idx = pos / block_size;
-        const int64_t block_offset = pos % block_size;
-        {
-          // Write Key
-          // Head elements should be packed as quand-words and stored in token
-          // groups with (quadword_stride/4) tokens
-          constexpr int64_t token_num_per_group = amx_tile_row_size / 4;
-          static_assert(head_dim % (4 / sizeof(scalar_t)) == 0);
-          constexpr int64_t quadword_num = head_dim / (4 / sizeof(scalar_t));
-          const int32_t* key_start_quadword_ptr =
-              reinterpret_cast<const int32_t*>(
-                  key + token_idx * key_token_num_stride +
-                  head_idx * key_head_num_stride);
-          const int64_t group_idx = block_offset / token_num_per_group;
-          const int64_t group_offset = block_offset % token_num_per_group;
-          constexpr int64_t quadword_num_per_group =
-              token_num_per_group * quadword_num;
-          int32_t* key_cache_start_ptr =
-              reinterpret_cast<int32_t*>(key_cache +
-                                         block_idx * num_blocks_stride +
-                                         head_idx * cache_head_num_stride) +
-              group_idx * quadword_num_per_group + group_offset;
-
-#pragma GCC unroll 8
-          for (int64_t i = 0, j = 0; j < quadword_num;
-               i += token_num_per_group, ++j) {
-            key_cache_start_ptr[i] = key_start_quadword_ptr[j];
-          }
-        }
-        {
-          // Write Value
-          // Different from Key, block_size dimension is packed rather than
-          // head_size dimension block_size dimension is packed as quand-words;
-          constexpr int64_t token_num_per_sub_group = 4 / sizeof(scalar_t);
-          const int64_t token_num_per_group = block_size;
-          constexpr int64_t head_elems_per_group = amx_b_tile_n_size;
-          const int64_t group_size = token_num_per_group * head_elems_per_group;
-          // For now suppose head_dim is divisible by amx_b_tile_n_size
-          static_assert(head_dim % head_elems_per_group == 0);
-          constexpr int64_t group_num = head_dim / head_elems_per_group;
-          const int64_t sub_group_idx = block_offset / token_num_per_sub_group;
-          const int64_t sub_group_offset =
-              block_offset % token_num_per_sub_group;
-
-          const scalar_t* value_start_ptr = value +
-                                            token_idx * value_token_num_stride +
-                                            head_idx * value_head_num_stride;
-          scalar_t* value_cache_start_ptr =
-              value_cache + block_idx * num_blocks_stride +
-              head_idx * cache_head_num_stride +
-              sub_group_idx * token_num_per_sub_group * amx_b_tile_n_size +
-              sub_group_offset;
-
-          for (int64_t i = 0; i < group_num; ++i) {
-#pragma GCC unroll head_elems_per_group
-            for (int64_t j = 0, k = 0; j < head_elems_per_group;
-                 ++j, k += token_num_per_sub_group) {
-              value_cache_start_ptr[k] = value_start_ptr[j];
-            }
-            value_start_ptr += head_elems_per_group;
-            value_cache_start_ptr += group_size;
-          }
-        }
-      }
-    }
-  }
-
- private:
-  alignas(64) __tilecfg amx_tile_config_;
-  int32_t current_q_head_num_;
-};
-}  // namespace cpu_attention
-
-#endif

csrc/cpu/cpu_attn_macros.h

@@ -1,63 +0,0 @@
-#ifndef CPU_ATTN_MACROS_H
-#define CPU_ATTN_MACROS_H
-
-// x86_64
-#ifdef __x86_64__
-  #define FAST_SPINNING _mm_pause();
-
-  #ifdef __AVX512F__
-    #define DEFINE_FAST_EXP                                                    \
-      const __m512 vec_factorial_1 = _mm512_set1_ps(0.999999701f);             \
-      const __m512 vec_factorial_2 = _mm512_set1_ps(0.499991506f);             \
-      const __m512 vec_factorial_3 = _mm512_set1_ps(0.166676521f);             \
-      const __m512 vec_factorial_4 = _mm512_set1_ps(0.0418978221f);            \
-      const __m512 vec_factorial_5 = _mm512_set1_ps(0.00828929059f);           \
-      const __m512 vec_exp_log2ef =                                            \
-          _mm512_castsi512_ps(_mm512_set1_epi32(0x3fb8aa3b));                  \
-      const __m512 vec_half = _mm512_set1_ps(0.5f);                            \
-      const __m512 vec_one = _mm512_set1_ps(1.f);                              \
-      const __m512 vec_zero = _mm512_set1_ps(0.f);                             \
-      const __m512 vec_two = _mm512_set1_ps(2.f);                              \
-      const __m512 vec_ln2f =                                                  \
-          _mm512_castsi512_ps(_mm512_set1_epi32(0x3f317218));                  \
-      const __m512 vec_ln_flt_min =                                            \
-          _mm512_castsi512_ps(_mm512_set1_epi32(0xc2aeac50));                  \
-      const __m512 vec_ln_flt_max =                                            \
-          _mm512_castsi512_ps(_mm512_set1_epi32(0x42b17218));                  \
-      const __m512i vec_127 = _mm512_set1_epi32(0x0000007f);                   \
-      const int n_mantissa_bits = 23;                                          \
-      auto fast_exp = [&](vec_op::FP32Vec16& vec) __attribute__((              \
-                          always_inline)) {                                    \
-        __m512 values = vec.reg;                                               \
-        auto less_ln_flt_min_mask =                                            \
-            _mm512_cmp_ps_mask(values, vec_ln_flt_min, 1 /*_CMP_LT_OS*/);      \
-        auto vec_src = _mm512_min_ps(values, vec_ln_flt_max);                  \
-        vec_src = _mm512_max_ps(vec_src, vec_ln_flt_min);                      \
-        auto vec_fx = _mm512_fmadd_ps(vec_src, vec_exp_log2ef, vec_half);      \
-        auto vec_fx_i = _mm512_cvt_roundps_epi32(                              \
-            vec_fx, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC);                \
-        vec_fx = _mm512_cvtepi32_ps(vec_fx_i);                                 \
-        auto vec_exp_poly = _mm512_fnmadd_ps(vec_fx, vec_ln2f, vec_src);       \
-        auto vec_res =                                                         \
-            _mm512_fmadd_ps(vec_exp_poly, vec_factorial_5, vec_factorial_4);   \
-        vec_res = _mm512_fmadd_ps(vec_exp_poly, vec_res, vec_factorial_3);     \
-        vec_res = _mm512_fmadd_ps(vec_exp_poly, vec_res, vec_factorial_2);     \
-        vec_res = _mm512_fmadd_ps(vec_exp_poly, vec_res, vec_factorial_1);     \
-        vec_res = _mm512_fmadd_ps(vec_exp_poly, vec_res, vec_one);             \
-        auto vec_exp_number = _mm512_sub_ps(vec_fx, vec_one);                  \
-        auto vec_exp_number_i = _mm512_cvtps_epi32(vec_exp_number);            \
-        auto vec_two_pow_n_i = _mm512_add_epi32(vec_exp_number_i, vec_127);    \
-        vec_two_pow_n_i = _mm512_slli_epi32(vec_two_pow_n_i, n_mantissa_bits); \
-        auto vec_two_pow_n = _mm512_castsi512_ps(vec_two_pow_n_i);             \
-        vec_two_pow_n = _mm512_mask_blend_ps(less_ln_flt_min_mask,             \
-                                             vec_two_pow_n, vec_zero);         \
-        vec_res = _mm512_mul_ps(vec_res, vec_two_pow_n);                       \
-        vec_res = _mm512_mul_ps(vec_res, vec_two);                             \
-        vec_op::FP32Vec16 res(vec_res);                                        \
-        return res;                                                            \
-      };
-  #endif
-
-#endif
-
-#endif

csrc/cpu/cpu_attn_vec.hpp

@@ -1,248 +0,0 @@
-#ifndef CPU_ATTN_VEC_HPP
-#define CPU_ATTN_VEC_HPP
-
-#include "cpu_attn_impl.hpp"
-
-namespace cpu_attention {
-
-namespace {
-// 8-2-16 pattern, 8 regs for A, 2 regs for B, 16 regs for C, [8, K] @ [k, 32]
-template <typename kv_cache_t>
-class TileGemm82 {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size,
-                                float* __restrict__ a_tile,
-                                kv_cache_t* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    switch (m_size) {
-      case 1:
-        gemm_micro<1>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 2:
-        gemm_micro<2>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 3:
-      case 4:
-        gemm_micro<4>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 5:
-      case 6:
-        gemm_micro<6>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 7:
-      case 8:
-        gemm_micro<8>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-    }
-  }
-
-  template <int32_t M>
-  static void gemm_micro(float* __restrict__ a_tile,
-                         kv_cache_t* __restrict__ b_tile,
-                         float* __restrict__ c_tile, const int64_t lda,
-                         const int64_t ldb, const int64_t ldc,
-                         const int32_t block_size, const int32_t dynamic_k_size,
-                         const bool accum_c) {
-    static_assert(0 < M <= 8);
-    using load_vec_t = typename VecTypeTrait<kv_cache_t>::vec_t;
-
-    kv_cache_t* __restrict__ curr_b_0 = b_tile;
-    kv_cache_t* __restrict__ curr_b_1 = b_tile + 16;
-    float* __restrict__ curr_c_0 = c_tile;
-    float* __restrict__ curr_c_1 = c_tile + 16;
-
-    vec_op::FP32Vec16 c_regs[M * 2];
-    if (accum_c) {
-      float* __restrict__ curr_m_c_0 = curr_c_0;
-      float* __restrict__ curr_m_c_1 = curr_c_1;
-      vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-        c_regs[i * 2] = vec_op::FP32Vec16(curr_m_c_0);
-        c_regs[i * 2 + 1] = vec_op::FP32Vec16(curr_m_c_1);
-
-        // update
-        curr_m_c_0 += ldc;
-        curr_m_c_1 += ldc;
-      });
-    }
-
-    float* __restrict__ curr_a = a_tile;
-    for (int32_t k = 0; k < dynamic_k_size; ++k) {
-      load_vec_t b_0_reg(curr_b_0);
-      vec_op::FP32Vec16 fp32_b_0_reg(b_0_reg);
-      load_vec_t b_1_reg(curr_b_1);
-      vec_op::FP32Vec16 fp32_b_1_reg(b_1_reg);
-
-      float* __restrict__ curr_m_a = curr_a;
-      vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-        float v = *curr_m_a;
-        vec_op::FP32Vec16 a_reg(v);
-        c_regs[i * 2] = c_regs[i * 2] + a_reg * fp32_b_0_reg;
-        c_regs[i * 2 + 1] = c_regs[i * 2 + 1] + a_reg * fp32_b_1_reg;
-
-        // update
-        curr_m_a += lda;
-      });
-
-      // update
-      curr_a += 1;
-      curr_b_0 += ldb;
-      curr_b_1 += ldb;
-    }
-
-    vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-      c_regs[i * 2].save(curr_c_0);
-      c_regs[i * 2 + 1].save(curr_c_1);
-
-      // update
-      curr_c_0 += ldc;
-      curr_c_1 += ldc;
-    });
-  }
-};
-}  // namespace
-
-// This is a general but naive implementation based on vector instructions
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::VEC, scalar_t, head_dim> {
- public:
-  using query_t = scalar_t;
-  using q_buffer_t = float;
-  using kv_cache_t = scalar_t;
-  using logits_buffer_t = float;
-  using partial_output_buffer_t = float;
-  using prob_buffer_t = float;
-
-  constexpr static int64_t BlockSizeAlignment =
-      32;  // KV token num unit of QK and PV phases
-  constexpr static int64_t HeadDimAlignment =
-      32;  // headdim num unit of PV phase
-  constexpr static int64_t MaxQHeadNumPerIteration = 8;
-  constexpr static int64_t HeadDim = head_dim;
-  constexpr static ISA ISAType = ISA::VEC;
-  constexpr static bool scale_on_logits = false;  // apply scale on q_buffer
-
- public:
-  template <template <typename tile_gemm_t> typename attention>
-  FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
-    attention<TileGemm82<kv_cache_t>> attention_iteration;
-    attention_iteration(CPU_ATTENTION_PARAMS);
-  }
-
-  // k_cache_token_group_stride: stride of K cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t k_cache_token_group_stride(
-      const int32_t block_size) {
-    return BlockSizeAlignment;  // layout of k_cache block is [head_dim,
-                                // block_size], row-major
-  }
-
-  // v_cache_token_group_stride: stride of V cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t v_cache_token_group_stride(
-      const int32_t block_size) {
-    return head_dim * BlockSizeAlignment;  // layout of v_cache is [block_size,
-                                           // head_dim], row-major
-  }
-
-  // v_cache_head_group_stride: stride of V cache when move to next
-  // HeadDimAlignment head dims in a block
-  constexpr static int64_t v_cache_head_group_stride(const int32_t block_size) {
-    return HeadDimAlignment;  // layout of v_cache is [block_size, head_dim],
-                              // row-major
-  }
-
-  // Copy q to q_buffer and cast it to fp32
-  static void copy_q_heads_tile(
-      scalar_t* __restrict__ src,  // [q_num, q_heads_per_kv, head_size]
-      float* __restrict__ q_buffer, const int32_t q_num,
-      const int32_t q_heads_per_kv, const int64_t q_num_stride,
-      const int64_t q_head_stride, float scale) {
-    static_assert(head_dim % 16 == 0);
-    constexpr int32_t unroll_size = head_dim / 16;
-    using load_vec_t = typename VecTypeTrait<scalar_t>::vec_t;
-
-    vec_op::FP32Vec16 scale_vec(scale);
-    for (int32_t q_num_idx = 0; q_num_idx < q_num; ++q_num_idx) {
-      for (int32_t q_head_idx = 0; q_head_idx < q_heads_per_kv; ++q_head_idx) {
-        scalar_t* __restrict__ curr_q =
-            src + q_num_idx * q_num_stride + q_head_idx * q_head_stride;
-        float* __restrict__ curr_q_buffer =
-            q_buffer + q_num_idx * q_heads_per_kv * head_dim +
-            q_head_idx * head_dim;
-
-        vec_op::unroll_loop<int32_t, unroll_size>([&](int32_t i) {
-          load_vec_t vec(curr_q);
-          vec_op::FP32Vec16 fp32_vec(vec);
-          fp32_vec = fp32_vec * scale_vec;
-          fp32_vec.save(curr_q_buffer);
-
-          curr_q += 16;
-          curr_q_buffer += 16;
-        });
-      }
-    }
-  }
-
-  // reshape K as column-major and V as row-major
-  static void reshape_and_cache(
-      const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
-      scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
-      const int64_t* __restrict__ slot_mapping, const int64_t token_num,
-      const int64_t key_token_num_stride, const int64_t value_token_num_stride,
-      const int64_t head_num, const int64_t key_head_num_stride,
-      const int64_t value_head_num_stride, const int64_t num_blocks,
-      const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
-      const int64_t block_size, const int64_t block_size_stride) {
-#pragma omp parallel for collapse(2)
-    for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
-      for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
-        const int64_t pos = slot_mapping[token_idx];
-        if (pos < 0) {
-          // skip
-          continue;
-        }
-
-        const int64_t block_idx = pos / block_size;
-        const int64_t block_offset = pos % block_size;
-        {
-          // Write Key as column-major
-          const scalar_t* key_start_ptr = key +
-                                          token_idx * key_token_num_stride +
-                                          head_idx * key_head_num_stride;
-          scalar_t* key_cache_start_ptr =
-              key_cache + block_idx * num_blocks_stride +
-              head_idx * cache_head_num_stride + block_offset;
-
-#pragma GCC unroll 8
-          for (int64_t i = 0, j = 0; i < head_dim; ++i, j += block_size) {
-            key_cache_start_ptr[j] = key_start_ptr[i];
-          }
-        }
-        {
-          // Write Value as row-major
-          const scalar_t* value_start_ptr = value +
-                                            token_idx * value_token_num_stride +
-                                            head_idx * value_head_num_stride;
-          scalar_t* value_cache_start_ptr =
-              value_cache + block_idx * num_blocks_stride +
-              head_idx * cache_head_num_stride + block_offset * head_dim;
-          std::memcpy(value_cache_start_ptr, value_start_ptr,
-                      sizeof(scalar_t) * head_dim);
-        }
-      }
-    }
-  }
-};
-}  // namespace cpu_attention
-
-#endif

