2.24.0

Signed-off-by: khluu <khluu000@gmail.com>

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

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

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

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

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

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

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

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

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

@@ -0,0 +1,205 @@
+#!/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.
+if [ -n "$RAY_COMPAT_SLACK_WEBHOOK_URL" ]; then
+    echo ">>> Sending Slack notification"
+    # Single quotes are intentional: the f-string expressions are Python, not shell.
+    # shellcheck disable=SC2016
+    PAYLOAD=$(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"
+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"
-while true; do
-        sleep 3
-        if grep -q clean /opt/amdgpu/etc/gpu_state; then
-                echo "GPUs state is \"clean\""
-                break
-        fi
-done
+###############################################################################
+# Helper Functions
+###############################################################################
 
-echo "--- ROCm info"
-rocminfo
+wait_for_clean_gpus() {
+  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 $((NUM_NODES-1))); do
-    if docker pr -a -q -f name="node${node}" | grep -q .; then
-      docker stop "node${node}"
+  local max_nodes=${NUM_NODES:-2}
+  for node in $(seq 0 $((max_nodes - 1))); do
+    if docker ps -a -q -f name="node${node}" | grep -q .; then
+      docker stop "node${node}" || true
     fi
   done
-  if docker network ls | grep docker-net; then
-    docker network rm docker-net
+  if docker network ls | grep -q docker-net; then
+    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
-        sleep 3
-        if grep -q clean /opt/amdgpu/etc/gpu_state; then
-                echo "GPUs state is \"clean\""
-                break
-        fi
-done
-
+# --- Pull test image ---
 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 [[ $commands == *"pytest -v -s models/test_registry.py"* ]]; then
-  commands=${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')
+commands=$(apply_rocm_test_overrides "$commands")
 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
-# proper access to GPUs
+# Verify GPU access
 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
-      prefix=$( echo "${BASH_REMATCH[1]}" | sed 's/;//g')
-      echo "PREFIX: ${prefix}"
-      export composite_command="(command rocm-smi || true)"
-      myIFS=$IFS
-      IFS=','
-      read -ra node0 <<< ${BASH_REMATCH[2]}
-      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}'"
-        echo "COMMANDS: ${commands}"
-        composite_command=$(echo "${composite_command} && ${commands}")
-      done
-      /bin/bash -c "${composite_command}"
-      cleanup_network
+  # Parse the bracket syntax:  prefix ; [node0_cmds] && [node1_cmds]
+  #   BASH_REMATCH[1] = prefix (everything before first bracket)
+  #   BASH_REMATCH[2] = comma-separated node0 commands
+  #   BASH_REMATCH[3] = comma-separated node1 commands
+  if [[ "$commands" =~ ^(.*)\[(.*)"] && ["(.*)\]$ ]]; then
+    prefix=$(echo "${BASH_REMATCH[1]}" | sed 's/;//g')
+    echo "PREFIX: ${prefix}"
+
+    export composite_command="(command rocm-smi || true)"
+    saved_IFS=$IFS
+    IFS=','
+    read -ra node0 <<< "${BASH_REMATCH[2]}"
+    read -ra node1 <<< "${BASH_REMATCH[3]}"
+    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."
-      cleanup_network
-      exit 111
+    echo "Multi-node job detected but failed to parse bracket command syntax."
+    echo "Expected format: prefix ; [node0_cmd1, node0_cmd2] && [node1_cmd1, node1_cmd2]"
+    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 $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
-          --network=host \
-          --shm-size=16gb \
-          --group-add "$render_gid" \
-          --rm \
-          -e HF_TOKEN \
-          -e AWS_ACCESS_KEY_ID \
-          -e AWS_SECRET_ACCESS_KEY \
-          -v "${HF_CACHE}:${HF_MOUNT}" \
-          -e "HF_HOME=${HF_MOUNT}" \
-          -e "PYTHONPATH=${MYPYTHONPATH}" \
-          --name "${container_name}" \
-          "${image_name}" \
-          /bin/bash -c "${commands}"
+    --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+    $RDMA_FLAGS \
+    --network=host \
+    --shm-size=16gb \
+    --group-add "$render_gid" \
+    --rm \
+    -e HF_TOKEN \
+    -e AWS_ACCESS_KEY_ID \
+    -e AWS_SECRET_ACCESS_KEY \
+    -v "${HF_CACHE}:${HF_MOUNT}" \
+    -e "HF_HOME=${HF_MOUNT}" \
+    -e "PYTHONPATH=${MYPYTHONPATH}" \
+    --name "${container_name}" \
+    "${image_name}" \
+    /bin/bash -c "${commands}"
+
+  exit_code=$?
+  handle_pytest_exit "$exit_code"
 fi

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

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

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

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

.buildkite/scripts/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/test-amd.yaml

@@ -156,8 +156,9 @@ steps:
 
 - label: Entrypoints Integration Test (API Server 1) # 100min
   timeout_in_minutes: 130
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   fast_check: true
@@ -173,8 +174,9 @@ steps:
 
 - label: Entrypoints Integration Test (API Server 2)
   timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   fast_check: true
@@ -192,8 +194,9 @@ steps:
 
 - label: Entrypoints Integration Test (Pooling)
   timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   fast_check: true
@@ -207,8 +210,9 @@ steps:
 
 - label: Entrypoints Integration Test (Responses API)
   timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   fast_check: true
@@ -222,8 +226,9 @@ steps:
 
 - label: Distributed Tests (4 GPUs) # 35min
   timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -278,14 +283,16 @@ steps:
   - popd
   # NEW rlhf examples
   - pushd ../examples/offline_inference/new_weight_syncing
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_nccl.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_ipc.py
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_async_new_apis.py
   - popd
 
 - label: Distributed Tests (8 GPUs) # 4min
   timeout_in_minutes: 10
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_8
+  optional: true
   # grade: Blocking
   gpu: h100
   num_gpus: 8
@@ -380,10 +387,9 @@ steps:
 
 - label: V1 Test e2e + engine # 65min
   timeout_in_minutes: 90
-  mirror_hardwares: [amdexperimental]
-  # The test uses 4 GPUs, but we schedule it on 8-GPU machines for stability.
-  # See discussion here: https://github.com/vllm-project/vllm/pull/31040
-  agent_pool: mi325_8
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   source_file_dependencies:
     - vllm/
@@ -394,6 +400,34 @@ steps:
     - pytest -v -s v1/e2e
     - pytest -v -s v1/engine
 
+- label: V1 Test e2e (2 GPUs) # 65min
+  timeout_in_minutes: 90
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_2
+  optional: true
+  # grade: Blocking
+  source_file_dependencies:
+    - vllm/
+    - tests/v1
+  commands:
+    # Only run tests that need exactly 2 GPUs
+    - pytest -v -s v1/e2e/test_spec_decode.py -k "tensor_parallelism"
+
+- label: V1 Test e2e (4 GPUs) # 65min
+  timeout_in_minutes: 90
+  mirror_hardwares: [amdexperimental, amdproduction]
+  # The test uses 4 GPUs, but we schedule it on 8-GPU machines for stability.
+  # See discussion here: https://github.com/vllm-project/vllm/pull/31040
+  agent_pool: mi325_4
+  optional: true
+  # grade: Blocking
+  source_file_dependencies:
+    - vllm/
+    - tests/v1
+  commands:
+    # Only run tests that need 4 GPUs
+    - pytest -v -s v1/e2e/test_spec_decode.py -k "eagle_correctness_heavy"
+
 - label: V1 Test entrypoints # 35min
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
@@ -407,8 +441,9 @@ steps:
 
 - label: V1 Test others # 42min
   timeout_in_minutes: 60
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   source_file_dependencies:
     - vllm/
@@ -435,8 +470,9 @@ steps:
 # TODO: Add the "V1 Test attetion (MI300)" test group
 
 - label: V1 Test attention (H100) # 10min
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   timeout_in_minutes: 30
   gpu: h100
@@ -540,8 +576,9 @@ steps:
 
 - label: Samplers Test # 56min
   timeout_in_minutes: 75
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   source_file_dependencies:
   - vllm/model_executor/layers
@@ -553,8 +590,9 @@ steps:
 
 - label: LoRA Test %N # 20min each
   timeout_in_minutes: 30
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   source_file_dependencies:
   - vllm/lora
@@ -572,6 +610,8 @@ steps:
       --ignore=lora/test_qwen3moe_tp.py
   parallelism: 4
 
+##### .buildkite/test_areas/pytorch.yaml #####
+# corresponds to .buildkite/test_areas/pytorch.yaml
 - label: PyTorch Compilation Unit Tests # 15min
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -589,6 +629,20 @@ steps:
   # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
   - "find compile/ -maxdepth 1 -name 'test_*.py' -exec pytest -s -v {} \\\\;"
 
+# corresponds to .buildkite/test_areas/pytorch.yaml
+- label: PyTorch Compilation Passes Unit Tests
+  timeout_in_minutes: 20
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_1
+  source_file_dependencies:
+    - vllm/
+    - tests/compile/passes
+  commands:
+  # TODO: clean up this comment if not needed. It is used to 
+  # keep track of the tests changes during vLLM IR Ops refactoring.
+  # Use `find` to launch multiple instances of pytest.
+  - "find compile/passes -maxdepth 1 -name 'test_*.py' -exec pytest -s -v {} \\\\;"
+
 - label: PyTorch Fullgraph Smoke Test # 15min
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -664,8 +718,9 @@ steps:
 
 - label: Kernels Quantization Test %N # 64min
   timeout_in_minutes: 90
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   source_file_dependencies:
   - csrc/quantization/
@@ -798,8 +853,9 @@ steps:
 
 - label: LM Eval Small Models # 53min
   timeout_in_minutes: 75
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   source_file_dependencies:
   - csrc/
@@ -860,8 +916,9 @@ steps:
 
 - label: Basic Models Tests (Other)
   timeout_in_minutes: 45
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
@@ -902,8 +959,9 @@ steps:
 
 - label: Language Models Tests (Extra Standard) %N
   timeout_in_minutes: 45
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
@@ -923,8 +981,9 @@ steps:
 
 - label: Language Models Tests (Hybrid) %N
   timeout_in_minutes: 75
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
@@ -944,7 +1003,7 @@ steps:
 
 - label: Language Models Test (Extended Generation) # 80min
   timeout_in_minutes: 110
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   optional: true
@@ -960,7 +1019,7 @@ steps:
 
 - label: Language Models Test (PPL)
   timeout_in_minutes: 110
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   optional: true
@@ -972,7 +1031,7 @@ steps:
 
 - label: Language Models Test (Extended Pooling)  # 36min
   timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   optional: true
@@ -984,7 +1043,7 @@ steps:
 
 - label: Language Models Test (MTEB)
   timeout_in_minutes: 110
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   optional: true
@@ -996,11 +1055,12 @@ steps:
 
 - label: Multi-Modal Processor Test (CPU)
   timeout_in_minutes: 60
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
+  - tests/models/registry.py
   no_gpu: true
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
@@ -1008,19 +1068,20 @@ steps:
 
 - label: Multi-Modal Processor Test # 44min
   timeout_in_minutes: 60
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   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
 
 - label: Multi-Modal Models Test (Standard) # 60min
   timeout_in_minutes: 100
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   torch_nightly: true
@@ -1053,7 +1114,7 @@ steps:
 
 - label: Multi-Modal Models Test (Extended) 1 # 60min
   timeout_in_minutes: 120
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   optional: true
@@ -1068,7 +1129,7 @@ steps:
 
 - label: Multi-Modal Models Test (Extended) 2 #60min
   timeout_in_minutes: 120
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   optional: true
@@ -1083,7 +1144,7 @@ steps:
 
 - label: Multi-Modal Models Test (Extended) 3 # 75min
   timeout_in_minutes: 150
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   optional: true
@@ -1108,7 +1169,7 @@ steps:
     - pytest -v -s models/quantization
 
 - label: Transformers Nightly Models Test
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   working_dir: "/vllm-workspace/"
@@ -1166,41 +1227,6 @@ steps:
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
     - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
 
-- label: Blackwell Fusion and Compile Tests # 30 min
-  timeout_in_minutes: 40
-  working_dir: "/vllm-workspace/"
-  gpu: b200
-  source_file_dependencies:
-  - csrc/quantization/fp4/
-  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-  - vllm/v1/attention/backends/flashinfer.py
-  - vllm/v1/worker/
-  - vllm/v1/cudagraph_dispatcher.py
-  - vllm/compilation/
-  # can affect pattern matching
-  - vllm/model_executor/layers/layernorm.py
-  - vllm/model_executor/layers/activation.py
-  - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/passes/test_fusion_attn.py
-  - tests/compile/passes/test_silu_mul_quant_fusion.py
-  - tests/compile/passes/distributed/test_fusion_all_reduce.py
-  - tests/compile/fullgraph/test_full_graph.py
-  commands:
-    - nvidia-smi
-    - pytest -v -s tests/compile/passes/test_fusion_attn.py
-    - pytest -v -s tests/compile/passes/test_silu_mul_quant_fusion.py
-    # this runner has 2 GPUs available even though num_gpus=2 is not set
-    - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
-
-    # # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
-    # # Wrap with quotes to escape yaml
-    # - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
-    # Old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293
-    # in favor of new tests in fusions_e2e. We avoid replicating the new jobs in this file as it's deprecated.
-
-    # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
-    - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
-
 - label: Blackwell GPT-OSS Eval
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/"
@@ -1263,8 +1289,9 @@ steps:
 
 - label: 2 Node Tests (4 GPUs in total) # 16min
   timeout_in_minutes: 30
-  mirror_hardwares: [amdexperimental, amdmultinode]
+  mirror_hardwares: [amdexperimental, amdproduction, amdmultinode]
   agent_pool: mi325_4
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
@@ -1290,8 +1317,9 @@ steps:
 
 - label: Distributed Tests (2 GPUs) # 68min
   timeout_in_minutes: 90
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_2
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
@@ -1311,6 +1339,7 @@ steps:
   - tests/v1/entrypoints/openai/test_multi_api_servers.py
   - tests/v1/shutdown
   - tests/v1/worker/test_worker_memory_snapshot.py
+  - examples/offline_inference/new_weight_syncing/
   commands:
   # Work around HIP bug tracked here: https://github.com/ROCm/hip/issues/3876
   # TODO: Remove when the bug is fixed in a future ROCm release
@@ -1324,14 +1353,14 @@ steps:
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
   - VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
-  - pytest -v -s compile/correctness_e2e/test_sequence_parallel.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
   - pytest -v -s v1/worker/test_worker_memory_snapshot.py
 
 - label: Distributed Model Tests (2 GPUs) # 37min
   timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_2
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
@@ -1370,6 +1399,10 @@ steps:
   - pip install -e ./plugins/prithvi_io_processor_plugin
   - pytest -v -s plugins_tests/test_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
@@ -1441,7 +1474,7 @@ steps:
     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-amd.txt
 
 - label: Weight Loading Multiple GPU Test - Large Models # optional
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_2
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
@@ -1485,7 +1518,7 @@ steps:
 ##### A100 test #####
 
 - label: Distributed Tests (A100) # optional
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
   # grade: Blocking
   gpu: a100
@@ -1508,7 +1541,7 @@ steps:
 - label: LM Eval Large Models # optional
   gpu: a100
   optional: true
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
   # grade: Blocking
   num_gpus: 4
@@ -1520,11 +1553,11 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
-##### H100 test #####
-- label: LM Eval Large Models (H100) # optional
+##### FP8 test #####
+- label: LM Eval Large Models (H100) # optional, still use H100 for consistency
   gpu: h100
   optional: true
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
   # grade: Blocking
   num_gpus: 4
@@ -1533,13 +1566,13 @@ steps:
   - csrc/
   - vllm/model_executor/layers/quantization
   commands:
-    - export VLLM_USE_DEEP_GEMM=0  # We found Triton is faster than DeepGEMM for H100
-    - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4
+    - export VLLM_USE_DEEP_GEMM=0 
+    - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm.txt --tp-size=4
 
 
 ##### H200 test #####
 - label: Distributed Tests (H200) # optional
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_2
   # grade: Blocking
   gpu: h200
@@ -1549,16 +1582,16 @@ steps:
   commands:
     - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/passes/distributed/test_async_tp.py
     - pytest -v -s tests/compile/passes/distributed/test_sequence_parallelism.py
-    - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
+    # TODO: this test is not supported on ROCm, there are aiter kernels for this.
+    # - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
     #- pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
     # - "VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'"
     # Old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293
     # in favor of new tests in fusions_e2e. We avoid replicating the new jobs in this file as it's deprecated.
-
-    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/correctness_e2e/test_sequence_parallel.py
     - pytest -v -s tests/distributed/test_context_parallel.py
     - HIP_VISIBLE_DEVICES=0,1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=allgather_reducescatter --disable-nccl-for-dp-synchronization
-    - pytest -v -s tests/v1/distributed/test_dbo.py
+    # this test is not supported on ROCm
+    # - pytest -v -s tests/v1/distributed/test_dbo.py
 
