[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/hardware_tests/amd.yaml

@@ -10,7 +10,7 @@ steps:
       docker build
       --build-arg max_jobs=16
       --build-arg REMOTE_VLLM=1
-      --build-arg ARG_PYTORCH_ROCM_ARCH='gfx90a;gfx942'
+      --build-arg ARG_PYTORCH_ROCM_ARCH='gfx942;gfx950'
       --build-arg VLLM_BRANCH=$BUILDKITE_COMMIT
       --tag "rocm/vllm-ci:${BUILDKITE_COMMIT}"
       -f docker/Dockerfile.rocm

.buildkite/image_build/image_build_cpu_arm64.sh

@@ -14,7 +14,7 @@ BUILDKITE_COMMIT=$3
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
 
 # skip build if image already exists
-if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu) ]]; then
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu) ]]; then
   echo "Image not found, proceeding with build..."
 else
   echo "Image found"
@@ -25,9 +25,9 @@ fi
 docker build --file docker/Dockerfile.cpu \
   --build-arg max_jobs=16 \
   --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
-  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu \
+  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu \
   --target vllm-test \
   --progress plain .
 
 # push
-docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu
+docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu

.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

@@ -1,25 +1,57 @@
 #!/bin/bash
 
-# This script runs test inside the corresponding ROCm docker container.
+# This script runs tests inside the corresponding ROCm docker container.
+# It handles both single-node and multi-node test configurations.
+#
+# Multi-node detection: Instead of matching on fragile group names, we detect
+# multi-node jobs structurally by looking for the bracket command syntax
+# "[node0_cmds] && [node1_cmds]" or via the NUM_NODES environment variable.
+#
+###############################################################################
+# QUOTING / COMMAND PASSING
+#
+# Passing commands as positional arguments ($*) is fragile when the command
+# string itself contains double quotes, e.g.:
+#
+#   bash run-amd-test.sh "export FLAGS="value" && pytest -m "not slow""
+#
+# The outer shell resolves the nested quotes *before* this script runs, so
+# the script receives mangled input it cannot fully recover.
+#
+# Preferred: pass commands via the VLLM_TEST_COMMANDS environment variable:
+#
+#   export VLLM_TEST_COMMANDS='export FLAGS="value" && pytest -m "not slow"'
+#   bash run-amd-test.sh
+#
+# Single-quoted assignment preserves all inner double quotes verbatim.
+# The $* path is kept for backward compatibility but callers should migrate.
+###############################################################################
 set -o pipefail
 
 # Export Python path
 export PYTHONPATH=".."
 
-# Print ROCm version
+###############################################################################
-echo "--- Confirming Clean Initial State"
+# Helper Functions
-while true; do
+###############################################################################
-        sleep 3
-        if grep -q clean /opt/amdgpu/etc/gpu_state; then
-                echo "GPUs state is \"clean\""
-                break
-        fi
-done
 
-echo "--- ROCm info"
+wait_for_clean_gpus() {
-rocminfo
+  local timeout=${1:-300}
+  local start=$SECONDS
+  echo "--- Waiting for clean GPU state (timeout: ${timeout}s)"
+  while true; do
+    if grep -q clean /opt/amdgpu/etc/gpu_state; then
+      echo "GPUs state is \"clean\""
+      return
+    fi
+    if (( SECONDS - start >= timeout )); then
+      echo "Error: GPUs did not reach clean state within ${timeout}s" >&2
+      exit 1
+    fi
+    sleep 3
+  done
+}
 
-# cleanup older docker images
 cleanup_docker() {
   # Get Docker's root directory
   docker_root=$(docker info -f '{{.DockerRootDir}}')
@@ -28,15 +60,12 @@ cleanup_docker() {
     exit 1
   fi
   echo "Docker root directory: $docker_root"
-  # Check disk usage of the filesystem where Docker's root directory is located
+
   disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
-  # Define the threshold
   threshold=70
   if [ "$disk_usage" -gt "$threshold" ]; then
     echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
-    # Remove dangling images (those that are not tagged and not used by any container)
     docker image prune -f
-    # Remove unused volumes / force the system prune for old images as well.
     docker volume prune -f && docker system prune --force --filter "until=72h" --all
     echo "Docker images and volumes cleanup completed."
   else
@@ -45,193 +74,431 @@ cleanup_docker() {
 }
 
 cleanup_network() {
-  for node in $(seq 0 $((NUM_NODES-1))); do
+  local max_nodes=${NUM_NODES:-2}
-    if docker pr -a -q -f name="node${node}" | grep -q .; then
+  for node in $(seq 0 $((max_nodes - 1))); do
-      docker stop "node${node}"
+    if docker ps -a -q -f name="node${node}" | grep -q .; then
+      docker stop "node${node}" || true
     fi
   done
-  if docker network ls | grep docker-net; then
+  if docker network ls | grep -q docker-net; then
-    docker network rm docker-net
+    docker network rm docker-net || true
   fi
 }
 
-# Call the cleanup docker function
+is_multi_node() {
+  local cmds="$1"
+  # Primary signal: NUM_NODES environment variable set by the pipeline
+  if [[ "${NUM_NODES:-1}" -gt 1 ]]; then
+    return 0
+  fi
+  # Fallback: detect the bracket syntax structurally
+  # Pattern: [...] && [...] (per-node command arrays)
+  if [[ "$cmds" =~ \[.*\].*\&\&.*\[.*\] ]]; then
+    return 0
+  fi
+  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
+#
+# When commands are passed through Buildkite -> shell -> $* -> bash -c,
+# quotes around multi-word pytest -m/-k expressions get stripped:
+#   pytest -v -s -m 'not cpu_test' v1/core
+# becomes:
+#   pytest -v -s -m not cpu_test v1/core
+#
+# pytest then interprets "cpu_test" as a file path, not part of the marker.
+#
+# This function detects unquoted expressions after -m/-k and re-quotes them
+# by collecting tokens until a recognizable boundary is reached:
+#   - test path (contains '/')
+#   - test file (ends with '.py')
+#   - another pytest flag (--xxx or -x single-char flags)
+#   - command separator (&& || ; |)
+#   - environment variable assignment (FOO=bar)
+#
+# Single-word markers (e.g. -m cpu_test, -m hybrid_model) pass through
+# unquoted since they have no spaces and work fine.
+#
+# Already-quoted expressions (containing literal single quotes) are passed
+# through untouched to avoid double-quoting values injected by
+# apply_rocm_test_overrides.
+#
+# NOTE: This ONLY fixes -m/-k flags. It cannot recover arbitrary inner
+# double-quotes stripped by the calling shell (see header comment).
+# Use VLLM_TEST_COMMANDS to avoid the problem entirely.
+###############################################################################
+re_quote_pytest_markers() {
+  local input="$1"
+  local output=""
+  local collecting=false
+  local marker_buf=""
+
+  # Strip backslash-newline continuations, then flatten remaining newlines
+  local flat="${input//$'\\\n'/ }"
+  flat="${flat//$'\n'/ }"
+
+  # Disable globbing to prevent *.py etc. from expanding during read -ra
+  local restore_glob
+  restore_glob="$(shopt -p -o noglob 2>/dev/null || true)"
+  set -o noglob
+  local -a words
+  read -ra words <<< "$flat"
+  eval "$restore_glob"
+
+  for word in "${words[@]}"; do
+    if $collecting; then
+      # If the token we're about to collect already contains a literal
+      # single quote, the expression was already quoted upstream.
+      # Flush and stop collecting.
+      if [[ "$word" == *"'"* ]]; then
+        if [[ -n "$marker_buf" ]]; then
+          # Should not normally happen (partial buf + quote), flush raw
+          output+="${marker_buf} "
+          marker_buf=""
+        fi
+        output+="${word} "
+        collecting=false
+        continue
+      fi
+
+      local is_boundary=false
+      case "$word" in
+        # Line-continuation artifact
+        "\\")
+          is_boundary=true ;;
+        # Command separators
+        "&&"|"||"|";"|"|")
+          is_boundary=true ;;
+        # Long flags (--ignore, --shard-id, etc.)
+        --*)
+          is_boundary=true ;;
+        # Short flags (-v, -s, -x, etc.) but NOT negative marker tokens
+        # like "not" which don't start with "-". Also skip -k/-m which
+        # would start a new marker (handled below).
+        -[a-zA-Z])
+          is_boundary=true ;;
+        # Test path (contains /)
+        */*)
+          is_boundary=true ;;
+        # Test file (ends with .py, possibly with ::method)
+        *.py|*.py::*)
+          is_boundary=true ;;
+        # Environment variable assignment preceding a command (FOO=bar)
+        *=*)
+          # Only treat as boundary if it looks like VAR=value, not
+          # pytest filter expressions like num_gpus=2 inside markers
+          if [[ "$word" =~ ^[A-Z_][A-Z0-9_]*= ]]; then
+            is_boundary=true
+          fi
+          ;;
+      esac
+
+      if $is_boundary; then
+        # Flush the collected marker expression
+        if [[ "$marker_buf" == *" "* || "$marker_buf" == *"("* ]]; then
+          output+="'${marker_buf}' "
+        else
+          output+="${marker_buf} "
+        fi
+        collecting=false
+        marker_buf=""
+        # Check if this boundary word itself starts a new -m/-k
+        if [[ "$word" == "-m" || "$word" == "-k" ]]; then
+          output+="${word} "
+          collecting=true
+        # Drop stray backslash tokens silently
+        elif [[ "$word" == "\\" ]]; then
+          :
+        else
+          output+="${word} "
+        fi
+      else
+        # Accumulate into marker buffer
+        if [[ -n "$marker_buf" ]]; then
+          marker_buf+=" ${word}"
+        else
+          marker_buf="${word}"
+        fi
+      fi
+    elif [[ "$word" == "-m" || "$word" == "-k" ]]; then
+      output+="${word} "
+      collecting=true
+      marker_buf=""
+    else
+      output+="${word} "
+    fi
+  done
+
+  # Flush any trailing marker expression (marker at end of command)
+  if $collecting && [[ -n "$marker_buf" ]]; then
+    if [[ "$marker_buf" == *" "* || "$marker_buf" == *"("* ]]; then
+      output+="'${marker_buf}'"
+    else
+      output+="${marker_buf}"
+    fi
+  fi
+
+  echo "${output% }"
+}
+
+###############################################################################
+# ROCm-specific pytest command rewrites
+#
+# These apply ignore flags and environment overrides for tests that are not
+# yet supported or behave differently on ROCm hardware. Kept as a single
+# function so new exclusions are easy to add in one place.
+###############################################################################
+
+apply_rocm_test_overrides() {
+  local cmds="$1"
+
+  # --- Model registry filter ---
+  if [[ $cmds == *"pytest -v -s models/test_registry.py"* ]]; then
+    cmds=${cmds//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
+  fi
+
+  # --- LoRA: disable custom paged attention ---
+  if [[ $cmds == *"pytest -v -s lora"* ]]; then
+    cmds=${cmds//"pytest -v -s lora"/"VLLM_ROCM_CUSTOM_PAGED_ATTN=0 pytest -v -s lora"}
+  fi
+
+  # --- Kernel ignores ---
+  if [[ $cmds == *" kernels/core"* ]]; then
+    cmds="${cmds} \
+    --ignore=kernels/core/test_fused_quant_layernorm.py \
+    --ignore=kernels/core/test_permute_cols.py"
+  fi
+
+  if [[ $cmds == *" kernels/attention"* ]]; then
+    cmds="${cmds} \
+    --ignore=kernels/attention/test_attention_selector.py \
+    --ignore=kernels/attention/test_encoder_decoder_attn.py \
+    --ignore=kernels/attention/test_flash_attn.py \
+    --ignore=kernels/attention/test_flashinfer.py \
+    --ignore=kernels/attention/test_prefix_prefill.py \
+    --ignore=kernels/attention/test_cascade_flash_attn.py \
+    --ignore=kernels/attention/test_mha_attn.py \
+    --ignore=kernels/attention/test_lightning_attn.py \
+    --ignore=kernels/attention/test_attention.py"
+  fi
+
+  if [[ $cmds == *" kernels/quantization"* ]]; then
+    cmds="${cmds} \
+    --ignore=kernels/quantization/test_int8_quant.py \
+    --ignore=kernels/quantization/test_machete_mm.py \
+    --ignore=kernels/quantization/test_block_fp8.py \
+    --ignore=kernels/quantization/test_block_int8.py \
+    --ignore=kernels/quantization/test_marlin_gemm.py \
+    --ignore=kernels/quantization/test_cutlass_scaled_mm.py \
+    --ignore=kernels/quantization/test_int8_kernel.py"
+  fi
+
+  if [[ $cmds == *" kernels/mamba"* ]]; then
+    cmds="${cmds} \
+    --ignore=kernels/mamba/test_mamba_mixer2.py \
+    --ignore=kernels/mamba/test_causal_conv1d.py \
+    --ignore=kernels/mamba/test_mamba_ssm_ssd.py"
+  fi
+
+  if [[ $cmds == *" kernels/moe"* ]]; then
+    cmds="${cmds} \
+    --ignore=kernels/moe/test_moe.py \
+    --ignore=kernels/moe/test_cutlass_moe.py \
+    --ignore=kernels/moe/test_triton_moe_ptpc_fp8.py"
+  fi
+
+  # --- Entrypoint ignores ---
+  if [[ $cmds == *" entrypoints/openai "* ]]; then
+    cmds=${cmds//" entrypoints/openai "/" entrypoints/openai \
+    --ignore=entrypoints/openai/test_audio.py \
+    --ignore=entrypoints/openai/test_shutdown.py \
+    --ignore=entrypoints/openai/test_completion.py \
+    --ignore=entrypoints/openai/test_models.py \
+    --ignore=entrypoints/openai/test_lora_adapters.py \
+    --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
+    --ignore=entrypoints/openai/test_root_path.py \
+    --ignore=entrypoints/openai/test_tokenization.py \
+    --ignore=entrypoints/openai/test_prompt_validation.py "}
+  fi
+
+  if [[ $cmds == *" entrypoints/llm "* ]]; then
+    cmds=${cmds//" entrypoints/llm "/" entrypoints/llm \
+    --ignore=entrypoints/llm/test_chat.py \
+    --ignore=entrypoints/llm/test_accuracy.py \
+    --ignore=entrypoints/llm/test_init.py \
+    --ignore=entrypoints/llm/test_prompt_validation.py "}
+  fi
+
+  # Clean up escaped newlines from --ignore appends
+  cmds=$(echo "$cmds" | sed 's/ \\ / /g')
+
+  echo "$cmds"
+}
+
+###############################################################################
+# Main
+###############################################################################
+
+# --- GPU initialization ---
+echo "--- Confirming Clean Initial State"
+wait_for_clean_gpus
+
+echo "--- ROCm info"
+rocminfo
+
+# --- Docker housekeeping ---
 cleanup_docker
 
 echo "--- Resetting GPUs"
-
 echo "reset" > /opt/amdgpu/etc/gpu_state
+wait_for_clean_gpus
 
-while true; do
+# --- Pull test image ---
-        sleep 3
-        if grep -q clean /opt/amdgpu/etc/gpu_state; then
-                echo "GPUs state is \"clean\""
-                break
-        fi
-done
-
 echo "--- Pulling container"
 image_name="rocm/vllm-ci:${BUILDKITE_COMMIT}"
 container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 docker pull "${image_name}"
 
 remove_docker_container() {
-   docker rm -f "${container_name}" || docker image rm -f "${image_name}" || true
+  docker rm -f "${container_name}" || docker image rm -f "${image_name}" || true
 }
 trap remove_docker_container EXIT
 
+# --- Prepare commands ---
 echo "--- Running container"
 
 HF_CACHE="$(realpath ~)/huggingface"
 mkdir -p "${HF_CACHE}"
 HF_MOUNT="/root/.cache/huggingface"
 
-commands=$@
+# ---- Command source selection ----
+# Prefer VLLM_TEST_COMMANDS (preserves all inner quoting intact).
+# Fall back to $* for backward compatibility, but warn that inner
+# double-quotes will have been stripped by the calling shell.
+if [[ -n "${VLLM_TEST_COMMANDS:-}" ]]; then
+  commands="${VLLM_TEST_COMMANDS}"
+  echo "Commands sourced from VLLM_TEST_COMMANDS (quoting preserved)"
+else
+  commands="$*"
+  if [[ -z "$commands" ]]; then
+    echo "Error: No test commands provided." >&2
+    echo "Usage:" >&2
+    echo "  Preferred:  VLLM_TEST_COMMANDS='...' bash $0" >&2
+    echo "  Legacy:     bash $0 \"commands here\"" >&2
+    exit 1
+  fi
+  echo "Commands sourced from positional args (legacy mode)"
+  echo "WARNING: Inner double-quotes in the command string may have been"
+  echo "  stripped by the calling shell. If you see syntax errors, switch to:"
+  echo "  export VLLM_TEST_COMMANDS='your commands here'"
+  echo "  bash $0"
+fi
+
 echo "Raw commands: $commands"
 
-commands=${commands//"pytest -v -s basic_correctness/test_basic_correctness.py"/"pytest -v -s basic_correctness/test_basic_correctness.py"}
+# Fix quoting before ROCm overrides (so overrides see correct structure)
+commands=$(re_quote_pytest_markers "$commands")
+echo "After re-quoting: $commands"
 
-if [[ $commands == *"pytest -v -s models/test_registry.py"* ]]; then
+commands=$(apply_rocm_test_overrides "$commands")
-  commands=${commands//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
-fi
-
-commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"pytest -v -s compile/test_basic_correctness.py"}
-
-if [[ $commands == *"pytest -v -s lora"* ]]; then
-  commands=${commands//"pytest -v -s lora"/"VLLM_ROCM_CUSTOM_PAGED_ATTN=0 pytest -v -s lora"}
-fi
-
-#ignore certain kernels tests
-if [[ $commands == *" kernels/core"* ]]; then
-  commands="${commands} \
-  --ignore=kernels/core/test_fused_quant_layernorm.py \
-  --ignore=kernels/core/test_permute_cols.py"
-fi
-
-if [[ $commands == *" kernels/attention"* ]]; then
-  commands="${commands} \
-  --ignore=kernels/attention/test_attention_selector.py \
-  --ignore=kernels/attention/test_encoder_decoder_attn.py \
-  --ignore=kernels/attention/test_flash_attn.py \
-  --ignore=kernels/attention/test_flashinfer.py \
-  --ignore=kernels/attention/test_prefix_prefill.py \
-  --ignore=kernels/attention/test_cascade_flash_attn.py \
-  --ignore=kernels/attention/test_mha_attn.py \
-  --ignore=kernels/attention/test_lightning_attn.py \
-  --ignore=kernels/attention/test_attention.py"
-fi
-
-if [[ $commands == *" kernels/quantization"* ]]; then
-  commands="${commands} \
-  --ignore=kernels/quantization/test_int8_quant.py \
-  --ignore=kernels/quantization/test_machete_mm.py \
-  --ignore=kernels/quantization/test_block_fp8.py \
-  --ignore=kernels/quantization/test_block_int8.py \
-  --ignore=kernels/quantization/test_marlin_gemm.py \
-  --ignore=kernels/quantization/test_cutlass_scaled_mm.py \
-  --ignore=kernels/quantization/test_int8_kernel.py"
-fi
-
-if [[ $commands == *" kernels/mamba"* ]]; then
-  commands="${commands} \
-  --ignore=kernels/mamba/test_mamba_mixer2.py \
-  --ignore=kernels/mamba/test_causal_conv1d.py \
-  --ignore=kernels/mamba/test_mamba_ssm_ssd.py"
-fi
-
-if [[ $commands == *" kernels/moe"* ]]; then
-  commands="${commands} \
-  --ignore=kernels/moe/test_moe.py \
-  --ignore=kernels/moe/test_cutlass_moe.py \
-  --ignore=kernels/moe/test_triton_moe_ptpc_fp8.py"
-fi
-
-#ignore certain Entrypoints/openai tests
-if [[ $commands == *" entrypoints/openai "* ]]; then
-  commands=${commands//" entrypoints/openai "/" entrypoints/openai \
-  --ignore=entrypoints/openai/test_audio.py \
-  --ignore=entrypoints/openai/test_shutdown.py \
-  --ignore=entrypoints/openai/test_completion.py \
-  --ignore=entrypoints/openai/test_models.py \
-  --ignore=entrypoints/openai/test_lora_adapters.py \
-  --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
-  --ignore=entrypoints/openai/test_root_path.py \
-  --ignore=entrypoints/openai/test_tokenization.py \
-  --ignore=entrypoints/openai/test_prompt_validation.py "}
-fi
-
-#ignore certain Entrypoints/llm tests
-if [[ $commands == *" entrypoints/llm "* ]]; then
-  commands=${commands//" entrypoints/llm "/" entrypoints/llm \
-  --ignore=entrypoints/llm/test_chat.py \
-  --ignore=entrypoints/llm/test_accuracy.py \
-  --ignore=entrypoints/llm/test_init.py \
-  --ignore=entrypoints/llm/test_prompt_validation.py "}
-fi
-
-commands=$(echo "$commands" | sed 's/ \\ / /g')
 echo "Final commands: $commands"
 
-# --ignore=entrypoints/openai/test_encoder_decoder.py \
-# --ignore=entrypoints/openai/test_embedding.py \
-# --ignore=entrypoints/openai/test_oot_registration.py
-# --ignore=entrypoints/openai/test_accuracy.py \
-# --ignore=entrypoints/openai/test_models.py <= Fails on MI250 but passes on MI300 as of 2025-03-13
-
-
 MYPYTHONPATH=".."
 