csrc/cpu/cpu_attn_vec16.hpp

@@ -1,171 +0,0 @@
-#ifndef CPU_ATTN_VEC16_HPP
-#define CPU_ATTN_VEC16_HPP
-
-#include "cpu_attn_vec.hpp"
-
-namespace cpu_attention {
-
-namespace {
-// 16-1-16 pattern, 16 regs for A, 1 regs for B, 16 regs for C, [16, K] @ [k,
-// 16]
-template <typename kv_cache_t>
-class TileGemm161 {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size,
-                                float* __restrict__ a_tile,
-                                kv_cache_t* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    switch (m_size) {
-      case 1:
-        gemm_micro<1>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 2:
-        gemm_micro<2>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 3:
-      case 4:
-        gemm_micro<4>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 5:
-      case 6:
-        gemm_micro<6>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 7:
-      case 8:
-        gemm_micro<8>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 9:
-      case 10:
-      case 11:
-      case 12:
-        gemm_micro<12>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                       dynamic_k_size, accum_c);
-        break;
-      case 13:
-      case 14:
-      case 15:
-      case 16:
-        gemm_micro<16>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                       dynamic_k_size, accum_c);
-        break;
-    }
-  }
-
-  template <int32_t M>
-  static void gemm_micro(float* __restrict__ a_tile,
-                         kv_cache_t* __restrict__ b_tile,
-                         float* __restrict__ c_tile, const int64_t lda,
-                         const int64_t ldb, const int64_t ldc,
-                         const int32_t block_size, const int32_t dynamic_k_size,
-                         const bool accum_c) {
-    static_assert(0 < M <= 16);
-    using load_vec_t = typename VecTypeTrait<kv_cache_t>::vec_t;
-
-    kv_cache_t* __restrict__ curr_b_0 = b_tile;
-    float* __restrict__ curr_c_0 = c_tile;
-
-    vec_op::FP32Vec16 c_regs[M];
-    if (accum_c) {
-      float* __restrict__ curr_m_c_0 = curr_c_0;
-      vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-        c_regs[i] = vec_op::FP32Vec16(curr_m_c_0);
-
-        // update
-        curr_m_c_0 += ldc;
-      });
-    }
-
-    float* __restrict__ curr_a = a_tile;
-    for (int32_t k = 0; k < dynamic_k_size; ++k) {
-      load_vec_t b_0_reg(curr_b_0);
-      vec_op::FP32Vec16 fp32_b_0_reg(b_0_reg);
-
-      float* __restrict__ curr_m_a = curr_a;
-      vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-        float v = *curr_m_a;
-        vec_op::FP32Vec16 a_reg(v);
-        c_regs[i] = c_regs[i] + a_reg * fp32_b_0_reg;
-
-        // update
-        curr_m_a += lda;
-      });
-
-      // update
-      curr_a += 1;
-      curr_b_0 += ldb;
-    }
-
-    vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-      c_regs[i].save(curr_c_0);
-
-      // update
-      curr_c_0 += ldc;
-    });
-  }
-};
-}  // namespace
-
-// This is a general but naive implementation based on vector instructions
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::VEC16, scalar_t, head_dim>
-    : public AttentionImpl<ISA::VEC, scalar_t, head_dim> {
- public:
-  using query_t = scalar_t;
-  using q_buffer_t = float;
-  using kv_cache_t = scalar_t;
-  using logits_buffer_t = float;
-  using partial_output_buffer_t = float;
-  using prob_buffer_t = float;
-
-  constexpr static int64_t BlockSizeAlignment =
-      16;  // KV token num unit of QK and PV phases
-  constexpr static int64_t HeadDimAlignment =
-      16;  // headdim num unit of PV phase
-  constexpr static int64_t MaxQHeadNumPerIteration = 16;
-  constexpr static int64_t HeadDim = head_dim;
-  constexpr static ISA ISAType = ISA::VEC16;
-  constexpr static bool scale_on_logits = false;  // apply scale on q_buffer
-
- public:
-  template <template <typename tile_gemm_t> typename attention>
-  FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
-    attention<TileGemm161<kv_cache_t>> attention_iteration;
-    attention_iteration(CPU_ATTENTION_PARAMS);
-  }
-
-  // k_cache_token_group_stride: stride of K cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t k_cache_token_group_stride(
-      const int32_t block_size) {
-    return BlockSizeAlignment;  // layout of k_cache block is [head_dim,
-                                // block_size], row-major
-  }
-
-  // v_cache_token_group_stride: stride of V cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t v_cache_token_group_stride(
-      const int32_t block_size) {
-    return head_dim * BlockSizeAlignment;  // layout of v_cache is [block_size,
-                                           // head_dim], row-major
-  }
-
-  // v_cache_head_group_stride: stride of V cache when move to next
-  // HeadDimAlignment head dims in a block
-  constexpr static int64_t v_cache_head_group_stride(const int32_t block_size) {
-    return HeadDimAlignment;  // layout of v_cache is [block_size, head_dim],
-                              // row-major
-  }
-};
-}  // namespace cpu_attention
-
-#endif

csrc/cpu/cpu_types_x86.hpp

@@ -40,23 +40,6 @@ namespace vec_op {
 
 #define FORCE_INLINE __attribute__((always_inline)) inline
 
-// Function to get the timestamp using RDTSCP
-FORCE_INLINE uint64_t bench_timestamp() {
-  unsigned int cycles_low, cycles_high;
-  asm volatile(
-      ".intel_syntax noprefix\n\t"
-      "CPUID\n\t"        // Serialize instruction stream to ensure previous
-                         // instructions complete
-      "RDTSCP\n\t"       // Read TSC and core ID
-      "mov %0, edx\n\t"  // Store high 32 bits of TSC
-      "mov %1, eax\n\t"  // Store low 32 bits of TSC
-      ".att_syntax"
-      : "=r"(cycles_high), "=r"(cycles_low)::"rax", "rbx", "rcx",
-        "rdx"  // Clobbered registers
-  );
-  return (uint64_t)cycles_high << 32 | cycles_low;
-}
-
 namespace {
 template <typename T, T... indexes, typename F>
 constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F&& f) {
@@ -424,8 +407,6 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   float reduce_min() const { return _mm512_reduce_min_ps(reg); }
 
-  float get_last_elem() const { return _mm512_cvtss_f32(reg); }
-
   template <int group_size>
   float reduce_sub_sum(int idx) {
     static_assert(VEC_ELEM_NUM % group_size == 0);
@@ -465,6 +446,9 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   explicit FP32Vec16(__m256 low, __m256 high) : reg_low(low), reg_high(high) {}
 
+  explicit FP32Vec16(const FP32Vec16& data)
+      : reg_low(data.reg_low), reg_high(data.reg_high) {}
+
   explicit FP32Vec16(const FP32Vec4& data)
       : reg_low((__m256)_mm256_inserti128_si256(
             _mm256_castsi128_si256((__m128i)data.reg), (__m128i)data.reg, 1)),
@@ -520,32 +504,6 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
                      _mm256_div_ps(reg_high, b.reg_high));
   }
 
-  FP32Vec16 max(const FP32Vec16& b) const {
-    return FP32Vec16(_mm256_max_ps(reg_low, b.reg_low),
-                     _mm256_max_ps(reg_high, b.reg_high));
-  }
-
-  float reduce_max() const {
-    __m256 v = _mm256_max_ps(reg_low, reg_high);
-    // Permute to compare elements within 128-bit lanes
-    __m256 v_shuffled = _mm256_permute_ps(
-        v, 0b00001011);  // Swap halves within each 128-bit lane
-    __m256 v_max = _mm256_max_ps(v, v_shuffled);
-
-    v_shuffled = _mm256_permute_ps(
-        v_max, 0b00000001);  // Shuffle elements within each 128-bit lane
-    v_max = _mm256_max_ps(v_max, v_shuffled);
-
-    // Permute to compare elements between 128-bit lanes
-    v_shuffled =
-        _mm256_permute2f128_ps(v_max, v_max, 0b00000001);  // Swap 128-bit lanes
-    v_max = _mm256_max_ps(v_max, v_shuffled);
-
-    // At this point, the maximum value is present in all elements of v_max.
-    // Extract the first element for the scalar result.
-    return _mm256_cvtss_f32(v_max);  // Extract the lowest 32-bit float
-  }
-
   float reduce_sum() const {
     FP32Vec8 low = FP32Vec8(reg_low);
     FP32Vec8 high = FP32Vec8(reg_high);
@@ -684,7 +642,7 @@ inline FP16Vec16::FP16Vec16(const FP32Vec16& v)
 inline FP16Vec16::FP16Vec16(const FP32Vec16& v)
     : reg(_mm256_insertf128_si256(
           _mm256_castsi128_si256(FP16Vec8(FP32Vec8(v.reg_low)).reg),
-          FP16Vec8(FP32Vec8(v.reg_high)).reg, 1)) {}
+          FP16Vec8(FP32Vec8(v.reg_low)).reg, 1)) {}
 #endif
 
 #ifdef __AVX512BF16__

csrc/cpu/dnnl_helper.cpp

@@ -5,7 +5,6 @@
 #include "common/memory.hpp"
 
 #include "dnnl_helper.h"
-#include "scratchpad_manager.h"
 
 static dnnl::engine& default_engine() {
   static dnnl::engine engine(dnnl::engine::kind::cpu, 0);
@@ -23,6 +22,23 @@ void release_dnnl_matmul_handler(int64_t handler) {
   delete ptr;
 }
 
+DNNLScratchPadManager::DNNLScratchPadManager() : size_(0), ptr_(nullptr) {
+  this->realloc(allocation_unit * 128);
+}
+
+void DNNLScratchPadManager::realloc(size_t new_size) {
+  new_size = round(new_size);
+  if (new_size > size_) {
+    ptr_ = std::aligned_alloc(64, new_size);
+    size_ = new_size;
+  }
+}
+
+DNNLScratchPadManager* DNNLScratchPadManager::get_dnnl_scratchpad_manager() {
+  static DNNLScratchPadManager manager;
+  return &manager;
+}
+
 template <typename KT, typename VT>
 class DNNLPrimitiveCache {
  public:

csrc/cpu/dnnl_helper.h

@@ -59,6 +59,30 @@ constexpr inline dnnl::memory::data_type get_dnnl_type() {
   return DNNLType<std::decay_t<T>>::type;
 }
 
+class DNNLScratchPadManager {
+ public:
+  static constexpr size_t allocation_unit = 4 * 1024 * 1024;  // 4KB
+
+  static DNNLScratchPadManager* get_dnnl_scratchpad_manager();
+
+  DNNLScratchPadManager();
+
+  template <typename T>
+  T* get_data() {
+    return reinterpret_cast<T*>(ptr_);
+  }
+
+  static size_t round(size_t size) {
+    return ((size + allocation_unit - 1) / allocation_unit) * allocation_unit;
+  }
+
+  void realloc(size_t new_size);
+
+ private:
+  size_t size_;
+  void* ptr_;
+};
+
 class DNNLMatMulPrimitiveHandler {
  public:
   virtual ~DNNLMatMulPrimitiveHandler() = default;

csrc/cpu/scratchpad_manager.cpp

@@ -1,23 +0,0 @@
-#include <cstdlib>
-
-#include "scratchpad_manager.h"
-
-DNNLScratchPadManager::DNNLScratchPadManager() : size_(0), ptr_(nullptr) {
-  this->realloc(allocation_unit * 128);
-}
-
-void DNNLScratchPadManager::realloc(size_t new_size) {
-  new_size = round(new_size);
-  if (new_size > size_) {
-    if (ptr_ != nullptr) {
-      std::free(ptr_);
-    }
-    ptr_ = std::aligned_alloc(64, new_size);
-    size_ = new_size;
-  }
-}
-
-DNNLScratchPadManager* DNNLScratchPadManager::get_dnnl_scratchpad_manager() {
-  static DNNLScratchPadManager manager;
-  return &manager;
-}

csrc/cpu/scratchpad_manager.h

@@ -1,31 +0,0 @@
-#ifndef SCRATCHPAD_MANAGER_H
-#define SCRATCHPAD_MANAGER_H
-
-#include <cstddef>
-#include <cstdio>
-
-class DNNLScratchPadManager {
- public:
-  static constexpr size_t allocation_unit = 4 * 1024;  // 4KB
-
-  static DNNLScratchPadManager* get_dnnl_scratchpad_manager();
-
-  DNNLScratchPadManager();
-
-  template <typename T>
-  T* get_data() {
-    return reinterpret_cast<T*>(ptr_);
-  }
-
-  static size_t round(size_t size) {
-    return ((size + allocation_unit - 1) / allocation_unit) * allocation_unit;
-  }
-
-  void realloc(size_t new_size);
-
- private:
-  size_t size_;
-  void* ptr_;
-};
-
-#endif

csrc/cpu/shm.cpp

@@ -192,7 +192,7 @@ class SHMManager {
                       const int group_size)
       : _rank(rank),
         _group_size(group_size),
-        _thread_num(omp_get_max_threads()),
+        _thread_num(torch::get_num_threads()),
         _shm_names({""}),
         _shared_mem_ptrs({nullptr}),
         _shm_ctx(nullptr) {

csrc/cpu/torch_bindings.cpp

@@ -74,38 +74,25 @@ at::Tensor int8_scaled_mm_with_quant(at::Tensor& mat1, at::Tensor& mat2,
                                      const std::optional<at::Tensor>& bias,
                                      at::ScalarType out_dtype, bool is_vnni);
 
-torch::Tensor get_scheduler_metadata(
-    const int64_t num_req, const int64_t num_heads_q,
-    const int64_t num_heads_kv, const int64_t head_dim,
-    const torch::Tensor& seq_lens, at::ScalarType dtype,
-    const torch::Tensor& query_start_loc, const bool casual,
-    const int64_t window_size, const std::string& isa_hint,
-    const bool enable_kv_split);
-
-void cpu_attn_reshape_and_cache(const torch::Tensor& key,
-                                const torch::Tensor& value,
-                                torch::Tensor& key_cache,
-                                torch::Tensor& value_cache,
-                                const torch::Tensor& slot_mapping,
-                                const std::string& isa);
-
-void cpu_attention_with_kv_cache(
-    const torch::Tensor& query, const torch::Tensor& key_cache,
-    const torch::Tensor& value_cache, torch::Tensor& output,
-    const torch::Tensor& query_start_loc, const torch::Tensor& seq_lens,
-    const double scale, const bool causal,
-    const std::optional<torch::Tensor>& alibi_slopes,
-    const int64_t sliding_window_left, const int64_t sliding_window_right,
-    const torch::Tensor& block_table, const double softcap,
-    const torch::Tensor& scheduler_metadata,
-    const std::optional<torch::Tensor>& s_aux);
-
-// Note: just for avoiding importing errors
-void placeholder_op() { TORCH_CHECK(false, "Unimplemented"); }
-
 TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // vLLM custom ops
 
+  // Attention ops
+  // Compute the attention between an input query and the cached keys/values
+  // using PagedAttention.
+  ops.def(
+      "paged_attention_v1("
+      "    Tensor! out, Tensor query, Tensor key_cache,"
+      "    Tensor value_cache, int num_kv_heads, float scale,"
+      "    Tensor block_tables, Tensor seq_lens, int block_size,"
+      "    int max_seq_len, Tensor? alibi_slopes,"
+      "    str kv_cache_dtype, Tensor k_scale, Tensor v_scale,"
+      "    int tp_rank, int blocksparse_local_blocks,"
+      "    int blocksparse_vert_stride, int blocksparse_block_size,"
+      "    int blocksparse_head_sliding_step) -> ()");
+
+  ops.impl("paged_attention_v1", torch::kCPU, &paged_attention_v1);
+
   ops.def(
       "dynamic_4bit_int_moe("
       "Tensor x, Tensor topk_ids, Tensor topk_weights,"
@@ -115,6 +102,20 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 
   ops.impl("dynamic_4bit_int_moe", torch::kCPU, &dynamic_4bit_int_moe_cpu);
 
+  // PagedAttention V2.
+  ops.def(
+      "paged_attention_v2("
+      "    Tensor! out, Tensor! exp_sums, Tensor! max_logits,"
+      "    Tensor! tmp_out, Tensor query, Tensor key_cache,"
+      "    Tensor value_cache, int num_kv_heads, float scale,"
+      "    Tensor block_tables, Tensor seq_lens, int block_size,"
+      "    int max_seq_len, Tensor? alibi_slopes,"
+      "    str kv_cache_dtype, Tensor k_scale, Tensor v_scale,"
+      "    int tp_rank, int blocksparse_local_blocks,"
+      "    int blocksparse_vert_stride, int blocksparse_block_size,"
+      "    int blocksparse_head_sliding_step) -> ()");
+  ops.impl("paged_attention_v2", torch::kCPU, &paged_attention_v2);
+
   // Activation ops
 
   // Activation function used in SwiGLU.
@@ -258,31 +259,37 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("int8_scaled_mm_with_quant", torch::kCPU,
            &int8_scaled_mm_with_quant);
 #endif
+}
 
-  // CPU attention kernels
-  ops.def(
-      "get_scheduler_metadata(int num_req, int num_heads_q, int num_heads_kv, "
-      "int head_dim, Tensor seq_lens, ScalarType dtype, Tensor "
-      "query_start_loc, bool casual, int window_size, str isa_hint, bool "
-      "enable_kv_split) -> Tensor",
-      &get_scheduler_metadata);
-  ops.def(
-      "cpu_attn_reshape_and_cache(Tensor key, Tensor value, Tensor(a2!) "
-      "key_cache, Tensor(a3!) value_cache, Tensor slot_mapping, str "
-      "isa) -> ()",
-      &cpu_attn_reshape_and_cache);
-  ops.def(
-      "cpu_attention_with_kv_cache(Tensor query, Tensor key_cache, Tensor "
-      "value_cache, Tensor(a3!) output, Tensor query_start_loc, Tensor "
-      "seq_lens, float scale, bool causal, Tensor? alibi_slopes, SymInt "
-      "sliding_window_left, SymInt sliding_window_right, Tensor block_table, "
-      "float softcap, Tensor sheduler_metadata, Tensor? s_aux) -> ()",
-      &cpu_attention_with_kv_cache);
+TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
+  // Cache ops
+  // Swap in (out) the cache blocks from src to dst.
+  cache_ops.def(
+      "swap_blocks(Tensor src, Tensor! dst, Tensor block_mapping) -> ()");
+  cache_ops.impl("swap_blocks", torch::kCPU, &swap_blocks);
 
-  // placeholders
-  ops.def("static_scaled_fp8_quant() -> ()", placeholder_op);
-  ops.def("dynamic_scaled_fp8_quant() -> ()", placeholder_op);
-  ops.def("dynamic_per_token_scaled_fp8_quant() -> ()", placeholder_op);
+  // Copy the cache blocks from src to dst.
+  cache_ops.def(
+      "copy_blocks(Tensor(a!)[] key_caches, Tensor[](b!) value_caches, "
+      "Tensor block_mapping) -> ()");
+  cache_ops.impl("copy_blocks", torch::kCPU, &copy_blocks);
+
+  // Reshape the key and value tensors and cache them.
+  cache_ops.def(
+      "reshape_and_cache(Tensor key, Tensor value,"
+      "                  Tensor! key_cache, Tensor! value_cache,"
+      "                  Tensor slot_mapping,"
+      "                  str kv_cache_dtype,"
+      "                  Tensor k_scale, Tensor v_scale) -> ()");
+  cache_ops.impl("reshape_and_cache", torch::kCPU, &reshape_and_cache);
+
+  cache_ops.def(
+      "concat_and_cache_mla(Tensor kv_c, Tensor k_pe,"
+      "                     Tensor! kv_cache,"
+      "                     Tensor slot_mapping,"
+      "                     str kv_cache_dtype,"
+      "                     Tensor scale) -> ()");
+  cache_ops.impl("concat_and_cache_mla", torch::kCPU, &concat_and_cache_mla);
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _utils), utils) {

csrc/cpu/utils.cpp

@@ -45,16 +45,6 @@ std::string init_cpu_threads_env(const std::string& cpu_ids) {
   // Memory node binding
   if (numa_available() != -1) {
     int mem_node_id = numa_node_of_cpu(omp_cpu_ids.front());
-    // Verify all CPUs are on the same NUMA node
-    for (size_t i = 1; i < omp_cpu_ids.size(); ++i) {
-      int node_id = numa_node_of_cpu(omp_cpu_ids[i]);
-      TORCH_CHECK(node_id == mem_node_id, "CPU ", omp_cpu_ids[i],
-                  " is on NUMA node ", node_id, ", but CPU ",
-                  omp_cpu_ids.front(), " is on NUMA node ", mem_node_id,
-                  ". All CPUs should be on the same NUMA node for optimal "
-                  "performance. Memory will be bound to NUMA node ",
-                  mem_node_id, ".");
-    }
     bitmask* mask = numa_parse_nodestring(std::to_string(mem_node_id).c_str());
     bitmask* src_mask = numa_get_membind();
 

csrc/cumem_allocator.cpp

@@ -3,58 +3,14 @@
 // need to be unsigned long long
 #include <iostream>
 
-#include "cumem_allocator_compat.h"
-
-#ifndef USE_ROCM
-static const char* PYARGS_PARSE = "KKKK";
-#else
-  #include <cstdlib>
-  #include <cerrno>
-  #include <climits>
-
-// Default chunk size 256MB for ROCm. Can be overridden at runtime by the
-// environment variable VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE, specified in megabytes
-// (MB). The env value is parsed with strtoull as an integer number of MB
-// (decimal or 0x hex). The parsed MB value is converted to bytes. If
-// parsing fails, the value is 0, or the multiplication would overflow,
-// the default (256MB) is used.
-static const unsigned long long DEFAULT_MEMCREATE_CHUNK_SIZE =
-    (256ULL * 1024ULL * 1024ULL);
-
-static unsigned long long get_memcreate_chunk_size() {
-  const char* env = getenv("VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE");
-  if (!env) return DEFAULT_MEMCREATE_CHUNK_SIZE;
-  char* endptr = nullptr;
-  errno = 0;
-  unsigned long long val_mb = strtoull(env, &endptr, 0);
-  if (endptr == env || errno != 0) {
-    // parsing failed, fallback to default
-    return DEFAULT_MEMCREATE_CHUNK_SIZE;
-  }
-  if (val_mb == 0) return DEFAULT_MEMCREATE_CHUNK_SIZE;
-
-  const unsigned long long MB = 1024ULL * 1024ULL;
-  // guard against overflow when converting MB -> bytes
-  if (val_mb > (ULLONG_MAX / MB)) {
-    return DEFAULT_MEMCREATE_CHUNK_SIZE;
-  }
-  return val_mb * MB;
-}
-
-static inline unsigned long long my_min(unsigned long long a,
-                                        unsigned long long b) {
-  return a < b ? a : b;
-}
-
-static const char* PYARGS_PARSE = "KKKO";
-#endif
-
 extern "C" {
 