 ##### B200 test #####
 - label: Distributed Tests (B200) # optional
@@ -1599,8 +1632,9 @@ steps:
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
 - label: ROCm LM Eval Large Models (8 Card)
-  mirror_hardwares: [amdproduction]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_8
+  optional: true
   num_gpus: 8
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   commands:
@@ -1659,7 +1693,7 @@ steps:
 
 - label: Qwen3-Next-80B-A3B-Instruct MTP Async EPLB Accuracy
   timeout_in_minutes: 60
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
   # grade: Blocking
   optional: true
@@ -1668,6 +1702,93 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh 0.8 1319 8040
 
+##### .buildkite/test_areas/compile.yaml #####
+# Slowly setting up the tests so that it is also easier for the 
+# CI team to review and upstream to the pipelinev2.
+# The following tests are important for vLLM IR Ops refactoring,
+# which affects fusion passes on ROCm. So we have to 
+# enable them as as soon as possible.
+
+## TODO: Enable the test in this group
+# # corresponds to .buildkite/test_areas/compile.yaml
+# - label: Fusion and Compile Unit Tests (2xMI325 GPUs)
+#   timeout_in_minutes: 20
+#   working_dir: "/vllm-workspace/"
+#   mirror_hardwares: [amdexperimental, amdproduction, tj]
+#   agent_pool: mi325_1 # changed to 1 GPU until the fusion all reduce is enabled then only revert back to 2 GPUs
+#   source_file_dependencies:
+#   - csrc/quantization/fp4/
+#   - vllm/model_executor/layers/quantization/
+#   - vllm/model_executor/layers/layernorm.py
+#   - vllm/model_executor/layers/activation.py
+#   - vllm/model_executor/layers/attention/attention.py
+#   - vllm/v1/attention/backends/flashinfer.py
+#   - vllm/compilation/ # TODO(luka) limit to vllm/compilation/passes
+#   - tests/compile/test_fusion_attn.py
+#   - tests/compile/test_silu_mul_quant_fusion.py
+#   - tests/compile/distributed/test_fusion_all_reduce.py
+#   - tests/compile/fullgraph/test_full_graph.py
+#   commands:
+#     - rocm-smi
+#     # we run all backend tests on ROCm
+#     # These two tests are covered in "PyTorch Compilation Passes Unit Tests"
+#     # - "pytest -v -s tests/compile/passes/test_fusion_attn.py"
+#     # - "pytest -v -s tests/compile/passes/test_silu_mul_quant_fusion.py"
+#     # TODO: this test is not supported on ROCm, there are aiter kernels for this.
+#     # - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
+#     # TODO: find out more details
+#     # - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
+
+# corresponds to .buildkite/test_areas/compile.yaml
+- label: Fusion E2E Quick (MI325)
+  timeout_in_minutes: 15
+  working_dir: "/vllm-workspace/"
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_1
+  num_devices: 1
+  source_file_dependencies:
+    - csrc/quantization/
+    - vllm/model_executor/
+    - vllm/v1/attention/
+    - vllm/compilation/
+    - tests/compile/fusions_e2e/
+  commands:
+    - rocm-smi
+    # Run all models and attn backends but only Inductor partition and native custom ops
+    - "pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k 'inductor_partition and not +rms_norm and not +quant_fp8'"
+    # Different from CUDA, Qwen requires +rms_norm and +quant_fp8 as rms+quant fusion is only supported on AITER
+    - "pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k 'inductor_partition and +rms_norm and +quant_fp8 and qwen3'"
+
+# corresponds to .buildkite/test_areas/compile.yaml
+- label: Fusion E2E Config Sweep (MI325)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/"
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_1
+  num_devices: 1
+  source_file_dependencies:
+    - csrc/quantization/
+    - vllm/compilation/
+    # can affect pattern matching
+    - vllm/model_executor/layers/layernorm.py
+    - vllm/model_executor/layers/activation.py
+    - vllm/model_executor/layers/attention/attention.py
+    - vllm/model_executor/layers/quantization/input_quant_fp8.py
+    - tests/compile/fusions_e2e/
+  commands:
+    - rocm-smi
+    # Run just llama3 (fp8) for all config combinations
+    - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "llama-3"
+
+## There are no ops on ROCm for these tests.
+## The test still passes but the logs are not useful.
+## fused ops just call torch.ops.symm_mem which 
+## exists in ROCm even though they don't work
+# - label: AsyncTP Correctness Tests  (2xMI325 GPUs)
+# - label: Fusion E2E TP2 Quick (MI325)
+# - label: Fusion E2E TP2 AsyncTP Config Sweep (MI325)
+# - label: Fusion E2E TP2 (MI325)
+# - label: Sequence Parallel Correctness Tests (2xMI325 GPUs)
 
 
 #####################################################################################################################################
@@ -1850,8 +1971,10 @@ steps:
 
 - label: Distributed Tests (4 GPUs) # 35min
   timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
+  optional: true
+  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -1905,7 +2028,8 @@ steps:
   - popd
   # NEW rlhf examples
   - pushd ../examples/offline_inference/new_weight_syncing
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_nccl.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_ipc.py
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_async_new_apis.py
   - popd
 
@@ -2869,8 +2993,10 @@ steps:
 
 - label: Distributed Tests (2 GPUs) # 68min
   timeout_in_minutes: 90
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_2
+  optional: true
+  # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -2946,6 +3072,10 @@ steps:
   - pip install -e ./plugins/prithvi_io_processor_plugin
   - pytest -v -s plugins_tests/test_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
@@ -3227,4 +3357,4 @@ steps:
   num_gpus: 4
   working_dir: "/vllm-workspace"
   commands:
-  - bash .buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh 0.8 1319 8040
+  - bash .buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh 0.8 1319 8040

.buildkite/test_areas/distributed.yaml

@@ -103,7 +103,8 @@ 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_ipc.py
 
 - label: Distributed Tests (8 GPUs)(H100)
   timeout_in_minutes: 10
@@ -145,7 +146,7 @@ steps:
   num_devices: 2
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
-    - cd examples/offline_inference/new_weight_syncing && VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_async_new_apis.py
+    # - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py --- failing, need to re-enable
     - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
     - pytest -v -s tests/v1/distributed/test_dbo.py
 

.buildkite/test_areas/e2e_integration.yaml

@@ -28,3 +28,12 @@ steps:
   working_dir: "/vllm-workspace"
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
+
+- label: DeepSeek V2-Lite Prefetch Offload Accuracy (H100)
+  timeout_in_minutes: 60
+  device: h100
+  optional: true
+  num_devices: 1
+  working_dir: "/vllm-workspace"
+  commands:
+  - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_prefetch_offload.sh 0.25 200 8030

.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

@@ -65,6 +65,11 @@ 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

.buildkite/test_areas/expert_parallelism.yaml

@@ -20,4 +20,19 @@ steps:
   - tests/distributed/test_eplb_execute.py
   commands:
   - pytest -v -s distributed/test_eplb_execute.py
-  - pytest -v -s distributed/test_eplb_spec_decode.py
+  - 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
@@ -154,9 +156,7 @@ 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
   device: b200

.buildkite/test_areas/lm_eval.yaml

@@ -11,17 +11,17 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
 
-- label: LM Eval Large Models (4 GPUs)(A100)
-  device: a100
-  optional: true
-  num_devices: 4
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
+# - label: LM Eval Large Models (4 GPUs)(A100)
+#   device: a100
+#   optional: true
+#   num_devices: 4
+#   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+#   source_file_dependencies:
+#   - csrc/
+#   - vllm/model_executor/layers/quantization
+#   commands:
+#   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+#   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
 - label: LM Eval Large Models (4 GPUs)(H100)
   device: h100
@@ -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:
@@ -147,33 +154,6 @@ steps:
   - pytest -v -s transformers_utils
   - pytest -v -s config
 
-- label: GPT-OSS Eval (H100)
-  timeout_in_minutes: 60
-  working_dir: "/vllm-workspace/"
-  device: h100
-  optional: true
-  source_file_dependencies:
-  - tests/evals/gpt_oss
-  - vllm/model_executor/models/gpt_oss.py
-  - vllm/model_executor/layers/quantization/mxfp4.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: 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

@@ -20,6 +20,7 @@ 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 +31,7 @@ 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
@@ -70,12 +72,3 @@ 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

@@ -19,6 +19,10 @@ steps:
   - pip install -e ./plugins/prithvi_io_processor_plugin
   - pytest -v -s plugins_tests/test_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

.buildkite/test_areas/ray_compat.yaml

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

.buildkite/test_areas/weight_loading.yaml

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

.github/.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/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

@@ -259,8 +259,7 @@ 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/openai_chat_completion_structured_outputs_with_reasoning.py
+      - files=examples/online_serving/structured_outputs/structured_outputs.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

.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/*

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 +771,33 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
+  # Expert-specialization MXFP8 blockscaled grouped kernels (SM100+).
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(ES_MXFP8_GROUPED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(ES_MXFP8_GROUPED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+  endif()
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND ES_MXFP8_GROUPED_MM_ARCHS)
+    set(SRCS
+      "csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm.cu"
+      "csrc/moe/mxfp8_moe/mxfp8_experts_quant.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${ES_MXFP8_GROUPED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_ES_MXFP8_GROUPED_MM_SM100=1")
+    message(STATUS "Building ES MXFP8 grouped kernels for archs: ${ES_MXFP8_GROUPED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8
+        AND ES_MXFP8_GROUPED_MM_ARCHS)
+      message(STATUS "Not building ES MXFP8 grouped kernels as CUDA Compiler version is "
+                     "not >= 12.8.")
+    else()
+      message(STATUS "Not building ES MXFP8 grouped kernels as no compatible archs found "
+                     "in CUDA target architectures.")
+    endif()
+  endif()
+
   # DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper and later)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
     cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")
@@ -783,7 +810,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       SRCS "${DSV3_FUSED_A_GEMM_SRC}"
       CUDA_ARCHS "${DSV3_FUSED_A_GEMM_ARCHS}")
     list(APPEND VLLM_EXT_SRC ${DSV3_FUSED_A_GEMM_SRC})
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_DSV3_FUSED_A_GEMM=1")
     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 "
@@ -972,7 +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")

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
@@ -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/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_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/kernels/benchmark_cutlass_moe_fp8.py

@@ -12,12 +12,12 @@ import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from tests.kernels.moe.utils import make_dummy_moe_config
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.fused_moe.activation import MoEActivation
+from vllm.model_executor.layers.fused_moe.all2all_utils import (
+    maybe_make_prepare_finalize,
+)
 from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
 from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassExpertsFp8
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
-from vllm.model_executor.layers.fused_moe.prepare_finalize import (
-    MoEPrepareAndFinalizeNoEP,
-)
 from vllm.platforms import current_platform
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.v1.worker.workspace import init_workspace_manager
@@ -137,15 +137,21 @@ def bench_run(
         per_out_ch_quant=per_out_ch,
     )
 
-    fn = mk.FusedMoEModularKernel(
-        MoEPrepareAndFinalizeNoEP(),
+    moe_config = make_dummy_moe_config(
+        num_experts=num_experts,
+        hidden_dim=k,
+        intermediate_size_per_partition=n,
+        in_dtype=a.dtype,
+    )
+    fn = mk.FusedMoEKernel(
+        maybe_make_prepare_finalize(
+            moe=moe_config,
+            quant_config=quant_config,
+            allow_new_interface=True,
+            use_monolithic=False,
+        ),
         CutlassExpertsFp8(
-            moe_config=make_dummy_moe_config(
-                num_experts=num_experts,
-                hidden_dim=k,
-                intermediate_size_per_partition=n,
-                in_dtype=a.dtype,
-            ),
+            moe_config=moe_config,
             quant_config=quant_config,
         ),
     )

benchmarks/kernels/benchmark_cutlass_moe_nvfp4.py

@@ -15,6 +15,9 @@ import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from tests.kernels.moe.utils import make_dummy_moe_config
 from vllm import _custom_ops as ops
 from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.fused_moe.all2all_utils import (
+    maybe_make_prepare_finalize,
+)
 from vllm.model_executor.layers.fused_moe.config import (
     fp8_w8a8_moe_quant_config,
     nvfp4_moe_quant_config,
@@ -23,9 +26,6 @@ from vllm.model_executor.layers.fused_moe.cutlass_moe import (
     CutlassExpertsFp4,
 )
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
-from vllm.model_executor.layers.fused_moe.prepare_finalize import (
-    MoEPrepareAndFinalizeNoEP,
-)
 from vllm.scalar_type import scalar_types
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.v1.worker.workspace import init_workspace_manager
@@ -196,10 +196,21 @@ def bench_run(
             g2_alphas=w2_gs,
         )
 
-        kernel = mk.FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+        moe_config = make_dummy_moe_config(
+            num_experts=num_experts,
+            hidden_dim=k,
+            intermediate_size_per_partition=n,
+            in_dtype=a.dtype,
+        )
+        kernel = mk.FusedMoEKernel(
+            maybe_make_prepare_finalize(
+                moe=moe_config,
+                quant_config=quant_config,
+                allow_new_interface=True,
+                use_monolithic=False,
+            ),
             CutlassExpertsFp4(
-                make_dummy_moe_config(),
+                moe_config=moe_config,
                 quant_config=quant_config,
             ),
         )
@@ -240,11 +251,17 @@ def bench_run(
             g1_alphas=w1_gs,
             g2_alphas=w2_gs,
         )
+        moe_config = make_dummy_moe_config()
 
-        kernel = mk.FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+        kernel = mk.FusedMoEKernel(
+            maybe_make_prepare_finalize(
+                moe=moe_config,
+                quant_config=quant_config,
+                allow_new_interface=True,
+                use_monolithic=False,
+            ),
             CutlassExpertsFp4(
-                make_dummy_moe_config(),
+                moe_config=moe_config,
                 quant_config=quant_config,
             ),
         )

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:
-            # Set environment variable if needed
-            if env_var is not None:
-                os.environ["VLLM_CUSTOM_ALLREDUCE_ALGO"] = env_var
-            else:
-                # Clear the environment variable to avoid interference
-                os.environ.pop("VLLM_CUSTOM_ALLREDUCE_ALGO", None)
-
-            latency = self.benchmark_allreduce_single(
-                sequence_length,
-                allreduce_fn,
-                should_use_fn,
-                context,
-                num_warmup,
-                num_trials,
-            )
-            if latency is not None:
-                results[name] = latency
+        for (
+            name,
+            allreduce_fn,
+            should_use_fn,
+            context,
+            env_dict,
+            destroy_fn,
+        ) in communicators:
+            # Save original values and apply new environment variables
+            saved_env = {key: os.environ.get(key) for key in env_dict}
+            for key, value in env_dict.items():
+                os.environ[key] = value
+            try:
+                latency = self.benchmark_allreduce_single(
+                    sequence_length,
+                    allreduce_fn,
+                    should_use_fn,
+                    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
 

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)
-2. Standard tensor_model_parallel_all_reduce + separate rmsnorm/quant operations
+1. FlashInfer's allreduce_fusion with trtllm backend
+   (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
-_FI_WORKSPACE = None
+# Global workspace tensors for FlashInfer (keyed by backend name)
+_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):
-    """Cleanup FlashInfer workspace."""
-    if flashinfer_comm is None or workspace is None:
+def cleanup_flashinfer_workspaces():
+    """Cleanup all FlashInfer workspaces."""
+    if flashinfer_comm is None:
         return
 
-    try:
-        workspace.destroy()
-    except Exception as e:
-        logger.error("Failed to cleanup FlashInfer workspace: %s", e)
+    for backend, workspace in _FI_WORKSPACES.items():
+        try:
+            workspace.destroy()
+        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."""
-    if flashinfer_comm is None or _FI_WORKSPACE is None:
+    """FlashInfer fused allreduce + rmsnorm + FP8 quantization.
+
+    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."""
-    if flashinfer_comm is None or _FI_WORKSPACE is None:
+    """FlashInfer fused allreduce + rmsnorm + FP4 quantization.
+
+    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(),
     )
 