-# Test that we're launching on the machine that has
+# Verify GPU access
-# proper access to GPUs
 render_gid=$(getent group render | cut -d: -f3)
 if [[ -z "$render_gid" ]]; then
   echo "Error: 'render' group not found. This is required for GPU access." >&2
   exit 1
 fi
 
-if [[ $commands == *"VLLM_TEST_GROUP_NAME=mi325_4-2-node-tests-4-gpus-in-total"* ]]; then
+# --- RDMA device passthrough (conditional) ---
+# If the host has RDMA devices, pass them through so tests like
+# test_moriio_connector can access ibverbs. On hosts without RDMA
+# hardware the tests will gracefully skip via _rdma_available().
+RDMA_FLAGS=""
+if [ -d /dev/infiniband ]; then
+  echo "RDMA devices detected on host, enabling passthrough"
+  RDMA_FLAGS="--device /dev/infiniband --cap-add=IPC_LOCK"
+else
+  echo "No RDMA devices found on host, RDMA tests will be skipped"
+fi
 
+# --- Route: multi-node vs single-node ---
+if is_multi_node "$commands"; then
+  echo "--- Multi-node job detected"
   export DCKR_VER=$(docker --version | sed 's/Docker version \(.*\), build .*/\1/')
 
-  if [[ "$commands" =~ ^(.*)"["(.*)"] && ["(.*)"]"$ ]]; then
+  # Parse the bracket syntax:  prefix ; [node0_cmds] && [node1_cmds]
-      prefix=$( echo "${BASH_REMATCH[1]}" | sed 's/;//g')
+  #   BASH_REMATCH[1] = prefix (everything before first bracket)
-      echo "PREFIX: ${prefix}"
+  #   BASH_REMATCH[2] = comma-separated node0 commands
-      export composite_command="(command rocm-smi || true)"
+  #   BASH_REMATCH[3] = comma-separated node1 commands
-      myIFS=$IFS
+  if [[ "$commands" =~ ^(.*)\[(.*)"] && ["(.*)\]$ ]]; then
-      IFS=','
+    prefix=$(echo "${BASH_REMATCH[1]}" | sed 's/;//g')
-      read -ra node0 <<< ${BASH_REMATCH[2]}
+    echo "PREFIX: ${prefix}"
-      read -ra node1 <<< ${BASH_REMATCH[3]}
-      IFS=$myIFS
-      for i in "${!node0[@]}";do 
-        command_node_0=$(echo ${node0[i]} | sed 's/\"//g')
-        command_node_1=$(echo ${node1[i]} | sed 's/\"//g')
 
-        export commands="./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 ${image_name} '${command_node_0}' '${command_node_1}'"
+    export composite_command="(command rocm-smi || true)"
-        echo "COMMANDS: ${commands}"
+    saved_IFS=$IFS
-        composite_command=$(echo "${composite_command} && ${commands}")
+    IFS=','
-      done
+    read -ra node0 <<< "${BASH_REMATCH[2]}"
-      /bin/bash -c "${composite_command}"
+    read -ra node1 <<< "${BASH_REMATCH[3]}"
-      cleanup_network
+    IFS=$saved_IFS
+
+    if [[ ${#node0[@]} -ne ${#node1[@]} ]]; then
+      echo "Warning: node0 has ${#node0[@]} commands, node1 has ${#node1[@]}. They will be paired by index."
+    fi
+
+    for i in "${!node0[@]}"; do
+      command_node_0=$(echo "${node0[i]}" | sed 's/\"//g')
+      command_node_1=$(echo "${node1[i]}" | sed 's/\"//g')
+
+      step_cmd="./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 ${image_name} '${command_node_0}' '${command_node_1}'"
+      echo "COMMANDS: ${step_cmd}"
+      composite_command="${composite_command} && ${step_cmd}"
+    done
+
+    /bin/bash -c "${composite_command}"
+    exit_code=$?
+    cleanup_network
+    handle_pytest_exit "$exit_code"
   else
-      echo "Failed to parse node commands! Exiting."
+    echo "Multi-node job detected but failed to parse bracket command syntax."
-      cleanup_network
+    echo "Expected format: prefix ; [node0_cmd1, node0_cmd2] && [node1_cmd1, node1_cmd2]"
-      exit 111
+    echo "Got: $commands"
+    cleanup_network
+    exit 111
   fi
 else
+  echo "--- Single-node job"
   echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
   docker run \
-          --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+    --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
-          --network=host \
+    $RDMA_FLAGS \
-          --shm-size=16gb \
+    --network=host \
-          --group-add "$render_gid" \
+    --shm-size=16gb \
-          --rm \
+    --group-add "$render_gid" \
-          -e HF_TOKEN \
+    --rm \
-          -e AWS_ACCESS_KEY_ID \
+    -e HF_TOKEN \
-          -e AWS_SECRET_ACCESS_KEY \
+    -e AWS_ACCESS_KEY_ID \
-          -v "${HF_CACHE}:${HF_MOUNT}" \
+    -e AWS_SECRET_ACCESS_KEY \
-          -e "HF_HOME=${HF_MOUNT}" \
+    -v "${HF_CACHE}:${HF_MOUNT}" \
-          -e "PYTHONPATH=${MYPYTHONPATH}" \
+    -e "HF_HOME=${HF_MOUNT}" \
-          --name "${container_name}" \
+    -e "PYTHONPATH=${MYPYTHONPATH}" \
-          "${image_name}" \
+    --name "${container_name}" \
-          /bin/bash -c "${commands}"
+    "${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/run-multi-node-test.sh

@@ -67,7 +67,7 @@ start_nodes() {
         # 3. map the huggingface cache directory to the container
         # 3. assign ip addresses to the containers (head node: 192.168.10.10, worker nodes:
         #    starting from 192.168.10.11)
-        docker run -d "$GPU_DEVICES" --shm-size=10.24gb -e HF_TOKEN \
+        docker run -d $GPU_DEVICES --shm-size=10.24gb -e HF_TOKEN \
             -v ~/.cache/huggingface:/root/.cache/huggingface --name "node$node" \
             --network docker-net --ip 192.168.10.$((10 + $node)) --rm "$DOCKER_IMAGE" \
             /bin/bash -c "tail -f /dev/null"

.buildkite/scripts/run-prime-rl-test.sh

@@ -1,64 +0,0 @@
-#!/bin/bash
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-# Setup script for Prime-RL integration tests
-# This script prepares the environment for running Prime-RL tests with nightly vLLM
-
-set -euo pipefail
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd "${SCRIPT_DIR}/../.." && pwd)"
-PRIME_RL_REPO="https://github.com/PrimeIntellect-ai/prime-rl.git"
-PRIME_RL_DIR="${REPO_ROOT}/prime-rl"
-
-if command -v rocm-smi &> /dev/null || command -v rocminfo &> /dev/null; then
-    echo "AMD GPU detected. Prime-RL currently only supports NVIDIA. Skipping..."
-    exit 0
-fi
-
-echo "Setting up Prime-RL integration test environment..."
-
-# Clean up any existing Prime-RL directory
-if [ -d "${PRIME_RL_DIR}" ]; then
-    echo "Removing existing Prime-RL directory..."
-    rm -rf "${PRIME_RL_DIR}"
-fi
-
-# Install UV if not available
-if ! command -v uv &> /dev/null; then
-    echo "Installing UV package manager..."
-    curl -LsSf https://astral.sh/uv/install.sh | sh
-    source "$HOME"/.local/bin/env
-fi
-
-# Clone Prime-RL repository at specific branch for reproducible tests
-PRIME_RL_BRANCH="integ-vllm-main"
-echo "Cloning Prime-RL repository at branch: ${PRIME_RL_BRANCH}..."
-git clone --branch "${PRIME_RL_BRANCH}" --single-branch "${PRIME_RL_REPO}" "${PRIME_RL_DIR}"
-cd "${PRIME_RL_DIR}"
-
-echo "Setting up UV project environment..."
-export UV_PROJECT_ENVIRONMENT=/usr/local
-ln -s /usr/bin/python3 /usr/local/bin/python
-
-# Remove vllm pin from pyproject.toml
-echo "Removing vllm pin from pyproject.toml..."
-sed -i '/vllm==/d' pyproject.toml
-
-# Sync Prime-RL dependencies
-echo "Installing Prime-RL dependencies..."
-uv sync --inexact && uv sync --inexact --all-extras
-
-# Verify installation
-echo "Verifying installations..."
-uv run python -c "import vllm; print(f'vLLM version: {vllm.__version__}')"
-uv run python -c "import prime_rl; print('Prime-RL imported successfully')"
-
-echo "Prime-RL integration test environment setup complete!"
-
-echo "Running Prime-RL integration tests..."
-export WANDB_MODE=offline # this makes this test not require a WANDB_API_KEY
-uv run pytest -vs tests/integration/test_rl.py -m gpu
-
-echo "Prime-RL integration tests completed!"

.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_prefetch_offload.sh

@@ -0,0 +1,57 @@
+#!/usr/bin/env bash
+set -euxo pipefail
+
+# Nightly e2e test for prefetch offloading with a MoE model.
+# Runs DeepSeek-V2-Lite with prefetch offloading of MoE expert weights
+# and validates GSM8K accuracy matches baseline (no offloading).
+#
+# args: [THRESHOLD] [NUM_QUESTIONS] [START_PORT]
+THRESHOLD=${1:-0.25}
+NUM_Q=${2:-1319}
+PORT=${3:-8030}
+OUT_DIR=${OUT_DIR:-/tmp/vllm-scheduled}
+mkdir -p "${OUT_DIR}"
+
+wait_for_server() {
+  local port=$1
+  timeout 600 bash -c '
+    until curl -sf "http://127.0.0.1:'"$port"'/health" > /dev/null; do
+      sleep 1
+    done'
+}
+
+MODEL="deepseek-ai/DeepSeek-V2-Lite"
+
+cleanup() {
+  if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
+    kill "${SERVER_PID}" 2>/dev/null || true
+    for _ in {1..20}; do
+      kill -0 "${SERVER_PID}" 2>/dev/null || break
+      sleep 0.5
+    done
+    kill -9 "${SERVER_PID}" 2>/dev/null || true
+  fi
+}
+trap cleanup EXIT
+
+vllm serve "$MODEL" \
+  --max-model-len 2048 \
+  --offload-group-size 8 \
+  --offload-num-in-group 2 \
+  --offload-prefetch-step 1 \
+  --offload-params w13_weight w2_weight \
+  --port "$PORT" &
+SERVER_PID=$!
+wait_for_server "$PORT"
+
+TAG=$(echo "$MODEL" | tr '/: \\n' '_____')
+OUT="${OUT_DIR}/${TAG}_prefetch_offload.json"
+python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port "$PORT" --num-questions "${NUM_Q}" --save-results "${OUT}"
+python3 - <<PY
+import json; acc=json.load(open('${OUT}'))['accuracy']
+print(f"${MODEL} prefetch_offload: accuracy {acc:.3f}")
+assert acc >= ${THRESHOLD}, f"${MODEL} prefetch_offload accuracy {acc}"
+PY
+
+cleanup
+SERVER_PID=

.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" $DIST_DIR
 echo "Wheels copied to local directory"
-# generate source tarball
+# generate source distribution using setup.py
-git archive --format=tar.gz --output="$DIST_DIR/vllm-${PURE_VERSION}.tar.gz" "$BUILDKITE_COMMIT"
+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 "$PYPI_WHEEL_FILES"
+python3 -m twine check "$PYPI_WHEEL_FILES" "$SDIST_FILE"
-python3 -m twine upload --non-interactive --verbose "$PYPI_WHEEL_FILES"
+python3 -m twine upload --non-interactive --verbose "$PYPI_WHEEL_FILES" "$SDIST_FILE"
-echo "Wheels uploaded to PyPI"
+echo "Wheels and source distribution uploaded to PyPI"

.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
@@ -103,8 +106,8 @@ steps:
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
   # NEW rlhf examples
   - cd 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_async_new_apis.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_ipc.py
 
 - label: Distributed Tests (8 GPUs)(H100)
   timeout_in_minutes: 10
@@ -146,6 +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 --- 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
 
@@ -165,6 +169,7 @@ steps:
   num_devices: 2
   num_nodes: 2
   no_plugin: true
+  optional: true # TODO: revert once infra issue solved
   source_file_dependencies:
   - vllm/distributed/
   - vllm/engine/
@@ -208,7 +213,20 @@ 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: Pipeline + Context Parallelism (4 GPUs))
+- 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"
   num_devices: 4

.buildkite/test_areas/e2e_integration.yaml

@@ -29,15 +29,11 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
 
-- label: Prime-RL Integration (2 GPUs)
+- label: DeepSeek V2-Lite Prefetch Offload Accuracy (H100)
-  timeout_in_minutes: 30
+  timeout_in_minutes: 60
+  device: h100
   optional: true
-  soft_fail: true
+  num_devices: 1
-  num_devices: 2
   working_dir: "/vllm-workspace"
-  source_file_dependencies:
-  - vllm/
-  - .buildkite/scripts/run-prime-rl-test.sh
   commands:
-    - nvidia-smi
+  - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_prefetch_offload.sh 0.25 200 8030
-    - bash .buildkite/scripts/run-prime-rl-test.sh

.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/
@@ -28,3 +28,43 @@ steps:
     - pytest -v -s v1/engine/test_preprocess_error_handling.py
     # Run the rest of v1/engine tests
     - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+      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
@@ -65,6 +75,11 @@ steps:
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/pooling
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Entrypoints Integration (Responses API)
   timeout_in_minutes: 50
@@ -82,6 +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/expert_parallelism.yaml

@@ -21,3 +21,18 @@ steps:
   commands:
   - pytest -v -s distributed/test_eplb_execute.py
   - pytest -v -s distributed/test_eplb_spec_decode.py
+
+- label: Elastic EP Scaling Test
+  timeout_in_minutes: 20
+  device: b200
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/executor/
+  - vllm/compilation/
+  - tests/distributed/
+  commands:
+  - pytest -v -s distributed/test_elastic_ep.py

.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
@@ -70,7 +71,7 @@ steps:
   - tests/kernels/moe/test_batched_deepgemm.py
   - tests/kernels/attention/test_deepgemm_attention.py
   commands:
-    - pytest -v -s kernels/quantization/test_block_fp8.py -k deep_gemm
+    - pytest -v -s kernels/quantization/test_block_fp8.py
     - pytest -v -s kernels/moe/test_deepgemm.py
     - pytest -v -s kernels/moe/test_batched_deepgemm.py
     - pytest -v -s kernels/attention/test_deepgemm_attention.py
@@ -115,6 +116,7 @@ steps:
     - 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_flashinfer_moe.py
     - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
     # e2e
     - pytest -v -s tests/models/quantization/test_nvfp4.py
@@ -154,8 +156,6 @@ steps:
   commands:
     - pytest -v -s kernels/moe/test_deepep_deepgemm_moe.py
     - pytest -v -s kernels/moe/test_deepep_moe.py
-    - pytest -v -s kernels/moe/test_pplx_cutlass_moe.py
-    # - pytest -v -s kernels/moe/test_pplx_moe.py - failing on main
 
 - label: Kernels Fp4 MoE Test (B200)
   timeout_in_minutes: 60

.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)
+# - label: LM Eval Large Models (4 GPUs)(A100)
-  device: a100
+#   device: a100
-  optional: true
+#   optional: true
-  num_devices: 4
+#   num_devices: 4
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+#   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  source_file_dependencies:
+#   source_file_dependencies:
-  - csrc/
+#   - csrc/
-  - vllm/model_executor/layers/quantization
+#   - vllm/model_executor/layers/quantization
-  commands:
+#   commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+#   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
+#   - 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
@@ -73,3 +73,29 @@ steps:
   num_devices: 2
   commands:
     - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor-dp-ep/config-b200.txt
+
+- label: GPQA Eval (GPT-OSS) (H100)
+  timeout_in_minutes: 120
+  device: h100
+  optional: true
+  num_devices: 2
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  - tests/evals/gpt_oss/
+  commands:
+    - uv pip install --system 'gpt-oss[eval]==0.0.5'
+    - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-h100.txt
+
+- label: GPQA Eval (GPT-OSS) (B200)
+  timeout_in_minutes: 120
+  device: b200
+  optional: true
+  num_devices: 2
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  - tests/evals/gpt_oss/
+  commands:
+    - uv pip install --system 'gpt-oss[eval]==0.0.5'
+    - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-b200.txt

.buildkite/test_areas/misc.yaml

@@ -9,6 +9,7 @@ steps:
     - tests/v1
   commands:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     # split the test to avoid interference
     - pytest -v -s -m 'not cpu_test' v1/core
     - pytest -v -s v1/executor
@@ -16,6 +17,7 @@ steps:
     - pytest -v -s v1/sample
     - pytest -v -s v1/logits_processors
     - pytest -v -s v1/worker
+    # TODO: create another `optional` test group for slow tests
     - pytest -v -s -m 'not slow_test' v1/spec_decode
     - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
     - pytest -v -s -m 'not cpu_test' v1/metrics
@@ -25,6 +27,11 @@ steps:
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: V1 Others (CPU)
   depends_on:
@@ -80,6 +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
@@ -108,9 +120,11 @@ steps:
   timeout_in_minutes: 50
   source_file_dependencies:
   - vllm/
+  - tests/detokenizer
   - tests/multimodal
   - tests/utils_
   commands:
+  - pytest -v -s detokenizer
   - pytest -v -s -m 'not cpu_test' multimodal
   - pytest -v -s utils_
 
@@ -145,20 +159,6 @@ steps:
   - pytest -v -s transformers_utils
   - pytest -v -s config
 
-- label: GPT-OSS Eval (B200)
-  timeout_in_minutes: 60
-  working_dir: "/vllm-workspace/"
-  device: b200
-  optional: true
-  source_file_dependencies:
-  - tests/evals/gpt_oss
-  - vllm/model_executor/models/gpt_oss.py
-  - vllm/model_executor/layers/quantization/mxfp4.py
-  - vllm/v1/attention/backends/flashinfer.py
-  commands:
-    - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
-
 - label: Batch Invariance (H100)
   timeout_in_minutes: 25
   device: h100

.buildkite/test_areas/models_language.yaml

@@ -55,6 +55,15 @@ steps:
     - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
     - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+      commands:
+      - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
+      - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
+      - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
 - label: Language Models Test (PPL)
   timeout_in_minutes: 110
@@ -73,6 +82,11 @@ steps:
   - tests/models/language/pooling
   commands:
     - pytest -v -s models/language/pooling -m 'not core_model'
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Language Models Test (MTEB)
   timeout_in_minutes: 110

.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

@@ -15,10 +15,17 @@ 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
+  - pytest -v -s plugins_tests/test_bge_m3_sparse_io_processor_plugins.py
+  - pip uninstall bge_m3_sparse_plugin -y
   # end io_processor plugins test
   # begin stat_logger plugins test
   - pip install -e ./plugins/vllm_add_dummy_stat_logger
@@ -32,3 +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/samplers.yaml

@@ -18,4 +18,4 @@ steps:
       depends_on:
       - image-build-amd
       commands:
-      - pytest -v -s -m 'not skip_v1' samplers
+      - pytest -v -s samplers

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

.github/.bc-linter.yml

@@ -1,24 +0,0 @@
-# doc: https://github.com/pytorch/test-infra/blob/main/tools/stronghold/docs/bc_linter_config.md
-version: 1
-paths:
-# We temporarily disable globally, and will only enable with `annotations.include`
-# include:
-#   - "vllm/v1/attetion/*.py"
-#   - "vllm/v1/core/*.py"
-exclude:
-  - "**/*.py"
-
-scan:
-  functions: true        # check free functions and methods
-  classes: true          # check classes/dataclasses
-  public_only: true      # ignore names starting with "_" at any level
-
-annotations:
-  include:               # decorators that force‑include a symbol
-    - name: "bc_linter_include"  # matched by simple name or dotted suffix
-      propagate_to_members: false # for classes, include methods/inner classes
-  exclude:               # decorators that force‑exclude a symbol
-    - name: "bc_linter_skip"     # matched by simple name or dotted suffix
-      propagate_to_members: true  # for classes, exclude methods/inner classes
-
-excluded_violations: []  # e.g. ["ParameterRenamed", "FieldTypeChanged"]

.github/CODEOWNERS

@@ -2,17 +2,17 @@
 # for more info about CODEOWNERS file
 
 # This lists cover the "core" components of vLLM that require careful review
-/vllm/compilation @zou3519 @youkaichao @ProExpertProg
+/vllm/compilation @zou3519 @youkaichao @ProExpertProg @BoyuanFeng
 /vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery
 /vllm/lora @jeejeelee
-/vllm/model_executor/layers/attention @LucasWilkinson
+/vllm/model_executor/layers/attention @LucasWilkinson @MatthewBonanni
 /vllm/model_executor/layers/fused_moe @mgoin @pavanimajety
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety
 /vllm/model_executor/layers/mamba @tdoublep
 /vllm/model_executor/model_loader @22quinn
 /vllm/model_executor/layers/batch_invariant.py @yewentao256 
 /vllm/multimodal @DarkLight1337 @ywang96 @NickLucche @tjtanaa
-/vllm/vllm_flash_attn @LucasWilkinson
+/vllm/vllm_flash_attn @LucasWilkinson @MatthewBonanni
 CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
@@ -43,22 +43,25 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /vllm/tool_parsers @aarnphm @chaunceyjiang
 
 # vLLM V1
-/vllm/v1/attention @LucasWilkinson
+/vllm/v1/attention @LucasWilkinson @MatthewBonanni
 /vllm/v1/attention/backend.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @njhill
 /vllm/v1/attention/backends/mla @pavanimajety
 /vllm/v1/attention/backends/flashinfer.py @mgoin @pavanimajety
 /vllm/v1/attention/backends/triton_attn.py @tdoublep
 /vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC @orozery
 /vllm/v1/sample @22quinn @houseroad @njhill
-/vllm/v1/spec_decode @benchislett @luccafong
+/vllm/v1/spec_decode @benchislett @luccafong @MatthewBonanni
 /vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
 /vllm/v1/kv_cache_interface.py @heheda12345
 /vllm/v1/kv_offload @ApostaC @orozery
-/vllm/v1/worker/gpu/kv_connector.py @orozery
+/vllm/v1/engine @njhill
-/vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery
+/vllm/v1/executor @njhill
+/vllm/v1/worker @njhill
+/vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery @NickLucche
 
 # Model runner V2
-/vllm/v1/worker/gpu @WoosukKwon
+/vllm/v1/worker/gpu @WoosukKwon @njhill
+/vllm/v1/worker/gpu/kv_connector.py @orozery
 