 #define PY_SSIZE_T_CLEAN
 #include <Python.h>
 
 #include <sys/types.h>
+#include <cuda_runtime_api.h>
+#include <cuda.h>
 
 char error_msg[10240];  // 10KB buffer to store error messages
 CUresult no_error = CUresult(0);
@@ -93,12 +49,7 @@ void ensure_context(unsigned long long device) {
 }
 
 void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
-#ifndef USE_ROCM
                     CUmemGenericAllocationHandle* p_memHandle) {
-#else
-                    CUmemGenericAllocationHandle** p_memHandle,
-                    unsigned long long* chunk_sizes, size_t num_chunks) {
-#endif
   ensure_context(device);
   // Define memory allocation properties
   CUmemAllocationProp prop = {};
@@ -107,7 +58,6 @@ void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
   prop.location.id = device;
   prop.allocFlags.compressionType = CU_MEM_ALLOCATION_COMP_NONE;
 
-#ifndef USE_ROCM
   // Allocate memory using cuMemCreate
   CUDA_CHECK(cuMemCreate(p_memHandle, size, &prop, 0));
   if (error_code != 0) {
@@ -117,39 +67,6 @@ void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
   if (error_code != 0) {
     return;
   }
-#else
-  for (auto i = 0; i < num_chunks; ++i) {
-    CUDA_CHECK(cuMemCreate(p_memHandle[i], chunk_sizes[i], &prop, 0));
-    if (error_code != 0) {
-      // Clean up previously created handles
-      for (auto j = 0; j < i; ++j) {
-        cuMemRelease(*(p_memHandle[j]));
-      }
-      return;
-    }
-  }
-  unsigned long long allocated_size = 0;
-  for (auto i = 0; i < num_chunks; ++i) {
-    void* map_addr = (void*)((uintptr_t)d_mem + allocated_size);
-    CUDA_CHECK(cuMemMap(map_addr, chunk_sizes[i], 0, *(p_memHandle[i]), 0));
-    if (error_code != 0) {
-      // unmap previously mapped chunks
-      unsigned long long unmapped_size = 0;
-      for (auto j = 0; j < i; ++j) {
-        void* unmap_addr = (void*)((uintptr_t)d_mem + unmapped_size);
-        cuMemUnmap(unmap_addr, chunk_sizes[j]);
-        unmapped_size += chunk_sizes[j];
-      }
-      // release all created handles
-      for (auto j = 0; j < num_chunks; ++j) {
-        cuMemRelease(*(p_memHandle[j]));
-      }
-      return;
-    }
-    allocated_size += chunk_sizes[i];
-  }
-#endif
-
   CUmemAccessDesc accessDesc = {};
   accessDesc.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
   accessDesc.location.id = device;
@@ -165,16 +82,10 @@ void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
 
 void unmap_and_release(unsigned long long device, ssize_t size,
                        CUdeviceptr d_mem,
-#ifndef USE_ROCM
                        CUmemGenericAllocationHandle* p_memHandle) {
-#else
-                       CUmemGenericAllocationHandle** p_memHandle,
-                       unsigned long long* chunk_sizes, size_t num_chunks) {
-#endif
   // std::cout << "unmap_and_release: device=" << device << ", size=" << size <<
   // ", d_mem=" << d_mem << ", p_memHandle=" << p_memHandle << std::endl;
   ensure_context(device);
-#ifndef USE_ROCM
   CUDA_CHECK(cuMemUnmap(d_mem, size));
   if (error_code != 0) {
     return;
@@ -183,30 +94,6 @@ void unmap_and_release(unsigned long long device, ssize_t size,
   if (error_code != 0) {
     return;
   }
-#else
-  unsigned long long allocated_size = 0;
-  CUresult first_error = no_error;
-
-  for (auto i = 0; i < num_chunks; ++i) {
-    void* map_addr = (void*)((uintptr_t)d_mem + allocated_size);
-    CUresult status = cuMemUnmap(map_addr, chunk_sizes[i]);
-    if (status != no_error && first_error == no_error) {
-      first_error = status;
-    }
-    allocated_size += chunk_sizes[i];
-  }
-
-  for (auto i = 0; i < num_chunks; ++i) {
-    CUresult status = cuMemRelease(*(p_memHandle[i]));
-    if (status != no_error && first_error == no_error) {
-      first_error = status;
-    }
-  }
-
-  if (first_error != no_error) {
-    CUDA_CHECK(first_error);
-  }
-#endif
 }
 
 PyObject* create_tuple_from_c_integers(unsigned long long a,
@@ -233,36 +120,6 @@ PyObject* create_tuple_from_c_integers(unsigned long long a,
   return tuple;  // Return the created tuple
 }
 
-PyObject* create_tuple_from_c_mixed(unsigned long long a, unsigned long long b,
-                                    unsigned long long c,
-                                    CUmemGenericAllocationHandle** vec,
-                                    unsigned long long* chunk_sizes,
-                                    size_t num_chunks) {
-  PyObject* tuple = PyTuple_New(4);
-  if (!tuple) {
-    return NULL;
-  }
-
-  // PyObject* list = PyList_New(vec.size());
-  PyObject* list = PyList_New(num_chunks);
-  for (auto i = 0; i < num_chunks; ++i) {
-    PyObject* addr_size_pair = PyTuple_New(2);
-    PyObject* addr = PyLong_FromUnsignedLongLong((unsigned long long)(vec[i]));
-    PyObject* size =
-        PyLong_FromUnsignedLongLong((unsigned long long)(chunk_sizes[i]));
-    PyTuple_SetItem(addr_size_pair, 0, addr);
-    PyTuple_SetItem(addr_size_pair, 1, size);
-    PyList_SetItem(list, i, addr_size_pair);
-  }
-
-  PyTuple_SetItem(tuple, 0, PyLong_FromUnsignedLongLong(a));
-  PyTuple_SetItem(tuple, 1, PyLong_FromUnsignedLongLong(b));
-  PyTuple_SetItem(tuple, 2, PyLong_FromUnsignedLongLong(c));
-  PyTuple_SetItem(tuple, 3, list);
-
-  return tuple;
-}
-
 // ---------------------------------------------------------------------------
 // Our exported C functions that call Python:
 
@@ -290,55 +147,14 @@ void* my_malloc(ssize_t size, int device, CUstream stream) {
   size_t alignedSize = ((size + granularity - 1) / granularity) * granularity;
 
   CUdeviceptr d_mem;
-#ifndef USE_ROCM
   CUDA_CHECK(cuMemAddressReserve(&d_mem, alignedSize, 0, 0, 0));
   if (error_code != 0) {
     return nullptr;
   }
-#else
-  CUDA_CHECK(cuMemAddressReserve(&d_mem, alignedSize, granularity, 0, 0));
-  if (error_code != 0) {
-    return nullptr;
-  }
-#endif
-
-#ifndef USE_ROCM
   // allocate the CUmemGenericAllocationHandle
   CUmemGenericAllocationHandle* p_memHandle =
       (CUmemGenericAllocationHandle*)malloc(
           sizeof(CUmemGenericAllocationHandle));
-#else
-  // Make sure chunk size is aligned with hardware granularity. The base
-  // chunk size can be configured via environment variable
-  // ``VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE``; otherwise
-  // DEFAULT_MEMCREATE_CHUNK_SIZE is used.
-  size_t base_chunk = (size_t)get_memcreate_chunk_size();
-  size_t aligned_chunk_size =
-      ((base_chunk + granularity - 1) / granularity) * granularity;
-  size_t num_chunks =
-      (alignedSize + aligned_chunk_size - 1) / aligned_chunk_size;
-  CUmemGenericAllocationHandle** p_memHandle =
-      (CUmemGenericAllocationHandle**)malloc(
-          num_chunks * sizeof(CUmemGenericAllocationHandle*));
-  unsigned long long* chunk_sizes =
-      (unsigned long long*)malloc(num_chunks * sizeof(unsigned long long));
-  for (auto i = 0; i < num_chunks; ++i) {
-    p_memHandle[i] = (CUmemGenericAllocationHandle*)malloc(
-        sizeof(CUmemGenericAllocationHandle));
-    if (p_memHandle[i] == nullptr) {
-      std::cerr << "ERROR: malloc failed for p_memHandle[" << i << "].\n";
-      for (auto j = 0; j < i; ++j) {
-        free(p_memHandle[j]);
-      }
-      free(p_memHandle);
-      free(chunk_sizes);
-      return nullptr;
-    }
-    chunk_sizes[i] = (unsigned long long)my_min(
-        (unsigned long long)(alignedSize - i * aligned_chunk_size),
-        (unsigned long long)aligned_chunk_size);
-  }
-#endif
 
   if (!g_python_malloc_callback) {
     std::cerr << "ERROR: g_python_malloc_callback not set.\n";
@@ -348,15 +164,9 @@ void* my_malloc(ssize_t size, int device, CUstream stream) {
   // Acquire GIL (not in stable ABI officially, but often works)
   PyGILState_STATE gstate = PyGILState_Ensure();
 
-#ifndef USE_ROCM
   PyObject* arg_tuple = create_tuple_from_c_integers(
       (unsigned long long)device, (unsigned long long)alignedSize,
       (unsigned long long)d_mem, (unsigned long long)p_memHandle);
-#else
-  PyObject* arg_tuple = create_tuple_from_c_mixed(
-      (unsigned long long)device, (unsigned long long)alignedSize,
-      (unsigned long long)d_mem, p_memHandle, chunk_sizes, num_chunks);
-#endif
 
   // Call g_python_malloc_callback
   PyObject* py_result =
@@ -372,27 +182,7 @@ void* my_malloc(ssize_t size, int device, CUstream stream) {
   PyGILState_Release(gstate);
 
   // do the final mapping
-#ifndef USE_ROCM
   create_and_map(device, alignedSize, d_mem, p_memHandle);
-#else
-  create_and_map(device, alignedSize, d_mem, p_memHandle, chunk_sizes,
-                 num_chunks);
-  free(chunk_sizes);
-#endif
-
-  if (error_code != 0) {
-    // free address and the handle
-    CUDA_CHECK(cuMemAddressFree(d_mem, alignedSize));
-#ifndef USE_ROCM
-    free(p_memHandle);
-#else
-    for (size_t i = 0; i < num_chunks; ++i) {
-      free(p_memHandle[i]);
-    }
-    free(p_memHandle);
-#endif
-    return nullptr;
-  }
 
   return (void*)d_mem;
 }
@@ -416,96 +206,36 @@ void my_free(void* ptr, ssize_t size, int device, CUstream stream) {
 
   if (!py_result || !PyTuple_Check(py_result) || PyTuple_Size(py_result) != 4) {
     PyErr_SetString(PyExc_TypeError, "Expected a tuple of size 4");
-    Py_XDECREF(py_result);
-    Py_XDECREF(py_ptr);
     return;
   }
 
   unsigned long long recv_device, recv_size;
-  unsigned long long recv_d_mem;
-#ifndef USE_ROCM
-  unsigned long long recv_p_memHandle;
-#else
-  PyObject* recv_p_memHandle;
-#endif
+  unsigned long long recv_d_mem, recv_p_memHandle;
   // Unpack the tuple into four C integers
-  if (!PyArg_ParseTuple(py_result, PYARGS_PARSE, &recv_device, &recv_size,
+  if (!PyArg_ParseTuple(py_result, "KKKK", &recv_device, &recv_size,
                         &recv_d_mem, &recv_p_memHandle)) {
     // PyArg_ParseTuple sets an error if it fails
-    Py_XDECREF(py_result);
-    Py_XDECREF(py_ptr);
     return;
   }
 
-  // For ROCm, copy the Python list of (addr,size) pairs into C arrays while
-  // holding the GIL. Then release the GIL and call the unmap/release helper
-  // using the copied arrays. This avoids calling PyList_* APIs without the
-  // GIL (which is undefined behavior and can crash when called from other
-  // threads).
+  PyGILState_Release(gstate);
+
+  // recv_size == size
+  // recv_device == device
+
+  // Free memory
+
   CUdeviceptr d_mem = (CUdeviceptr)recv_d_mem;
-#ifdef USE_ROCM
-  Py_ssize_t num_chunks = PyList_Size(recv_p_memHandle);
-  CUmemGenericAllocationHandle** p_memHandle =
-      (CUmemGenericAllocationHandle**)malloc(
-          num_chunks * sizeof(CUmemGenericAllocationHandle*));
-  if (p_memHandle == nullptr) {
-    Py_DECREF(py_ptr);
-    Py_DECREF(py_result);
-    PyGILState_Release(gstate);
-    std::cerr << "ERROR: malloc failed for p_memHandle in my_free."
-              << std::endl;
-    return;
-  }
-  unsigned long long* chunk_sizes =
-      (unsigned long long*)malloc(num_chunks * sizeof(unsigned long long));
-  if (chunk_sizes == nullptr) {
-    free(p_memHandle);
-    Py_DECREF(py_ptr);
-    Py_DECREF(py_result);
-    PyGILState_Release(gstate);
-    std::cerr << "ERROR: malloc failed for chunk_sizes in my_free."
-              << std::endl;
-    return;
-  }
-  for (Py_ssize_t i = 0; i < num_chunks; ++i) {
-    PyObject* item = PyList_GetItem(recv_p_memHandle, i);
-    PyObject* addr_py = PyTuple_GetItem(item, 0);
-    PyObject* size_py = PyTuple_GetItem(item, 1);
-    p_memHandle[i] =
-        (CUmemGenericAllocationHandle*)PyLong_AsUnsignedLongLong(addr_py);
-    chunk_sizes[i] = (unsigned long long)PyLong_AsUnsignedLongLong(size_py);
-  }
-
-  // Drop temporary Python refs, then release the GIL before calling into
-  // non-Python APIs.
-  Py_DECREF(py_ptr);
-  Py_DECREF(py_result);
-  PyGILState_Release(gstate);
-
-  unmap_and_release(device, size, d_mem, p_memHandle, chunk_sizes, num_chunks);
-#else
-  // Non-ROCm path: simple integer handle already extracted; drop temporary
-  // Python refs while still holding the GIL, then release it.
-  Py_DECREF(py_ptr);
-  Py_DECREF(py_result);
-  PyGILState_Release(gstate);
-
   CUmemGenericAllocationHandle* p_memHandle =
       (CUmemGenericAllocationHandle*)recv_p_memHandle;
   unmap_and_release(device, size, d_mem, p_memHandle);
-#endif
 
   // free address and the handle
   CUDA_CHECK(cuMemAddressFree(d_mem, size));
-#ifndef USE_ROCM
-  free(p_memHandle);
-#else
-  for (auto i = 0; i < num_chunks; ++i) {
-    free(p_memHandle[i]);
+  if (error_code != 0) {
+    return;
   }
   free(p_memHandle);
-  free(chunk_sizes);
-#endif
 }
 
 // ---------------------------------------------------------------------------
@@ -541,87 +271,19 @@ static PyObject* python_unmap_and_release(PyObject* self, PyObject* args) {
   }
 
   unsigned long long recv_device, recv_size;
-  unsigned long long recv_d_mem;
-#ifndef USE_ROCM
-  unsigned long long recv_p_memHandle;
-#else
-  PyObject* recv_p_memHandle;
-#endif
+  unsigned long long recv_d_mem, recv_p_memHandle;
   // Unpack the tuple into four C integers
-  if (!PyArg_ParseTuple(args, PYARGS_PARSE, &recv_device, &recv_size,
-                        &recv_d_mem, &recv_p_memHandle)) {
+  if (!PyArg_ParseTuple(args, "KKKK", &recv_device, &recv_size, &recv_d_mem,
+                        &recv_p_memHandle)) {
     // PyArg_ParseTuple sets an error if it fails
     return nullptr;
   }
 
