[Misc] Add EPLB debug logging for balance diagnostics

Add comprehensive debug logging to the EPLB system to help diagnose
expert load balancing issues in wideEP deployments:

- Per-step balance breakdown: worst/best layer indices, min/max rank
  token counts for the worst layer, replica distribution stats
- Pre-rearrange diagnostics: window utilization, load distribution
  across logical experts, top-5 hottest experts
- Post-rearrange diagnostics: number of changed slots, replica count
  stats, predicted post-rearrange balancedness (simulates expected
  balance with the new mapping applied to current load data)
- Warning when window_size > step_interval (stale data risk)

Signed-off-by: Travis Shears <travis@neuralmagic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>

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

@@ -51,5 +51,56 @@
             "max-model-len": 256,
             "async-scheduling": ""
         }
+    },
+    {
+        "test_name": "latency_deepseek_r1",
+        "environment_variables": {
+            "PT_HPU_LAZY_MODE": 1,
+            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
+            "VLLM_CONTIGUOUS_PA": 1,
+            "VLLM_DEFRAG": 1
+        },
+        "parameters": {
+            "model": "deepseek-ai/DeepSeek-R1",
+            "tensor_parallel_size": 8,
+            "load_format": "dummy",
+            "max-model-len": 2048,
+            "dtype": "bfloat16"
+        }
+    },
+    {
+        "test_name": "latency_llama4_maverick_17b128e_instruct_fp8",
+        "environment_variables": {
+            "PT_HPU_LAZY_MODE": 1,
+            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
+            "VLLM_CONTIGUOUS_PA": 1,
+            "VLLM_DEFRAG": 1
+        },
+        "parameters": {
+            "model": "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
+            "tensor_parallel_size": 8,
+            "max-model-len": 512,
+            "max-num-seqs": 128,
+            "async-scheduling": "",
+            "gpu-memory-utilization": 0.95,
+            "enable_expert_parallel": ""
+        }
+    },
+    {
+        "test_name": "latency_qwen3_8b",
+        "environment_variables": {
+            "PT_HPU_LAZY_MODE": 1,
+            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
+            "VLLM_CONTIGUOUS_PA": 1,
+            "VLLM_DEFRAG": 1
+        },
+        "parameters": {
+            "model": "Qwen/Qwen3-8B",
+            "tensor_parallel_size": 1,
+            "max-model-len": 2048,
+            "max-num-seqs": 128,
+            "dtype": "bfloat16",
+            "async-scheduling": ""
+        }
     }
 ]

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

@@ -78,5 +78,84 @@
             "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
             "num_prompts": 200
         }
+    },
+    {
+        "test_name": "serving_deepseek_r1",
+        "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": "deepseek-ai/DeepSeek-R1",
+            "tensor_parallel_size": 8,
+            "swap_space": 16,
+            "disable_log_stats": "",
+            "load_format": "dummy",
+            "max-model-len": 2048,
+            "max-num-seqs": 200,
+            "async-scheduling": "",
+            "dtype": "bfloat16"
+        },
+        "client_parameters": {
+            "model": "deepseek-ai/DeepSeek-R1",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_llama4_maverick_17b128e_instruct_fp8",
+        "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/Llama-4-Maverick-17B-128E-Instruct-FP8",
+            "tensor_parallel_size": 8,
+            "disable_log_stats": "",
+            "max-model-len": 2048,
+            "max-num-seqs": 128,
+            "async-scheduling": "",
+            "enable_expert_parallel": "",
+            "max-num-batched-tokens": 4096
+        },
+        "client_parameters": {
+            "model": "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
+    },
+    {
+        "test_name": "serving_qwen3_8b",
+        "qps_list": [1, 4, 10, "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": "Qwen/Qwen-3-8B",
+            "tensor_parallel_size": 1,
+            "dtype": "bfloat16",
+            "disable_log_stats": "",
+            "async-scheduling": ""
+        },
+        "client_parameters": {
+            "model": "Qwen/Qwen-3-8B",
+            "backend": "vllm",
+            "dataset_name": "sharegpt",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200
+        }
     }
 ]

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

@@ -57,5 +57,67 @@
             "max-num-seqs": 512,
             "async-scheduling": ""
         }
+    },
+    {
+        "test_name": "throughput_deepseek_r1",
+        "environment_variables": {
+            "PT_HPU_LAZY_MODE": 1,
+            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
+            "VLLM_CONTIGUOUS_PA": 1,
+            "VLLM_DEFRAG": 1
+        },
+        "parameters": {
+            "model": "deepseek-ai/DeepSeek-R1",
+            "tensor_parallel_size": 8,
+            "load_format": "dummy",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "dataset_name": "sharegpt",
+            "num_prompts": 1000,
+            "backend": "vllm",
+            "max-model-len": 2048,
+            "max-num-seqs": 384,
+            "async-scheduling": ""
+        }
+    },
+    {
+        "test_name": "throughput_llama4_maverick_17b128e_instruct_fp8",
+        "environment_variables": {
+            "PT_HPU_LAZY_MODE": 1,
+            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
+            "VLLM_CONTIGUOUS_PA": 1,
+            "VLLM_DEFRAG": 1
+        },
+        "parameters": {
+            "model": "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
+            "tensor_parallel_size": 8,
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "dataset_name": "sharegpt",
+            "num_prompts": 1000,
+            "backend": "vllm",
+            "max-model-len": 2048,
+            "max-num-seqs": 512,
+            "async-scheduling": "",
+            "enable_expert_parallel": ""
+        }
+    },
+    {
+        "test_name": "throughput_qwen3_8b",
+        "environment_variables": {
+            "PT_HPU_LAZY_MODE": 1,
+            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
+            "VLLM_CONTIGUOUS_PA": 1,
+            "VLLM_DEFRAG": 1
+        },
+        "parameters": {
+            "model": "Qwen/Qwen-3-8B",
+            "tensor_parallel_size": 1,
+            "load_format": "dummy",
+            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "dataset_name": "sharegpt",
+            "num_prompts": 1000,
+            "max-num-seqs": 512,
+            "backend": "vllm",
+            "async-scheduling": ""
+        }
     }
 ]

.buildkite/scripts/annotate-rocm-release.sh

@@ -68,7 +68,7 @@ aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/triton
 aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/torchvision-*.whl .
 aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/torchaudio-*.whl .
 aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/amdsmi-*.whl .
-aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/aiter-*.whl .
+aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/amd_aiter-*.whl .
 aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/flash-attn-*.whl .
 \`\`\`
 
@@ -80,7 +80,7 @@ aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/flash-
 - **torchvision**: TorchVision for ROCm PyTorch
 - **torchaudio**: Torchaudio for ROCm PyTorch
 - **amdsmi**: AMD SMI Python bindings
-- **aiter**: Aiter for ROCm
+- **amd_aiter**: Aiter for ROCm
 - **flash-attn**: Flash Attention for ROCm
 
 ### :warning: Notes

.buildkite/scripts/check-ray-compatibility.sh

@@ -0,0 +1,213 @@
+#!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Check if Ray LLM can generate lock files that are compatible with this
+# version of vllm. Downloads Ray's requirement files and runs a full
+# dependency resolution with the installed vllm's constraints to see if
+# a valid lock file can be produced.
+#
+# See: https://github.com/vllm-project/vllm/issues/33599
+
+set -eo pipefail
+
+RAY_BASE_URL="https://raw.githubusercontent.com/ray-project/ray/master/python"
+
+WORK_DIR=$(mktemp -d)
+trap 'rm -rf "$WORK_DIR"' EXIT
+
+# Fetch all Ray requirement files used in the LLM depset pipeline
+echo ">>> Fetching Ray requirement files"
+RAY_FILES=(
+    "requirements.txt"
+    "requirements/cloud-requirements.txt"
+    "requirements/base-test-requirements.txt"
+    "requirements/llm/llm-requirements.txt"
+    "requirements/llm/llm-test-requirements.txt"
+)
+for FILE in "${RAY_FILES[@]}"; do
+    LOCAL_PATH="${WORK_DIR}/$(basename "$FILE")"
+    echo "    ${FILE}"
+    curl -fsSL -o "$LOCAL_PATH" "${RAY_BASE_URL}/${FILE}"
+done
+
+# Extract installed vllm deps
+echo ">>> Extracting installed vllm dependency constraints"
+python3 - "${WORK_DIR}/vllm-constraints.txt" <<'PYEOF'
+"""Write out the installed vllm's dependencies as pip constraint lines.
+
+Ray uses vllm[audio], so audio-extra deps are included with their extra
+markers stripped. The resolver cannot evaluate extra markers for a
+package that is not itself being resolved from an index, so we activate
+them manually here.
+"""
+import importlib.metadata
+import re
+import sys
+
+out_path = sys.argv[1]
+raw_reqs = importlib.metadata.requires("vllm") or []
+
+# Ray uses vllm[audio] – activate that extra.
+ACTIVE_EXTRAS = {"audio"}
+EXTRA_RE = re.compile(r"""extra\s*==\s*['"]([^'"]+)['"]""")
+
+lines = []
+for r in raw_reqs:
+    if ";" not in r:
+        # Unconditional dep — always include.
+        lines.append(r.strip())
+        continue
+
+    req_part, _, marker_part = r.partition(";")
+    marker_part = marker_part.strip()
+
+    extra_matches = EXTRA_RE.findall(marker_part)
+    if not extra_matches:
+        # Non-extra marker (python_version, etc.) — keep as-is.
+        lines.append(r.strip())
+        continue
+
+    if not ACTIVE_EXTRAS.intersection(extra_matches):
+        continue  # Skip inactive extras (tensorizer, bench, …).
+
+    # Strip the extra== conditions but keep any remaining markers
+    # (e.g. python_version).
+    cleaned = EXTRA_RE.sub("", marker_part)
+    cleaned = re.sub(r"\band\b\s*\band\b", "and", cleaned)
+    cleaned = re.sub(r"^\s*and\s+|\s+and\s*$", "", cleaned).strip()
+
+    if cleaned:
+        lines.append(f"{req_part.strip()} ; {cleaned}")
+    else:
+        lines.append(req_part.strip())
+
+with open(out_path, "w") as f:
+    for line in lines:
+        f.write(line + "\n")
+
+print(f"Wrote {len(lines)} constraints to {out_path}")
+PYEOF
+
+echo ">>> Installed vllm deps (first 20 lines):"
+head -20 "${WORK_DIR}/vllm-constraints.txt"
+
+# Remove Ray's vllm pin — the installed vllm's transitive deps
+# (written above) replace it in the resolution. vllm itself cannot
+# be resolved from PyPI for in-development versions, so we test
+# whether Ray's requirements can coexist with vllm's dependency
+# constraints instead.
+sed -i '/^vllm/d' "${WORK_DIR}/llm-requirements.txt"
+
+# Install uv if needed
+if ! command -v uv &>/dev/null; then
+    echo ">>> Installing uv"
+    pip install uv -q
+fi
+
+# Resolve: given vllm's constraints, can Ray compile a lock file?
+#
+# vllm's dependency constraints are the fixed side — Ray is flexible and
+# can regenerate its lock files. We pass vllm's constraints via -c so
+# the resolver treats them as non-negotiable bounds, then check whether
+# Ray's own requirements can still be satisfied within those bounds.
+echo ""
+echo "============================================================"
+echo ">>> Resolving: Can Ray generate compatible lock files?"
+echo "============================================================"
+
+set +e
+uv pip compile \
+    "${WORK_DIR}/requirements.txt" \
+    "${WORK_DIR}/cloud-requirements.txt" \
+    "${WORK_DIR}/base-test-requirements.txt" \
+    "${WORK_DIR}/llm-requirements.txt" \
+    "${WORK_DIR}/llm-test-requirements.txt" \
+    -c "${WORK_DIR}/vllm-constraints.txt" \
+    --python-version 3.12 \
+    --python-platform x86_64-manylinux_2_31 \
+    --extra-index-url https://download.pytorch.org/whl/cu129 \
+    --index-strategy unsafe-best-match \
+    --unsafe-package setuptools \
+    --unsafe-package ray \
+    --no-header \
+    -o "${WORK_DIR}/resolved.txt" \
+    2>&1
+EXIT_CODE=$?
+set -e
+
+echo ""
+echo "=========================================="
+if [ $EXIT_CODE -eq 0 ]; then
+    echo "SUCCESS: Ray can generate lock files compatible with this vllm."
+    echo ""
+    echo "Key resolved versions:"
+    grep -E '^(protobuf|torch|numpy|transformers)==' \
+        "${WORK_DIR}/resolved.txt" | sort || true
+    echo "=========================================="
+    exit 0
+fi
+
+echo "FAILURE: Ray cannot generate lock files compatible with this vllm."
+echo "This means a fundamental dependency conflict exists that Ray"
+echo "cannot resolve by regenerating its lock files."
+echo "See: https://github.com/vllm-project/vllm/issues/33599"
+echo "=========================================="
+
+# Buildkite annotation
+if [ -f /usr/bin/buildkite-agent ]; then
+    buildkite-agent annotate --style 'warning' --context 'ray-compat' << EOF
+### :warning: Ray Dependency Compatibility Warning
+This PR introduces dependencies that **cannot** be resolved with Ray's requirements.
+Ray would not be able to regenerate its lock files to accommodate this vllm version.
+
+Please check the **Ray Dependency Compatibility Check** step logs for details.
+See [issue #33599](https://github.com/vllm-project/vllm/issues/33599) for context.
+EOF
+fi
+
+# Notify Slack if webhook is configured and PR/branch are valid.
+if [ -n "$RAY_COMPAT_SLACK_WEBHOOK_URL" ]; then
+    PR="${BUILDKITE_PULL_REQUEST:-}"
+    BRANCH="${BUILDKITE_BRANCH:-}"
+
+    # Skip notification if PR is invalid or branch is empty
+    if [[ "$PR" = "false" || -z "$PR" || -z "$BRANCH" ]]; then
+        echo ">>> Skipping Slack notification (invalid PR or empty branch: PR=$PR, branch=$BRANCH)"
+    else
+        echo ">>> Sending Slack notification"
+        # Single quotes are intentional: the f-string expressions are Python, not shell.
+        # shellcheck disable=SC2016
+        PAYLOAD=$(python3 -c '
+import json, os, sys
+pr = os.getenv("BUILDKITE_PULL_REQUEST", "N/A")
+branch = os.getenv("BUILDKITE_BRANCH", "unknown")
+url = os.getenv("BUILDKITE_BUILD_URL", "#")
+data = {
+    "text": ":warning: Ray Dependency Compatibility Check Failed",
+    "blocks": [{
+        "type": "section",
+        "text": {
+            "type": "mrkdwn",
+            "text": (
+                "*:warning: Ray Dependency Compatibility Check Failed*\n"
+                f"PR #{pr} on branch `{branch}` introduces dependencies "
+                f"that cannot be resolved with Ray'\''s requirements.\n"
+                f"<{url}|View Build>"
+            ),
+        },
+    }],
+}
+print(json.dumps(data))
+')
+
+        HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" -X POST "$RAY_COMPAT_SLACK_WEBHOOK_URL" \
+            -H 'Content-type: application/json' \
+            -d "$PAYLOAD")
+        echo "    Slack webhook response: $HTTP_CODE"
+    fi
+else
+    echo ">>> Skipping Slack notification (RAY_COMPAT_SLACK_WEBHOOK_URL not set)"
+fi
+
+exit 1

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

@@ -99,6 +99,15 @@ is_multi_node() {
   return 1
 }
 
+handle_pytest_exit() {
+  local exit_code=$1
+  if [ "$exit_code" -eq 5 ]; then
+    echo "Pytest exit code 5 (no tests collected) - treating as success."
+    exit 0
+  fi
+  exit "$exit_code"
+}
+
 ###############################################################################
 # Pytest marker/keyword re-quoting
 #
@@ -135,8 +144,9 @@ re_quote_pytest_markers() {
   local collecting=false
   local marker_buf=""
 
-  # Flatten newlines for consistent tokenization
-  local flat="${input//$'\n'/ }"
+  # Strip backslash-newline continuations, then flatten remaining newlines
+  local flat="${input//$'\\\n'/ }"
+  flat="${flat//$'\n'/ }"
 
   # Disable globbing to prevent *.py etc. from expanding during read -ra
   local restore_glob
@@ -164,6 +174,9 @@ re_quote_pytest_markers() {
 
       local is_boundary=false
       case "$word" in
+        # Line-continuation artifact
+        "\\")
+          is_boundary=true ;;
         # Command separators
         "&&"|"||"|";"|"|")
           is_boundary=true ;;
@@ -204,6 +217,9 @@ re_quote_pytest_markers() {
         if [[ "$word" == "-m" || "$word" == "-k" ]]; then
           output+="${word} "
           collecting=true
+        # Drop stray backslash tokens silently
+        elif [[ "$word" == "\\" ]]; then
+          :
         else
           output+="${word} "
         fi
@@ -453,7 +469,9 @@ if is_multi_node "$commands"; then
     done
 
     /bin/bash -c "${composite_command}"
+    exit_code=$?
     cleanup_network
+    handle_pytest_exit "$exit_code"
   else
     echo "Multi-node job detected but failed to parse bracket command syntax."
     echo "Expected format: prefix ; [node0_cmd1, node0_cmd2] && [node1_cmd1, node1_cmd2]"
@@ -480,4 +498,7 @@ else
     --name "${container_name}" \
     "${image_name}" \
     /bin/bash -c "${commands}"
+
+  exit_code=$?
+  handle_pytest_exit "$exit_code"
 fi

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

@@ -1,26 +1,43 @@
 #!/bin/bash
 set -euox pipefail
+export VLLM_CPU_CI_ENV=0
 
 echo "--- PP+TP"
 vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -pp=2 &
 server_pid=$!
-timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
+timeout 600 bash -c "until curl localhost:8000/v1/models > /dev/null 2>&1; do sleep 1; done" || exit 1
 vllm bench serve \
     --backend vllm \
     --dataset-name random \
     --model meta-llama/Llama-3.2-3B-Instruct \
     --num-prompts 20 \
+    --result-dir ./test_results \
+    --result-filename tp_pp.json \
+    --save-result \
     --endpoint /v1/completions
-kill -s SIGTERM $server_pid &
+kill -s SIGTERM $server_pid; wait $server_pid || true
+failed_req=$(jq '.failed' ./test_results/tp_pp.json)
+if [ "$failed_req" -ne 0 ]; then
+  echo "Some requests were failed!"
+  exit 1
+fi
 
 echo "--- DP+TP"
 vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -dp=2 &
 server_pid=$!
-timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
+timeout 600 bash -c "until curl localhost:8000/v1/models > /dev/null 2>&1; do sleep 1; done" || exit 1
 vllm bench serve \
     --backend vllm \
     --dataset-name random \
     --model meta-llama/Llama-3.2-3B-Instruct \
     --num-prompts 20 \
+    --result-dir ./test_results \
+    --result-filename dp_pp.json \
+    --save-result \
     --endpoint /v1/completions
-kill -s SIGTERM $server_pid &
+kill -s SIGTERM $server_pid; wait $server_pid || true
+failed_req=$(jq '.failed' ./test_results/dp_pp.json)
+if [ "$failed_req" -ne 0 ]; then
+  echo "Some requests were failed!"
+  exit 1
+fi

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

@@ -1,9 +1,27 @@
 #!/bin/bash
 
-# This script build the CPU docker image and run the offline inference inside the container.
+# This script builds the HPU docker image and runs the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
+#
+# vllm-gaudi compatibility pinning:
+#   The vllm-gaudi plugin is installed on top of the vllm upstream checkout used by this CI job.
+#   When upstream vllm changes its API, the plugin may break before it has been updated.
+#   To handle this, the vllm-gaudi repository maintains a file:
+#     vllm/last-good-commit-for-vllm-gaudi/VLLM_COMMUNITY_COMMIT
+#   The first line of that file controls what version of vllm is used inside the Docker image:
+#     - "latest"        : no checkout override; the current Buildkite CI commit is used as-is.
+#     - "<commit SHA>"  : vllm is checked out to that specific commit before building, pinning
+#                         the test to a known-compatible baseline.
+#   To unpin (resume testing against the live vllm tip), set the file content back to "latest".
 set -exuo pipefail
 
+# Fetch the vllm community commit reference from vllm-gaudi (first line only).
+VLLM_COMMUNITY_COMMIT=$(curl -s \
+  https://raw.githubusercontent.com/vllm-project/vllm-gaudi/vllm/last-good-commit-for-vllm-gaudi/VLLM_COMMUNITY_COMMIT \
+  | head -1 | tr -d '\n')
+
+echo "Using vllm community commit: ${VLLM_COMMUNITY_COMMIT}"
+
 # Try building the docker image
 image_name="hpu/upstream-vllm-ci:${BUILDKITE_COMMIT}"
 container_name="hpu-upstream-vllm-ci-${BUILDKITE_COMMIT}-container"
@@ -12,6 +30,13 @@ FROM gaudi-base-image:latest
 
 COPY ./ /workspace/vllm
 
+# If VLLM_COMMUNITY_COMMIT is a specific commit (not "latest"), check it out to pin vllm
+# to the version known to be compatible with vllm-gaudi. When the value is "latest",
+# the current checkout (the Buildkite CI commit) is used unchanged.
+RUN if [ "${VLLM_COMMUNITY_COMMIT}" != "latest" ]; then \
+      cd /workspace/vllm && git fetch --unshallow 2>/dev/null || true && git checkout ${VLLM_COMMUNITY_COMMIT}; \
+    fi
+
 WORKDIR /workspace/vllm
 
 ENV no_proxy=localhost,127.0.0.1

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

@@ -44,7 +44,7 @@ docker run \
     python3 examples/offline_inference/basic/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2
     python3 examples/offline_inference/basic/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
     cd tests
-    pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py
+    pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py
     pytest -v -s v1/engine
     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

.buildkite/scripts/upload-nightly-wheels.sh

@@ -72,7 +72,7 @@ obj_json="objects.json"
 aws s3api list-objects-v2 --bucket "$BUCKET" --prefix "$SUBPATH/" --delimiter / --output json > "$obj_json"
 mkdir -p "$INDICES_OUTPUT_DIR"
 
-# call script to generate indicies for all existing wheels
+# call script to generate indices for all existing wheels
 # this indices have relative paths that could work as long as it is next to the wheel directory in s3
 # i.e., the wheels are always in s3://vllm-wheels/<commit>/
 # and indices can be placed in /<commit>/, or /nightly/, or /<version>/

.buildkite/scripts/upload-release-wheels-pypi.sh

@@ -54,10 +54,13 @@ mkdir -p $DIST_DIR
 # include only wheels for the release version, ignore all files with "dev" or "rc" in the name (without excluding 'aarch64')
 aws s3 cp --recursive --exclude "*" --include "vllm-${PURE_VERSION}*.whl" --exclude "*dev*" --exclude "*rc[0-9]*" "$S3_COMMIT_PREFIX" $DIST_DIR
 echo "Wheels copied to local directory"
-# generate source tarball
-git archive --format=tar.gz --output="$DIST_DIR/vllm-${PURE_VERSION}.tar.gz" "$BUILDKITE_COMMIT"
+# generate source distribution using setup.py
+python setup.py sdist --dist-dir=$DIST_DIR
 ls -la $DIST_DIR
 
+SDIST_FILE=$(find $DIST_DIR -name "vllm*.tar.gz")
+echo "Found sdist: $SDIST_FILE"
+
 # upload wheels to PyPI (only default variant, i.e. files without '+' in the name)
 PYPI_WHEEL_FILES=$(find $DIST_DIR -name "vllm-${PURE_VERSION}*.whl" -not -name "*+*")
 if [[ -z "$PYPI_WHEEL_FILES" ]]; then
@@ -65,6 +68,6 @@ if [[ -z "$PYPI_WHEEL_FILES" ]]; then
   exit 1
 fi
 
-python3 -m twine check "$PYPI_WHEEL_FILES"
-python3 -m twine upload --non-interactive --verbose "$PYPI_WHEEL_FILES"
-echo "Wheels uploaded to PyPI"
+python3 -m twine check "$PYPI_WHEEL_FILES" "$SDIST_FILE"
+python3 -m twine upload --non-interactive --verbose "$PYPI_WHEEL_FILES" "$SDIST_FILE"
+echo "Wheels and source distribution uploaded to PyPI"

.buildkite/test-amd.yaml

@@ -388,9 +388,7 @@ steps:
 - label: V1 Test e2e + engine # 65min
   timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental, amdproduction]
-  # The test uses 4 GPUs, but we schedule it on 8-GPU machines for stability.
-  # See discussion here: https://github.com/vllm-project/vllm/pull/31040
-  agent_pool: mi325_8
+  agent_pool: mi325_1
   optional: true
   # grade: Blocking
   source_file_dependencies:
@@ -402,6 +400,34 @@ steps:
     - pytest -v -s v1/e2e
     - pytest -v -s v1/engine
 
+- label: V1 Test e2e (2 GPUs) # 65min
+  timeout_in_minutes: 90
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_2
+  optional: true
+  # grade: Blocking
+  source_file_dependencies:
+    - vllm/
+    - tests/v1
+  commands:
+    # Only run tests that need exactly 2 GPUs
+    - pytest -v -s v1/e2e/test_spec_decode.py -k "tensor_parallelism"
+
+- label: V1 Test e2e (4 GPUs) # 65min
+  timeout_in_minutes: 90
+  mirror_hardwares: [amdexperimental, amdproduction]
+  # The test uses 4 GPUs, but we schedule it on 8-GPU machines for stability.
+  # See discussion here: https://github.com/vllm-project/vllm/pull/31040
+  agent_pool: mi325_4
+  optional: true
+  # grade: Blocking
+  source_file_dependencies:
+    - vllm/
+    - tests/v1
+  commands:
+    # Only run tests that need 4 GPUs
+    - pytest -v -s v1/e2e/test_spec_decode.py -k "eagle_correctness_heavy"
+
 - label: V1 Test entrypoints # 35min
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
@@ -441,7 +467,7 @@ steps:
     - 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
+# TODO: Add the "V1 Test attention (MI300)" test group
 
 - label: V1 Test attention (H100) # 10min
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -473,17 +499,6 @@ steps:
     - pytest -v -s v1/determinism/test_batch_invariance.py
     - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
 
-- label: V1 Test attention (B200) # 10min
-  timeout_in_minutes: 30
-  gpu: b200
-  source_file_dependencies:
-    - vllm/config/attention.py
-    - vllm/model_executor/layers/attention
-    - vllm/v1/attention
-    - tests/v1/attention
-  commands:
-    - pytest -v -s v1/attention
-
 - label: V1 Test others (CPU) # 5 mins
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi325_1
@@ -584,6 +599,8 @@ steps:
       --ignore=lora/test_qwen3moe_tp.py
   parallelism: 4
 
+##### .buildkite/test_areas/pytorch.yaml #####
+# corresponds to .buildkite/test_areas/pytorch.yaml
 - label: PyTorch Compilation Unit Tests # 15min
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -601,6 +618,20 @@ steps:
   # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
   - "find compile/ -maxdepth 1 -name 'test_*.py' -exec pytest -s -v {} \\\\;"
 
+# corresponds to .buildkite/test_areas/pytorch.yaml
+- label: PyTorch Compilation Passes Unit Tests
+  timeout_in_minutes: 20
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_1
+  source_file_dependencies:
+    - vllm/
+    - tests/compile/passes
+  commands:
+  # TODO: clean up this comment if not needed. It is used to 
+  # keep track of the tests changes during vLLM IR Ops refactoring.
+  # Use `find` to launch multiple instances of pytest.
+  - "find compile/passes -maxdepth 1 -name 'test_*.py' -exec pytest -s -v {} \\\\;"
+
 - label: PyTorch Fullgraph Smoke Test # 15min
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -1018,6 +1049,7 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
+  - tests/models/registry.py
   no_gpu: true
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
@@ -1031,6 +1063,7 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
+  - tests/models/registry.py
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/processing
@@ -1141,48 +1174,6 @@ steps:
     # Whisper needs spawn method to avoid deadlock
     - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
 
-- label: Blackwell Test # 21 min
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/"
-  gpu: b200
-  # optional: true
-  source_file_dependencies:
-  - csrc/quantization/fp4/
-  - csrc/attention/mla/
-  - csrc/quantization/cutlass_w8a8/moe/
-  - vllm/model_executor/layers/fused_moe/cutlass_moe.py
-  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py
-  - vllm/model_executor/layers/fused_moe/flashinfer_a2a_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/v1/attention/selector.py
-  - vllm/platforms/cuda.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
-    - pytest -v -s tests/kernels/attention/test_flashinfer_mla_decode.py
-    # Quantization
-    - pytest -v -s tests/kernels/quantization/test_cutlass_scaled_mm.py -k 'fp8'
-    - pytest -v -s tests/kernels/quantization/test_nvfp4_quant.py
-    - pytest -v -s tests/kernels/quantization/test_silu_mul_nvfp4_quant.py
-    - pytest -v -s tests/kernels/quantization/test_nvfp4_scaled_mm.py
-    - pytest -v -s tests/kernels/quantization/test_flashinfer_scaled_mm.py
-    - pytest -v -s tests/kernels/quantization/test_flashinfer_nvfp4_scaled_mm.py
-    - pytest -v -s tests/kernels/quantization/test_nvfp4_qutlass.py
-    - pytest -v -s tests/kernels/quantization/test_mxfp4_qutlass.py
-    - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
-    - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
-    - pytest -v -s tests/kernels/moe/test_flashinfer.py
-    - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
-
 - label: Blackwell Fusion and Compile Tests # 30 min
   timeout_in_minutes: 40
   working_dir: "/vllm-workspace/"
@@ -1249,16 +1240,6 @@ steps:
   commands:
     - pytest -s -v tests/quantization/test_blackwell_moe.py
 
-- label: Blackwell LM Eval Small Models
-  timeout_in_minutes: 120
-  gpu: b200
-  optional: true # run on nightlies
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt
-
 #####  1 GPU test  #####
 #####  multi gpus test  #####
 
@@ -1330,6 +1311,7 @@ steps:
   - tests/v1/entrypoints/openai/test_multi_api_servers.py
   - tests/v1/shutdown
   - tests/v1/worker/test_worker_memory_snapshot.py
+  - examples/offline_inference/new_weight_syncing/
   commands:
   # Work around HIP bug tracked here: https://github.com/ROCm/hip/issues/3876
   # TODO: Remove when the bug is fixed in a future ROCm release
@@ -1343,7 +1325,6 @@ steps:
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
   - VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
-  - pytest -v -s compile/correctness_e2e/test_sequence_parallel.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
   - pytest -v -s v1/worker/test_worker_memory_snapshot.py
 