 # Test ownership
 /.buildkite/lm-eval-harness @mgoin 

.github/mergify.yml

@@ -3,6 +3,7 @@ pull_request_rules:
   description: Automatically apply documentation label
   conditions:
     - label != stale
+    - -closed
     - or:
       - files~=^[^/]+\.md$
       - files~=^docs/
@@ -259,8 +260,7 @@ pull_request_rules:
       - files=benchmarks/run_structured_output_benchmark.sh
       - files=docs/features/structured_outputs.md
       - files=examples/offline_inference/structured_outputs.py
-      - files=examples/online_serving/openai_chat_completion_structured_outputs.py
+      - files=examples/online_serving/structured_outputs/structured_outputs.py
-      - files=examples/online_serving/openai_chat_completion_structured_outputs_with_reasoning.py
       - files~=^tests/v1/structured_output/
       - files=tests/v1/entrypoints/llm/test_struct_output_generate.py
       - files~=^vllm/v1/structured_output/

.github/workflows/bc-lint.yml

@@ -1,29 +0,0 @@
-name: BC Lint
-
-on:
-  pull_request:
-    types:
-      - opened
-      - synchronize
-      - reopened
-      - labeled
-      - unlabeled
-
-jobs:
-  bc_lint:
-    if: github.repository_owner == 'vllm-project'
-    runs-on: ubuntu-latest
-    steps:
-      - name: Run BC Lint Action
-        uses: pytorch/test-infra/.github/actions/bc-lint@main
-        with:
-          repo: ${{ github.event.pull_request.head.repo.full_name }}
-          base_sha: ${{ github.event.pull_request.base.sha }}
-          head_sha: ${{ github.event.pull_request.head.sha }}
-          suppression: ${{ contains(github.event.pull_request.labels.*.name, 'suppress-bc-linter') }}
-          docs_link: 'https://github.com/pytorch/test-infra/wiki/BC-Linter'
-          config_dir: .github
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}
-  cancel-in-progress: true

.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

.gitignore

@@ -3,6 +3,8 @@
 
 # vllm-flash-attn built from source
 vllm/vllm_flash_attn/*
+!vllm/vllm_flash_attn/__init__.py
+!vllm/vllm_flash_attn/flash_attn_interface.py
 
 # OpenAI triton kernels copied from source
 vllm/third_party/triton_kernels/*

.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

@@ -725,7 +725,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # CUTLASS MoE kernels
 
   # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and ONLY works
-  # on Hopper). get_cutlass_(pplx_)moe_mm_data should only be compiled
+  # on Hopper). get_cutlass_(batched_)moe_mm_data should only be compiled
   # if it's possible to compile MoE kernels that use its output.
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
@@ -771,6 +771,51 @@ 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" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+  endif()
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND DSV3_FUSED_A_GEMM_ARCHS)
+    set(DSV3_FUSED_A_GEMM_SRC "csrc/dsv3_fused_a_gemm.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${DSV3_FUSED_A_GEMM_SRC}"
+      CUDA_ARCHS "${DSV3_FUSED_A_GEMM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC ${DSV3_FUSED_A_GEMM_SRC})
+    message(STATUS "Building dsv3_fused_a_gemm for archs: ${DSV3_FUSED_A_GEMM_ARCHS}")
+  else()
+    message(STATUS "Not building dsv3_fused_a_gemm as no compatible archs found "
+                   "in CUDA target architectures.")
+  endif()
+
   # moe_data.cu is used by all CUTLASS MoE kernels.
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
     cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
@@ -953,7 +998,8 @@ set(VLLM_MOE_EXT_SRC
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_MOE_EXT_SRC
     "csrc/moe/moe_wna16.cu"
-    "csrc/moe/grouped_topk_kernels.cu")
+    "csrc/moe/grouped_topk_kernels.cu"
+    "csrc/moe/router_gemm.cu")
 endif()
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
@@ -1082,6 +1128,27 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
                    " in CUDA target architectures")
   endif()
+
+  # DeepSeek V3 router GEMM kernel - requires SM90+
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+  endif()
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND DSV3_ROUTER_GEMM_ARCHS)
+    set(DSV3_ROUTER_GEMM_SRC
+      "csrc/moe/dsv3_router_gemm_entry.cu"
+      "csrc/moe/dsv3_router_gemm_float_out.cu"
+      "csrc/moe/dsv3_router_gemm_bf16_out.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${DSV3_ROUTER_GEMM_SRC}"
+      CUDA_ARCHS "${DSV3_ROUTER_GEMM_ARCHS}")
+    list(APPEND VLLM_MOE_EXT_SRC "${DSV3_ROUTER_GEMM_SRC}")
+    message(STATUS "Building DSV3 router GEMM kernel for archs: ${DSV3_ROUTER_GEMM_ARCHS}")
+  else()
+    message(STATUS "Not building DSV3 router GEMM kernel as no compatible archs found"
+                   " (requires SM90+ and CUDA >= 12.0)")
+  endif()
 endif()
 
 message(STATUS "Enabling moe extension.")

benchmarks/attention_benchmarks/__init__.py

@@ -15,7 +15,6 @@ from .common import (
     BenchmarkConfig,
     BenchmarkResult,
     MockLayer,
-    MockModelConfig,
     ResultsFormatter,
     get_attention_scale,
     is_mla_backend,
@@ -36,7 +35,6 @@ __all__ = [
     "ResultsFormatter",
     # Mock objects
     "MockLayer",
-    "MockModelConfig",
     # Utilities
     "setup_mla_dims",
     "get_attention_scale",

benchmarks/attention_benchmarks/common.py

@@ -10,7 +10,6 @@ from dataclasses import asdict, dataclass
 from pathlib import Path
 from typing import Any
 
-import numpy as np
 import torch
 from batch_spec import get_batch_type, parse_batch_spec
 from rich.console import Console
@@ -31,7 +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)
 
 
@@ -62,10 +61,7 @@ class MockHfConfig:
 # Import AttentionLayerBase at module level to avoid circular dependencies
 try:
     from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
-
-    _HAS_ATTENTION_LAYER_BASE = True
 except ImportError:
-    _HAS_ATTENTION_LAYER_BASE = False
     AttentionLayerBase = object  # Fallback
 
 
@@ -167,95 +163,6 @@ class MockLayer(AttentionLayerBase):
         return self._kv_cache_spec
 
 
-class MockModelConfig:
-    """Mock model configuration."""
-
-    def __init__(
-        self,
-        num_q_heads: int,
-        num_kv_heads: int,
-        head_dim: int,
-        dtype: torch.dtype = torch.float16,
-        max_model_len: int = 32768,
-    ):
-        self._n_q = num_q_heads
-        self._n_kv = num_kv_heads
-        self._d = head_dim
-        self.dtype = dtype
-        self.max_model_len = max_model_len
-
-    def get_num_attention_heads(self, _=None) -> int:
-        return self._n_q
-
-    def get_num_kv_heads(self, _=None) -> int:
-        return self._n_kv
-
-    def get_head_size(self) -> int:
-        return self._d
-
-    def get_num_layers(self) -> int:
-        """Mock method for layer count queries."""
-        return 1
-
-    def get_sliding_window_for_layer(self, _layer_idx: int):
-        """Mock method for sliding window queries."""
-        return None
-
-    def get_logits_soft_cap_for_layer(self, _layer_idx: int):
-        """Mock method for logits soft cap queries."""
-        return None
-
-    def get_sm_scale_for_layer(self, _layer_idx: int) -> float:
-        """Mock method for SM scale queries."""
-        return 1.0 / (self.get_head_size() ** 0.5)
-
-
-class MockParallelConfig:
-    """Mock parallel configuration."""
-
-    pass
-
-
-class MockCompilationConfig:
-    """Mock compilation configuration."""
-
-    def __init__(self):
-        self.full_cuda_graph = False
-        self.static_forward_context = {}
-
-
-class MockVLLMConfig:
-    """Mock VLLM configuration."""
-
-    def __init__(self):
-        self.compilation_config = MockCompilationConfig()
-
-
-class MockRunner:
-    """Mock GPU runner for metadata builders."""
-
-    def __init__(
-        self,
-        seq_lens: np.ndarray,
-        query_start_locs: np.ndarray,
-        device: torch.device,
-        num_q_heads: int,
-        num_kv_heads: int,
-        head_dim: int,
-        dtype: torch.dtype,
-    ):
-        self.model_config = MockModelConfig(num_q_heads, num_kv_heads, head_dim, dtype)
-        self.parallel_config = MockParallelConfig()
-        self.vllm_config = MockVLLMConfig()
-        self.seq_lens_np = seq_lens
-        self.query_start_loc_np = query_start_locs
-        self.device = device
-        self.attention_chunk_size = None
-        self.num_query_heads = num_q_heads
-        self.num_kv_heads = num_kv_heads
-        self.dtype = dtype
-
-
 @dataclass
 class ParameterSweep:
     """Configuration for sweeping a backend parameter."""

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/backend_request_func.py

@@ -649,9 +649,3 @@ ASYNC_REQUEST_FUNCS = {
     "sglang": async_request_openai_completions,
     "llama.cpp": async_request_openai_completions,
 }
-
-OPENAI_COMPATIBLE_BACKENDS = [
-    k
-    for k, v in ASYNC_REQUEST_FUNCS.items()
-    if v in (async_request_openai_completions, async_request_openai_chat_completions)
-]

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/benchmark_utils.py

@@ -1,78 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import argparse
-import json
-import math
-import os
 import time
 from types import TracebackType
-from typing import Any
-
-
-def convert_to_pytorch_benchmark_format(
-    args: argparse.Namespace, metrics: dict[str, list], extra_info: dict[str, Any]
-) -> list:
-    """
-    Save the benchmark results in the format used by PyTorch OSS benchmark with
-    on metric per record
-    https://github.com/pytorch/pytorch/wiki/How-to-integrate-with-PyTorch-OSS-benchmark-database
-    """
-    records = []
-    if not os.environ.get("SAVE_TO_PYTORCH_BENCHMARK_FORMAT", False):
-        return records
-
-    for name, benchmark_values in metrics.items():
-        record = {
-            "benchmark": {
-                "name": "vLLM benchmark",
-                "extra_info": {
-                    "args": vars(args),
-                },
-            },
-            "model": {
-                "name": args.model,
-            },
-            "metric": {
-                "name": name,
-                "benchmark_values": benchmark_values,
-                "extra_info": extra_info,
-            },
-        }
-
-        tp = record["benchmark"]["extra_info"]["args"].get("tensor_parallel_size")
-        # Save tensor_parallel_size parameter if it's part of the metadata
-        if not tp and "tensor_parallel_size" in extra_info:
-            record["benchmark"]["extra_info"]["args"]["tensor_parallel_size"] = (
-                extra_info["tensor_parallel_size"]
-            )
-
-        records.append(record)
-
-    return records
-
-
-class InfEncoder(json.JSONEncoder):
-    def clear_inf(self, o: Any):
-        if isinstance(o, dict):
-            return {k: self.clear_inf(v) for k, v in o.items()}
-        elif isinstance(o, list):
-            return [self.clear_inf(v) for v in o]
-        elif isinstance(o, float) and math.isinf(o):
-            return "inf"
-        return o
-
-    def iterencode(self, o: Any, *args, **kwargs) -> Any:
-        return super().iterencode(self.clear_inf(o), *args, **kwargs)
-
-
-def write_to_json(filename: str, records: list) -> None:
-    with open(filename, "w") as f:
-        json.dump(
-            records,
-            f,
-            cls=InfEncoder,
-            default=lambda o: f"<{type(o).__name__} object is not JSON serializable>",
-        )
 
 
 # Collect time and generate time metrics

benchmarks/cutlass_benchmarks/utils.py

@@ -2,7 +2,6 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 # Cutlass bench utils
-from collections.abc import Iterable
 
 import torch
 
@@ -86,15 +85,3 @@ def make_rand_sparse_tensors(
 
     # Compressed B, Metadata, Original A, B
     return b_compressed, e, a, b
-
-
-def make_n_rand_sparse_tensors(
-    num_tensors: int, dtype: torch.dtype, m: int, n: int, k: int
-) -> tuple[Iterable[torch.Tensor], Iterable[torch.Tensor]]:
-    ABs = []
-    for _ in range(num_tensors):
-        b_comp, e, a, b = make_rand_sparse_tensors(dtype, m, n, k)
-        if b_comp is not None:
-            ABs.append(make_rand_sparse_tensors(dtype, m, n, k))
-    BComps, Es, As, Bs = zip(*ABs)
-    return list(BComps), list(Es), list(As), list(Bs)

benchmarks/disagg_benchmarks/rate_limiter.py

@@ -1,45 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import asyncio
-import time
-
-
-class RateLimiter:
-    """Token bucket rate limiter implementation"""
-
-    def __init__(self, rate_limit):
-        self.rate_limit = rate_limit  # Requests per second
-        self.num_available_tokens = rate_limit  # Available tokens
-        self.last_refill = time.monotonic()  # Last token refill time
-        self.lock = asyncio.Lock()  # Synchronization lock
-
-    async def acquire(self):
-        """Acquire a token from the rate limiter"""
-        while True:
-            async with self.lock:
-                current_time = time.monotonic()
-                elapsed = current_time - self.last_refill
-
-                # Refill num_available_tokens if more than 1 second has passed
-                if elapsed > 1.0:
-                    self.num_available_tokens = self.rate_limit
-                    self.last_refill = current_time
-
-                # Check if num_available_tokens are available
-                if self.num_available_tokens > 0:
-                    self.num_available_tokens -= 1
-                    return True
-
-                # Calculate wait time if no num_available_tokens available
-                wait_time = 1.0 - elapsed
-            await asyncio.sleep(wait_time)
-
-    async def __aenter__(self):
-        """Enter async context manager - acquire token"""
-        await self.acquire()
-        return self
-
-    async def __aexit__(self, exc_type, exc_value, traceback):
-        """Exit async context manager - no cleanup needed"""
-        pass

benchmarks/disagg_benchmarks/request_queue.py

@@ -1,39 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import asyncio
-from collections import deque
-
-
-class RequestQueue:
-    """Request queue manager with concurrency control"""
-
-    def __init__(self, max_concurrent, max_queue_size):
-        # Maximum concurrent requests
-        self.max_concurrent = max_concurrent
-        self.max_queue_size = max_queue_size  # Maximum queue size
-        # Concurrency control
-        self.semaphore = asyncio.Semaphore(max_concurrent)
-        self.queue = deque()  # Request queue
-        self.queue_size = 0  # Current queue size
-        self.lock = asyncio.Lock()  # Sync queue Lock
-
-    async def enqueue(self, task):
-        """Add a request task to the queue"""
-        async with self.lock:
-            if self.queue_size >= self.max_queue_size:
-                return False
-
-            self.queue.append(task)
-            self.queue_size += 1
-            return True
-
-    async def process(self):
-        """Process queued requests using semaphore for concurrency control"""
-        while True:
-            if self.queue:
-                async with self.semaphore, self.lock:
-                    task = self.queue.popleft()
-                    self.queue_size -= 1
-                    await task
-            await asyncio.sleep(0.01)  # Yield control to event loop

benchmarks/fused_kernels/layernorm_rms_benchmarks.py

@@ -13,6 +13,7 @@ from torch.utils.benchmark import Measurement as TMeasurement
 from tqdm import tqdm
 
 import vllm._custom_ops as ops
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8,
@@ -291,6 +292,7 @@ def print_timers(timers: Iterable[TMeasurement]):
     compare.print()
 
 
+@default_vllm_config()
 def main():
     torch.set_default_device("cuda")
     bench_params = get_bench_params()

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_activation.py

@@ -7,6 +7,7 @@ import itertools
 import torch
 
 import vllm.model_executor.layers.activation  # noqa F401
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.custom_op import op_registry
 from vllm.triton_utils import triton
 from vllm.utils.argparse_utils import FlexibleArgumentParser
@@ -18,6 +19,7 @@ intermediate_size = [3072, 9728, 12288]
 configs = list(itertools.product(batch_size_range, seq_len_range, intermediate_size))
 
 
+@default_vllm_config()
 def benchmark_activation(
     batch_size: int,
     seq_len: int,

benchmarks/kernels/benchmark_block_fp8_gemm.py

@@ -8,6 +8,7 @@ os.environ["VLLM_USE_DEEP_GEMM"] = "0"
 
 import torch
 
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     W8A8BlockFp8LinearOp,
 )
@@ -40,6 +41,7 @@ DEEPSEEK_V3_SHAPES = [
 ]
 
 
+@default_vllm_config()
 def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
     """Build runner function for w8a8 block fp8 matmul."""
     factor_for_scale = 1e-2

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(
+    moe_config = make_dummy_moe_config(
-        MoEPrepareAndFinalizeNoEP(),
+        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(
+            moe_config=moe_config,
-                num_experts=num_experts,
-                hidden_dim=k,
-                intermediate_size_per_partition=n,
-                in_dtype=a.dtype,
-            ),
             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(
+        moe_config = make_dummy_moe_config(
-            MoEPrepareAndFinalizeNoEP(),
+            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(
+        kernel = mk.FusedMoEKernel(
-            MoEPrepareAndFinalizeNoEP(),
+            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

@@ -30,6 +30,9 @@ import torch.distributed as dist
 from torch.distributed import ProcessGroup
 
 from vllm.distributed.device_communicators.custom_all_reduce import CustomAllreduce
+from vllm.distributed.device_communicators.flashinfer_all_reduce import (
+    FlashInferAllReduce,
+)
 from vllm.distributed.device_communicators.pynccl import (
     PyNcclCommunicator,
     register_nccl_symmetric_ops,
@@ -44,7 +47,7 @@ from vllm.utils.argparse_utils import FlexibleArgumentParser
 logger = init_logger(__name__)
 
 # Default sequence lengths to benchmark
-DEFAULT_SEQUENCE_LENGTHS = [128, 512, 1024, 2048, 4096, 8192]
+DEFAULT_SEQUENCE_LENGTHS = [16, 64, 128, 512, 1024, 2048, 4096, 8192]
 
 # Fixed hidden size and dtype for all benchmarks
 HIDDEN_SIZE = 8192
@@ -81,6 +84,7 @@ class CommunicatorBenchmark:
         self.symm_mem_comm = None
         self.symm_mem_comm_multimem = None
         self.symm_mem_comm_two_shot = None
+        self.fi_ar_comm = None
 
         self._init_communicators()
 
@@ -161,6 +165,22 @@ class CommunicatorBenchmark:
             )
             self.symm_mem_comm_two_shot = None
 
+        try:
+            self.fi_ar_comm = FlashInferAllReduce(
+                group=self.cpu_group,
+                device=self.device,
+            )
+            if not self.fi_ar_comm.disabled:
+                logger.info("Rank %s: FlashInferAllReduce initialized", self.rank)
+            else:
+                logger.info("Rank %s: FlashInferAllReduce disabled", self.rank)
+                self.fi_ar_comm = None
+        except Exception as e:
+            logger.warning(
+                "Rank %s: Failed to initialize FlashInferAllReduce: %s", self.rank, e
+            )
+            self.fi_ar_comm = None
+
     def benchmark_allreduce(
         self, sequence_length: int, num_warmup: int, num_trials: int
     ) -> dict[str, float]:
@@ -180,7 +200,8 @@ class CommunicatorBenchmark:
                     lambda t, c=comm: c.custom_all_reduce(t),
                     lambda t, c=comm: c.should_custom_ar(t),
                     comm.capture(),
-                    "1stage",  # env variable value
+                    {"VLLM_CUSTOM_ALLREDUCE_ALGO": "1stage"},
+                    None,  # no destroy function
                 )
             )
             # CustomAllreduce two-shot
@@ -190,7 +211,8 @@ class CommunicatorBenchmark:
                     lambda t, c=comm: c.custom_all_reduce(t),
                     lambda t, c=comm: c.should_custom_ar(t),
                     comm.capture(),
-                    "2stage",  # env variable value
+                    {"VLLM_CUSTOM_ALLREDUCE_ALGO": "2stage"},
+                    None,  # no destroy function
                 )
             )
 
@@ -202,7 +224,8 @@ class CommunicatorBenchmark:
                     lambda t, c=comm: c.all_reduce(t),
                     lambda t: True,  # Always available if initialized
                     nullcontext(),
-                    None,  # no env variable needed
+                    {},  # no env variable needed
+                    None,  # no destroy function
                 )
             )
             communicators.append(
@@ -211,7 +234,8 @@ class CommunicatorBenchmark:
                     lambda t: torch.ops.vllm.all_reduce_symmetric_with_copy(t),
                     lambda t: True,  # Always available if initialized
                     nullcontext(),
-                    None,  # no env variable needed
+                    {},  # no env variable needed
+                    None,  # no destroy function
                 )
             )
 
@@ -223,7 +247,8 @@ class CommunicatorBenchmark:
                     lambda t, c=comm: c.all_reduce(t),
                     lambda t, c=comm: c.should_use_symm_mem(t),
                     nullcontext(),
-                    None,  # no env variable needed
+                    {},  # no env variable needed
+                    None,  # no destroy function
                 )
             )
 
@@ -235,29 +260,67 @@ class CommunicatorBenchmark:
                     lambda t, c=comm: c.all_reduce(t),
                     lambda t, c=comm: c.should_use_symm_mem(t),
                     nullcontext(),
-                    None,  # no env variable needed
+                    {},  # no env variable needed
+                    None,  # no destroy function needed
+                )
+            )
+
+        if self.fi_ar_comm is not None:
+            comm = self.fi_ar_comm
+            communicators.append(
+                (
+                    "flashinfer_trtllm",
+                    lambda t, c=comm: c.all_reduce(t),
+                    lambda t, c=comm: c.should_use_fi_ar(t),
+                    nullcontext(),
+                    {"VLLM_FLASHINFER_ALLREDUCE_BACKEND": "trtllm"},
+                    lambda c=comm: c.destroy(),
+                )
+            )
+            communicators.append(
+                (
+                    "flashinfer_mnnvl",
+                    lambda t, c=comm: c.all_reduce(t),
+                    lambda t, c=comm: c.should_use_fi_ar(t),
+                    nullcontext(),
+                    {"VLLM_FLASHINFER_ALLREDUCE_BACKEND": "mnnvl"},
+                    lambda c=comm: c.destroy(),
                 )
             )
 
         # Benchmark each communicator
-        for name, allreduce_fn, should_use_fn, context, env_var in communicators:
+        for (
-            # Set environment variable if needed
+            name,
-            if env_var is not None:
+            allreduce_fn,
-                os.environ["VLLM_CUSTOM_ALLREDUCE_ALGO"] = env_var
+            should_use_fn,
-            else:
+            context,
-                # Clear the environment variable to avoid interference
+            env_dict,
-                os.environ.pop("VLLM_CUSTOM_ALLREDUCE_ALGO", None)
+            destroy_fn,
-
+        ) in communicators:
-            latency = self.benchmark_allreduce_single(
+            # Save original values and apply new environment variables
-                sequence_length,
+            saved_env = {key: os.environ.get(key) for key in env_dict}
-                allreduce_fn,
+            for key, value in env_dict.items():
-                should_use_fn,
+                os.environ[key] = value
-                context,
+            try:
-                num_warmup,
+                latency = self.benchmark_allreduce_single(
-                num_trials,
+                    sequence_length,
-            )
+                    allreduce_fn,
-            if latency is not None:
+                    should_use_fn,
-                results[name] = latency
+                    context,
+                    num_warmup,
+                    num_trials,
+                )
+                if latency is not None:
+                    results[name] = latency
+            finally:
+                if destroy_fn is not None:
+                    destroy_fn()
+                # Restore environment variables to their original state
+                for key, original_value in saved_env.items():
+                    if original_value is None:
+                        os.environ.pop(key, None)
+                    else:
+                        os.environ[key] = original_value
 
         return results
 
@@ -279,7 +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()
@@ -297,17 +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

@@ -5,8 +5,11 @@
 Benchmark for FlashInfer fused collective operations vs standard operations.
 