   CUdeviceptr d_mem_ptr = (CUdeviceptr)recv_d_mem;
-#ifndef USE_ROCM
   CUmemGenericAllocationHandle* p_memHandle =
       (CUmemGenericAllocationHandle*)recv_p_memHandle;
 
   unmap_and_release(recv_device, recv_size, d_mem_ptr, p_memHandle);
-#else
-  if (!PyList_Check(recv_p_memHandle)) {
-    PyErr_SetString(PyExc_TypeError,
-                    "Expected a list for the 4th argument on ROCm");
-    return nullptr;
-  }
-  Py_ssize_t num_chunks = PyList_Size(recv_p_memHandle);
-  if (num_chunks < 0) {
-    return nullptr;  // PyList_Size sets an exception on error.
-  }
-  CUmemGenericAllocationHandle** p_memHandle =
-      (CUmemGenericAllocationHandle**)malloc(
-          num_chunks * sizeof(CUmemGenericAllocationHandle*));
-  if (p_memHandle == nullptr) {
-    PyErr_SetString(PyExc_MemoryError, "malloc failed for p_memHandle");
-    return nullptr;
-  }
-  unsigned long long* chunk_sizes =
-      (unsigned long long*)malloc(num_chunks * sizeof(unsigned long long));
-  if (chunk_sizes == nullptr) {
-    free(p_memHandle);
-    PyErr_SetString(PyExc_MemoryError, "malloc failed for chunk_sizes");
-    return nullptr;
-  }
-  for (Py_ssize_t i = 0; i < num_chunks; ++i) {
-    PyObject* item = PyList_GetItem(recv_p_memHandle, i);
-    if (item == nullptr || !PyTuple_Check(item) || PyTuple_Size(item) != 2) {
-      free(p_memHandle);
-      free(chunk_sizes);
-      PyErr_SetString(
-          PyExc_TypeError,
-          "List items must be tuples of size 2 (handle_addr, size)");
-      return nullptr;
-    }
-    PyObject* addr_py = PyTuple_GetItem(item, 0);
-    PyObject* size_py = PyTuple_GetItem(item, 1);
-    if (addr_py == nullptr || size_py == nullptr) {
-      free(p_memHandle);
-      free(chunk_sizes);
-      return nullptr;  // PyTuple_GetItem sets an exception
-    }
-    p_memHandle[i] =
-        (CUmemGenericAllocationHandle*)PyLong_AsUnsignedLongLong(addr_py);
-    if (PyErr_Occurred()) {
-      free(p_memHandle);
-      free(chunk_sizes);
-      return nullptr;
-    }
-    chunk_sizes[i] = (unsigned long long)PyLong_AsUnsignedLongLong(size_py);
-    if (PyErr_Occurred()) {
-      free(p_memHandle);
-      free(chunk_sizes);
-      return nullptr;
-    }
-  }
-
-  unmap_and_release(recv_device, recv_size, d_mem_ptr, p_memHandle, chunk_sizes,
-                    num_chunks);
-
-  free(p_memHandle);
-  free(chunk_sizes);
-#endif
 
   if (error_code != 0) {
     error_code = no_error;
@@ -639,56 +301,19 @@ static PyObject* python_create_and_map(PyObject* self, PyObject* args) {
   }
 
   unsigned long long recv_device, recv_size;
-  unsigned long long recv_d_mem;
-#ifndef USE_ROCM
-  unsigned long long recv_p_memHandle;
-#else
-  PyObject* recv_p_memHandle;
-#endif
+  unsigned long long recv_d_mem, recv_p_memHandle;
   // Unpack the tuple into four C integers
-  if (!PyArg_ParseTuple(args, PYARGS_PARSE, &recv_device, &recv_size,
-                        &recv_d_mem, &recv_p_memHandle)) {
+  if (!PyArg_ParseTuple(args, "KKKK", &recv_device, &recv_size, &recv_d_mem,
+                        &recv_p_memHandle)) {
     // PyArg_ParseTuple sets an error if it fails
     return nullptr;
   }
 
   CUdeviceptr d_mem_ptr = (CUdeviceptr)recv_d_mem;
-#ifndef USE_ROCM
   CUmemGenericAllocationHandle* p_memHandle =
       (CUmemGenericAllocationHandle*)recv_p_memHandle;
 
   create_and_map(recv_device, recv_size, d_mem_ptr, p_memHandle);
-#else
-  Py_ssize_t num_chunks = PyList_Size(recv_p_memHandle);
-  CUmemGenericAllocationHandle** p_memHandle =
-      (CUmemGenericAllocationHandle**)malloc(
-          num_chunks * sizeof(CUmemGenericAllocationHandle*));
-  if (p_memHandle == nullptr) {
-    PyErr_SetString(PyExc_MemoryError, "malloc failed for p_memHandle");
-    return nullptr;
-  }
-  unsigned long long* chunk_sizes =
-      (unsigned long long*)malloc(num_chunks * sizeof(unsigned long long));
-  if (chunk_sizes == nullptr) {
-    free(p_memHandle);
-    PyErr_SetString(PyExc_MemoryError, "malloc failed for chunk_sizes");
-    return nullptr;
-  }
-  for (auto i = 0; i < num_chunks; ++i) {
-    PyObject* item = PyList_GetItem(recv_p_memHandle, i);
-    PyObject* addr_py = PyTuple_GetItem(item, 0);
-    PyObject* size_py = PyTuple_GetItem(item, 1);
-    p_memHandle[i] =
-        (CUmemGenericAllocationHandle*)PyLong_AsUnsignedLongLong(addr_py);
-    chunk_sizes[i] = PyLong_AsUnsignedLongLong(size_py);
-  }
-
-  create_and_map(recv_device, recv_size, d_mem_ptr, p_memHandle, chunk_sizes,
-                 num_chunks);
-
-  free(p_memHandle);
-  free(chunk_sizes);
-#endif
 
   if (error_code != 0) {
     error_code = no_error;

csrc/cumem_allocator_compat.h

@@ -1,109 +0,0 @@
-#pragma once
-
-#ifdef USE_ROCM
-////////////////////////////////////////
-// For compatibility with CUDA and ROCm
-////////////////////////////////////////
-  #include <hip/hip_runtime_api.h>
-
-extern "C" {
-  #ifndef CUDA_SUCCESS
-    #define CUDA_SUCCESS hipSuccess
-  #endif  // CUDA_SUCCESS
-
-// https://rocm.docs.amd.com/projects/HIPIFY/en/latest/tables/CUDA_Driver_API_functions_supported_by_HIP.html
-typedef unsigned long long CUdevice;
-typedef hipDeviceptr_t CUdeviceptr;
-typedef hipError_t CUresult;
-typedef hipCtx_t CUcontext;
-typedef hipStream_t CUstream;
-typedef hipMemGenericAllocationHandle_t CUmemGenericAllocationHandle;
-typedef hipMemAllocationGranularity_flags CUmemAllocationGranularity_flags;
-typedef hipMemAllocationProp CUmemAllocationProp;
-typedef hipMemAccessDesc CUmemAccessDesc;
-
-  #define CU_MEM_ALLOCATION_TYPE_PINNED hipMemAllocationTypePinned
-  #define CU_MEM_LOCATION_TYPE_DEVICE hipMemLocationTypeDevice
-  #define CU_MEM_ACCESS_FLAGS_PROT_READWRITE hipMemAccessFlagsProtReadWrite
-  #define CU_MEM_ALLOC_GRANULARITY_MINIMUM hipMemAllocationGranularityMinimum
-
-  // https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__TYPES.html
-  #define CU_MEM_ALLOCATION_COMP_NONE 0x0
-
-// Error Handling
-// https://docs.nvidia.com/cuda/archive/11.4.4/cuda-driver-api/group__CUDA__ERROR.html
-CUresult cuGetErrorString(CUresult hipError, const char** pStr) {
-  *pStr = hipGetErrorString(hipError);
-  return CUDA_SUCCESS;
-}
-
-// Context Management
-// https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__CTX.html
-CUresult cuCtxGetCurrent(CUcontext* ctx) {
-  // This API is deprecated on the AMD platform, only for equivalent cuCtx
-  // driver API on the NVIDIA platform.
-  return hipCtxGetCurrent(ctx);
-}
-
-CUresult cuCtxSetCurrent(CUcontext ctx) {
-  // This API is deprecated on the AMD platform, only for equivalent cuCtx
-  // driver API on the NVIDIA platform.
-  return hipCtxSetCurrent(ctx);
-}
-
-// Primary Context Management
-// https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__PRIMARY__CTX.html
-CUresult cuDevicePrimaryCtxRetain(CUcontext* ctx, CUdevice dev) {
-  return hipDevicePrimaryCtxRetain(ctx, dev);
-}
-
-// Virtual Memory Management
-// https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__VA.html
-CUresult cuMemAddressFree(CUdeviceptr ptr, size_t size) {
-  return hipMemAddressFree(ptr, size);
-}
-
-CUresult cuMemAddressReserve(CUdeviceptr* ptr, size_t size, size_t alignment,
-                             CUdeviceptr addr, unsigned long long flags) {
-  return hipMemAddressReserve(ptr, size, alignment, addr, flags);
-}
-
-CUresult cuMemCreate(CUmemGenericAllocationHandle* handle, size_t size,
-                     const CUmemAllocationProp* prop,
-                     unsigned long long flags) {
-  return hipMemCreate(handle, size, prop, flags);
-}
-
-CUresult cuMemGetAllocationGranularity(
-    size_t* granularity, const CUmemAllocationProp* prop,
-    CUmemAllocationGranularity_flags option) {
-  return hipMemGetAllocationGranularity(granularity, prop, option);
-}
-
-CUresult cuMemMap(CUdeviceptr dptr, size_t size, size_t offset,
-                  CUmemGenericAllocationHandle handle,
-                  unsigned long long flags) {
-  return hipMemMap(dptr, size, offset, handle, flags);
-}
-
-CUresult cuMemRelease(CUmemGenericAllocationHandle handle) {
-  return hipMemRelease(handle);
-}
-
-CUresult cuMemSetAccess(CUdeviceptr ptr, size_t size,
-                        const CUmemAccessDesc* desc, size_t count) {
-  return hipMemSetAccess(ptr, size, desc, count);
-}
-
-CUresult cuMemUnmap(CUdeviceptr ptr, size_t size) {
-  return hipMemUnmap(ptr, size);
-}
-}  // extern "C"
-
-#else
-////////////////////////////////////////
-// Import CUDA headers for NVIDIA GPUs
-////////////////////////////////////////
-  #include <cuda_runtime_api.h>
-  #include <cuda.h>
-#endif

csrc/dispatch_utils.h

@@ -88,32 +88,3 @@
 #define VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(TYPE, NAME, ...) \
   AT_DISPATCH_SWITCH(                                              \
       TYPE, NAME, VLLM_DISPATCH_CASE_INTEGRAL_AND_UNSIGNED_TYPES(__VA_ARGS__))
-
-#define VLLM_DISPATCH_VEC_SIZE(VEC_SIZE, ...) \
-  switch (VEC_SIZE) {                         \
-    case 16: {                                \
-      constexpr int vec_size = 16;            \
-      __VA_ARGS__();                          \
-      break;                                  \
-    }                                         \
-    case 8: {                                 \
-      constexpr int vec_size = 8;             \
-      __VA_ARGS__();                          \
-      break;                                  \
-    }                                         \
-    case 4: {                                 \
-      constexpr int vec_size = 4;             \
-      __VA_ARGS__();                          \
-      break;                                  \
-    }                                         \
-    case 2: {                                 \
-      constexpr int vec_size = 2;             \
-      __VA_ARGS__();                          \
-      break;                                  \
-    }                                         \
-    default: {                                \
-      constexpr int vec_size = 1;             \
-      __VA_ARGS__();                          \
-      break;                                  \
-    }                                         \
-  }