@@ -1544,8 +1525,8 @@ 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
+##### FP8 test #####
+- label: LM Eval Large Models (H100) # optional, still use H100 for consistency
   gpu: h100
   optional: true
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -1557,8 +1538,8 @@ steps:
   - 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
+    - export VLLM_USE_DEEP_GEMM=0 
+    - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm.txt --tp-size=4
 
 
 ##### H200 test #####
@@ -1573,16 +1554,16 @@ steps:
   commands:
     - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/passes/distributed/test_async_tp.py
     - pytest -v -s tests/compile/passes/distributed/test_sequence_parallelism.py
-    - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
+    # TODO: this test is not supported on ROCm, there are aiter kernels for this.
+    # - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
     #- pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
     # - "VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'"
     # Old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293
     # in favor of new tests in fusions_e2e. We avoid replicating the new jobs in this file as it's deprecated.
-
-    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/correctness_e2e/test_sequence_parallel.py
     - pytest -v -s tests/distributed/test_context_parallel.py
     - HIP_VISIBLE_DEVICES=0,1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=allgather_reducescatter --disable-nccl-for-dp-synchronization
-    - pytest -v -s tests/v1/distributed/test_dbo.py
+    # this test is not supported on ROCm
+    # - pytest -v -s tests/v1/distributed/test_dbo.py
 
 ##### B200 test #####
 - label: Distributed Tests (B200) # optional
@@ -1672,16 +1653,6 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020
 
-- label: Qwen3-30B-A3B-FP8-block Accuracy (B200)
-  timeout_in_minutes: 60
-  gpu: b200
-  optional: true
-  num_gpus: 2
-  working_dir: "/vllm-workspace"
-  commands:
-  - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
-
-
 - label: Qwen3-Next-80B-A3B-Instruct MTP Async EPLB Accuracy
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -1693,6 +1664,93 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh 0.8 1319 8040
 
+##### .buildkite/test_areas/compile.yaml #####
+# Slowly setting up the tests so that it is also easier for the 
+# CI team to review and upstream to the pipelinev2.
+# The following tests are important for vLLM IR Ops refactoring,
+# which affects fusion passes on ROCm. So we have to 
+# enable them as as soon as possible.
+
+## TODO: Enable the test in this group
+# # corresponds to .buildkite/test_areas/compile.yaml
+# - label: Fusion and Compile Unit Tests (2xMI325 GPUs)
+#   timeout_in_minutes: 20
+#   working_dir: "/vllm-workspace/"
+#   mirror_hardwares: [amdexperimental, amdproduction, tj]
+#   agent_pool: mi325_1 # changed to 1 GPU until the fusion all reduce is enabled then only revert back to 2 GPUs
+#   source_file_dependencies:
+#   - csrc/quantization/fp4/
+#   - vllm/model_executor/layers/quantization/
+#   - vllm/model_executor/layers/layernorm.py
+#   - vllm/model_executor/layers/activation.py
+#   - vllm/model_executor/layers/attention/attention.py
+#   - vllm/v1/attention/backends/flashinfer.py
+#   - vllm/compilation/ # TODO(luka) limit to vllm/compilation/passes
+#   - tests/compile/test_fusion_attn.py
+#   - tests/compile/test_silu_mul_quant_fusion.py
+#   - tests/compile/distributed/test_fusion_all_reduce.py
+#   - tests/compile/fullgraph/test_full_graph.py
+#   commands:
+#     - rocm-smi
+#     # we run all backend tests on ROCm
+#     # These two tests are covered in "PyTorch Compilation Passes Unit Tests"
+#     # - "pytest -v -s tests/compile/passes/test_fusion_attn.py"
+#     # - "pytest -v -s tests/compile/passes/test_silu_mul_quant_fusion.py"
+#     # TODO: this test is not supported on ROCm, there are aiter kernels for this.
+#     # - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
+#     # TODO: find out more details
+#     # - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
+
+# corresponds to .buildkite/test_areas/compile.yaml
+- label: Fusion E2E Quick (MI325)
+  timeout_in_minutes: 15
+  working_dir: "/vllm-workspace/"
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_1
+  num_devices: 1
+  source_file_dependencies:
+    - csrc/quantization/
+    - vllm/model_executor/
+    - vllm/v1/attention/
+    - vllm/compilation/
+    - tests/compile/fusions_e2e/
+  commands:
+    - rocm-smi
+    # Run all models and attn backends but only Inductor partition and native custom ops
+    - "pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k 'inductor_partition and not +rms_norm and not +quant_fp8'"
+    # Different from CUDA, Qwen requires +rms_norm and +quant_fp8 as rms+quant fusion is only supported on AITER
+    - "pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k 'inductor_partition and +rms_norm and +quant_fp8 and qwen3'"
+
+# corresponds to .buildkite/test_areas/compile.yaml
+- label: Fusion E2E Config Sweep (MI325)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/"
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_1
+  num_devices: 1
+  source_file_dependencies:
+    - csrc/quantization/
+    - vllm/compilation/
+    # can affect pattern matching
+    - vllm/model_executor/layers/layernorm.py
+    - vllm/model_executor/layers/activation.py
+    - vllm/model_executor/layers/attention/attention.py
+    - vllm/model_executor/layers/quantization/input_quant_fp8.py
+    - tests/compile/fusions_e2e/
+  commands:
+    - rocm-smi
+    # Run just llama3 (fp8) for all config combinations
+    - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "llama-3"
+
+## There are no ops on ROCm for these tests.
+## The test still passes but the logs are not useful.
+## fused ops just call torch.ops.symm_mem which 
+## exists in ROCm even though they don't work
+# - label: AsyncTP Correctness Tests  (2xMI325 GPUs)
+# - label: Fusion E2E TP2 Quick (MI325)
+# - label: Fusion E2E TP2 AsyncTP Config Sweep (MI325)
+# - label: Fusion E2E TP2 (MI325)
+# - label: Sequence Parallel Correctness Tests (2xMI325 GPUs)
 
 
 #####################################################################################################################################
@@ -1875,8 +1933,10 @@ steps:
 
 - label: Distributed Tests (4 GPUs) # 35min
   timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
+  optional: true
+  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -1930,7 +1990,8 @@ steps:
   - popd
   # NEW rlhf examples
   - pushd ../examples/offline_inference/new_weight_syncing
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_nccl.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_ipc.py
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_async_new_apis.py
   - popd
 
@@ -2075,20 +2136,7 @@ steps:
     - 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: mi355_1
-  timeout_in_minutes: 30
-  gpu: h100
-  source_file_dependencies:
-    - vllm/config/attention.py
-    - vllm/model_executor/layers/attention
-    - vllm/v1/attention
-    - tests/v1/attention
-  commands:
-    - pytest -v -s v1/attention
+# TODO: Add the "V1 Test attention (MI300)" test group
 
 - label: Batch Invariance Tests (H100) # 10min
   mirror_hardwares: [amdexperimental]
@@ -2106,6 +2154,8 @@ steps:
     - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
 
 - label: V1 Test attention (B200) # 10min
+  mirror_hardwares: [amdexperimental, amdmi355]
+  agent_pool: mi355_1
   timeout_in_minutes: 30
   gpu: b200
   source_file_dependencies:
@@ -2730,7 +2780,9 @@ steps:
     # Whisper needs spawn method to avoid deadlock
     - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
 
-- label: Blackwell Test # 21 min
+- label: Blackwell Test (MI355) # 21 min
+  mirror_hardwares: [amdexperimental, amdmi355]
+  agent_pool: mi355_1
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
   gpu: b200
@@ -2749,28 +2801,28 @@ steps:
   - vllm/v1/attention/selector.py
   - vllm/platforms/cuda.py
   commands:
-    - nvidia-smi
+    rocm-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
-    - pytest -v -s tests/kernels/attention/test_flashinfer_mla_decode.py
-    # Quantization
-    - pytest -v -s tests/kernels/quantization/test_cutlass_scaled_mm.py -k 'fp8'
-    - pytest -v -s tests/kernels/quantization/test_nvfp4_quant.py
-    - pytest -v -s tests/kernels/quantization/test_silu_mul_nvfp4_quant.py
-    - pytest -v -s tests/kernels/quantization/test_nvfp4_scaled_mm.py
-    - pytest -v -s tests/kernels/quantization/test_flashinfer_scaled_mm.py
-    - pytest -v -s tests/kernels/quantization/test_flashinfer_nvfp4_scaled_mm.py
-    - pytest -v -s tests/kernels/quantization/test_nvfp4_qutlass.py
-    - pytest -v -s tests/kernels/quantization/test_mxfp4_qutlass.py
-    - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
-    - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
-    - pytest -v -s tests/kernels/moe/test_flashinfer.py
-    - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
+    - 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
+    #- pytest -v -s tests/kernels/attention/test_flashinfer_mla_decode.py
+    ## Quantization
+    #- pytest -v -s tests/kernels/quantization/test_cutlass_scaled_mm.py -k 'fp8'
+    #- pytest -v -s tests/kernels/quantization/test_nvfp4_quant.py
+    #- pytest -v -s tests/kernels/quantization/test_silu_mul_nvfp4_quant.py
+    #- pytest -v -s tests/kernels/quantization/test_nvfp4_scaled_mm.py
+    #- pytest -v -s tests/kernels/quantization/test_flashinfer_scaled_mm.py
+    #- pytest -v -s tests/kernels/quantization/test_flashinfer_nvfp4_scaled_mm.py
+    #- pytest -v -s tests/kernels/quantization/test_nvfp4_qutlass.py
+    #- pytest -v -s tests/kernels/quantization/test_mxfp4_qutlass.py
+    #- pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
+    #- pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
+    #- pytest -v -s tests/kernels/moe/test_flashinfer.py
+    #- pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
 
 - label: Blackwell Fusion and Compile Tests # 30 min
   timeout_in_minutes: 40
@@ -2840,13 +2892,15 @@ steps:
 
 - label: Blackwell LM Eval Small Models
   timeout_in_minutes: 120
+  mirror_hardwares: [amdexperimental, amdproduction, amdmi355]
+  agent_pool: mi355_2
   gpu: b200
   optional: true # run on nightlies
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
   commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi355.txt
 
 #####  1 GPU test  #####
 #####  multi gpus test  #####
@@ -2894,8 +2948,10 @@ steps:
 
 - label: Distributed Tests (2 GPUs) # 68min
   timeout_in_minutes: 90
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_2
+  optional: true
+  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -3227,18 +3283,9 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
 
-- label: Qwen3-30B-A3B-FP8-block Accuracy (H100)
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi355_4
-  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_eplb.sh 0.8 200 8020
-
-- label: Qwen3-30B-A3B-FP8-block Accuracy (B200)
+- label: Qwen3-30B-A3B-FP8-block Accuracy (B200/MI355)
+  mirror_hardwares: [amdexperimental, amdproduction, amdmi355]
+  agent_pool: mi355_2
   timeout_in_minutes: 60
   gpu: b200
   optional: true
@@ -3257,3 +3304,18 @@ steps:
   working_dir: "/vllm-workspace"
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh 0.8 1319 8040
+
+- label: Attention Benchmarks Smoke Test (B200/MI355)
+  device: b200
+  mirror_hardwares: [amdexperimental, amdmi355]
+  agent_pool: mi355_2
+  num_gpus: 2
+  optional: true
+  working_dir: "/vllm-workspace/"
+  timeout_in_minutes: 10
+  source_file_dependencies:
+  - benchmarks/attention_benchmarks/
+  - vllm/v1/attention/
+  commands:
+  - python3 benchmarks/attention_benchmarks/benchmark.py --backends ROCM_ATTN ROCM_AITER_FA ROCM_AITER_UNIFIED_ATTN --batch-specs "8q1s1k" --repeats 1 --warmup-iters 1
+

.buildkite/test_areas/compile.yaml

@@ -36,6 +36,16 @@ steps:
   - export VLLM_TEST_CLEAN_GPU_MEMORY=1
   - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py
 
+- label: AsyncTP Correctness Tests (B200)
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/"
+  device: b200
+  optional: true
+  num_devices: 2
+  commands:
+  - export VLLM_TEST_CLEAN_GPU_MEMORY=1
+  - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py
+
 - label: Distributed Compile Unit Tests (2xH100)
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/"

.buildkite/test_areas/distributed.yaml

@@ -67,6 +67,7 @@ steps:
   - tests/v1/distributed
   - tests/v1/engine/test_engine_core_client.py
   - tests/distributed/test_symm_mem_allreduce.py
+  - tests/distributed/test_multiproc_executor.py
   commands:
   # https://github.com/NVIDIA/nccl/issues/1838
   - export NCCL_CUMEM_HOST_ENABLE=0
@@ -95,6 +96,8 @@ steps:
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s distributed/test_events.py
   - pytest -v -s distributed/test_symm_mem_allreduce.py
+  # test multi-node TP with multiproc executor (simulated on single node)
+  - pytest -v -s distributed/test_multiproc_executor.py::test_multiproc_executor_multi_node
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
   # OLD rlhf examples
@@ -146,7 +149,7 @@ steps:
   num_devices: 2
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
-    - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py
+    # - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py --- failing, need to re-enable
     - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
     - pytest -v -s tests/v1/distributed/test_dbo.py
 
@@ -210,6 +213,19 @@ steps:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
     - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
+- label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
+  timeout_in_minutes: 30
+  device: a100
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - vllm/v1/worker/kv_connector_model_runner_mixin.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
+
 - label: Pipeline + Context Parallelism (4 GPUs)
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"

.buildkite/test_areas/engine.yaml

@@ -14,7 +14,7 @@ steps:
   commands:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
 
-- label: V1 e2e + engine
+- label: V1 e2e + engine (1 GPU)
   timeout_in_minutes: 45
   source_file_dependencies:
     - vllm/
@@ -36,3 +36,35 @@ steps:
       commands:
       - pytest -v -s v1/e2e
       - pytest -v -s v1/engine
+
+- label: V1 e2e (2 GPUs)
+  timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
+  optional: true
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/
+    - tests/v1/e2e
+  commands:
+    # Only run tests that need exactly 2 GPUs
+    - pytest -v -s v1/e2e/test_spec_decode.py -k "tensor_parallelism"
+  mirror:
+    amd:
+      device: mi325_2
+      depends_on:
+      - image-build-amd
+
+- label: V1 e2e (4 GPUs)
+  timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
+  optional: true
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/
+    - tests/v1/e2e
+  commands:
+    # Only run tests that need 4 GPUs
+    - pytest -v -s v1/e2e/test_spec_decode.py -k "eagle_correctness_heavy"
+  mirror:
+    amd:
+      device: mi325_4
+      depends_on:
+      - image-build-amd

.buildkite/test_areas/entrypoints.yaml

@@ -41,6 +41,11 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/  --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses
   - pytest -v -s entrypoints/test_chat_utils.py
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Entrypoints Integration (API Server 2)
   timeout_in_minutes: 130
@@ -55,6 +60,11 @@ steps:
   - pytest -v -s entrypoints/instrumentator
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
   - pytest -v -s tool_use
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Entrypoints Integration (Pooling)
   timeout_in_minutes: 50
@@ -87,6 +97,11 @@ steps:
     - tests/v1
   commands:
     - pytest -v -s v1/entrypoints
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: OpenAI API Correctness
   timeout_in_minutes: 30

.buildkite/test_areas/kernels.yaml

@@ -44,7 +44,8 @@ steps:
   - vllm/envs.py
   - vllm/config
   commands:
-    - pytest -v -s kernels/moe --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+    - pytest -v -s kernels/moe --ignore=kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+    - pytest -v -s kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 2
 
 - label: Kernels Mamba Test

.buildkite/test_areas/lm_eval.yaml

@@ -11,17 +11,17 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
 
-- label: LM Eval Large Models (4 GPUs)(A100)
-  device: a100
-  optional: true
-  num_devices: 4
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
+# - label: LM Eval Large Models (4 GPUs)(A100)
+#   device: a100
+#   optional: true
+#   num_devices: 4
+#   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+#   source_file_dependencies:
+#   - csrc/
+#   - vllm/model_executor/layers/quantization
+#   commands:
+#   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+#   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
 - label: LM Eval Large Models (4 GPUs)(H100)
   device: h100

.buildkite/test_areas/misc.yaml

@@ -87,6 +87,11 @@ steps:
     - python3 offline_inference/spec_decode.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
     # https://github.com/vllm-project/vllm/pull/26682 uses slightly more memory in PyTorch 2.9+ causing this test to OOM in 1xL4 GPU
     - python3 offline_inference/spec_decode.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Metrics, Tracing (2 GPUs)
   timeout_in_minutes: 20

.buildkite/test_areas/models_multimodal.yaml

@@ -12,6 +12,11 @@ steps:
     - pip freeze | grep -E 'torch'
     - 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
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Multi-Modal Processor Test (CPU)
   depends_on: 
@@ -20,6 +25,7 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
+  - tests/models/registry.py
   device: cpu
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
@@ -30,6 +36,7 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
+  - tests/models/registry.py
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/processing/test_tensor_schema.py
@@ -52,6 +59,11 @@ steps:
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal -m 'not core_model' --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/processing
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Multi-Modal Models (Extended) 2
   optional: true
@@ -70,12 +82,3 @@ steps:
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
-
-# This test is used only in PR development phase to test individual models and should never run on main
-- label: Custom Models
-  optional: true
-  commands:
-    - echo 'Testing custom models...'
-    # PR authors can temporarily add commands below to test individual models
-    # e.g. pytest -v -s models/encoder_decoder/vision_language/test_mllama.py
-    # *To avoid merge conflicts, remember to REMOVE (not just comment out) them before merging the PR*

.buildkite/test_areas/plugins.yaml

@@ -15,9 +15,12 @@ steps:
   - pytest -v -s plugins_tests/test_platform_plugins.py
   - pip uninstall vllm_add_dummy_platform -y
   # end platform plugin tests
-  # begin io_processor plugins test, all the code in between uses the prithvi_io_processor plugin
+  # begin io_processor plugins test
+  # test generic io_processor plugins functions
+  - pytest -v -s ./plugins_tests/test_io_processor_plugins.py
+  # test Terratorch io_processor plugins
   - pip install -e ./plugins/prithvi_io_processor_plugin
-  - pytest -v -s plugins_tests/test_io_processor_plugins.py
+  - pytest -v -s plugins_tests/test_terratorch_io_processor_plugins.py
   - pip uninstall prithvi_io_processor_plugin -y
   # test bge_m3_sparse io_processor plugin
   - pip install -e ./plugins/bge_m3_sparse_plugin
@@ -36,3 +39,8 @@ steps:
   - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
   - pytest -v -s models/test_oot_registration.py # it needs a clean process
   - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
+  mirror:
+    amd:
+      device: mi325_2
+      depends_on:
+      - image-build-amd

.buildkite/test_areas/ray_compat.yaml

@@ -0,0 +1,16 @@
+group: Ray Compatibility
+depends_on:
+  - image-build
+steps:
+- label: Ray Dependency Compatibility Check
+  # Informational only — does not block the pipeline.
+  # If this fails, it means the PR introduces a dependency that
+  # conflicts with Ray's dependency constraints.
+  # See https://github.com/vllm-project/vllm/issues/33599
+  soft_fail: true
+  timeout_in_minutes: 10
+  source_file_dependencies:
+  - requirements/
+  - setup.py
+  commands:
+  - bash /vllm-workspace/.buildkite/scripts/check-ray-compatibility.sh

.buildkite/test_areas/weight_loading.yaml

@@ -13,13 +13,13 @@ steps:
   commands:
     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
 
-- label: Weight Loading Multiple GPU - Large Models # optional
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  device: a100
-  optional: true
-  source_file_dependencies:
-  - vllm/
-  - tests/weight_loading
-  commands:
-    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt
+# - label: Weight Loading Multiple GPU - Large Models # optional
+#   working_dir: "/vllm-workspace/tests"
+#   num_devices: 2
+#   device: a100
+#   optional: true
+#   source_file_dependencies:
+#   - vllm/
+#   - tests/weight_loading
+#   commands:
+#     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt

.github/mergify.yml

@@ -3,6 +3,7 @@ pull_request_rules:
   description: Automatically apply documentation label
   conditions:
     - label != stale
+    - -closed
     - or:
       - files~=^[^/]+\.md$
       - files~=^docs/

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

@@ -6,6 +6,9 @@ on:
       - main
   workflow_dispatch:  # Manual trigger
 
+permissions:
+  contents: read
+
 jobs:
   macos-m1-smoke-test:
     runs-on: macos-latest

.pre-commit-config.yaml

@@ -13,7 +13,7 @@ repos:
     args: [--output-format, github, --fix]
   - id: ruff-format
 - repo: https://github.com/crate-ci/typos
-  rev: v1.38.1
+  rev: v1.43.5
   hooks:
   - id: typos
     args: [--force-exclude]
@@ -55,7 +55,7 @@ repos:
       language: python
       types_or: [python, pyi]
       require_serial: true
-      additional_dependencies: [mypy==1.11.1, regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
+      additional_dependencies: ["mypy[faster-cache]==1.15.0", regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
   - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.10
     entry: python tools/pre_commit/mypy.py 1 "3.10"
@@ -127,6 +127,13 @@ repos:
     language: python
     types: [python]
     additional_dependencies: [regex]
+  # prevent use torch.cuda APIs
+  - id: check-torch-cuda-call
+    name: "Prevent new 'torch.cuda' APIs call"
+    entry: python tools/pre_commit/check_torch_cuda.py
+    language: python
+    types: [python]
+    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

.readthedocs.yaml

@@ -9,6 +9,7 @@ build:
     python: "3.12"
   jobs:
     post_checkout:
+      - bash docs/maybe_skip_pr_build.sh
       - git fetch origin main --unshallow --no-tags --filter=blob:none || true
     pre_create_environment:
       - pip install uv

CMakeLists.txt

@@ -771,6 +771,33 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
+  # Expert-specialization MXFP8 blockscaled grouped kernels (SM100+).
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(ES_MXFP8_GROUPED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(ES_MXFP8_GROUPED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+  endif()
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND ES_MXFP8_GROUPED_MM_ARCHS)
+    set(SRCS
+      "csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm.cu"
+      "csrc/moe/mxfp8_moe/mxfp8_experts_quant.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${ES_MXFP8_GROUPED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_ES_MXFP8_GROUPED_MM_SM100=1")
+    message(STATUS "Building ES MXFP8 grouped kernels for archs: ${ES_MXFP8_GROUPED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8
+        AND ES_MXFP8_GROUPED_MM_ARCHS)
+      message(STATUS "Not building ES MXFP8 grouped kernels as CUDA Compiler version is "
+                     "not >= 12.8.")
+    else()
+      message(STATUS "Not building ES MXFP8 grouped kernels as no compatible archs found "
+                     "in CUDA target architectures.")
+    endif()
+  endif()
+
   # DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper and later)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
     cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")

benchmarks/attention_benchmarks/common.py

@@ -30,7 +30,7 @@ def batch_spec_sort_key(spec: str) -> tuple[int, int, int]:
         max_kv_len = max(r.kv_len for r in requests) if requests else 0
         return (batch_size, max_q_len, max_kv_len)
     except Exception:
-        # Fallback for unparseable specs
+        # Fallback for unparsable specs
         return (0, 0, 0)
 
 

benchmarks/attention_benchmarks/mla_runner.py

@@ -701,7 +701,7 @@ def _run_single_benchmark(
     # Warmup
     for _ in range(config.warmup_iters):
         forward_fn()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Benchmark
     times = []
@@ -714,7 +714,7 @@ def _run_single_benchmark(
             forward_fn()
         end.record()
 
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         elapsed_ms = start.elapsed_time(end)
         times.append(elapsed_ms / 1000.0 / config.num_layers)
 

benchmarks/attention_benchmarks/runner.py

@@ -391,7 +391,7 @@ def _run_single_benchmark(
                 attn_metadata,
                 output=out,
             )
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Benchmark
     times = []
@@ -412,7 +412,7 @@ def _run_single_benchmark(
             )
         end.record()
 
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         elapsed_ms = start.elapsed_time(end)
         times.append(elapsed_ms / 1000.0 / config.num_layers)  # seconds per layer
 

benchmarks/auto_tune/auto_tune.sh

@@ -85,7 +85,6 @@ start_server() {
     # Each argument and its value are separate elements.
     local common_args_array=(
         "$MODEL"
-        "--disable-log-requests"
         "--port" "8004"
         "--host" "$HOSTNAME"
         "--gpu-memory-utilization" "$gpu_memory_utilization"

benchmarks/benchmark_topk_topp.py

@@ -94,7 +94,7 @@ def create_logits(
 
 def measure_memory() -> tuple[int, int]:
     """Return (allocated, reserved) memory in bytes."""
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
     return torch.cuda.memory_allocated(), torch.cuda.max_memory_allocated()
 
 
@@ -102,7 +102,7 @@ def reset_memory_stats():
     """Reset peak memory statistics."""
     reset_buffer_cache()
     torch.cuda.reset_peak_memory_stats()
-    torch.cuda.empty_cache()
+    torch.accelerator.empty_cache()
     gc.collect()
 
 
@@ -123,7 +123,7 @@ def benchmark_function(
     for _ in range(warmup_iters):
         logits_copy = logits.clone()
         func(logits_copy, k, p)
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Reset memory stats before benchmark
     reset_memory_stats()
@@ -140,7 +140,7 @@ def benchmark_function(
         func(logits_copy, k, p)
         end_events[i].record()
 
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Calculate timing
     times = [

benchmarks/kernels/benchmark_2d_silu_mul_fp8_quant.py

@@ -168,7 +168,7 @@ def bench_impl(
     # warmup
     for kwargs in kwargs_list:
         impl_type.get_impl()(**kwargs)
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Merge into a single kwargs and qualify arguments as ArgPool
     kwargs = {k: ArgPool([]) for k in kwargs_list[0]}
@@ -202,7 +202,7 @@ def test_correctness(T: int, N: int):
     # reference output
     ref_out_q, ref_out_s = output_from_impl(ImplType.REFERENCE)
 
-    # test ouptut
+    # test output
     out_q, out_s = output_from_impl(
         ImplType.SILU_MUL_PER_TOKEN_GROUP_QUANT_FP8_COLMAJOR
     )

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -12,12 +12,12 @@ import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from tests.kernels.moe.utils import make_dummy_moe_config
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.fused_moe.activation import MoEActivation
+from vllm.model_executor.layers.fused_moe.all2all_utils import (
+    maybe_make_prepare_finalize,
+)
 from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
 from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassExpertsFp8
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
-from vllm.model_executor.layers.fused_moe.prepare_finalize import (
-    MoEPrepareAndFinalizeNoEP,
-)
 from vllm.platforms import current_platform
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.v1.worker.workspace import init_workspace_manager
@@ -137,15 +137,21 @@ def bench_run(
         per_out_ch_quant=per_out_ch,
     )
 
-    fn = mk.FusedMoEModularKernel(
-        MoEPrepareAndFinalizeNoEP(),
+    moe_config = make_dummy_moe_config(
+        num_experts=num_experts,
+        hidden_dim=k,
+        intermediate_size_per_partition=n,
+        in_dtype=a.dtype,
+    )
+    fn = mk.FusedMoEKernel(
+        maybe_make_prepare_finalize(
+            moe=moe_config,
+            quant_config=quant_config,
+            allow_new_interface=True,
+            use_monolithic=False,
+        ),
         CutlassExpertsFp8(
-            moe_config=make_dummy_moe_config(
-                num_experts=num_experts,
-                hidden_dim=k,
-                intermediate_size_per_partition=n,
-                in_dtype=a.dtype,
-            ),
+            moe_config=moe_config,
             quant_config=quant_config,
         ),
     )
@@ -165,7 +171,7 @@ def bench_run(
                 activation=MoEActivation.SILU,
                 global_num_experts=num_experts,
             )
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Create CUDA graphs for Triton (match benchmark_moe.py pattern exactly)
     triton_stream = torch.cuda.Stream()
@@ -181,14 +187,14 @@ def bench_run(
                 topk_ids,
                 quant_config=quant_config,
             )
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     def bench_cuda_graph(graph, num_warmup=5, num_iters=100):
         """Benchmark CUDA graph using events like benchmark_moe.py"""
         # Warmup
         for _ in range(num_warmup):
             graph.replay()
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
 
         # Timing
         start_event = torch.Event(enable_timing=True)
@@ -196,7 +202,7 @@ def bench_run(
 
         latencies = []
         for _ in range(num_iters):
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
             start_event.record()
             graph.replay()
             end_event.record()

benchmarks/kernels/benchmark_cutlass_moe_nvfp4.py

@@ -15,6 +15,9 @@ import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from tests.kernels.moe.utils import make_dummy_moe_config
 from vllm import _custom_ops as ops
 from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.fused_moe.all2all_utils import (
+    maybe_make_prepare_finalize,
+)
 from vllm.model_executor.layers.fused_moe.config import (
     fp8_w8a8_moe_quant_config,
     nvfp4_moe_quant_config,
@@ -23,9 +26,6 @@ from vllm.model_executor.layers.fused_moe.cutlass_moe import (
     CutlassExpertsFp4,
 )
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
-from vllm.model_executor.layers.fused_moe.prepare_finalize import (
-    MoEPrepareAndFinalizeNoEP,
-)
 from vllm.scalar_type import scalar_types
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.v1.worker.workspace import init_workspace_manager
@@ -196,10 +196,21 @@ def bench_run(
             g2_alphas=w2_gs,
         )
 
-        kernel = mk.FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+        moe_config = make_dummy_moe_config(
+            num_experts=num_experts,
+            hidden_dim=k,
+            intermediate_size_per_partition=n,
+            in_dtype=a.dtype,
+        )
+        kernel = mk.FusedMoEKernel(
+            maybe_make_prepare_finalize(
+                moe=moe_config,
+                quant_config=quant_config,
+                allow_new_interface=True,
+                use_monolithic=False,
+            ),
             CutlassExpertsFp4(
-                make_dummy_moe_config(),
+                moe_config=moe_config,
                 quant_config=quant_config,
             ),
         )
@@ -240,11 +251,17 @@ def bench_run(
             g1_alphas=w1_gs,
             g2_alphas=w2_gs,
         )
+        moe_config = make_dummy_moe_config()
 