 This benchmark compares:
-1. FlashInfer's allreduce_fusion (fused allreduce + rmsnorm + optional quant)
+1. FlashInfer's allreduce_fusion with trtllm backend
-2. Standard tensor_model_parallel_all_reduce + separate rmsnorm/quant operations
+   (fused allreduce + rmsnorm + optional FP8/FP4 quant)
+2. FlashInfer's allreduce_fusion with mnnvl backend
+   (fused allreduce + rmsnorm only, no quantization support)
+3. Standard tensor_model_parallel_all_reduce + separate rmsnorm/quant operations
 
 Usage with torchrun:
     torchrun --nproc_per_node=2 benchmark_fused_collective.py
@@ -48,8 +51,12 @@ SCALED_FP4_QUANT_OP = torch.ops._C.scaled_fp4_quant
 logger = init_logger(__name__)
 
 # Try to import FlashInfer
+TorchDistBackend = None
 try:
     import flashinfer.comm as flashinfer_comm  # type: ignore
+    from flashinfer.comm.mnnvl import (  # type: ignore
+        TorchDistBackend,
+    )
 
     if not (
         hasattr(flashinfer_comm, "allreduce_fusion")
@@ -74,11 +81,15 @@ _FI_MAX_SIZES = {
     8: 64 * MiB,  # 64MB
 }
 
-# Global workspace tensor for FlashInfer
+# Global workspace tensors for FlashInfer (keyed by backend name)
-_FI_WORKSPACE = None
+_FI_WORKSPACES: dict = {}
+
+# Backends to benchmark
+FLASHINFER_BACKENDS = ["trtllm", "mnnvl"]
 
 
 def setup_flashinfer_workspace(
+    backend: str,
     world_size: int,
     rank: int,
     hidden_dim: int,
@@ -86,41 +97,54 @@ def setup_flashinfer_workspace(
     dtype: torch.dtype,
 ):
     """Setup FlashInfer workspace for fused allreduce operations."""
-    global _FI_WORKSPACE
+    global FI_WORKSPACES
 
     if flashinfer_comm is None:
-        return None, None
+        return None
 
     if world_size not in _FI_MAX_SIZES:
         logger.warning("FlashInfer not supported for world size %s", world_size)
-        return None, None
+        return None
 
     try:
+        kwargs = {}
+        if TorchDistBackend is not None:
+            kwargs["comm_backend"] = TorchDistBackend(group=dist.group.WORLD)
+
         workspace = flashinfer_comm.create_allreduce_fusion_workspace(
-            backend="trtllm",
+            backend=backend,
             world_size=world_size,
             rank=rank,
             max_token_num=max_token_num,
             hidden_dim=hidden_dim,
             dtype=dtype,
+            **kwargs,
         )
 
-        _FI_WORKSPACE = workspace
+        _FI_WORKSPACES[backend] = workspace
         return workspace
     except Exception as e:
-        logger.error("Failed to setup FlashInfer workspace: %s", e)
+        logger.error(
+            "Failed to setup FlashInfer workspace (backend=%s): %s", backend, e
+        )
         return None
 
 
-def cleanup_flashinfer_workspace(workspace):
+def cleanup_flashinfer_workspaces():
-    """Cleanup FlashInfer workspace."""
+    """Cleanup all FlashInfer workspaces."""
-    if flashinfer_comm is None or workspace is None:
+    if flashinfer_comm is None:
         return
 
-    try:
+    for backend, workspace in _FI_WORKSPACES.items():
-        workspace.destroy()
+        try:
-    except Exception as e:
+            workspace.destroy()
-        logger.error("Failed to cleanup FlashInfer workspace: %s", e)
+        except Exception as e:
+            logger.error(
+                "Failed to cleanup FlashInfer workspace (backend=%s): %s",
+                backend,
+                e,
+            )
+    _FI_WORKSPACES.clear()
 
 
 class FlashInferFusedAllReduceParams:
@@ -134,7 +158,7 @@ class FlashInferFusedAllReduceParams:
         self.fp32_acc = True
         self.max_token_num = max_token_num
 
-    def get_trtllm_fused_allreduce_kwargs(self):
+    def get_flashinfer_fused_allreduce_kwargs(self):
         return {
             "launch_with_pdl": self.launch_with_pdl,
             "fp32_acc": self.fp32_acc,
@@ -147,11 +171,12 @@ def flashinfer_fused_allreduce_rmsnorm(
     rms_gamma: torch.Tensor,
     rms_eps: float,
     allreduce_params: "FlashInferFusedAllReduceParams",
+    workspace: object,
     use_oneshot: bool,
     norm_out: torch.Tensor | None = None,
 ):
     """FlashInfer fused allreduce + rmsnorm operation."""
-    if flashinfer_comm is None or _FI_WORKSPACE is None:
+    if flashinfer_comm is None or workspace is None:
         raise RuntimeError("FlashInfer not available or workspace not initialized")
 
     if norm_out is None:
@@ -160,9 +185,13 @@ def flashinfer_fused_allreduce_rmsnorm(
     else:
         residual_out = input_tensor
 
+    layout_code = None
+    if workspace.backend == "trtllm":
+        layout_code = flashinfer_comm.QuantizationSFLayout.SWIZZLED_128x4
+
     flashinfer_comm.allreduce_fusion(
         input=input_tensor,
-        workspace=_FI_WORKSPACE,
+        workspace=workspace,
         pattern=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNorm,
         residual_in=residual,
         residual_out=residual_out,
@@ -171,10 +200,10 @@ def flashinfer_fused_allreduce_rmsnorm(
         rms_eps=rms_eps,
         quant_out=None,
         scale_out=None,
-        layout_code=flashinfer_comm.QuantizationSFLayout.SWIZZLED_128x4,
+        layout_code=layout_code,
         scale_factor=None,
         use_oneshot=use_oneshot,
-        **allreduce_params.get_trtllm_fused_allreduce_kwargs(),
+        **allreduce_params.get_flashinfer_fused_allreduce_kwargs(),
     )
 
 
@@ -185,12 +214,16 @@ def flashinfer_fused_allreduce_rmsnorm_fp8_quant(
     rms_eps: float,
     scale_factor: torch.Tensor,
     allreduce_params: FlashInferFusedAllReduceParams,
+    workspace: object,
     use_oneshot: bool = True,
     norm_out: torch.Tensor | None = None,
     quant_out: torch.Tensor | None = None,
 ):
-    """FlashInfer fused allreduce + rmsnorm + FP8 quantization."""
+    """FlashInfer fused allreduce + rmsnorm + FP8 quantization.
-    if flashinfer_comm is None or _FI_WORKSPACE is None:
+
+    Note: Only supported by the trtllm backend.
+    """
+    if flashinfer_comm is None or workspace is None:
         raise RuntimeError("FlashInfer not available or workspace not initialized")
 
     if norm_out is None:
@@ -201,7 +234,7 @@ def flashinfer_fused_allreduce_rmsnorm_fp8_quant(
 
     flashinfer_comm.allreduce_fusion(
         input=input_tensor,
-        workspace=_FI_WORKSPACE,
+        workspace=workspace,
         pattern=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNormFP8Quant,
         residual_in=residual,
         residual_out=residual_out,
@@ -213,7 +246,7 @@ def flashinfer_fused_allreduce_rmsnorm_fp8_quant(
         layout_code=flashinfer_comm.QuantizationSFLayout.SWIZZLED_128x4,
         scale_factor=scale_factor,
         use_oneshot=use_oneshot,
-        **allreduce_params.get_trtllm_fused_allreduce_kwargs(),
+        **allreduce_params.get_flashinfer_fused_allreduce_kwargs(),
     )
 
 
@@ -224,13 +257,17 @@ def flashinfer_fused_allreduce_rmsnorm_fp4_quant(
     rms_eps: float,
     input_global_scale: torch.Tensor,
     allreduce_params: FlashInferFusedAllReduceParams,
+    workspace: object,
     quant_out: torch.Tensor,
     use_oneshot: bool,
     output_scale: torch.Tensor,
     norm_out: torch.Tensor | None = None,
 ):
-    """FlashInfer fused allreduce + rmsnorm + FP4 quantization."""
+    """FlashInfer fused allreduce + rmsnorm + FP4 quantization.
-    if flashinfer_comm is None or _FI_WORKSPACE is None:
+
+    Note: Only supported by the trtllm backend.
+    """
+    if flashinfer_comm is None or workspace is None:
         raise RuntimeError("FlashInfer not available or workspace not initialized")
 
     if norm_out is None:
@@ -241,7 +278,7 @@ def flashinfer_fused_allreduce_rmsnorm_fp4_quant(
 
     flashinfer_comm.allreduce_fusion(
         input=input_tensor,
-        workspace=_FI_WORKSPACE,
+        workspace=workspace,
         pattern=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNormFP4Quant,
         residual_in=residual,
         residual_out=residual_out,
@@ -253,7 +290,7 @@ def flashinfer_fused_allreduce_rmsnorm_fp4_quant(
         layout_code=flashinfer_comm.QuantizationSFLayout.SWIZZLED_128x4,
         scale_factor=input_global_scale,
         use_oneshot=use_oneshot,
-        **allreduce_params.get_trtllm_fused_allreduce_kwargs(),
+        **allreduce_params.get_flashinfer_fused_allreduce_kwargs(),
     )
 
 
@@ -348,7 +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()
@@ -361,19 +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
@@ -386,13 +423,16 @@ def run_benchmarks(
     dtype: torch.dtype,
     use_residual: bool,
     allreduce_params: FlashInferFusedAllReduceParams | None,
+    workspaces: dict,
     quant_modes: set[str],
     no_oneshot: bool,
 ):
     """Run all benchmarks for given configuration.
 
     Args:
-        quant_mode: "none", "fp8_only", "fp4_only", or "all"
+        allreduce_params: Shared parameters for FlashInfer fused allreduce.
+        workspaces: Dict mapping backend name ("trtllm", "mnnvl") to workspace.
+        quant_modes: Set of quantization modes: "none", "fp8", "fp4".
     """
     (
         input_tensor,
@@ -408,18 +448,18 @@ def run_benchmarks(
 
     rms_eps = 1e-6
     results = {}
-    vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
     use_oneshot_options = [False] if no_oneshot else [True, False]
 
-    # Create RMSNorm and QuantFP8 layers once for native benchmarks
-
     if "none" in quant_modes:
         # Standard AllReduce + RMSNorm
+        # Re-create VllmFusedAllreduce per config so CustomOp binds the
+        # correct forward method (native vs custom kernel).
         for custom_op in ["-rms_norm", "+rms_norm"]:
             with set_current_vllm_config(
                 VllmConfig(compilation_config=CompilationConfig(custom_ops=[custom_op]))
             ):
                 try:
+                    vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                     suffix = (
                         "_custom_rms_norm" if "+" in custom_op else "_native_rms_norm"
                     )
@@ -438,6 +478,7 @@ def run_benchmarks(
             VllmConfig(compilation_config=CompilationConfig(custom_ops=["-rms_norm"]))
         ):
             try:
+                vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                 standard_allreduce_rmsnorm_native_compiled = torch.compile(
                     vllm_fused_allreduce.allreduce_rmsnorm,
                     fullgraph=True,
@@ -453,10 +494,11 @@ def run_benchmarks(
                 logger.error("Standard AllReduce+RMSNorm Native Compiled failed: %s", e)
                 results["standard_allreduce_rmsnorm_native_compiled"] = float("inf")
 
-        # FlashInfer Fused AllReduce + RMSNorm Oneshot/Twoshot
+        # FlashInfer Fused AllReduce + RMSNorm (all backends)
-        if flashinfer_comm is not None and allreduce_params is not None:
+        for backend, workspace in workspaces.items():
             for use_oneshot in use_oneshot_options:
                 suffix = "_oneshot" if use_oneshot else "_twoshot"
+                key = f"flashinfer_{backend}_fused_allreduce_rmsnorm{suffix}"
                 try:
                     time_ms = benchmark_operation(
                         flashinfer_fused_allreduce_rmsnorm,
@@ -466,14 +508,17 @@ def run_benchmarks(
                         rms_gamma=rms_gamma,
                         rms_eps=rms_eps,
                         allreduce_params=allreduce_params,
+                        workspace=workspace,
                         use_oneshot=use_oneshot,
                     )
-                    results[f"flashinfer_fused_allreduce_rmsnorm{suffix}"] = time_ms
+                    results[key] = time_ms
                 except Exception as e:
-                    logger.error("FlashInfer Fused AllReduce+RMSNorm failed: %s", e)
+                    logger.error(
-                    results[f"flashinfer_fused_allreduce_rmsnorm{suffix}"] = float(
+                        "FlashInfer (%s) Fused AllReduce+RMSNorm failed: %s",
-                        "inf"
+                        backend,
+                        e,
                     )
+                    results[key] = float("inf")
 
     if "fp8" in quant_modes:
         # Standard AllReduce + RMSNorm + FP8 Quant
@@ -482,7 +527,7 @@ def run_benchmarks(
                 "_custom_rms_norm" if "+" in rms_norm_custom_op else "_native_rms_norm"
             )
             for quant_fp8_custom_op in ["-quant_fp8", "+quant_fp8"]:
-                suffix += (
+                op_suffix = suffix + (
                     "_custom_quant_fp8"
                     if "+" in quant_fp8_custom_op
                     else "_native_quant_fp8"
@@ -495,16 +540,17 @@ def run_benchmarks(
                     )
                 ):
                     try:
+                        vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                         time_ms = benchmark_operation(
                             vllm_fused_allreduce.allreduce_rmsnorm_fp8_quant,
                             input_tensor,
                             residual=residual,
                             scale_factor=scale_fp8,
                         )
-                        results[f"standard_allreduce{suffix}"] = time_ms
+                        results[f"standard_allreduce{op_suffix}"] = time_ms
                     except Exception as e:
                         logger.error("Standard AllReduce+RMSNorm+FP8 failed: %s", e)
-                        results[f"standard_allreduce{suffix}"] = float("inf")
+                        results[f"standard_allreduce{op_suffix}"] = float("inf")
 
         # Standard AllReduce + RMSNorm + FP8 Quant Native Compiled
         with set_current_vllm_config(
@@ -515,6 +561,7 @@ def run_benchmarks(
             )
         ):
             try:
+                vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                 standard_allreduce_rmsnorm_fp8_quant_native_compiled = torch.compile(
                     vllm_fused_allreduce.allreduce_rmsnorm_fp8_quant,
                     fullgraph=True,
@@ -537,10 +584,12 @@ def run_benchmarks(
                     "inf"
                 )
 
-        # FlashInfer Fused AllReduce + RMSNorm + FP8 Quant Oneshot
+        # FlashInfer Fused AllReduce + RMSNorm + FP8 Quant (trtllm only)
-        if flashinfer_comm is not None and allreduce_params is not None:
+        if "trtllm" in workspaces:
+            trtllm_ws = workspaces["trtllm"]
             for use_oneshot in use_oneshot_options:
                 suffix = "_oneshot" if use_oneshot else "_twoshot"
+                key = f"flashinfer_trtllm_fused_allreduce_rmsnorm_fp8_quant{suffix}"
                 try:
                     time_ms = benchmark_operation(
                         flashinfer_fused_allreduce_rmsnorm_fp8_quant,
@@ -552,19 +601,16 @@ def run_benchmarks(
                         scale_factor=scale_fp8,
                         quant_out=quant_out_fp8,
                         allreduce_params=allreduce_params,
+                        workspace=trtllm_ws,
                         use_oneshot=use_oneshot,
                     )
-                    results[f"flashinfer_fused_allreduce_rmsnorm_fp8_quant{suffix}"] = (
+                    results[key] = time_ms
-                        time_ms
-                    )
                 except Exception as e:
                     logger.error(
-                        "FlashInfer Fused AllReduce+RMSNorm+FP8 Oneshot failed: %s",
+                        "FlashInfer (trtllm) Fused AllReduce+RMSNorm+FP8 failed: %s",
                         e,
                     )
-                    results[f"flashinfer_fused_allreduce_rmsnorm_fp8_quant{suffix}"] = (
+                    results[key] = float("inf")
-                        float("inf")
-                    )
 
     if "fp4" in quant_modes and current_platform.has_device_capability(100):
         # Standard AllReduce + RMSNorm + FP4 Quant
@@ -580,6 +626,7 @@ def run_benchmarks(
                 )
             ):
                 try:
+                    vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                     time_ms = benchmark_operation(
                         vllm_fused_allreduce.allreduce_rmsnorm_fp4_quant,
                         input_tensor,
@@ -598,6 +645,7 @@ def run_benchmarks(
             VllmConfig(compilation_config=CompilationConfig(custom_ops=["-rms_norm"]))
         ):
             try:
+                vllm_fused_allreduce = VllmFusedAllreduce(hidden_dim, dtype)
                 standard_allreduce_rmsnorm_fp4_quant_native_compiled = torch.compile(
                     vllm_fused_allreduce.allreduce_rmsnorm_fp4_quant,
                     fullgraph=True,
@@ -622,10 +670,12 @@ def run_benchmarks(
                     "inf"
                 )
 
-        # FlashInfer Fused AllReduce + RMSNorm + FP4 Quant Oneshot
+        # FlashInfer Fused AllReduce + RMSNorm + FP4 Quant (trtllm only)
-        if flashinfer_comm is not None and allreduce_params is not None:
+        if "trtllm" in workspaces:
+            trtllm_ws = workspaces["trtllm"]
             for use_oneshot in use_oneshot_options:
                 suffix = "_oneshot" if use_oneshot else "_twoshot"
+                key = f"flashinfer_trtllm_fused_allreduce_rmsnorm_fp4_quant{suffix}"
                 try:
                     time_ms = benchmark_operation(
                         flashinfer_fused_allreduce_rmsnorm_fp4_quant,
@@ -636,49 +686,18 @@ def run_benchmarks(
                         rms_eps=rms_eps,
                         input_global_scale=scale_fp4,
                         allreduce_params=allreduce_params,
+                        workspace=trtllm_ws,
                         quant_out=fp4_quant_out,
                         output_scale=fp4_output_scale,
                         use_oneshot=use_oneshot,
                     )
-                    results[f"flashinfer_fused_allreduce_rmsnorm_fp4_quant{suffix}"] = (
+                    results[key] = time_ms
-                        time_ms
-                    )
                 except Exception as e:
                     logger.error(
-                        "FlashInfer Fused AllReduce+RMSNorm+FP4 Oneshot failed: %s",
+                        "FlashInfer (trtllm) Fused AllReduce+RMSNorm+FP4 failed: %s",
                         e,
                     )
-                    results[f"flashinfer_fused_allreduce_rmsnorm_fp4_quant{suffix}"] = (
+                    results[key] = float("inf")
-                        float("inf")
-                    )
-
-        # FlashInfer Fused AllReduce + RMSNorm + FP4 Quant Two-shot
-        if flashinfer_comm is not None and allreduce_params is not None:
-            try:
-                time_ms = benchmark_operation(
-                    flashinfer_fused_allreduce_rmsnorm_fp4_quant,
-                    input_tensor,
-                    residual=residual,
-                    norm_out=norm_out,
-                    rms_gamma=rms_gamma,
-                    rms_eps=rms_eps,
-                    input_global_scale=scale_fp4,
-                    allreduce_params=allreduce_params,
-                    quant_out=fp4_quant_out,
-                    output_scale=fp4_output_scale,
-                    use_oneshot=False,
-                )
-                results["flashinfer_fused_allreduce_rmsnorm_fp4_quant_twoshot"] = (
-                    time_ms
-                )
-            except Exception as e:
-                logger.error(
-                    "FlashInfer Fused AllReduce+RMSNorm+FP4 Two-shot failed: %s",
-                    e,
-                )
-                results["flashinfer_fused_allreduce_rmsnorm_fp4_quant_twoshot"] = float(
-                    "inf"
-                )
 
     return results
 
@@ -1016,8 +1035,7 @@ def main():
 
     configs = list(itertools.product(args.num_tokens, dtypes, residual_options))
 
-    # Setup FlashInfer workspace if available
+    # Setup FlashInfer workspaces for all backends
-    workspace = None
     allreduce_params = None
 
     if flashinfer_comm is not None:
@@ -1032,15 +1050,17 @@ def main():
             args.hidden_dim * max_element_size
         )
 
-        workspace = setup_flashinfer_workspace(
+        for backend in FLASHINFER_BACKENDS:
-            world_size,
+            setup_flashinfer_workspace(
-            rank,
+                backend=backend,
-            args.hidden_dim,
+                world_size=world_size,
-            max_num_token,
+                rank=rank,
-            dtype=workspace_dtype,
+                hidden_dim=args.hidden_dim,
-        )
+                max_token_num=max_num_token,
+                dtype=workspace_dtype,
+            )
 