csrc/fused_qknorm_rope_kernel.cu

@@ -1,428 +0,0 @@
-/*
- * Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <cmath>
-#include <cuda_runtime.h>
-#include <type_traits>
-
-#include <torch/cuda.h>
-#include <c10/cuda/CUDAGuard.h>
-
-#include "cuda_compat.h"
-#include "dispatch_utils.h"
-#include "type_convert.cuh"
-
-#define CHECK_TYPE(x, st)                                              \
-  TORCH_CHECK(x.scalar_type() == st, #x " dtype is ", x.scalar_type(), \
-              ", while ", st, " is expected")
-#define CHECK_TH_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor")
-#define CHECK_CONTIGUOUS(x) \
-  TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
-#define CHECK_INPUT(x) \
-  CHECK_TH_CUDA(x);    \
-  CHECK_CONTIGUOUS(x)
-
-#ifdef USE_ROCM
-  #define FINAL_MASK 0xffffffffffffffffULL
-
-  #if defined(HIP_VERSION) && HIP_VERSION < 70000000
-// On ROCm versions before 7.0, __syncwarp isn't defined. The below
-// implementation is copy/pasted from the implementation in ROCm 7.0
-__device__ inline void __syncwarp() {
-  __builtin_amdgcn_fence(__ATOMIC_RELEASE, "wavefront");
-  __builtin_amdgcn_wave_barrier();
-  __builtin_amdgcn_fence(__ATOMIC_ACQUIRE, "wavefront");
-}
-  #endif
-#else
-  #define FINAL_MASK 0xffffffff
-#endif
-
-namespace tensorrt_llm::common {
-template <typename T, int num>
-struct packed_as;
-// Specialization for packed_as used in this kernel.
-template <>
-struct packed_as<uint, 1> {
-  using type = uint;
-};
-
-template <>
-struct packed_as<uint, 2> {
-  using type = uint2;
-};
-
-template <>
-struct packed_as<uint, 4> {
-  using type = uint4;
-};
-
-template <typename T>
-__inline__ __device__ T warpReduceSum(T val) {
-#pragma unroll
-  for (int mask = 16; mask > 0; mask >>= 1)
-    val += __shfl_xor_sync(FINAL_MASK, val, mask, 32);
-  return val;
-}
-
-template <typename T>
-inline __device__ __host__ T divUp(T m, T n) {
-  return (m + n - 1) / n;
-}
-
-}  // namespace tensorrt_llm::common
-
-namespace tensorrt_llm::kernels {
-// NOTE(zhuhaoran): This kernel is adapted from TensorRT-LLM implementation,
-// with added support for passing the cos_sin_cache as an input.
-// https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/kernels/fusedQKNormRopeKernel.cu
-
-// Perform per-head QK Norm and RoPE in a single kernel.
-// scalar_t_in: data type of QKV and RMSNorm weights
-// scalar_t_cache: data type of cos/sin cache
-// head_dim: the dimension of each head
-// interleave: interleave=!is_neox.
-template <typename scalar_t_in, typename scalar_t_cache, int head_dim,
-          bool interleave>
-__global__ void fusedQKNormRopeKernel(
-    void* qkv_void,                  // Combined QKV tensor
-    int const num_heads_q,           // Number of query heads
-    int const num_heads_k,           // Number of key heads
-    int const num_heads_v,           // Number of value heads
-    float const eps,                 // Epsilon for RMS normalization
-    void const* q_weight_void,       // RMSNorm weights for query
-    void const* k_weight_void,       // RMSNorm weights for key
-    void const* cos_sin_cache_void,  // Pre-computed cos/sin cache
-    int64_t const* position_ids,     // Position IDs for RoPE
-    int const num_tokens             // Number of tokens
-) {
-#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
-  if constexpr ((std::is_same_v<scalar_t_in, c10::BFloat16>) ||
-                std::is_same_v<scalar_t_cache, c10::BFloat16>) {
-    return;
-  } else {
-#endif
-
-    using Converter = vllm::_typeConvert<scalar_t_in>;
-    static_assert(Converter::exists,
-                  "Input QKV data type is not supported for this CUDA "
-                  "architecture or toolkit version.");
-    using T_in = typename Converter::hip_type;
-    using T2_in = typename Converter::packed_hip_type;
-
-    using CacheConverter = vllm::_typeConvert<scalar_t_cache>;
-    static_assert(CacheConverter::exists,
-                  "Cache data type is not supported for this CUDA architecture "
-                  "or toolkit version.");
-    using T_cache = typename CacheConverter::hip_type;
-
-    T_in* qkv = reinterpret_cast<T_in*>(qkv_void);
-    T_in const* q_weight = reinterpret_cast<T_in const*>(q_weight_void);
-    T_in const* k_weight = reinterpret_cast<T_in const*>(k_weight_void);
-    T_cache const* cos_sin_cache =
-        reinterpret_cast<T_cache const*>(cos_sin_cache_void);
-
-    int const warpsPerBlock = blockDim.x / 32;
-    int const warpId = threadIdx.x / 32;
-    int const laneId = threadIdx.x % 32;
-
-    // Calculate global warp index to determine which head/token this warp
-    // processes
-    int const globalWarpIdx = blockIdx.x * warpsPerBlock + warpId;
-
-    // Total number of attention heads (Q and K)
-    int const total_qk_heads = num_heads_q + num_heads_k;
-
-    // Determine which token and head type (Q or K) this warp processes
-    int const tokenIdx = globalWarpIdx / total_qk_heads;
-    int const localHeadIdx = globalWarpIdx % total_qk_heads;
-
-    // Skip if this warp is assigned beyond the number of tokens
-    if (tokenIdx >= num_tokens) return;
-
-    bool const isQ = localHeadIdx < num_heads_q;
-    int const headIdx = isQ ? localHeadIdx : localHeadIdx - num_heads_q;
-
-    int const num_heads = num_heads_q + num_heads_k + num_heads_v;
-
-    static_assert(head_dim % (32 * 2) == 0,
-                  "head_dim must be divisible by 64 (each warp processes one "
-                  "head, and each thread gets even number of "
-                  "elements)");
-    constexpr int numElemsPerThread = head_dim / 32;
-    float elements[numElemsPerThread];
-    constexpr int elemSizeBytes = numElemsPerThread * sizeof(__nv_bfloat16);
-    static_assert(elemSizeBytes % 4 == 0,
-                  "numSizeBytes must be a multiple of 4");
-    constexpr int vecSize =
-        elemSizeBytes /
-        4;  // Use packed_as<uint, vecSize> to perform loading/saving.
-    using vec_T = typename tensorrt_llm::common::packed_as<uint, vecSize>::type;
-
-    int offsetWarp;  // Offset for the warp
-    if (isQ) {
-      // Q segment: token offset + head offset within Q segment
-      offsetWarp = tokenIdx * num_heads * head_dim + headIdx * head_dim;
-    } else {
-      // K segment: token offset + entire Q segment + head offset within K
-      // segment
-      offsetWarp = tokenIdx * num_heads * head_dim + num_heads_q * head_dim +
-                   headIdx * head_dim;
-    }
-    int offsetThread = offsetWarp + laneId * numElemsPerThread;
-
-    // Sum of squares for RMSNorm
-    float sumOfSquares = 0.0f;
-
-    // Load.
-    {
-      vec_T vec = *reinterpret_cast<vec_T const*>(&qkv[offsetThread]);
-      constexpr int num_packed_elems = elemSizeBytes / sizeof(T2_in);
-#pragma unroll
-      for (int i = 0; i < num_packed_elems; i++) {
-        // Interpret the generic vector chunk as the specific packed type
-        T2_in packed_val = *(reinterpret_cast<T2_in*>(&vec) + i);
-        // Convert to float2 for computation
-        float2 vals = Converter::convert(packed_val);
-        sumOfSquares += vals.x * vals.x;
-        sumOfSquares += vals.y * vals.y;
-
-        elements[2 * i] = vals.x;
-        elements[2 * i + 1] = vals.y;
-      }
-    }
-
-    // Reduce sum across warp using the utility function
-    sumOfSquares = tensorrt_llm::common::warpReduceSum(sumOfSquares);
-
-    // Compute RMS normalization factor
-    float rms_rcp = rsqrtf(sumOfSquares / static_cast<float>(head_dim) + eps);
-
-    // Normalize elements
-#pragma unroll
-    for (int i = 0; i < numElemsPerThread; i++) {
-      int dim = laneId * numElemsPerThread + i;
-      float weight = isQ ? Converter::convert(q_weight[dim])
-                         : Converter::convert(k_weight[dim]);
-      elements[i] *= rms_rcp * weight;
-    }
-
-    // Apply RoPE to normalized elements
-    float elements2[numElemsPerThread];  // Additional buffer required for RoPE.
-
-    int64_t pos_id = position_ids[tokenIdx];
-
-    // Calculate cache pointer for this position - similar to
-    // pos_encoding_kernels.cu
-    T_cache const* cache_ptr = cos_sin_cache + pos_id * head_dim;
-    int const embed_dim = head_dim / 2;
-    T_cache const* cos_ptr = cache_ptr;
-    T_cache const* sin_ptr = cache_ptr + embed_dim;
-
-    if constexpr (interleave) {
-      // Perform interleaving. Use pre-computed cos/sin values.
-#pragma unroll
-      for (int i = 0; i < numElemsPerThread / 2; ++i) {
-        int const idx0 = 2 * i;
-        int const idx1 = 2 * i + 1;
-
-        float const val0 = elements[idx0];
-        float const val1 = elements[idx1];
-
-        int const dim_idx = laneId * numElemsPerThread + idx0;
-        int const half_dim = dim_idx / 2;
-        float const cos_val =
-            CacheConverter::convert(VLLM_LDG(cos_ptr + half_dim));
-        float const sin_val =
-            CacheConverter::convert(VLLM_LDG(sin_ptr + half_dim));
-
-        elements[idx0] = val0 * cos_val - val1 * sin_val;
-        elements[idx1] = val0 * sin_val + val1 * cos_val;
-      }
-    } else {
-      // Before data exchange with in warp, we need to sync.
-      __syncwarp();
-      // Get the data from the other half of the warp. Use pre-computed cos/sin
-      // values.
-#pragma unroll
-      for (int i = 0; i < numElemsPerThread; i++) {
-        elements2[i] = __shfl_xor_sync(FINAL_MASK, elements[i], 16);
-        if (laneId < 16) {
-          elements2[i] = -elements2[i];
-        }
-
-        int dim_idx = laneId * numElemsPerThread + i;
-        dim_idx = (dim_idx * 2) % head_dim;
-        int half_dim = dim_idx / 2;
-        // Use pre-computed cos/sin from cache
-        float cos_val = CacheConverter::convert(VLLM_LDG(cos_ptr + half_dim));
-        float sin_val = CacheConverter::convert(VLLM_LDG(sin_ptr + half_dim));
-
-        elements[i] = elements[i] * cos_val + elements2[i] * sin_val;
-      }
-      // __shfl_xor_sync does not provide memfence. Need to sync again.
-      __syncwarp();
-    }
-
-    // Store.
-    {
-      vec_T vec;
-      constexpr int num_packed_elems = elemSizeBytes / sizeof(T2_in);
-#pragma unroll
-      for (int i = 0; i < num_packed_elems; i++) {
-        // Convert from float2 back to the specific packed type
-        T2_in packed_val = Converter::convert(
-            make_float2(elements[2 * i], elements[2 * i + 1]));
-        // Place it into the generic vector
-        *(reinterpret_cast<T2_in*>(&vec) + i) = packed_val;
-      }
-      *reinterpret_cast<vec_T*>(&qkv[offsetThread]) = vec;
-    }
-
-#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
-  }
-#endif
-}
-
-// Borrowed from
-// https://github.com/flashinfer-ai/flashinfer/blob/8125d079a43e9a0ba463a4ed1b639cefd084cec9/include/flashinfer/pos_enc.cuh#L568
-#define DISPATCH_INTERLEAVE(interleave, INTERLEAVE, ...) \
-  if (interleave) {                                      \
-    const bool INTERLEAVE = true;                        \
-    __VA_ARGS__                                          \
-  } else {                                               \
-    const bool INTERLEAVE = false;                       \
-    __VA_ARGS__                                          \
-  }
-
-template <typename scalar_t_in, typename scalar_t_cache>
-void launchFusedQKNormRope(void* qkv, int const num_tokens,
-                           int const num_heads_q, int const num_heads_k,
-                           int const num_heads_v, int const head_dim,
-                           float const eps, void const* q_weight,
-                           void const* k_weight, void const* cos_sin_cache,
-                           bool const interleave, int64_t const* position_ids,
-                           cudaStream_t stream) {
-  constexpr int blockSize = 256;
-
-  int const warpsPerBlock = blockSize / 32;
-  int const totalQKHeads = num_heads_q + num_heads_k;
-  int const totalWarps = num_tokens * totalQKHeads;
-
-  int const gridSize = common::divUp(totalWarps, warpsPerBlock);
-  dim3 gridDim(gridSize);
-  dim3 blockDim(blockSize);
-
-  switch (head_dim) {
-    case 64:
-      DISPATCH_INTERLEAVE(interleave, INTERLEAVE, {
-        fusedQKNormRopeKernel<scalar_t_in, scalar_t_cache, 64, INTERLEAVE>
-            <<<gridDim, blockDim, 0, stream>>>(
-                qkv, num_heads_q, num_heads_k, num_heads_v, eps, q_weight,
-                k_weight, cos_sin_cache, position_ids, num_tokens);
-      });
-      break;
-    case 128:
-      DISPATCH_INTERLEAVE(interleave, INTERLEAVE, {
-        fusedQKNormRopeKernel<scalar_t_in, scalar_t_cache, 128, INTERLEAVE>
-            <<<gridDim, blockDim, 0, stream>>>(
-                qkv, num_heads_q, num_heads_k, num_heads_v, eps, q_weight,
-                k_weight, cos_sin_cache, position_ids, num_tokens);
-      });
-      break;
-    case 256:
-      DISPATCH_INTERLEAVE(interleave, INTERLEAVE, {
-        fusedQKNormRopeKernel<scalar_t_in, scalar_t_cache, 256, INTERLEAVE>
-            <<<gridDim, blockDim, 0, stream>>>(
-                qkv, num_heads_q, num_heads_k, num_heads_v, eps, q_weight,
-                k_weight, cos_sin_cache, position_ids, num_tokens);
-      });
-      break;
-    default:
-      TORCH_CHECK(false,
-                  "Unsupported head dimension for fusedQKNormRope: ", head_dim);
-  }
-}
-}  // namespace tensorrt_llm::kernels
-
-void fused_qk_norm_rope(
-    torch::Tensor& qkv,       // Combined QKV tensor [num_tokens,
-                              // (num_heads_q+num_heads_k+num_heads_v)*head_dim]
-    int64_t num_heads_q,      // Number of query heads
-    int64_t num_heads_k,      // Number of key heads
-    int64_t num_heads_v,      // Number of value heads
-    int64_t head_dim,         // Dimension per head
-    double eps,               // Epsilon for RMS normalization
-    torch::Tensor& q_weight,  // RMSNorm weights for query [head_dim]
-    torch::Tensor& k_weight,  // RMSNorm weights for key [head_dim]
-    torch::Tensor& cos_sin_cache,  // Cos/sin cache [max_position, head_dim]
-    bool is_neox,                  // Whether RoPE is applied in Neox style
-    torch::Tensor& position_ids    // Position IDs for RoPE [num_tokens]
-) {
-  // Input validation
-  CHECK_INPUT(qkv);
-  CHECK_INPUT(position_ids);
-  CHECK_INPUT(q_weight);
-  CHECK_INPUT(k_weight);
-  CHECK_INPUT(cos_sin_cache);
-  CHECK_TYPE(position_ids, torch::kInt64);
-
-  TORCH_CHECK(qkv.dim() == 2,
-              "QKV tensor must be 2D: [num_tokens, "
-              "(num_heads_q+num_heads_k+num_heads_v)*head_dim]");
-  TORCH_CHECK(position_ids.dim() == 1, "Position IDs must be 1D: [num_tokens]");
-  TORCH_CHECK(q_weight.dim() == 1, "Query weights must be 1D: [head_dim]");
-  TORCH_CHECK(k_weight.dim() == 1, "Key weights must be 1D: [head_dim]");
-  TORCH_CHECK(cos_sin_cache.dim() == 2,
-              "Cos/sin cache must be 2D: [max_position, head_dim]");
-  TORCH_CHECK(q_weight.size(0) == head_dim,
-              "Query weights size must match head dimension");
-  TORCH_CHECK(k_weight.size(0) == head_dim,
-              "Key weights size must match head dimension");
-  TORCH_CHECK(cos_sin_cache.size(1) == head_dim,
-              "Cos/sin cache dimension must match head_dim");
-  TORCH_CHECK(qkv.scalar_type() == q_weight.scalar_type() &&
-                  qkv.scalar_type() == k_weight.scalar_type(),
-              "qkv, q_weight and k_weight must have the same dtype");
-
-  int64_t num_tokens = qkv.size(0);
-  TORCH_CHECK(position_ids.size(0) == num_tokens,
-              "Number of tokens in position_ids must match QKV");
-
-  int64_t total_heads = num_heads_q + num_heads_k + num_heads_v;
-  TORCH_CHECK(
-      qkv.size(1) == total_heads * head_dim,
-      "QKV tensor size must match total number of heads and head dimension");
-
-  auto stream = at::cuda::getCurrentCUDAStream(qkv.get_device());
-
-  VLLM_DISPATCH_HALF_TYPES(qkv.scalar_type(), "fused_qk_norm_rope_kernel", [&] {
-    using qkv_scalar_t = scalar_t;
-    VLLM_DISPATCH_FLOATING_TYPES(
-        cos_sin_cache.scalar_type(), "fused_qk_norm_rope_kernel", [&] {
-          using cache_scalar_t = scalar_t;
-          tensorrt_llm::kernels::launchFusedQKNormRope<qkv_scalar_t,
-                                                       cache_scalar_t>(
-              qkv.data_ptr(), static_cast<int>(num_tokens),
-              static_cast<int>(num_heads_q), static_cast<int>(num_heads_k),
-              static_cast<int>(num_heads_v), static_cast<int>(head_dim),
-              static_cast<float>(eps), q_weight.data_ptr(), k_weight.data_ptr(),
-              cos_sin_cache.data_ptr(), !is_neox,
-              reinterpret_cast<int64_t const*>(position_ids.data_ptr()),
-              stream);
-        });
-  });
-}

csrc/layernorm_kernels.cu

@@ -10,7 +10,7 @@
 namespace vllm {
 
 // TODO(woosuk): Further optimize this kernel.
-template <typename scalar_t, int VEC_SIZE>
+template <typename scalar_t>
 __global__ void rms_norm_kernel(
     scalar_t* __restrict__ out,          // [..., hidden_size]
     const scalar_t* __restrict__ input,  // [..., hidden_size]
@@ -21,6 +21,7 @@ __global__ void rms_norm_kernel(
   float variance = 0.0f;
   const scalar_t* input_row = input + blockIdx.x * input_stride;
 
+  constexpr int VEC_SIZE = 8;
   auto vec_op = [&variance](const vec_n_t<scalar_t, VEC_SIZE>& vec) {
 #pragma unroll
     for (int i = 0; i < VEC_SIZE; ++i) {
@@ -44,20 +45,10 @@ __global__ void rms_norm_kernel(
   }
   __syncthreads();
 
-  scalar_t* out_row = out + blockIdx.x * hidden_size;
-  auto* v_in = reinterpret_cast<const vec_n_t<scalar_t, VEC_SIZE>*>(input_row);
-  auto* v_w = reinterpret_cast<const vec_n_t<scalar_t, VEC_SIZE>*>(weight);
-  auto* v_out = reinterpret_cast<vec_n_t<scalar_t, VEC_SIZE>*>(out_row);
-  for (int i = threadIdx.x; i < hidden_size / VEC_SIZE; i += blockDim.x) {
-    vec_n_t<scalar_t, VEC_SIZE> dst;
-    vec_n_t<scalar_t, VEC_SIZE> src1 = v_in[i];
-    vec_n_t<scalar_t, VEC_SIZE> src2 = v_w[i];
-#pragma unroll
-    for (int j = 0; j < VEC_SIZE; j++) {
-      float x = static_cast<float>(src1.val[j]);
-      dst.val[j] = ((scalar_t)(x * s_variance)) * src2.val[j];
-    }
-    v_out[i] = dst;
+  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
+    float x = (float)input[blockIdx.x * input_stride + idx];
+    out[blockIdx.x * hidden_size + idx] =
+        ((scalar_t)(x * s_variance)) * weight[idx];
   }
 }
 
@@ -177,24 +168,16 @@ void rms_norm(torch::Tensor& out,     // [..., hidden_size]
   int num_tokens = input_view.numel() / hidden_size;
   int64_t input_stride = input_view.stride(-2);
 
-  // For large num_tokens, use smaller blocks to increase SM concurrency.
-  const int max_block_size = (num_tokens < 256) ? 1024 : 256;
   dim3 grid(num_tokens);
+  dim3 block(std::min(hidden_size, 1024));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input_view));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   VLLM_DISPATCH_FLOATING_TYPES(
       input_view.scalar_type(), "rms_norm_kernel", [&] {
-        const int calculated_vec_size =
-            std::gcd(16 / sizeof(scalar_t), hidden_size);
-        const int block_size =
-            std::min(hidden_size / calculated_vec_size, max_block_size);
-        dim3 block(block_size);
-        VLLM_DISPATCH_VEC_SIZE(calculated_vec_size, [&] {
-          vllm::rms_norm_kernel<scalar_t, vec_size><<<grid, block, 0, stream>>>(
-              out.data_ptr<scalar_t>(), input_view.data_ptr<scalar_t>(),
-              input_stride, weight.data_ptr<scalar_t>(), epsilon, num_tokens,
-              hidden_size);
-        });
+        vllm::rms_norm_kernel<scalar_t><<<grid, block, 0, stream>>>(
+            out.data_ptr<scalar_t>(), input_view.data_ptr<scalar_t>(),
+            input_stride, weight.data_ptr<scalar_t>(), epsilon, num_tokens,
+            hidden_size);
       });
 }
 

csrc/layernorm_quant_kernels.cu

@@ -18,7 +18,7 @@
 namespace vllm {
 
 // TODO(woosuk): Further optimize this kernel.
-template <typename scalar_t, typename fp8_type, int VEC_SIZE>
+template <typename scalar_t, typename fp8_type>
 __global__ void rms_norm_static_fp8_quant_kernel(
     fp8_type* __restrict__ out,          // [..., hidden_size]
     const scalar_t* __restrict__ input,  // [..., hidden_size]
@@ -31,6 +31,7 @@ __global__ void rms_norm_static_fp8_quant_kernel(
 
   const scalar_t* input_row = input + blockIdx.x * input_stride;
 
+  constexpr int VEC_SIZE = 8;
   auto vec_op = [&variance](const vec_n_t<scalar_t, VEC_SIZE>& vec) {
 #pragma unroll
     for (int i = 0; i < VEC_SIZE; ++i) {
@@ -57,18 +58,11 @@ __global__ void rms_norm_static_fp8_quant_kernel(
   // invert scale to avoid division
   float const scale_inv = 1.0f / *scale;
 
-  auto* v_in = reinterpret_cast<const vec_n_t<scalar_t, VEC_SIZE>*>(input_row);
-  auto* v_w = reinterpret_cast<const vec_n_t<scalar_t, VEC_SIZE>*>(weight);
-  for (int idx = threadIdx.x; idx < hidden_size / VEC_SIZE; idx += blockDim.x) {
-    vec_n_t<scalar_t, VEC_SIZE> src1 = v_in[idx];
-    vec_n_t<scalar_t, VEC_SIZE> src2 = v_w[idx];
-#pragma unroll
-    for (int j = 0; j < VEC_SIZE; j++) {
-      float x = static_cast<float>(src1.val[j]);
-      float const out_norm = ((scalar_t)(x * s_variance)) * src2.val[j];
-      out[blockIdx.x * hidden_size + idx * VEC_SIZE + j] =
-          scaled_fp8_conversion<true, fp8_type>(out_norm, scale_inv);
-    }
+  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
+    float x = (float)input[blockIdx.x * input_stride + idx];
+    float const out_norm = ((scalar_t)(x * s_variance)) * weight[idx];
+    out[blockIdx.x * hidden_size + idx] =
+        scaled_fp8_conversion<true, fp8_type>(out_norm, scale_inv);
   }
 }
 
@@ -194,29 +188,20 @@ void rms_norm_static_fp8_quant(torch::Tensor& out,     // [..., hidden_size]
   int input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
 
-  // For large num_tokens, use smaller blocks to increase SM concurrency.
-  const int max_block_size = (num_tokens < 256) ? 1024 : 256;
   dim3 grid(num_tokens);
+  dim3 block(std::min(hidden_size, 1024));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   VLLM_DISPATCH_FLOATING_TYPES(
       input.scalar_type(), "rms_norm_kernel_scalar_type", [&] {
         VLLM_DISPATCH_FP8_TYPES(
             out.scalar_type(), "rms_norm_kernel_fp8_type", [&] {
-              const int calculated_vec_size =
-                  std::gcd(16 / sizeof(scalar_t), hidden_size);
-              const int block_size =
-                  std::min(hidden_size / calculated_vec_size, max_block_size);
-              dim3 block(block_size);
-              VLLM_DISPATCH_VEC_SIZE(calculated_vec_size, [&] {
-                vllm::rms_norm_static_fp8_quant_kernel<scalar_t, fp8_t,
-                                                       vec_size>
-                    <<<grid, block, 0, stream>>>(
-                        out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),
-                        input_stride, weight.data_ptr<scalar_t>(),
-                        scale.data_ptr<float>(), epsilon, num_tokens,
-                        hidden_size);
-              });
+              vllm::rms_norm_static_fp8_quant_kernel<scalar_t, fp8_t>
+                  <<<grid, block, 0, stream>>>(
+                      out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),
+                      input_stride, weight.data_ptr<scalar_t>(),
+                      scale.data_ptr<float>(), epsilon, num_tokens,
+                      hidden_size);
             });
       });
 }

csrc/mamba/mamba_ssm/selective_scan.h

@@ -24,8 +24,6 @@ struct SSMParamsBase {
     int64_t pad_slot_id;
 
     bool delta_softplus;
-    bool cache_enabled;
-    int block_size;
 
     index_t A_d_stride;
     index_t A_dstate_stride;
@@ -48,9 +46,8 @@ struct SSMParamsBase {
     index_t out_z_batch_stride;
     index_t out_z_d_stride;
     index_t ssm_states_batch_stride;
-    index_t ssm_states_dim_stride;
+    index_t ssm_states_dim_stride;  
     index_t ssm_states_dstate_stride;
-    index_t cache_indices_stride;
 