 
@@ -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,
@@ -454,10 +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
-        if flashinfer_comm is not None and allreduce_params is not None:
+        # FlashInfer Fused AllReduce + RMSNorm (all backends)
+        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,
@@ -467,14 +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)
-                    results[f"flashinfer_fused_allreduce_rmsnorm{suffix}"] = float(
-                        "inf"
+                    logger.error(
+                        "FlashInfer (%s) Fused AllReduce+RMSNorm failed: %s",
+                        backend,
+                        e,
                     )
+                    results[key] = float("inf")
 
     if "fp8" in quant_modes:
         # Standard AllReduce + RMSNorm + FP8 Quant
@@ -540,10 +584,12 @@ def run_benchmarks(
                     "inf"
                 )
 
-        # FlashInfer Fused AllReduce + RMSNorm + FP8 Quant Oneshot
-        if flashinfer_comm is not None and allreduce_params is not None:
+        # FlashInfer Fused AllReduce + RMSNorm + FP8 Quant (trtllm only)
+        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,
@@ -555,19 +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}"] = (
-                        time_ms
-                    )
+                    results[key] = 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}"] = (
-                        float("inf")
-                    )
+                    results[key] = float("inf")
 
     if "fp4" in quant_modes and current_platform.has_device_capability(100):
         # Standard AllReduce + RMSNorm + FP4 Quant
@@ -627,10 +670,12 @@ def run_benchmarks(
                     "inf"
                 )
 
-        # FlashInfer Fused AllReduce + RMSNorm + FP4 Quant Oneshot
-        if flashinfer_comm is not None and allreduce_params is not None:
+        # FlashInfer Fused AllReduce + RMSNorm + FP4 Quant (trtllm only)
+        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,
@@ -641,49 +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}"] = (
-                        time_ms
-                    )
+                    results[key] = 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}"] = (
-                        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"
-                )
+                    results[key] = float("inf")
 
     return results
 
@@ -1021,8 +1035,7 @@ def main():
 
     configs = list(itertools.product(args.num_tokens, dtypes, residual_options))
 
-    # Setup FlashInfer workspace if available
-    workspace = None
+    # Setup FlashInfer workspaces for all backends
     allreduce_params = None
 
     if flashinfer_comm is not None:
@@ -1037,15 +1050,17 @@ def main():
             args.hidden_dim * max_element_size
         )
 
-        workspace = setup_flashinfer_workspace(
-            world_size,
-            rank,
-            args.hidden_dim,
-            max_num_token,
-            dtype=workspace_dtype,
-        )
+        for backend in FLASHINFER_BACKENDS:
+            setup_flashinfer_workspace(
+                backend=backend,
+                world_size=world_size,
+                rank=rank,
+                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,
             )
@@ -1071,6 +1086,7 @@ def main():
                 dtype,
                 use_residual,
                 allreduce_params,
+                workspaces=_FI_WORKSPACES,
                 quant_modes=quant_modes,
                 no_oneshot=args.no_oneshot,
             )
@@ -1109,11 +1125,13 @@ def main():
 
     finally:
         # Cleanup
-        if workspace is not None:
-            cleanup_flashinfer_workspace(workspace)
+        cleanup_flashinfer_workspaces()
 
         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(
-            MoEPrepareAndFinalizeNoEP(),
+        fn = mk.FusedMoEKernel(
+            maybe_make_prepare_finalize(
+                moe=moe_config,
+                quant_config=quant_config,
+                allow_new_interface=True,
+                use_monolithic=False,
+            ),
             CutlassExpertsFp8(
-                moe_config=make_dummy_moe_config(
-                    num_experts=w2.shape[0],
-                    hidden_dim=w2.shape[1],
-                    intermediate_size_per_partition=w2.shape[2],
-                    in_dtype=a.dtype,
-                ),
+                moe_config=moe_config,
                 quant_config=quant_config,
             ),
         )
@@ -163,16 +169,22 @@ def bench_run(
             w2_scale=w2_scale,
             per_act_token_quant=per_act_token,
         )
+        moe_config = make_dummy_moe_config(
+            num_experts=w2.shape[0],
+            hidden_dim=w2.shape[1],
+            intermediate_size_per_partition=w2.shape[2],
+            in_dtype=a.dtype,
+        )
 
-        fn = mk.FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+        fn = mk.FusedMoEKernel(
+            maybe_make_prepare_finalize(
+                moe=moe_config,
+                quant_config=quant_config,
+                allow_new_interface=True,
+                use_monolithic=False,
+            ),
             CutlassExpertsFp8(
-                moe_config=make_dummy_moe_config(
-                    num_experts=w2.shape[0],
-                    hidden_dim=w2.shape[1],
-                    intermediate_size_per_partition=w2.shape[2],
-                    in_dtype=a.dtype,
-                ),
+                moe_config=moe_config,
                 quant_config=quant_config,
             ),
         )

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,
@@ -242,24 +245,33 @@ def benchmark_config(
 
         deep_gemm_experts = None
         if use_deep_gemm:
-            deep_gemm_experts = mk.FusedMoEModularKernel(
-                prepare_finalize=MoEPrepareAndFinalizeNoEP(),
+            moe_config = (
+                FusedMoEConfig(
+                    num_experts=num_experts,
+                    experts_per_token=topk,
+                    hidden_dim=hidden_size,
+                    intermediate_size_per_partition=shard_intermediate_size,
+                    num_local_experts=num_experts,
+                    num_logical_experts=num_experts,
+                    activation=MoEActivation.SILU,
+                    moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(),
+                    in_dtype=init_dtype,
+                    routing_method=RoutingMethodType.TopK,
+                    device="cuda",
+                ),
+            )
+            deep_gemm_experts = mk.FusedMoEKernel(
+                prepare_finalize=maybe_make_prepare_finalize(
+                    moe=moe_config,
+                    quant_config=quant_config,
+                    allow_new_interface=True,
+                    use_monolithic=False,
+                ),
                 fused_experts=TritonOrDeepGemmExperts(
-                    moe_config=FusedMoEConfig(
-                        num_experts=num_experts,
-                        experts_per_token=topk,
-                        hidden_dim=hidden_size,
-                        intermediate_size_per_partition=shard_intermediate_size,
-                        num_local_experts=num_experts,
-                        num_logical_experts=num_experts,
-                        activation=MoEActivation.SILU,
-                        moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(),
-                        in_dtype=init_dtype,
-                        routing_method=RoutingMethodType.TopK,
-                        device="cuda",
-                    ),
+                    moe_config=moe_config,
                     quant_config=quant_config,
                 ),
+                inplace=not disable_inplace(),
             )
 
         with override_config(config):
@@ -269,8 +281,16 @@ def benchmark_config(
 
             inplace = not disable_inplace()
             if use_deep_gemm:
-                return deep_gemm_experts(
-                    x, w1, w2, topk_weights, topk_ids, inplace=inplace
+                return deep_gemm_experts.apply(
+                    x,
+                    w1,
+                    w2,
+                    topk_weights,
+                    topk_ids,
+                    activation=MoEActivation.SILU,
+                    global_num_experts=num_experts,
+                    apply_router_weight_on_input=False,
+                    expert_map=False,
                 )
             return fused_experts(
                 x,

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_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_X86_ISA $ENV{VLLM_CPU_X86})
 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)
-    list(APPEND CXX_COMPILE_FLAGS
+if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64" OR ENABLE_X86_ISA)
+    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 ${CXX_COMPILE_FLAGS})
+    list(APPEND CXX_COMPILE_FLAGS_AVX512
         "-mavx512f"
         "-mavx512vl"
         "-mavx512bw"
-        "-mavx512dq")
-
-    find_isa(${CPUINFO} "avx512_bf16" AVX512BF16_FOUND)
-    if (AVX512BF16_FOUND OR ENABLE_AVX512BF16)
-        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
-            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
-            list(APPEND CXX_COMPILE_FLAGS "-mavx512bf16")
-            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")
-    
+        "-mavx512dq"
+        "-mavx512bf16"
+        "-mavx512vnni"
+        "-mamx-bf16"
+        "-mamx-tile")
+    list(APPEND CXX_COMPILE_FLAGS_AVX2
+        "-mavx2")
 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)
-if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
+# Build oneDNN for GEMM kernels
+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 "")
@@ -329,13 +258,21 @@ 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_CPU_ISA_HINTS "OFF")
-    set(ONEDNN_VERBOSE "OFF")
+    set(ONEDNN_ENABLE_MAX_CPU_ISA "ON")
+    set(ONEDNN_ENABLE_CPU_ISA_HINTS "ON")
+    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 +285,20 @@ 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 +333,6 @@ 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 +345,83 @@ 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") 
 
-#
-# Define extension targets
-#
+    set(VLLM_EXT_SRC_AVX2 
+        "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(
-    _C
-    DESTINATION vllm
-    LANGUAGE CXX
-    SOURCES ${VLLM_EXT_SRC}
-    LIBRARIES ${LIBS}
-    COMPILE_FLAGS ${CXX_COMPILE_FLAGS}
-    USE_SABI 3
-    WITH_SOABI
-)
+    message(STATUS "CPU extension (AVX512) source files: ${VLLM_EXT_SRC_AVX512}")
+    message(STATUS "CPU extension (AVX2) source files: ${VLLM_EXT_SRC_AVX2}")
+
+    define_extension_target(
+        _C
+        DESTINATION vllm
+        LANGUAGE CXX
+        SOURCES ${VLLM_EXT_SRC_AVX512}
+        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.
-# setup.py installs every component separately but uses the same prefix for all.
-# ALL_COMPONENTS is used to avoid duplication for FA2 and FA3,
-# and these statements don't hurt when installing neither component.
+# ALL_COMPONENTS ensures the save/modify/restore runs exactly once regardless
+# of how many components are being installed, avoiding double-append of /vllm/.
 install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" ALL_COMPONENTS)
 install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
 install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" ALL_COMPONENTS)
@@ -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
-# case only one is built, in the case both are built redundant work is done)
-install(
-  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm/vllm_flash_attn
-  COMPONENT _vllm_fa2_C
-  FILES_MATCHING PATTERN "*.py"
-)
+# Install shared Python files for both FA2 and FA3 components
+foreach(_FA_COMPONENT _vllm_fa2_C _vllm_fa3_C)
+  # Ensure the vllm/vllm_flash_attn directory exists before installation
+  install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")"
+    COMPONENT ${_FA_COMPONENT})
 
-install(
-  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm/vllm_flash_attn
-  COMPONENT _vllm_fa3_C
-  FILES_MATCHING PATTERN "*.py"
-)
+  # Copy vllm_flash_attn python files (except __init__.py and flash_attn_interface.py
+  # which are source-controlled in vllm)
+  install(
+    DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
+    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

@@ -5,117 +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 && \
-    defined(CUDA_VERSION) && CUDA_VERSION >= 12090
-  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 && \
-    defined(CUDA_VERSION) && CUDA_VERSION >= 12090
-  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,
@@ -131,16 +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&),
@@ -155,36 +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 vec_t = typename VecTraits<use_256b>::vec_t;
-    constexpr int ARCH_MAX_VEC_SIZE = VecTraits<use_256b>::ARCH_MAX_VEC_SIZE;
-    constexpr int VEC_SIZE = ARCH_MAX_VEC_SIZE / sizeof(packed_t);
+    using cuda_t = typename CUDATypeConverter<scalar_t>::Type;
+    using pvec_t = PackedVec<cuda_t, use_256b>;
 
-    const vec_t* x_vec = reinterpret_cast<const vec_t*>(x_ptr);
-    const vec_t* y_vec = reinterpret_cast<const vec_t*>(y_ptr);
-    vec_t* out_vec = reinterpret_cast<vec_t*>(out_ptr);
-    const int num_vecs = d / 2 / VEC_SIZE;
+    const pvec_t* x_vec = reinterpret_cast<const pvec_t*>(x_ptr);
+    const pvec_t* y_vec = reinterpret_cast<const pvec_t*>(y_ptr);
+    pvec_t* out_vec = reinterpret_cast<pvec_t*>(out_ptr);
+    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]);
-        y = VLLM_LDG(&y_vec[i]);
+        ld128(x, &x_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++) {
-        xp[j] =
-            packed_compute<packed_t, PACKED_ACT_FN, act_first>(xp[j], yp[j]);
+      for (int j = 0; j < pvec_t::NUM_ELTS; j++) {
+        x.elts[j] = packed_compute<packed_t, PACKED_ACT_FN, act_first>(
+            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 {
@@ -272,51 +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)     \
-  auto dtype = input.scalar_type();                                         \
-  int d = input.size(-1) / 2;                                               \
-  int64_t num_tokens = input.numel() / input.size(-1);                      \
-  if (num_tokens == 0) {                                                    \
-    return;                                                                 \
-  }                                                                         \
-  dim3 grid(num_tokens);                                                    \
-  int cc_major = at::cuda::getCurrentDeviceProperties()->major;             \
-  int support_vec = (cc_major >= 10 && num_tokens > 128) ? 32 : 16;         \
-  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) {                               \
-      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {       \
-        vllm::act_and_mul_kernel<                                           \
-            scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,      \
-            KERNEL<scalar_t>,                                               \
-            PACKED_KERNEL<typename vllm::PackedTraits<scalar_t>::packed_t>, \
-            ACT_FIRST, true, true><<<grid, block, 0, stream>>>(             \
-            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);       \
-      });                                                                   \
-    } else {                                                                \
-      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {       \
-        vllm::act_and_mul_kernel<                                           \
-            scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,      \
-            KERNEL<scalar_t>,                                               \
-            PACKED_KERNEL<typename vllm::PackedTraits<scalar_t>::packed_t>, \
-            ACT_FIRST, true, false><<<grid, block, 0, stream>>>(            \
-            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);       \
-      });                                                                   \
-    }                                                                       \
-  } else {                                                                  \
-    dim3 block(std::min(d, 1024));                                          \
-    VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {         \
-      vllm::act_and_mul_kernel<                                             \
-          scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,        \
-          KERNEL<scalar_t>,                                                 \
-          PACKED_KERNEL<typename vllm::PackedTraits<scalar_t>::packed_t>,   \
-          ACT_FIRST, false><<<grid, block, 0, stream>>>(                    \
-          out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);         \
-    });                                                                     \
+#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, PACKED_KERNEL, ACT_FIRST)        \
+  auto dtype = input.scalar_type();                                            \
+  int d = input.size(-1) / 2;                                                  \
+  int64_t num_tokens = input.numel() / input.size(-1);                         \
+  if (num_tokens == 0) {                                                       \
+    return;                                                                    \
+  }                                                                            \
+  dim3 grid(num_tokens);                                                       \
+  int cc_major = at::cuda::getCurrentDeviceProperties()->major;                \
+  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 (CUDA_VERSION >= 12090 && cc_major >= 10 && num_tokens > 128) {         \
+      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {          \
+        vllm::act_and_mul_kernel<                                              \
+            scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,      \
+            KERNEL<scalar_t>,                                                  \
+            PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
+            ACT_FIRST, true, true><<<grid, block, 0, stream>>>(                \
+            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);          \
+      });                                                                      \
+    } else {                                                                   \
+      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {          \
+        vllm::act_and_mul_kernel<                                              \
+            scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,      \
+            KERNEL<scalar_t>,                                                  \
+            PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
+            ACT_FIRST, true, false><<<grid, block, 0, stream>>>(               \
+            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);          \
+      });                                                                      \
+    }                                                                          \
+  } else {                                                                     \
+    dim3 block(std::min(d, 1024));                                             \
+    VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {            \
+      vllm::act_and_mul_kernel<                                                \
+          scalar_t, typename vllm::PackedTypeConverter<scalar_t>::Type,        \
+          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]
@@ -378,35 +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 vec_t = typename VecTraits<use_256b>::vec_t;
-    constexpr int ARCH_MAX_VEC_SIZE = VecTraits<use_256b>::ARCH_MAX_VEC_SIZE;
-    constexpr int VEC_SIZE = ARCH_MAX_VEC_SIZE / sizeof(packed_t);
+    using cuda_t = typename CUDATypeConverter<scalar_t>::Type;
+    using pvec_t = PackedVec<cuda_t, use_256b>;
 