-        kernel = mk.FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+        kernel = mk.FusedMoEKernel(
+            maybe_make_prepare_finalize(
+                moe=moe_config,
+                quant_config=quant_config,
+                allow_new_interface=True,
+                use_monolithic=False,
+            ),
             CutlassExpertsFp4(
-                make_dummy_moe_config(),
+                moe_config=moe_config,
                 quant_config=quant_config,
             ),
         )
@@ -290,7 +307,7 @@ def bench_run(
     def replay_graph(graph, num_repeats):
         for _ in range(num_repeats):
             graph.replay()
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
 
     cutlass_stream = torch.cuda.Stream()
     cutlass_graph = torch.cuda.CUDAGraph()
@@ -313,7 +330,7 @@ def bench_run(
             e=num_experts,
             device=device,
         )
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     triton_stream = torch.cuda.Stream()
     triton_graph = torch.cuda.CUDAGraph()
@@ -328,7 +345,7 @@ def bench_run(
             w2_fp8scale,
             a_fp8_scale,
         )
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     min_run_time = 5
     num_warmup = 5

benchmarks/kernels/benchmark_device_communicators.py

@@ -342,7 +342,7 @@ class CommunicatorBenchmark:
             if not should_use_fn(tensor):
                 return None
 
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
             stream = torch.cuda.Stream()
             with torch.cuda.stream(stream):
                 graph_input = tensor.clone()
@@ -360,17 +360,17 @@ class CommunicatorBenchmark:
                         for _ in range(CUDA_GRAPH_CAPTURE_CYCLES):
                             allreduce_fn(graph_input)
 
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
             for _ in range(num_warmup):
                 graph.replay()
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
 
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
             start_time = time.perf_counter()
 
             for _ in range(num_trials):
                 graph.replay()
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
 
             end_time = time.perf_counter()
 

benchmarks/kernels/benchmark_fused_collective.py

@@ -385,7 +385,7 @@ def benchmark_operation(
     # Warmup before graph capture
     for _ in range(warmup):
         operation_func(*args, **kwargs)
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Create CUDA graph
     graph = torch.cuda.CUDAGraph()
@@ -398,19 +398,19 @@ def benchmark_operation(
             operation_func(*args, **kwargs)
 
     # Graph warmup
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
     for _ in range(warmup):
         graph.replay()
 
     # Benchmark with CUDA graph
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
     start_time = time.perf_counter()
 
     for _ in range(trials // num_op_per_cudagraph):
         # operation_func(*args, **kwargs)
         graph.replay()
 
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
     end_time = time.perf_counter()
 
     avg_time_ms = ((end_time - start_time) / trials) * 1000

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -9,15 +9,15 @@ import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from tests.kernels.moe.utils import make_dummy_moe_config
 from vllm import _custom_ops as ops
 from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.fused_moe.all2all_utils import (
+    maybe_make_prepare_finalize,
+)
 from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
 from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassExpertsFp8
 from vllm.model_executor.layers.fused_moe.fused_moe import (
     fused_experts,
     fused_topk,
 )
-from vllm.model_executor.layers.fused_moe.prepare_finalize import (
-    MoEPrepareAndFinalizeNoEP,
-)
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.v1.worker.workspace import init_workspace_manager
 
@@ -131,16 +131,22 @@ def bench_run(
             w2_scale=w2_scale,
             per_act_token_quant=per_act_token,
         )
+        moe_config = make_dummy_moe_config(
+            num_experts=w2.shape[0],
+            hidden_dim=w2.shape[1],
+            intermediate_size_per_partition=w2.shape[2],
+            in_dtype=a.dtype,
+        )
 
-        fn = mk.FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+        fn = mk.FusedMoEKernel(
+            maybe_make_prepare_finalize(
+                moe=moe_config,
+                quant_config=quant_config,
+                allow_new_interface=True,
+                use_monolithic=False,
+            ),
             CutlassExpertsFp8(
-                moe_config=make_dummy_moe_config(
-                    num_experts=w2.shape[0],
-                    hidden_dim=w2.shape[1],
-                    intermediate_size_per_partition=w2.shape[2],
-                    in_dtype=a.dtype,
-                ),
+                moe_config=moe_config,
                 quant_config=quant_config,
             ),
         )
@@ -163,16 +169,22 @@ def bench_run(
             w2_scale=w2_scale,
             per_act_token_quant=per_act_token,
         )
+        moe_config = make_dummy_moe_config(
+            num_experts=w2.shape[0],
+            hidden_dim=w2.shape[1],
+            intermediate_size_per_partition=w2.shape[2],
+            in_dtype=a.dtype,
+        )
 
-        fn = mk.FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+        fn = mk.FusedMoEKernel(
+            maybe_make_prepare_finalize(
+                moe=moe_config,
+                quant_config=quant_config,
+                allow_new_interface=True,
+                use_monolithic=False,
+            ),
             CutlassExpertsFp8(
-                moe_config=make_dummy_moe_config(
-                    num_experts=w2.shape[0],
-                    hidden_dim=w2.shape[1],
-                    intermediate_size_per_partition=w2.shape[2],
-                    in_dtype=a.dtype,
-                ),
+                moe_config=moe_config,
                 quant_config=quant_config,
             ),
         )
@@ -212,7 +224,7 @@ def bench_run(
     def replay_graph(graph, num_repeats):
         for _ in range(num_repeats):
             graph.replay()
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
 
     cutlass_stream = torch.cuda.Stream()
     cutlass_graph = torch.cuda.CUDAGraph()
@@ -227,7 +239,7 @@ def bench_run(
             topk_weights,
             topk_ids,
         )
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     triton_stream = torch.cuda.Stream()
     triton_graph = torch.cuda.CUDAGraph()
@@ -242,7 +254,7 @@ def bench_run(
             w2_scale,
             a_scale,
         )
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     min_run_time = 5
     num_warmup = 5

benchmarks/kernels/benchmark_layernorm.py

@@ -34,14 +34,14 @@ def main(
     residual = torch.randn_like(x) * scale if add_residual else None
 
     def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         if profile:
             torch.cuda.cudart().cudaProfilerStart()
         start_time = time.perf_counter()
 
         for _ in range(num_iters):
             layer(x, residual)
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
 
         end_time = time.perf_counter()
         if profile:

benchmarks/kernels/benchmark_lora.py

@@ -1035,7 +1035,7 @@ def bench_optype(
     # 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)
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Merge into a single kwargs and qualify arguments as ArgPool
     kwargs = {k: ArgPool([]) for k in kwargs_list[0]}

benchmarks/kernels/benchmark_mla_k_concat.py

@@ -47,13 +47,13 @@ def benchmark_method(
     # Warmup
     for _ in range(num_warmup):
         _ = method(k_nope, k_pe)
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Benchmark
     start = time.perf_counter()
     for _ in range(num_iters):
         _ = method(k_nope, k_pe)
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
     end = time.perf_counter()
 
     return (end - start) / num_iters * 1000  # Convert to ms

benchmarks/kernels/benchmark_moe.py

@@ -17,6 +17,9 @@ from ray.experimental.tqdm_ray import tqdm
 
 from vllm.model_executor.layers.fused_moe import fused_topk
 from vllm.model_executor.layers.fused_moe.activation import MoEActivation
+from vllm.model_executor.layers.fused_moe.all2all_utils import (
+    maybe_make_prepare_finalize,
+)
 from vllm.model_executor.layers.fused_moe.config import (
     FusedMoEConfig,
     FusedMoEParallelConfig,
@@ -51,7 +54,7 @@ def clear_triton_cache():
 
     # Clear CUDA memory cache
     if torch.cuda.is_available():
-        torch.cuda.empty_cache()
+        torch.accelerator.empty_cache()
 
     # Try to clear Triton's runtime cache
     try:
@@ -242,24 +245,33 @@ def benchmark_config(
 
         deep_gemm_experts = None
         if use_deep_gemm:
-            deep_gemm_experts = mk.FusedMoEModularKernel(
-                prepare_finalize=MoEPrepareAndFinalizeNoEP(),
+            moe_config = (
+                FusedMoEConfig(
+                    num_experts=num_experts,
+                    experts_per_token=topk,
+                    hidden_dim=hidden_size,
+                    intermediate_size_per_partition=shard_intermediate_size,
+                    num_local_experts=num_experts,
+                    num_logical_experts=num_experts,
+                    activation=MoEActivation.SILU,
+                    moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(),
+                    in_dtype=init_dtype,
+                    routing_method=RoutingMethodType.TopK,
+                    device="cuda",
+                ),
+            )
+            deep_gemm_experts = mk.FusedMoEKernel(
+                prepare_finalize=maybe_make_prepare_finalize(
+                    moe=moe_config,
+                    quant_config=quant_config,
+                    allow_new_interface=True,
+                    use_monolithic=False,
+                ),
                 fused_experts=TritonOrDeepGemmExperts(
-                    moe_config=FusedMoEConfig(
-                        num_experts=num_experts,
-                        experts_per_token=topk,
-                        hidden_dim=hidden_size,
-                        intermediate_size_per_partition=shard_intermediate_size,
-                        num_local_experts=num_experts,
-                        num_logical_experts=num_experts,
-                        activation=MoEActivation.SILU,
-                        moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(),
-                        in_dtype=init_dtype,
-                        routing_method=RoutingMethodType.TopK,
-                        device="cuda",
-                    ),
+                    moe_config=moe_config,
                     quant_config=quant_config,
                 ),
+                inplace=not disable_inplace(),
             )
 
         with override_config(config):
@@ -269,8 +281,16 @@ def benchmark_config(
 
             inplace = not disable_inplace()
             if use_deep_gemm:
-                return deep_gemm_experts(
-                    x, w1, w2, topk_weights, topk_ids, inplace=inplace
+                return deep_gemm_experts.apply(
+                    x,
+                    w1,
+                    w2,
+                    topk_weights,
+                    topk_ids,
+                    activation=MoEActivation.SILU,
+                    global_num_experts=num_experts,
+                    apply_router_weight_on_input=False,
+                    expert_map=False,
                 )
             return fused_experts(
                 x,
@@ -284,19 +304,19 @@ def benchmark_config(
 
     # JIT compilation & warmup
     run()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Capture 10 invocations with CUDA graph
     graph = torch.cuda.CUDAGraph()
     with torch.cuda.graph(graph):
         for _ in range(10):
             run()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Warmup
     for _ in range(5):
         graph.replay()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     start_event = torch.Event(enable_timing=True)
     end_event = torch.Event(enable_timing=True)
@@ -304,7 +324,7 @@ def benchmark_config(
     latencies: list[float] = []
     for i in range(num_iters):
         prepare(i)
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
 
         start_event.record()
         graph.replay()

benchmarks/kernels/benchmark_moe_defaults.py

@@ -131,7 +131,7 @@ def benchmark_config(
                 topk_ids,
                 quant_config=quant_config,
             )
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Benchmark
     start = torch.cuda.Event(enable_timing=True)
@@ -149,7 +149,7 @@ def benchmark_config(
                 quant_config=quant_config,
             )
     end.record()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
     return start.elapsed_time(end) / num_iters * 1000  # ms -> us
 
 

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -69,19 +69,19 @@ def benchmark_permute(
 
     # JIT compilation & warmup
     run()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Capture 10 invocations with CUDA graph
     graph = torch.cuda.CUDAGraph()
     with torch.cuda.graph(graph):
         for _ in range(10):
             run()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Warmup
     for _ in range(5):
         graph.replay()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     start_event = torch.Event(enable_timing=True)
     end_event = torch.Event(enable_timing=True)
@@ -89,7 +89,7 @@ def benchmark_permute(
     latencies: list[float] = []
     for i in range(num_iters):
         prepare(i)
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
 
         start_event.record()
         graph.replay()
@@ -159,26 +159,26 @@ def benchmark_unpermute(
     # JIT compilation & warmup
     input = prepare()
     run(input)
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Capture 10 invocations with CUDA graph
     graph = torch.cuda.CUDAGraph()
     with torch.cuda.graph(graph):
         for _ in range(10):
             run(input)
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Warmup
     for _ in range(5):
         graph.replay()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     start_event = torch.Event(enable_timing=True)
     end_event = torch.Event(enable_timing=True)
 
     latencies: list[float] = []
     for i in range(num_iters):
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         start_event.record()
         graph.replay()
         end_event.record()

benchmarks/kernels/benchmark_mrope.py

@@ -135,14 +135,14 @@ def benchmark_mrope(
             key.clone(),
         )
 
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     # Time reference implementation
     torch_times = []
     for _ in range(benchmark_iter):
         query_clone = query.clone()
         key_clone = key.clone()
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         start_time = time.time()
 
         mrope_helper_class.forward_native(
@@ -151,7 +151,7 @@ def benchmark_mrope(
             key_clone,
         )
 
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         torch_times.append(time.time() - start_time)
 
     # Time triton kernel implementation
@@ -159,14 +159,14 @@ def benchmark_mrope(
     for _ in range(benchmark_iter):
         query_clone = query.clone()
         key_clone = key.clone()
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         start_time = time.time()
         mrope_helper_class.forward_cuda(
             positions,
             query_clone,
             key_clone,
         )
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         triton_times.append(time.time() - start_time)
 
     # Calculate statistics

benchmarks/kernels/benchmark_paged_attention.py

@@ -103,7 +103,7 @@ def main(
         max_logits = torch.empty_like(exp_sums)
 
     def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         if profile:
             torch.cuda.cudart().cudaProfilerStart()
         start_time = time.perf_counter()
@@ -173,7 +173,7 @@ def main(
                     )
             else:
                 raise ValueError(f"Invalid version: {version}")
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
 
         end_time = time.perf_counter()
         if profile:

benchmarks/kernels/benchmark_per_token_group_quant.py

@@ -28,7 +28,7 @@ def _time_cuda(
     # warmup
     for _ in range(warmup_iters):
         fn()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     start = torch.Event(enable_timing=True)
     end = torch.Event(enable_timing=True)
@@ -37,7 +37,7 @@ def _time_cuda(
     for _ in range(bench_iters):
         fn()
     end.record()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     return start.elapsed_time(end) / bench_iters  # ms/iter
 

benchmarks/kernels/benchmark_quant.py

@@ -29,7 +29,7 @@ def main(
     scale = torch.randn(1, 1, dtype=torch.float32) if static_scale else None
 
     def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         if profile:
             torch.cuda.cudart().cudaProfilerStart()
         start_time = time.perf_counter()
@@ -39,7 +39,7 @@ def main(
                 ops.scaled_int8_quant(x, scale)
             else:
                 ops.scaled_fp8_quant(x, scale)
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
 
         end_time = time.perf_counter()
         if profile:

benchmarks/kernels/benchmark_reshape_and_cache.py

@@ -84,16 +84,16 @@ def run_benchmark(
         g = torch.cuda.CUDAGraph()
         with torch.cuda.graph(g):
             function_under_test()
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         function_under_test = lambda: g.replay()
 
     def run_cuda_benchmark(n_iters: int) -> float:
         nonlocal key, value, key_cache, value_cache, slot_mapping
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         start = time.perf_counter()
         for _ in range(n_iters):
             function_under_test()
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
         end = time.perf_counter()
         return (end - start) / n_iters
 
@@ -104,7 +104,7 @@ def run_benchmark(
 
     # free tensors to mitigate OOM when sweeping
     del key, value, key_cache, value_cache, slot_mapping
-    torch.cuda.empty_cache()
+    torch.accelerator.empty_cache()
 
     return lat
 

benchmarks/kernels/benchmark_reshape_and_cache_flash.py

@@ -109,16 +109,16 @@ def run_benchmark(
         g = torch.cuda.CUDAGraph()
         with torch.cuda.graph(g):
             function_under_test()
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         function_under_test = lambda: g.replay()
 
     def run_cuda_benchmark(n_iters: int) -> float:
         nonlocal key, value, key_cache, value_cache, slot_mapping
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         start = time.perf_counter()
         for _ in range(n_iters):
             function_under_test()
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
         end = time.perf_counter()
         return (end - start) / n_iters
 
@@ -129,7 +129,7 @@ def run_benchmark(
 
     # free tensors to mitigate OOM when sweeping
     del key, value, key_cache, value_cache, slot_mapping
-    torch.cuda.empty_cache()
+    torch.accelerator.empty_cache()
 
     return lat
 

benchmarks/kernels/benchmark_silu_mul_fp8_quant.py

@@ -251,7 +251,7 @@ def benchmark(
         kernel(
             y, tokens_per_expert, num_parallel_tokens=num_parallel_tokens, group_size=G
         )
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     start_event = torch.Event(enable_timing=True)
     end_event = torch.Event(enable_timing=True)
@@ -259,7 +259,7 @@ def benchmark(
     # Benchmark
     latencies: list[float] = []
     for _ in range(runs):
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
 
         start_event.record()
         for i in range(iterations_per_run):

benchmarks/kernels/benchmark_trtllm_decode_attention.py

@@ -126,7 +126,7 @@ def benchmark_decode(
     )
 
     def time_fn(fn, warmup=10, trials=20):
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         start = torch.Event(enable_timing=True)
         end = torch.Event(enable_timing=True)
         times = []
@@ -136,7 +136,7 @@ def benchmark_decode(
             start.record()
             fn()
             end.record()
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
             times.append(start.elapsed_time(end))  # ms
         return sum(times) / len(times), torch.std(torch.tensor(times))
 

benchmarks/kernels/benchmark_trtllm_prefill_attention.py

@@ -138,7 +138,7 @@ def benchmark_prefill(
     )
 
     def time_fn(fn, warmup=10, trials=20):
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         start = torch.Event(enable_timing=True)
         end = torch.Event(enable_timing=True)
         times = []
@@ -148,7 +148,7 @@ def benchmark_prefill(
             start.record()
             fn()
             end.record()
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
             times.append(start.elapsed_time(end))  # ms
         return sum(times) / len(times), torch.std(torch.tensor(times))
 

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -177,18 +177,18 @@ def benchmark_config(
     def run():
         w8a8_block_matmul(A, B, As, Bs, block_size, config, out_dtype)
 
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
     # JIT complication & warmup
     for _ in range(5):
         run()
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
 
     start_event = torch.Event(enable_timing=True)
     end_event = torch.Event(enable_timing=True)
 
     latencies: list[float] = []
     for i in range(num_iters):
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         start_event.record()
         run()
         end_event.record()

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -35,7 +35,7 @@ def benchmark_shape(
     B = torch.randn((n, k), device="cuda", dtype=torch.bfloat16)
 
     # Reference result in BF16
-    torch.cuda.synchronize()
+    torch.accelerator.synchronize()
     C_ref = A @ B.t()
 
     # Pre-quantize B for all implementations
@@ -121,14 +121,14 @@ def benchmark_shape(
         # Warmup
         for _ in range(warmup):
             func()
-            torch.cuda.synchronize()
+            torch.accelerator.synchronize()
 
         # Timing loop
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         start = time.time()
         for _ in range(repeat):
             func()
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         end = time.time()
 
         # Calculate timing and TFLOPS

benchmarks/multi_turn/README.md

@@ -7,7 +7,7 @@ First start serving your model
 ```bash
 export MODEL_PATH=/models/meta-llama/Meta-Llama-3.1-8B-Instruct/
 
-vllm serve $MODEL_PATH --served-model-name Llama --disable-log-requests
+vllm serve $MODEL_PATH --served-model-name Llama
 ```
 
 The variable `MODEL_PATH` should be a path to the model files (e.g. downloaded from huggingface).

cmake/cpu_extension.cmake

@@ -242,13 +242,24 @@ if (ENABLE_X86_ISA OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND
         )
     else()
         message(STATUS "Downloading oneDNN from GitHub")
-        FetchContent_Declare(
-            oneDNN
-            GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-            GIT_TAG v3.10
-            GIT_PROGRESS TRUE
-            GIT_SHALLOW TRUE
-        )
+        if(ASIMD_FOUND AND NOT APPLE_SILICON_FOUND)
+            message(STATUS "aarch64 detected: using pinned oneDNN commit 9c5be1cc59e368aebf0909e6cf20f981ea61462a")
+            FetchContent_Declare(
+                oneDNN
+                GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
+                GIT_TAG        9c5be1cc59e368aebf0909e6cf20f981ea61462a
+                GIT_PROGRESS   TRUE
+                GIT_SHALLOW    FALSE
+            )
+        else()
+            FetchContent_Declare(
+                oneDNN
+                GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
+                GIT_TAG        v3.10
+                GIT_PROGRESS   TRUE
+                GIT_SHALLOW    TRUE
+            )
+        endif()
     endif()
 
     set(ONEDNN_LIBRARY_TYPE "STATIC")

cmake/external_projects/vllm_flash_attn.cmake

@@ -46,24 +46,20 @@ else()
   )
 endif()
 
-
-# Install rules for FA components need the install prefix nested under vllm/
-# These run at install time, before the FA library's own install rules
-foreach(_FA_COMPONENT _vllm_fa2_C _vllm_fa3_C)
-  install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" COMPONENT ${_FA_COMPONENT})
-  install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" COMPONENT ${_FA_COMPONENT})
-  install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" COMPONENT ${_FA_COMPONENT})
-endforeach()
+# Make sure vllm-flash-attn install rules are nested under vllm/
+# ALL_COMPONENTS ensures the save/modify/restore runs exactly once regardless
+# of how many components are being installed, avoiding double-append of /vllm/.
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" ALL_COMPONENTS)
+install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
+install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" ALL_COMPONENTS)
 
 # Fetch the vllm-flash-attn library
 FetchContent_MakeAvailable(vllm-flash-attn)
 message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
 
 # Restore the install prefix after FA's install rules
-foreach(_FA_COMPONENT _vllm_fa2_C _vllm_fa3_C)
-  install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" COMPONENT ${_FA_COMPONENT})
-  install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" COMPONENT ${_FA_COMPONENT})
-endforeach()
+install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
 
 # Install shared Python files for both FA2 and FA3 components
 foreach(_FA_COMPONENT _vllm_fa2_C _vllm_fa3_C)

csrc/cpu/cpu_attn_amx.hpp

@@ -420,7 +420,7 @@ class AttentionImpl<ISA::AMX, scalar_t, head_dim> {
       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
+    // For AMX B matrix, 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

csrc/cpu/dnnl_helper.cpp

@@ -237,13 +237,10 @@ W8A8MatMulPrimitiveHandler::W8A8MatMulPrimitiveHandler(const Args& args)
   };
   dnnl::memory::desc original_b_md({b_k_size_, b_n_size_}, b_type_,
                                    {b_k_stride_, b_n_stride_});
-#ifdef __aarch64__
+
   // dummy M size for prepacking weights
   // Prepacking weights improves performance and avoid runtime reorders
   constexpr dnnl_dim_t kProbeM = 128;
-#else
-  constexpr dnnl_dim_t kProbeM = DNNL_RUNTIME_DIM_VAL;
-#endif
 
   prepack_weight(args.b_ptr, original_b_md,
                  create_primitive_desc(
@@ -411,21 +408,19 @@ MatMulPrimitiveHandler::MatMulPrimitiveHandler(const Args& args)
   dnnl::memory::desc original_b_md({b_k_size_, b_n_size_}, b_type_,
                                    {b_k_stride_, b_n_stride_});
 
+  // dummy M size for prepacking weights
+  // Prepacking weights improves performance and avoid runtime reorders
+  constexpr dnnl_dim_t kProbeM = 128;
+
   prepack_weight(args.b_ptr, original_b_md,
                  create_primitive_desc(
-                     MSizeCacheKey{
-#ifdef VLLM_USE_ACL
-                         // Arm Compute Library (ACL) backend for oneDNN does
-                         // not support runtime
-                         // dimensions, so we set M to a default value
-                         .a_m_size = 128,
-                         .a_m_stride = b_k_size_,
-#else
-                         .a_m_size = DNNL_RUNTIME_DIM_VAL,
-                         .a_m_stride = DNNL_RUNTIME_DIM_VAL,
-#endif
-                         .use_bias = false,
-                         .bias_type = dnnl::memory::data_type::undef},
+                     MSizeCacheKey{// Use a concrete M so oneDNN's kernel
+                                   // selector can choose an optimally blocked
+                                   // weight layout.
+                                   .a_m_size = kProbeM,
+                                   .a_m_stride = b_k_size_,
+                                   .use_bias = false,
+                                   .bias_type = dnnl::memory::data_type::undef},
                      true)
                      .weights_desc());
   init_runtime_memory_cache(args);

csrc/cpu/torch_bindings.cpp

@@ -4,7 +4,7 @@
 
 #include <torch/library.h>
 
-// Note: overwrite the external defination for sharing same name between
+// Note: overwrite the external definition for sharing same name between
 // libraries use different ISAs.
 #define TORCH_EXTENSION_NAME _C
 

csrc/moe/moe_align_sum_kernels.cu

@@ -35,11 +35,11 @@ __global__ void batched_moe_align_block_size_kernel(
   int32_t const block_ids_size = sorted_ids_size / block_size;
   int32_t const SENTINEL =
       num_batches * max_tokens_per_batch;  // To denote invalid entries.
-  // Intialize sorted_ids
+  // Initialize sorted_ids
   for (size_t i = threadIdx.x; i < sorted_ids_size; i += stride) {
     sorted_ids[i] = SENTINEL;
   }
-  // Intialize expert_ids with -1
+  // Initialize expert_ids with -1
   for (size_t i = threadIdx.x; i < block_ids_size; i += stride) {
     block_ids[i] = -1;
   }

csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm.cu

@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+// Adapted from SGLang:
+// https://github.com/sgl-project/sglang/blob/ded068a76e00878881d52d5bfb791e0f60d7311b/sgl-kernel/csrc/expert_specialization/es_sm100_mxfp8_blockscaled.cu
+
+#include <torch/all.h>
+
+#include "cutlass_mxfp8_grouped_mm_launcher.cuh"
+
+void cutlass_mxfp8_grouped_mm(const torch::Tensor& a, const torch::Tensor& b,
+                              const torch::Tensor& sfa,
+                              const torch::Tensor& sfb, torch::Tensor& d,
+                              const torch::Tensor& problem_sizes,
+                              const torch::Tensor& expert_offsets,
+                              const torch::Tensor& blockscale_offsets) {
+#if defined(CUTLASS_ARCH_MMA_SM100_SUPPORTED)
+  TORCH_CHECK(problem_sizes.dim() == 2, "problem_sizes must be 2D tensor");
+  TORCH_CHECK(problem_sizes.size(1) == 3,
+              "problem_sizes must have shape (num_experts, 3)");
+  TORCH_CHECK(problem_sizes.size(0) == expert_offsets.size(0),
+              "Number of experts in problem_sizes must match expert_offsets");
+  TORCH_CHECK(problem_sizes.dtype() == torch::kInt32,
+              "problem_sizes must be int32");
+  TORCH_CHECK(expert_offsets.dtype() == torch::kInt32,
+              "expert_offsets must be int32");
+  TORCH_CHECK(blockscale_offsets.dtype() == torch::kInt32,
+              "blockscale_offsets must be int32");
+  TORCH_CHECK(a.dim() == 2, "a must be a 2D tensor of shape (num_tokens, k)");
+  TORCH_CHECK(b.dim() == 3,
+              "b must be a 3D tensor of shape (num_experts, k, n)");
+  TORCH_CHECK(a.size(1) == b.size(1) && a.size(1) % 128 == 0,
+              "k should align 128");
+  TORCH_CHECK(b.size(2) % 128 == 0, "n should align 128");
+  TORCH_CHECK(a.strides()[1] == 1, "a must be row major");
+  TORCH_CHECK(b.strides()[1] == 1, "b must be column major");
+
+  auto stream = at::cuda::getCurrentCUDAStream();
+  if (d.dtype() == torch::kBFloat16) {
+    expert_specialization::cutlass_mxfp8_grouped_mm_dispatch_out_dtype<
+        cutlass::bfloat16_t>(a, b, sfa, sfb, d, problem_sizes, expert_offsets,
+                             blockscale_offsets, stream);
+  } else if (d.dtype() == torch::kFloat16) {
+    expert_specialization::cutlass_mxfp8_grouped_mm_dispatch_out_dtype<
+        cutlass::half_t>(a, b, sfa, sfb, d, problem_sizes, expert_offsets,
+                         blockscale_offsets, stream);
+  } else {
+    TORCH_CHECK(false, "dtype must be kFloat16 or kBFloat16");
+  }
+#else
+  TORCH_CHECK(false,
+              "No implemented cutlass_mxfp8_grouped_mm for "
+              "current device");
+#endif
+}
+
+#include "core/registration.h"
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("cutlass_mxfp8_grouped_mm", cutlass_mxfp8_grouped_mm);
+}

csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm_functor.cuh

@@ -0,0 +1,141 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+// Adapted from SGLang:
+// https://github.com/sgl-project/sglang/blob/ded068a76e00878881d52d5bfb791e0f60d7311b/sgl-kernel/csrc/expert_specialization/es_sm100_mxfp8_blockscaled_functor.cuh
+
+#pragma once
+#include <cuda.h>
+
+#include "cute/tensor.hpp"
+#include "cutlass/util/packed_stride.hpp"
+#include "cutlass_mxfp8_grouped_mm_traits.cuh"
+
+namespace expert_specialization {
+
+using namespace cute;
+
+template <typename GemmTraits>
+struct CutlassMxfp8GroupedMmOffsetFunctor {
+  using Gemm = typename GemmTraits::Gemm;
+  using ElementA = typename Gemm::ElementA;
+  using ElementB = typename Gemm::ElementB;
+  using ElementSF = typename GemmTraits::ElementSF;
+  using ElementD = typename GemmTraits::ElementOutput;
+  // Input
+  int* expert_offsets{nullptr};
+  int* blockscale_offsets{nullptr};
+  // Output
+  ElementA* a_base{nullptr};
+  ElementB* b_base{nullptr};
+  ElementSF* sfa_base{nullptr};
+  ElementSF* sfb_base{nullptr};
+  ElementD* d_base{nullptr};
+  ElementA** a_offsets{nullptr};
+  ElementB** b_offsets{nullptr};
+  ElementSF** sfa_offsets{nullptr};
+  ElementSF** sfb_offsets{nullptr};
+  ElementD** d_offsets{nullptr};
+
+  CutlassMxfp8GroupedMmOffsetFunctor() = default;
+  CutlassMxfp8GroupedMmOffsetFunctor(
+      int* _expert_offsets, int* _blockscale_offsets, ElementA* _a_base,
+      ElementB* _b_base, ElementSF* _sfa_base, ElementSF* _sfb_base,
+      ElementD* _d_base, ElementA** _a_offsets, ElementB** _b_offsets,
+      ElementSF** _sfa_offsets, ElementSF** _sfb_offsets, ElementD** _d_offsets)
+      : expert_offsets{_expert_offsets},
+        blockscale_offsets{_blockscale_offsets},
+        a_base(_a_base),
+        b_base(_b_base),
+        sfa_base(_sfa_base),
+        sfb_base(_sfb_base),
+        d_base(_d_base),
+        a_offsets(_a_offsets),
+        b_offsets(_b_offsets),
+        sfa_offsets(_sfa_offsets),
+        sfb_offsets(_sfb_offsets),
+        d_offsets(_d_offsets) {}
+
+  void CUTE_DEVICE operator()(int64_t expert_id, int m, int n, int k) {
+    int64_t expert_offset = static_cast<int64_t>(expert_offsets[expert_id]);
+    int64_t blockscale_offset =
+        static_cast<int64_t>(blockscale_offsets[expert_id]);
+    int64_t a_stride = expert_offset * k;
+    int64_t b_stride = expert_id * k * n;
+    int64_t d_stride = expert_offset * n;
+    int64_t sfa_stride = blockscale_offset * (k / 32);
+    int64_t sfb_stride = expert_id * n * (k / 32);
+
+    a_offsets[expert_id] = a_base + a_stride;
+    b_offsets[expert_id] = b_base + b_stride;
+    sfa_offsets[expert_id] = sfa_base + sfa_stride;
+    sfb_offsets[expert_id] = sfb_base + sfb_stride;
+    d_offsets[expert_id] = d_base + d_stride;
+  }
+};
+
+template <typename GemmTraits>
+struct CutlassMxfp8GroupedMmLayoutFunctor {
+  using Sm1xxBlkScaledConfig = typename GemmTraits::Sm1xxBlkScaledConfig;
+  using LayoutSFA = typename GemmTraits::LayoutSFA;
+  using LayoutSFB = typename GemmTraits::LayoutSFB;
+  LayoutSFA* layout_sfa_base{nullptr};
+  LayoutSFB* layout_sfb_base{nullptr};
+
+  CutlassMxfp8GroupedMmLayoutFunctor() = default;
+  CutlassMxfp8GroupedMmLayoutFunctor(LayoutSFA* _layout_sfa_base,
+                                     LayoutSFB* _layout_sfb_base)
+      : layout_sfa_base(_layout_sfa_base), layout_sfb_base(_layout_sfb_base) {}
+
+  void CUTE_DEVICE operator()(int64_t expert_id, int m, int n, int k) {
+    LayoutSFA* layout_sfa_ptr = layout_sfa_base + expert_id;
+    LayoutSFB* layout_sfb_ptr = layout_sfb_base + expert_id;
+    *layout_sfa_ptr = Sm1xxBlkScaledConfig::tile_atom_to_shape_SFA(
+        cute::make_shape(m, n, k, 1));
+    *layout_sfb_ptr = Sm1xxBlkScaledConfig::tile_atom_to_shape_SFB(
+        cute::make_shape(m, n, k, 1));
+  }
+};
+
+template <typename GemmTraits>
+struct CutlassMxfp8GroupedMmStrideFunctor {
+  using StrideA = typename GemmTraits::StrideA;
+  using StrideB = typename GemmTraits::StrideB;
+  using StrideD = typename GemmTraits::StrideD;
+  StrideA* stride_A_base{nullptr};
+  StrideB* stride_B_base{nullptr};
+  StrideD* stride_D_base{nullptr};
+
+  CutlassMxfp8GroupedMmStrideFunctor() = default;
+  CutlassMxfp8GroupedMmStrideFunctor(StrideA* _stride_A_base,
+                                     StrideB* _stride_B_base,
+                                     StrideD* _stride_D_base)
+      : stride_A_base(_stride_A_base),
+        stride_B_base(_stride_B_base),
+        stride_D_base(_stride_D_base) {}
+
+  void CUTE_DEVICE operator()(int64_t expert_id, int m, int n, int k) {
+    StrideA* stride_A = stride_A_base + expert_id;
+    StrideB* stride_B = stride_B_base + expert_id;
+    StrideD* stride_D = stride_D_base + expert_id;
+    *stride_A = cutlass::make_cute_packed_stride(StrideA{}, {m, k, 1});
+    *stride_B = cutlass::make_cute_packed_stride(StrideB{}, {n, k, 1});
+    *stride_D = cutlass::make_cute_packed_stride(StrideD{}, {m, n, 1});
+  }
+};
+
+template <typename OffsetFunctor, typename LayoutFunctor,
+          typename StrideFunctor>
+__global__ void cutlassMxfp8GroupedMmPreComputeKernel(
+    int* problem_sizes, OffsetFunctor offset_functor,
+    LayoutFunctor layout_functor, StrideFunctor stride_functor) {
+  int64_t expert_id = static_cast<int64_t>(threadIdx.x);
+  int m = problem_sizes[expert_id * 3 + 0];
+  int n = problem_sizes[expert_id * 3 + 1];
+  int k = problem_sizes[expert_id * 3 + 2];
+
+  offset_functor(expert_id, m, n, k);
+  layout_functor(expert_id, m, n, k);
+  stride_functor(expert_id, m, n, k);
+}
+
+}  // namespace expert_specialization

csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm_launcher.cuh

@@ -0,0 +1,179 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+// Adapted from SGLang:
+// https://github.com/sgl-project/sglang/blob/ded068a76e00878881d52d5bfb791e0f60d7311b/sgl-kernel/csrc/expert_specialization/es_sm100_mxfp8_blockscaled_launcher.cuh
+
+#pragma once
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <torch/all.h>
+
+#include <cassert>
+#include <iostream>
+#include <string>
+
+#include "cute/tensor.hpp"
+#include "cutlass_mxfp8_grouped_mm_functor.cuh"
+#include "cutlass_mxfp8_grouped_mm_traits.cuh"
+
+namespace expert_specialization {
+
+template <typename GemmTraits>
+void cutlass_mxfp8_grouped_mm_pre_compute(
+    torch::Tensor& a_ptrs, torch::Tensor& b_ptrs, torch::Tensor& sfa_ptrs,
+    torch::Tensor& sfb_ptrs, torch::Tensor& d_ptrs, torch::Tensor& stride_a,
+    torch::Tensor& stride_b, torch::Tensor& stride_d, torch::Tensor& layout_sfa,
+    torch::Tensor& layout_sfb, const torch::Tensor& a, const torch::Tensor& b,
+    const torch::Tensor& sfa, const torch::Tensor& sfb, const torch::Tensor& d,
+    const torch::Tensor& problem_sizes, const torch::Tensor& expert_offsets,
+    const torch::Tensor& blockscale_offsets, cudaStream_t stream) {
+  using OffsetFunctor = CutlassMxfp8GroupedMmOffsetFunctor<GemmTraits>;
+  using ElementA = typename OffsetFunctor::ElementA;
+  using ElementB = typename OffsetFunctor::ElementB;
+  using ElementSF = typename OffsetFunctor::ElementSF;
+  using ElementD = typename OffsetFunctor::ElementD;
+
+  using LayoutFunctor = CutlassMxfp8GroupedMmLayoutFunctor<GemmTraits>;
+  using LayoutSFA = typename LayoutFunctor::LayoutSFA;
+  using LayoutSFB = typename LayoutFunctor::LayoutSFB;
+
+  using StrideFunctor = CutlassMxfp8GroupedMmStrideFunctor<GemmTraits>;
+  using StrideA = typename StrideFunctor::StrideA;
+  using StrideB = typename StrideFunctor::StrideB;
+  using StrideD = typename StrideFunctor::StrideD;
+
+  int num_experts = (int)expert_offsets.size(0);
+  TORCH_CHECK(num_experts <= 1024,
+              "Number of experts cannot exceed 1024, the maximum number of "
+              "threads per block.");
+
+  OffsetFunctor offset_functor(
+      reinterpret_cast<int*>(expert_offsets.data_ptr()),
+      reinterpret_cast<int*>(blockscale_offsets.data_ptr()),
+      reinterpret_cast<ElementA*>(a.data_ptr()),
+      reinterpret_cast<ElementB*>(b.data_ptr()),
+      reinterpret_cast<ElementSF*>(sfa.data_ptr()),
+      reinterpret_cast<ElementSF*>(sfb.data_ptr()),
+      reinterpret_cast<ElementD*>(d.data_ptr()),
+      reinterpret_cast<ElementA**>(a_ptrs.data_ptr()),
+      reinterpret_cast<ElementB**>(b_ptrs.data_ptr()),
+      reinterpret_cast<ElementSF**>(sfa_ptrs.data_ptr()),
+      reinterpret_cast<ElementSF**>(sfb_ptrs.data_ptr()),
+      reinterpret_cast<ElementD**>(d_ptrs.data_ptr()));
+  LayoutFunctor layout_functor(
+      reinterpret_cast<LayoutSFA*>(layout_sfa.data_ptr()),
+      reinterpret_cast<LayoutSFB*>(layout_sfb.data_ptr()));
+  StrideFunctor stride_functor(reinterpret_cast<StrideA*>(stride_a.data_ptr()),
+                               reinterpret_cast<StrideB*>(stride_b.data_ptr()),
+                               reinterpret_cast<StrideD*>(stride_d.data_ptr()));
+  cutlassMxfp8GroupedMmPreComputeKernel<<<1, num_experts, 0, stream>>>(
+      static_cast<int*>(problem_sizes.data_ptr()), offset_functor,
+      layout_functor, stride_functor);
+}
+
+template <typename GemmTraits>
+void cutlass_mxfp8_grouped_mm(
+    const torch::Tensor& a_ptrs, const torch::Tensor& b_ptrs,
+    const torch::Tensor& sfa_ptrs, const torch::Tensor& sfb_ptrs,
+    const torch::Tensor& d_ptrs, const torch::Tensor& stride_a,
+    const torch::Tensor& stride_b, const torch::Tensor& stride_d,
+    const torch::Tensor& layout_sfa, const torch::Tensor& layout_sfb,
+    const torch::Tensor& problem_sizes, cudaStream_t stream) {
+  using Gemm = typename GemmTraits::Gemm;
+  using ElementA = typename Gemm::ElementA;
+  using ElementB = typename Gemm::ElementB;
+  using ElementSF = typename GemmTraits::ElementSF;
+  using ElementD = typename GemmTraits::ElementOutput;
+  using StrideA = typename GemmTraits::StrideA;
+  using StrideB = typename GemmTraits::StrideB;
+  using StrideD = typename GemmTraits::StrideD;
+  using LayoutSFA = typename GemmTraits::LayoutSFA;
+  using LayoutSFB = typename GemmTraits::LayoutSFB;
+  using UnderlyingProblemShape =
+      typename GemmTraits::ProblemShape::UnderlyingProblemShape;
+
+  cutlass::KernelHardwareInfo hw_info;
+  hw_info.device_id = c10::cuda::current_device();
+  hw_info.sm_count =
+      at::cuda::getCurrentDeviceProperties()->multiProcessorCount;
+  hw_info.cluster_shape = GemmTraits::MMAConfig::preferred_cluster;
+  hw_info.cluster_shape_fallback = GemmTraits::MMAConfig::fallback_cluster;
+
+  int num_experts = (int)problem_sizes.size(0);
+
+  UnderlyingProblemShape* underlying_problem_shape =
+      reinterpret_cast<UnderlyingProblemShape*>(problem_sizes.data_ptr());
+
+  typename Gemm::Arguments arguments = {
+      cutlass::gemm::GemmUniversalMode::kGrouped,
+      {num_experts, underlying_problem_shape, nullptr},
+      {reinterpret_cast<const ElementA**>(a_ptrs.data_ptr()),
+       reinterpret_cast<StrideA*>(stride_a.data_ptr()),
+       reinterpret_cast<const ElementB**>(b_ptrs.data_ptr()),
+       reinterpret_cast<StrideB*>(stride_b.data_ptr()),
+       reinterpret_cast<const ElementSF**>(sfa_ptrs.data_ptr()),
+       reinterpret_cast<LayoutSFA*>(layout_sfa.data_ptr()),
+       reinterpret_cast<const ElementSF**>(sfb_ptrs.data_ptr()),
+       reinterpret_cast<LayoutSFB*>(layout_sfb.data_ptr())},
+      {{},
+       nullptr,
+       nullptr,
+       reinterpret_cast<ElementD**>(d_ptrs.data_ptr()),
+       reinterpret_cast<StrideD*>(stride_d.data_ptr())},
+      hw_info,
+      {}  // Scheduler
+  };
+
+  Gemm gemm;
+
+  auto can_implement_status = gemm.can_implement(arguments);
+  TORCH_CHECK(can_implement_status == cutlass::Status::kSuccess,
+              "Failed to implement GEMM");
+
+  torch::TensorOptions options_uint8 =
+      torch::TensorOptions().dtype(torch::kUInt8).device(d_ptrs.device());
+  size_t workspace_size = gemm.get_workspace_size(arguments);
+  torch::Tensor workspace = torch::empty(workspace_size, options_uint8);
+
+  auto status = gemm.initialize(arguments, workspace.data_ptr(), stream);
+  TORCH_CHECK(status == cutlass::Status::kSuccess, "Failed to initialize GEMM");
+
+  status = gemm.run(stream, nullptr, true);  // Enable PDL
+  TORCH_CHECK(status == cutlass::Status::kSuccess, "Failed to run GEMM");
+}
+
+template <typename OutType>
+void cutlass_mxfp8_grouped_mm_dispatch_out_dtype(
+    const torch::Tensor& a, const torch::Tensor& b, const torch::Tensor& sfa,
+    const torch::Tensor& sfb, torch::Tensor& d,
+    const torch::Tensor& problem_sizes, const torch::Tensor& expert_offsets,
+    const torch::Tensor& blockscale_offsets, cudaStream_t stream) {
+  int num_experts = (int)problem_sizes.size(0);
+  torch::TensorOptions options_int64 =
+      torch::TensorOptions().dtype(torch::kInt64).device(a.device());
+  torch::TensorOptions options_int32 =
+      torch::TensorOptions().dtype(torch::kInt32).device(a.device());
+
+  torch::Tensor a_ptrs = torch::empty(num_experts, options_int64);
+  torch::Tensor b_ptrs = torch::empty(num_experts, options_int64);
+  torch::Tensor sfa_ptrs = torch::empty(num_experts, options_int64);
+  torch::Tensor sfb_ptrs = torch::empty(num_experts, options_int64);
+  torch::Tensor d_ptrs = torch::empty(num_experts, options_int64);
+
+  torch::Tensor stride_a = torch::empty(num_experts, options_int64);
+  torch::Tensor stride_b = torch::empty(num_experts, options_int64);
+  torch::Tensor stride_d = torch::empty(num_experts, options_int64);
+  torch::Tensor layout_sfa = torch::empty({num_experts, 5}, options_int32);
+  torch::Tensor layout_sfb = torch::empty({num_experts, 5}, options_int32);
+
+  using GemmTraits = CutlassMxfp8GroupedMmGemmTraits<MMA1SMConfig, OutType>;
+  cutlass_mxfp8_grouped_mm_pre_compute<GemmTraits>(
+      a_ptrs, b_ptrs, sfa_ptrs, sfb_ptrs, d_ptrs, stride_a, stride_b, stride_d,
+      layout_sfa, layout_sfb, a, b, sfa, sfb, d, problem_sizes, expert_offsets,
+      blockscale_offsets, stream);
+  cutlass_mxfp8_grouped_mm<GemmTraits>(
+      a_ptrs, b_ptrs, sfa_ptrs, sfb_ptrs, d_ptrs, stride_a, stride_b, stride_d,
+      layout_sfa, layout_sfb, problem_sizes, stream);
+}
+
+}  // namespace expert_specialization

csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm_traits.cuh

@@ -0,0 +1,127 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+// Adapted from SGLang:
+// https://github.com/sgl-project/sglang/blob/ded068a76e00878881d52d5bfb791e0f60d7311b/sgl-kernel/csrc/expert_specialization/es_sm100_mxfp8_blockscaled_traits.cuh
+
+#pragma once
+
+// Misc
+#include "cute/tensor.hpp"
+#include "cutlass/arch/arch.h"
+#include "cutlass/arch/mma.h"
+#include "cutlass/cutlass.h"
+#include "cutlass/detail/sm100_blockscaled_layout.hpp"
+#include "cutlass/epilogue/dispatch_policy.hpp"
+#include "cutlass/gemm/dispatch_policy.hpp"
+#include "cutlass/gemm/group_array_problem_shape.hpp"
+#include "cutlass/layout/layout.h"
+#include "cutlass/numeric_conversion.h"
+#include "cutlass/numeric_size.h"
+
+// Collective Builder
+#include "cutlass/epilogue/collective/collective_builder.hpp"
+#include "cutlass/epilogue/fusion/sm90_callbacks_tma_warpspecialized.hpp"
+#include "cutlass/epilogue/thread/activation.h"
+#include "cutlass/gemm/collective/collective_builder.hpp"
+
+// Integration
+#include "cutlass/gemm/device/gemm_universal_adapter.h"
+#include "cutlass/gemm/kernel/gemm_universal.hpp"
+
+namespace expert_specialization {
+
+using namespace cute;
+
+// Different configs for 1SM and 2SM MMA kernel
+struct MMA1SMConfig {
+  using MmaTileShape = Shape<_128, _128, _128>;
+  using KernelSchedule =
+      cutlass::gemm::KernelPtrArrayTmaWarpSpecialized1SmMxf8f6f4Sm100;
+  using EpilogueSchedule = cutlass::epilogue::PtrArrayTmaWarpSpecialized1Sm;
+  const static dim3 preferred_cluster;
+  const static dim3 fallback_cluster;
+};
+const dim3 MMA1SMConfig::preferred_cluster(1, 4, 1);
+const dim3 MMA1SMConfig::fallback_cluster(1, 2, 1);
+
+template <typename _MMAConfig, typename OutputDtype>
+struct CutlassMxfp8GroupedMmGemmTraits {
+  using MMAConfig = _MMAConfig;
+  using ElementInput = cutlass::float_e4m3_t;
+  using ElementOutput = OutputDtype;
+  using ProblemShape = cutlass::gemm::GroupProblemShape<Shape<int, int, int>>;
+
+  // A matrix configuration
+  using ElementA = cutlass::mx_float8_t<ElementInput>;
+  using LayoutA = cutlass::layout::RowMajor;
+  constexpr static int AlignmentA = 32;
+
+  // B matrix configuration
+  using ElementB = cutlass::mx_float8_t<ElementInput>;
+  using LayoutB = cutlass::layout::ColumnMajor;
+  constexpr static int AlignmentB = 32;
+
+  // C/D matrix configuration
+  using ElementC = void;
+  using ElementD = ElementOutput;
+  using LayoutC = cutlass::layout::RowMajor;
+  using LayoutD = cutlass::layout::RowMajor;
+  constexpr static int AlignmentC = 128 / cutlass::sizeof_bits<ElementD>::value;
+  constexpr static int AlignmentD = 128 / cutlass::sizeof_bits<ElementD>::value;
+  using ElementAccumulator = float;
+
+  static constexpr auto RoundStyle = cutlass::FloatRoundStyle::round_to_nearest;
+  using CustomEVTIdentity =  // acc
+      cutlass::epilogue::fusion::Sm90EVT<
+          cutlass::epilogue::fusion::Sm90Compute<
+              cutlass::epilogue::thread::Identity, ElementD, ElementAccumulator,
+              RoundStyle>,
+          cutlass::epilogue::fusion::Sm90AccFetch>;
+
+  // Core kernel configurations
+  using ArchTag = cutlass::arch::Sm100;
+  using OperatorClass = cutlass::arch::OpClassBlockScaledTensorOp;
+  using StageCountType = cutlass::gemm::collective::StageCountAuto;
+
+  // Runtime Cluster Shape
+  using ClusterShape = Shape<int32_t, int32_t, _1>;
+
+  // Define Epilogue
+  using CollectiveEpilogue =
+      typename cutlass::epilogue::collective::CollectiveBuilder<
+          ArchTag, OperatorClass, typename MMAConfig::MmaTileShape,
+          ClusterShape, Shape<_64, _64>, ElementAccumulator, ElementAccumulator,
+          ElementC, LayoutC*, AlignmentC, ElementD, LayoutD*, AlignmentD,
+          typename MMAConfig::EpilogueSchedule,
+          CustomEVTIdentity>::CollectiveOp;
+
+  // Define Mainloop
+  using CollectiveMainloop =
+      typename cutlass::gemm::collective::CollectiveBuilder<
+          ArchTag, OperatorClass, ElementA, LayoutA*, AlignmentA, ElementB,
+          LayoutB*, AlignmentB, ElementAccumulator,
+          typename MMAConfig::MmaTileShape, ClusterShape,
+          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
+              sizeof(typename CollectiveEpilogue::SharedStorage))>,
+          typename MMAConfig::KernelSchedule>::CollectiveOp;
+
+  // Define GemmKernel
+  using GemmKernel =
+      cutlass::gemm::kernel::GemmUniversal<ProblemShape, CollectiveMainloop,
+                                           CollectiveEpilogue>;
+  using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
+
+  using ElementSF = typename Gemm::GemmKernel::ElementSF;
+  using StrideA = typename Gemm::GemmKernel::InternalStrideA;
+  using StrideB = typename Gemm::GemmKernel::InternalStrideB;
+  using StrideC = typename Gemm::GemmKernel::InternalStrideC;
+  using StrideD = typename Gemm::GemmKernel::InternalStrideD;
+  using LayoutSFA =
+      typename Gemm::GemmKernel::CollectiveMainloop::InternalLayoutSFA;
+  using LayoutSFB =
+      typename Gemm::GemmKernel::CollectiveMainloop::InternalLayoutSFB;
+  using Sm1xxBlkScaledConfig =
+      typename Gemm::GemmKernel::CollectiveMainloop::Sm1xxBlkScaledConfig;
+};
+
+}  // namespace expert_specialization

csrc/moe/mxfp8_moe/mxfp8_experts_quant.cu

@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+// Adapted from SGLang:
+// https://github.com/sgl-project/sglang/blob/ded068a76e00878881d52d5bfb791e0f60d7311b/sgl-kernel/csrc/expert_specialization/es_sm100_mxfp8_blockscaled_group_quant.cu
+
+#include <torch/all.h>
+
+#include "mxfp8_experts_quant.cuh"
+
+void mxfp8_experts_quant(const torch::Tensor& input,
+                         const torch::Tensor& problem_sizes,
+                         const torch::Tensor& expert_offsets,
+                         const torch::Tensor& blockscale_offsets,
+                         torch::Tensor& quant_output,
+                         torch::Tensor& scale_factor) {
+#if defined(CUTLASS_ARCH_MMA_SM100_SUPPORTED)
+  TORCH_CHECK(input.dim() == 2, "input must be 2D tensor");
+  TORCH_CHECK(input.size(1) % 128 == 0, "k must align to 128");
+  TORCH_CHECK(input.strides()[1] == 1, "input must be row major");
+  TORCH_CHECK(problem_sizes.dim() == 2, "problem_sizes must be 2D tensor");
+  TORCH_CHECK(problem_sizes.dtype() == torch::kInt32,
+              "problem_sizes must be int32");
+  TORCH_CHECK(expert_offsets.dtype() == torch::kInt32,
+              "expert_offsets must be int32");
+  TORCH_CHECK(blockscale_offsets.dtype() == torch::kInt32,
+              "blockscale_offsets must be int32");
+
+  auto groups = problem_sizes.size(0);
+  TORCH_CHECK(
+      expert_offsets.dim() == 1 && expert_offsets.size(0) == groups,
+      "expert_offsets must be 1D and have size equal to the number of groups");
+  TORCH_CHECK(
+      blockscale_offsets.dim() == 1 && blockscale_offsets.size(0) == groups,
+      "blockscale_offsets must be 1D and have size equal to the number of "
+      "groups");
+
+  auto stream = at::cuda::getCurrentCUDAStream();
+  if (input.dtype() == torch::kBFloat16) {
+    expert_specialization::launch_mxfp8_experts_quant<__nv_bfloat16>(
+        input, problem_sizes, expert_offsets, blockscale_offsets, quant_output,
+        scale_factor);
+  } else if (input.dtype() == torch::kFloat16) {
+    expert_specialization::launch_mxfp8_experts_quant<__half>(
+        input, problem_sizes, expert_offsets, blockscale_offsets, quant_output,
+        scale_factor);
+  } else {
+    TORCH_CHECK(false, "dtype must be kFloat16 or kBFloat16");
+  }
+#else
+  TORCH_CHECK(false,
+              "No implemented mxfp8_experts_quant for "
+              "current device");
+#endif
+}
+
+#include "core/registration.h"
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("mxfp8_experts_quant", mxfp8_experts_quant);
+}