     // Common data pointers.
     void *__restrict__ A_ptr;
@@ -69,9 +66,6 @@ struct SSMParamsBase {
     void *__restrict__ cache_indices_ptr;
     void *__restrict__ has_initial_state_ptr;
 
-    void *__restrict__ block_idx_first_scheduled_token_ptr;  // (batch,) - first block to write
-    void *__restrict__ block_idx_last_scheduled_token_ptr;   // (batch,) - last block to write
-    void *__restrict__ initial_state_idx_ptr;  // (batch,) - index of the initial state to use
 };
 
 

csrc/mamba/mamba_ssm/selective_scan_fwd.cu

@@ -119,7 +119,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
 
     const int* cache_indices = params.cache_indices_ptr == nullptr ? nullptr
         : reinterpret_cast<int *>(params.cache_indices_ptr);
-    const int cache_index = cache_indices == nullptr ? batch_id : cache_indices[batch_id]; 
+    const int cache_index = cache_indices == nullptr ? batch_id : cache_indices[batch_id];
     // cache_index == params.pad_slot_id is defined as padding, so we exit early
     if (cache_index == params.pad_slot_id){
         return;
@@ -133,18 +133,9 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
     input_t *Bvar = reinterpret_cast<input_t *>(params.B_ptr) + sequence_start_index * params.B_batch_stride + group_id * params.B_group_stride;
     weight_t *C = reinterpret_cast<weight_t *>(params.C_ptr) + dim_id * kNRows * params.C_d_stride;
     input_t *Cvar = reinterpret_cast<input_t *>(params.C_ptr) + sequence_start_index * params.C_batch_stride + group_id * params.C_group_stride;
-
-    typename Ktraits::state_t *ssm_states;
-    if (params.cache_enabled) {
-        // APC mode: ssm_states points to the base, we'll use absolute cache slots later
-        ssm_states = reinterpret_cast<typename Ktraits::state_t *>(params.ssm_states_ptr) +
-            dim_id * kNRows * params.ssm_states_dim_stride;
-    } else {
-        // Non-APC mode: offset by cache_index as before
-        ssm_states = reinterpret_cast<typename Ktraits::state_t *>(params.ssm_states_ptr) +
-            cache_index * params.ssm_states_batch_stride +
-            dim_id * kNRows * params.ssm_states_dim_stride;
-    }
+    typename Ktraits::state_t *ssm_states = reinterpret_cast<typename Ktraits::state_t *>(params.ssm_states_ptr) + 
+    cache_index * params.ssm_states_batch_stride + 
+    dim_id * kNRows * params.ssm_states_dim_stride;
     
     float D_val[kNRows] = {0};
     if (params.D_ptr != nullptr) {
@@ -168,22 +159,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
     // }
 
     constexpr int kChunkSize = kNThreads * kNItems;
-
-    // Use block_size for chunking when APC is enabled, otherwise use 2048 for backwards compatibility
-    const int iteration_chunk_size = params.cache_enabled ? params.block_size : 2048;
-    const int n_chunks = (seqlen + iteration_chunk_size - 1) / iteration_chunk_size;
-
-    const int* batch_cache_indices = cache_indices != nullptr ?
-                                     cache_indices + batch_id * params.cache_indices_stride : nullptr;
-    const int* block_idx_first_scheduled = params.block_idx_first_scheduled_token_ptr != nullptr ?
-                                           reinterpret_cast<const int*>(params.block_idx_first_scheduled_token_ptr) : nullptr;
-    const int* block_idx_last_scheduled = params.block_idx_last_scheduled_token_ptr != nullptr ?
-                                          reinterpret_cast<const int*>(params.block_idx_last_scheduled_token_ptr) : nullptr;
-    const int* initial_state_idx = params.initial_state_idx_ptr != nullptr ?
-                                   reinterpret_cast<const int*>(params.initial_state_idx_ptr) : nullptr;
-
-    const size_t load_cache_slot = params.cache_enabled && batch_cache_indices != nullptr ? batch_cache_indices[initial_state_idx[batch_id]] : cache_index;
-
+    const int n_chunks = (seqlen + 2048 - 1) / 2048;
     for (int chunk = 0; chunk < n_chunks; ++chunk) {
         input_t u_vals[kNRows][kNItems], delta_vals_load[kNRows][kNItems];
 
@@ -243,7 +219,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
             if constexpr (kIsVariableC) {
                 auto &smem_load_weight_C = !kIsVariableB ? smem_load_weight : smem_load_weight1;
                 load_weight<Ktraits>(Cvar + state_idx * params.C_dstate_stride, C_vals,
-                    smem_load_weight_C, (seqlen - chunk * kChunkSize) * (1));
+                    smem_load_weight_C, (seqlen - chunk * kChunkSize) * (1 ));
                 if constexpr (!kIsVariableB) {
                     #pragma unroll
                     for (int r = 0; r < kNRows; ++r) {
@@ -266,6 +242,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
                 for (int i = 0; i < kNItems; ++i) {
                     thread_data[i] = make_float2(exp2f(delta_vals[r][i] * A_val[r]),
                                                  !kIsVariableB ? delta_u_vals[r][i] : B_vals[i] * delta_u_vals[r][i]);
+                    
                     if (seqlen % (kNItems * kNThreads) != 0) {  // So that the last state is correct
                         if (threadIdx.x * kNItems + i >= seqlen - chunk * kChunkSize) {
                             thread_data[i] = make_float2(1.f, 0.f);
@@ -273,24 +250,8 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
                     }
                 }
                 // Initialize running total
-                scan_t running_prefix;
-                if (chunk > 0) {
-                    running_prefix = smem_running_prefix[state_idx + r * MAX_DSTATE];
-                } else {
-                    // Load initial state
-                    if (params.cache_enabled && has_initial_state && batch_cache_indices != nullptr) {
-                        size_t state_offset = load_cache_slot * params.ssm_states_batch_stride +
-                                             r * params.ssm_states_dim_stride +
-                                             state_idx * params.ssm_states_dstate_stride;
-                        running_prefix = make_float2(1.0, float(ssm_states[state_offset]));
-                    } else if (has_initial_state) {
-                        // Non-APC mode: load from current batch position
-                        running_prefix = make_float2(1.0, float(ssm_states[state_idx * params.ssm_states_dstate_stride]));
-                    } else {
-                        // No initial state
-                        running_prefix = make_float2(1.0, 0.0);
-                    }
-                }
+
+                scan_t running_prefix = chunk > 0 ? smem_running_prefix[state_idx + r * MAX_DSTATE] : make_float2(1.0, has_initial_state ? float(ssm_states[state_idx * params.ssm_states_dstate_stride]): 0.0);
 
                 SSMScanPrefixCallbackOp<weight_t> prefix_op(running_prefix);
                 typename Ktraits::BlockScanT(smem_scan).InclusiveScan(
@@ -299,25 +260,8 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
                 // There's a syncthreads in the scan op, so we don't need to sync here.
                 // Unless there's only 1 warp, but then it's the same thread (0) reading and writing.
                 if (threadIdx.x == 0) {
-                    smem_running_prefix[state_idx + r * MAX_DSTATE] = prefix_op.running_prefix;
-
-                    // Store state at the end of each chunk when cache is enabled
-                    if (params.cache_enabled && batch_cache_indices != nullptr) {
-
-                        size_t cache_slot;
-                        if (chunk == n_chunks - 1) {
-                            cache_slot = batch_cache_indices[block_idx_last_scheduled[batch_id]];
-                        } else {
-                            cache_slot = batch_cache_indices[block_idx_first_scheduled[batch_id] + chunk];
-                        }
-
-                        size_t state_offset = cache_slot * params.ssm_states_batch_stride +
-                                             r * params.ssm_states_dim_stride +
-                                             state_idx * params.ssm_states_dstate_stride;
-
-                        ssm_states[state_offset] = typename Ktraits::state_t(prefix_op.running_prefix.y);
-                    } else if (!params.cache_enabled && chunk == n_chunks - 1) {
-                        // Non-APC mode: store only final state at current batch position
+                    smem_running_prefix[state_idx] = prefix_op.running_prefix;
+                    if (chunk == n_chunks - 1) {
                         ssm_states[state_idx * params.ssm_states_dstate_stride] = typename Ktraits::state_t(prefix_op.running_prefix.y);
                     }
                 }
@@ -330,6 +274,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
                 }
             }
         }
+        
         input_t *out = reinterpret_cast<input_t *>(params.out_ptr) + sequence_start_index * params.out_batch_stride
             + dim_id * kNRows * params.out_d_stride + chunk * kChunkSize;
         __syncthreads();
@@ -401,9 +346,7 @@ template<typename input_t, typename weight_t, typename state_t>
 void selective_scan_fwd_cuda(SSMParamsBase &params, cudaStream_t stream) {
 
     #ifndef USE_ROCM
-        if (params.cache_enabled && params.block_size == 1024) {
-            selective_scan_fwd_launch<64, 16, input_t, weight_t, state_t>(params, stream);
-        } else if (params.seqlen <= 128) {
+        if (params.seqlen <= 128) {           
             selective_scan_fwd_launch<32, 4, input_t, weight_t, state_t>(params, stream);
         } else if (params.seqlen <= 256) {
             selective_scan_fwd_launch<32, 8, input_t, weight_t, state_t>(params, stream);
@@ -415,9 +358,7 @@ void selective_scan_fwd_cuda(SSMParamsBase &params, cudaStream_t stream) {
             selective_scan_fwd_launch<128, 16, input_t, weight_t, state_t>(params, stream);
         }
     #else
-        if (params.cache_enabled && params.block_size == 1024) {
-            selective_scan_fwd_launch<64, 16, input_t, weight_t, state_t>(params, stream);
-        } else if (params.seqlen <= 256) {
+        if (params.seqlen <= 256) {
             selective_scan_fwd_launch<64, 4, input_t, weight_t, state_t>(params, stream);
         } else if (params.seqlen <= 512) {
             selective_scan_fwd_launch<64, 8, input_t, weight_t, state_t>(params, stream);
@@ -496,17 +437,13 @@ void set_ssm_params_fwd(SSMParamsBase &params,
                         const std::optional<at::Tensor>& D,
                         const std::optional<at::Tensor>& delta_bias,
                         const torch::Tensor ssm_states,
-                        bool has_z,
+                        bool has_z, 
                         bool delta_softplus,
                         const std::optional<at::Tensor>& query_start_loc,
                         const std::optional<at::Tensor>& cache_indices,
                         const std::optional<at::Tensor>& has_initial_state,
                         bool varlen,
-                        int64_t pad_slot_id,
-                        int64_t block_size,
-                        const std::optional<torch::Tensor> &block_idx_first_scheduled_token,
-                        const std::optional<torch::Tensor> &block_idx_last_scheduled_token,
-                        const std::optional<torch::Tensor> &initial_state_idx) {
+                        int64_t pad_slot_id) {
 
     // Reset the parameters
     memset(&params, 0, sizeof(params));
@@ -540,14 +477,6 @@ void set_ssm_params_fwd(SSMParamsBase &params,
     params.cache_indices_ptr = cache_indices.has_value() ? cache_indices.value().data_ptr() : nullptr;
     params.has_initial_state_ptr = has_initial_state.has_value() ? has_initial_state.value().data_ptr() : nullptr;
 
-    // Set cache parameters - cache is enabled if we have direct cache writing params
-    params.cache_enabled = block_idx_first_scheduled_token.has_value();
-    params.block_size = static_cast<int>(block_size);
-
-    // Set direct cache writing pointers
-    params.block_idx_first_scheduled_token_ptr = block_idx_first_scheduled_token.has_value() ? block_idx_first_scheduled_token.value().data_ptr() : nullptr;
-    params.block_idx_last_scheduled_token_ptr = block_idx_last_scheduled_token.has_value() ? block_idx_last_scheduled_token.value().data_ptr() : nullptr;
-    params.initial_state_idx_ptr = initial_state_idx.has_value() ? initial_state_idx.value().data_ptr() : nullptr;
 
     // All stride are in elements, not bytes.
     params.A_d_stride = A.stride(0);
@@ -575,11 +504,9 @@ void set_ssm_params_fwd(SSMParamsBase &params,
         params.out_d_stride = out.stride(0);
 
         params.ssm_states_batch_stride = ssm_states.stride(0);
-        params.ssm_states_dim_stride = ssm_states.stride(1);
+        params.ssm_states_dim_stride = ssm_states.stride(1);  
         params.ssm_states_dstate_stride = ssm_states.stride(2);
 
-        params.cache_indices_stride = cache_indices.has_value() ? cache_indices.value().stride(0) : 0;
-
     }
     else{
         if (!is_variable_B) {
@@ -610,10 +537,8 @@ void set_ssm_params_fwd(SSMParamsBase &params,
         params.out_d_stride = out.stride(1);
         
         params.ssm_states_batch_stride = ssm_states.stride(0);
-        params.ssm_states_dim_stride = ssm_states.stride(1);
+        params.ssm_states_dim_stride = ssm_states.stride(1);  
         params.ssm_states_dstate_stride = ssm_states.stride(2);
-
-        params.cache_indices_stride = cache_indices.has_value() ? cache_indices.value().stride(0) : 0;
     }
 }
 
@@ -629,11 +554,7 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
                   const torch::Tensor &ssm_states,
                   // used to identify padding entries if cache_indices provided
                   // in case of padding, the kernel will return early
-                  int64_t pad_slot_id,
-                  int64_t block_size,
-                  const std::optional<torch::Tensor> &block_idx_first_scheduled_token,
-                  const std::optional<torch::Tensor> &block_idx_last_scheduled_token,
-                  const std::optional<torch::Tensor> &initial_state_idx) {
+                  int64_t pad_slot_id) {
     auto input_type = u.scalar_type();
     auto weight_type = A.scalar_type();
     TORCH_CHECK(input_type == at::ScalarType::Float || input_type == at::ScalarType::Half || input_type == at::ScalarType::BFloat16);
@@ -725,16 +646,7 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
         auto cache_indices_ = cache_indices.value();
         TORCH_CHECK(cache_indices_.scalar_type() == at::ScalarType::Int);
         TORCH_CHECK(cache_indices_.is_cuda());
-
-        // cache_indices can be either 1D (batch_size,) for non-APC mode
-        // or 2D (batch_size, max_positions) for APC mode
-        const bool is_apc_mode = block_idx_first_scheduled_token.has_value();
-        if (is_apc_mode) {
-            TORCH_CHECK(cache_indices_.dim() == 2, "cache_indices must be 2D for APC mode");
-            TORCH_CHECK(cache_indices_.size(0) == batch_size, "cache_indices first dimension must match batch_size");
-        } else {
-            CHECK_SHAPE(cache_indices_, batch_size);
-        }
+        CHECK_SHAPE(cache_indices_, batch_size);
     }
    
 
@@ -774,11 +686,7 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
                        cache_indices,
                        has_initial_state,
                        varlen,
-                       pad_slot_id,
-                       block_size,
-                       block_idx_first_scheduled_token,
-                       block_idx_last_scheduled_token,
-                       initial_state_idx
+                       pad_slot_id
                        );
 
     

csrc/moe/dynamic_4bit_int_moe_cpu.cpp

@@ -87,23 +87,30 @@ torch::Tensor dynamic_4bit_int_moe_cpu(
   const int64_t g_eff_13 = (group_size != -1) ? group_size : H;
   const int64_t g_eff_2 = (group_size != -1) ? group_size : I;
 
-  auto X_all = x_c.index_select(/*dim=*/0, expert_tokens);
-  if (apply_router_weight_on_input) {
-    X_all = X_all.mul(expert_gates.unsqueeze(1));
-  }
-  auto Y_all = at::empty({offsets[E], H}, x_c.options());
+  // Per-expert outputs filled in parallel
+  std::vector<torch::Tensor> y_list(E);
+  y_list.resize(E);
 
   at::parallel_for(0, E, 1, [&](int64_t e_begin, int64_t e_end) {
-    c10::InferenceMode guard;
     for (int64_t e = e_begin; e < e_end; ++e) {
       const int64_t te = counts[e];
       if (te == 0) {
+        y_list[e] = at::empty({0, H}, x_c.options());
         continue;
       }
 
       const int64_t start = offsets[e];
 
-      auto x_e = X_all.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);
+      auto sel_tokens =
+          expert_tokens.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);
+      auto gates_e =
+          expert_gates.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);
+
+      auto x_e = x_c.index_select(/*dim=*/0, sel_tokens);
+
+      if (apply_router_weight_on_input) {
+        x_e = x_e.mul(gates_e.unsqueeze(1));
+      }
 
       auto w13_e = w13_packed.select(/*dim=*/0, e);
       auto w2_e = w2_packed.select(/*dim=*/0, e);
@@ -130,15 +137,17 @@ torch::Tensor dynamic_4bit_int_moe_cpu(
       // W2
       auto y = mm(act, w2_e, g_eff_2, /*in_features=*/I, /*out_features=*/H);
 
+      if (!apply_router_weight_on_input) {
+        y = y.mul(gates_e.unsqueeze(1));
+      }
+
       // Store per-expert result
-      Y_all.narrow(/*dim=*/0, /*start=*/start, /*length=*/te).copy_(y);
+      y_list[e] = y;
     }
   });
 