-        if workspace is not None:
+        if _FI_WORKSPACES:
             allreduce_params = FlashInferFusedAllReduceParams(
                 max_token_num=max_num_token,
             )
@@ -1066,6 +1086,7 @@ def main():
                 dtype,
                 use_residual,
                 allreduce_params,
+                workspaces=_FI_WORKSPACES,
                 quant_modes=quant_modes,
                 no_oneshot=args.no_oneshot,
             )
@@ -1104,11 +1125,13 @@ def main():
 
     finally:
         # Cleanup
-        if workspace is not None:
+        cleanup_flashinfer_workspaces()
-            cleanup_flashinfer_workspace(workspace)
 
         dist.barrier()
 
 
 if __name__ == "__main__":
-    main()
+    from vllm.config import VllmConfig, set_current_vllm_config
+
+    with set_current_vllm_config(VllmConfig()):
+        main()

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(
+        fn = mk.FusedMoEKernel(
-            MoEPrepareAndFinalizeNoEP(),
+            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(
+                moe_config=moe_config,
-                    num_experts=w2.shape[0],
-                    hidden_dim=w2.shape[1],
-                    intermediate_size_per_partition=w2.shape[2],
-                    in_dtype=a.dtype,
-                ),
                 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(
+        fn = mk.FusedMoEKernel(
-            MoEPrepareAndFinalizeNoEP(),
+            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(
+                moe_config=moe_config,
-                    num_experts=w2.shape[0],
-                    hidden_dim=w2.shape[1],
-                    intermediate_size_per_partition=w2.shape[2],
-                    in_dtype=a.dtype,
-                ),
                 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

@@ -5,12 +5,14 @@ import time
 
 import torch
 
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed
 
 
 @torch.inference_mode()
+@default_vllm_config()
 def main(
     num_tokens: int,
     hidden_size: int,
@@ -32,14 +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(
+            moe_config = (
-                prepare_finalize=MoEPrepareAndFinalizeNoEP(),
+                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(
+                    moe_config=moe_config,
-                        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",
-                    ),
                     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(
+                return deep_gemm_experts.apply(
-                    x, w1, w2, topk_weights, topk_ids, inplace=inplace
+                    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

@@ -0,0 +1,278 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Benchmark comparing old vs new default fused MoE configs.
+
+Runs the triton fused_moe kernel with three configurations for each scenario:
+  1. Tuned config (from JSON file, if available) — the target to match
+  2. Old default (the hardcoded defaults before this change)
+  3. New default (the improved defaults)
+
+Usage:
+    python benchmarks/kernels/benchmark_moe_defaults.py
+
+Produces a table showing kernel time (us) and speedup of new vs old defaults.
+"""
+
+import torch
+
+from vllm.model_executor.layers.fused_moe import fused_topk, override_config
+from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
+from vllm.model_executor.layers.fused_moe.fused_moe import (
+    fused_experts,
+    get_default_config,
+    get_moe_configs,
+)
+from vllm.platforms import current_platform
+from vllm.triton_utils import triton
+from vllm.utils.torch_utils import set_random_seed
+
+FP8_DTYPE = current_platform.fp8_dtype()
+
+
+def old_default_config(M, E, N, K, topk, dtype=None, block_shape=None):
+    """The original defaults before https://github.com/vllm-project/vllm/pull/34846,
+    for comparison."""
+    if dtype == "fp8_w8a8" and block_shape is not None:
+        return {
+            "BLOCK_SIZE_M": 64,
+            "BLOCK_SIZE_N": block_shape[0],
+            "BLOCK_SIZE_K": block_shape[1],
+            "GROUP_SIZE_M": 32,
+            "SPLIT_K": 1,
+            "num_warps": 4,
+            "num_stages": 3 if not current_platform.is_rocm() else 2,
+        }
+    elif M <= E:
+        return {
+            "BLOCK_SIZE_M": 16,
+            "BLOCK_SIZE_N": 32,
+            "BLOCK_SIZE_K": 64,
+            "GROUP_SIZE_M": 1,
+            "SPLIT_K": 1,
+        }
+    else:
+        return {
+            "BLOCK_SIZE_M": 64,
+            "BLOCK_SIZE_N": 64,
+            "BLOCK_SIZE_K": 32,
+            "GROUP_SIZE_M": 8,
+            "SPLIT_K": 1,
+        }
+
+
+def benchmark_config(
+    config,
+    M,
+    E,
+    N,
+    K,
+    topk,
+    dtype,
+    use_fp8=False,
+    block_shape=None,
+    num_iters=100,
+):
+    """Time a single kernel config. Returns kernel time in microseconds."""
+    init_dtype = torch.float16 if use_fp8 else dtype
+
+    a = torch.randn(M, K, device="cuda", dtype=init_dtype) / 10
+    w1 = torch.randn(E, 2 * N, K, device="cuda", dtype=init_dtype) / 10
+    w2 = torch.randn(E, K, N, device="cuda", dtype=init_dtype) / 10
+
+    w1_scale = None
+    w2_scale = None
+    a1_scale = None
+    a2_scale = None
+    if use_fp8:
+        if block_shape is not None:
+            bsn, bsk = block_shape
+            n_tiles_w1 = triton.cdiv(2 * N, bsn)
+            k_tiles_w1 = triton.cdiv(K, bsk)
+            n_tiles_w2 = triton.cdiv(K, bsn)
+            k_tiles_w2 = triton.cdiv(N, bsk)
+            w1_scale = torch.rand(
+                E, n_tiles_w1, k_tiles_w1, device="cuda", dtype=torch.float32
+            )
+            w2_scale = torch.rand(
+                E, n_tiles_w2, k_tiles_w2, device="cuda", dtype=torch.float32
+            )
+        else:
+            w1_scale = torch.rand(E, device="cuda", dtype=torch.float32)
+            w2_scale = torch.rand(E, device="cuda", dtype=torch.float32)
+        a1_scale = torch.rand(1, device="cuda", dtype=torch.float32)
+        a2_scale = torch.rand(1, device="cuda", dtype=torch.float32)
+        # Only weights are stored in fp8; activations stay in bf16/fp16
+        # and get dynamically quantized inside the kernel.
+        w1 = w1.to(FP8_DTYPE)
+        w2 = w2.to(FP8_DTYPE)
+
+    quant_config = FusedMoEQuantConfig.make(
+        quant_dtype=torch.float8_e4m3fn if use_fp8 else None,
+        w1_scale=w1_scale,
+        w2_scale=w2_scale,
+        a1_scale=a1_scale,
+        a2_scale=a2_scale,
+        block_shape=block_shape,
+    )
+
+    gating = torch.randn(M, E, device="cuda", dtype=torch.float32)
+
+    # Warmup
+    for _ in range(20):
+        with override_config(config):
+            topk_weights, topk_ids, _ = fused_topk(a, gating, topk, renormalize=True)
+            fused_experts(
+                a,
+                w1,
+                w2,
+                topk_weights,
+                topk_ids,
+                quant_config=quant_config,
+            )
+    torch.accelerator.synchronize()
+
+    # Benchmark
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+    start.record()
+    for _ in range(num_iters):
+        with override_config(config):
+            topk_weights, topk_ids, _ = fused_topk(a, gating, topk, renormalize=True)
+            fused_experts(
+                a,
+                w1,
+                w2,
+                topk_weights,
+                topk_ids,
+                quant_config=quant_config,
+            )
+    end.record()
+    torch.accelerator.synchronize()
+    return start.elapsed_time(end) / num_iters * 1000  # ms -> us
+
+
+# Model configurations: (name, E, N, K, topk, dtype_str, use_fp8, block_shape)
+# N = moe_intermediate_size // tp_size (the value used in config file lookup)
+MODELS = [
+    # --- Few experts ---
+    ("Mixtral bf16", 8, 7168, 4096, 2, None, False, None),
+    ("Mixtral fp8", 8, 7168, 4096, 2, "fp8_w8a8", True, None),
+    # --- Many experts: real model shapes at tp=1 ---
+    # Qwen2-MoE-57B: E=60, topk=4, N=1408, K=2048
+    ("Qwen2-MoE bf16", 60, 1408, 2048, 4, None, False, None),
+    # DeepSeek-V2: E=64, topk=6, N=1407, K=4096
+    # (use 1408 to avoid odd alignment; real model is 1407)
+    ("DeepSeek-V2 bf16", 64, 1408, 4096, 6, None, False, None),
+    # OLMoE-7B: E=64, topk=8, N=2048, K=2048
+    ("OLMoE bf16", 64, 2048, 2048, 8, None, False, None),
+    # GLM-4-100B-A10B: E=128, topk=8, N=1408, K=4096
+    ("GLM-4-MoE bf16", 128, 1408, 4096, 8, None, False, None),
+    # Qwen3-30B-A3B: E=128, topk=8, N=768, K=2048
+    ("Qwen3-MoE bf16", 128, 768, 2048, 8, None, False, None),
+    # DeepSeek-V3 / MiMo-V2-Flash: E=256, topk=8, N=2048, K=7168
+    ("DeepSeek-V3 bf16", 256, 2048, 7168, 8, None, False, None),
+    # Qwen3.5-70B-A22B (Qwen3-Next): E=512, topk=10, N=512, K=2048
+    ("Qwen3-Next bf16", 512, 512, 2048, 10, None, False, None),
+    # E=128 N=1856 bf16
+    ("E128 N1856 bf16", 128, 1856, 4096, 8, None, False, None),
+    # E=256 N=512 bf16 (DS-V3 tp=4)
+    ("DS-V3 tp4 bf16", 256, 512, 7168, 8, None, False, None),
+    # E=512 N=512 bf16 (Qwen3-Next tp=1)
+    ("Qwen3-Next bf16", 512, 512, 2048, 10, None, False, None),
+    # E=512 N=256 bf16 (Qwen3-Next tp=2)
+    ("Qwen3-Next tp2", 512, 256, 2048, 10, None, False, None),
+    # --- FP8 block quant (many experts) ---
+    # DS-V3 tp=4: E=256, N=512, fp8 block
+    ("DS-V3 tp4 fp8blk", 256, 512, 7168, 8, "fp8_w8a8", True, [128, 128]),
+    # DS-V3 tp=8: E=256, N=256, fp8 block
+    ("DS-V3 tp8 fp8blk", 256, 256, 7168, 8, "fp8_w8a8", True, [128, 128]),
+    # Qwen3-Next tp=2 fp8 block
+    ("Qwen3-Next tp2 fp8blk", 512, 256, 2048, 10, "fp8_w8a8", True, [128, 128]),
+]
+
+BATCH_SIZES = [1, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096]
+
+
+def main():
+    set_random_seed(0)
+    torch.set_default_device("cuda")
+    dtype = torch.bfloat16
+
+    for name, E, N, K, topk, dtype_str, use_fp8, block_shape in MODELS:
+        print(f"\n{'=' * 90}")
+        print(f"  {name}  (E={E}, N={N}, K={K}, topk={topk})")
+        print(f"{'=' * 90}")
+
+        # Try to load tuned config
+        block_n = block_shape[0] if block_shape else None
+        block_k = block_shape[1] if block_shape else None
+        tuned = get_moe_configs(E, N, dtype_str, block_n, block_k)
+        has_tuned = tuned is not None
+        print(f"  Tuned config available: {has_tuned}")
+
+        hdr = (
+            f"{'Batch':>6} | {'Tuned (us)':>11} | {'Old (us)':>11} | "
+            f"{'New (us)':>11} | {'New/Old':>8} | {'New/Tuned':>10}"
+        )
+        print(f"  {hdr}")
+        print(f"  {'-' * len(hdr)}")
+
+        for M in BATCH_SIZES:
+            old_cfg = old_default_config(M, E, N, K, topk, dtype_str, block_shape)
+            new_cfg = get_default_config(M, E, N, K, topk, dtype_str, block_shape)
+
+            if has_tuned:
+                tuned_cfg = tuned[min(tuned.keys(), key=lambda x: abs(x - M))]
+                t_tuned = benchmark_config(
+                    tuned_cfg,
+                    M,
+                    E,
+                    N,
+                    K,
+                    topk,
+                    dtype,
+                    use_fp8=use_fp8,
+                    block_shape=block_shape,
+                )
+            else:
+                t_tuned = None
+
+            t_old = benchmark_config(
+                old_cfg,
+                M,
+                E,
+                N,
+                K,
+                topk,
+                dtype,
+                use_fp8=use_fp8,
+                block_shape=block_shape,
+            )
+            t_new = benchmark_config(
+                new_cfg,
+                M,
+                E,
+                N,
+                K,
+                topk,
+                dtype,
+                use_fp8=use_fp8,
+                block_shape=block_shape,
+            )
+
+            ratio_new_old = t_new / t_old
+            tuned_str = f"{t_tuned:11.2f}" if t_tuned else f"{'N/A':>11}"
+            ratio_tuned = f"{t_new / t_tuned:10.2f}x" if t_tuned else f"{'N/A':>10}"
+            # flag regressions where new default is >5% slower than old
+            marker = " <--" if ratio_new_old > 1.05 else ""
+
+            print(
+                f"  {M:>6} | {tuned_str} | {t_old:11.2f} | {t_new:11.2f} "
+                f"| {ratio_new_old:7.2f}x | {ratio_tuned}{marker}"
+            )
+
+
+if __name__ == "__main__":
+    main()

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

@@ -36,6 +36,7 @@ from typing import Any
 import numpy as np
 import torch
 
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.transformers_utils.config import get_config
 from vllm.utils.argparse_utils import FlexibleArgumentParser
@@ -78,6 +79,7 @@ def calculate_stats(times: list[float]) -> dict[str, float]:
     }
 
 
+@default_vllm_config()
 def benchmark_mrope(
     model_name: str,
     num_tokens: int,
@@ -133,14 +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(
@@ -149,7 +151,7 @@ def benchmark_mrope(
             key_clone,
         )
 
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
         torch_times.append(time.time() - start_time)
 
     # Time triton kernel implementation
@@ -157,14 +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_per_token_quant_fp8.py

@@ -7,6 +7,7 @@ from unittest.mock import patch
 import pandas as pd
 import torch
 
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
 from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
 from vllm.triton_utils import triton
@@ -84,6 +85,7 @@ def calculate_diff(
 configs = []
 
 
+@default_vllm_config()
 def benchmark_quantization(
     batch_size,
     hidden_size,

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_rope.py

@@ -5,6 +5,7 @@ import itertools
 
 import torch
 
+from vllm.benchmarks.lib.utils import default_vllm_config
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.triton_utils import triton
 from vllm.utils.argparse_utils import FlexibleArgumentParser
@@ -29,6 +30,7 @@ def get_benchmark(head_size, rotary_dim, is_neox_style, device):
             args={},
         )
     )
+    @default_vllm_config()
     def benchmark(batch_size, seq_len, num_heads, provider):
         dtype = torch.bfloat16
         max_position = 8192

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

@@ -13,28 +13,16 @@ endif()
 #
 # Define environment variables for special configurations
 #
-set(ENABLE_AVX2 $ENV{VLLM_CPU_AVX2})
+set(ENABLE_X86_ISA $ENV{VLLM_CPU_X86})
-set(ENABLE_AVX512 $ENV{VLLM_CPU_AVX512})
-set(ENABLE_AVX512BF16 $ENV{VLLM_CPU_AVX512BF16})
-set(ENABLE_AVX512VNNI $ENV{VLLM_CPU_AVX512VNNI})
-set(ENABLE_AMXBF16 $ENV{VLLM_CPU_AMXBF16})
 set(ENABLE_ARM_BF16 $ENV{VLLM_CPU_ARM_BF16})
 
 include_directories("${CMAKE_SOURCE_DIR}/csrc")
 
-
 set (ENABLE_NUMA TRUE)
 
 #
 # Check the compile flags
 #
-
-if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64")
-    list(APPEND CXX_COMPILE_FLAGS
-        "-mf16c"
-    )
-endif()
-
 if(MACOSX_FOUND)
     list(APPEND CXX_COMPILE_FLAGS
         "-DVLLM_CPU_EXTENSION")
@@ -78,18 +66,6 @@ function(check_sysctl TARGET OUT)
     endif()
 endfunction()
 
-
-function (is_avx512_disabled OUT)
-    set(DISABLE_AVX512 $ENV{VLLM_CPU_DISABLE_AVX512})
-    if(DISABLE_AVX512 AND DISABLE_AVX512 STREQUAL "true")
-        set(${OUT} ON PARENT_SCOPE)
-    else()
-        set(${OUT} OFF PARENT_SCOPE)
-    endif()
-endfunction()
-
-is_avx512_disabled(AVX512_DISABLED)
-
 if (MACOSX_FOUND AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
     message(STATUS "Apple Silicon Detected")
     set(APPLE_SILICON_FOUND TRUE)
@@ -97,8 +73,6 @@ if (MACOSX_FOUND AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
     check_sysctl(hw.optional.neon ASIMD_FOUND)
     check_sysctl(hw.optional.arm.FEAT_BF16 ARM_BF16_FOUND)
 else()
-    find_isa(${CPUINFO} "avx2" AVX2_FOUND)
-    find_isa(${CPUINFO} "avx512f" AVX512_FOUND)
     find_isa(${CPUINFO} "Power11" POWER11_FOUND)
     find_isa(${CPUINFO} "POWER10" POWER10_FOUND)
     find_isa(${CPUINFO} "POWER9" POWER9_FOUND)
@@ -108,77 +82,32 @@ else()
     find_isa(${CPUINFO} "v" RVV_FOUND) # Check for RISC-V RVV support
 
     # Support cross-compilation by allowing override via environment variables
-    if (ENABLE_AVX2)
-        set(AVX2_FOUND ON)
-        message(STATUS "AVX2 support enabled via VLLM_CPU_AVX2 environment variable")
-    endif()
-    if (ENABLE_AVX512)
-        set(AVX512_FOUND ON)
-        message(STATUS "AVX512 support enabled via VLLM_CPU_AVX512 environment variable")
-    endif()
     if (ENABLE_ARM_BF16)
         set(ARM_BF16_FOUND ON)
         message(STATUS "ARM BF16 support enabled via VLLM_CPU_ARM_BF16 environment variable")
     endif()
 endif()
 
-if (AVX512_FOUND AND NOT AVX512_DISABLED)
+if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64" OR ENABLE_X86_ISA)
-    list(APPEND CXX_COMPILE_FLAGS
+    set(ENABLE_X86_ISA ON)
+    if (NOT (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
+            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3))
+        message(FATAL_ERROR "X86 backend requires gcc/g++ >= 12.3")
+    endif()
+    list(APPEND CXX_COMPILE_FLAGS "-mf16c")
+    list(APPEND CXX_COMPILE_FLAGS_AVX512 ${CXX_COMPILE_FLAGS})
+    list(APPEND CXX_COMPILE_FLAGS_AVX2 ${CXX_COMPILE_FLAGS})
+    list(APPEND CXX_COMPILE_FLAGS_AVX512
         "-mavx512f"
         "-mavx512vl"
         "-mavx512bw"
-        "-mavx512dq")
+        "-mavx512dq"
-
+        "-mavx512bf16"
-    find_isa(${CPUINFO} "avx512_bf16" AVX512BF16_FOUND)
+        "-mavx512vnni"
-    if (AVX512BF16_FOUND OR ENABLE_AVX512BF16)
+        "-mamx-bf16"
-        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
+        "-mamx-tile")
-            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
+    list(APPEND CXX_COMPILE_FLAGS_AVX2
-            list(APPEND CXX_COMPILE_FLAGS "-mavx512bf16")
+        "-mavx2")
-            set(ENABLE_AVX512BF16 ON)
-        else()
-            set(ENABLE_AVX512BF16 OFF)
-            message(WARNING "Disable AVX512-BF16 ISA support, requires gcc/g++ >= 12.3")
-        endif()
-    else()
-        set(ENABLE_AVX512BF16 OFF)
-        message(WARNING "Disable AVX512-BF16 ISA support, no avx512_bf16 found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512BF16=1.")
-    endif()
-
-    find_isa(${CPUINFO} "avx512_vnni" AVX512VNNI_FOUND)
-    if (AVX512VNNI_FOUND OR ENABLE_AVX512VNNI)
-        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
-            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
-            list(APPEND CXX_COMPILE_FLAGS "-mavx512vnni")
-            set(ENABLE_AVX512VNNI ON)
-        else()
-            set(ENABLE_AVX512VNNI OFF)
-            message(WARNING "Disable AVX512-VNNI ISA support, requires gcc/g++ >= 12.3")
-        endif()
-    else()
-        set(ENABLE_AVX512VNNI OFF)
-        message(WARNING "Disable AVX512-VNNI ISA support, no avx512_vnni found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512VNNI=1.")
-    endif()
-
-    find_isa(${CPUINFO} "amx_bf16" AMXBF16_FOUND)
-    if (AMXBF16_FOUND OR ENABLE_AMXBF16)
-        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
-            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
-            list(APPEND CXX_COMPILE_FLAGS "-mamx-bf16" "-mamx-tile")
-            set(ENABLE_AMXBF16 ON)
-            add_compile_definitions(-DCPU_CAPABILITY_AMXBF16)
-        else()
-            set(ENABLE_AMXBF16 OFF)
-            message(WARNING "Disable AMX_BF16 ISA support, requires gcc/g++ >= 12.3")
-        endif()
-    else()
-        set(ENABLE_AMXBF16 OFF)
-        message(WARNING "Disable AMX_BF16 ISA support, no amx_bf16 found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AMXBF16=1.")
-    endif()
-    
-elseif (AVX2_FOUND)
-    list(APPEND CXX_COMPILE_FLAGS "-mavx2")
-    message(WARNING "vLLM CPU backend using AVX2 ISA")
-    
 elseif (POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
     message(STATUS "PowerPC detected")
     if (POWER9_FOUND)
@@ -219,12 +148,12 @@ elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
         list(APPEND CXX_COMPILE_FLAGS "-march=rv64gc")
     endif()
 else()
-    message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA, S390X ISA, ARMv8 or RISC-V support.")
+    message(FATAL_ERROR "vLLM CPU backend requires X86, Power9+ ISA, S390X ISA, ARMv8 or RISC-V support.")
 endif()
 