csrc/moe/mxfp8_moe/mxfp8_experts_quant.cuh

@@ -0,0 +1,414 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+// Adapted from SGLang:
+// https://github.com/sgl-project/sglang/blob/ded068a76e00878881d52d5bfb791e0f60d7311b/sgl-kernel/csrc/expert_specialization/es_sm100_mxfp8_blockscaled_group_quant.cuh
+
+#pragma once
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <cuda.h>
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+#include <torch/all.h>
+
+#include <cuda/ptx>
+
+#include "cute/tensor.hpp"
+
+namespace expert_specialization {
+
+using namespace cute;
+
+constexpr uint32_t THREAD_BLOCK_SIZE = 128;
+constexpr uint32_t WARP_SIZE = 32;
+constexpr int BLOCK_M = 128;
+constexpr int BLOCK_K = 128;
+using ThrLayout = Layout<Shape<_16, _8>, Stride<_8, _1>>;
+using ValLayout = Layout<Shape<_1, _16>>;
+using SfR2SThrLayout = Layout<Shape<_16, _4>, Stride<_4, _1>>;
+using SfR2SValLayout = Layout<Shape<_1, _1>>;
+using ScaleFactorTileLayout =
+    Layout<Shape<Shape<_32, _4>, _4>, Stride<Stride<_16, _4>, _1>>;
+
+// Fast reciprocal.
+inline __device__ float reciprocal_approximate_ftz(float a) {
+  float b;
+  asm volatile("rcp.approx.ftz.f32 %0, %1;\n" : "=f"(b) : "f"(a));
+  return b;
+}
+
+// Some code references TRT-LLM:
+// https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/kernels/quantization.cuh
+template <typename FragmentS, typename FragmentD>
+__inline__ __device__ uint8_t cvt_warp_fp16_to_mxfp8(FragmentS& fragment_s,
+                                                     FragmentD& fragment_d) {
+  using FragmentSLayout = typename FragmentS::layout_type;
+  using FragmentDLayout = typename FragmentD::layout_type;
+  FragmentSLayout fragment_s_layout;
+  FragmentDLayout fragment_d_layout;
+  static_assert(is_static<FragmentSLayout>::value &&
+                size(fragment_s_layout) == 16);
+  static_assert(is_static<FragmentDLayout>::value &&
+                size(fragment_d_layout) == 16);
+
+  constexpr int eles_per_thr = 16;
+  using ValType = typename FragmentS::element_type;
+  using VecType = std::conditional_t<std::is_same_v<ValType, __nv_bfloat16>,
+                                     __nv_bfloat162, __half2>;
+  VecType vec[8];
+  // Assign vals
+  vec[0].x = fragment_s(Int<0>{});
+  vec[0].y = fragment_s(Int<1>{});
+  vec[1].x = fragment_s(Int<2>{});
+  vec[1].y = fragment_s(Int<3>{});
+  vec[2].x = fragment_s(Int<4>{});
+  vec[2].y = fragment_s(Int<5>{});
+  vec[3].x = fragment_s(Int<6>{});
+  vec[3].y = fragment_s(Int<7>{});
+  vec[4].x = fragment_s(Int<8>{});
+  vec[4].y = fragment_s(Int<9>{});
+  vec[5].x = fragment_s(Int<10>{});
+  vec[5].y = fragment_s(Int<11>{});
+  vec[6].x = fragment_s(Int<12>{});
+  vec[6].y = fragment_s(Int<13>{});
+  vec[7].x = fragment_s(Int<14>{});
+  vec[7].y = fragment_s(Int<15>{});
+
+  auto local_max = __habs2(vec[0]);
+  for (int i = 1; i < eles_per_thr / 2; i++) {
+    local_max = __hmax2(__habs2(vec[i]), local_max);
+  }
+  local_max = __hmax2(__shfl_xor_sync(uint32_t(-1), local_max, 1), local_max);
+
+  // Get the final absolute maximum values.
+  float block_max(0.0f);
+  if constexpr (std::is_same_v<ValType, __nv_bfloat16>) {
+    block_max = __bfloat162float(__hmax(local_max.x, local_max.y));
+  } else {
+    block_max = __half2float(__hmax(local_max.x, local_max.y));
+  }
+  // Get the SF (max value of the vector / max value of mxfp8).
+  float sf_val = block_max * reciprocal_approximate_ftz(448.0f);
+  // 8 bits representation of the SF.
+  uint8_t fp8_sf_val;
+
+  __nv_fp8_e8m0 tmp_sf_val;
+  tmp_sf_val.__x =
+      __nv_cvt_float_to_e8m0(sf_val, __NV_SATFINITE, cudaRoundPosInf);
+  sf_val = static_cast<float>(tmp_sf_val);
+  fp8_sf_val = tmp_sf_val.__x;
+  // Get the output scale (reciprocal of the SFValue).
+  float output_scale =
+      block_max != 0.f ? reciprocal_approximate_ftz(sf_val) : 0.0f;
+
+  // Convert the input to float.
+  float2 fp2_vals[eles_per_thr / 2];
+
+#pragma unroll
+  for (int i = 0; i < eles_per_thr / 2; i++) {
+    if constexpr (std::is_same_v<ValType, __half>) {
+      fp2_vals[i] = __half22float2(vec[i]);
+    } else {
+      fp2_vals[i] = __bfloat1622float2(vec[i]);
+    }
+    fp2_vals[i].x *= output_scale;
+    fp2_vals[i].y *= output_scale;
+  }
+  union {
+    uint8_t bytes[16];
+    __nv_fp8x2_e4m3 elts[8];
+  } u;
+  u.elts[0] = __nv_fp8x2_e4m3(fp2_vals[0]);
+  u.elts[1] = __nv_fp8x2_e4m3(fp2_vals[1]);
+  u.elts[2] = __nv_fp8x2_e4m3(fp2_vals[2]);
+  u.elts[3] = __nv_fp8x2_e4m3(fp2_vals[3]);
+  u.elts[4] = __nv_fp8x2_e4m3(fp2_vals[4]);
+  u.elts[5] = __nv_fp8x2_e4m3(fp2_vals[5]);
+  u.elts[6] = __nv_fp8x2_e4m3(fp2_vals[6]);
+  u.elts[7] = __nv_fp8x2_e4m3(fp2_vals[7]);
+  fragment_d(Int<0>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[0]);
+  fragment_d(Int<1>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[1]);
+  fragment_d(Int<2>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[2]);
+  fragment_d(Int<3>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[3]);
+  fragment_d(Int<4>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[4]);
+  fragment_d(Int<5>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[5]);
+  fragment_d(Int<6>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[6]);
+  fragment_d(Int<7>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[7]);
+  fragment_d(Int<8>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[8]);
+  fragment_d(Int<9>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[9]);
+  fragment_d(Int<10>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[10]);
+  fragment_d(Int<11>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[11]);
+  fragment_d(Int<12>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[12]);
+  fragment_d(Int<13>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[13]);
+  fragment_d(Int<14>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[14]);
+  fragment_d(Int<15>{}) = cutlass::float_e4m3_t::bitcast(u.bytes[15]);
+  return fp8_sf_val;
+}
+
+template <typename TensorS, typename TensorP, typename TensorD,
+          typename TensorSharedSF, typename TensorSF, typename TiledCopyG2R,
+          typename TiledCopyR2G, typename TiledCopyR2S>
+__inline__ __device__ void mxfp8_experts_quant_tile(
+    TensorS& tensor_s, TensorP& tensor_p, TensorD& tensor_d,
+    TensorSharedSF& tensor_shared_sf, TensorSF& tensor_sf, int m,
+    TiledCopyG2R& tiled_copy_g2r, TiledCopyR2G& tiled_copy_r2g,
+    TiledCopyR2S& tiled_copy_r2s) {
+  static_assert(size(get<0>(typename TensorS::layout_type{})) == 128 &&
+                size(get<1>(typename TensorS::layout_type{})) == 128 &&
+                stride(get<1>(typename TensorS::layout_type{})) == 1);
+  static_assert(size(get<0>(typename TensorD::layout_type{})) == 128 &&
+                size(get<1>(typename TensorD::layout_type{})) == 128 &&
+                stride(get<1>(typename TensorD::layout_type{})) == 1);
+  static_assert(size(get<0>(typename TensorP::layout_type{})) == 128 &&
+                size(get<1>(typename TensorP::layout_type{})) == 128);
+  static_assert(size(get<0>(typename TensorSharedSF::layout_type{})) == 128 &&
+                size(get<1>(typename TensorSharedSF::layout_type{})) == 4);
+  static_assert(size(get<0>(typename TensorSF::layout_type{})) == 128 &&
+                size(get<1>(typename TensorSF::layout_type{})) == 4);
+
+  using Tiler_MN = typename TiledCopyG2R::Tiler_MN;
+  auto tiler_mn = Tiler_MN{};
+  static_assert(size<0>(tiler_mn) == 16 && size<1>(tiler_mn) == 128);
+
+  auto tiled_tensor_s = tiled_divide(tensor_s, tiler_mn);
+  auto tiled_tensor_p = tiled_divide(tensor_p, tiler_mn);
+  auto tiled_tensor_d = tiled_divide(tensor_d, tiler_mn);
+  static_assert(size<2>(tiled_tensor_s) == 1);
+  static_assert(size<2>(tiled_tensor_p) == 1);
+  static_assert(size<2>(tiled_tensor_d) == 1);
+  auto squeeze_tiled_tensor_s = take<0, 2>(tiled_tensor_s);
+  auto squeeze_tiled_tensor_p = take<0, 2>(tiled_tensor_p);
+  auto squeeze_tiled_tensor_d = take<0, 2>(tiled_tensor_d);
+
+  using SF_Tiler_MN = typename TiledCopyR2S::Tiler_MN;
+  auto sf_tiler_mn = SF_Tiler_MN{};
+  static_assert(size<0>(sf_tiler_mn) == 16 && size<1>(sf_tiler_mn) == 4);
+
+  auto tiled_tensor_sf = tiled_divide(tensor_sf, sf_tiler_mn);
+  auto tiled_tensor_shared_sf = tiled_divide(tensor_shared_sf, sf_tiler_mn);
+  auto squeeze_tiled_tensor_sf = take<0, 2>(tiled_tensor_sf);
+  auto squeeze_tiled_tensor_shared_sf = take<0, 2>(tiled_tensor_shared_sf);
+
+  constexpr int tile_loop_count = size<1>(tiled_tensor_s);
+  constexpr int rows_in_tile = 16;
+  // We don't need to clear shared memory
+  // clear(squeeze_tiled_tensor_shared_sf);
+#pragma unroll 4
+  for (int t = 0; t < tile_loop_count; t++) {
+    if (t * rows_in_tile >= m) {
+      break;
+    }
+    auto current_copy_tile_s = tensor<0>(squeeze_tiled_tensor_s(_, t));
+    auto current_copy_tile_p = tensor<0>(squeeze_tiled_tensor_p(_, t));
+    auto current_copy_tile_d = tensor<0>(squeeze_tiled_tensor_d(_, t));
+    auto current_copy_tile_sf = tensor<0>(squeeze_tiled_tensor_sf(_, t));
+    auto current_copy_tile_shared_sf =
+        tensor<0>(squeeze_tiled_tensor_shared_sf(_, t));
+
+    // Global to Register copy
+    auto thr_copy_g2r = tiled_copy_g2r.get_thread_slice(threadIdx.x);
+    auto thr_tile_g2r_s = thr_copy_g2r.partition_S(current_copy_tile_s);
+    auto thr_tile_g2r_p = thr_copy_g2r.partition_S(current_copy_tile_p);
+    auto input_fragment = make_fragment_like(thr_tile_g2r_s);
+
+    // Register to Global copy
+    auto thr_copy_r2g = tiled_copy_r2g.get_thread_slice(threadIdx.x);
+    auto thr_tile_r2g_d = thr_copy_r2g.partition_D(current_copy_tile_d);
+    auto thr_tile_r2g_p = thr_copy_r2g.partition_D(current_copy_tile_p);
+    auto output_fragment = make_fragment_like(thr_tile_r2g_d);
+
+    // Register to Shared copy
+    auto thr_copy_r2s = tiled_copy_r2s.get_thread_slice(threadIdx.x / 2);
+    auto thr_tile_r2s_shared_sf =
+        thr_copy_r2s.partition_D(current_copy_tile_shared_sf);
+    auto shared_sf_fragment = make_fragment_like(thr_tile_r2s_shared_sf);
+
+    // CopyG2R & convert & CopyR2G
+    copy_if(tiled_copy_g2r, thr_tile_g2r_p, thr_tile_g2r_s, input_fragment);
+    uint8_t fp8_sf_val =
+        cvt_warp_fp16_to_mxfp8(input_fragment, output_fragment);
+    copy_if(tiled_copy_r2g, thr_tile_r2g_p, output_fragment, thr_tile_r2g_d);
+    shared_sf_fragment[0] = fp8_sf_val;
+
+    // Before first copy r2s, clear shared memory and wait previous group
+    if (t == 0 && threadIdx.x == 0) {
+      // Wait for the group to have completed reading from shared memory.
+      cuda::ptx::cp_async_bulk_wait_group_read(cuda::ptx::n32_t<0>());
+    }
+    __syncthreads();
+
+    if (threadIdx.x % 2 == 0) {
+      copy(tiled_copy_r2s, shared_sf_fragment, thr_tile_r2s_shared_sf);
+    }
+    __syncthreads();
+  }
+
+  // Wait for shared memory writes to be visible to TMA engine.
+  cuda::ptx::fence_proxy_async(cuda::ptx::space_shared);  // b)
+  __syncthreads();
+
+  if (threadIdx.x == 0) {
+    cuda::ptx::cp_async_bulk(cuda::ptx::space_global, cuda::ptx::space_shared,
+                             squeeze_tiled_tensor_sf.data().get(),
+                             squeeze_tiled_tensor_shared_sf.data().get(), 512);
+    // Wait for TMA transfer to have finished reading shared memory.
+    // Create a "bulk async-group" out of the previous bulk copy operation.
+    cuda::ptx::cp_async_bulk_commit_group();
+  }
+  __syncthreads();
+}
+
+template <typename T_IN, typename TiledCopyG2R, typename TiledCopyR2G,
+          typename TiledCopyR2S>
+__global__ void mxfp8_experts_quant_kernel(
+    const T_IN* input, const int* problem_sizes, const int* expert_offsets,
+    const int* blockscale_offsets, cutlass::float_e4m3_t* quant_output,
+    uint8_t* scale_factor, int groups, TiledCopyG2R tiled_copy_g2r,
+    TiledCopyR2G tiled_copy_r2g, TiledCopyR2S tiled_copy_r2s) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000
+  __shared__ __align__(512) uint8_t shared_memory[512];
+  ScaleFactorTileLayout scale_factor_tile_layout{};
+  auto scale_factor_shared =
+      make_tensor(make_smem_ptr(shared_memory),
+                  scale_factor_tile_layout);  // ((_32,_4), _4):((_16,_4), _1)
+  // TODO: Transform Groupwise Schedule into a more efficient Schedule
+  for (int g = 0; g < groups; g++) {
+    int m = problem_sizes[g * 3 + 0];
+    int k = problem_sizes[g * 3 + 2];
+    int64_t expert_offset = static_cast<int64_t>(expert_offsets[g]);
+    int64_t blockscale_offset = static_cast<int64_t>(blockscale_offsets[g]);
+
+    auto input_tensor = make_tensor(
+        make_gmem_ptr(input + expert_offset * k),
+        make_layout(make_shape(m, k),
+                    LayoutRight{}));  // (M, K):(K, 1) half_t/bfloat16_t
+
+    auto quant_output_tensor = make_tensor(
+        make_gmem_ptr(quant_output + expert_offset * k),
+        make_layout(make_shape(m, k),
+                    LayoutRight{}));  // (M, K):(K, 1) cutlass::float_e4m3_t
+
+    auto scale_factor_shape = make_shape(ceil_div(m, 128) * 128, k / 32);
+    auto scale_factor_layout = tile_to_shape(scale_factor_tile_layout,
+                                             scale_factor_shape, LayoutRight{});
+    // layout<0>(layout<0>(scale_factor_layout))  (_32,_4):(_16,_4) -- static
+    // layout<1>(layout<0>(scale_factor_layout))  M_align_128 / 128 -- dynamic
+    // shape dynamic stride layout<0>(layout<1>(scale_factor_layout))  _4:_1 --
+    // static layout<1>(layout<1>(scale_factor_layout))  (K / 32) / 4 : _512 --
+    // dynamic shape static stride
+
+    // Reshape to zipped layout for 1D indexing
+    auto zipped_scale_factor_layout = make_layout(
+        make_layout(layout<0>(layout<0>(scale_factor_layout)),
+                    layout<0>(layout<1>(scale_factor_layout))),
+        make_layout(
+            layout<1>(layout<0>(scale_factor_layout)),
+            layout<1>(layout<1>(
+                scale_factor_layout))));  // (((_32,_4),_4),(M_align_128 /
+                                          // 128,(K / 32) /
+                                          // 4)):(((_16,_4),_1),(?,_512))
+
+    auto scale_factor_tensor =
+        make_tensor(make_gmem_ptr(scale_factor + blockscale_offset * (k / 32)),
+                    zipped_scale_factor_layout);
+
+    // Used for cases where M is not divisible by 128 (most scenarios).
+    auto input_shape = shape(input_tensor);  // (M, K):(K, 1)
+    auto identity_tensor = make_identity_tensor(input_shape);
+    auto predict_tensor = cute::lazy::transform(
+        identity_tensor, [&](auto c) { return elem_less(c, input_shape); });
+
+    // (_128, _128)
+    auto tiler = make_shape(Int<BLOCK_M>{}, Int<BLOCK_K>{});
+
+    auto tiled_input_tensor = zipped_divide(
+        input_tensor, tiler);  // ((128, 128), (cdiv(M, 128), cdiv(K, 128)))
+    auto tiled_quant_output_tensor =
+        zipped_divide(quant_output_tensor,
+                      tiler);  // ((128, 128), (cdiv(M, 128), cdiv(K, 128)))
+    auto tiled_predict_tensor = zipped_divide(
+        predict_tensor, tiler);  // ((128, 128), (cdiv(M, 128), cdiv(K, 128)))
+
+    auto total_tiles =
+        size<1>(tiled_input_tensor);  // cdiv(M, 128) * cdiv(K, 128)
+    decltype(total_tiles) blk_offset = blockIdx.x;
+    while (blk_offset < total_tiles) {
+      auto current_input_tile = tensor<0>(tiled_input_tensor(_, blk_offset));
+      auto current_quant_output_tile =
+          tensor<0>(tiled_quant_output_tensor(_, blk_offset));
+      auto current_predict_tile =
+          tensor<0>(tiled_predict_tensor(_, blk_offset));
+      auto current_scale_factor_tile =
+          tensor<0>(scale_factor_tensor(_, blk_offset));
+
+      mxfp8_experts_quant_tile<
+          decltype(current_input_tile), decltype(current_predict_tile),
+          decltype(current_quant_output_tile), decltype(scale_factor_shared),
+          decltype(current_scale_factor_tile), TiledCopyG2R, TiledCopyR2G,
+          TiledCopyR2S>(current_input_tile, current_predict_tile,
+                        current_quant_output_tile, scale_factor_shared,
+                        current_scale_factor_tile, m, tiled_copy_g2r,
+                        tiled_copy_r2g, tiled_copy_r2s);
+      blk_offset += gridDim.x;
+    }
+  }
+#endif
+}
+
+template <typename T_IN>
+void launch_mxfp8_experts_quant(const torch::Tensor& input,
+                                const torch::Tensor& problem_sizes,
+                                const torch::Tensor& expert_offsets,
+                                const torch::Tensor& blockscale_offsets,
+                                torch::Tensor& quant_output,
+                                torch::Tensor& scale_factor) {
+  ThrLayout thr_layout{};
+  ValLayout val_layout{};
+  SfR2SThrLayout r2s_thr_layout{};
+  SfR2SValLayout r2s_val_layout{};
+
+  using CopyOpG2R =
+      UniversalCopy<cutlass::AlignedArray<T_IN, size(val_layout)>>;
+  using CopyAtomG2R = cute::Copy_Atom<CopyOpG2R, T_IN>;
+  auto tiled_copy_g2r = cute::make_tiled_copy(
+      CopyAtomG2R{}, thr_layout, val_layout);  // Tiler_MN: (16, 128)
+
+  using CopyOpR2G = UniversalCopy<
+      cutlass::AlignedArray<cutlass::float_e4m3_t, size(val_layout)>>;
+  using CopyAtomR2G = cute::Copy_Atom<CopyOpR2G, cutlass::float_e4m3_t>;
+  auto tiled_copy_r2g = cute::make_tiled_copy(
+      CopyAtomR2G{}, thr_layout, val_layout);  // Tiler_MN: (16, 128)
+
+  using CopyOpR2S =
+      UniversalCopy<cutlass::AlignedArray<uint8_t, size(r2s_val_layout)>>;
+  using CopyAtomR2S = cute::Copy_Atom<CopyOpR2S, uint8_t>;
+  auto tiled_copy_r2s = cute::make_tiled_copy(
+      CopyAtomR2S{}, r2s_thr_layout, r2s_val_layout);  // Tiler_MN: (16, 4)
+
+  int max_active_blocks_per_sm = -1;
+  AT_CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(
+      &max_active_blocks_per_sm,
+      mxfp8_experts_quant_kernel<T_IN, decltype(tiled_copy_g2r),
+                                 decltype(tiled_copy_r2g),
+                                 decltype(tiled_copy_r2s)>,
+      THREAD_BLOCK_SIZE, 0));
+
+  dim3 grid(at::cuda::getCurrentDeviceProperties()->multiProcessorCount *
+                max_active_blocks_per_sm,
+            1, 1);
+  dim3 block(THREAD_BLOCK_SIZE, 1, 1);
+  int num_experts = (int)problem_sizes.size(0);
+  auto stream = at::cuda::getCurrentCUDAStream();
+  mxfp8_experts_quant_kernel<T_IN, decltype(tiled_copy_g2r),
+                             decltype(tiled_copy_r2g), decltype(tiled_copy_r2s)>
+      <<<grid, block, 0, stream>>>(
+          reinterpret_cast<const T_IN*>(input.data_ptr()),
+          reinterpret_cast<const int*>(problem_sizes.data_ptr()),
+          reinterpret_cast<const int*>(expert_offsets.data_ptr()),
+          reinterpret_cast<const int*>(blockscale_offsets.data_ptr()),
+          reinterpret_cast<cutlass::float_e4m3_t*>(quant_output.data_ptr()),
+          reinterpret_cast<uint8_t*>(scale_factor.data_ptr()), num_experts,
+          tiled_copy_g2r, tiled_copy_r2g, tiled_copy_r2s);
+}
+
+}  // namespace expert_specialization

csrc/quantization/activation_kernels.cu

@@ -542,7 +542,7 @@ __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
       if (!lane_id) {
         // Store scales.
         if constexpr (std::is_same<scale_t, uint8_t>::value) {
-          // Packed UE8MO format. Remove Mantissa.
+          // Packed UE8M0 format. Remove Mantissa.
           *y_s_ptr = reinterpret_cast<int16_t&>(y_s) >> 7;
 
           bool const jump_pack = (current_group_id + 1) % 4 == 0;

csrc/rocm/skinny_gemms.cu

@@ -1476,7 +1476,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   #endif
 
     // B[] staging is cooperative across GrpsShrB, so sync here before reading
-    // back. This wait is currently inserted by compiler, but not gauranteed.
+    // back. This wait is currently inserted by compiler, but not guaranteed.
     asm volatile("s_waitcnt 0");
     __syncthreads();
 

csrc/torch_bindings.cpp

@@ -426,6 +426,22 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       " Tensor problem_sizes, Tensor expert_offsets, Tensor sf_offsets) -> ()");
   // conditionally compiled so impl registration is in source file
 
+  // Expert-specialization mxfp8 blockscaled grouped quantization (SM100+).
+  ops.def(
+      "mxfp8_experts_quant("
+      " Tensor input, Tensor problem_sizes, Tensor expert_offsets,"
+      " Tensor blockscale_offsets, Tensor! quant_output, Tensor! scale_factor)"
+      " -> ()");
+  // conditionally compiled so impl registration is in source file
+
+  // Expert-specialization mxfp8 blockscaled grouped GEMM (SM100+).
+  ops.def(
+      "cutlass_mxfp8_grouped_mm("
+      " Tensor a, Tensor b, Tensor sfa, Tensor sfb, Tensor! out,"
+      " Tensor problem_sizes, Tensor expert_offsets, Tensor blockscale_offsets)"
+      " -> ()");
+  // conditionally compiled so impl registration is in source file
+
   // CUTLASS w8a8 GEMM, supporting symmetric per-tensor or per-row/column
   // quantization, as well as bias
   ops.def(

docker/Dockerfile

@@ -262,7 +262,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 # Build the vLLM wheel
 # if USE_SCCACHE is set, use sccache to speed up compilation
+# AWS credentials mounted at ~/.aws/credentials for sccache S3 auth (optional)
 RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=secret,id=aws-credentials,target=/root/.aws/credentials,required=false \
     if [ "$USE_SCCACHE" = "1" ]; then \
         echo "Installing sccache..." \
         && case "${TARGETPLATFORM}" in \

docker/Dockerfile.xpu

@@ -115,9 +115,57 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # install development dependencies (for testing)
 RUN uv pip install -e tests/vllm_test_utils
 
-# install nixl from source code
-ENV NIXL_VERSION=0.7.0
-RUN python /workspace/vllm/tools/install_nixl_from_source_ubuntu.py
+# install NIXL and UCX from source code
+ARG UCX_VERSION=e5d98879705239d254ede40b4a52891850cb5349
+ARG NIXL_VERSION=0.7.0
+
+RUN apt-get update && apt-get install -y \
+    pciutils \
+    net-tools \
+    iproute2 \
+    hwloc \
+    numactl \
+    wget \
+    curl \
+    git \
+    build-essential \
+    autoconf \
+    automake \
+    libtool \
+    pkg-config \
+    rdma-core \
+    libibverbs-dev \
+    ibverbs-utils \
+    libibverbs1 \
+    librdmacm-dev \
+    librdmacm1 \
+    libibumad-dev \
+    libibumad3 \
+    libibmad-dev \
+    libibmad5 \
+    infiniband-diags \
+    perftest \
+    ibutils \
+    libmlx5-1 \
+    libmlx4-1 \
+    ibverbs-providers \
+    librdmacm1t64
+
+ENV PKG_CONFIG_PATH=/tmp/ucx_install/lib/pkgconfig:${PKG_CONFIG_PATH}
+ENV LD_LIBRARY_PATH=/tmp/ucx_install/lib:${LD_LIBRARY_PATH}
+RUN --mount=type=cache,target=/root/.cache/uv \
+    git clone https://github.com/openucx/ucx /tmp/ucx_source && \
+    cd /tmp/ucx_source && git checkout "${UCX_VERSION}" && \
+    bash autogen.sh && \
+    ./configure --prefix=/tmp/ucx_install --with-ze=yes --enable-examples --enable-mt && \
+    make CFLAGS="-Wno-error=incompatible-pointer-types" -j8 && make install && \
+    git clone https://github.com/ai-dynamo/nixl /tmp/nixl_source && \
+    cd /tmp/nixl_source && git checkout "${NIXL_VERSION}" && \
+    cd /tmp/nixl_source && \
+    uv pip install --upgrade meson pybind11 patchelf && \
+    uv pip install -r requirements.txt && \
+    uv pip install . && \
+    rm -rf /tmp/ucx_source /tmp/nixl_source
 