-  if (!apply_router_weight_on_input) {
-    Y_all = Y_all.mul(expert_gates.unsqueeze(1));
-  }
-
+  // Concatenate all expert outputs to match expert_tokens order
+  auto Y_all = at::cat(y_list, /*dim=*/0);
   auto out = at::zeros({T, H}, x.options());
   out =
       at::index_add(out, /*dim=*/0, /*index=*/expert_tokens, /*source=*/Y_all);

csrc/moe/grouped_topk_kernels.cu

@@ -427,29 +427,11 @@ __device__ inline bool is_finite(const T val) {
 #endif
 }
 
-// Scoring function enums
-enum ScoringFunc {
-  SCORING_NONE = 0,    // no activation function
-  SCORING_SIGMOID = 1  // apply sigmoid
-};
-
-// Efficient sigmoid approximation from TensorRT-LLM
-__device__ inline float sigmoid_accurate(float x) {
-  return 0.5f * tanhf(0.5f * x) + 0.5f;
-}
-
 template <typename T>
-__device__ inline T apply_sigmoid(T val) {
-  float f = cuda_cast<float, T>(val);
-  return cuda_cast<T, float>(sigmoid_accurate(f));
-}
-
-template <typename T>
-__device__ void topk_with_k2(T* output, T const* input, T const* bias,
+__device__ void topk_with_k2(T* output, T const* input,
                              cg::thread_block_tile<32> const& tile,
                              int32_t const lane_id,
-                             int const num_experts_per_group,
-                             int const scoring_func) {
+                             int const num_experts_per_group) {
   // Get the top2 per thread
   T largest = neg_inf<T>();
   T second_largest = neg_inf<T>();
@@ -457,12 +439,6 @@ __device__ void topk_with_k2(T* output, T const* input, T const* bias,
   if (num_experts_per_group > WARP_SIZE) {
     for (int i = lane_id; i < num_experts_per_group; i += WARP_SIZE) {
       T value = input[i];
-      // Apply scoring function if needed
-      if (scoring_func == SCORING_SIGMOID) {
-        value = apply_sigmoid(value);
-      }
-      value = value + bias[i];
-
       if (value > largest) {
         second_largest = largest;
         largest = value;
@@ -472,13 +448,7 @@ __device__ void topk_with_k2(T* output, T const* input, T const* bias,
     }
   } else {
     for (int i = lane_id; i < num_experts_per_group; i += WARP_SIZE) {
-      T value = input[i];
-      // Apply scoring function if needed
-      if (scoring_func == SCORING_SIGMOID) {
-        value = apply_sigmoid(value);
-      }
-      value = value + bias[i];
-      largest = value;
+      largest = input[i];
     }
   }
 
@@ -502,21 +472,17 @@ __device__ void topk_with_k2(T* output, T const* input, T const* bias,
 }
 
 template <typename T>
-__global__ void topk_with_k2_kernel(T* output, T* input, T const* bias,
+__global__ void topk_with_k2_kernel(T* output, T* input,
                                     int64_t const num_tokens,
                                     int64_t const num_cases,
                                     int64_t const n_group,
-                                    int64_t const num_experts_per_group,
-                                    int const scoring_func) {
+                                    int64_t const num_experts_per_group) {
   int32_t warp_id = threadIdx.x / WARP_SIZE;
   int32_t lane_id = threadIdx.x % WARP_SIZE;
 
   int32_t case_id = blockIdx.x * NUM_WARPS_PER_BLOCK + warp_id;
   if (case_id < num_cases) {
     input += case_id * num_experts_per_group;
-    // bias is per expert group, offset to current group
-    int32_t group_id = case_id % n_group;
-    T const* group_bias = bias + group_id * num_experts_per_group;
     output += case_id;
 
     cg::thread_block block = cg::this_thread_block();
@@ -525,8 +491,7 @@ __global__ void topk_with_k2_kernel(T* output, T* input, T const* bias,
 #if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
     asm volatile("griddepcontrol.wait;");
 #endif
-    topk_with_k2(output, input, group_bias, tile, lane_id,
-                 num_experts_per_group, scoring_func);
+    topk_with_k2(output, input, tile, lane_id, num_experts_per_group);
   }
 #if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
   asm volatile("griddepcontrol.launch_dependents;");
@@ -535,15 +500,16 @@ __global__ void topk_with_k2_kernel(T* output, T* input, T const* bias,
 
 template <typename T, typename IdxT>
 __global__ void group_idx_and_topk_idx_kernel(
-    T* scores, T const* group_scores, float* topk_values, IdxT* topk_indices,
-    T const* bias, int64_t const num_tokens, int64_t const n_group,
+    T* scores, T const* group_scores, T* topk_values, IdxT* topk_indices,
+    T* scores_with_bias, int64_t const num_tokens, int64_t const n_group,
     int64_t const topk_group, int64_t const topk, int64_t const num_experts,
     int64_t const num_experts_per_group, bool renormalize,
-    double routed_scaling_factor, int scoring_func) {
+    double routed_scaling_factor) {
   int32_t warp_id = threadIdx.x / WARP_SIZE;
   int32_t lane_id = threadIdx.x % WARP_SIZE;
   int32_t case_id =
       blockIdx.x * NUM_WARPS_PER_BLOCK + warp_id;  // one per token
+  scores_with_bias += case_id * num_experts;
   scores += case_id * num_experts;
   group_scores += case_id * n_group;
   topk_values += case_id * topk;
@@ -611,16 +577,10 @@ __global__ void group_idx_and_topk_idx_kernel(
         int32_t offset = i_group * num_experts_per_group;
         for (int32_t i = lane_id; i < align_num_experts_per_group;
              i += WARP_SIZE) {
-          T candidates = neg_inf<T>();
-          if (i < num_experts_per_group) {
-            // Apply scoring function (if any) and add bias
-            T input = scores[offset + i];
-            if (is_finite(input)) {
-              T score = (scoring_func == SCORING_SIGMOID) ? apply_sigmoid(input)
-                                                          : input;
-              candidates = score + bias[offset + i];
-            }
-          }
+          T candidates = (i < num_experts_per_group) &&
+                                 is_finite(scores_with_bias[offset + i])
+                             ? scores_with_bias[offset + i]
+                             : neg_inf<T>();
           queue.add(candidates, offset + i);
         }
         if (group_scores[i_group] == topk_group_value) {
@@ -642,12 +602,11 @@ __global__ void group_idx_and_topk_idx_kernel(
     for (int i = lane_id;
          i < warp_topk::round_up_to_multiple_of<WARP_SIZE>(topk);
          i += WARP_SIZE) {
-      T value = cuda_cast<T, float>(0.0f);
+      T value =
+          i < topk
+              ? scores[s_topk_idx[i]]
+              : cuda_cast<T, float>(0.0f);  // Load the valid value of expert
       if (i < topk) {
-        // Load the score value (without bias) for normalization
-        T input = scores[s_topk_idx[i]];
-        value =
-            (scoring_func == SCORING_SIGMOID) ? apply_sigmoid(input) : input;
         s_topk_value[i] = value;
       }
       topk_sum +=
@@ -668,12 +627,12 @@ __global__ void group_idx_and_topk_idx_kernel(
           value = cuda_cast<float, T>(s_topk_value[i]) * routed_scaling_factor;
         }
         topk_indices[i] = s_topk_idx[i];
-        topk_values[i] = value;
+        topk_values[i] = cuda_cast<T, float>(value);
       }
     } else {
       for (int i = lane_id; i < topk; i += WARP_SIZE) {
         topk_indices[i] = i;
-        topk_values[i] = 1.0f / topk;
+        topk_values[i] = cuda_cast<T, float>(1.0f / topk);
       }
     }
     // Note: when if_proceed_next_topk==false, choose the first 8 experts as the
@@ -685,12 +644,12 @@ __global__ void group_idx_and_topk_idx_kernel(
 }
 
 template <typename T, typename IdxT>
-void invokeNoAuxTc(T* scores, T* group_scores, float* topk_values,
-                   IdxT* topk_indices, T const* bias, int64_t const num_tokens,
-                   int64_t const num_experts, int64_t const n_group,
-                   int64_t const topk_group, int64_t const topk,
-                   bool const renormalize, double const routed_scaling_factor,
-                   int const scoring_func, bool enable_pdl = false,
+void invokeNoAuxTc(T* scores, T* group_scores, T* topk_values,
+                   IdxT* topk_indices, T* scores_with_bias,
+                   int64_t const num_tokens, int64_t const num_experts,
+                   int64_t const n_group, int64_t const topk_group,
+                   int64_t const topk, bool const renormalize,
+                   double const routed_scaling_factor, bool enable_pdl = false,
                    cudaStream_t const stream = 0) {
   int64_t num_cases = num_tokens * n_group;
   int64_t topk_with_k2_num_blocks = (num_cases - 1) / NUM_WARPS_PER_BLOCK + 1;
@@ -705,9 +664,8 @@ void invokeNoAuxTc(T* scores, T* group_scores, float* topk_values,
   attrs[0].val.programmaticStreamSerializationAllowed = enable_pdl;
   config.numAttrs = 1;
   config.attrs = attrs;
-  cudaLaunchKernelEx(&config, kernel_instance1, group_scores, scores, bias,
-                     num_tokens, num_cases, n_group, num_experts / n_group,
-                     scoring_func);
+  cudaLaunchKernelEx(&config, kernel_instance1, group_scores, scores_with_bias,
+                     num_tokens, num_cases, n_group, num_experts / n_group);
 
   int64_t topk_with_k_group_num_blocks =
       (num_tokens - 1) / NUM_WARPS_PER_BLOCK + 1;
@@ -724,18 +682,19 @@ void invokeNoAuxTc(T* scores, T* group_scores, float* topk_values,
   config.numAttrs = 1;
   config.attrs = attrs;
   cudaLaunchKernelEx(&config, kernel_instance2, scores, group_scores,
-                     topk_values, topk_indices, bias, num_tokens, n_group,
-                     topk_group, topk, num_experts, num_experts / n_group,
-                     renormalize, routed_scaling_factor, scoring_func);
+                     topk_values, topk_indices, scores_with_bias, num_tokens,
+                     n_group, topk_group, topk, num_experts,
+                     num_experts / n_group, renormalize, routed_scaling_factor);
 }
 
 #define INSTANTIATE_NOAUX_TC(T, IdxT)                                       \
   template void invokeNoAuxTc<T, IdxT>(                                     \
-      T * scores, T * group_scores, float* topk_values, IdxT* topk_indices, \
-      T const* bias, int64_t const num_tokens, int64_t const num_experts,   \
-      int64_t const n_group, int64_t const topk_group, int64_t const topk,  \
-      bool const renormalize, double const routed_scaling_factor,           \
-      int const scoring_func, bool enable_pdl, cudaStream_t const stream);
+      T * scores, T * group_scores, T * topk_values, IdxT * topk_indices,   \
+      T * scores_with_bias, int64_t const num_tokens,                       \
+      int64_t const num_experts, int64_t const n_group,                     \
+      int64_t const topk_group, int64_t const topk, bool const renormalize, \
+      double const routed_scaling_factor, bool enable_pdl,                  \
+      cudaStream_t const stream);
 
 INSTANTIATE_NOAUX_TC(float, int32_t);
 INSTANTIATE_NOAUX_TC(half, int32_t);
@@ -744,32 +703,28 @@ INSTANTIATE_NOAUX_TC(__nv_bfloat16, int32_t);
 }  // namespace vllm
 
 std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
-    torch::Tensor const& scores, int64_t n_group, int64_t topk_group,
-    int64_t topk, bool renormalize, double routed_scaling_factor,
-    torch::Tensor const& bias, int64_t scoring_func = 0) {
-  auto data_type = scores.scalar_type();
-  auto input_size = scores.sizes();
+    torch::Tensor const& scores, torch::Tensor const& scores_with_bias,
+    int64_t n_group, int64_t topk_group, int64_t topk, bool renormalize,
+    double routed_scaling_factor) {
+  auto data_type = scores_with_bias.scalar_type();
+  auto input_size = scores_with_bias.sizes();
   int64_t num_tokens = input_size[0];
   int64_t num_experts = input_size[1];
-  TORCH_CHECK(input_size.size() == 2, "scores must be a 2D Tensor");
+  TORCH_CHECK(input_size.size() == 2, "scores_with_bias must be a 2D Tensor");
   TORCH_CHECK(num_experts % n_group == 0,
               "num_experts should be divisible by n_group");
   TORCH_CHECK(n_group <= 32,
               "n_group should be smaller than or equal to 32 for now");
   TORCH_CHECK(topk <= 32, "topk should be smaller than or equal to 32 for now");
-  TORCH_CHECK(scoring_func == vllm::moe::SCORING_NONE ||
-                  scoring_func == vllm::moe::SCORING_SIGMOID,
-              "scoring_func must be SCORING_NONE (0) or SCORING_SIGMOID (1)");
 
   torch::Tensor group_scores = torch::empty(
       {num_tokens, n_group}, torch::dtype(data_type).device(torch::kCUDA));
-  // Always output float32 for topk_values (eliminates Python-side conversion)
   torch::Tensor topk_values = torch::empty(
-      {num_tokens, topk}, torch::dtype(torch::kFloat32).device(torch::kCUDA));
+      {num_tokens, topk}, torch::dtype(data_type).device(torch::kCUDA));
   torch::Tensor topk_indices = torch::empty(
       {num_tokens, topk}, torch::dtype(torch::kInt32).device(torch::kCUDA));
 
-  auto stream = c10::cuda::getCurrentCUDAStream(scores.get_device());
+  auto stream = c10::cuda::getCurrentCUDAStream(scores_with_bias.get_device());
 
   switch (data_type) {
     case torch::kFloat16:
@@ -777,11 +732,11 @@ std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
       vllm::moe::invokeNoAuxTc<half, int32_t>(
           reinterpret_cast<half*>(scores.mutable_data_ptr()),
           reinterpret_cast<half*>(group_scores.mutable_data_ptr()),
-          reinterpret_cast<float*>(topk_values.mutable_data_ptr()),
+          reinterpret_cast<half*>(topk_values.mutable_data_ptr()),
           reinterpret_cast<int32_t*>(topk_indices.mutable_data_ptr()),
-          reinterpret_cast<half const*>(bias.data_ptr()), num_tokens,
+          reinterpret_cast<half*>(scores_with_bias.data_ptr()), num_tokens,
           num_experts, n_group, topk_group, topk, renormalize,
-          routed_scaling_factor, static_cast<int>(scoring_func), false, stream);
+          routed_scaling_factor, false, stream);
       break;
     case torch::kFloat32:
       // Handle Float32
@@ -790,20 +745,20 @@ std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
           reinterpret_cast<float*>(group_scores.mutable_data_ptr()),
           reinterpret_cast<float*>(topk_values.mutable_data_ptr()),
           reinterpret_cast<int32_t*>(topk_indices.mutable_data_ptr()),
-          reinterpret_cast<float const*>(bias.data_ptr()), num_tokens,
+          reinterpret_cast<float*>(scores_with_bias.data_ptr()), num_tokens,
           num_experts, n_group, topk_group, topk, renormalize,
-          routed_scaling_factor, static_cast<int>(scoring_func), false, stream);
+          routed_scaling_factor, false, stream);
       break;
     case torch::kBFloat16:
       // Handle BFloat16
       vllm::moe::invokeNoAuxTc<__nv_bfloat16, int32_t>(
           reinterpret_cast<__nv_bfloat16*>(scores.mutable_data_ptr()),
           reinterpret_cast<__nv_bfloat16*>(group_scores.mutable_data_ptr()),
-          reinterpret_cast<float*>(topk_values.mutable_data_ptr()),
+          reinterpret_cast<__nv_bfloat16*>(topk_values.mutable_data_ptr()),
           reinterpret_cast<int32_t*>(topk_indices.mutable_data_ptr()),
-          reinterpret_cast<__nv_bfloat16 const*>(bias.data_ptr()), num_tokens,
-          num_experts, n_group, topk_group, topk, renormalize,
-          routed_scaling_factor, static_cast<int>(scoring_func), false, stream);
+          reinterpret_cast<__nv_bfloat16*>(scores_with_bias.data_ptr()),
+          num_tokens, num_experts, n_group, topk_group, topk, renormalize,
+          routed_scaling_factor, false, stream);
       break;
     default:
       // Handle other data types

csrc/moe/moe_lora_align_sum_kernels.cu

@@ -28,16 +28,11 @@ __global__ void moe_lora_align_sum_kernel(
     int64_t block_size, int num_experts, int max_loras, size_t numel,
     int max_num_tokens_padded, int max_num_m_blocks,
     int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ expert_ids,
-    int topk_num, int32_t* total_tokens_post_pad, int32_t* adapter_enabled,
-    int32_t* lora_ids) {
+    int topk_num, int32_t* total_tokens_post_pad) {
   const size_t tokens_per_thread = div_ceil(numel, blockDim.x);
   const size_t start_idx = threadIdx.x * tokens_per_thread;
 
-  int lora_idx = blockIdx.x;
-  int lora_id = lora_ids[lora_idx];
-  if (lora_id == -1 || adapter_enabled[lora_id] == 0) {
-    return;
-  }
+  int lora_id = blockIdx.x;
   extern __shared__ int32_t shared_mem[];
   int32_t* cumsum = shared_mem;
   token_cnts_t* tokens_cnts = (token_cnts_t*)(shared_mem + num_experts + 1);
@@ -126,13 +121,14 @@ __global__ void moe_lora_align_sum_kernel(
   }
 }
 
-void moe_lora_align_block_size(
-    torch::Tensor topk_ids, torch::Tensor token_lora_mapping,
-    int64_t num_experts, int64_t block_size, int64_t max_loras,
-    int64_t max_num_tokens_padded, int64_t max_num_m_blocks,
-    torch::Tensor sorted_token_ids, torch::Tensor expert_ids,
-    torch::Tensor num_tokens_post_pad, torch::Tensor adapter_enabled,
-    torch::Tensor lora_ids) {
+void moe_lora_align_block_size(torch::Tensor topk_ids,
+                               torch::Tensor token_lora_mapping,
+                               int64_t num_experts, int64_t block_size,
+                               int64_t max_loras, int64_t max_num_tokens_padded,
+                               int64_t max_num_m_blocks,
+                               torch::Tensor sorted_token_ids,
+                               torch::Tensor expert_ids,
+                               torch::Tensor num_tokens_post_pad) {
   const int topk_num = topk_ids.size(1);
 
   TORCH_CHECK(block_size > 0, "block_size should be greater than 0. ");
@@ -168,7 +164,6 @@ void moe_lora_align_block_size(
             max_loras, topk_ids.numel(), max_num_tokens_padded,
             max_num_m_blocks, sorted_token_ids.data_ptr<int32_t>(),
             expert_ids.data_ptr<int32_t>(), topk_num,
-            num_tokens_post_pad.data_ptr<int32_t>(),
-            adapter_enabled.data_ptr<int32_t>(), lora_ids.data_ptr<int32_t>());
+            num_tokens_post_pad.data_ptr<int32_t>());
       });
 }

csrc/moe/moe_ops.h

@@ -20,13 +20,14 @@ void batched_moe_align_block_size(int64_t max_tokens_per_batch,
                                   torch::Tensor expert_ids,
                                   torch::Tensor num_tokens_post_pad);
 