-    const vec_t* x_vec = reinterpret_cast<const vec_t*>(x_ptr);
-    const vec_t* y_vec = reinterpret_cast<const vec_t*>(y_ptr);
-    vec_t* out_vec = reinterpret_cast<vec_t*>(out_ptr);
-    const int num_vecs = d / 2 / VEC_SIZE;
+    const pvec_t* x_vec = reinterpret_cast<const pvec_t*>(x_ptr);
+    const pvec_t* y_vec = reinterpret_cast<const pvec_t*>(y_ptr);
+    pvec_t* out_vec = reinterpret_cast<pvec_t*>(out_ptr);
+    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]);
-        y = VLLM_LDG(&y_vec[i]);
+        ld128(x, &x_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++) {
-        xp[j] = packed_mul(PACKED_ACT_FN(xp[j], param), yp[j]);
+      for (int j = 0; j < pvec_t::NUM_ELTS; 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 {
@@ -499,21 +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);                                                        \
@@ -522,10 +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);                                                        \
@@ -535,9 +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);     \
     });                                                                        \
@@ -629,14 +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/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_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/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 defination 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
-}
-
-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(
+  ops.def("init_cpu_threads_env(str cpu_ids) -> str", &init_cpu_threads_env);
+  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

csrc/dsv3_fused_a_gemm.cu

@@ -745,3 +745,7 @@ void dsv3_fused_a_gemm(torch::Tensor& output, torch::Tensor const& mat_a,
         stream);
   }
 }
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("dsv3_fused_a_gemm", &dsv3_fused_a_gemm);
+}

csrc/mamba/mamba_ssm/selective_scan.h

@@ -15,9 +15,9 @@
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 
 struct SSMParamsBase {
-    using index_t = uint32_t;
+    using index_t = size_t;
 
-    int batch, dim, seqlen, dstate, n_groups, n_chunks;
+    int batch, dim, seqlen, dstate, n_groups;
     int dim_ngroups_ratio;
     bool is_variable_B;
     bool is_variable_C;
@@ -72,6 +72,8 @@ struct SSMParamsBase {
     void *__restrict__ block_idx_first_scheduled_token_ptr;  // (batch,) - first block to write
     void *__restrict__ block_idx_last_scheduled_token_ptr;   // (batch,) - last block to write
     void *__restrict__ initial_state_idx_ptr;  // (batch,) - index of the initial state to use
+    void *__restrict__ cu_chunk_seqlen_ptr;      // (nchunks+1,) - cumulative chunk token offsets
+    void *__restrict__ last_chunk_indices_ptr;   // (batch,) - index of last chunk per sequence
 };
 
 

csrc/mamba/mamba_ssm/selective_scan_fwd.cu

@@ -81,7 +81,6 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
     constexpr bool kIsVariableC = Ktraits::kIsVariableC;
     constexpr bool kHasZ = Ktraits::kHasZ;
     constexpr bool kVarlen = Ktraits::kVarlen;
-    constexpr int kNThreads = Ktraits::kNThreads;
     constexpr int kNItems = Ktraits::kNItems;
     constexpr int kNRows = Ktraits::kNRows;
     constexpr bool kDirectIO = Ktraits::kDirectIO;
@@ -161,17 +160,8 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
         }
     }
 
-
-    // for (int state_idx = threadIdx.x; state_idx < params.dstate; state_idx += blockDim.x) {
-    //     smem_a[state_idx] = A[state_idx * params.A_dstate_stride];
-    //     smem_bc[state_idx] = B[state_idx * params.B_dstate_stride] * C[state_idx * params.C_dstate_stride];
-    // }
-
-    constexpr int kChunkSize = kNThreads * kNItems;
-
     // Use block_size for chunking when APC is enabled, otherwise use 2048 for backwards compatibility
-    const int iteration_chunk_size = params.cache_enabled ? params.block_size : 2048;
-    const int n_chunks = (seqlen + iteration_chunk_size - 1) / iteration_chunk_size;
+    const int block_size = params.cache_enabled ? params.block_size : 2048;
 
     const int* batch_cache_indices = cache_indices != nullptr ?
                                      cache_indices + batch_id * params.cache_indices_stride : nullptr;
@@ -181,10 +171,44 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
                                           reinterpret_cast<const int*>(params.block_idx_last_scheduled_token_ptr) : nullptr;
     const int* initial_state_idx = params.initial_state_idx_ptr != nullptr ?
                                    reinterpret_cast<const int*>(params.initial_state_idx_ptr) : nullptr;
+    const int* cu_chunk_seqlen = params.cu_chunk_seqlen_ptr != nullptr ?
+                                 reinterpret_cast<const int*>(params.cu_chunk_seqlen_ptr) : nullptr;
+    const int* last_chunk_indices = params.last_chunk_indices_ptr != nullptr ?
+                                    reinterpret_cast<const int*>(params.last_chunk_indices_ptr) : nullptr;
 
     const size_t load_cache_slot = params.cache_enabled && batch_cache_indices != nullptr ? batch_cache_indices[initial_state_idx[batch_id]] : cache_index;
 
+    const int block_idx_first = (params.cache_enabled && block_idx_first_scheduled != nullptr) ?
+                                 block_idx_first_scheduled[batch_id] : 0;
+
+    // Determine chunk boundaries from pre-computed metadata (APC mode)
+    // or fall back to simple block_size chunking.
+    int first_chunk_idx, n_chunks;
+    int current_position;
+
+    if (cu_chunk_seqlen != nullptr && last_chunk_indices != nullptr) {
+        const int last_chunk_idx = last_chunk_indices[batch_id];
+        first_chunk_idx = (batch_id == 0) ? 0 : last_chunk_indices[batch_id - 1] + 1;
+        n_chunks = last_chunk_idx - first_chunk_idx + 1;
+        // Derive current_position: if the first chunk is partial (fills remainder
+        // of a started block), offset into the block accordingly.
+        const int first_chunk_tokens = cu_chunk_seqlen[first_chunk_idx + 1] - cu_chunk_seqlen[first_chunk_idx];
+        const int chunk_start_offset = (n_chunks > 1 && first_chunk_tokens < block_size)
+                                        ? (block_size - first_chunk_tokens) : 0;
+        current_position = block_idx_first * block_size + chunk_start_offset;
+    } else {
+        first_chunk_idx = 0;
+        n_chunks = (seqlen + block_size - 1) / block_size;
+        current_position = 0;
+    }
+
+    int tokens_processed = 0;
+
     for (int chunk = 0; chunk < n_chunks; ++chunk) {
+        const int chunk_tokens = (cu_chunk_seqlen != nullptr)
+            ? cu_chunk_seqlen[first_chunk_idx + chunk + 1] - cu_chunk_seqlen[first_chunk_idx + chunk]
+            : min(block_size, seqlen - tokens_processed);
+        if (chunk_tokens <= 0) break;
         input_t u_vals[kNRows][kNItems], delta_vals_load[kNRows][kNItems];
 
         __syncthreads();
@@ -193,12 +217,12 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
             if constexpr (!kDirectIO) {
                 if (r > 0) { __syncthreads(); }
             }
-            load_input<Ktraits>(u + r * params.u_d_stride, u_vals[r], smem_load, seqlen - chunk * kChunkSize);
+            load_input<Ktraits>(u + r * params.u_d_stride, u_vals[r], smem_load, chunk_tokens);
             if constexpr (!kDirectIO) { __syncthreads(); }
-            load_input<Ktraits>(delta + r * params.delta_d_stride, delta_vals_load[r], smem_load, seqlen - chunk * kChunkSize);
+            load_input<Ktraits>(delta + r * params.delta_d_stride, delta_vals_load[r], smem_load, chunk_tokens);
         }
-        u += kChunkSize;
-        delta += kChunkSize;
+        u += chunk_tokens;
+        delta += chunk_tokens;
     
         float delta_vals[kNRows][kNItems], delta_u_vals[kNRows][kNItems], out_vals[kNRows][kNItems];
         #pragma unroll
@@ -232,7 +256,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
             weight_t B_vals[kNItems], C_vals[kNItems];
             if constexpr (kIsVariableB) {
                 load_weight<Ktraits>(Bvar + state_idx * params.B_dstate_stride, B_vals,
-                    smem_load_weight, (seqlen - chunk * kChunkSize) * (1));
+                    smem_load_weight, chunk_tokens);
                 if constexpr (!kIsVariableC) {
                     #pragma unroll
                     for (int r = 0; r < kNRows; ++r) {
@@ -243,7 +267,7 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
             if constexpr (kIsVariableC) {
                 auto &smem_load_weight_C = !kIsVariableB ? smem_load_weight : smem_load_weight1;
                 load_weight<Ktraits>(Cvar + state_idx * params.C_dstate_stride, C_vals,
-                    smem_load_weight_C, (seqlen - chunk * kChunkSize) * (1));
+                    smem_load_weight_C, chunk_tokens);
                 if constexpr (!kIsVariableB) {
                     #pragma unroll
                     for (int r = 0; r < kNRows; ++r) {
@@ -266,10 +290,8 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
                 for (int i = 0; i < kNItems; ++i) {
                     thread_data[i] = make_float2(exp2f(delta_vals[r][i] * A_val[r]),
                                                  !kIsVariableB ? delta_u_vals[r][i] : B_vals[i] * delta_u_vals[r][i]);
-                    if (seqlen % (kNItems * kNThreads) != 0) {  // So that the last state is correct
-                        if (threadIdx.x * kNItems + i >= seqlen - chunk * kChunkSize) {
-                            thread_data[i] = make_float2(1.f, 0.f);
-                        }
+                    if (threadIdx.x * kNItems + i >= chunk_tokens) {
+                        thread_data[i] = make_float2(1.f, 0.f);
                     }
                 }
                 // Initialize running total
@@ -301,14 +323,14 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
                 if (threadIdx.x == 0) {
                     smem_running_prefix[state_idx + r * MAX_DSTATE] = prefix_op.running_prefix;
 
-                    // Store state at the end of each chunk when cache is enabled
+                    // Store state at the end of each aligned chunk when cache is enabled
                     if (params.cache_enabled && batch_cache_indices != nullptr) {
-
                         size_t cache_slot;
                         if (chunk == n_chunks - 1) {
                             cache_slot = batch_cache_indices[block_idx_last_scheduled[batch_id]];
                         } else {
-                            cache_slot = batch_cache_indices[block_idx_first_scheduled[batch_id] + chunk];
+                            const int block_idx_completed = (current_position + chunk_tokens - 1) / block_size;
+                            cache_slot = batch_cache_indices[block_idx_completed];
                         }
 
                         size_t state_offset = cache_slot * params.ssm_states_batch_stride +
@@ -331,38 +353,41 @@ void selective_scan_fwd_kernel(SSMParamsBase params) {
             }
         }
         input_t *out = reinterpret_cast<input_t *>(params.out_ptr) + sequence_start_index * params.out_batch_stride
-            + dim_id * kNRows * params.out_d_stride + chunk * kChunkSize;
+            + dim_id * kNRows * params.out_d_stride + tokens_processed;
         __syncthreads();
         #pragma unroll
         for (int r = 0; r < kNRows; ++r) {
             if constexpr (!kDirectIO) {
                 if (r > 0) { __syncthreads(); }
             }
-            store_output<Ktraits>(out + r * params.out_d_stride, out_vals[r], smem_store, seqlen - chunk * kChunkSize);
+            store_output<Ktraits>(out + r * params.out_d_stride, out_vals[r], smem_store, chunk_tokens);
         }
 
         if constexpr (kHasZ) {
             input_t *z = reinterpret_cast<input_t *>(params.z_ptr) + sequence_start_index * params.z_batch_stride
-                + dim_id * kNRows * params.z_d_stride + chunk * kChunkSize;
+                + dim_id * kNRows * params.z_d_stride + tokens_processed;
             input_t *out_z = reinterpret_cast<input_t *>(params.out_z_ptr) + sequence_start_index * params.out_z_batch_stride
-                + dim_id * kNRows * params.out_z_d_stride + chunk * kChunkSize;
+                + dim_id * kNRows * params.out_z_d_stride + tokens_processed;
             #pragma unroll
             for (int r = 0; r < kNRows; ++r) {
                 input_t z_vals[kNItems];
                 __syncthreads();
-                load_input<Ktraits>(z + r * params.z_d_stride, z_vals, smem_load, seqlen - chunk * kChunkSize);
+                load_input<Ktraits>(z + r * params.z_d_stride, z_vals, smem_load, chunk_tokens);
                 #pragma unroll
                 for (int i = 0; i < kNItems; ++i) {
                     float z_val = z_vals[i];
                     out_vals[r][i] *= z_val / (1 + expf(-z_val));
                 }
                 __syncthreads();
-                store_output<Ktraits>(out_z + r * params.out_z_d_stride, out_vals[r], smem_store, seqlen - chunk * kChunkSize);
+                store_output<Ktraits>(out_z + r * params.out_z_d_stride, out_vals[r], smem_store, chunk_tokens);
             }
         }
 
-        Bvar += kChunkSize * 1;
-        Cvar += kChunkSize * 1;
+        Bvar += chunk_tokens;
+        Cvar += chunk_tokens;
+
+        tokens_processed += chunk_tokens;
+        current_position += chunk_tokens;
     }
 }
 
@@ -506,7 +531,9 @@ void set_ssm_params_fwd(SSMParamsBase &params,
                         int64_t block_size,
                         const std::optional<torch::Tensor> &block_idx_first_scheduled_token,
                         const std::optional<torch::Tensor> &block_idx_last_scheduled_token,
-                        const std::optional<torch::Tensor> &initial_state_idx) {
+                        const std::optional<torch::Tensor> &initial_state_idx,
+                        const std::optional<torch::Tensor> &cu_chunk_seqlen,
+                        const std::optional<torch::Tensor> &last_chunk_indices) {
 
     // Reset the parameters
     memset(&params, 0, sizeof(params));
@@ -548,6 +575,8 @@ void set_ssm_params_fwd(SSMParamsBase &params,
     params.block_idx_first_scheduled_token_ptr = block_idx_first_scheduled_token.has_value() ? block_idx_first_scheduled_token.value().data_ptr() : nullptr;
     params.block_idx_last_scheduled_token_ptr = block_idx_last_scheduled_token.has_value() ? block_idx_last_scheduled_token.value().data_ptr() : nullptr;
     params.initial_state_idx_ptr = initial_state_idx.has_value() ? initial_state_idx.value().data_ptr() : nullptr;
+    params.cu_chunk_seqlen_ptr = cu_chunk_seqlen.has_value() ? cu_chunk_seqlen.value().data_ptr() : nullptr;
+    params.last_chunk_indices_ptr = last_chunk_indices.has_value() ? last_chunk_indices.value().data_ptr() : nullptr;
 
     // All stride are in elements, not bytes.
     params.A_d_stride = A.stride(0);
@@ -633,7 +662,9 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
                   int64_t block_size,
                   const std::optional<torch::Tensor> &block_idx_first_scheduled_token,
                   const std::optional<torch::Tensor> &block_idx_last_scheduled_token,
-                  const std::optional<torch::Tensor> &initial_state_idx) {
+                  const std::optional<torch::Tensor> &initial_state_idx,
+                  const std::optional<torch::Tensor> &cu_chunk_seqlen,
+                  const std::optional<torch::Tensor> &last_chunk_indices) {
     auto input_type = u.scalar_type();
     auto weight_type = A.scalar_type();
     TORCH_CHECK(input_type == at::ScalarType::Float || input_type == at::ScalarType::Half || input_type == at::ScalarType::BFloat16);
@@ -778,7 +809,9 @@ void selective_scan_fwd(const torch::Tensor &u, const torch::Tensor &delta,
                        block_size,
                        block_idx_first_scheduled_token,
                        block_idx_last_scheduled_token,
-                       initial_state_idx
+                       initial_state_idx,
+                       cu_chunk_seqlen,
+                       last_chunk_indices
                        );
 
     

csrc/moe/dsv3_router_gemm_entry.cu

@@ -20,10 +20,12 @@
 
 #include <ATen/ATen.h>
 #include <ATen/cuda/CUDAContext.h>
+#include <torch/all.h>
 
 #include <cuda_bf16.h>
 #include <cuda_runtime.h>
 
+#include "core/registration.h"
 #include "dsv3_router_gemm_utils.h"
 
 static constexpr int DEFAULT_NUM_EXPERTS = 256;
@@ -161,3 +163,7 @@ void dsv3_router_gemm(at::Tensor& output,       // [num_tokens, num_experts]
     }
   }
 }
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("dsv3_router_gemm", &dsv3_router_gemm);
+}

csrc/moe/moe_align_sum_kernels.cu

@@ -172,7 +172,7 @@ __device__ void _moe_align_block_size(
     }
   }
 
-  // Fill remaining expert_ids with 0
+  // Fill remaining expert_ids with -1
   const size_t fill_start_idx =
       cumsum[cumsum_offset + num_experts] / block_size + threadIdx.x;
   for (size_t i = fill_start_idx; i < max_num_m_blocks; i += blockDim.x) {
@@ -265,7 +265,7 @@ __device__ void _moe_align_block_size_small_batch_expert(
     }
   }
 