 # FIX triton
 RUN --mount=type=cache,target=/root/.cache/uv \

docs/benchmarking/cli.md

@@ -25,7 +25,7 @@ th {
 | BurstGPT | ✅ | ✅ | `wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv` |
 | Sonnet (deprecated) | ✅ | ✅ | Local file: `benchmarks/sonnet.txt` |
 | Random | ✅ | ✅ | `synthetic` |
-| RandomMultiModal (Image/Video) | 🟡 | 🚧 | `synthetic` |
+| RandomMultiModal (Image/Video) | ✅ | ✅ | `synthetic` |
 | RandomForReranking | ✅ | ✅ | `synthetic` |
 | Prefix Repetition | ✅ | ✅ | `synthetic` |
 | HuggingFace-VisionArena | ✅ | ✅ | `lmarena-ai/VisionArena-Chat` |
@@ -545,6 +545,24 @@ vllm bench throughput \
   --lora-path yard1/llama-2-7b-sql-lora-test
 ```
 
+#### Synthetic Random Multimodal (random-mm)
+
+Generate synthetic multimodal inputs for offline throughput testing without external datasets.
+Use `--backend vllm-chat` so that image tokens are counted correctly.
+
+```bash
+vllm bench throughput \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name random-mm \
+  --num-prompts 100 \
+  --random-input-len 300 \
+  --random-output-len 40 \
+  --random-mm-base-items-per-request 2 \
+  --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \
+  --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}'
+```
+
 </details>
 
 ### 🛠️ Structured Output Benchmark
@@ -846,8 +864,8 @@ Generate synthetic image inputs alongside random text prompts to stress-test vis
 
 Notes:
 
-- Works only with online benchmark via the OpenAI backend (`--backend openai-chat`) and endpoint `/v1/chat/completions`.
-- Video sampling is not yet implemented.
+- For online benchmarks, use `--backend openai-chat` with endpoint `/v1/chat/completions`.
+- For offline benchmarks, use `--backend vllm-chat` (see [Offline Throughput Benchmark](#-offline-throughput-benchmark) for an example).
 
 Start the server (example):
 
@@ -913,6 +931,74 @@ This should be seen as an edge case, and if this behavior can be avoided by sett
 
 </details>
 
+### 🔬 Multimodal Processor Benchmark
+
+Benchmark per-stage latency of the multimodal (MM) input processor pipeline, including the encoder forward pass. This is useful for profiling preprocessing bottlenecks in vision-language models.
+
+<details class="admonition abstract" markdown="1">
+<summary>Show more</summary>
+
+The benchmark measures the following stages for each request:
+
+| Stage | Description |
+|-------|-------------|
+| `get_mm_hashes_secs` | Time spent hashing multimodal inputs |
+| `get_cache_missing_items_secs` | Time spent looking up the processor cache |
+| `apply_hf_processor_secs` | Time spent in the HuggingFace processor |
+| `merge_mm_kwargs_secs` | Time spent merging multimodal kwargs |
+| `apply_prompt_updates_secs` | Time spent updating prompt tokens |
+| `preprocessor_total_secs` | Total preprocessing time |
+| `encoder_forward_secs` | Time spent in the encoder model forward pass |
+| `num_encoder_calls` | Number of encoder invocations per request |
+
+The benchmark also reports end-to-end latency (TTFT + decode time) per
+request. Use `--metric-percentiles` to select which percentiles to report
+(default: p99) and `--output-json` to save results.
+
+#### Basic Example with Synthetic Data (random-mm)
+
+```bash
+vllm bench mm-processor \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --dataset-name random-mm \
+  --num-prompts 50 \
+  --random-input-len 300 \
+  --random-output-len 40 \
+  --random-mm-base-items-per-request 2 \
+  --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \
+  --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}'
+```
+
+#### Using a HuggingFace Dataset
+
+```bash
+vllm bench mm-processor \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --hf-split train \
+  --num-prompts 100
+```
+
+#### Warmup, Custom Percentiles, and JSON Output
+
+```bash
+vllm bench mm-processor \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --dataset-name random-mm \
+  --num-prompts 200 \
+  --num-warmups 5 \
+  --random-input-len 300 \
+  --random-output-len 40 \
+  --random-mm-base-items-per-request 1 \
+  --metric-percentiles 50,90,95,99 \
+  --output-json results.json
+```
+
+See [`vllm bench mm-processor`](../cli/bench/mm_processor.md) for the full argument reference.
+
+</details>
+
 ### Embedding Benchmark
 
 Benchmark the performance of embedding requests in vLLM.

docs/cli/bench/mm_processor.md

@@ -1,5 +1,51 @@
 # vllm bench mm-processor
 
+## Overview
+
+`vllm bench mm-processor` profiles the multimodal input processor pipeline of
+vision-language models. It measures per-stage latency from the HuggingFace
+processor through to the encoder forward pass, helping you identify
+preprocessing bottlenecks and understand how different image resolutions or
+item counts affect end-to-end request time.
+
+The benchmark supports two data sources: synthetic random multimodal inputs
+(`random-mm`) and HuggingFace datasets (`hf`). Warmup requests are run before
+measurement to ensure stable results.
+
+## Quick Start
+
+```bash
+vllm bench mm-processor \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --dataset-name random-mm \
+  --num-prompts 50 \
+  --random-input-len 300 \
+  --random-output-len 40 \
+  --random-mm-base-items-per-request 2 \
+  --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \
+  --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}'
+```
+
+## Measured Stages
+
+| Stage | Description |
+|-------|-------------|
+| `get_mm_hashes_secs` | Time spent hashing multimodal inputs |
+| `get_cache_missing_items_secs` | Time spent looking up the processor cache |
+| `apply_hf_processor_secs` | Time spent in the HuggingFace processor |
+| `merge_mm_kwargs_secs` | Time spent merging multimodal kwargs |
+| `apply_prompt_updates_secs` | Time spent updating prompt tokens |
+| `preprocessor_total_secs` | Total preprocessing time |
+| `encoder_forward_secs` | Time spent in the encoder model forward pass |
+| `num_encoder_calls` | Number of encoder invocations per request |
+
+The benchmark also reports end-to-end latency (TTFT + decode time) per
+request. Use `--metric-percentiles` to select which percentiles to report
+(default: p99) and `--output-json` to save results.
+
+For more examples (HF datasets, warmup, JSON output), see
+[Benchmarking CLI — Multimodal Processor Benchmark](../../benchmarking/cli.md#multimodal-processor-benchmark).
+
 ## JSON CLI Arguments
 
 --8<-- "docs/cli/json_tip.inc.md"

docs/configuration/optimization.md

@@ -5,6 +5,17 @@ This guide covers optimization strategies and performance tuning for vLLM V1.
 !!! tip
     Running out of memory? Consult [this guide](./conserving_memory.md) on how to conserve memory.
 
+## Optimization Levels
+
+vLLM provides 4 optimization levels (`-O0`, `-O1`, `-O2`, `-O3`) that allow users to trade off startup time for performance:
+
+- `-O0`: No optimizations. Fastest startup time, but lowest performance.
+- `-O1`: Fast optimization. Simple compilation and fast fusions, and PIECEWISE cudagraphs.
+- `-O2`: Default optimization. Additional compilation ranges, additional fusions, FULL_AND_PIECEWISE cudagraphs.
+- `-O3`: Aggressive optimization. Currently equal to `-O2`, but may include additional time-consuming or experimental optimizations in the future.
+
+For more information, see the [optimization level documentation](../design/optimization_levels.md).
+
 ## Preemption
 
 Due to the autoregressive nature of transformer architecture, there are times when KV cache space is insufficient to handle all batched requests.

docs/contributing/profiling.md

@@ -5,8 +5,12 @@
 
 ## Profile with PyTorch Profiler
 
-We support tracing vLLM workers using the `torch.profiler` module. You can enable the torch profiler by setting `--profiler-config`
-when launching the server, and setting the entries `profiler` to `'torch'` and `torch_profiler_dir` to the directory where you want to save the traces. Additionally, you can control the profiling content by specifying the following additional arguments in the config:
+We support tracing vLLM workers using different profilers. You can enable profiling by setting the `--profiler-config` flag when launching the server.
+
+!!! note
+    The `--profiler-config` flag is available in vLLM v0.13.0 and later. If you are using an earlier version, please upgrade to use this feature.
+
+To use the `torch.profiler` module, set the `profiler` entry to `'torch'` and `torch_profiler_dir` to the directory where you want to save the traces. Additionally, you can control the profiling content by specifying the following additional arguments in the config:
 
 - `torch_profiler_record_shapes` to enable recording Tensor Shapes, off by default
 - `torch_profiler_with_memory` to record memory, off by default

docs/deployment/frameworks/runpod.md

@@ -0,0 +1,87 @@
+# RunPod
+
+vLLM can be deployed on [RunPod](https://www.runpod.io/), a cloud GPU platform that provides on-demand and serverless GPU instances for AI inference workloads.
+
+## Prerequisites
+
+- A RunPod account with GPU pod access
+- A GPU pod running a CUDA-compatible template (e.g., `runpod/pytorch`)
+
+## Starting the Server
+
+SSH into your RunPod pod and launch the vLLM OpenAI-compatible server:
+
+```bash
+python -m vllm.entrypoints.openai.api_server \
+    --model <model-name> \
+    --host 0.0.0.0 \
+    --port 8000
+```
+
+!!! note
+
+    Use `--host 0.0.0.0` to bind to all interfaces so the server is reachable from outside the container.
+
+## Exposing Port 8000
+
+RunPod exposes HTTP services through its proxy. To make port 8000 accessible:
+
+1. In the RunPod dashboard, navigate to your pod settings.
+2. Add `8000` to the list of exposed HTTP ports.
+3. After the pod restarts, RunPod provides a public URL in the format:
+
+    ```text
+    https://<pod-id>-8000.proxy.runpod.net
+    ```
+
+## Troubleshooting 502 Bad Gateway
+
+A `502 Bad Gateway` error from the RunPod proxy typically means the server is not yet listening. Common causes:
+
+- **Model still loading** — Large models take time to download and load into GPU memory. Check the pod logs for progress.
+- **Wrong host binding** — Ensure you passed `--host 0.0.0.0`. Binding to `127.0.0.1` (the default) makes the server unreachable from the proxy.
+- **Port mismatch** — Verify the `--port` value matches the port exposed in the RunPod dashboard.
+- **Out of GPU memory** — The model may be too large for the allocated GPU. Check logs for CUDA OOM errors and consider using a larger instance or adding `--tensor-parallel-size` for multi-GPU pods.
+
+## Verifying the Deployment
+
+Once the server is running, test it with a curl request:
+
+!!! console "Command"
+
+    ```bash
+    curl https://<pod-id>-8000.proxy.runpod.net/v1/chat/completions \
+        -H "Content-Type: application/json" \
+        -d '{
+            "model": "<model-name>",
+            "messages": [
+                {"role": "user", "content": "Hello, how are you?"}
+            ],
+            "max_tokens": 50
+        }'
+    ```
+
+!!! console "Response"
+
+    ```json
+    {
+        "id": "chat-abc123",
+        "object": "chat.completion",
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "content": "I'm doing well, thank you for asking! How can I help you today?"
+                },
+                "index": 0,
+                "finish_reason": "stop"
+            }
+        ]
+    }
+    ```
+
+You can also check the server health endpoint:
+
+```bash
+curl https://<pod-id>-8000.proxy.runpod.net/health
+```

docs/deployment/integrations/aibrix.md

@@ -0,0 +1,5 @@
+# AIBrix
+
+[AIBrix](https://github.com/vllm-project/aibrix) is a cloud-native control plane that integrates with vLLM to simplify Kubernetes deployment, scaling, routing, and LoRA adapter management for large language model inference.
+
+For installation and usage instructions, please refer to the [AIBrix documentation](https://aibrix.readthedocs.io/).

docs/deployment/integrations/dynamo.md

@@ -0,0 +1,7 @@
+# NVIDIA Dynamo
+
+[NVIDIA Dynamo](https://github.com/ai-dynamo/dynamo) is an open-source framework for distributed LLM inference that can run vLLM on Kubernetes with flexible serving architectures (e.g. aggregated/disaggregated, optional router/planner).
+
+For Kubernetes deployment instructions and examples (including vLLM), see the [Deploying Dynamo on Kubernetes](https://github.com/ai-dynamo/dynamo/blob/main/docs/kubernetes/README.md) guide.
+
+Background reading: InfoQ news coverage — [NVIDIA Dynamo simplifies Kubernetes deployment for LLM inference](https://www.infoq.com/news/2025/12/nvidia-dynamo-kubernetes/).

docs/deployment/integrations/kubeai.md

@@ -5,6 +5,7 @@
 Please see the Installation Guides for environment specific instructions:
 
 - [Any Kubernetes Cluster](https://www.kubeai.org/installation/any/)
+- [AKS](https://www.kubeai.org/installation/aks/)
 - [EKS](https://www.kubeai.org/installation/eks/)
 - [GKE](https://www.kubeai.org/installation/gke/)
 

docs/deployment/k8s.md

@@ -11,6 +11,7 @@ Deploying vLLM on Kubernetes is a scalable and efficient way to serve machine le
 Alternatively, you can deploy vLLM to Kubernetes using any of the following:
 
 - [Helm](frameworks/helm.md)
+- [NVIDIA Dynamo](integrations/dynamo.md)
 - [InftyAI/llmaz](integrations/llmaz.md)
 - [llm-d](integrations/llm-d.md)
 - [KAITO](integrations/kaito.md)
@@ -20,7 +21,7 @@ Alternatively, you can deploy vLLM to Kubernetes using any of the following:
 - [kubernetes-sigs/lws](frameworks/lws.md)
 - [meta-llama/llama-stack](integrations/llamastack.md)
 - [substratusai/kubeai](integrations/kubeai.md)
-- [vllm-project/aibrix](https://github.com/vllm-project/aibrix)
+- [vllm-project/AIBrix](integrations/aibrix.md)
 - [vllm-project/production-stack](integrations/production-stack.md)
 
 ## Deployment with CPUs

docs/design/arch_overview.md

@@ -122,7 +122,7 @@ For a deployment with `N` GPUs, `TP` tensor parallel size, `DP` data parallel si
 |---|---|---|
 | API Server | `A` (default `DP`) | Handles HTTP requests and input processing |
 | Engine Core | `DP` (default 1) | Scheduler and KV cache management |
-| GPU Worker | `N` (= `DP x TP`) | One per GPU, executes model forward passes |
+| GPU Worker | `N` (= `DP x PP x TP`) | One per GPU, executes model forward passes |
 | DP Coordinator | 1 if `DP > 1`, else 0 | Load balancing across DP ranks |
 | **Total** | **`A + DP + N` (+ 1 if DP > 1)** | |
 

docs/design/attention_backends.md

@@ -171,9 +171,9 @@ Priority is **1 = highest** (tried first).
 | `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥10.0 |
 | `FLASH_ATTN_DIFFKV` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ✅ | Decoder | Any |
 | `FLEX_ATTENTION` |  | fp16, bf16, fp32 | `auto`, `bfloat16` | Any | Any | ❌ | ✅ | ❌ | Decoder, Encoder Only | Any |
-| `ROCM_AITER_FA` |  | fp16, bf16 | `auto` | 16, 32 | 64, 128, 256 | ❌ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_UNIFIED_ATTN` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | All | N/A |
-| `ROCM_ATTN` |  | fp16, bf16, fp32 | `auto` | 16, 32, 544 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | All | N/A |
+| `ROCM_AITER_FA` |  | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32 | 64, 128, 256 | ❌ | ❌ | ❌ | Decoder, Enc-Dec | N/A |
+| `ROCM_AITER_UNIFIED_ATTN` |  | fp16, bf16 | `auto` | %16 | Any | ✅ | ✅ | ❌ | All | N/A |
+| `ROCM_ATTN` |  | fp16, bf16, fp32 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 544 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ✅ | ✅ | ❌ | All | N/A |
 | `TREE_ATTN` |  | fp16, bf16 | `auto` | %16 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | Any |
 | `TRITON_ATTN` |  | fp16, bf16, fp32 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
 
@@ -210,7 +210,7 @@ configuration.
 | `FLASHMLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 64 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x-10.x |
 | `FLASHMLA_SPARSE` | bf16 | `auto`, `bfloat16`, `fp8_ds_mla` | 64 | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
 | `FLASH_ATTN_MLA` | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x |
-| `ROCM_AITER_MLA` | fp16, bf16 | `auto` | 1 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto` | Any | 576 | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
+| `ROCM_AITER_MLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 1 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
+| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto`, `bfloat16` | 1 | Any | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
 | `ROCM_AITER_TRITON_MLA` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
 | `TRITON_MLA` | fp16, bf16 | `auto`, `bfloat16` | Any | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |

docs/design/cuda_graphs.md

@@ -98,7 +98,7 @@ The goal of this structure is to uniquely identify a (padded) batch with minimal
 
 ### `CudagraphDispatcher`
 
-The [CudagraphDispatcher][vllm.v1.cudagraph_dispatcher.CudagraphDispatcher] takes responsibility for maintaining two sets of valid dispatching keys, one set for `FULL` runtime mode and one set for `PIECEWISE` runtime mode, and dispatches the correct runtime mode and the dispatching keys before executing the model's forwards. It will take in the initial key (a rough batch_descriptor for the padded input) and return the selected runtime mode and the final batch_descriptor, then tell the CUDAGraphWarpper instances that decision through forward contexts. Notice that `CudagraphDispatcher` is the only source of truth for available CUDA Graph keys and `CUDAGraphWrapper` instances can blindly trust the forward context on what CUDA Graphs to dispatch to. This lets us simplify the wrapper code and centralize the logic in the dispatcher.
+The [CudagraphDispatcher][vllm.v1.cudagraph_dispatcher.CudagraphDispatcher] takes responsibility for maintaining two sets of valid dispatching keys, one set for `FULL` runtime mode and one set for `PIECEWISE` runtime mode, and dispatches the correct runtime mode and the dispatching keys before executing the model's forwards. It will take in the initial key (a rough batch_descriptor for the padded input) and return the selected runtime mode and the final batch_descriptor, then tell the CUDAGraphWrapper instances that decision through forward contexts. Notice that `CudagraphDispatcher` is the only source of truth for available CUDA Graph keys and `CUDAGraphWrapper` instances can blindly trust the forward context on what CUDA Graphs to dispatch to. This lets us simplify the wrapper code and centralize the logic in the dispatcher.
 
 The dispatching keys are initialized through the dispatcher's `initialize_cudagraph_keys` method, which is called by the gpu_model_runner after all possible attention backends are initialized. This is where we can get much fancier in the future and “prepare” all kinds of CUDA Graphs combinations. For now, we just append available keys based on the valid combos of `decode_mode`/`mixed_mode` of `cudagraph_mode` and `cudagraph_capture_sizes` in the compilation config.
 

docs/design/custom_op.md

@@ -54,6 +54,8 @@ For example:
 --8<-- "vllm/model_executor/layers/attention/mm_encoder_attention.py:mm_encoder_attn"
 
 --8<-- "vllm/model_executor/layers/mla.py:multi_head_latent_attention"
+
+--8<-- "vllm/model_executor/models/deepencoder.py:rel_pos_attention"
 ```
 
 **2. Activation:**

docs/design/dbo.md

@@ -81,7 +81,7 @@ The current implementation has all `dbo_yield` and `dbo_maybe_run_recv_hook` cal
 
 The `make_ubatch_context` function initializes two `UBatchContexts`, one for each UBatch thread. It takes two CUDA streams, the preexisting `ForwardContexts` and a CPU thread barrier. This function should be used exclusively to instantiate `UBatchContexts`. It will handle all of the event initialization.
 
-The `dbo_register_recv_hook` method registers a callback that can be returned by the `FusedMoEPrepareAndFinalize` class in the other UBatch thread’s `UBatchContext`. The callback will be run when the other thread calls `dbo_maybe_run_recv_hook`. This is typically used to wait on an all-to-all kernel.
+The `dbo_register_recv_hook` method registers a callback that can be returned by the `FusedMoEPrepareAndFinalizeModular` class in the other UBatch thread’s `UBatchContext`. The callback will be run when the other thread calls `dbo_maybe_run_recv_hook`. This is typically used to wait on an all-to-all kernel.
 
 The `dbo_maybe_run_recv_hook` method runs a callback that’s set by the `dbo_register_recv_hook` function if that callback exists.
 

docs/design/dcp_communication_patterns.md

@@ -1,350 +0,0 @@
-# DCP Communication Patterns
-
-This document describes the communication patterns for Decode Context Parallelism (DCP) with various configurations of Tensor Parallelism (TP) and Prefill Context Parallelism (PCP).
-
-## Background
-
-- **TP (Tensor Parallelism)**: Splits attention heads across ranks. Each rank has `H/TP` heads.
-- **PCP (Prefill Context Parallelism)**: Splits prefill tokens across ranks. Each PCP slice has its own TP group.
-- **DCP (Decode Context Parallelism)**: Splits KV cache context across ranks for decode.
-
-### Rank Layout
-
-Ranks are laid out as `(ep, dp, pp, pcp, tp)`. For simplicity, we assume `ep=dp=pp=1`.
-
-For **PCP=2, TP=4** (8 ranks):
-```
-        TP=0  TP=1  TP=2  TP=3
-PCP=0    0     1     2     3
-PCP=1    4     5     6     7
-```
-
-For **PCP=2, TP=2** (4 ranks):
-```
-        TP=0  TP=1
-PCP=0    0     1
-PCP=1    2     3
-```
-
-For **PCP=1, TP=4** (4 ranks):
-```
-        TP=0  TP=1  TP=2  TP=3
-PCP=0    0     1     2     3
-```
-
-### DCP Group Formation
-
-DCP groups are formed by spanning **PCP first, then TP**:
-1. Transpose layout to `(tp, pcp)`
-2. Flatten and reshape to `(-1, dcp_size)`
-
----
-
-## Case 1: PCP=1, TP=4, DCP=4
-
-**Groups:**
-| Group Type | Ranks |
-|------------|-------|
-| TP group | `[0, 1, 2, 3]` |
-| DCP group | `[0, 1, 2, 3]` (same as TP) |
-
-**Head distribution:**
-| Rank | TP Position | Heads |
-|------|-------------|-------|
-| 0 | 0 | `[0, H/4)` |
-| 1 | 1 | `[H/4, H/2)` |
-| 2 | 2 | `[H/2, 3H/4)` |
-| 3 | 3 | `[3H/4, H)` |
-
-**DCP Decode Communication:**
-
-```
-Step 1: TP All-Gather (query)
-┌─────────────────────────────────────────────────────┐
-│  All ranks gather with TP group [0,1,2,3]           │
-│  Each rank: H/4 heads → H heads                     │
-└─────────────────────────────────────────────────────┘
-
-Step 2: Attention
-┌─────────────────────────────────────────────────────┐
-│  Each rank computes attention with ALL H heads      │
-│  against its local KV slice (1/4 of context)        │
-└─────────────────────────────────────────────────────┘
-
-Step 3: DCP Reduce (reduce-scatter)
-┌─────────────────────────────────────────────────────┐
-│  DCP == TP, so reduce-scatter works directly        │
-│  Each rank gets its original H/4 heads back         │
-│  Rank 0: heads [0, H/4)                             │
-│  Rank 1: heads [H/4, H/2)                           │
-│  Rank 2: heads [H/2, 3H/4)                          │
-│  Rank 3: heads [3H/4, H)                            │
-└─────────────────────────────────────────────────────┘
-```
-
-**Optimization:** None needed. DCP == TP is optimal.
-
----
-
-## Case 2: PCP=1, TP=4, DCP=2
-
-**Groups:**
-| Group Type | Ranks |
-|------------|-------|
-| TP group | `[0, 1, 2, 3]` |
-| DCP groups | `[0, 1]`, `[2, 3]` |
-
-**Head distribution:**
-| Rank | TP Position | DCP Group | Heads |
-|------|-------------|-----------|-------|
-| 0 | 0 | 0 | `[0, H/4)` |
-| 1 | 1 | 0 | `[H/4, H/2)` |
-| 2 | 2 | 1 | `[H/2, 3H/4)` |
-| 3 | 3 | 1 | `[3H/4, H)` |
-
-**DCP Decode Communication:**
-
-```
-Step 1: TP All-Gather (query)
-┌─────────────────────────────────────────────────────┐
-│  All ranks gather with TP group [0,1,2,3]           │
-│  Each rank: H/4 heads → H heads                     │
-│  (Gathers more than needed!)                        │
-└─────────────────────────────────────────────────────┘
-
-Step 2: Attention
-┌─────────────────────────────────────────────────────┐
-│  DCP group [0,1]: each computes with H heads        │
-│    against 1/2 of context                           │
-│  DCP group [2,3]: each computes with H heads        │
-│    against 1/2 of context                           │
-└─────────────────────────────────────────────────────┘
-
-Step 3: DCP Reduce (all-reduce + slice)
-┌─────────────────────────────────────────────────────┐
-│  DCP ⊂ TP, so need all-reduce + manual slice        │
-│  Within [0,1]: all-reduce, then each slices to H/4  │
-│  Within [2,3]: all-reduce, then each slices to H/4  │
-└─────────────────────────────────────────────────────┘
-```
-
-**Optimization:** Use partial-TP all-gather
-- DCP group `[0,1]` covers TP positions `{0, 1}` → only need `H/2` heads
-- DCP group `[2,3]` covers TP positions `{2, 3}` → only need `H/2` heads
-- Partial-TP groups: `[0,1]` and `[2,3]` (same as DCP groups in this case)
-
-```
-Optimized Step 1: Partial-TP All-Gather
-┌─────────────────────────────────────────────────────┐
-│  Ranks [0,1] gather within [0,1]: H/4 → H/2 heads   │
-│  Ranks [2,3] gather within [2,3]: H/4 → H/2 heads   │
-│  (Half the communication!)                          │
-└─────────────────────────────────────────────────────┘
-```
-
----
-
-## Case 3: PCP=2, TP=2, DCP=4
-
-**Groups:**
-| Group Type | Ranks |
-|------------|-------|
-| TP groups | `[0, 1]` (PCP=0), `[2, 3]` (PCP=1) |
-| DCP group | `[0, 2, 1, 3]` (all ranks) |
-| PCP group | `[0, 1, 2, 3]` |
-
-**Head distribution:**
-| Rank | TP Position | PCP Slice | Heads |
-|------|-------------|-----------|-------|
-| 0 | 0 | 0 | `[0, H/2)` |
-| 1 | 1 | 0 | `[H/2, H)` |
-| 2 | 0 | 1 | `[0, H/2)` |
-| 3 | 1 | 1 | `[H/2, H)` |
-
-**DCP Decode Communication:**
-
-```
-Step 1: TP All-Gather (query)
-┌─────────────────────────────────────────────────────┐
-│  Rank 0 gathers with [0,1] → H heads                │
-│  Rank 1 gathers with [0,1] → H heads                │
-│  Rank 2 gathers with [2,3] → H heads                │
-│  Rank 3 gathers with [2,3] → H heads                │
-└─────────────────────────────────────────────────────┘
-
-Step 2: Attention
-┌─────────────────────────────────────────────────────┐
-│  Each rank computes with ALL H heads                │
-│  against its local KV slice (1/4 of context)        │
-└─────────────────────────────────────────────────────┘
-
-Step 3: DCP Reduce (all-reduce + slice)
-┌─────────────────────────────────────────────────────┐
-│  DCP=4 > TP=2, so need all-reduce + slice           │
-│  All-reduce across [0,2,1,3]                        │
-│  Each rank slices to its TP-local H/2 heads         │
-└─────────────────────────────────────────────────────┘
-```
-
-**Note:** DCP spans both PCP and TP dimensions. All ranks are in one DCP group.
-
----
-
-## Case 4: PCP=2, TP=2, DCP=2
-
-**Groups:**
-| Group Type | Ranks |
-|------------|-------|
-| TP groups | `[0, 1]` (PCP=0), `[2, 3]` (PCP=1) |
-| DCP groups | `[0, 2]`, `[1, 3]` (span PCP, same TP!) |
-
-**Head distribution:**
-| Rank | TP Position | PCP Slice | DCP Group | Heads |
-|------|-------------|-----------|-----------|-------|
-| 0 | 0 | 0 | 0 | `[0, H/2)` |
-| 2 | 0 | 1 | 0 | `[0, H/2)` |
-| 1 | 1 | 0 | 1 | `[H/2, H)` |
-| 3 | 1 | 1 | 1 | `[H/2, H)` |
-
-**Key insight:** Ranks in the same DCP group have the **same TP position** (same heads)!
-
-**DCP Decode Communication:**
-
-```
-Step 1: TP All-Gather (query)
-┌─────────────────────────────────────────────────────┐
-│  Rank 0 gathers with [0,1] → H heads                │
-│  Rank 2 gathers with [2,3] → H heads                │
-│  (Both DCP peers do redundant work!)                │
-└─────────────────────────────────────────────────────┘
-
-Step 2: Attention
-┌─────────────────────────────────────────────────────┐
-│  DCP group [0,2]: each computes with H heads        │
-│    Rank 0: against KV slice A                       │
-│    Rank 2: against KV slice B                       │
-└─────────────────────────────────────────────────────┘
-
-Step 3: DCP Reduce (all-reduce, no scatter needed)
-┌─────────────────────────────────────────────────────┐
-│  All-reduce within [0,2] and [1,3]                  │
-│  Each rank keeps its original H/2 heads             │
-│  (No redistribution needed - same TP position!)    │
-└─────────────────────────────────────────────────────┘
-```
-
-**Optimization:** Skip TP all-gather entirely!
-- Ranks 0 and 2 already have the same heads `[0, H/2)`
-- They can compute attention with `H/2` heads directly
-- Partial-TP group size = 1 (no all-gather needed)
-
-```
-Optimized Step 1: No All-Gather!
-┌─────────────────────────────────────────────────────┐
-│  Each rank keeps its original H/2 heads             │
-│  (Zero communication!)                              │
-└─────────────────────────────────────────────────────┘
-```
-
----
-
-## Case 5: PCP=2, TP=4, DCP=4
-
-**Groups:**
-| Group Type | Ranks |
-|------------|-------|
-| TP groups | `[0,1,2,3]` (PCP=0), `[4,5,6,7]` (PCP=1) |
-| DCP groups | `[0,4,1,5]`, `[2,6,3,7]` |
-
-**Head distribution:**
-| Rank | TP Position | PCP Slice | DCP Group | Heads |
-|------|-------------|-----------|-----------|-------|
-| 0 | 0 | 0 | 0 | `[0, H/4)` |
-| 4 | 0 | 1 | 0 | `[0, H/4)` |
-| 1 | 1 | 0 | 0 | `[H/4, H/2)` |
-| 5 | 1 | 1 | 0 | `[H/4, H/2)` |
-| 2 | 2 | 0 | 1 | `[H/2, 3H/4)` |
-| 6 | 2 | 1 | 1 | `[H/2, 3H/4)` |
-| 3 | 3 | 0 | 1 | `[3H/4, H)` |
-| 7 | 3 | 1 | 1 | `[3H/4, H)` |
-
-**Key insight:** Each DCP group covers **2 unique TP positions** (half of TP=4).
-
-**DCP Decode Communication (current):**
-
-```
-Step 1: TP All-Gather (query)
-┌─────────────────────────────────────────────────────┐
-│  Rank 0 gathers with [0,1,2,3] → H heads            │
-│  Rank 4 gathers with [4,5,6,7] → H heads            │
-│  (Gathers H heads but only needs H/2!)              │
-└─────────────────────────────────────────────────────┘
-
-Step 2: Attention
-┌─────────────────────────────────────────────────────┐
-│  Each rank computes with ALL H heads                │
-│  against its local KV slice (1/4 of context)        │
-└─────────────────────────────────────────────────────┘
-
-Step 3: DCP Reduce (all-reduce + slice)
-┌─────────────────────────────────────────────────────┐
-│  DCP=4, TP=4, but PCP>1 so can't reduce-scatter     │
-│  All-reduce within DCP group, slice to H/4 heads    │
-└─────────────────────────────────────────────────────┘
-```
-
-**Optimization:** Use partial-TP all-gather
-
-DCP group analysis:
-- `[0,4,1,5]` covers TP positions `{0, 1}` → needs `H/2` heads
-- `[2,6,3,7]` covers TP positions `{2, 3}` → needs `H/2` heads
-
-Partial-TP groups (per PCP slice):
-| Partial-TP Group | Ranks | TP Positions | For DCP Group |
-|------------------|-------|--------------|---------------|
-| `[0, 1]` | PCP=0 | {0, 1} | 0 |
-| `[4, 5]` | PCP=1 | {0, 1} | 0 |
-| `[2, 3]` | PCP=0 | {2, 3} | 1 |
-| `[6, 7]` | PCP=1 | {2, 3} | 1 |
-
-```
-Optimized Step 1: Partial-TP All-Gather
-┌─────────────────────────────────────────────────────┐
-│  Ranks [0,1] gather within [0,1]: H/4 → H/2 heads   │
-│  Ranks [4,5] gather within [4,5]: H/4 → H/2 heads   │
-│  Ranks [2,3] gather within [2,3]: H/4 → H/2 heads   │
-│  Ranks [6,7] gather within [6,7]: H/4 → H/2 heads   │
-│  (Half the communication vs full TP all-gather!)   │
-└─────────────────────────────────────────────────────┘
-```
-
----
-
-## Summary: When to Use Each Pattern
-
-| Condition | All-Gather | Reduce |
-|-----------|------------|--------|
-| `DCP == TP` and `PCP == 1` | Full TP | reduce-scatter |
-| `DCP < TP` and `PCP == 1` | Partial-TP (DCP group) | all-reduce + slice |
-| `DCP == TP * PCP` | Full TP | all-reduce + slice |
-| `DCP < TP * PCP` and `PCP > 1` | Partial-TP | all-reduce + slice |
-
-### Partial-TP Group Formula
-
-```python
-unique_tp_per_dcp = dcp_size // pcp_size
-num_dcp_groups = (tp_size * pcp_size) // dcp_size
-
-# For each DCP group i, for each PCP slice p:
-#   Partial-TP group = ranks at TP positions [i * unique_tp_per_dcp, (i+1) * unique_tp_per_dcp)
-#                      within PCP slice p
-```
-
-### Communication Savings
-
-| Config | Full TP All-Gather | Partial-TP All-Gather | Savings |
-|--------|-------------------|----------------------|---------|
-| PCP=1, TP=4, DCP=4 | H | H | 0% |
-| PCP=1, TP=4, DCP=2 | H | H/2 | 50% |
-| PCP=2, TP=2, DCP=2 | H | 0 (skip!) | 100% |
-| PCP=2, TP=4, DCP=4 | H | H/2 | 50% |

docs/design/fused_moe_modular_kernel.md

@@ -37,31 +37,31 @@ The rest of the document will focus on the Contiguous / Non-Batched case. Extrap
 FusedMoEModularKernel splits the FusedMoE operation into 3 parts,
 
 1. TopKWeightAndReduce
-2. FusedMoEPrepareAndFinalize
-3. FusedMoEPermuteExpertsUnpermute
+2. FusedMoEPrepareAndFinalizeModular
+3. FusedMoEExpertsModular
 
 ### TopKWeightAndReduce
 
-The TopK Weight Application and Reduction components happen right after the Unpermute operation and before the All2All Combine. Note that the `FusedMoEPermuteExpertsUnpermute` is responsible for the Unpermute and `FusedMoEPrepareAndFinalize` is responsible for the All2All Combine. There is value in doing the TopK Weight Application and Reduction in the `FusedMoEPermuteExpertsUnpermute`. But some implementations choose to do it `FusedMoEPrepareAndFinalize`. In order to enable this flexibility, we have a TopKWeightAndReduce abstract class.
+The TopK Weight Application and Reduction components happen right after the Unpermute operation and before the All2All Combine. Note that the `FusedMoEExpertsModular` is responsible for the Unpermute and `FusedMoEPrepareAndFinalizeModular` is responsible for the All2All Combine. There is value in doing the TopK Weight Application and Reduction in the `FusedMoEExpertsModular`. But some implementations choose to do it `FusedMoEPrepareAndFinalizeModular`. In order to enable this flexibility, we have a TopKWeightAndReduce abstract class.
 
 Please find the implementations of TopKWeightAndReduce [here](../../vllm/model_executor/layers/fused_moe/topk_weight_and_reduce.py).
 
-`FusedMoEPrepareAndFinalize::finalize()` method accepts a `TopKWeightAndReduce` argument that is invoked inside the method.
-The `FusedMoEModularKernel` acts as a bridge between the `FusedMoEPermuteExpertsUnpermute` and `FusedMoEPerpareAndFinalize` implementations to determine where the TopK Weight Application and Reduction happens.
+`FusedMoEPrepareAndFinalizeModular::finalize()` method accepts a `TopKWeightAndReduce` argument that is invoked inside the method.
+The `FusedMoEModularKernel` acts as a bridge between the `FusedMoEExpertsModular` and `FusedMoEPrepareAndFinalize` implementations to determine where the TopK Weight Application and Reduction happens.
 
-* `FusedMoEPermuteExpertsUnpermute::finalize_weight_and_reduce_impl` method returns `TopKWeightAndReduceNoOp` if the `FusedMoEPermuteExpertsUnpermute` implementation does the weight application and reduction itself.
-* `FusedMoEPermuteExpertsUnpermute::finalize_weight_and_reduce_impl` method returns `TopKWeightAndReduceContiguous` / `TopKWeightAndReduceNaiveBatched` / `TopKWeightAndReduceDelegate` if the `FusedMoEPermuteExpertsUnpermute` implementation needs the `FusedMoEPrepareAndFinalize::finalize()` to do the weight application and reduction.
+* `FusedMoEExpertsModular::finalize_weight_and_reduce_impl` method returns `TopKWeightAndReduceNoOp` if the `FusedMoEExpertsModular` implementation does the weight application and reduction itself.
+* `FusedMoEExpertsModular::finalize_weight_and_reduce_impl` method returns `TopKWeightAndReduceContiguous` / `TopKWeightAndReduceNaiveBatched` / `TopKWeightAndReduceDelegate` if the `FusedMoEExpertsModular` implementation needs the `FusedMoEPrepareAndFinalizeModular::finalize()` to do the weight application and reduction.
 