 
-# Build oneDNN for GEMM kernels (only for x86-AVX512 /ARM platforms)
+# Build oneDNN for GEMM kernels
-if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
+if (ENABLE_X86_ISA OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
     # Fetch and build Arm Compute Library (ACL) as oneDNN's backend for AArch64
     # TODO [fadara01]: remove this once ACL can be fetched and built automatically as a dependency of oneDNN
     set(ONEDNN_AARCH64_USE_ACL OFF CACHE BOOL "")
@@ -313,13 +242,24 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
         )
     else()
         message(STATUS "Downloading oneDNN from GitHub")
-        FetchContent_Declare(
+        if(ASIMD_FOUND AND NOT APPLE_SILICON_FOUND)
-            oneDNN
+            message(STATUS "aarch64 detected: using pinned oneDNN commit 9c5be1cc59e368aebf0909e6cf20f981ea61462a")
-            GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
+            FetchContent_Declare(
-            GIT_TAG v3.10
+                oneDNN
-            GIT_PROGRESS TRUE
+                GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-            GIT_SHALLOW TRUE
+                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")
@@ -329,13 +269,21 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
     set(ONEDNN_ENABLE_WORKLOAD "INFERENCE")
     set(ONEDNN_ENABLE_PRIMITIVE "MATMUL;REORDER")
     set(ONEDNN_BUILD_GRAPH "OFF")
-    set(ONEDNN_ENABLE_JIT_PROFILING "OFF")
+    set(ONEDNN_ENABLE_JIT_PROFILING "ON")
     set(ONEDNN_ENABLE_ITT_TASKS "OFF")
-    set(ONEDNN_ENABLE_MAX_CPU_ISA "OFF")
+    set(ONEDNN_ENABLE_MAX_CPU_ISA "ON")
-    set(ONEDNN_ENABLE_CPU_ISA_HINTS "OFF")
+    set(ONEDNN_ENABLE_CPU_ISA_HINTS "ON")
-    set(ONEDNN_VERBOSE "OFF")
+    set(ONEDNN_VERBOSE "ON")
     set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
 
+    # TODO: Refactor this
+    if (ENABLE_X86_ISA)
+        # Note: only enable oneDNN for AVX512
+        list(APPEND DNNL_COMPILE_FLAGS ${CXX_COMPILE_FLAGS_AVX512})
+    else()
+        list(APPEND DNNL_COMPILE_FLAGS ${CXX_COMPILE_FLAGS})
+    endif()
+
     set(VLLM_BUILD_TYPE ${CMAKE_BUILD_TYPE})
     set(CMAKE_BUILD_TYPE "Release") # remove oneDNN debug symbols to reduce size
     FetchContent_MakeAvailable(oneDNN)
@@ -348,14 +296,20 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
         PRIVATE ${oneDNN_SOURCE_DIR}/src
     )
     target_link_libraries(dnnl_ext dnnl torch)
-    target_compile_options(dnnl_ext PRIVATE ${CXX_COMPILE_FLAGS} -fPIC)
+    target_compile_options(dnnl_ext PRIVATE ${DNNL_COMPILE_FLAGS} -fPIC)
     list(APPEND LIBS dnnl_ext)
     set(USE_ONEDNN ON)
 else()
     set(USE_ONEDNN OFF)
 endif()
 
-message(STATUS "CPU extension compile flags: ${CXX_COMPILE_FLAGS}")
+# TODO: Refactor this
+if (ENABLE_X86_ISA)
+    message(STATUS "CPU extension (AVX512) compile flags: ${CXX_COMPILE_FLAGS_AVX512}")
+    message(STATUS "CPU extension (AVX2) compile flags: ${CXX_COMPILE_FLAGS_AVX2}")
+else()
+    message(STATUS "CPU extension compile flags: ${CXX_COMPILE_FLAGS}")
+endif()
 
 if(ENABLE_NUMA)
     list(APPEND LIBS numa)
@@ -390,25 +344,6 @@ set(VLLM_EXT_SRC
     "csrc/cpu/cpu_attn.cpp"
     "csrc/cpu/torch_bindings.cpp")
 
-if (AVX512_FOUND AND NOT AVX512_DISABLED)
-    set(VLLM_EXT_SRC
-        "csrc/cpu/shm.cpp"
-        "csrc/cpu/cpu_wna16.cpp"
-        "csrc/cpu/cpu_fused_moe.cpp"
-        ${VLLM_EXT_SRC})
-    if (ENABLE_AVX512BF16 AND ENABLE_AVX512VNNI)
-        set(VLLM_EXT_SRC
-            "csrc/cpu/sgl-kernels/gemm.cpp"
-            "csrc/cpu/sgl-kernels/gemm_int8.cpp"
-            "csrc/cpu/sgl-kernels/gemm_fp8.cpp"
-            "csrc/cpu/sgl-kernels/moe.cpp"
-            "csrc/cpu/sgl-kernels/moe_int8.cpp"
-            "csrc/cpu/sgl-kernels/moe_fp8.cpp"
-            ${VLLM_EXT_SRC})
-        add_compile_definitions(-DCPU_CAPABILITY_AVX512)
-    endif()
-endif()
-
 if (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND)
     set(VLLM_EXT_SRC
         "csrc/cpu/shm.cpp"
@@ -421,21 +356,83 @@ if(USE_ONEDNN)
         ${VLLM_EXT_SRC})
 endif()
 
-message(STATUS "CPU extension source files: ${VLLM_EXT_SRC}")
+if (ENABLE_X86_ISA)
+    set(VLLM_EXT_SRC_AVX512
+        "csrc/cpu/sgl-kernels/gemm.cpp"
+        "csrc/cpu/sgl-kernels/gemm_int8.cpp"
+        "csrc/cpu/sgl-kernels/gemm_fp8.cpp"
+        "csrc/cpu/sgl-kernels/moe.cpp"
+        "csrc/cpu/sgl-kernels/moe_int8.cpp"
+        "csrc/cpu/sgl-kernels/moe_fp8.cpp"
+        "csrc/cpu/shm.cpp"
+        "csrc/cpu/cpu_wna16.cpp"
+        "csrc/cpu/cpu_fused_moe.cpp"
+        "csrc/cpu/utils.cpp"
+        "csrc/cpu/cpu_attn.cpp"
+        "csrc/cpu/dnnl_kernels.cpp"
+        "csrc/cpu/torch_bindings.cpp"
+        # TODO: Remove these files
+        "csrc/cpu/activation.cpp"
+        "csrc/cpu/layernorm.cpp"
+        "csrc/cpu/mla_decode.cpp"
+        "csrc/cpu/pos_encoding.cpp"
+        "csrc/moe/dynamic_4bit_int_moe_cpu.cpp") 
 
-#
+    set(VLLM_EXT_SRC_AVX2 
-# Define extension targets
+        "csrc/cpu/utils.cpp"
-#
+        "csrc/cpu/cpu_attn.cpp"
+        "csrc/cpu/torch_bindings.cpp"
+        # TODO: Remove these files
+        "csrc/cpu/activation.cpp"
+        "csrc/cpu/layernorm.cpp"
+        "csrc/cpu/mla_decode.cpp"
+        "csrc/cpu/pos_encoding.cpp"
+        "csrc/moe/dynamic_4bit_int_moe_cpu.cpp") 
 
-define_extension_target(
+    message(STATUS "CPU extension (AVX512) source files: ${VLLM_EXT_SRC_AVX512}")
-    _C
+    message(STATUS "CPU extension (AVX2) source files: ${VLLM_EXT_SRC_AVX2}")
-    DESTINATION vllm
+
-    LANGUAGE CXX
+    define_extension_target(
-    SOURCES ${VLLM_EXT_SRC}
+        _C
-    LIBRARIES ${LIBS}
+        DESTINATION vllm
-    COMPILE_FLAGS ${CXX_COMPILE_FLAGS}
+        LANGUAGE CXX
-    USE_SABI 3
+        SOURCES ${VLLM_EXT_SRC_AVX512}
-    WITH_SOABI
+        LIBRARIES ${LIBS}
-)
+        COMPILE_FLAGS ${CXX_COMPILE_FLAGS_AVX512}
+        USE_SABI 3
+        WITH_SOABI
+    )
+
+    # For SGL kernels
+    target_compile_definitions(_C PRIVATE "-DCPU_CAPABILITY_AVX512")
+    # For AMX kernels
+    target_compile_definitions(_C PRIVATE "-DCPU_CAPABILITY_AMXBF16")
+
+    define_extension_target(
+        _C_AVX2
+        DESTINATION vllm
+        LANGUAGE CXX
+        SOURCES ${VLLM_EXT_SRC_AVX2}
+        LIBRARIES ${LIBS}
+        COMPILE_FLAGS ${CXX_COMPILE_FLAGS_AVX2}
+        USE_SABI 3
+        WITH_SOABI
+    )
+else()
+    message(STATUS "CPU extension source files: ${VLLM_EXT_SRC}")
+    #
+    # Define extension targets
+    #
+    define_extension_target(
+        _C
+        DESTINATION vllm
+        LANGUAGE CXX
+        SOURCES ${VLLM_EXT_SRC}
+        LIBRARIES ${LIBS}
+        COMPILE_FLAGS ${CXX_COMPILE_FLAGS}
+        USE_SABI 3
+        WITH_SOABI
+    )
+endif()
 
 message(STATUS "Enabling C extension.")

cmake/external_projects/vllm_flash_attn.cmake

@@ -17,7 +17,8 @@ endif()
 # They should be identical but if they aren't, this is a massive footgun.
 #
 # The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
-# To only install vllm-flash-attn, use --component _vllm_fa2_C (for FA2) or --component _vllm_fa3_C (for FA3).
+# To only install vllm-flash-attn, use --component _vllm_fa2_C (for FA2), --component _vllm_fa3_C (for FA3),
+# or --component _vllm_fa4_cutedsl_C (for FA4 CuteDSL Python files).
 # If no component is specified, vllm-flash-attn is still installed.
 
 # If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
@@ -38,22 +39,16 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 5824e6e2008271063c3229ab3e7032bd74abbbc6
+          GIT_TAG 140c00c0241bb60cc6e44e7c1be9998d4b20d8d2
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
   )
 endif()
 
-
-# Ensure the vllm/vllm_flash_attn directory exists before installation
-install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")" ALL_COMPONENTS)
-
 # Make sure vllm-flash-attn install rules are nested under vllm/
-# This is here to support installing all components under the same prefix with cmake --install.
+# ALL_COMPONENTS ensures the save/modify/restore runs exactly once regardless
-# setup.py installs every component separately but uses the same prefix for all.
+# of how many components are being installed, avoiding double-append of /vllm/.
-# ALL_COMPONENTS is used to avoid duplication for FA2 and FA3,
-# and these statements don't hurt when installing neither component.
 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)
@@ -62,22 +57,48 @@ install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" ALL_
 FetchContent_MakeAvailable(vllm-flash-attn)
 message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
 
-# Restore the install prefix
+# Restore the install prefix after FA's install rules
 install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
 install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
 
-# Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in
+# Install shared Python files for both FA2 and FA3 components
-# case only one is built, in the case both are built redundant work is done)
+foreach(_FA_COMPONENT _vllm_fa2_C _vllm_fa3_C)
-install(
+  # Ensure the vllm/vllm_flash_attn directory exists before installation
-  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
+  install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")"
-  DESTINATION vllm/vllm_flash_attn
+    COMPONENT ${_FA_COMPONENT})
-  COMPONENT _vllm_fa2_C
-  FILES_MATCHING PATTERN "*.py"
-)
 
-install(
+  # Copy vllm_flash_attn python files (except __init__.py and flash_attn_interface.py
-  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
+  # which are source-controlled in vllm)
-  DESTINATION vllm/vllm_flash_attn
+  install(
-  COMPONENT _vllm_fa3_C
+    DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  FILES_MATCHING PATTERN "*.py"
+    DESTINATION vllm/vllm_flash_attn
-)
+    COMPONENT ${_FA_COMPONENT}
+    FILES_MATCHING PATTERN "*.py"
+    PATTERN "__init__.py" EXCLUDE
+    PATTERN "flash_attn_interface.py" EXCLUDE
+  )
+
+endforeach()
+
+#
+# FA4 CuteDSL component
+# This is a Python-only component that copies the flash_attn/cute directory
+# and transforms imports to match our package structure.
+#
+add_custom_target(_vllm_fa4_cutedsl_C)
+
+# Copy flash_attn/cute directory (needed for FA4) and transform imports
+# The cute directory uses flash_attn.cute imports internally, which we replace
+# with vllm.vllm_flash_attn.cute to match our package structure.
+install(CODE "
+  file(GLOB_RECURSE CUTE_PY_FILES \"${vllm-flash-attn_SOURCE_DIR}/flash_attn/cute/*.py\")
+  foreach(SRC_FILE \${CUTE_PY_FILES})
+    file(RELATIVE_PATH REL_PATH \"${vllm-flash-attn_SOURCE_DIR}/flash_attn/cute\" \${SRC_FILE})
+    set(DST_FILE \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn/cute/\${REL_PATH}\")
+    get_filename_component(DST_DIR \${DST_FILE} DIRECTORY)
+    file(MAKE_DIRECTORY \${DST_DIR})
+    file(READ \${SRC_FILE} FILE_CONTENTS)
+    string(REPLACE \"flash_attn.cute\" \"vllm.vllm_flash_attn.cute\" FILE_CONTENTS \"\${FILE_CONTENTS}\")
+    file(WRITE \${DST_FILE} \"\${FILE_CONTENTS}\")
+  endforeach()
+" COMPONENT _vllm_fa4_cutedsl_C)

csrc/activation_kernels.cu

@@ -5,115 +5,11 @@
 #include <cmath>
 
 #include "cuda_compat.h"
+#include "cuda_vec_utils.cuh"
 #include "dispatch_utils.h"
 
 namespace vllm {
 
-struct alignas(32) u32x8_t {
-  uint32_t u0, u1, u2, u3, u4, u5, u6, u7;
-};
-
-__device__ __forceinline__ void ld256(u32x8_t& val, const u32x8_t* ptr) {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000
-  asm volatile("ld.global.nc.v8.u32 {%0,%1,%2,%3,%4,%5,%6,%7}, [%8];\n"
-               : "=r"(val.u0), "=r"(val.u1), "=r"(val.u2), "=r"(val.u3),
-                 "=r"(val.u4), "=r"(val.u5), "=r"(val.u6), "=r"(val.u7)
-               : "l"(ptr));
-#else
-  const uint4* uint_ptr = reinterpret_cast<const uint4*>(ptr);
-  uint4 top_half = __ldg(&uint_ptr[0]);
-  uint4 bottom_half = __ldg(&uint_ptr[1]);
-  val.u0 = top_half.x;
-  val.u1 = top_half.y;
-  val.u2 = top_half.z;
-  val.u3 = top_half.w;
-  val.u4 = bottom_half.x;
-  val.u5 = bottom_half.y;
-  val.u6 = bottom_half.z;
-  val.u7 = bottom_half.w;
-#endif
-}
-
-__device__ __forceinline__ void st256(u32x8_t& val, u32x8_t* ptr) {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000
-  asm volatile("st.global.v8.u32 [%0], {%1,%2,%3,%4,%5,%6,%7,%8};\n"
-               :
-               : "l"(ptr), "r"(val.u0), "r"(val.u1), "r"(val.u2), "r"(val.u3),
-                 "r"(val.u4), "r"(val.u5), "r"(val.u6), "r"(val.u7)
-               : "memory");
-#else
-  uint4* uint_ptr = reinterpret_cast<uint4*>(ptr);
-  uint_ptr[0] = make_uint4(val.u0, val.u1, val.u2, val.u3);
-  uint_ptr[1] = make_uint4(val.u4, val.u5, val.u6, val.u7);
-#endif
-}
-
-template <bool support_256>
-struct VecTraits;
-
-template <>
-struct VecTraits<true> {
-  static constexpr int ARCH_MAX_VEC_SIZE = 32;
-  using vec_t = u32x8_t;
-};
-
-template <>
-struct VecTraits<false> {
-  static constexpr int ARCH_MAX_VEC_SIZE = 16;
-  using vec_t = int4;
-};
-
-template <typename T>
-struct PackedTraits;
-
-template <>
-struct PackedTraits<c10::BFloat16> {
-  using packed_t = __nv_bfloat162;
-};
-
-template <>
-struct PackedTraits<c10::Half> {
-  using packed_t = __half2;
-};
-
-template <>
-struct PackedTraits<float> {
-  using packed_t = float2;
-};
-
-template <typename packed_t>
-__device__ __forceinline__ float2 cast_to_float2(const packed_t& val) {
-  if constexpr (std::is_same_v<packed_t, __nv_bfloat162>) {
-    return __bfloat1622float2(val);
-  } else if constexpr (std::is_same_v<packed_t, __half2>) {
-    return __half22float2(val);
-  } else if constexpr (std::is_same_v<packed_t, float2>) {
-    return float2(val);
-  }
-}
-
-template <typename packed_t>
-__device__ __forceinline__ packed_t cast_to_packed(const float2& val) {
-  if constexpr (std::is_same_v<packed_t, __nv_bfloat162>) {
-    return __float22bfloat162_rn(val);
-  } else if constexpr (std::is_same_v<packed_t, __half2>) {
-    return __float22half2_rn(val);
-  } else if constexpr (std::is_same_v<packed_t, float2>) {
-    return float2(val);
-  }
-}
-
-template <typename packed_t>
-__device__ __forceinline__ packed_t packed_mul(const packed_t& x,
-                                               const packed_t& y) {
-  if constexpr (std::is_same_v<packed_t, __nv_bfloat162> ||
-                std::is_same_v<packed_t, __half2>) {
-    return __hmul2(x, y);
-  } else if constexpr (std::is_same_v<packed_t, float2>) {
-    return make_float2(x.x * y.x, x.y * y.y);
-  }
-}
-
 template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
           bool act_first>
 __device__ __forceinline__ scalar_t compute(const scalar_t& x,
@@ -129,16 +25,6 @@ __device__ __forceinline__ packed_t packed_compute(const packed_t& x,
                    : packed_mul(x, PACKED_ACT_FN(y));
 }
 
-// Check if all pointers are 16-byte aligned for int4 vectorized access
-__host__ __device__ __forceinline__ bool is_16byte_aligned(const void* ptr) {
-  return (reinterpret_cast<uintptr_t>(ptr) & 15) == 0;
-}
-
-// Check if all pointers are 16-byte aligned for longlong4_32a vectorized access
-__host__ __device__ __forceinline__ bool is_32byte_aligned(const void* ptr) {
-  return (reinterpret_cast<uintptr_t>(ptr) & 31) == 0;
-}
-
 // Activation and gating kernel template.
 template <typename scalar_t, typename packed_t,
           scalar_t (*ACT_FN)(const scalar_t&),
@@ -153,36 +39,32 @@ __global__ void act_and_mul_kernel(
   scalar_t* out_ptr = out + blockIdx.x * d;
 
   if constexpr (use_vec) {
-    // Fast path: 128-bit/256-bit vectorized loop
+    using cuda_t = typename CUDATypeConverter<scalar_t>::Type;
-    using vec_t = typename VecTraits<use_256b>::vec_t;
+    using pvec_t = PackedVec<cuda_t, use_256b>;
-    constexpr int ARCH_MAX_VEC_SIZE = VecTraits<use_256b>::ARCH_MAX_VEC_SIZE;
-    constexpr int VEC_SIZE = ARCH_MAX_VEC_SIZE / sizeof(packed_t);
 
-    const vec_t* x_vec = reinterpret_cast<const vec_t*>(x_ptr);
+    const pvec_t* x_vec = reinterpret_cast<const pvec_t*>(x_ptr);
-    const vec_t* y_vec = reinterpret_cast<const vec_t*>(y_ptr);
+    const pvec_t* y_vec = reinterpret_cast<const pvec_t*>(y_ptr);
-    vec_t* out_vec = reinterpret_cast<vec_t*>(out_ptr);
+    pvec_t* out_vec = reinterpret_cast<pvec_t*>(out_ptr);
-    const int num_vecs = d / 2 / VEC_SIZE;
+    const int num_vecs = d / 2 / pvec_t::NUM_ELTS;
 