-void moe_lora_align_block_size(
-    torch::Tensor topk_ids, torch::Tensor token_lora_mapping,
-    int64_t num_experts, int64_t block_size, int64_t max_loras,
-    int64_t max_num_tokens_padded, int64_t max_num_m_blocks,
-    torch::Tensor sorted_token_ids, torch::Tensor expert_ids,
-    torch::Tensor num_tokens_post_pad, torch::Tensor adapter_enabled,
-    torch::Tensor lora_ids);
+void moe_lora_align_block_size(torch::Tensor topk_ids,
+                               torch::Tensor token_lora_mapping,
+                               int64_t num_experts, int64_t block_size,
+                               int64_t max_loras, int64_t max_num_tokens_padded,
+                               int64_t max_num_m_blocks,
+                               torch::Tensor sorted_token_ids,
+                               torch::Tensor expert_ids,
+                               torch::Tensor num_tokens_post_pad);
 #ifndef USE_ROCM
 torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
                              torch::Tensor b_qweight, torch::Tensor b_scales,
@@ -39,9 +40,9 @@ torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
                              int64_t BLOCK_SIZE_K, int64_t bit);
 
 std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
-    torch::Tensor const& scores, int64_t n_group, int64_t topk_group,
-    int64_t topk, bool renormalize, double routed_scaling_factor,
-    torch::Tensor const& bias, int64_t scoring_func);
+    torch::Tensor const& scores, torch::Tensor const& scores_with_bias,
+    int64_t n_group, int64_t topk_group, int64_t topk, bool renormalize,
+    double routed_scaling_factor);
 #endif
 
 bool moe_permute_unpermute_supported();

csrc/moe/torch_bindings.cpp

@@ -44,9 +44,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
       "                     int max_num_m_blocks, "
       "                     Tensor !sorted_token_ids,"
       "                     Tensor !experts_ids,"
-      "                     Tensor !num_tokens_post_pad,"
-      "                     Tensor !adapter_enabled,"
-      "                     Tensor !lora_ids) -> () ");
+      "                     Tensor !num_tokens_post_pad) -> () ");
   m.impl("moe_lora_align_block_size", torch::kCUDA, &moe_lora_align_block_size);
 
 #ifndef USE_ROCM
@@ -107,10 +105,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
 
   // Apply grouped topk routing to select experts.
   m.def(
-      "grouped_topk(Tensor scores, int n_group, int "
+      "grouped_topk(Tensor scores, Tensor scores_with_bias, int n_group, int "
       "topk_group, int topk, bool renormalize, float "
-      "routed_scaling_factor, Tensor bias, int scoring_func) -> (Tensor, "
-      "Tensor)");
+      "routed_scaling_factor) -> (Tensor, Tensor)");
   m.impl("grouped_topk", torch::kCUDA, &grouped_topk);
 #endif
 }

csrc/ops.h

@@ -92,12 +92,6 @@ void rms_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
 void fused_add_rms_norm(torch::Tensor& input, torch::Tensor& residual,
                         torch::Tensor& weight, double epsilon);
 
-void fused_qk_norm_rope(torch::Tensor& qkv, int64_t num_heads_q,
-                        int64_t num_heads_k, int64_t num_heads_v,
-                        int64_t head_dim, double eps, torch::Tensor& q_weight,
-                        torch::Tensor& k_weight, torch::Tensor& cos_sin_cache,
-                        bool is_neox, torch::Tensor& position_ids);
-
 void apply_repetition_penalties_(torch::Tensor& logits,
                                  const torch::Tensor& prompt_mask,
                                  const torch::Tensor& output_mask,
@@ -327,19 +321,17 @@ void dynamic_per_token_scaled_fp8_quant(
     torch::Tensor& out, torch::Tensor const& input, torch::Tensor& scale,
     std::optional<torch::Tensor> const& scale_ub);
 
-void selective_scan_fwd(
-    const torch::Tensor& u, const torch::Tensor& delta, const torch::Tensor& A,
-    const torch::Tensor& B, const torch::Tensor& C,
-    const std::optional<torch::Tensor>& D_,
-    const std::optional<torch::Tensor>& z_,
-    const std::optional<torch::Tensor>& delta_bias_, bool delta_softplus,
-    const std::optional<torch::Tensor>& query_start_loc,
-    const std::optional<torch::Tensor>& cache_indices,
-    const std::optional<torch::Tensor>& has_initial_state,
-    const torch::Tensor& ssm_states, int64_t pad_slot_id, int64_t block_size,
-    const std::optional<torch::Tensor>& block_idx_first_scheduled_token,
-    const std::optional<torch::Tensor>& block_idx_last_scheduled_token,
-    const std::optional<torch::Tensor>& initial_state_idx);
+void selective_scan_fwd(const torch::Tensor& u, const torch::Tensor& delta,
+                        const torch::Tensor& A, const torch::Tensor& B,
+                        const torch::Tensor& C,
+                        const std::optional<torch::Tensor>& D_,
+                        const std::optional<torch::Tensor>& z_,
+                        const std::optional<torch::Tensor>& delta_bias_,
+                        bool delta_softplus,
+                        const std::optional<torch::Tensor>& query_start_loc,
+                        const std::optional<torch::Tensor>& cache_indices,
+                        const std::optional<torch::Tensor>& has_initial_state,
+                        const torch::Tensor& ssm_states, int64_t pad_slot_id);
 
 torch::Tensor dynamic_4bit_int_moe_cpu(
     torch::Tensor x, torch::Tensor topk_ids, torch::Tensor topk_weights,

csrc/quantization/activation_kernels.cu

@@ -279,17 +279,17 @@ __device__ __forceinline__ void token_bounds(int32_t n_tokens,
 }
 
 template <int BLOCK_COUNT, int SMEM_SIZE_BYTES_Y, typename fp8_type,
-          typename scale_t, int THREADS, typename Idx_t, bool CEIL_UE8M0,
-          int GROUP_SIZE = 128, int NUM_STAGES = 3>
+          int THREADS, typename Idx_t, bool USE_UE8M0, int GROUP_SIZE = 128,
+          int NUM_STAGES = 3>
 __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
     const __nv_bfloat16* __restrict__ _input, fp8_type* __restrict__ _y_q,
-    scale_t* __restrict__ _y_s, const int32_t* __restrict__ tokens_per_expert,
+    float* __restrict__ _y_s, const int32_t* __restrict__ tokens_per_expert,
     // sizes
     Idx_t E, Idx_t T, Idx_t H,
     // strides (in elements)
     Idx_t stride_i_e, Idx_t stride_i_t, Idx_t stride_i_h, Idx_t stride_yq_e,
     Idx_t stride_yq_t, Idx_t stride_yq_h, Idx_t stride_ys_e, Idx_t stride_ys_t,
-    Idx_t stride_ys_g, Idx_t stride_ys_p, Idx_t stride_counts_e) {
+    Idx_t stride_ys_g, Idx_t stride_counts_e) {
 #ifndef USE_ROCM
   static constexpr int NUM_WARPS = THREADS / WARP_SIZE;
 
@@ -466,22 +466,9 @@ __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
 
   __nv_fp8x4_e4m3* y_q_base_ptr =
       reinterpret_cast<__nv_fp8x4_e4m3*>(_y_q) + lane_id;
-
-  Idx_t scale_group_offset = 0;
-  if constexpr (std::is_same<scale_t, uint8_t>::value) {
-    // packed int32_t format
-    int pack_id = warp_position_scales / 4;
-    int scale_in_pack = warp_position_scales % 4;
-    scale_group_offset = pack_id * stride_ys_p + scale_in_pack * stride_ys_g;
-  } else {
-    scale_group_offset = warp_position_scales * stride_ys_g;
-  }
-
-  scale_t* const y_scale_base_ptr = _y_s + scale_group_offset;
+  auto y_scale_base_ptr = _y_s + warp_position_scales * stride_ys_g;
 
   for (auto j = tokens_lower; j < tokens_upper; j++) {
-    int current_group_id = warp_position_scales;  // Running count of which
-                                                  // group is being processed
     const Idx_t base_ys = expert_id * stride_ys_e;
     auto y_s_ptr = y_scale_base_ptr + base_ys + token_offset * stride_ys_t;
     __nv_fp8x4_e4m3* y_q_ptr =
@@ -522,7 +509,7 @@ __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
 
       __nv_bfloat16 y_s = __hmul(warp_max(_y_max2.x), fp8_inv);
 
-      if constexpr (CEIL_UE8M0) {
+      if constexpr (USE_UE8M0) {
         y_s = hexp2(hceil(hlog2(y_s)));
       }
 
@@ -540,24 +527,8 @@ __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
       y_q_ptr += WARP_SIZE * stride_yq_h;
 
       if (!lane_id) {
-        // Store scales.
-        if constexpr (std::is_same<scale_t, uint8_t>::value) {
-          // Packed UE8MO format. Remove Mantissa.
-          *y_s_ptr = reinterpret_cast<int16_t&>(y_s) >> 7;
-
-          bool const jump_pack = (current_group_id + 1) % 4 == 0;
-          // Minus 3 because we need to get to the first group in the
-          // next pack.
-          y_s_ptr += jump_pack ? (stride_ys_p - 3) : stride_ys_g;
-
-        } else {
-          // float32 format
-          static_assert(std::is_same<scale_t, float>::value);
-          *y_s_ptr = y_s;
-          y_s_ptr += stride_ys_g;
-        }
-
-        current_group_id += 1;
+        *y_s_ptr = y_s;
+        y_s_ptr += stride_ys_g;
       }
     }
   }
@@ -602,21 +573,16 @@ void persistent_masked_m_silu_mul_quant(
     const at::Tensor& tokens_per_expert,  // (E)
     at::Tensor& y_q,                      // (E, T, H) [OUT]
     at::Tensor& y_s,                      // (E, T, H//group_size) [OUT]
-    bool cast_scale_ue8m0) {
+    bool use_ue8m0) {
 #ifndef USE_ROCM
 
   // This kernel currently only supports H % 128 == 0 and assumes a
   // fixed GROUP_SIZE of 128.
-  static constexpr int GROUP_SIZE = 128;
-
   TORCH_CHECK(input.dtype() == torch::kBFloat16);
   TORCH_CHECK(y_q.dtype() == torch::kFloat8_e4m3fn ||
               y_q.dtype() == torch::kFloat8_e4m3fnuz);
-  TORCH_CHECK(input.size(-1) % (GROUP_SIZE * 2) == 0);
-
-  bool const is_packed_ue8m0 =
-      (y_s.dtype() == torch::kInt32 && cast_scale_ue8m0);
-  TORCH_CHECK(y_s.dtype() == torch::kFloat32 || is_packed_ue8m0);
+  TORCH_CHECK(y_s.dtype() == torch::kFloat32);
+  TORCH_CHECK(input.size(-1) % 256 == 0);
 
   using Idx_t = int64_t;
 
@@ -629,18 +595,17 @@ void persistent_masked_m_silu_mul_quant(
   Idx_t stride_yq_e = y_q.stride(0);
   Idx_t stride_yq_t = y_q.stride(1);
   Idx_t stride_yq_h = y_q.stride(2);
+  Idx_t stride_ys_e = y_s.stride(0);
+  Idx_t stride_ys_t = y_s.stride(1);
+  Idx_t stride_ys_g = y_s.stride(2);
 
   Idx_t stride_counts_e = tokens_per_expert.stride(0);
 
-  int const NUM_GROUPS = H / GROUP_SIZE;
+  static constexpr int GROUP_SIZE = 128;
 
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  // TODO: Get this from cuda_arch ?
-  static constexpr int SILU_V2_BLOCK_COUNT = 132 * 32;
-
-  #define KERNEL(BLOCK_COUNT, scale_t, STRIDE_YS_E, STRIDE_YS_T, STRIDE_YS_G,  \
-                 STRIDE_YS_P, CEIL_UE8M0, THREAD_COUNT, STAGES)                \
+  #define KERNEL(BLOCK_COUNT, USE_UE8M0, THREAD_COUNT, STAGES)                 \
     static constexpr int NUM_WARPS = THREAD_COUNT / WARP_SIZE;                 \
     int sms = SILU_V2_BLOCK_COUNT;                                             \
     static constexpr int max_shared_mem_bytes =                                \
@@ -650,86 +615,38 @@ void persistent_masked_m_silu_mul_quant(
     VLLM_DISPATCH_FP8_TYPES(                                                   \
         y_q.scalar_type(), "silu_mul_fp8_quant_deep_gemm_kernel", [&] {        \
           vllm::silu_mul_fp8_quant_deep_gemm_kernel<                           \
-              BLOCK_COUNT, max_shared_mem_bytes, fp8_t, scale_t, THREAD_COUNT, \
-              Idx_t, CEIL_UE8M0, GROUP_SIZE, STAGES>                           \
+              BLOCK_COUNT, max_shared_mem_bytes, fp8_t, THREAD_COUNT, Idx_t,   \
+              USE_UE8M0, GROUP_SIZE, STAGES>                                   \
               <<<grid, block, max_shared_mem_bytes + (E + 1) * 16, stream>>>(  \
                   reinterpret_cast<__nv_bfloat16*>(input.data_ptr()),          \
-                  (fp8_t*)y_q.data_ptr(),                                      \
-                  reinterpret_cast<scale_t*>(y_s.data_ptr()),                  \
+                  (fp8_t*)y_q.data_ptr(), y_s.data_ptr<float>(),               \
                   reinterpret_cast<int32_t*>(tokens_per_expert.data_ptr()), E, \
                   T, H, stride_i_e, stride_i_t, stride_i_h, stride_yq_e,       \
-                  stride_yq_t, stride_yq_h, STRIDE_YS_E, STRIDE_YS_T,          \
-                  STRIDE_YS_G, STRIDE_YS_P, stride_counts_e);                  \
+                  stride_yq_t, stride_yq_h, stride_ys_e, stride_ys_t,          \
+                  stride_ys_g, stride_counts_e);                               \
         });
 
-  #define LAUNCH_ON_H(scale_t, STRIDE_YS_E, STRIDE_YS_T, STRIDE_YS_G,         \
-                      STRIDE_YS_P, CEIL_UE8M0)                                \
-    if (H >= 4096 && (NUM_GROUPS % 8) == 0) {                                 \
-      /* 8 warp config */                                                     \
-      static constexpr int NUM_STAGES = 4;                                    \
-      static constexpr int THREAD_COUNT = 256;                                \
-      KERNEL(SILU_V2_BLOCK_COUNT, scale_t, STRIDE_YS_E, STRIDE_YS_T,          \
-             STRIDE_YS_G, STRIDE_YS_P, CEIL_UE8M0, THREAD_COUNT, NUM_STAGES); \
-    } else {                                                                  \
-      /* 1 warp config */                                                     \
-      static constexpr int THREAD_COUNT = 32;                                 \
-      KERNEL(SILU_V2_BLOCK_COUNT, scale_t, STRIDE_YS_E, STRIDE_YS_T,          \
-             STRIDE_YS_G, STRIDE_YS_P, CEIL_UE8M0, THREAD_COUNT, 2);          \
+  static constexpr int SILU_V2_BLOCK_COUNT = 132 * 32;
+
+  if (!use_ue8m0) {
+    if (H >= 4096) {
+      static constexpr int NUM_STAGES = 4;
+      static constexpr int THREAD_COUNT = 256;
+      KERNEL(SILU_V2_BLOCK_COUNT, false, THREAD_COUNT, NUM_STAGES);
+    } else {
+      static constexpr int THREAD_COUNT = 32;
+      KERNEL(SILU_V2_BLOCK_COUNT, false, THREAD_COUNT, 2);
+    }
+  } else {
+    if (H >= 4096) {
+      static constexpr int NUM_STAGES = 4;
+      static constexpr int THREAD_COUNT = 256;
+      KERNEL(SILU_V2_BLOCK_COUNT, true, THREAD_COUNT, NUM_STAGES);
+    } else {
+      static constexpr int THREAD_COUNT = 32;
+      KERNEL(SILU_V2_BLOCK_COUNT, true, THREAD_COUNT, 2);
     }
-
-  Idx_t stride_ys_e = y_s.stride(0);
-  Idx_t stride_ys_t = y_s.stride(1);
-  Idx_t stride_ys_g = y_s.stride(2);
-  Idx_t stride_ys_p = 0;
-  if (!cast_scale_ue8m0) {
-    TORCH_CHECK(!is_packed_ue8m0);
-    LAUNCH_ON_H(float, stride_ys_e, stride_ys_t, stride_ys_g, stride_ys_p,
-                false);
-    return;
   }
 
-  if (!is_packed_ue8m0) {
-    // UE8M0 but not packed
-    LAUNCH_ON_H(float, stride_ys_e, stride_ys_t, stride_ys_g, stride_ys_p,
-                true);
-    return;
-  }
-
-  TORCH_CHECK(cast_scale_ue8m0 && is_packed_ue8m0);
-  TORCH_CHECK(y_s.dtype() == torch::kInt32);
-
-  // Int32 packed ue8m0 scales tensor.
-  // Let E, T, G be the number to experts, number of tokens and number of groups
-  // respectively. Let, E = 2, T = 4, G = 6, in this case the int32 scales
-  // tensor are of shape [1, 4, 2] and stride [8, 1, 4]. The scales are expected
-  // to be arranged as follows,
-  // [[T0G0-T0G1-T0G2-T0G3, T0G4-T0G5-X-X,],
-  //  [T1G0-T1G1-T1G2-T1G3, T1G4-T1G5-X-X,]
-  //  [T2G0-T2G1-T2G2-T2G3, T2G4-T2G5-X-X,]
-  //  [T3G0-T3G1-T3G2-T3G3, T3G4-T3G5-X-X,]]
-  // where, TxGy is the scale ue8m0 scale value of Token x, Group y.
-  //
-  // In memory (in bytes) the scale values are arranged as,
-  //  [T0G0, T0G1, T0G2, T0G3, T1G0, T1G2, T1G3, T1G4, T2G0, T2G1, T2G3, T2G4,
-  //   T3G0, T3G1, T3G2, T3G3, T0G4, T0G5, X, X, T1G4, T1G5, X, X, T2G4, T2G5,
-  //   X, X, T3G4, T3G5, X, X]
-  //
-  // An Int32 tensor of size [1, 4, 2] and stride [8, 1, 4] can be represented
-  // as an uint8 tensor of shape [1, 2, 4, 4] and stride [32, 16, 4, 1]. In
-  // english, ignoring the Experts dimension, the original int32 tensor is
-  // simply treated as two packed [4, 4] uint8 tensor (or two [4, 1] int32
-  // tensor). The following strides setting reflects this change. Caveat: This
-  // means that the G dimension is no longer contiguous. i.e. Note that to move
-  // from G3 to G4, we need to jump along the packing dimension. The kernel
-  // handles this case.
-
-  stride_ys_e *= sizeof(int32_t);
-  stride_ys_p = T * sizeof(int32_t);  // Packing dimension
-  stride_ys_t = sizeof(int32_t);
-  stride_ys_g = 1;
-
-  LAUNCH_ON_H(uint8_t, stride_ys_e, stride_ys_t, stride_ys_g, stride_ys_p,
-              true);
-
 #endif
 }