-### FusedMoEPrepareAndFinalize
+### FusedMoEPrepareAndFinalizeModular
 
-The `FusedMoEPrepareAndFinalize` abstract class exposes `prepare`, `prepare_no_receive`  and `finalize` functions.
-The `prepare` function is responsible for input activation Quantization and All2All Dispatch. If implemented, The `prepare_no_receive` is like `prepare` except it does not wait to receive results from other workers.  Instead it returns a "receiver" callback that must be invoked to wait for the final results of worker. It is not required that this method is supported by all `FusedMoEPrepareAndFinalize` classes, but if it is available, it can be used to interleave work with the initial all to all communication, e.g. interleaving shared experts with fused experts.  The `finalize` function is responsible for invoking the All2All Combine. Additionally the `finalize` function may or may not do the TopK weight application and reduction (Please refer to the TopKWeightAndReduce section)
+The `FusedMoEPrepareAndFinalizeModular` abstract class exposes `prepare`, `prepare_no_receive`  and `finalize` functions.
+The `prepare` function is responsible for input activation Quantization and All2All Dispatch. If implemented, The `prepare_no_receive` is like `prepare` except it does not wait to receive results from other workers.  Instead it returns a "receiver" callback that must be invoked to wait for the final results of worker. It is not required that this method is supported by all `FusedMoEPrepareAndFinalizeModular` classes, but if it is available, it can be used to interleave work with the initial all to all communication, e.g. interleaving shared experts with fused experts.  The `finalize` function is responsible for invoking the All2All Combine. Additionally the `finalize` function may or may not do the TopK weight application and reduction (Please refer to the TopKWeightAndReduce section)
 
-![FusedMoEPrepareAndFinalize Blocks](../assets/design/fused_moe_modular_kernel/prepare_and_finalize_blocks.png)
+![FusedMoEPrepareAndFinalizeModular Blocks](../assets/design/fused_moe_modular_kernel/prepare_and_finalize_blocks.png)
 
-### FusedMoEPermuteExpertsUnpermute
+### FusedMoEExpertsModular
 
-The `FusedMoEPermuteExpertsUnpermute` class is where the crux of the MoE operations happen. The `FusedMoEPermuteExpertsUnpermute` abstract class exposes a few important functions,
+The `FusedMoEExpertsModular` class is where the crux of the MoE operations happen. The `FusedMoEExpertsModular` abstract class exposes a few important functions,
 
 * apply()
 * workspace_shapes()
@@ -81,25 +81,25 @@ The `apply` method is where the implementations perform
 
 #### workspace_shapes()
 
-The core FusedMoE implementation performs a series of operations. It would be inefficient to create output memory for each of these operations separately. To that effect, implementations are required to declare 2 workspace shapes, the workspace datatype and the FusedMoE output shape as outputs of the workspace_shapes() method. This information is used to allocate the workspace tensors and the output tensor in `FusedMoEModularKernel::forward()` and passed on to the `FusedMoEPermuteExpertsUnpermute::apply()` method. The workspaces could then be used as intermediate buffers in the FusedMoE implementation.
+The core FusedMoE implementation performs a series of operations. It would be inefficient to create output memory for each of these operations separately. To that effect, implementations are required to declare 2 workspace shapes, the workspace datatype and the FusedMoE output shape as outputs of the workspace_shapes() method. This information is used to allocate the workspace tensors and the output tensor in `FusedMoEModularKernel::forward()` and passed on to the `FusedMoEExpertsModular::apply()` method. The workspaces could then be used as intermediate buffers in the FusedMoE implementation.
 
 #### finalize_weight_and_reduce_impl()
 
-It is sometimes efficient to perform TopK weight application and Reduction inside the `FusedMoEPermuteExpertsUnpermute::apply()`. Find an example [here](https://github.com/vllm-project/vllm/pull/20228). We have a `TopKWeightAndReduce` abstract class to facilitate such implementations. Please refer to the TopKWeightAndReduce section.
-`FusedMoEPermuteExpertsUnpermute::finalize_weight_and_reduce_impl()` returns the `TopKWeightAndReduce` object that the implementation wants the `FusedMoEPrepareAndFinalize::finalize()` to use.
+It is sometimes efficient to perform TopK weight application and Reduction inside the `FusedMoEExpertsModular::apply()`. Find an example [here](https://github.com/vllm-project/vllm/pull/20228). We have a `TopKWeightAndReduce` abstract class to facilitate such implementations. Please refer to the TopKWeightAndReduce section.
+`FusedMoEExpertsModular::finalize_weight_and_reduce_impl()` returns the `TopKWeightAndReduce` object that the implementation wants the `FusedMoEPrepareAndFinalizeModular::finalize()` to use.
 
-![FusedMoEPermuteExpertsUnpermute Blocks](../assets/design/fused_moe_modular_kernel/fused_experts_blocks.png)
+![FusedMoEExpertsModular Blocks](../assets/design/fused_moe_modular_kernel/fused_experts_blocks.png)
 