-  // Fill remaining expert_ids with 0
+  // Fill remaining expert_ids with -1
   const size_t fill_start_idx = cumsum[num_experts] / block_size + tid;
   for (size_t i = fill_start_idx; i < max_num_m_blocks; i += stride) {
     expert_ids[expert_ids_offset + i] = inactive_expert_id;
@@ -332,7 +332,7 @@ __global__ void moe_align_block_size_kernel(
       topk_ids, sorted_token_ids, expert_ids, total_tokens_post_pad, expert_map,
       num_experts, padded_num_experts, experts_per_warp, block_size, numel,
       cumsum, max_num_tokens_padded, CEILDIV(max_num_tokens_padded, block_size),
-      0, 0, topk_num, nullptr, has_expert_map);
+      0, -1, topk_num, nullptr, has_expert_map);
 }
 
 template <typename scalar_t>
@@ -373,7 +373,7 @@ __global__ void moe_align_block_size_small_batch_expert_kernel(
   _moe_align_block_size_small_batch_expert<scalar_t, fill_threads>(
       topk_ids, sorted_token_ids, expert_ids, total_tokens_post_pad, expert_map,
       num_experts, block_size, numel, max_num_tokens_padded,
-      CEILDIV(max_num_tokens_padded, block_size), 0, 0, topk_num, nullptr,
+      CEILDIV(max_num_tokens_padded, block_size), -1, 0, topk_num, nullptr,
       has_expert_map);
 }
 

csrc/moe/moe_ops.h

@@ -58,6 +58,10 @@ void shuffle_rows(const torch::Tensor& input_tensor,
                   torch::Tensor& output_tensor);
 
 #ifndef USE_ROCM
+// cuBLAS bf16 x bf16 -> fp32 router GEMM (fallback for non-SM90 / batch > 16)
+torch::Tensor router_gemm_bf16_fp32(torch::Tensor const& input,
+                                    torch::Tensor const& weight);
+
 // DeepSeek V3 optimized router GEMM kernel for SM90+
 // Computes output = mat_a @ mat_b.T where:
 //   mat_a: [num_tokens, hidden_dim] in bf16

csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm.cu

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

csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm_functor.cuh

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

csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm_launcher.cuh

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

csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm_traits.cuh

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

csrc/moe/mxfp8_moe/mxfp8_experts_quant.cu

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

csrc/moe/mxfp8_moe/mxfp8_experts_quant.cuh

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

csrc/moe/router_gemm.cu

@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+// bf16 x bf16 -> fp32 router GEMM via cuBLAS.
+// Uses CUBLAS_COMPUTE_32F so bf16 operands accumulate into fp32,
+// matching TRT-LLM's cuBLAS fallback behaviour in dsv3RouterGemmOp.
+
+#include <torch/all.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <cublas_v2.h>
+
+// cuBLAS column-major math for row-major PyTorch tensors:
+//   weight[N,K]_row  lda=K  -> cuBLAS sees (K,N) col-major; CUBLAS_OP_T ->
+//   (N,K) input[M,K]_row   ldb=K  -> cuBLAS sees (K,M) col-major; CUBLAS_OP_N
+//   -> (K,M) out[M,N]_row     ldc=N  -> cuBLAS sees (N,M) col-major (written as
+//   output^T)
+// cuBLAS: C(N,M) = weight(N,K) @ input(K,M)  =>  C^T = output[M,N]
+// params: m=N, n=M, k=K, lda=K (weight), ldb=K (input), ldc=N (output)
+
+torch::Tensor router_gemm_bf16_fp32(torch::Tensor const& input,
+                                    torch::Tensor const& weight) {
+  TORCH_CHECK(input.dtype() == torch::kBFloat16,
+              "router_gemm_bf16_fp32: input must be bfloat16");
+  TORCH_CHECK(weight.dtype() == torch::kBFloat16,
+              "router_gemm_bf16_fp32: weight must be bfloat16");
+  TORCH_CHECK(input.dim() == 2 && weight.dim() == 2,
+              "router_gemm_bf16_fp32: input and weight must be 2-D");
+  TORCH_CHECK(input.size(1) == weight.size(1),
+              "router_gemm_bf16_fp32: inner dimensions must match");
+
+  int64_t const M = input.size(0);
+  int64_t const N = weight.size(0);
+  int64_t const K = input.size(1);
+
+  auto out = torch::empty({M, N}, input.options().dtype(torch::kFloat32));
+
+  cublasHandle_t handle = at::cuda::getCurrentCUDABlasHandle();
+  TORCH_CUDABLAS_CHECK(
+      cublasSetStream(handle, at::cuda::getCurrentCUDAStream()));
+
+  float const alpha = 1.0f;
+  float const beta = 0.0f;
+
+  TORCH_CUDABLAS_CHECK(cublasGemmEx(
+      handle, CUBLAS_OP_T, CUBLAS_OP_N, static_cast<int>(N),
+      static_cast<int>(M), static_cast<int>(K), &alpha, weight.data_ptr(),
+      CUDA_R_16BF, static_cast<int>(K), input.data_ptr(), CUDA_R_16BF,
+      static_cast<int>(K), &beta, out.data_ptr(), CUDA_R_32F,
+      static_cast<int>(N), CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
+
+  return out;
+}

csrc/moe/torch_bindings.cpp

@@ -125,9 +125,13 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
       "Tensor)");
   m.impl("grouped_topk", torch::kCUDA, &grouped_topk);
 
+  // cuBLAS bf16 x bf16 -> fp32 router GEMM (fallback for non-SM90 / batch > 16)
+  m.def("router_gemm_bf16_fp32(Tensor input, Tensor weight) -> Tensor");
+  m.impl("router_gemm_bf16_fp32", torch::kCUDA, &router_gemm_bf16_fp32);
+
   // DeepSeek V3 optimized router GEMM for SM90+
   m.def("dsv3_router_gemm(Tensor! output, Tensor mat_a, Tensor mat_b) -> ()");
-  m.impl("dsv3_router_gemm", torch::kCUDA, &dsv3_router_gemm);
+  // conditionally compiled so impl registration is in source file
 #endif
 }
 

csrc/ops.h

@@ -269,13 +269,13 @@ void get_cutlass_moe_mm_problem_sizes_from_expert_offsets(
     torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
     const int64_t n, const int64_t k, const bool swap_ab);
 
-void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
-                                  torch::Tensor& problem_sizes1,
-                                  torch::Tensor& problem_sizes2,
-                                  const torch::Tensor& expert_num_tokens,
-                                  const int64_t num_local_experts,
-                                  const int64_t padded_m, const int64_t n,
-                                  const int64_t k);
+void get_cutlass_batched_moe_mm_data(torch::Tensor& expert_offsets,
+                                     torch::Tensor& problem_sizes1,
+                                     torch::Tensor& problem_sizes2,
+                                     const torch::Tensor& expert_num_tokens,
+                                     const int64_t num_local_experts,
+                                     const int64_t padded_m, const int64_t n,
+                                     const int64_t k);
 
 void cutlass_scaled_mm_azp(torch::Tensor& out, torch::Tensor const& a,
                            torch::Tensor const& b,
@@ -371,7 +371,9 @@ void selective_scan_fwd(
     const torch::Tensor& ssm_states, int64_t pad_slot_id, int64_t block_size,
     const std::optional<torch::Tensor>& block_idx_first_scheduled_token,
     const std::optional<torch::Tensor>& block_idx_last_scheduled_token,
-    const std::optional<torch::Tensor>& initial_state_idx);
+    const std::optional<torch::Tensor>& initial_state_idx,
+    const std::optional<torch::Tensor>& cu_chunk_seqlen,
+    const std::optional<torch::Tensor>& last_chunk_indices);
 
 torch::Tensor dynamic_4bit_int_moe_cpu(
     torch::Tensor x, torch::Tensor topk_ids, torch::Tensor topk_weights,

csrc/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu

@@ -39,12 +39,12 @@ namespace vllm {
 template <class Type, bool UE8M0_SF = false>
 __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
     silu_mul_cvt_fp16_to_fp4(int32_t numRows, int32_t numCols,
-                             int32_t num_padded_cols,
+                             int32_t num_packed_cols,
                              Type const* __restrict__ in,
                              float const* __restrict__ SFScale,
                              uint32_t* __restrict__ out,
                              uint32_t* __restrict__ SFout) {
-  using PackedVec = vllm::PackedVec<Type>;
+  using PackedVec = vllm::PackedVec<Type, CVT_FP4_PACK16>;
   static constexpr int CVT_FP4_NUM_THREADS_PER_SF =
       (CVT_FP4_SF_VEC_SIZE / CVT_FP4_ELTS_PER_THREAD);
   static_assert(sizeof(PackedVec) == sizeof(Type) * CVT_FP4_ELTS_PER_THREAD,
@@ -63,7 +63,7 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
 
   // Input tensor row/col loops.
   for (int rowIdx = blockIdx.x; rowIdx < numRows; rowIdx += gridDim.x) {
-    if (colIdx < num_padded_cols) {
+    if (colIdx < num_packed_cols) {
       PackedVec in_vec;
       PackedVec in_vec2;
       int64_t inOffset =
@@ -73,19 +73,19 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
 
       bool valid = (rowIdx < numRows) && (elem_idx < numCols);
       if constexpr (CVT_FP4_PACK16) {
-        ld256_or_zero_cg_u32<Type>(
-            in_vec, &reinterpret_cast<const uint32_t*>(in)[inOffset * 8],
-            valid);
-        ld256_or_zero_cg_u32<Type>(
-            in_vec2, &reinterpret_cast<const uint32_t*>(in)[inOffset2 * 8],
-            valid);
+        ld256_cg_or_zero(reinterpret_cast<u32x8_t&>(in_vec),
+                         &reinterpret_cast<const uint32_t*>(in)[inOffset * 8],
+                         valid);
+        ld256_cg_or_zero(reinterpret_cast<u32x8_t&>(in_vec2),
+                         &reinterpret_cast<const uint32_t*>(in)[inOffset2 * 8],
+                         valid);
       } else {
-        ld128_or_zero_cg_u32<Type>(
-            in_vec, &reinterpret_cast<const uint32_t*>(in)[inOffset * 4],
-            valid);
-        ld128_or_zero_cg_u32<Type>(
-            in_vec2, &reinterpret_cast<const uint32_t*>(in)[inOffset2 * 4],
-            valid);
+        ld128_cg_or_zero(reinterpret_cast<uint4&>(in_vec),
+                         &reinterpret_cast<const uint32_t*>(in)[inOffset * 4],
+                         valid);
+        ld128_cg_or_zero(reinterpret_cast<uint4&>(in_vec2),
+                         &reinterpret_cast<const uint32_t*>(in)[inOffset2 * 4],
+                         valid);
       }
 
       // Compute silu and mul
@@ -107,7 +107,9 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
               (uint64_t(out_val.hi) << 32) | uint64_t(out_val.lo);
           reinterpret_cast<uint64_t*>(out)[outOffset >> 1] = packed64;
         } else {
-          out[inOffset] = out_val;
+          int64_t outOffset =
+              rowIdx * (numCols / CVT_FP4_ELTS_PER_THREAD) + colIdx;
+          out[outOffset] = out_val;
         }
       }
     }
@@ -140,9 +142,9 @@ void silu_and_mul_nvfp4_quant_sm1xxa(torch::Tensor& output,  // [..., d]
   int const numBlocksPerSM =
       vllm_runtime_blocks_per_sm(static_cast<int>(block.x));
 
-  int sf_n_unpadded = int(n / CVT_FP4_SF_VEC_SIZE);
+  int num_packed_cols = int(n / CVT_FP4_ELTS_PER_THREAD);
 
-  int grid_y = vllm::div_round_up(sf_n_unpadded, static_cast<int>(block.x));
+  int grid_y = vllm::div_round_up(num_packed_cols, static_cast<int>(block.x));
   int grid_x = std::min(
       int(m), std::max(1, (multiProcessorCount * numBlocksPerSM) / grid_y));
   dim3 grid(grid_x, grid_y);
@@ -152,7 +154,7 @@ void silu_and_mul_nvfp4_quant_sm1xxa(torch::Tensor& output,  // [..., d]
         using cuda_type = vllm::CUDATypeConverter<scalar_t>::Type;
         auto input_ptr = static_cast<cuda_type const*>(input.data_ptr());
         vllm::silu_mul_cvt_fp16_to_fp4<cuda_type><<<grid, block, 0, stream>>>(
-            m, n, sf_n_unpadded, input_ptr, input_sf_ptr,
+            m, n, num_packed_cols, input_ptr, input_sf_ptr,
             reinterpret_cast<uint32_t*>(output_ptr),
             reinterpret_cast<uint32_t*>(sf_out));
       });

csrc/quantization/fp4/nvfp4_experts_quant.cu

@@ -43,7 +43,7 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
                     uint32_t* input_offset_by_experts,
                     uint32_t* output_scale_offset_by_experts, int n_experts,
                     bool low_latency) {
-  using PackedVec = PackedVec<Type>;
+  using PackedVec = PackedVec<Type, CVT_FP4_PACK16>;
   static constexpr int CVT_FP4_NUM_THREADS_PER_SF =
       (CVT_FP4_SF_VEC_SIZE / CVT_FP4_ELTS_PER_THREAD);
   static_assert(sizeof(PackedVec) == sizeof(Type) * CVT_FP4_ELTS_PER_THREAD,
@@ -155,7 +155,7 @@ __global__ void __launch_bounds__(1024, VLLM_BLOCKS_PER_SM(1024))
                     float const* SFScale, uint32_t* out, uint32_t* SFout,
                     uint32_t* input_offset_by_experts,
                     uint32_t* output_scale_offset_by_experts, int n_experts) {
-  using PackedVec = PackedVec<Type>;
+  using PackedVec = PackedVec<Type, CVT_FP4_PACK16>;
   static constexpr int CVT_FP4_NUM_THREADS_PER_SF =
       (CVT_FP4_SF_VEC_SIZE / CVT_FP4_ELTS_PER_THREAD);
   static_assert(sizeof(PackedVec) == sizeof(Type) * CVT_FP4_ELTS_PER_THREAD,

csrc/quantization/fp4/nvfp4_quant_kernels.cu

@@ -42,7 +42,7 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
                     Type const* __restrict__ in,
                     float const* __restrict__ SFScale,
                     uint32_t* __restrict__ out, uint32_t* __restrict__ SFout) {
-  using PackedVec = vllm::PackedVec<Type>;
+  using PackedVec = vllm::PackedVec<Type, CVT_FP4_PACK16>;
 
   static constexpr int CVT_FP4_NUM_THREADS_PER_SF =
       (CVT_FP4_SF_VEC_SIZE / CVT_FP4_ELTS_PER_THREAD);
@@ -71,13 +71,13 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
       // If we are outside valid rows OR outside valid columns -> Use Zeros
       bool valid = (rowIdx < numRows) && (elem_idx < numCols);
       if constexpr (CVT_FP4_PACK16) {
-        ld256_or_zero_cg_u32<Type>(
-            in_vec, &reinterpret_cast<const uint32_t*>(in)[inOffset * 8],
-            valid);
+        ld256_cg_or_zero(reinterpret_cast<u32x8_t&>(in_vec),
+                         &reinterpret_cast<const uint32_t*>(in)[inOffset * 8],
+                         valid);
       } else {
-        ld128_or_zero_cg_u32<Type>(
-            in_vec, &reinterpret_cast<const uint32_t*>(in)[inOffset * 4],
-            valid);
+        ld128_cg_or_zero(reinterpret_cast<uint4&>(in_vec),
+                         &reinterpret_cast<const uint32_t*>(in)[inOffset * 4],
+                         valid);
       }
 
       auto sf_out =
@@ -109,11 +109,12 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
 template <class Type, bool UE8M0_SF = false>
 __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
     cvt_fp16_to_fp4_sf_major(int32_t numRows, int32_t numCols,
-                             int32_t sf_n_unpadded, Type const* __restrict__ in,
+                             int32_t sf_n_unpadded, int32_t num_packed_cols,
+                             Type const* __restrict__ in,
                              float const* __restrict__ SFScale,
                              uint32_t* __restrict__ out,
                              uint32_t* __restrict__ SFout) {
-  using PackedVec = PackedVec<Type>;
+  using PackedVec = PackedVec<Type, CVT_FP4_PACK16>;
 
   static constexpr int CVT_FP4_NUM_THREADS_PER_SF =
       (CVT_FP4_SF_VEC_SIZE / CVT_FP4_ELTS_PER_THREAD);
@@ -131,20 +132,20 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
   // Iterate over all rows and cols including padded ones -
   //  ensures we visit every single scale factor address to initialize it.
   for (int rowIdx = blockIdx.x; rowIdx < numRows; rowIdx += gridDim.x) {
-    if (colIdx < sf_n_unpadded) {
+    if (colIdx < num_packed_cols) {
       PackedVec in_vec;
       int64_t inOffset = rowIdx * (numCols / CVT_FP4_ELTS_PER_THREAD) + colIdx;
 