     for (int i = threadIdx.x; i < num_vecs; i += blockDim.x) {
-      vec_t x, y;
+      pvec_t x, y;
       if constexpr (use_256b) {
         ld256(x, &x_vec[i]);
         ld256(y, &y_vec[i]);
       } else {
-        x = VLLM_LDG(&x_vec[i]);
+        ld128(x, &x_vec[i]);
-        y = VLLM_LDG(&y_vec[i]);
+        ld128(y, &y_vec[i]);
       }
-      auto* xp = reinterpret_cast<packed_t*>(&x);
-      auto* yp = reinterpret_cast<packed_t*>(&y);
 #pragma unroll
-      for (int j = 0; j < VEC_SIZE; j++) {
+      for (int j = 0; j < pvec_t::NUM_ELTS; j++) {
-        xp[j] =
+        x.elts[j] = packed_compute<packed_t, PACKED_ACT_FN, act_first>(
-            packed_compute<packed_t, PACKED_ACT_FN, act_first>(xp[j], yp[j]);
+            x.elts[j], y.elts[j]);
       }
       if constexpr (use_256b) {
         st256(x, &out_vec[i]);
       } else {
-        out_vec[i] = x;
+        st128(x, &out_vec[i]);
       }
     }
   } else {
@@ -270,51 +152,54 @@ packed_gelu_tanh_kernel(const packed_t& val) {
 // Launch activation and gating kernel.
 // Use ACT_FIRST (bool) indicating whether to apply the activation function
 // first.
-#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, PACKED_KERNEL, ACT_FIRST)     \
+#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, PACKED_KERNEL, ACT_FIRST)        \
-  auto dtype = input.scalar_type();                                         \
+  auto dtype = input.scalar_type();                                            \
-  int d = input.size(-1) / 2;                                               \
+  int d = input.size(-1) / 2;                                                  \
-  int64_t num_tokens = input.numel() / input.size(-1);                      \
+  int64_t num_tokens = input.numel() / input.size(-1);                         \
-  if (num_tokens == 0) {                                                    \
+  if (num_tokens == 0) {                                                       \
-    return;                                                                 \
+    return;                                                                    \
-  }                                                                         \
+  }                                                                            \
-  dim3 grid(num_tokens);                                                    \
+  dim3 grid(num_tokens);                                                       \
-  int cc_major = at::cuda::getCurrentDeviceProperties()->major;             \
+  int cc_major = at::cuda::getCurrentDeviceProperties()->major;                \
-  int support_vec = (cc_major >= 10 && num_tokens > 128) ? 32 : 16;         \
+  int support_vec =                                                            \
-  int vec_size = support_vec / at::elementSize(dtype);                      \
+      (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128)            \
-  const bool use_vec = (d % vec_size == 0);                                 \
+          ? vllm::VecTraits<true>::ARCH_MAX_VEC_SIZE                           \
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));         \
+          : vllm::VecTraits<false>::ARCH_MAX_VEC_SIZE;                         \
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();             \
+  int vec_size = support_vec / at::elementSize(dtype);                         \
-  if (use_vec) {                                                            \
+  const bool use_vec = (d % vec_size == 0);                                    \
-    dim3 block(std::min(d / vec_size, 1024));                               \
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));            \
-    if (cc_major >= 10 && num_tokens > 128) {                               \
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                \
-      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {       \
+  if (use_vec) {                                                               \
-        vllm::act_and_mul_kernel<                                           \
+    dim3 block(std::min(d / vec_size, 1024));                                  \
-            scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,      \
+    if (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128) {         \
-            KERNEL<scalar_t>,                                               \
+      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {          \
-            PACKED_KERNEL<typename vllm::PackedTraits<scalar_t>::packed_t>, \
+        vllm::act_and_mul_kernel<                                              \
-            ACT_FIRST, true, true><<<grid, block, 0, stream>>>(             \
+            scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,      \
-            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);       \
+            KERNEL<scalar_t>,                                                  \
-      });                                                                   \
+            PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
-    } else {                                                                \
+            ACT_FIRST, true, true><<<grid, block, 0, stream>>>(                \
-      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {       \
+            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);          \
-        vllm::act_and_mul_kernel<                                           \
+      });                                                                      \
-            scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,      \
+    } else {                                                                   \
-            KERNEL<scalar_t>,                                               \
+      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {          \
-            PACKED_KERNEL<typename vllm::PackedTraits<scalar_t>::packed_t>, \
+        vllm::act_and_mul_kernel<                                              \
-            ACT_FIRST, true, false><<<grid, block, 0, stream>>>(            \
+            scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,      \
-            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);       \
+            KERNEL<scalar_t>,                                                  \
-      });                                                                   \
+            PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
-    }                                                                       \
+            ACT_FIRST, true, false><<<grid, block, 0, stream>>>(               \
-  } else {                                                                  \
+            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);          \
-    dim3 block(std::min(d, 1024));                                          \
+      });                                                                      \
-    VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {         \
+    }                                                                          \
-      vllm::act_and_mul_kernel<                                             \
+  } else {                                                                     \
-          scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,        \
+    dim3 block(std::min(d, 1024));                                             \
-          KERNEL<scalar_t>,                                                 \
+    VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {            \
-          PACKED_KERNEL<typename vllm::PackedTraits<scalar_t>::packed_t>,   \
+      vllm::act_and_mul_kernel<                                                \
-          ACT_FIRST, false><<<grid, block, 0, stream>>>(                    \
+          scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,        \
-          out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);         \
+          KERNEL<scalar_t>,                                                    \
-    });                                                                     \
+          PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>,   \
+          ACT_FIRST, false><<<grid, block, 0, stream>>>(                       \
+          out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);            \
+    });                                                                        \
   }
 
 void silu_and_mul(torch::Tensor& out,    // [..., d]
@@ -376,35 +261,31 @@ __global__ void act_and_mul_kernel_with_param(
   scalar_t* out_ptr = out + blockIdx.x * d;
 
   if constexpr (use_vec) {
-    // Fast path: 128-bit/256-bit vectorized loop
+    using cuda_t = typename CUDATypeConverter<scalar_t>::Type;
-    using vec_t = typename VecTraits<use_256b>::vec_t;
+    using pvec_t = PackedVec<cuda_t, use_256b>;
-    constexpr int ARCH_MAX_VEC_SIZE = VecTraits<use_256b>::ARCH_MAX_VEC_SIZE;
-    constexpr int VEC_SIZE = ARCH_MAX_VEC_SIZE / sizeof(packed_t);
 
-    const vec_t* x_vec = reinterpret_cast<const vec_t*>(x_ptr);
+    const pvec_t* x_vec = reinterpret_cast<const pvec_t*>(x_ptr);
-    const vec_t* y_vec = reinterpret_cast<const vec_t*>(y_ptr);
+    const pvec_t* y_vec = reinterpret_cast<const pvec_t*>(y_ptr);
-    vec_t* out_vec = reinterpret_cast<vec_t*>(out_ptr);
+    pvec_t* out_vec = reinterpret_cast<pvec_t*>(out_ptr);
-    const int num_vecs = d / 2 / VEC_SIZE;
+    const int num_vecs = d / 2 / pvec_t::NUM_ELTS;
 
     for (int i = threadIdx.x; i < num_vecs; i += blockDim.x) {
-      vec_t x, y;
+      pvec_t x, y;
       if constexpr (use_256b) {
         ld256(x, &x_vec[i]);
         ld256(y, &y_vec[i]);
       } else {
-        x = VLLM_LDG(&x_vec[i]);
+        ld128(x, &x_vec[i]);
-        y = VLLM_LDG(&y_vec[i]);
+        ld128(y, &y_vec[i]);
       }
-      auto* xp = reinterpret_cast<packed_t*>(&x);
-      auto* yp = reinterpret_cast<packed_t*>(&y);
 #pragma unroll
-      for (int j = 0; j < VEC_SIZE; j++) {
+      for (int j = 0; j < pvec_t::NUM_ELTS; j++) {
-        xp[j] = packed_mul(PACKED_ACT_FN(xp[j], param), yp[j]);
+        x.elts[j] = packed_mul(PACKED_ACT_FN(x.elts[j], param), y.elts[j]);
       }
       if constexpr (use_256b) {
         st256(x, &out_vec[i]);
       } else {
-        out_vec[i] = x;
+        st128(x, &out_vec[i]);
       }
     }
   } else {
@@ -497,21 +378,24 @@ __global__ void swigluoai_and_mul_kernel(
   }                                                                            \
   dim3 grid(num_tokens);                                                       \
   int cc_major = at::cuda::getCurrentDeviceProperties()->major;                \
-  int support_vec = (cc_major >= 10 && num_tokens > 128) ? 32 : 16;            \
+  int support_vec =                                                            \
+      (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128)            \
+          ? vllm::VecTraits<true>::ARCH_MAX_VEC_SIZE                           \
+          : vllm::VecTraits<false>::ARCH_MAX_VEC_SIZE;                         \
   int vec_size = support_vec / at::elementSize(dtype);                         \
   const bool use_vec = (d % vec_size == 0);                                    \
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));            \
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                \
   if (use_vec) {                                                               \
     dim3 block(std::min(d / vec_size, 1024));                                  \
-    if (cc_major >= 10 && num_tokens > 128) {                                  \
+    if (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128) {         \
       VLLM_DISPATCH_FLOATING_TYPES(                                            \
           dtype, "act_and_mul_kernel_with_param", [&] {                        \
             vllm::act_and_mul_kernel_with_param<                               \
-                scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,     \
+                scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,  \
                 KERNEL<scalar_t>,                                              \
                 PACKED_KERNEL<                                                 \
-                    typename vllm::PackedTraits<scalar_t>::packed_t>,          \
+                    typename vllm::PackedTypeConverter<scalar_t>::Type>,       \
                 true, true><<<grid, block, 0, stream>>>(                       \
                 out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d,       \
                 PARAM);                                                        \
@@ -520,10 +404,10 @@ __global__ void swigluoai_and_mul_kernel(
       VLLM_DISPATCH_FLOATING_TYPES(                                            \
           dtype, "act_and_mul_kernel_with_param", [&] {                        \
             vllm::act_and_mul_kernel_with_param<                               \
-                scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,     \
+                scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,  \
                 KERNEL<scalar_t>,                                              \
                 PACKED_KERNEL<                                                 \
-                    typename vllm::PackedTraits<scalar_t>::packed_t>,          \
+                    typename vllm::PackedTypeConverter<scalar_t>::Type>,       \
                 true, false><<<grid, block, 0, stream>>>(                      \
                 out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d,       \
                 PARAM);                                                        \
@@ -533,9 +417,9 @@ __global__ void swigluoai_and_mul_kernel(
     dim3 block(std::min(d, 1024));                                             \
     VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel_with_param", [&] { \
       vllm::act_and_mul_kernel_with_param<                                     \
-          scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,           \
+          scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,        \
           KERNEL<scalar_t>,                                                    \
-          PACKED_KERNEL<typename vllm::PackedTraits<scalar_t>::packed_t>,      \
+          PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>,   \
           false><<<grid, block, 0, stream>>>(                                  \
           out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, PARAM);     \
     });                                                                        \
@@ -627,14 +511,17 @@ __global__ void activation_kernel(
   }                                                                      \
   dim3 grid(num_tokens);                                                 \
   int cc_major = at::cuda::getCurrentDeviceProperties()->major;          \
-  int support_vec = (cc_major >= 10 && num_tokens > 128) ? 32 : 16;      \
+  int support_vec =                                                      \
+      (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128)      \
+          ? vllm::VecTraits<true>::ARCH_MAX_VEC_SIZE                     \
+          : vllm::VecTraits<false>::ARCH_MAX_VEC_SIZE;                   \
   int vec_size = support_vec / at::elementSize(dtype);                   \
   const bool use_vec = (d % vec_size == 0);                              \
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));      \
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();          \
   if (use_vec) {                                                         \
     dim3 block(std::min(d / vec_size, 1024));                            \
-    if (cc_major >= 10 && num_tokens > 128) {                            \
+    if (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128) {   \
       VLLM_DISPATCH_FLOATING_TYPES(dtype, "activation_kernel", [&] {     \
         vllm::activation_kernel<scalar_t, KERNEL<scalar_t>, true, true>  \
             <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),       \

csrc/cache_kernels.cu

@@ -1305,7 +1305,8 @@ void indexer_k_quant_and_cache(
   const at::cuda::OptionalCUDAGuard device_guard(device_of(k));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  DISPATCH_BY_KV_CACHE_DTYPE(k.dtype(), "fp8_e4m3",
+  static const std::string kv_cache_dtype = "fp8_e4m3";
+  DISPATCH_BY_KV_CACHE_DTYPE(k.dtype(), kv_cache_dtype,
                              CALL_INDEXER_K_QUANT_AND_CACHE);
 }
 

csrc/cpu/cpu_attn.cpp

@@ -16,6 +16,8 @@ torch::Tensor get_scheduler_metadata(
     isa = cpu_attention::ISA::VEC16;
   } else if (isa_hint == "neon") {
     isa = cpu_attention::ISA::NEON;
+  } else if (isa_hint == "vxe") {
+    isa = cpu_attention::ISA::VXE;
   } else {
     TORCH_CHECK(false, "Unsupported CPU attention ISA hint: " + isa_hint);
   }
@@ -100,6 +102,8 @@ void cpu_attn_reshape_and_cache(
       return cpu_attention::ISA::VEC16;
     } else if (isa == "neon") {
       return cpu_attention::ISA::NEON;
+    } else if (isa == "vxe") {
+      return cpu_attention::ISA::VXE;
     } else {
       TORCH_CHECK(false, "Invalid ISA type: " + isa);
     }

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/cpu_attn_impl.hpp

@@ -12,7 +12,7 @@
 #include "cpu/utils.hpp"
 
 namespace cpu_attention {
-enum class ISA { AMX, VEC, VEC16, NEON };
+enum class ISA { AMX, VEC, VEC16, NEON, VXE };
 
 template <ISA isa, typename scalar_t, int64_t head_dim>
 class AttentionImpl {};

csrc/cpu/cpu_attn_vxe.hpp

@@ -0,0 +1,386 @@
+#ifndef CPU_ATTN_VXE_HPP
+#define CPU_ATTN_VXE_HPP
+
+#include "cpu_attn_impl.hpp"
+#include <vecintrin.h>
+#include <type_traits>
+
+namespace cpu_attention {
+
+namespace {
+
+// s390x Vector = 16 bytes (128 bits)
+#define BLOCK_SIZE_ALIGNMENT 32
+#define HEAD_SIZE_ALIGNMENT 32
+#define MAX_Q_HEAD_NUM_PER_ITER 16
+
+template <typename kv_cache_t>
+FORCE_INLINE void load_row8_B_as_f32(const kv_cache_t* p, __vector float& b0,
+                                     __vector float& b1);
+
+// [1] Float Specialization
+template <>
+FORCE_INLINE void load_row8_B_as_f32<float>(const float* p, __vector float& b0,
+                                            __vector float& b1) {
+  // Explicitly cast to long long for offset, and float* for pointer
+  b0 = vec_xl((long long)0, const_cast<float*>(p));
+  b1 = vec_xl((long long)0, const_cast<float*>(p + 4));
+}
+
+// [2] BFloat16 Specialization (Big Endian Fix)
+template <>
+FORCE_INLINE void load_row8_B_as_f32<c10::BFloat16>(const c10::BFloat16* p,
+                                                    __vector float& b0,
+                                                    __vector float& b1) {
+  // 1. Load 8 BF16s (16 bytes) into one vector
+  // Explicit cast to unsigned short* for vec_xl to return vector unsigned short
+  __vector unsigned short raw = vec_xl((long long)0, (unsigned short*)p);
+
+  // 2. Prepare Zero vector
+  __vector unsigned short zeros = vec_splat_u16(0);
+
+  // 3. Merge High/Low to expand BF16 -> Float32
+  // On Big Endian, a float is [BF16_bits | 16_zero_bits]
+  b0 = (__vector float)vec_mergeh(raw, zeros);
+  b1 = (__vector float)vec_mergel(raw, zeros);
+}
+
+template <>
+FORCE_INLINE void load_row8_B_as_f32<c10::Half>(const c10::Half* p,
+                                                __vector float& b0,
+                                                __vector float& b1) {
+  alignas(16) float tmp[8];
+
+  // Manual unroll / conversion
+  tmp[0] = static_cast<float>(p[0]);
+  tmp[1] = static_cast<float>(p[1]);
+  tmp[2] = static_cast<float>(p[2]);
+  tmp[3] = static_cast<float>(p[3]);
+  tmp[4] = static_cast<float>(p[4]);
+  tmp[5] = static_cast<float>(p[5]);
+  tmp[6] = static_cast<float>(p[6]);
+  tmp[7] = static_cast<float>(p[7]);
+
+  // Explicit arguments for intrinsic: (long long offset, float* ptr)
+  b0 = vec_xl((long long)0, (float*)tmp);
+  b1 = vec_xl((long long)0, (float*)(tmp + 4));
+}
+
+template <int32_t M, typename kv_cache_t>
+FORCE_INLINE void gemm_micro_s390x_Mx8_Ku4(
+    const float* __restrict A,       // [M x K]
+    const kv_cache_t* __restrict B,  // [K x 8]
+    float* __restrict C,             // [M x 8]
+    int64_t lda, int64_t ldb, int64_t ldc, int32_t K, bool accumulate) {
+  static_assert(1 <= M && M <= 8, "M must be in [1,8]");
+
+// Helper macros to unroll codegen for M rows
+#define ROWS_APPLY(OP) OP(0) OP(1) OP(2) OP(3) OP(4) OP(5) OP(6) OP(7)
+#define IF_M(i) if constexpr (M > (i))
+
+  // 1. Define A pointers
+#define DECL_A(i) const float* a##i = A + (i) * lda;
+  ROWS_APPLY(DECL_A)
+#undef DECL_A
+
+  // 2. Define Accumulators (2 vectors covers 8 columns)
+#define DECL_ACC(i) __vector float acc##i##_0, acc##i##_1;
+  ROWS_APPLY(DECL_ACC)
+#undef DECL_ACC
+
+  // 3. Initialize Accumulators (Load C or Zero)
+#define INIT_ACC(i)                                                    \
+  IF_M(i) {                                                            \
+    if (accumulate) {                                                  \
+      acc##i##_0 =                                                     \
+          vec_xl((long long)0, const_cast<float*>(C + (i) * ldc + 0)); \
+      acc##i##_1 =                                                     \
+          vec_xl((long long)0, const_cast<float*>(C + (i) * ldc + 4)); \
+    } else {                                                           \
+      acc##i##_0 = vec_splats(0.0f);                                   \
+      acc##i##_1 = vec_splats(0.0f);                                   \
+    }                                                                  \
+  }
+  ROWS_APPLY(INIT_ACC)
+#undef INIT_ACC
+
+  int32_t k = 0;
+
+  for (; k + 3 < K; k += 4) {
+    // Load 4 values of A for each Row M: A[k...k+3]
+#define LOAD_A4(i)        \
+  __vector float a##i##v; \
+  IF_M(i) a##i##v = vec_xl((long long)0, const_cast<float*>(a##i + k));
+    ROWS_APPLY(LOAD_A4)
+#undef LOAD_A4
+
+    // Helper: FMA for specific lane L of A
+    // s390x: vec_madd(b, vec_splat(a, lane), acc)
+#define FMAS_LANE(i, aiv, L)                        \
+  IF_M(i) {                                         \
+    __vector float a_broad = vec_splat(aiv, L);     \
+    acc##i##_0 = vec_madd(b0, a_broad, acc##i##_0); \
+    acc##i##_1 = vec_madd(b1, a_broad, acc##i##_1); \
+  }
+
+    // Unroll K=0..3
+    {
+      __vector float b0, b1;
+      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 0) * ldb, b0, b1);
+#define STEP_K0(i) FMAS_LANE(i, a##i##v, 0)
+      ROWS_APPLY(STEP_K0)
+#undef STEP_K0
+    }
+    {
+      __vector float b0, b1;
+      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 1) * ldb, b0, b1);
+#define STEP_K1(i) FMAS_LANE(i, a##i##v, 1)
+      ROWS_APPLY(STEP_K1)
+#undef STEP_K1
+    }
+    {
+      __vector float b0, b1;
+      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 2) * ldb, b0, b1);
+#define STEP_K2(i) FMAS_LANE(i, a##i##v, 2)
+      ROWS_APPLY(STEP_K2)
+#undef STEP_K2
+    }
+
+    {
+      __vector float b0, b1;
+      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 3) * ldb, b0, b1);
+#define STEP_K3(i) FMAS_LANE(i, a##i##v, 3)
+      ROWS_APPLY(STEP_K3)
+#undef STEP_K3
+    }
+#undef FMAS_LANE
+  }
+
+  for (; k < K; ++k) {
+    __vector float b0, b1;
+    load_row8_B_as_f32<kv_cache_t>(B + (int64_t)k * ldb, b0, b1);
+#define TAIL_ROW(i)                              \
+  IF_M(i) {                                      \
+    __vector float ai = vec_splats(*(a##i + k)); \
+    acc##i##_0 = vec_madd(b0, ai, acc##i##_0);   \
+    acc##i##_1 = vec_madd(b1, ai, acc##i##_1);   \
+  }
+    ROWS_APPLY(TAIL_ROW)
+#undef TAIL_ROW
+  }
+
+#define STORE_ROW(i)                           \
+  IF_M(i) {                                    \
+    vec_xst(acc##i##_0, 0, C + (i) * ldc + 0); \
+    vec_xst(acc##i##_1, 0, C + (i) * ldc + 4); \
+  }
+  ROWS_APPLY(STORE_ROW)
+#undef STORE_ROW
+
+#undef ROWS_APPLY
+#undef IF_M
+}
+
+template <int32_t N, typename kv_cache_t>
+FORCE_INLINE void gemm_macro_s390x_Mx8_Ku4(const float* __restrict A,
+                                           const kv_cache_t* __restrict B,
+                                           float* __restrict C, int32_t M,
+                                           int32_t K, int64_t lda, int64_t ldb,
+                                           int64_t ldc, bool accumulate) {
+  static_assert(N % 8 == 0, "N must be a multiple of 8");
+  for (int32_t m = 0; m < M;) {
+    int32_t mb = (M - m >= 8) ? 8 : (M - m >= 4) ? 4 : (M - m >= 2) ? 2 : 1;
+    const float* Ab = A + m * lda;
+    float* Cb = C + m * ldc;
+
+    for (int32_t n = 0; n < N; n += 8) {
+      const kv_cache_t* Bn = B + n;
+      float* Cn = Cb + n;
+      switch (mb) {
+        case 8:
+          gemm_micro_s390x_Mx8_Ku4<8, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc, K,
+                                                  accumulate);
+          break;
+        case 4:
+          gemm_micro_s390x_Mx8_Ku4<4, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc, K,
+                                                  accumulate);
+          break;
+        case 2:
+          gemm_micro_s390x_Mx8_Ku4<2, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc, K,
+                                                  accumulate);
+          break;
+        default:
+          gemm_micro_s390x_Mx8_Ku4<1, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc, K,
+                                                  accumulate);
+          break;
+      }
+    }
+    m += mb;
+  }
+}
+
+template <typename kv_cache_t>
+class TileGemmS390X {
+ public:
+  template <AttentionGemmPhase phase, int32_t k_size>
+  FORCE_INLINE static void gemm(const int32_t m_size,
+                                float* __restrict__ a_tile,
+                                kv_cache_t* __restrict__ b_tile,
+                                float* __restrict__ c_tile, const int64_t lda,
+                                const int64_t ldb, const int64_t ldc,
+                                const int32_t block_size,
+                                const int32_t dynamic_k_size,
+                                const bool accum_c) {
+    if constexpr (phase == AttentionGemmPhase::QK) {
+      gemm_macro_s390x_Mx8_Ku4<BLOCK_SIZE_ALIGNMENT, kv_cache_t>(
+          a_tile, b_tile, c_tile, m_size, k_size, lda, ldb, ldc, accum_c);
+    } else {
+      gemm_macro_s390x_Mx8_Ku4<HEAD_SIZE_ALIGNMENT, kv_cache_t>(
+          a_tile, b_tile, c_tile, m_size, dynamic_k_size, lda, ldb, ldc,
+          accum_c);
+    }
+  }
+};
+
+}  // namespace
+
+template <typename scalar_t, int64_t head_dim>
+class AttentionImpl<ISA::VXE, scalar_t, head_dim> {
+ public:
+  using query_t = scalar_t;
+  using q_buffer_t = float;
+  using kv_cache_t = scalar_t;
+  using logits_buffer_t = float;
+  using partial_output_buffer_t = float;
+  using prob_buffer_t = float;
+
+  constexpr static int64_t BlockSizeAlignment = BLOCK_SIZE_ALIGNMENT;
+  constexpr static int64_t HeadDimAlignment = HEAD_SIZE_ALIGNMENT;
+  constexpr static int64_t MaxQHeadNumPerIteration = MAX_Q_HEAD_NUM_PER_ITER;
+  constexpr static int64_t HeadDim = head_dim;
+  constexpr static ISA ISAType = ISA::VXE;
+  constexpr static bool scale_on_logits =
+      false;  // Scale is applied to Q during copy
+
+ public:
+  AttentionImpl() {}
+
+  template <template <typename tile_gemm_t> typename attention>
+  FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
+    attention<TileGemmS390X<kv_cache_t>> attention_iteration;
+    attention_iteration(CPU_ATTENTION_PARAMS);
+  }
+
+  // Strides for Memory Layout
+  constexpr static int64_t k_cache_token_group_stride(
+      const int32_t block_size) {
+    return BlockSizeAlignment;  // [head_dim, block_size] layout
+  }
+
+  constexpr static int64_t v_cache_token_group_stride(
+      const int32_t block_size) {
+    return head_dim * BlockSizeAlignment;
+  }
+
+  constexpr static int64_t v_cache_head_group_stride(const int32_t block_size) {
+    return HeadDimAlignment;
+  }
+
+  static void copy_q_heads_tile(scalar_t* __restrict__ src,
+                                float* __restrict__ q_buffer,
+                                const int32_t q_num,
+                                const int32_t q_heads_per_kv,
+                                const int64_t q_num_stride,
+                                const int64_t q_head_stride, float scale) {
+    __vector float scale_vec = vec_splats(scale);
+    constexpr bool is_bf16 = std::is_same<scalar_t, c10::BFloat16>::value;
+
+    // Process 8 elements at a time (32 bytes of float output)
+    for (int32_t i = 0; i < q_num; ++i) {
+      for (int32_t h = 0; h < q_heads_per_kv; ++h) {
+        scalar_t* curr_src = src + i * q_num_stride + h * q_head_stride;
+        float* curr_dst =
+            q_buffer + i * q_heads_per_kv * head_dim + h * head_dim;
+
+        int32_t d = 0;
+        for (; d <= head_dim - 8; d += 8) {
+          if constexpr (is_bf16) {
+            __vector float v0, v1;
+            // Reuse our Big-Endian-Safe loader
+            load_row8_B_as_f32<scalar_t>(curr_src + d, v0, v1);
+
+            v0 = vec_mul(v0, scale_vec);
+            v1 = vec_mul(v1, scale_vec);
+
+            vec_xst(v0, 0, curr_dst + d);
+            vec_xst(v1, 0, curr_dst + d + 4);
+          } else {
+            __vector float v0 = vec_xl((long long)0, (float*)curr_src + d);
+            __vector float v1 = vec_xl((long long)0, (float*)curr_src + d + 4);
+
+            v0 = vec_mul(v0, scale_vec);
+            v1 = vec_mul(v1, scale_vec);
+
+            vec_xst(v0, 0, curr_dst + d);
+            vec_xst(v1, 0, curr_dst + d + 4);
+          }
+        }
+
+        for (; d < head_dim; ++d) {
+          float val = static_cast<float>(curr_src[d]);
+          curr_dst[d] = val * scale;
+        }
+      }
+    }
+  }
+
+  static void reshape_and_cache(
+      const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
+      scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
+      const int64_t* __restrict__ slot_mapping, const int64_t token_num,
+      const int64_t key_token_num_stride, const int64_t value_token_num_stride,
+      const int64_t head_num, const int64_t key_head_num_stride,
+      const int64_t value_head_num_stride, const int64_t num_blocks,
+      const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
+      const int64_t block_size, const int64_t block_size_stride) {
+#pragma omp parallel for collapse(2)
+    for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
+      for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
+        const int64_t pos = slot_mapping[token_idx];
+        if (pos < 0) continue;
+
+        const int64_t block_idx = pos / block_size;
+        const int64_t block_offset = pos % block_size;
+
+        {
+          const scalar_t* key_src = key + token_idx * key_token_num_stride +
+                                    head_idx * key_head_num_stride;
+          scalar_t* key_dst = key_cache + block_idx * num_blocks_stride +
+                              head_idx * cache_head_num_stride + block_offset;
+
+          for (int64_t i = 0, j = 0; i < head_dim; ++i, j += block_size) {
+            key_dst[j] = key_src[i];
+          }
+        }
+
+        {
+          const scalar_t* val_src = value + token_idx * value_token_num_stride +
+                                    head_idx * value_head_num_stride;
+          scalar_t* val_dst = value_cache + block_idx * num_blocks_stride +
+                              head_idx * cache_head_num_stride +
+                              block_offset * head_dim;
+
+          std::memcpy(val_dst, val_src, sizeof(scalar_t) * head_dim);
+        }
+      }
+    }
+  }
+};
+
+}  // namespace cpu_attention
+
+#undef BLOCK_SIZE_ALIGNMENT
+#undef HEAD_SIZE_ALIGNMENT
+#undef MAX_Q_HEAD_NUM_PER_ITER
+
+#endif