 ### FusedMoEModularKernel
 
-`FusedMoEModularKernel` is composed of the `FusedMoEPrepareAndFinalize` and `FusedMoEPermuteExpertsUnpermute` objects.
+`FusedMoEModularKernel` is composed of the `FusedMoEPrepareAndFinalizeModular` and `FusedMoEExpertsModular` objects.
 `FusedMoEModularKernel` pseudocode/sketch,
 
 ```py
 class FusedMoEModularKernel:
     def __init__(self,
-                 prepare_finalize: FusedMoEPrepareAndFinalize,
-                 fused_experts: FusedMoEPermuteExpertsUnpermute):
+                 prepare_finalize: FusedMoEPrepareAndFinalizeModular,
+                 fused_experts: FusedMoEExpertsModular):
 
         self.prepare_finalize = prepare_finalize
         self.fused_experts = fused_experts
@@ -128,53 +128,53 @@ class FusedMoEModularKernel:
 
 ## How-To
 
-### How To Add a FusedMoEPrepareAndFinalize Type
+### How To Add a FusedMoEPrepareAndFinalizeModular Type
 
-Typically a FusedMoEPrepareAndFinalize type is backed by an All2All Dispatch & Combine implementation / kernel. For example,
+Typically a FusedMoEPrepareAndFinalizeModular type is backed by an All2All Dispatch & Combine implementation / kernel. For example,
 
 * DeepEPHTPrepareAndFinalize type is backed by DeepEP High-Throughput All2All kernels, and
 * DeepEPLLPrepareAndFinalize type is backed by DeepEP Low-Latency All2All kernels.
 
 #### Step 1: Add an All2All manager
 
-The purpose of the All2All Manager is to set up the All2All kernel implementations. The `FusedMoEPrepareAndFinalize` implementations typically fetch a kernel-implementation "handle" from the All2All Manager to invoke the Dispatch and Combine functions. Please look at the All2All Manager implementations [here](../../vllm/distributed/device_communicators/all2all.py).
+The purpose of the All2All Manager is to set up the All2All kernel implementations. The `FusedMoEPrepareAndFinalizeModular` implementations typically fetch a kernel-implementation "handle" from the All2All Manager to invoke the Dispatch and Combine functions. Please look at the All2All Manager implementations [here](../../vllm/distributed/device_communicators/all2all.py).
 
-#### Step 2: Add a FusedMoEPrepareAndFinalize Type
+#### Step 2: Add a FusedMoEPrepareAndFinalizeModular Type
 
-This section describes the significance of the various functions exposed by the `FusedMoEPrepareAndFinalize` abstract class.
+This section describes the significance of the various functions exposed by the `FusedMoEPrepareAndFinalizeModular` abstract class.
 
-`FusedMoEPrepareAndFinalize::prepare()`: The prepare method implements the Quantization and All2All Dispatch. Typically the Dispatch function from the relevant All2All Manager is invoked.
+`FusedMoEPrepareAndFinalizeModular::prepare()`: The prepare method implements the Quantization and All2All Dispatch. Typically the Dispatch function from the relevant All2All Manager is invoked.
 
-`FusedMoEPrepareAndFinalize::has_prepare_no_receive()`: Indicates whether or not this subclass implements `prepare_no_receive`. Defaults to False.
+`FusedMoEPrepareAndFinalizeModular::has_prepare_no_receive()`: Indicates whether or not this subclass implements `prepare_no_receive`. Defaults to False.
 
-`FusedMoEPrepareAndFinalize::prepare_no_receive()`: The prepare_no_receive method implements the Quantization and All2All Dispatch. It does not wait for the result of the dispatch operation but instead returns a thunk that can be invoked to wait for the final results. Typically the Dispatch function from the relevant All2All Manager is invoked.
+`FusedMoEPrepareAndFinalizeModular::prepare_no_receive()`: The prepare_no_receive method implements the Quantization and All2All Dispatch. It does not wait for the result of the dispatch operation but instead returns a thunk that can be invoked to wait for the final results. Typically the Dispatch function from the relevant All2All Manager is invoked.
 
-`FusedMoEPrepareAndFinalize::finalize()`: Maybe perform TopK Weight Application and Reduction and All2All Combine. Typically the Combine function from the relevant All2AllManager is invoked.
+`FusedMoEPrepareAndFinalizeModular::finalize()`: Maybe perform TopK Weight Application and Reduction and All2All Combine. Typically the Combine function from the relevant All2AllManager is invoked.
 
-`FusedMoEPrepareAndFinalize::activation_format()`: Return `FusedMoEActivationFormat.BatchedExperts` if the output of the prepare method (i.e. the All2All dispatch) is Batched. Return `FusedMoEActivationFormat.Standard` otherwise.
+`FusedMoEPrepareAndFinalizeModular::activation_format()`: Return `FusedMoEActivationFormat.BatchedExperts` if the output of the prepare method (i.e. the All2All dispatch) is Batched. Return `FusedMoEActivationFormat.Standard` otherwise.
 
-`FusedMoEPrepareAndFinalize::topk_indices_dtype()`: Data type of the TopK ids. Some All2All kernels have strict requirements pertaining to the data type of the TopK ids. This requirement is passed on to the `FusedMoe::select_experts` function so it could be respected. If there are no strict requirements return None.
+`FusedMoEPrepareAndFinalizeModular::topk_indices_dtype()`: Data type of the TopK ids. Some All2All kernels have strict requirements pertaining to the data type of the TopK ids. This requirement is passed on to the `FusedMoe::select_experts` function so it could be respected. If there are no strict requirements return None.
 
-`FusedMoEPrepareAndFinalize::max_num_tokens_per_rank()`: This is the maximum number of tokens that would be submitted to the All2All Dispatch at once.
+`FusedMoEPrepareAndFinalizeModular::max_num_tokens_per_rank()`: This is the maximum number of tokens that would be submitted to the All2All Dispatch at once.
 
-`FusedMoEPrepareAndFinalize::num_dispatchers()`: Total number of dispatching units. This value determines the size of the Dispatch output. The Dispatch output is of shape (num_local_experts, max_num_tokens, K). Here max_num_tokens = num_dispatchers() * max_num_tokens_per_rank().
+`FusedMoEPrepareAndFinalizeModular::num_dispatchers()`: Total number of dispatching units. This value determines the size of the Dispatch output. The Dispatch output is of shape (num_local_experts, max_num_tokens, K). Here max_num_tokens = num_dispatchers() * max_num_tokens_per_rank().
 
-We suggest picking an already existing `FusedMoEPrepareAndFinalize` implementation that matches your All2All implementation closely and using it as a reference.
+We suggest picking an already existing `FusedMoEPrepareAndFinalizeModular` implementation that matches your All2All implementation closely and using it as a reference.
 
-### How To Add a FusedMoEPermuteExpertsUnpermute Type
+### How To Add a FusedMoEExpertsModular Type
 
-FusedMoEPermuteExpertsUnpermute performs the core of the FusedMoE operations. The various functions exposed by the abstract class and their significance is as follows,
+FusedMoEExpertsModular performs the core of the FusedMoE operations. The various functions exposed by the abstract class and their significance is as follows,
 
-`FusedMoEPermuteExpertsUnpermute::activation_formats()`: Return the supported Input and Output activation formats. i.e. Contiguous / Batched format.
+`FusedMoEExpertsModular::activation_formats()`: Return the supported Input and Output activation formats. i.e. Contiguous / Batched format.
 
-`FusedMoEPermuteExpertsUnpermute::supports_chunking()`: Return True if the implementation supports chunking. Typically
+`FusedMoEExpertsModular::supports_chunking()`: Return True if the implementation supports chunking. Typically
 implementations that input `FusedMoEActivationFormat.Standard` support chunking and `FusedMoEActivationFormat.BatchedExperts` do not.
 
-`FusedMoEPermuteExpertsUnpermute::supports_expert_map()`: Return True if the implementation supports expert map.
+`FusedMoEExpertsModular::supports_expert_map()`: Return True if the implementation supports expert map.
 
-`FusedMoEPermuteExpertsUnpermute::workspace_shapes()` /
-`FusedMoEPermuteExpertsUnpermute::finalize_weight_and_reduce_impl` /
-`FusedMoEPermuteExpertsUnpermute::apply`: Refer to `FusedMoEPermuteExpertsUnpermute` section above.
+`FusedMoEExpertsModular::workspace_shapes()` /
+`FusedMoEExpertsModular::finalize_weight_and_reduce_impl` /
+`FusedMoEExpertsModular::apply`: Refer to `FusedMoEExpertsModular` section above.
 
 ### FusedMoEModularKernel Initialization
 
@@ -186,14 +186,14 @@ implementations that input `FusedMoEActivationFormat.Standard` support chunking
 
 #### maybe_make_prepare_finalize
 
-The `maybe_make_prepare_finalize` method is responsible for constructing an instance of `FusedMoEPrepareAndFinalize` when appropriate based on the current all2all backend, e.g. when EP + DP is enabled.  The base class method currently constructs all the `FusedMoEPrepareAndFinalize` objects for the EP+DP case.  Derived classes can override this method to construct prepare/finalize objects for different scenarios, e.g. `ModelOptNvFp4FusedMoE` can construct a `FlashInferCutlassMoEPrepareAndFinalize` for the EP+TP case.
+The `maybe_make_prepare_finalize` method is responsible for constructing an instance of `FusedMoEPrepareAndFinalizeModular` when appropriate based on the current all2all backend, e.g. when EP + DP is enabled.  The base class method currently constructs all the `FusedMoEPrepareAndFinalizeModular` objects for the EP+DP case.  Derived classes can override this method to construct prepare/finalize objects for different scenarios, e.g. `ModelOptNvFp4FusedMoE` can construct a `FlashInferCutlassMoEPrepareAndFinalize` for the EP+TP case.
 Please refer to the implementations in,
 
 * `ModelOptNvFp4FusedMoE`
 
 #### select_gemm_impl
 
-The `select_gemm_impl` method is undefined in the base class. It is the responsibility of the derived class to implement a method that constructs a valid/appropriate `FusedMoEPermuteExpertsUnpermute` object.
+The `select_gemm_impl` method is undefined in the base class. It is the responsibility of the derived class to implement a method that constructs a valid/appropriate `FusedMoEExpertsModular` object.
 Please refer to the implementations in,
 
 * `UnquantizedFusedMoEMethod`
@@ -205,7 +205,7 @@ derived classes.
 
 #### init_prepare_finalize
 
-Based on the input and env settings, the `init_prepare_finalize` method creates the appropriate `FusedMoEPrepareAndFinalize` object. The method then queries `select_gemm_impl` for the appropriate `FusedMoEPermuteExpertsUnpermute` object and builds the `FusedMoEModularKernel` object
+Based on the input and env settings, the `init_prepare_finalize` method creates the appropriate `FusedMoEPrepareAndFinalizeModular` object. The method then queries `select_gemm_impl` for the appropriate `FusedMoEExpertsModular` object and builds the `FusedMoEModularKernel` object
 
 Please take a look at [init_prepare_finalize](https://github.com/vllm-project/vllm/blob/1cbf951ba272c230823b947631065b826409fa62/vllm/model_executor/layers/fused_moe/layer.py#L188).
 **Important**: The `FusedMoEMethodBase` derived classes use the `FusedMoEMethodBase::fused_experts` object in their `apply` methods. When settings permit the construction of a valid `FusedMoEModularKernel` object, we override `FusedMoEMethodBase::fused_experts` with it. This essentially makes the derived classes agnostic to what FusedMoE implementation is used.
@@ -214,9 +214,9 @@ Please take a look at [init_prepare_finalize](https://github.com/vllm-project/vl
 
 We have `FusedMoEModularKernel` unit tests at [test_modular_kernel_combinations.py](../../tests/kernels/moe/test_modular_kernel_combinations.py).
 
-The unit test iterates through all combinations of `FusedMoEPrepareAndFinalize` and `FusedMoEPremuteExpertsUnpermute` types and if they are
+The unit test iterates through all combinations of `FusedMoEPrepareAndFinalizeModular` and `FusedMoEPremuteExpertsUnpermute` types and if they are
 compatible, runs some correctness tests.
-If you are adding some `FusedMoEPrepareAndFinalize` / `FusedMoEPermuteExpertsUnpermute` implementations,
+If you are adding some `FusedMoEPrepareAndFinalizeModular` / `FusedMoEExpertsModular` implementations,
 
 1. Add the implementation type to `MK_ALL_PREPARE_FINALIZE_TYPES` and `MK_FUSED_EXPERT_TYPES` in [mk_objects.py](../../tests/kernels/moe/modular_kernel_tools/mk_objects.py) respectively.
 2. Update `Config::is_batched_prepare_finalize()`, `Config::is_batched_fused_experts()`, `Config::is_standard_fused_experts()`,
@@ -225,24 +225,24 @@ If you are adding some `FusedMoEPrepareAndFinalize` / `FusedMoEPermuteExpertsUnp
 
 Doing this will add the new implementation to the test suite.
 
-### How To Check `FusedMoEPrepareAndFinalize` & `FusedMoEPermuteExpertsUnpermute` Compatibility
+### How To Check `FusedMoEPrepareAndFinalizeModular` & `FusedMoEExpertsModular` Compatibility
 
 The unit test file [test_modular_kernel_combinations.py](../../tests/kernels/moe/test_modular_kernel_combinations.py) can also be executed as a standalone script.
 Example: `python3 -m tests.kernels.moe.test_modular_kernel_combinations --pf-type DeepEPLLPrepareAndFinalize --experts-type BatchedTritonExperts`
-As a side effect, this script can be used to test `FusedMoEPrepareAndFinalize` & `FusedMoEPermuteExpertsUnpermute` compatibility. When invoked
+As a side effect, this script can be used to test `FusedMoEPrepareAndFinalizeModular` & `FusedMoEExpertsModular` compatibility. When invoked
 with incompatible types, the script will error.
 
 ### How To Profile
 
 Please take a look at [profile_modular_kernel.py](../../tests/kernels/moe/modular_kernel_tools/profile_modular_kernel.py)
 The script can be used to generate Torch traces for a single `FusedMoEModularKernel::forward()` call for any compatible
-`FusedMoEPrepareAndFinalize` and `FusedMoEPermuteExpertsUnpermute` types.
+`FusedMoEPrepareAndFinalizeModular` and `FusedMoEExpertsModular` types.
 Example: `python3 -m tests.kernels.moe.modular_kernel_tools.profile_modular_kernel --pf-type DeepEPLLPrepareAndFinalize --experts-type BatchedTritonExperts`
 
-## FusedMoEPrepareAndFinalize Implementations
+## FusedMoEPrepareAndFinalizeModular Implementations
 
 See [Fused MoE Kernel features](./moe_kernel_features.md#fused-moe-modular-all2all-backends) for a list of all the available modular prepare and finalize subclasses.
 
-## FusedMoEPermuteExpertsUnpermute
+## FusedMoEExpertsModular
 
 See [Fused MoE Kernel features](./moe_kernel_features.md#fused-moe-experts-kernels) for a list of all the available modular experts.

docs/design/fusions.md

@@ -0,0 +1,339 @@
+# Fusion torch.compile passes
+
+vLLM applies a set of kernel/operator fusions at compile time (via custom [`torch.compile`](torch_compile.md) Inductor passes)
+to separate optimizations from model definitions and avoid breaking layer abstractions in model code.
+These fusions are controlled by fields in [`PassConfig`][vllm.config.compilation.PassConfig] and are automatically enabled
+at appropriate [optimization levels](optimization_levels.md).
+
+## Quick Reference
+
+The table below maps each fusion to its controlling flag/config knob, the
+operations it fuses, what level enables it by default, and an indicative speedup.
+The Fullgraph column indicates whether the fusion requires the entire model graph to be
+visible (either via Inductor partition or `splitting_ops=[]`),
+and the last column indicates whether the fusion activates for all `num_tokens`
+or just on the low or high end.
+
+!!! info
+    Speedup depends heavily on the exact model, batch size, and hardware.
+    If tuning performance by hand, always benchmark your exact use-case with and without the fusion to verify the impact.
+
+| Fusion                                                                         | `PassConfig` flag            | Fused operations                               | Default at                     | E2E Speedup        | Fullgraph | `num_tokens` |
+|--------------------------------------------------------------------------------|------------------------------|------------------------------------------------|--------------------------------|--------------------|-----------|--------------|
+| [AllReduce + RMSNorm](#allreduce--rmsnorm-fuse_allreduce_rms)                  | `fuse_allreduce_rms`         | All-reduce → RMSNorm (+residual_add) (→ quant) | O2 (Hopper/Blackwell + TP > 1) | 5-20%              | No        | Low          |
+| [Attention + Quant](#attention--quantization-fuse_attn_quant)                  | `fuse_attn_quant`            | Attention output → FP8/NVFP4 quant             | Off by default                 | 3-7%               | Yes       | Always       |
+| [RoPE + KV-Cache Update](#rope--kv-cache-update-fuse_rope_kvcache)             | `fuse_rope_kvcache`          | Rotary embedding → KV cache write              | O1 (ROCm/AITER only)           | TBD                | No        | Low          |
+| [QK Norm + RoPE](#qk-norm--rope-enable_qk_norm_rope_fusion)                    | `enable_qk_norm_rope_fusion` | Q/K RMSNorm → rotary embedding                 | Off by default                 | 2-3%               | No        | Low          |
+| [Sequence Parallelism](#sequence-parallelism-enable_sp)                        | `enable_sp`                  | AllReduce → ReduceScatter + AllGather          | Off by default                 | Prereq for AsyncTP | Yes       | High         |
+| [AsyncTP GEMM + collective](#asynctp-gemm--collective-overlap-fuse_gemm_comms) | `fuse_gemm_comms`            | GEMM → reduce-scatter / all-gather → GEMM      | Off by default                 | 7-10%              | Yes       | High         |
+| [RMSNorm + Quant](#rmsnorm--quantization-fuse_norm_quant)                      | `fuse_norm_quant`            | RMSNorm (+residual add) → FP8/FP4 quant        | O1 (conditional)               | 1-4%               | No        | Always       |
+| [SiLU+Mul + Quant](#silumul--quantization-fuse_act_quant)                      | `fuse_act_quant`             | SiLU+Mul activation → FP8/FP4 quant            | O1 (conditional)               | 1-4%               | No        | Always       |
+| [RMSNorm + Padding](#rmsnorm--padding-fuse_act_padding)                        | `fuse_act_padding`           | Residual add + RMSNorm → padding               | O1 (ROCm/AITER only)           | TBD                | No        | Always       |
+
+## Support Matrix
+
+The table below lists the quantization schemes supported by each fusion on each platform.
+**—** means the fusion is not available on that platform. The latest and in-progress work is available in the tracking issue:
+[#36066](https://github.com/vllm-project/vllm/issues/36066)
+
+| Fusion                       | SM100 (Blackwell)                        | SM90 (Hopper)                            | SM89 (Ada)                               | SM80 (Ampere) | ROCm                                     |
+|------------------------------|------------------------------------------|------------------------------------------|------------------------------------------|---------------|------------------------------------------|
+| `fuse_allreduce_rms`         | FP16/BF16, FP8 static, NVFP4             | FP16/BF16, FP8 static                    | —                                        | —             | —                                        |
+| `fuse_attn_quant`\*          | FP8 static\*, NVFP4\*                    | FP8 static\*                             | FP8 static\*                             | —             | FP8 static\*                             |
+| `fuse_rope_kvcache`          | —                                        | —                                        | —                                        | —             | FP16/BF16                                |
+| `enable_qk_norm_rope_fusion` | FP16/BF16                                | FP16/BF16                                | FP16/BF16†                               | FP16/BF16†    | —                                        |
+| `enable_sp`                  | FP16/BF16, FP8 static†                   | FP16/BF16, FP8 static                    | FP16/BF16†                               | FP16/BF16†    | —                                        |
+| `fuse_gemm_comms`            | FP16/BF16, FP8 static†                   | FP16/BF16, FP8 static                    | FP16/BF16†                               | FP16/BF16†    | —                                        |
+| `fuse_norm_quant`            | FP8 static, FP8 per-token, FP8 per-group | FP8 static, FP8 per-token, FP8 per-group | FP8 static, FP8 per-token, FP8 per-group | —             | FP8 static, FP8 per-token, FP8 per-group |
+| `fuse_act_quant`             | FP8 static, NVFP4                        | FP8 static                               | FP8 static                               | —             | FP8 per-group                            |
+| `fuse_act_padding`           | —                                        | —                                        | —                                        | —             | FP16/BF16                                |
+
+\* `fuse_attn_quant` support depends on the attention backend in use; not all backends support
+fused quantization output. See the [`fuse_attn_quant` section](#attention--quantization-fuse_attn_quant)
+for per-backend details.
+
+† `enable_sp` and `fuse_gemm_comms` are only autoconfigured for SM90 today;
+other architectures support requires setting `PassConfig.sp_min_token_num` explicitly.
+SM100 support also requires setting `VLLM_DISABLED_KERNELS=FlashInferFP8ScaledMMLinearKernel`.
+
+## Enabling / Disabling Fusions
+
+Fusions are exposed through `PassConfig`, which is nested inside `CompilationConfig`:
+
+```python
+from vllm import LLM
+from vllm.config import CompilationConfig, PassConfig
+
+llm = LLM(
+    model="...",
+    optimization_level=2, # Default optimization level
+    compilation_config=CompilationConfig(
+        pass_config=PassConfig(
+            fuse_norm_quant=True,
+            fuse_act_quant=True,
+            fuse_allreduce_rms=False,  # disable a specific fusion
+        )
+    ),
+)
+```
+
+Fusions can also be enabled using command-line flags with any `vllm ...` command:
+
+```bash
+# Enable O2 defaults, but turn off allreduce fusion
+vllm serve meta-llama/Llama-3.1-8B-Instruct -O2 -cc.pass_config.fuse_allreduce_rms=False
+
+# The above is equivalent to the more verbose:
+vllm serve meta-llama/Llama-3.1-8B-Instruct -O2 --compilation-config '{"pass_config": {"fuse_allreduce_rms": false}}'
+
+# Same syntax in other commands, e.g. vllm bench:
+vllm bench latency --model=meta-llama/Llama-3.1-8B-Instruct -O2 -cc.pass_config.fuse_allreduce_rms=False
+```
+
+Fields set explicitly by the user always take precedence over optimization-level defaults.
+
+## Fusion Details
+
+### AllReduce + RMSNorm (`fuse_allreduce_rms`)
+
+!!! warning
+    TP+DP and TP+PP combinations are currently broken
+    ([#34458](https://github.com/vllm-project/vllm/issues/34458) and
+    [#35426](https://github.com/vllm-project/vllm/issues/35426)).
+    Only supported on NVIDIA Hopper (SM90) and Blackwell (SM100) with FlashInfer installed.
+
+**What it fuses.** Fuses the tensor-parallel all-reduce collective with the subsequent residual add,
+RMSNorm, and optionally a quantization step into a single FlashInfer / TRT-LLM communication kernel.
+This fusion is only profitable for small `num_tokens`,
+so the fusion is only performed in the lower compiled range.
+
+Patterns covered:
+
+- `AllReduce → RMSNorm(+residual_add)`: CUDA sm90+ with FlashInfer
+- `AllReduce → RMSNorm(+residual_add) → FP8 static quant`: CUDA sm90+ with FlashInfer
+- `AllReduce → RMSNorm(+residual_add) → NVFP4 dynamic quant`: CUDA sm100+ with FlashInfer
+
+The maximum tensor size below which the fused kernel is used is hardware-dependent (64 MB for TP=2
+on SM90/SM100) and configurable via `PassConfig.fi_allreduce_fusion_max_size_mb`.
+
+**Code locations.**
+
+- Pass: [`vllm/compilation/passes/fusion/allreduce_rms_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/allreduce_rms_fusion.py)
+- FlashInfer all-reduce: [`vllm/distributed/device_communicators/flashinfer_all_reduce.py`](https://github.com/vllm-project/vllm/blob/main/vllm/distributed/device_communicators/flashinfer_all_reduce.py)
+- Benchmark: [`benchmarks/kernels/benchmark_fused_collective.py`](https://github.com/vllm-project/vllm/blob/main/benchmarks/kernels/benchmark_fused_collective.py)
+
+### Attention + Quantization (`fuse_attn_quant`)
+
+!!! info
+    `fuse_attn_quant` is currently not enabled at any optimization level by default and must be set
+    explicitly. It requires the full model graph to be visible (Inductor partition or `splitting_ops=[]`).
+
+**What it fuses.** Fuses the attention output quantization directly after the attention computation,
+eliminating a full-precision memory round-trip of the attention output. Patterns covered:
+
+`Attention → FP8 static quant`:
+
+- `TRITON_ATTN`: CUDA, ROCm
+- `FLASHINFER`: CUDA sm100+ with FlashInfer installed
+- `ROCM_ATTN`: ROCm
+- `ROCM_AITER_UNIFIED_ATTN`: ROCm with AITER
+
+`Attention → NVFP4 dynamic quant`:
+
+- `FLASHINFER`: CUDA sm100+ with FlashInfer installed
+
+Other attention backends do not support fused output quantization yet.
+
+**Code locations.**
+
+- Pass: [`vllm/compilation/passes/fusion/attn_quant_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/attn_quant_fusion.py)
+- Attention backends: [`vllm/v1/attention/backends/`](https://github.com/vllm-project/vllm/blob/main/vllm/v1/attention/backends/)
+
+### RoPE + KV-Cache Update (`fuse_rope_kvcache`)
+
+!!! info
+    ROCm/AITER-only. Not available on NVIDIA CUDA or CPU. The fusion is only enabled for
+    `num_tokens ≤ 256` by default due to AITER fused kernel performance issues.
+    This threshold is configurable via `PassConfig.rope_kvcache_fusion_max_token_num`.
+
+**What it fuses.** Fuses the rotary positional embedding kernel with the KV-cache scatter/write into
+a single kernel, avoiding separate reads and writes of the key and value tensors.
+
+Requires: AMD ROCm with AITER enabled, the `rotary_embedding` custom op active (automatic),
+and the `kv_cache` update op visible in the graph: either by using Inductor graph partition
+or removed from `splitting_ops`.
+If these conditions are set, the fusion is enabled automatically for optimization level O1 and above.
+
+**Code locations.**
+
+- Pass: [`vllm/compilation/passes/fusion/rope_kvcache_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/rope_kvcache_fusion.py)
+
+### Sequence Parallelism (`enable_sp`)
+
+**What it fuses.** Replaces all-reduce collectives with reduce-scatter + local RMSNorm + all-gather,
+splitting the sequence dimension across TP ranks. This restructures the graph so the subsequent AsyncTP
+pass can fuse the reduce-scatter / all-gather with the surrounding GEMMs.
+
+Sequence Parallelism itself does not directly improve performance; it is a prerequisite for the
+AsyncTP pass (`fuse_gemm_comms`). SP is only applied above a minimum token threshold that is
+autoconfigured based on device capability and model `hidden_size`. Currently only active on
+H100/SM90 for models with `hidden_size >= 8192`. The threshold is configurable via
+`PassConfig.sp_min_token_num`.
+
+The general transformation:
+
+```text
+Input → AllReduce → RMSNorm → Output
+becomes:
+Input → ReduceScatter → local RMSNorm → AllGather → Output
+```
+
+Patterns covered:
+
+- First block: `AllReduce → RMSNorm` → `ReduceScatter → RMSNorm → AllGather`
+- Middle blocks: `AllReduce → fused_add_RMSNorm` → `ReduceScatter → fused_add_RMSNorm → AllGather`
+- Both with optional `→ FP8 static quant` suffix
+
+Requires: `use_inductor_graph_partition=True` **or** piecewise compilation with static sizes
+divisible by `tensor_parallel_size`.
+
+Supported hardware: Only tested on NVIDIA CUDA, possibly works on ROCm. FP8 all-gather requires sm90+.
+
+**Code locations.**
+
+- Pass: [`vllm/compilation/passes/fusion/sequence_parallelism.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/sequence_parallelism.py)
+
+### AsyncTP GEMM + Collective Overlap (`fuse_gemm_comms`)
+
+!!! info
+    Requires `enable_sp=True` (enabled automatically). This pass is a no-op if Sequence Parallelism has not been applied.
+
+**What it fuses.** After Sequence Parallelism transforms the graph, fuses GEMM kernels with the
+surrounding reduce-scatter (output projection) and all-gather (input projection) using
+`torch.ops.symm_mem` symmetric-memory primitives, overlapping communication and computation.
+This overlap is only profitable for large `num_tokens`, so the fusion (and preceding SP)
+is only performed in the higher compiled range above `PassConfig.sp_min_token_num`.
+
+Patterns covered:
+
+- `GEMM → reduce-scatter` → `fused_matmul_reduce_scatter`
+- `all-gather → GEMM` → `all_gather_matmul`
+- FP8 scaled variants of both patterns
+
+Supported hardware: NVIDIA CUDA with symmetric-memory (`torch.distributed._symmetric_memory`) support.
+
+On B200, pattern-matching fp8 FlashInfer scaled MM is not supported, so it must be disabled
+([#27893](https://github.com/vllm-project/vllm/issues/27893))
+
+```shell
+VLLM_DISABLED_KERNELS=FlashInferFP8ScaledMMLinearKernel ...
+```
+
+**Code locations.**
+
+- Pass: [`vllm/compilation/passes/fusion/collective_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/collective_fusion.py)
+- Sequence parallelism pass: [`vllm/compilation/passes/fusion/sequence_parallelism.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/sequence_parallelism.py)
+
+### QK Norm + RoPE (`enable_qk_norm_rope_fusion`)
+
+!!! info
+    Only applicable to models that apply per-head RMSNorm to Q and K before rotary positional
+    embedding (e.g. Qwen). Not enabled by default at any optimization level due to perf issues on H100:
+    [#34391](https://github.com/vllm-project/vllm/issues/34391)
+
+**What it fuses.** Fuses the sequence: split QKV → reshape → Q/K RMSNorm → reshape → rotary
+embedding into a single `fused_qk_norm_rope` CUDA kernel.
+
+```text
+# Unfused:
+q, k, v = split(qkv)
+q_norm = rms_norm(q.view(heads))
+k_norm = rms_norm(k.view(kv_heads))
+q_rope, k_rope = rotary_embedding(q_norm, k_norm, ...)
+
+# Fused:
+fused_qk_norm_rope(qkv, ...)
+```
+
+Supported hardware: CUDA (sm80+) only, tested only on sm90 and sm100.
+
+**Code locations.**
+
+- Pass: [`vllm/compilation/passes/fusion/qk_norm_rope_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/qk_norm_rope_fusion.py)
+- CUDA kernel: [`csrc/ops.h`](https://github.com/vllm-project/vllm/blob/main/csrc/ops.h) (`fused_qk_norm_rope`)
+
+### RMSNorm + Quantization (`fuse_norm_quant`)
+
+!!! warning
+    On NVIDIA, Inductor actually generates a faster fused kernel than our custom CUDA kernel.
+    Hence, this fusion is only enabled when either `rms_norm` or `quant_fp8` is using a custom kernel.
+
+**What it fuses.** Combines the custom `rms_norm` / `fused_add_rms_norm`
+operations with subsequent quantization into a single fused kernel,
+eliminating an intermediate read/write of the full-precision activation tensor.
+Two variants are fused:
+
+- *Plain RMSNorm + quant*: `rms_norm(x) → quant_fp8(y)`
+- *Fused-add RMSNorm + quant*: `fused_add_rms_norm(x, residual) → quant_fp8(y)` — also updates the residual in-place.
+
+Note that AITER fusions are currently in a separate pass in `vllm.compilation.passes.fusion.rocm_aiter_fusion`.
+
+Supported quantization scheme/hardware combinations:
+
+- FP8 static per-tensor: CUDA & HIP kernel
+- FP8 dynamic per-token: CUDA & HIP kernel, AITER
+- FP8 dynamic per-token-group (128/64): CUDA & HIP kernel, AITER
+
+**Code locations.**
+
+- Pass: [`vllm/compilation/passes/fusion/rms_quant_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/rms_quant_fusion.py)
+- ROCm AITER pass: [`vllm/compilation/passes/fusion/rocm_aiter_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/rocm_aiter_fusion.py)
+- CUDA/HIP kernels: [`csrc/layernorm_quant_kernels.cu`](https://github.com/vllm-project/vllm/blob/main/csrc/layernorm_quant_kernels.cu)
+
+### SiLU+Mul + Quantization (`fuse_act_quant`)
+
+!!! warning
+    Same as `fuse_norm_quant`: on NVIDIA, Inductor generates a faster fused kernel than our custom ops.
+    This fusion is only enabled when either `silu_and_mul` or `quant_fp8` are using a custom kernel,
+    or for NVFP4-quantized models (where FP4 quant is always a custom op).
+
+**What it fuses.** Fuses the `silu_and_mul` gate-up projection activation with subsequent quantization into a single kernel,
+avoiding materialization of the full-precision post-activation tensor.
+
+Note that AITER fusions are in a separate pass in `vllm.compilation.passes.fusion.rocm_aiter_fusion`.
+
+Supported quantization scheme/hardware combinations:
+
+- FP8 static per-tensor: CUDA & HIP kernel
+- NVFP4 dynamic: CUDA sm100+ only with FlashInfer
+- FP8 per-token-group (128): ROCm AITER only
+
+**Code locations.**
+
+- Pass: [`vllm/compilation/passes/fusion/act_quant_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/act_quant_fusion.py)
+- ROCm AITER pass: [`vllm/compilation/passes/fusion/rocm_aiter_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/rocm_aiter_fusion.py)
+- CUDA/HIP kernels: [`csrc/quantization/`](https://github.com/vllm-project/vllm/blob/main/csrc/quantization/)
+
+### RMSNorm + Padding (`fuse_act_padding`)
+
+!!! info
+    ROCm/AITER-only. Targeted at GPT-OSS models.
+
+**What it fuses.** Fuses a residual add + RMSNorm with a subsequent padding operation that pads
+the hidden dimension to a multiple required by downstream AITER Triton GEMM kernels.
+
+Requires: AMD ROCm with AITER RMSNorm enabled. Enabled by default in optimization level O1 and above
+when the hidden size is 2880 and AITER Triton GEMMs *not* enabled.
+
+**Code locations.**
+
+- Pass: [`vllm/compilation/passes/fusion/rocm_aiter_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/rocm_aiter_fusion.py) (`RocmAiterTritonAddRMSNormPadFusionPass`)
+
+## See Also
+
+- [Optimization Levels](optimization_levels.md) — high-level presets that set
+  fusion defaults.
+- [torch.compile in vLLM](torch_compile.md) — how the Inductor pass pipeline
+  works.
+- [Attention Backends](attention_backends.md) — attention-specific kernel
+  selection.

docs/design/logits_processors.md

@@ -352,7 +352,7 @@ The `BatchUpdate` abstraction models the persistent batch as a list of requests,
         (s, d, UNIDIRECTIONAL or SWAP)
         ```
 
-    * If the Move specifies `UNIDRECTIONAL`:
+    * If the Move specifies `UNIDIRECTIONAL`:
 
         * The request at index `s` is moved to index `d`; index `s` becomes an empty slot
 

docs/design/model_runner_v2.md

@@ -0,0 +1,198 @@
+# Model Runner V2 Design Document
+
+## Introduction
+
+Since vLLM V1 was first implemented, we discovered several fundamental design mistakes and accumulated significant technical debt. Many features were bolted on that were not considered in the original design. We also gained valuable insights into sampling techniques (for example, Gumbel-max sampling), tools (for example, Triton), and CUDA features (for example, UVA). With this knowledge, we implemented Model Runner V2 (MRV2) from first principles to be cleaner, more efficient, and more modular.
+
+In hindsight, many of V1's design choices were suboptimal. While MRV2 is not yet feature-complete, not rigorously tested, and still has open design decisions, we believe it is a substantial improvement over V1.
+
+This document describes the design of MRV2.
+
+## 1. Persistent Batch
+
+One significant source of friction in V1 is its persistent batch implementation.
+
+### Background
+
+V1 introduced persistent batches to minimize CPU overhead during input preparation. When requests are scheduled for a step, the model runner must construct contiguous input tensors (for example, block tables and per-request temperature values) to feed into the model. Building these tensors from scratch each step is often very slow in Python, especially for large tensors like block tables.
+
+The persistent batch optimization exploits the fact that request batches in consecutive steps are mostly identical. Only a few requests (if any) join or finish per step. By maintaining persistent state tensors and applying incremental diffs instead of reconstructing inputs from scratch, CPU overhead can be reduced significantly.
+
+### Problems with V1's Approach
+
+While efficient, V1's persistent batch design introduced unnecessary complexity due to coupling persistent state with input tensors. V1 uses persistent state tensors directly as model and sampler inputs, which imposes strict layout and ordering requirements. When requests join or finish, this often requires complex tensor-wide reordering rather than simple row insertion/removal.
+
+V1 also had to maintain `CachedRequestState`, a redundant backup copy of request state, because rows in persistent tensors can be overwritten while requests are still active.
+
+The result is complex bookkeeping that becomes more difficult under async scheduling.
+
+![Persistent Batch in V1](../assets/design/model_runner_v2/persistent_batch_v1.png)
+
+### MRV2's Solution
+
+MRV2 decouples persistent state tensors from per-step input tensors. Given request ordering for the step (usually determined by the attention backend), MRV2 gathers input tensors from persistent state.
+
+1. Pre-allocate a fixed-size tensor with `max_num_reqs` rows (1024 by default on most platforms).
+2. Assign each request a permanent row for its active lifetime (until finish or preemption).
+3. Treat preemption as completion. On resume, re-add request data as fresh state.
+
+This removes the need for `CachedRequestState` and simplifies bookkeeping. Large state tensors are mostly stored on GPU memory, so gather runs in parallel on the GPU with low overhead.
+
+![Persistent Batch in MRV2](../assets/design/model_runner_v2/persistent_batch_mrv2.png)
+
+## 2. Async-First
+
+vLLM now relies heavily on asynchronous scheduling. The scheduler and worker prepare inputs for step `N+1` while the GPU executes step `N`, overlapping CPU and GPU work to maximize utilization.
+
+V1 was not originally designed with async scheduling in mind, and support required retrofitted behavior and hacks. MRV2 instead assumes the core model execution loop is a CUDA stream with no CPU synchronization points. CPU entrypoints queue work onto the stream.
+
+![Async execution timeline](../assets/design/model_runner_v2/async_sched.png)
+
+## 3. Removing Async Barrier
+
+A key requirement for async execution is that CPU operations remain non-blocking. Both explicit sync (for example, `torch.accelerator.synchronize`) and implicit sync (for example, unpinned `.to("cuda")`) must be avoided.
+
+However, async execution can introduce race conditions when CPU and GPU concurrently touch the same memory.
+
+Example (unsafe):
+
+```python
+class ModelRunner:
+    def __init__(self, ...):
+        # Pinned buffer
+        self.states = torch.zeros(
+            max_num_reqs, dtype=torch.int32, device="cpu", pin_memory=True
+        )
+
+    def execute_step(self, ...):
+        self.states[req_idx] = new_req.data
+        states = self.states.to("cuda", non_blocking=True)
+```
+
+The CPU may modify `self.states` while GPU is still reading from it via async copy.
+
+V1 addresses this with an async barrier around critical sections. That avoids races but has drawbacks:
+
+1. Easy to miss protected buffers (bug-prone).
+2. Inflexible organization (all CPU work must stay inside barrier).
+3. Potentially less overlap due to synchronization.
+
+![Race condition with shared CPU buffer](../assets/design/model_runner_v2/async_race_condition.png)
+
+### MRV2's Solution: Eliminate the Race
+
+MRV2 separates persistent CPU state from the copied tensor:
+
+```python
+class ModelRunner:
+    def __init__(self, ...):
+        # Not pinned
+        self.states = torch.zeros(
+            max_num_reqs, dtype=torch.int32, device="cpu", pin_memory=False
+        )
+
+    def execute_step(self, ...):
+        self.states[req_idx] = new_req.data
+        tmp_states = self.states.pin_memory()
+        states = tmp_states.to("cuda", non_blocking=True)
+```
+
+Now CPU writes to `self.states` while GPU reads from `tmp_states`, eliminating the race without explicit synchronization.
+
+![No race with temporary pinned copy](../assets/design/model_runner_v2/async_no_race_condition.png)
+
+## 4. StagedWriteTensor
+
+For large tensors like block tables, MRV2 avoids full CPU-to-GPU copies each step by using `StagedWriteTensor`:
+
+1. Keep the base tensor on GPU.
+2. Stage diffs on CPU.
+3. Pack diffs into contiguous buffers.
+4. Copy packed diffs to GPU.
+5. Launch one kernel to apply diffs.
+
+Example usage:
+
+```python
+# Initialize state on GPU
+state = StagedWriteTensor(size=(1024, 1000), dtype=torch.int32, device="cuda")
+
+# Write [3, 1, 2] into row 2, starting at index 3
+state.stage_write(row=2, start=3, value=[3, 1, 2])
+
+# Write [-1, -2, -5] into row 0, starting at index 1
+state.stage_write(row=0, start=1, value=[-1, -2, -5])
+
+# Apply staged changes
+state.apply_write()
+```
+
+This supports ragged updates with no CPU-GPU synchronization and minimal kernel launches. It is especially useful for block tables and mixed CPU/GPU-written states such as `num_computed_tokens`.
+
+## 5. GPU-Native Input Metadata Preparation and Output Processing
+
+MRV2 uses Triton kernels to prepare inputs such as `input_ids`, `positions`, `query_start_loc`, and `seq_lens`.
+
+Benefits:
+
+1. Better async behavior: GPU can derive values (for example with speculative decoding) that CPU may not know yet.
+2. Lower CPU overhead: input prep is very cheap on GPU and avoids Python bottlenecks.
+
+### Universal Virtual Addressing (UVA)
+
+MRV2 uses UVA in some paths to let GPU kernels access large CPU-resident tensors directly (for example `prefill_token_ids`) without duplicating those tensors into GPU memory.
+
+## 6. Triton-Native Sampler
+
+MRV2 reimplements sampling mostly in Triton for better numeric/memory control and optimization.
+
+### Gumbel Sampling Kernel
+
+MRV2 introduces a Triton Gumbel sampling kernel that avoids explicit softmax materialization and uses stateless in-kernel RNG from seed input.
+
+### Efficient Top-K Logprobs
+
+V1 materializes full-vocabulary logprobs before top-k. MRV2 identifies top-k tokens from logits first, then computes logprobs only for selected tokens. This reduces peak GPU memory usage.
+
+### Memory-Efficient Prompt Logprobs
+
+MRV2 supports finer-grained chunking, including chunking inside a single prompt, to avoid memory spikes on long prompts.
+
+### Better Compatibility with Speculative Decoding
+
+Instead of expanding per-request sampling states to match per-logit shapes, MRV2 uses indirection (`idx_mapping`) inside kernels to map each logits vector to the right request state. This simplifies support for complex sampling parameters and logits processors.
+
+## 7. Modularity
+
+MRV2 emphasizes modularity. Compared to V1's large, entangled `gpu_model_runner.py`, MRV2 splits feature logic across dedicated files (for example, `mrope_utils.py`, `penalties.py`, and many others).
+
+It also consolidates model inputs into an `InputBatch` class and reduces direct model-runner attribute coupling.
+
+## 8. No Abuse of `dummy_run`
+
+In V1, `dummy_run` handled too many responsibilities:
+
+- Initial memory profiling and `torch.compile`
+- CUDA graph capture
+- Warmups
+- Empty DP forward passes for EP+DP
+
+MRV2 simplifies this:
+
+1. `execute_model` supports dummy runs without affecting state.
+2. `dummy_run` delegates to `execute_model` for profiling, warmup, and empty DP forward passes.
+3. CUDA graph capture uses a separate dedicated path.
+
+This reduces complexity and removes bugs caused by divergence between `execute_model` and `dummy_run` behavior.
+
+## 9. Explicit CUDA Graph Management
+
+V1's CUDA graph handling is implicit and hard to reason about. MRV2 uses a `CUDAGraphManager` that explicitly captures and launches full CUDA graphs through standard PyTorch APIs.
+
+This makes graph lifecycle and execution mode decisions more understandable and easier to extend. Example: MRV2 can capture multiple draft-model forward passes into one CUDA graph.
+
+## Development Philosophy
+
+MRV2 changes should meet a higher code quality bar. As feature gaps with V1 are filled, features should be reconsidered from first principles in the MRV2 design context instead of quickly porting V1 behavior.
+
+A key requirement is preserving modularity and clean abstraction boundaries, even if that requires more upfront design iteration.

docs/design/moe_kernel_features.md

@@ -4,17 +4,17 @@ The purpose of this document is to provide an overview of the various MoE kernel
 
 ## Fused MoE Modular All2All backends
 
-There are a number of all2all communication backends that are used to implement expert parallelism (EP) for the `FusedMoE` layer. The different `FusedMoEPrepareAndFinalize` subclasses provide an interface for each all2all backend.
+There are a number of all2all communication backends that are used to implement expert parallelism (EP) for the `FusedMoE` layer. The different `FusedMoEPrepareAndFinalizeModular` subclasses provide an interface for each all2all backend.
 
 The following table describes the relevant features of each backend, i.e. activation format, supported quantization schemes and async support.
 
-The output activation format (standard or batched) corresponds to the output of the prepare step of the `FusedMoEPrepareAndFinalize` subclass, and the finalize step requires the same format. All the backend `prepare` methods expect activations in the standard format and all the `finalize` methods return activations in standard format. More details on the formats can be found in the [Fused MoE Modular Kernel](./fused_moe_modular_kernel.md) document.
+The output activation format (standard or batched) corresponds to the output of the prepare step of the `FusedMoEPrepareAndFinalizeModular` subclass, and the finalize step requires the same format. All the backend `prepare` methods expect activations in the standard format and all the `finalize` methods return activations in standard format. More details on the formats can be found in the [Fused MoE Modular Kernel](./fused_moe_modular_kernel.md) document.
 
-The quantization types and formats enumerate which quantization schemes are supported by each `FusedMoEPrepareAndFinalize` class. The quantization can happen before or after the dispatch based on the format the all2all backend supports, e.g. deepep_high_throughput supports only block-quantized fp8 format. Any other format will result in dispatching in higher precision and quantizing afterwards. The output of the prepare step for each backend is the quantized type. The finalize step generally requires the same input type as the original activations, e.g. if the original input is bfloat16 and the quantization scheme is fp8 with per-tensor scales, `prepare` will return fp8/per-tensor scale activations and `finalize` will take bfloat16 activations. See the diagrams in [Fused MoE Modular Kernel](./fused_moe_modular_kernel.md) for more details on the types and formats of activations at each step of the MoE process. If no quantization type is specified, the kernel operates on float16 and/or bfloat16.
+The quantization types and formats enumerate which quantization schemes are supported by each `FusedMoEPrepareAndFinalizeModular` class. The quantization can happen before or after the dispatch based on the format the all2all backend supports, e.g. deepep_high_throughput supports only block-quantized fp8 format. Any other format will result in dispatching in higher precision and quantizing afterwards. The output of the prepare step for each backend is the quantized type. The finalize step generally requires the same input type as the original activations, e.g. if the original input is bfloat16 and the quantization scheme is fp8 with per-tensor scales, `prepare` will return fp8/per-tensor scale activations and `finalize` will take bfloat16 activations. See the diagrams in [Fused MoE Modular Kernel](./fused_moe_modular_kernel.md) for more details on the types and formats of activations at each step of the MoE process. If no quantization type is specified, the kernel operates on float16 and/or bfloat16.
 
 Async backends support the use of DBO (Dual Batch Overlap) and shared expert overlap (where shared experts are computed during the combine step).
 
-Certain models require the topk weights to be applied to the input activations rather than the output activations when topk==1, e.g. Llama. For modular kernels, this feature is supported by the `FusedMoEPrepareAndFinalize` subclass. For non-modular kernels, it is up to the experts function to deal with this flag.
+Certain models require the topk weights to be applied to the input activations rather than the output activations when topk==1, e.g. Llama. For modular kernels, this feature is supported by the `FusedMoEPrepareAndFinalizeModular` subclass. For non-modular kernels, it is up to the experts function to deal with this flag.
 
 Unless otherwise specified, backends are controlled via the `--all2all-backend` command-line argument (or the `all2all_backend` parameter in `ParallelConfig`). All backends except `flashinfer` only work with EP+DP or EP+TP. `Flashinfer` can work with EP or DP without EP.
 
@@ -36,8 +36,6 @@ th {
 | deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize.DeepEPHTPrepareAndFinalize] |
 | deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize.DeepEPLLPrepareAndFinalize] |
 | flashinfer_all2allv | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferA2APrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_a2a_prepare_finalize.FlashInferA2APrepareAndFinalize] |
-| MoEPrepareAndFinalizeNoEP<sup>5</sup> | standard | fp8,int8 | G,A,T | N | Y | [`MoEPrepareAndFinalizeNoEP`][vllm.model_executor.layers.fused_moe.prepare_finalize.MoEPrepareAndFinalizeNoEP] |
-| BatchedPrepareAndFinalize<sup>5</sup> | batched | fp8,int8 | G,A,T | N | Y | [`BatchedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.fused_batched_moe.BatchedPrepareAndFinalize] |
 
 !!! info "Table key"
     1. All types: mxfp4, nvfp4, int4, int8, fp8
@@ -75,9 +73,9 @@ Each experts kernel supports one or more activation functions, e.g. silu or gelu
 
 As with the backends, some experts support applying topk weights on the input activations. The entries in the column in this table only apply to the non-modular experts.
 
-Most experts flavors include an equivalent modular interface which will be a subclass of `FusedMoEPermuteExpertsUnpermute`.
+Most experts flavors include an equivalent modular interface which will be a subclass of `FusedMoEExpertsModular`.
 
-To be used with a particular `FusedMoEPrepareAndFinalize` subclass, MoE kernels must have compatible activation formats, quantization types and quantization formats.
+To be used with a particular `FusedMoEPrepareAndFinalizeModular` subclass, MoE kernels must have compatible activation formats, quantization types and quantization formats.
 
 | Kernel | Input act. format | Quant. types | Quant. format | Activation function | Apply Weight On Input | Modular | Source |
 |--------|-------------------|--------------|---------------|---------------------|-----------------------|---------|--------|
@@ -106,7 +104,7 @@ To be used with a particular `FusedMoEPrepareAndFinalize` subclass, MoE kernels
 
 The following table shows "families" of modular kernels that are intended to work together. There are some combinations which may work but have not yet been tested, e.g. flashinfer with other fp8 experts. Note that the "naive" backend will work with any non-modular experts.
 
-| backend | `FusedMoEPrepareAndFinalize` subclasses | `FusedMoEPermuteExpertsUnpermute` subclasses |
+| backend | `FusedMoEPrepareAndFinalizeModular` subclasses | `FusedMoEExpertsModular` subclasses |
 |---------|-----------------------------------------|----------------------------------------------|
 | deepep_high_throughput | `DeepEPHTPrepareAndFinalize` |  `DeepGemmExperts`,</br>`TritonExperts`,</br>`TritonOrDeepGemmExperts`,</br>`CutlassExpertsFp8`, </br>`MarlinExperts` |
 | deepep_low_latency | `DeepEPLLPrepareAndFinalize` |  `BatchedDeepGemmExperts`,</br>`BatchedTritonExperts`,</br>`CutlassBatchedExpertsFp8`,</br>`BatchedMarlinExperts` |

docs/design/multiprocessing.md

@@ -12,9 +12,8 @@ page for information on known issues and how to solve them.
 
 The use of Python multiprocessing in vLLM is complicated by:
 
-- The use of vLLM as a library and the inability to control the code using vLLM
-- Varying levels of incompatibilities between multiprocessing methods and vLLM
-  dependencies
+- using vLLM as a library, which limits control over its internal code;
+- incompatibilities between certain multiprocessing methods and vLLM dependencies.
 
 This document describes how vLLM deals with these challenges.
 
@@ -22,11 +21,9 @@ This document describes how vLLM deals with these challenges.
 
 [Python multiprocessing methods](https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods) include:
 
-- `spawn` - spawn a new Python process. The default on Windows and macOS.
-
+- `spawn` - Spawn a new Python process. The default on Windows and macOS.
 - `fork` - Use `os.fork()` to fork the Python interpreter. The default on
   Linux for Python versions prior to 3.14.
-
 - `forkserver` - Spawn a server process that will fork a new process on request.
   The default on Linux for Python version 3.14 and newer.
 
@@ -36,8 +33,8 @@ This document describes how vLLM deals with these challenges.
 threads. If you are under macOS, using `fork` may cause the process to crash.
 
 `spawn` is more compatible with dependencies, but can be problematic when vLLM
-is used as a library. If the consuming code does not use a `__main__` guard (`if
-__name__ == "__main__":`), the code will be inadvertently re-executed when vLLM
+is used as a library. If the consuming code does not use a `__main__` guard
+(`if __name__ == "__main__":`), the code will be inadvertently re-executed when vLLM
 spawns a new process. This can lead to infinite recursion, among other problems.
 
 `forkserver` will spawn a new server process that will fork new processes on
@@ -57,8 +54,7 @@ Multiple vLLM dependencies indicate either a preference or requirement for using
 - <https://pytorch.org/docs/stable/multiprocessing.html#sharing-cuda-tensors>
 - <https://docs.habana.ai/en/latest/PyTorch/Getting_Started_with_PyTorch_and_Gaudi/Getting_Started_with_PyTorch.html?highlight=multiprocessing#torch-multiprocessing-for-dataloaders>
 
-It is perhaps more accurate to say that there are known problems with using
-`fork` after initializing these dependencies.
+Known issues exist when using `fork` after initializing these dependencies.
 
 ## Current State (v0)
 
@@ -66,8 +62,8 @@ The environment variable `VLLM_WORKER_MULTIPROC_METHOD` can be used to control w
 
 - <https://github.com/vllm-project/vllm/blob/d05f88679bedd73939251a17c3d785a354b2946c/vllm/envs.py#L339-L342>
 
-When we know we own the process because the `vllm` command was used, we use
-`spawn` because it's the most widely compatible.
+If the main process is controlled via the `vllm` command,
+`spawn` is used because it's the most widely compatible.
 
 - <https://github.com/vllm-project/vllm/blob/d05f88679bedd73939251a17c3d785a354b2946c/vllm/scripts.py#L123-L140>
 
@@ -104,8 +100,8 @@ dependencies and code using vLLM as a library.
 ### Changes Made in v1
 
 There is not an easy solution with Python's `multiprocessing` that will work
-everywhere. As a first step, we can get v1 into a state where it does "best
-effort" choice of multiprocessing method to maximize compatibility.
+everywhere. As a first step, we can get v1 into a state where it does
+"best effort" choice of multiprocessing method to maximize compatibility.
 
 - Default to `fork`.
 - Use `spawn` when we know we control the main process (`vllm` was executed).
@@ -154,8 +150,8 @@ RuntimeError:
 ### Detect if a `__main__` guard is present
 
 It has been suggested that we could behave better if we could detect whether
-code using vLLM as a library has a `__main__` guard in place. This [post on
-stackoverflow](https://stackoverflow.com/questions/77220442/multiprocessing-pool-in-a-python-class-without-name-main-guard)
+code using vLLM as a library has a `__main__` guard in place. This
+[post on Stack Overflow](https://stackoverflow.com/questions/77220442/multiprocessing-pool-in-a-python-class-without-name-main-guard)
 was from a library author facing the same question.
 
 It is possible to detect whether we are in the original, `__main__` process, or
@@ -192,4 +188,4 @@ that works around these challenges.
 2. We can explore other libraries that may better suit our needs. Examples to
    consider:
 
-- <https://github.com/joblib/loky>
+    - <https://github.com/joblib/loky>