       // If we are outside valid rows OR outside valid columns -> Use Zeros
       bool valid = (rowIdx < numRows) && (elem_idx < numCols);
       if constexpr (CVT_FP4_PACK16) {
-        ld256_or_zero_cg_u32<Type>(
-            in_vec, &reinterpret_cast<const uint32_t*>(in)[inOffset * 8],
-            valid);
+        ld256_cg_or_zero(reinterpret_cast<u32x8_t&>(in_vec),
+                         &reinterpret_cast<const uint32_t*>(in)[inOffset * 8],
+                         valid);
       } else {
-        ld128_or_zero_cg_u32<Type>(
-            in_vec, &reinterpret_cast<const uint32_t*>(in)[inOffset * 4],
-            valid);
+        ld128_cg_or_zero(reinterpret_cast<uint4&>(in_vec),
+                         &reinterpret_cast<const uint32_t*>(in)[inOffset * 4],
+                         valid);
       }
 
       auto sf_out =
@@ -222,7 +223,8 @@ void scaled_fp4_quant_sm1xxa(torch::Tensor const& output,
           reinterpret_cast<uint32_t*>(sf_out));
     });
   } else {
-    int grid_y = vllm::div_round_up(sf_n_unpadded, static_cast<int>(block.x));
+    int num_packed_cols = n / CVT_FP4_ELTS_PER_THREAD;
+    int grid_y = vllm::div_round_up(num_packed_cols, static_cast<int>(block.x));
     int grid_x = std::min(
         m, std::max(1, (multiProcessorCount * numBlocksPerSM) / grid_y));
     dim3 grid(grid_x, grid_y);
@@ -232,8 +234,8 @@ void scaled_fp4_quant_sm1xxa(torch::Tensor const& output,
       auto input_ptr = static_cast<cuda_type const*>(input.data_ptr());
       // NOTE: We don't support e8m0 scales at this moment.
       vllm::cvt_fp16_to_fp4_sf_major<cuda_type, false>
-          <<<grid, block, 0, stream>>>(m, n, sf_n_unpadded, input_ptr,
-                                       input_sf_ptr,
+          <<<grid, block, 0, stream>>>(m, n, sf_n_unpadded, num_packed_cols,
+                                       input_ptr, input_sf_ptr,
                                        reinterpret_cast<uint32_t*>(output_ptr),
                                        reinterpret_cast<uint32_t*>(sf_out));
     });

csrc/quantization/fp4/nvfp4_utils.cuh

@@ -19,8 +19,10 @@
 #include <cuda_runtime.h>
 #include <cuda_fp8.h>
 
-#if (defined(NVFP4_ENABLE_ELTS16) && (CUDART_VERSION >= 12090) && \
-     defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100)
+#include "../../cuda_vec_utils.cuh"
+
+#if defined(NVFP4_ENABLE_ELTS16) && defined(CUDA_VERSION) && \
+    CUDA_VERSION >= 12090
   #define ELTS_PER_THREAD 16
 constexpr int CVT_FP4_ELTS_PER_THREAD = 16;
 constexpr bool CVT_FP4_PACK16 = true;
@@ -34,68 +36,6 @@ constexpr int CVT_FP4_SF_VEC_SIZE = 16;
 
 namespace vllm {
 
-// Convert PyTorch cpp type to CUDA type
-template <typename T>
-struct CUDATypeConverter {
-  using Type = T;
-};
-
-template <>
-struct CUDATypeConverter<at::Half> {
-  using Type = half;
-};
-
-template <>
-struct CUDATypeConverter<at::BFloat16> {
-  using Type = __nv_bfloat16;
-};
-
-// Get type2 from type or vice versa (applied to half and bfloat16)
-template <typename T>
-struct TypeConverter {
-  using Type = half2;
-};  // keep for generality
-
-template <>
-struct TypeConverter<half2> {
-  using Type = half;
-};
-
-template <>
-struct TypeConverter<half> {
-  using Type = half2;
-};
-
-template <>
-struct TypeConverter<__nv_bfloat162> {
-  using Type = __nv_bfloat16;
-};
-
-template <>
-struct TypeConverter<__nv_bfloat16> {
-  using Type = __nv_bfloat162;
-};
-
-#if (defined(NVFP4_ENABLE_ELTS16) && (CUDART_VERSION >= 12090) && \
-     defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100)
-// Define a 32 bytes packed data type.
-template <class Type>
-struct alignas(32) PackedVec {
-  typename TypeConverter<Type>::Type elts[8];
-};
-#else
-// Define a 16 bytes packed data type.
-template <class Type>
-struct alignas(16) PackedVec {
-  typename TypeConverter<Type>::Type elts[4];
-};
-#endif
-
-template <>
-struct PackedVec<__nv_fp8_e4m3> {
-  __nv_fp8x2_e4m3 elts[8];
-};
-
 template <typename Int>
 __host__ __device__ inline Int round_up(Int x, Int y) {
   static_assert(std::is_integral_v<Int>,
@@ -208,56 +148,6 @@ __device__ __forceinline__ float reciprocal_approximate_ftz(float a) {
   return b;
 }
 
-template <class Type>
-__device__ __forceinline__ void ld128_or_zero_cg_u32(PackedVec<Type>& out,
-                                                     const void* ptr,
-                                                     bool pred) {
-  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));
-
-  *reinterpret_cast<uint4*>(&out) = uint4{r0, r1, r2, r3};
-}
-
-template <class Type>
-__device__ __forceinline__ void ld256_or_zero_cg_u32(PackedVec<Type>& out,
-                                                     const void* ptr,
-                                                     bool pred) {
-  uint32_t r0, r1, r2, r3, r4, r5, r6, r7;
-
-  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"(r0), "=r"(r1), "=r"(r2), "=r"(r3), "=r"(r4), "=r"(r5), "=r"(r6),
-        "=r"(r7)
-      : "r"((int)pred), "l"(ptr));
-
-  reinterpret_cast<uint4*>(&out)[0] = uint4{r0, r1, r2, r3};
-  reinterpret_cast<uint4*>(&out)[1] = uint4{r4, r5, r6, r7};
-}
-
 // Compute SF output offset for swizzled tensor core layout.
 // SF layout: [numMTiles, numKTiles, 32, 4, 4]
 // Caller must precompute: numKTiles = (numCols + 63) / 64
@@ -315,8 +205,8 @@ __device__ __forceinline__ uint8_t* sf_out_rowmajor_u8(int row, int pack,
 
 // Quantizes the provided PackedVec into the uint32_t output
 template <class Type, int CVT_FP4_NUM_THREADS_PER_SF, bool UE8M0_SF = false>
-__device__ __forceinline__ fp4_packed_t
-cvt_warp_fp16_to_fp4(PackedVec<Type>& vec, float SFScaleVal, uint8_t* SFout) {
+__device__ __forceinline__ fp4_packed_t cvt_warp_fp16_to_fp4(
+    PackedVec<Type, CVT_FP4_PACK16>& vec, float SFScaleVal, uint8_t* SFout) {
   // Get absolute maximum values among the local 8 values.
   auto localMax = __habs2(vec.elts[0]);
 
@@ -372,11 +262,7 @@ cvt_warp_fp16_to_fp4(PackedVec<Type>& vec, float SFScaleVal, uint8_t* SFout) {
 
 #pragma unroll
   for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) {
-    if constexpr (std::is_same_v<Type, half>) {
-      fp2Vals[i] = __half22float2(vec.elts[i]);
-    } else {
-      fp2Vals[i] = __bfloat1622float2(vec.elts[i]);
-    }
+    fp2Vals[i] = cast_to_float2(vec.elts[i]);
     fp2Vals[i].x *= outputScale;
     fp2Vals[i].y *= outputScale;
   }
@@ -395,22 +281,19 @@ __device__ __forceinline__ float2 silu2(float2 x) {
 }
 
 template <class Type>
-__inline__ __device__ PackedVec<Type> compute_silu_mul(
-    const PackedVec<Type>& x_vec, const PackedVec<Type>& y_vec) {
-  PackedVec<Type> result;
+__inline__ __device__ PackedVec<Type, CVT_FP4_PACK16> compute_silu_mul(
+    const PackedVec<Type, CVT_FP4_PACK16>& x_vec,
+    const PackedVec<Type, CVT_FP4_PACK16>& y_vec) {
+  PackedVec<Type, CVT_FP4_PACK16> result;
 
 #pragma unroll
   for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 2; ++i) {
     // silu_mul in float32
-    if constexpr (std::is_same_v<Type, half>) {
-      float2 silu_vec = silu2(__half22float2(x_vec.elts[i]));
-      result.elts[i] = __float22half2_rn(
-          __fmul2_rn(silu_vec, __half22float2(y_vec.elts[i])));
-    } else {
-      float2 silu_vec = silu2(__bfloat1622float2(x_vec.elts[i]));
-      result.elts[i] = __float22bfloat162_rn(
-          __fmul2_rn(silu_vec, __bfloat1622float2(y_vec.elts[i])));
-    }
+    using packed_t = typename PackedTypeConverter<Type>::Type;
+    float2 silu_vec = silu2(cast_to_float2(x_vec.elts[i]));
+    float2 y_f2 = cast_to_float2(y_vec.elts[i]);
+    result.elts[i] = cast_to_packed<packed_t>(
+        make_float2(silu_vec.x * y_f2.x, silu_vec.y * y_f2.y));
   }
   return result;
 }

csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm120_fp8_dispatch.cuh

@@ -12,6 +12,68 @@ namespace vllm {
 
 using c3x::cutlass_gemm_caller;
 
+// Custom wrapper to allow specifying EpilogueTile for small M
+template <typename ElementAB_, typename ElementD_,
+          template <typename, typename, typename> typename Epilogue_,
+          typename TileShape, typename ClusterShape, typename KernelSchedule,
+          typename EpilogueSchedule, typename EpilogueTile>
+struct cutlass_3x_gemm_sm120_custom {
+  using ElementAB = ElementAB_;
+  using LayoutA = cutlass::layout::RowMajor;
+  static constexpr int AlignmentA =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+
+  using LayoutB = cutlass::layout::ColumnMajor;
+  static constexpr int AlignmentB =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+
+  using ElementC = void;
+  using LayoutC = cutlass::layout::RowMajor;
+  static constexpr int AlignmentC =
+      128 / cutlass::sizeof_bits<ElementD_>::value;
+
+  using ElementD = ElementD_;
+  using LayoutD = cutlass::layout::RowMajor;
+  static constexpr int AlignmentD = AlignmentC;
+
+  using ElementAcc =
+      typename std::conditional<std::is_same_v<ElementAB, int8_t>, int32_t,
+                                float>::type;
+  using Epilogue = Epilogue_<ElementAcc, ElementD, TileShape>;
+
+  // MMA type
+  using ElementAccumulator = float;
+
+  // Epilogue types
+  using ElementBias = cutlass::half_t;
+  using ElementCompute = float;
+  using ElementAux = ElementD;
+  using LayoutAux = LayoutD;
+  using ElementAmax = float;
+
+  using EVTCompute = typename Epilogue::EVTCompute;
+
+  using CollectiveEpilogue =
+      typename cutlass::epilogue::collective::CollectiveBuilder<
+          cutlass::arch::Sm120, cutlass::arch::OpClassTensorOp, TileShape,
+          ClusterShape, EpilogueTile,  // Use custom EpilogueTile
+          ElementAccumulator, ElementCompute, ElementC, LayoutC, AlignmentC,
+          ElementD, LayoutD, AlignmentD, EpilogueSchedule,
+          EVTCompute>::CollectiveOp;
+
+  using CollectiveMainloop =
+      typename cutlass::gemm::collective::CollectiveBuilder<
+          cutlass::arch::Sm120, cutlass::arch::OpClassTensorOp, ElementAB,
+          LayoutA, AlignmentA, ElementAB, LayoutB, AlignmentB,
+          ElementAccumulator, TileShape, ClusterShape,
+          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
+              sizeof(typename CollectiveEpilogue::SharedStorage))>,
+          KernelSchedule, void>::CollectiveOp;
+
+  using GemmKernel = enable_sm120_only<cutlass::gemm::kernel::GemmUniversal<
+      Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>>;
+};
+
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue>
 struct sm120_fp8_config_default {
@@ -25,6 +87,54 @@ struct sm120_fp8_config_default {
                             KernelSchedule, EpilogueSchedule>;
 };
 
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm120_fp8_config_M64 {
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  // SM120 Cooperative kernel requires Tile M >= 128.
+  // For M=64 tile, we use Pingpong schedule which is more flexible with small
+  // tiles.
+  using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecializedPingpong;
+  using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
+  using TileShape = Shape<_64, _64, _128>;
+  // CUTLASS 3.x on SM120 currently restricts programmatic multicast (Cluster >
+  // 1) for certain schedules/types. Reverting to 1x1x1 to ensure compilation.
+  using ClusterShape = Shape<_1, _1, _1>;
+  using Cutlass3xGemm =
+      cutlass_3x_gemm_sm120<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm120_fp8_config_M32 {
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecializedPingpong;
+  using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
+  using TileShape = Shape<_32, _64, _128>;
+  using ClusterShape = Shape<_1, _1, _1>;
+  // Use custom gemm to specify EpilogueTile M=32
+  using Cutlass3xGemm =
+      cutlass_3x_gemm_sm120_custom<InType, OutType, Epilogue, TileShape,
+                                   ClusterShape, KernelSchedule,
+                                   EpilogueSchedule, Shape<_32, _32>>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm120_fp8_config_M16 {
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule = cutlass::gemm::KernelTmaWarpSpecializedPingpong;
+  using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
+  using TileShape = Shape<_16, _64, _128>;
+  using ClusterShape = Shape<_1, _1, _1>;
+  // Use custom gemm to specify EpilogueTile M=16
+  using Cutlass3xGemm =
+      cutlass_3x_gemm_sm120_custom<InType, OutType, Epilogue, TileShape,
+                                   ClusterShape, KernelSchedule,
+                                   EpilogueSchedule, Shape<_16, _32>>;
+};
+
 template <typename InType, typename OutType,
           template <typename, typename, typename> typename Epilogue,
           typename... EpilogueArgs>
@@ -36,6 +146,28 @@ inline void cutlass_gemm_sm120_fp8_dispatch(torch::Tensor& out,
   TORCH_CHECK(a.dtype() == torch::kFloat8_e4m3fn);
   TORCH_CHECK(b.dtype() == torch::kFloat8_e4m3fn);
 
+  int M = a.size(0);
+
+  if (M <= 16) {
+    using Cutlass3xGemmM16 =
+        typename sm120_fp8_config_M16<InType, OutType, Epilogue>::Cutlass3xGemm;
+    return cutlass_gemm_caller<Cutlass3xGemmM16>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  }
+  if (M <= 32) {
+    using Cutlass3xGemmM32 =
+        typename sm120_fp8_config_M32<InType, OutType, Epilogue>::Cutlass3xGemm;
+    return cutlass_gemm_caller<Cutlass3xGemmM32>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  }
+
+  if (M <= 256) {
+    using Cutlass3xGemmM64 =
+        typename sm120_fp8_config_M64<InType, OutType, Epilogue>::Cutlass3xGemm;
+    return cutlass_gemm_caller<Cutlass3xGemmM64>(
+        out, a, b, std::forward<EpilogueArgs>(args)...);
+  }
+
   using Cutlass3xGemmDefault =
       typename sm120_fp8_config_default<InType, OutType,
                                         Epilogue>::Cutlass3xGemm;
@@ -64,4 +196,4 @@ void cutlass_scaled_mm_sm120_fp8_epilogue(torch::Tensor& out,
   }
 }
 