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{
+                     MSizeCacheKey{// Use a concrete M so oneDNN's kernel
-#ifdef VLLM_USE_ACL
+                                   // selector can choose an optimally blocked
-                         // Arm Compute Library (ACL) backend for oneDNN does
+                                   // weight layout.
-                         // not support runtime
+                                   .a_m_size = kProbeM,
-                         // dimensions, so we set M to a default value
+                                   .a_m_stride = b_k_size_,
-                         .a_m_size = 128,
+                                   .use_bias = false,
-                         .a_m_stride = b_k_size_,
+                                   .bias_type = dnnl::memory::data_type::undef},
-#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},
                      true)
                      .weights_desc());
   init_runtime_memory_cache(args);

csrc/cpu/generate_cpu_attn_dispatch.py

@@ -19,10 +19,11 @@ ISA_TYPES = {
     "VEC": 1,
     "VEC16": 2,
     "NEON": 3,
+    "VXE": 4,
 }
 
 # ISAs supported for head_dims divisible by 32
-ISA_FOR_32 = ["AMX", "NEON", "VEC", "VEC16"]
+ISA_FOR_32 = ["AMX", "NEON", "VEC", "VEC16", "VXE"]
 
 # ISAs supported for head_dims divisible by 16 only
 ISA_FOR_16 = ["VEC16"]
@@ -118,6 +119,10 @@ def generate_header_file() -> str:
   #include "cpu_attn_neon.hpp"
 #endif
 
+#ifdef __s390x__
+  #include "cpu_attn_vxe.hpp"
+#endif
+
 """
 
     header += generate_helper_function()
@@ -163,6 +168,25 @@ def generate_header_file() -> str:
     } \\
   }()
 
+"""
+
+    # s390x with VXE
+    header += """#elif defined(__s390x__)
+#define CPU_ATTN_DISPATCH(HEAD_DIM, ISA_TYPE, ...) \\
+  [&] { \\
+    int64_t encoded_params = encode_cpu_attn_params(HEAD_DIM, ISA_TYPE); \\
+    switch (encoded_params) { \\
+"""
+    header += generate_cases_for_isa_group(["VXE", "VEC", "VEC16"])
+    header += """
+      default: { \\
+        TORCH_CHECK(false, "Unsupported CPU attention configuration: head_dim=" + \\
+                    std::to_string(HEAD_DIM) + " isa=" + \\
+                    std::to_string(static_cast<int>(ISA_TYPE))); \\
+      } \\
+    } \\
+  }()
+
 """
 
     # Fallback: VEC and VEC16 only
@@ -182,7 +206,7 @@ def generate_header_file() -> str:
     } \\
   }()
 
-#endif  /* CPU_CAPABILITY_AMXBF16 / __aarch64__ */
+#endif  /* CPU_CAPABILITY_AMXBF16 / __aarch64__ / __s390x__ */
 
 #endif  // CPU_ATTN_DISPATCH_GENERATED_H
 """

csrc/cpu/torch_bindings.cpp

@@ -4,6 +4,10 @@
 
 #include <torch/library.h>
 
+// Note: overwrite the external definition for sharing same name between
+// libraries use different ISAs.
+#define TORCH_EXTENSION_NAME _C
+
 std::string init_cpu_threads_env(const std::string& cpu_ids);
 
 void release_dnnl_matmul_handler(int64_t handler);
@@ -324,19 +328,12 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "str act, str isa) -> ()");
   ops.impl("cpu_fused_moe", torch::kCPU, &cpu_fused_moe);
 #endif
-}
+  ops.def("init_cpu_threads_env(str cpu_ids) -> str", &init_cpu_threads_env);
-
+  ops.def(
-TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _utils), utils) {
-  // CPU utils
-  utils.def("init_cpu_threads_env(str cpu_ids) -> str", &init_cpu_threads_env);
-}
-
-TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cpu), cpu_ops) {
-  cpu_ops.def(
       "mla_decode_kvcache("
       "   Tensor! out, Tensor query, Tensor kv_cache,"
       "   float scale, Tensor block_tables, Tensor seq_lens) -> ()");
-  cpu_ops.impl("mla_decode_kvcache", torch::kCPU, &mla_decode_kvcache);
+  ops.impl("mla_decode_kvcache", torch::kCPU, &mla_decode_kvcache);
 }
 
 REGISTER_EXTENSION(TORCH_EXTENSION_NAME)

csrc/cuda_vec_utils.cuh

@@ -0,0 +1,334 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+#pragma once
+
+#include <c10/util/BFloat16.h>
+#include <c10/util/Half.h>
+#include <cassert>
+
+#ifdef USE_ROCM
+  #include <hip/hip_runtime.h>
+#else
+  #include <cuda_bf16.h>
+  #include <cuda_fp16.h>
+  #include <cuda_runtime.h>
+#endif
+
+// Device-side: SM100+ architecture with CUDA 12.9+ toolkit, which
+// together enable 256-bit (v8.u32) PTX load/store instructions.
+// Use for PTX instruction selection with architecture fallback paths.
+#if !defined(USE_ROCM) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000 && \
+    defined(CUDA_VERSION) && CUDA_VERSION >= 12090
+  #define VLLM_256B_PTX_ENABLED 1
+#else
+  #define VLLM_256B_PTX_ENABLED 0
+#endif
+
+namespace vllm {
+
+// ============================================================
+// Types and traits
+// ============================================================
+
+// 256-bit (32-byte) aligned vector type: 8 x uint32_t
+struct alignas(32) u32x8_t {
+  uint32_t d[8];
+};
+
+// VecTraits — select between 128-bit (int4) and 256-bit
+// (u32x8_t) vector types at compile time.
+template <bool support_256>
+struct VecTraits;
+
+template <>
+struct VecTraits<true> {
+  static constexpr int ARCH_MAX_VEC_SIZE = 32;
+  using vec_t = u32x8_t;
+};
+
+template <>
+struct VecTraits<false> {
+  static constexpr int ARCH_MAX_VEC_SIZE = 16;
+  using vec_t = int4;
+};
+
+// PackedTypeConverter — map between CUDA scalar and packed types
+//   half  <-> half2,  __nv_bfloat16 <-> __nv_bfloat162, etc.
+template <typename T>
+struct PackedTypeConverter {
+  static_assert(sizeof(T) == 0,
+                "PackedTypeConverter is not specialized for this type.");
+};
+
+template <>
+struct PackedTypeConverter<half2> {
+  using Type = half;
+};
+
+template <>
+struct PackedTypeConverter<half> {
+  using Type = half2;
+};
+
+template <>
+struct PackedTypeConverter<__nv_bfloat162> {
+  using Type = __nv_bfloat16;
+};
+
+template <>
+struct PackedTypeConverter<__nv_bfloat16> {
+  using Type = __nv_bfloat162;
+};
+
+template <>
+struct PackedTypeConverter<float> {
+  using Type = float2;
+};
+
+template <>
+struct PackedTypeConverter<float2> {
+  using Type = float;
+};
+
+template <>
+struct PackedTypeConverter<c10::Half> {
+  using Type = half2;
+};
+
+template <>
+struct PackedTypeConverter<c10::BFloat16> {
+  using Type = __nv_bfloat162;
+};
+
+// CUDATypeConverter — map PyTorch scalar types to CUDA scalar
+//   c10::Half -> half,  c10::BFloat16 -> __nv_bfloat16
+template <typename T>
+struct CUDATypeConverter {
+  using Type = T;
+};
+
+template <>
+struct CUDATypeConverter<c10::Half> {
+  using Type = half;
+};
+
+template <>
+struct CUDATypeConverter<c10::BFloat16> {
+  using Type = __nv_bfloat16;
+};
+
+// PackedVec — typed vector container for packed element access.
+//   Derives alignment and element count from VecTraits.
+//   Type is the CUDA scalar type (e.g. half, __nv_bfloat16).
+template <class Type, bool use_256b>
+struct alignas(VecTraits<use_256b>::ARCH_MAX_VEC_SIZE) PackedVec {
+  static constexpr int NUM_ELTS =
+      VecTraits<use_256b>::ARCH_MAX_VEC_SIZE /
+      sizeof(typename PackedTypeConverter<Type>::Type);
+  typename PackedTypeConverter<Type>::Type elts[NUM_ELTS];
+};
+
+// ============================================================
+// Load / store primitives
+// ============================================================
+
+// 256-bit load / store — SM100+ only (PTX v8 instructions).
+__device__ __forceinline__ void ld256(u32x8_t& val, const u32x8_t* ptr) {
+#if VLLM_256B_PTX_ENABLED
+  asm volatile("ld.global.nc.v8.u32 {%0,%1,%2,%3,%4,%5,%6,%7}, [%8];\n"
+               : "=r"(val.d[0]), "=r"(val.d[1]), "=r"(val.d[2]), "=r"(val.d[3]),
+                 "=r"(val.d[4]), "=r"(val.d[5]), "=r"(val.d[6]), "=r"(val.d[7])
+               : "l"(ptr));
+#else
+  assert(false && "ld256 requires SM100+ with CUDA 12.9+");
+#endif
+}
+
+__device__ __forceinline__ void st256(u32x8_t& val, u32x8_t* ptr) {
+#if VLLM_256B_PTX_ENABLED
+  asm volatile("st.global.v8.u32 [%0], {%1,%2,%3,%4,%5,%6,%7,%8};\n"
+               :
+               : "l"(ptr), "r"(val.d[0]), "r"(val.d[1]), "r"(val.d[2]),
+                 "r"(val.d[3]), "r"(val.d[4]), "r"(val.d[5]), "r"(val.d[6]),
+                 "r"(val.d[7])
+               : "memory");
+#else
+  assert(false && "st256 requires SM100+ with CUDA 12.9+");
+#endif
+}
+
+// Generic ld256 / st256 for any 32-byte aligned type (e.g. PackedVec).
+// Non-template overloads above are preferred for u32x8_t.
+template <typename T>
+__device__ __forceinline__ void ld256(T& val, const T* ptr) {
+  static_assert(sizeof(T) == 32, "ld256 requires a 32-byte type");
+  ld256(reinterpret_cast<u32x8_t&>(val), reinterpret_cast<const u32x8_t*>(ptr));
+}
+
+template <typename T>
+__device__ __forceinline__ void st256(T& val, T* ptr) {
+  static_assert(sizeof(T) == 32, "st256 requires a 32-byte type");
+  st256(reinterpret_cast<u32x8_t&>(val), reinterpret_cast<u32x8_t*>(ptr));
+}
+
+// 128-bit load / store via __ldg (read-only cache hint).
+template <typename T>
+__device__ __forceinline__ void ld128(T& val, const T* ptr) {
+  static_assert(sizeof(T) == 16, "ld128 requires a 16-byte type");
+  *reinterpret_cast<int4*>(&val) = __ldg(reinterpret_cast<const int4*>(ptr));
+}
+
+template <typename T>
+__device__ __forceinline__ void st128(T& val, T* ptr) {
+  static_assert(sizeof(T) == 16, "st128 requires a 16-byte type");
+  *reinterpret_cast<int4*>(ptr) = *reinterpret_cast<int4*>(&val);
+}
+
+// 256-bit cache-streaming (.cs) load / store  — SM100+ only.
+__forceinline__ __device__ u32x8_t ld256_cs(const u32x8_t* addr) {
+#if VLLM_256B_PTX_ENABLED
+  u32x8_t val;
+  asm volatile("ld.global.cs.v8.u32 {%0,%1,%2,%3,%4,%5,%6,%7}, [%8];"
+               : "=r"(val.d[0]), "=r"(val.d[1]), "=r"(val.d[2]), "=r"(val.d[3]),
+                 "=r"(val.d[4]), "=r"(val.d[5]), "=r"(val.d[6]), "=r"(val.d[7])
+               : "l"(addr));
+  return val;
+#else
+  assert(false && "ld256_cs requires SM100+ with CUDA 12.9+");
+  return {};
+#endif
+}
+
+__forceinline__ __device__ void st256_cs(u32x8_t* addr, u32x8_t val) {
+#if VLLM_256B_PTX_ENABLED
+  asm volatile(
+      "st.global.cs.v8.u32 [%0], {%1,%2,%3,%4,%5,%6,%7,%8};" ::"l"(addr),
+      "r"(val.d[0]), "r"(val.d[1]), "r"(val.d[2]), "r"(val.d[3]), "r"(val.d[4]),
+      "r"(val.d[5]), "r"(val.d[6]), "r"(val.d[7]));
+#else
+  assert(false && "st256_cs requires SM100+ with CUDA 12.9+");
+#endif
+}
+
+// 32-bit cache-streaming (.cs) load / store  — SM100+ only.
+__forceinline__ __device__ int ld32_cs(const int* addr) {
+#if VLLM_256B_PTX_ENABLED
+  int val;
+  asm volatile("ld.global.cs.b32 %0, [%1];" : "=r"(val) : "l"(addr));
+  return val;
+#else
+  assert(false && "ld32_cs requires SM100+ with CUDA 12.9+");
+  return 0;
+#endif
+}
+
+__forceinline__ __device__ void st32_cs(int* addr, int val) {
+#if VLLM_256B_PTX_ENABLED
+  asm volatile("st.global.cs.b32 [%0], %1;" ::"l"(addr), "r"(val));
+#else
+  assert(false && "st32_cs requires SM100+ with CUDA 12.9+");
+#endif
+}
+
+// Predicated 256-bit / 128-bit cache-global (.cg) loads.
+// Returns zero if pred is false.  SM100+ only.
+__device__ __forceinline__ void ld256_cg_or_zero(u32x8_t& val, const void* ptr,
+                                                 bool pred) {
+#if VLLM_256B_PTX_ENABLED
+  asm volatile(
+      "{\n"
+      "  .reg .pred pr;\n"
+      "  setp.ne.u32 pr, %8, 0;\n"
+      "  mov.u32 %0, 0;\n"
+      "  mov.u32 %1, 0;\n"
+      "  mov.u32 %2, 0;\n"
+      "  mov.u32 %3, 0;\n"
+      "  mov.u32 %4, 0;\n"
+      "  mov.u32 %5, 0;\n"
+      "  mov.u32 %6, 0;\n"
+      "  mov.u32 %7, 0;\n"
+      "  @pr ld.global.cg.v8.u32 {%0,%1,%2,%3,%4,%5,%6,%7}, [%9];\n"
+      "}\n"
+      : "=r"(val.d[0]), "=r"(val.d[1]), "=r"(val.d[2]), "=r"(val.d[3]),
+        "=r"(val.d[4]), "=r"(val.d[5]), "=r"(val.d[6]), "=r"(val.d[7])
+      : "r"((int)pred), "l"(ptr));
+#else
+  assert(false && "ld256_cg_or_zero requires SM100+ with CUDA 12.9+");
+#endif
+}
+
+__device__ __forceinline__ void ld128_cg_or_zero(uint4& val, const void* ptr,
+                                                 bool pred) {
+#if VLLM_256B_PTX_ENABLED
+  uint32_t r0, r1, r2, r3;
+
+  asm volatile(
+      "{\n"
+      "  .reg .pred pr;\n"
+      "  setp.ne.u32 pr, %4, 0;\n"
+      "  mov.u32 %0, 0;\n"
+      "  mov.u32 %1, 0;\n"
+      "  mov.u32 %2, 0;\n"
+      "  mov.u32 %3, 0;\n"
+      "  @pr ld.global.cg.v4.u32 {%0,%1,%2,%3}, [%5];\n"
+      "}\n"
+      : "=r"(r0), "=r"(r1), "=r"(r2), "=r"(r3)
+      : "r"((int)pred), "l"(ptr));
+
+  val = uint4{r0, r1, r2, r3};
+#else
+  assert(false && "ld128_cg_or_zero requires SM100+ with CUDA 12.9+");
+#endif
+}
+
+// ============================================================
+// Alignment helpers
+// ============================================================
+
+__host__ __device__ __forceinline__ bool is_16byte_aligned(const void* ptr) {
+  return (reinterpret_cast<uintptr_t>(ptr) & 15) == 0;
+}
+
+__host__ __device__ __forceinline__ bool is_32byte_aligned(const void* ptr) {
+  return (reinterpret_cast<uintptr_t>(ptr) & 31) == 0;
+}
+
+// ============================================================
+// Packed type conversion and arithmetic
+// ============================================================
+
+template <typename packed_t>
+__device__ __forceinline__ float2 cast_to_float2(const packed_t& val) {
+  if constexpr (std::is_same_v<packed_t, __nv_bfloat162>) {
+    return __bfloat1622float2(val);
+  } else if constexpr (std::is_same_v<packed_t, __half2>) {
+    return __half22float2(val);
+  } else if constexpr (std::is_same_v<packed_t, float2>) {
+    return float2(val);
+  }
+}
+
+template <typename packed_t>
+__device__ __forceinline__ packed_t cast_to_packed(const float2& val) {
+  if constexpr (std::is_same_v<packed_t, __nv_bfloat162>) {
+    return __float22bfloat162_rn(val);
+  } else if constexpr (std::is_same_v<packed_t, __half2>) {
+    return __float22half2_rn(val);
+  } else if constexpr (std::is_same_v<packed_t, float2>) {
+    return float2(val);
+  }
+}
+
+template <typename packed_t>
+__device__ __forceinline__ packed_t packed_mul(const packed_t& x,
+                                               const packed_t& y) {
+  if constexpr (std::is_same_v<packed_t, __nv_bfloat162> ||
+                std::is_same_v<packed_t, __half2>) {
+    return __hmul2(x, y);
+  } else if constexpr (std::is_same_v<packed_t, float2>) {
+    return make_float2(x.x * y.x, x.y * y.y);
+  }
+}
+
+}  // namespace vllm