-}  // namespace vllm
+}  // namespace vllm

csrc/quantization/w8a8/cutlass/moe/moe_data.cu

@@ -263,12 +263,10 @@ void get_cutlass_moe_mm_data_caller(
 }
 
 template <bool SWAP_AB>
-__global__ void compute_pplx_data(int32_t* expert_offsets,
-                                  int32_t* problem_sizes1,
-                                  int32_t* problem_sizes2,
-                                  const int32_t* __restrict__ expert_num_tokens,
-                                  const int padded_m, const int n,
-                                  const int k) {
+__global__ void compute_batched_moe_data(
+    int32_t* expert_offsets, int32_t* problem_sizes1, int32_t* problem_sizes2,
+    const int32_t* __restrict__ expert_num_tokens, const int padded_m,
+    const int n, const int k) {
   int expert_idx = threadIdx.x;
   expert_offsets[expert_idx] = expert_idx * padded_m;
 
@@ -289,24 +287,22 @@ __global__ void compute_pplx_data(int32_t* expert_offsets,
   }
 }
 
-void get_cutlass_pplx_moe_mm_data_caller(torch::Tensor& expert_offsets,
-                                         torch::Tensor& problem_sizes1,
-                                         torch::Tensor& problem_sizes2,
-                                         const torch::Tensor& expert_num_tokens,
-                                         const int64_t num_local_experts,
-                                         const int64_t padded_m,
-                                         const int64_t n, const int64_t k) {
+void get_cutlass_batched_moe_mm_data_caller(
+    torch::Tensor& expert_offsets, torch::Tensor& problem_sizes1,
+    torch::Tensor& problem_sizes2, const torch::Tensor& expert_num_tokens,
+    const int64_t num_local_experts, const int64_t padded_m, const int64_t n,
+    const int64_t k) {
   auto stream = at::cuda::getCurrentCUDAStream(expert_offsets.device().index());
 
   if (num_local_experts * padded_m > SWAP_AB_THRESHOLD) {
-    compute_pplx_data<false><<<1, num_local_experts, 0, stream>>>(
+    compute_batched_moe_data<false><<<1, num_local_experts, 0, stream>>>(
         static_cast<int32_t*>(expert_offsets.data_ptr()),
         static_cast<int32_t*>(problem_sizes1.data_ptr()),
         static_cast<int32_t*>(problem_sizes2.data_ptr()),
         static_cast<const int32_t*>(expert_num_tokens.data_ptr()), padded_m, n,
         k);
   } else {
-    compute_pplx_data<true><<<1, num_local_experts, 0, stream>>>(
+    compute_batched_moe_data<true><<<1, num_local_experts, 0, stream>>>(
         static_cast<int32_t*>(expert_offsets.data_ptr()),
         static_cast<int32_t*>(problem_sizes1.data_ptr()),
         static_cast<int32_t*>(problem_sizes2.data_ptr()),

csrc/quantization/w8a8/cutlass/scaled_mm_entry.cu

@@ -82,13 +82,11 @@ void get_cutlass_moe_mm_problem_sizes_from_expert_offsets_caller(
     torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
     const int64_t n, const int64_t k, const bool swap_ab);
 
-void get_cutlass_pplx_moe_mm_data_caller(torch::Tensor& expert_offsets,
-                                         torch::Tensor& problem_sizes1,
-                                         torch::Tensor& problem_sizes2,
-                                         const torch::Tensor& expert_num_tokens,
-                                         const int64_t num_local_experts,
-                                         const int64_t padded_m,
-                                         const int64_t n, const int64_t k);
+void get_cutlass_batched_moe_mm_data_caller(
+    torch::Tensor& expert_offsets, torch::Tensor& problem_sizes1,
+    torch::Tensor& problem_sizes2, const torch::Tensor& expert_num_tokens,
+    const int64_t num_local_experts, const int64_t padded_m, const int64_t n,
+    const int64_t k);
 #endif
 
 void cutlass_scaled_mm_azp_sm75(torch::Tensor& c, torch::Tensor const& a,
@@ -319,29 +317,30 @@ void get_cutlass_moe_mm_problem_sizes_from_expert_offsets(
       version_num, ". Required capability: 90, 100, or 120");
 }
 
-void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
-                                  torch::Tensor& problem_sizes1,
-                                  torch::Tensor& problem_sizes2,
-                                  const torch::Tensor& expert_num_tokens,
-                                  const int64_t num_local_experts,
-                                  const int64_t padded_m, const int64_t n,
-                                  const int64_t k) {
+void get_cutlass_batched_moe_mm_data(torch::Tensor& expert_offsets,
+                                     torch::Tensor& problem_sizes1,
+                                     torch::Tensor& problem_sizes2,
+                                     const torch::Tensor& expert_num_tokens,
+                                     const int64_t num_local_experts,
+                                     const int64_t padded_m, const int64_t n,
+                                     const int64_t k) {
   // This function currently gets compiled only if we have a valid cutlass moe
   // mm to run it for.
   int32_t version_num = get_sm_version_num();
 #if (defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90) ||   \
     (defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100) || \
     (defined ENABLE_CUTLASS_MOE_SM120 && ENABLE_CUTLASS_MOE_SM120)
-  get_cutlass_pplx_moe_mm_data_caller(expert_offsets, problem_sizes1,
-                                      problem_sizes2, expert_num_tokens,
-                                      num_local_experts, padded_m, n, k);
+  get_cutlass_batched_moe_mm_data_caller(expert_offsets, problem_sizes1,
+                                         problem_sizes2, expert_num_tokens,
+                                         num_local_experts, padded_m, n, k);
   return;
 #endif
-  TORCH_CHECK_NOT_IMPLEMENTED(
-      false,
-      "No compiled get_cutlass_pplx_moe_mm_data: no cutlass_scaled_mm kernel "
-      "for CUDA device capability: ",
-      version_num, ". Required capability: 90, 100, or 120");
+  TORCH_CHECK_NOT_IMPLEMENTED(false,
+                              "No compiled get_cutlass_batched_moe_mm_data: no "
+                              "cutlass_scaled_mm kernel "
+                              "for CUDA device capability: ",
+                              version_num,
+                              ". Required capability: 90, 100, or 120");
 }
 
 void cutlass_scaled_mm_azp(torch::Tensor& c, torch::Tensor const& a,

csrc/torch_bindings.cpp

@@ -242,7 +242,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // DeepSeek V3 fused A GEMM (SM 9.0+, bf16 only, 1-16 tokens).
   ops.def(
       "dsv3_fused_a_gemm(Tensor! output, Tensor mat_a, Tensor mat_b) -> ()");
-  ops.impl("dsv3_fused_a_gemm", torch::kCUDA, &dsv3_fused_a_gemm);
+  // conditionally compiled so impl registration is in source file
 
   // Quantized GEMM for AWQ.
   ops.def(
@@ -426,6 +426,22 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       " Tensor problem_sizes, Tensor expert_offsets, Tensor sf_offsets) -> ()");
   // conditionally compiled so impl registration is in source file
 
+  // Expert-specialization mxfp8 blockscaled grouped quantization (SM100+).
+  ops.def(
+      "mxfp8_experts_quant("
+      " Tensor input, Tensor problem_sizes, Tensor expert_offsets,"
+      " Tensor blockscale_offsets, Tensor! quant_output, Tensor! scale_factor)"
+      " -> ()");
+  // conditionally compiled so impl registration is in source file
+
+  // Expert-specialization mxfp8 blockscaled grouped GEMM (SM100+).
+  ops.def(
+      "cutlass_mxfp8_grouped_mm("
+      " Tensor a, Tensor b, Tensor sfa, Tensor sfb, Tensor! out,"
+      " Tensor problem_sizes, Tensor expert_offsets, Tensor blockscale_offsets)"
+      " -> ()");
+  // conditionally compiled so impl registration is in source file
+
   // CUTLASS w8a8 GEMM, supporting symmetric per-tensor or per-row/column
   // quantization, as well as bias
   ops.def(
@@ -489,19 +505,19 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
            &get_cutlass_moe_mm_problem_sizes_from_expert_offsets);
 
   // A function that computes data required to run fused MoE with w8a8 grouped
-  // GEMM and PPLX. It takes expert_num_tokens and non_zero_expert_idxs
+  // GEMM in batched expert format. It takes expert_num_tokens
   // as an input, and computes expert_offsets (token start indices of each
   // expert). In addition to this, it computes problem sizes for each expert's
   // multiplication used by the two mms called from fused MoE operation.
   ops.def(
-      "get_cutlass_pplx_moe_mm_data(Tensor! expert_offsets, "
+      "get_cutlass_batched_moe_mm_data(Tensor! expert_offsets, "
       "                             Tensor! problem_sizes1, "
       "                             Tensor! problem_sizes2, "
       "                             Tensor expert_num_tokens, "
       "                             int num_local_experts, int padded_m, "
       "                             int n, int k) -> ()");
-  ops.impl("get_cutlass_pplx_moe_mm_data", torch::kCUDA,
-           &get_cutlass_pplx_moe_mm_data);
+  ops.impl("get_cutlass_batched_moe_mm_data", torch::kCUDA,
+           &get_cutlass_batched_moe_mm_data);
 
   // Check if cutlass scaled_mm supports block quantization (used by DeepSeekV3)
   ops.def(
@@ -640,7 +656,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "int block_size,"
       "Tensor? block_idx_first_scheduled_token,"
       "Tensor? block_idx_last_scheduled_token,"
-      "Tensor? initial_state_idx) -> ()");
+      "Tensor? initial_state_idx,"
+      "Tensor? cu_chunk_seqlen,"
+      "Tensor? last_chunk_indices) -> ()");
   ops.impl("selective_scan_fwd", torch::kCUDA, &selective_scan_fwd);
 
   // Hadamard transforms

docker/Dockerfile

@@ -132,8 +132,10 @@ ENV UV_LINK_MODE=copy
 # Verify GCC version
 RUN gcc --version
 
-# Ensure CUDA compatibility library is loaded
-RUN echo "/usr/local/cuda-$(echo "$CUDA_VERSION" | cut -d. -f1,2)/compat/" > /etc/ld.so.conf.d/cuda-compat.conf && ldconfig
+# Enable CUDA forward compatibility by setting '-e VLLM_ENABLE_CUDA_COMPATIBILITY=1'
+# Only needed for datacenter/professional GPUs with older drivers.
+# See: https://docs.nvidia.com/deploy/cuda-compatibility/
+ENV VLLM_ENABLE_CUDA_COMPATIBILITY=0
 
 # ============================================================
 # SLOW-CHANGING DEPENDENCIES BELOW
@@ -260,7 +262,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 # Build the vLLM wheel
 # if USE_SCCACHE is set, use sccache to speed up compilation
+# AWS credentials mounted at ~/.aws/credentials for sccache S3 auth (optional)
 RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=secret,id=aws-credentials,target=/root/.aws/credentials,required=false \
     if [ "$USE_SCCACHE" = "1" ]; then \
         echo "Installing sccache..." \
         && case "${TARGETPLATFORM}" in \
@@ -306,7 +310,7 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
 #################### CSRC BUILD IMAGE ####################
 
 #################### EXTENSIONS BUILD IMAGE ####################
-# Build DeepGEMM, pplx-kernels, DeepEP - runs in PARALLEL with csrc-build
+# Build DeepGEMM, DeepEP - runs in PARALLEL with csrc-build
 # This stage is independent and doesn't affect csrc cache
 FROM base AS extensions-build
 ARG CUDA_VERSION
@@ -333,10 +337,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Ensure the wheel dir exists so COPY won't fail when DeepGEMM is skipped
 RUN mkdir -p /tmp/deepgemm/dist && touch /tmp/deepgemm/dist/.deepgemm_skipped
 
-# Build pplx-kernels and DeepEP wheels
+# Build DeepEP wheels
 COPY tools/ep_kernels/install_python_libraries.sh /tmp/install_python_libraries.sh
 # Defaults moved here from tools/ep_kernels/install_python_libraries.sh for centralized version management
-ARG PPLX_COMMIT_HASH=12cecfd
 ARG DEEPEP_COMMIT_HASH=73b6ea4
 ARG NVSHMEM_VER
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -345,7 +348,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     /tmp/install_python_libraries.sh \
         --workspace /tmp/ep_kernels_workspace \
         --mode wheel \
-        ${PPLX_COMMIT_HASH:+--pplx-ref "$PPLX_COMMIT_HASH"} \
         ${DEEPEP_COMMIT_HASH:+--deepep-ref "$DEEPEP_COMMIT_HASH"} \
         ${NVSHMEM_VER:+--nvshmem-ver "$NVSHMEM_VER"} && \
     find /tmp/ep_kernels_workspace/nvshmem -name '*.a' -delete
@@ -560,8 +562,10 @@ ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 ENV UV_LINK_MODE=copy
 
-# Ensure CUDA compatibility library is loaded
-RUN echo "/usr/local/cuda-$(echo "$CUDA_VERSION" | cut -d. -f1,2)/compat/" > /etc/ld.so.conf.d/cuda-compat.conf && ldconfig
+# Enable CUDA forward compatibility by setting '-e VLLM_ENABLE_CUDA_COMPATIBILITY=1'
+# Only needed for datacenter/professional GPUs with older drivers.
+# See: https://docs.nvidia.com/deploy/cuda-compatibility/
+ENV VLLM_ENABLE_CUDA_COMPATIBILITY=0
 
 # ============================================================
 # SLOW-CHANGING DEPENDENCIES BELOW
@@ -672,7 +676,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Pytorch now installs NVSHMEM, setting LD_LIBRARY_PATH
 ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
 
-# Install EP kernels wheels (pplx-kernels and DeepEP) that have been built in the `build` stage
+# Install EP kernels wheels (DeepEP) that have been built in the `build` stage
 RUN --mount=type=bind,from=build,src=/tmp/ep_kernels_workspace/dist,target=/vllm-workspace/ep_kernels/dist \
     --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system ep_kernels/dist/*.whl --verbose \

docker/Dockerfile.rocm

@@ -305,6 +305,14 @@ RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
 RUN --mount=type=bind,from=build_rixl,src=/app/install,target=/rixl_install \
     uv pip install --system /rixl_install/*.whl
 
+# RIXL/MoRIIO runtime dependencies (RDMA userspace libraries)
+RUN apt-get update -q -y && apt-get install -q -y \
+    librdmacm1 \
+    libibverbs1 \
+    ibverbs-providers \
+    ibverbs-utils \
+    && rm -rf /var/lib/apt/lists/*
+
 WORKDIR /vllm-workspace
 ARG COMMON_WORKDIR
 COPY --from=build_vllm ${COMMON_WORKDIR}/vllm /vllm-workspace
@@ -330,6 +338,11 @@ RUN bash /tmp/install_torchcodec.sh \
 # Copy in the v1 package (for python-only install test group)
 COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1
 
+# Set MIOPEN ENVS to resolve performance regressions in MIOpen 3D convolution kernel
+# See: https://github.com/pytorch/pytorch/issues/169857
+ENV MIOPEN_DEBUG_CONV_DIRECT=0
+ENV MIOPEN_DEBUG_CONV_GEMM=0
+
 # Source code is used in the `python_only_compile.sh` test
 # We hide it inside `src/` so that this source code
 # will not be imported by other tests

docker/Dockerfile.xpu

@@ -6,8 +6,7 @@ ARG PYTHON_VERSION=3.12
 ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/xpu"
 
 RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
-    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \
-    add-apt-repository -y ppa:kobuk-team/intel-graphics
+    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list
 
 RUN apt clean && apt-get update -y && \
     apt-get install -y --no-install-recommends --fix-missing \
@@ -28,9 +27,22 @@ RUN apt clean && apt-get update -y && \
     python3-pip
 
 RUN apt update && apt upgrade -y && \
-    apt install -y libze1 libze-dev libze-intel-gpu1 intel-opencl-icd libze-intel-gpu-raytracing intel-ocloc && \
     apt install -y intel-oneapi-compiler-dpcpp-cpp-2025.3
 
+# Install UMD
+RUN mkdir neo && \
+    cd neo && \
+    wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.24.8/intel-igc-core-2_2.24.8+20344_amd64.deb && \
+    wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.24.8/intel-igc-opencl-2_2.24.8+20344_amd64.deb && \
+    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/intel-ocloc_25.48.36300.8-0_amd64.deb && \
+    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/intel-opencl-icd_25.48.36300.8-0_amd64.deb && \
+    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/libigdgmm12_22.8.2_amd64.deb && \
+    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/libze-intel-gpu1_25.48.36300.8-0_amd64.deb && \
+    wget https://github.com/oneapi-src/level-zero/releases/download/v1.26.0/level-zero_1.26.0+u24.04_amd64.deb && \
+    dpkg -i *.deb && \
+    cd .. && \
+    rm -rf neo
+
 ENV PATH="/root/.local/bin:$PATH"
 ENV VIRTUAL_ENV="/opt/venv"
 ENV UV_PYTHON_INSTALL_DIR=/opt/uv/python
@@ -103,9 +115,57 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # install development dependencies (for testing)
 RUN uv pip install -e tests/vllm_test_utils
 
-# install nixl from source code
-ENV NIXL_VERSION=0.7.0
-RUN python /workspace/vllm/tools/install_nixl_from_source_ubuntu.py
+# install NIXL and UCX from source code
+ARG UCX_VERSION=e5d98879705239d254ede40b4a52891850cb5349
+ARG NIXL_VERSION=0.7.0
+
+RUN apt-get update && apt-get install -y \
+    pciutils \
+    net-tools \
+    iproute2 \
+    hwloc \
+    numactl \
+    wget \
+    curl \
+    git \
+    build-essential \
+    autoconf \
+    automake \
+    libtool \
+    pkg-config \
+    rdma-core \
+    libibverbs-dev \
+    ibverbs-utils \
+    libibverbs1 \
+    librdmacm-dev \
+    librdmacm1 \
+    libibumad-dev \
+    libibumad3 \
+    libibmad-dev \
+    libibmad5 \
+    infiniband-diags \
+    perftest \
+    ibutils \
+    libmlx5-1 \
+    libmlx4-1 \
+    ibverbs-providers \
+    librdmacm1t64
+
+ENV PKG_CONFIG_PATH=/tmp/ucx_install/lib/pkgconfig:${PKG_CONFIG_PATH}
+ENV LD_LIBRARY_PATH=/tmp/ucx_install/lib:${LD_LIBRARY_PATH}
+RUN --mount=type=cache,target=/root/.cache/uv \
+    git clone https://github.com/openucx/ucx /tmp/ucx_source && \
+    cd /tmp/ucx_source && git checkout "${UCX_VERSION}" && \
+    bash autogen.sh && \
+    ./configure --prefix=/tmp/ucx_install --with-ze=yes --enable-examples --enable-mt && \
+    make CFLAGS="-Wno-error=incompatible-pointer-types" -j8 && make install && \
+    git clone https://github.com/ai-dynamo/nixl /tmp/nixl_source && \
+    cd /tmp/nixl_source && git checkout "${NIXL_VERSION}" && \
+    cd /tmp/nixl_source && \
+    uv pip install --upgrade meson pybind11 patchelf && \
+    uv pip install -r requirements.txt && \
+    uv pip install . && \
+    rm -rf /tmp/ucx_source /tmp/nixl_source
 
 # FIX triton
 RUN --mount=type=cache,target=/root/.cache/uv \

docker/versions.json

@@ -52,9 +52,6 @@
     "DEEPGEMM_GIT_REF": {
       "default": "477618cd51baffca09c4b0b87e97c03fe827ef03"
     },
-    "PPLX_COMMIT_HASH": {
-      "default": "12cecfd"
-    },
     "DEEPEP_COMMIT_HASH": {
       "default": "73b6ea4"
     },

docs/benchmarking/cli.md

@@ -4,6 +4,11 @@ This section guides you through running benchmark tests with the extensive datas
 
 It's a living document, updated as new features and datasets become available.
 
+!!! tip
+    The benchmarks described on this page are mainly for evaluating specific vLLM features as well as regression testing.
+
+    For benchmarking production vLLM servers, we recommend [GuideLLM](https://github.com/vllm-project/guidellm), an established performance benchmarking framework with live progress updates and automatic report generation. It is also more flexible than `vllm bench serve` in terms of dataset loading, request formatting, and workload patterns.
+
 ## Dataset Overview
 
 <style>

docs/benchmarking/sweeps.md

@@ -1,10 +1,15 @@
 # Parameter Sweeps
 
+`vllm bench sweep` is a suite of commands designed to run benchmarks across multiple configurations and compare them by visualizing the results.
+
 ## Online Benchmark
 
 ### Basic
 
-`vllm bench sweep serve` automatically starts `vllm serve` and runs `vllm bench serve` to evaluate vLLM over multiple configurations.
+`vllm bench sweep serve` starts `vllm serve` and iteratively runs `vllm bench serve` for each server configuration.
+
+!!! tip
+    If you only need to run benchmarks for a single server configuration, consider using [GuideLLM](https://github.com/vllm-project/guidellm), an established performance benchmarking framework with live progress updates and automatic report generation. It is also more flexible than `vllm bench serve` in terms of dataset loading, request formatting, and workload patterns.
 
 Follow these steps to run the script:
 
@@ -50,21 +55,24 @@ Follow these steps to run the script:
     ```json
     [
         {
+            "_benchmark_name": "scenario_A",
             "random_input_len": 128,
             "random_output_len": 32
         },
         {
+            "_benchmark_name": "scenario_B",
             "random_input_len": 256,
             "random_output_len": 64
         },
         {
+            "_benchmark_name": "scenario_C",
             "random_input_len": 512,
             "random_output_len": 128
         }
     ]
     ```
 
-5. Determine where you want to save the results, and pass that to `--output-dir`.
+5. Set `--output-dir` and optionally `--experiment-name` to control where to save the results.
 
 Example command:
 
@@ -74,9 +82,12 @@ vllm bench sweep serve \
     --bench-cmd 'vllm bench serve --model meta-llama/Llama-2-7b-chat-hf --backend vllm --endpoint /v1/completions --dataset-name sharegpt --dataset-path benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json' \
     --serve-params benchmarks/serve_hparams.json \
     --bench-params benchmarks/bench_hparams.json \
-    -o benchmarks/results
+    --output-dir benchmarks/results \
+    --experiment-name demo
 ```
 
+By default, each parameter combination is benchmarked 3 times to make the results more reliable. You can adjust the number of runs by setting `--num-runs`.
+
 !!! important
     If both `--serve-params` and `--bench-params` are passed, the script will iterate over the Cartesian product between them.
     You can use `--dry-run` to preview the commands to be run.
@@ -86,60 +97,48 @@ vllm bench sweep serve \
     In case you are using a custom `--serve-cmd`, you can override the commands used for resetting the state by setting `--after-bench-cmd`.
 
 !!! note
-    By default, each parameter combination is run 3 times to make the results more reliable. You can adjust the number of runs by setting `--num-runs`.
+    You should set `_benchmark_name` to provide a human-readable name for parameter combinations involving many variables.
+    This becomes mandatory if the file name would otherwise exceed the maximum path length allowed by the filesystem.
 
 !!! tip
-    You can use the `--resume` option to continue the parameter sweep if one of the runs failed.
-  
-### SLA auto-tuner
+    You can use the `--resume` option to continue the parameter sweep if an unexpected error occurs, e.g., timeout when connecting to HF Hub.
 
-`vllm bench sweep serve_sla` is a wrapper over `vllm bench sweep serve` that tunes either the request rate or concurrency (choose using `--sla-variable`) in order to satisfy the SLA constraints given by `--sla-params`.
+### Workload Explorer
 
-For example, to ensure E2E latency within different target values for 99% of requests:
+`vllm bench sweep serve_workload` is a variant of `vllm bench sweep serve` that explores different workload levels in order to find the tradeoff between latency and throughput. The results can also be [visualized](#visualization) to determine the feasible SLAs.
 
-```json
-[
-    {
-        "p99_e2el_ms": "<=200"
-    },
-    {
-        "p99_e2el_ms": "<=500"
-    },
-    {
-        "p99_e2el_ms": "<=1000"
-    },
-    {
-        "p99_e2el_ms": "<=2000"
-    }
-]
-```
+The workload can be expressed in terms of request rate or concurrency (choose using `--workload-var`).
 
 Example command:
 
 ```bash
-vllm bench sweep serve_sla \
+vllm bench sweep serve_workload \
     --serve-cmd 'vllm serve meta-llama/Llama-2-7b-chat-hf' \
-    --bench-cmd 'vllm bench serve --model meta-llama/Llama-2-7b-chat-hf --backend vllm --endpoint /v1/completions --dataset-name sharegpt --dataset-path benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json' \
+    --bench-cmd 'vllm bench serve --model meta-llama/Llama-2-7b-chat-hf --backend vllm --endpoint /v1/completions --dataset-name sharegpt --dataset-path benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 100' \
+    --workload-var max_concurrency \
     --serve-params benchmarks/serve_hparams.json \
     --bench-params benchmarks/bench_hparams.json \
-    --sla-params benchmarks/sla_hparams.json \
-    --sla-variable max_concurrency \
-    -o benchmarks/results
+    --num-runs 1 \
+    --output-dir benchmarks/results \
+    --experiment-name demo
 ```
 
-The algorithm for adjusting the SLA variable is as follows:
+The algorithm for exploring different workload levels can be summarized as follows:
 
-1. Run the benchmark once with maximum possible QPS, and once with minimum possible QPS. For each run, calculate the distance of the SLA metrics from their targets, resulting in data points of QPS vs SLA distance.
-2. Perform spline interpolation between the data points to estimate the QPS that results in zero SLA distance.
-3. Run the benchmark with the estimated QPS and add the resulting data point to the history.
-4. Repeat Steps 2 and 3 until the maximum QPS that passes SLA and the minimum QPS that fails SLA in the history are close enough to each other.
+1. Run the benchmark by sending requests one at a time (serial inference, lowest workload). This results in the lowest possible latency and throughput.
+2. Run the benchmark by sending all requests at once (batch inference, highest workload). This results in the highest possible latency and throughput.
+3. Estimate the value of `workload_var` corresponding to Step 2.
+4. Run the benchmark over intermediate values of `workload_var` uniformly using the remaining iterations.
 
-!!! important
-    SLA tuning is applied over each combination of `--serve-params`, `--bench-params`, and `--sla-params`.
+You can override the number of iterations in the algorithm by setting `--workload-iters`.
 
-    For a given combination of `--serve-params` and `--bench-params`, we share the benchmark results across `--sla-params` to avoid rerunning benchmarks with the same SLA variable value.
+!!! tip
+    This is our equivalent of [GuideLLM's `--profile sweep`](https://github.com/vllm-project/guidellm/blob/v0.5.3/src/guidellm/benchmark/profiles.py#L575).
 
-### Startup
+    In general, `--workload-var max_concurrency` produces more reliable results because it directly controls the workload imposed on the vLLM engine.
+    Nevertheless, we default to `--workload-var request_rate` to maintain similar behavior as GuideLLM.
+
+## Startup Benchmark
 
 `vllm bench sweep startup` runs `vllm bench startup` across parameter combinations to compare cold/warm startup time for different engine settings.
 
@@ -189,7 +188,8 @@ vllm bench sweep startup \
     --startup-cmd 'vllm bench startup --model Qwen/Qwen3-0.6B' \
     --serve-params benchmarks/serve_hparams.json \
     --startup-params benchmarks/startup_hparams.json \
-    -o benchmarks/results
+    --output-dir benchmarks/results \
+    --experiment-name demo
 ```
 
 !!! important
@@ -202,15 +202,36 @@ vllm bench sweep startup \
 
 `vllm bench sweep plot` can be used to plot performance curves from parameter sweep results.
 
-Example command:
+Control the variables to plot via `--var-x` and `--var-y`, optionally applying `--filter-by` and `--bin-by` to the values. The plot is organized according to `--fig-by`, `--row-by`, `--col-by`, and `--curve-by`.
+
+Example commands for visualizing [Workload Explorer](#workload-explorer) results:
 
 ```bash
-vllm bench sweep plot benchmarks/results/<timestamp> \
+EXPERIMENT_DIR=${1:-"benchmarks/results/demo"}
+
+# Latency increases as the workload increases
+vllm bench sweep plot $EXPERIMENT_DIR \
     --var-x max_concurrency \
-    --row-by random_input_len \
-    --col-by random_output_len \
-    --curve-by api_server_count,max_num_batched_tokens \
-    --filter-by 'max_concurrency<=1024'
+    --var-y median_ttft_ms \
+    --col-by _benchmark_name \
+    --curve-by max_num_seqs,max_num_batched_tokens \
+    --fig-name latency_curve
+
+# Throughput saturates as workload increases
+vllm bench sweep plot $EXPERIMENT_DIR \
+    --var-x max_concurrency \
+    --var-y total_token_throughput \
+    --col-by _benchmark_name \
+    --curve-by max_num_seqs,max_num_batched_tokens \
+    --fig-name throughput_curve
+
+# Tradeoff between latency and throughput
+vllm bench sweep plot $EXPERIMENT_DIR \
+    --var-x total_token_throughput \
+    --var-y median_ttft_ms \
+    --col-by _benchmark_name \
+    --curve-by max_num_seqs,max_num_batched_tokens \
+    --fig-name latency_throughput
 ```
 
 !!! tip
@@ -230,6 +251,11 @@ Higher concurrency or batch size can raise GPU efficiency (per-GPU), but can add
 Example:
 
 ```bash
-vllm bench sweep plot_pareto benchmarks/results/<timestamp> \
+EXPERIMENT_DIR=${1:-"benchmarks/results/demo"}
+
+vllm bench sweep plot_pareto $EXPERIMENT_DIR \
   --label-by max_concurrency,tensor_parallel_size,pipeline_parallel_size
 ```
+
+!!! tip
+    You can use `--dry-run` to preview the figures to be plotted.

docs/cli/bench/sweep/serve_sla.md

@@ -1,9 +0,0 @@
-# vllm bench sweep serve_sla
-
-## JSON CLI Arguments
-
---8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
---8<-- "docs/generated/argparse/bench_sweep_serve_sla.inc.md"

docs/cli/bench/sweep/serve_workload.md

@@ -0,0 +1,9 @@
+# vllm bench sweep serve_workload
+
+## JSON CLI Arguments
+
+--8<-- "docs/cli/json_tip.inc.md"
+
+## Arguments
+
+--8<-- "docs/generated/argparse/bench_sweep_serve_workload.inc.md"

docs/contributing/README.md

@@ -49,7 +49,13 @@ If you are developing vLLM's Python and CUDA/C++ code, install Pytorch first:
 uv pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu129
 ```
 
-then install vLLM using:
+Then install the necessary build dependencies from `requirements/build.txt`, skipping `torch` as it was installed in the previous step:
+
+```bash
+grep -v '^torch==' requirements/build.txt | uv pip install -r -
+```
+
+Finally install vLLM using:
 
 ```bash
 uv pip install -e . --no-build-isolation

docs/design/arch_overview.md

@@ -208,9 +208,7 @@ configurations affect the class we ultimately get.
 
 The following figure shows the class hierarchy of vLLM:
 
-> <figure markdown="span">
->   ![](../assets/design/hierarchy.png){ align="center" alt="query" width="100%" }
-> </figure>
+![Class Hierarchy](../assets/design/hierarchy.png)
 
 There are several important design choices behind this class hierarchy:
 

docs/design/attention_backends.md

@@ -168,17 +168,18 @@ Priority is **1 = highest** (tried first).
 | `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ✅ | ❌ | ✅ | Decoder | 10.x |
 | `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥8.0 |
 | `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ❌ | ✅ | All | 9.x |
+| `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥10.0 |
 | `FLASH_ATTN_DIFFKV` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ✅ | Decoder | Any |
 | `FLEX_ATTENTION` |  | fp16, bf16, fp32 | `auto`, `bfloat16` | Any | Any | ❌ | ✅ | ❌ | Decoder, Encoder Only | Any |
 | `ROCM_AITER_FA` |  | fp16, bf16 | `auto` | 16, 32 | 64, 128, 256 | ❌ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_UNIFIED_ATTN` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_ATTN` |  | fp16, bf16, fp32 | `auto` | 16, 32, 544 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | N/A |
+| `ROCM_AITER_UNIFIED_ATTN` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | All | N/A |
+| `ROCM_ATTN` |  | fp16, bf16, fp32 | `auto` | 16, 32, 544 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | All | N/A |
 | `TREE_ATTN` |  | fp16, bf16 | `auto` | %16 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | Any |
 | `TRITON_ATTN` |  | fp16, bf16, fp32 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
 
 > **†** FlashInfer uses TRTLLM attention on Blackwell (SM100), which supports sinks. Disable via `--attention-config.use_trtllm_attention=0`.
 >
-> **\*** Specify the FlashAttention version via `--attention-config.flash_attn_version=2` or `3`. Default is FA3 on SM90, FA2 otherwise.
+> **\*** Specify the FlashAttention version via `--attention-config.flash_attn_version=2`, `3`, or `4`. Default is FA4 on SM100+ (Blackwell), FA3 on SM90 (Hopper), FA2 otherwise.
 
 ## MLA (Multi-head Latent Attention) Backends
 

docs/design/custom_op.md

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

docs/design/dbo.md

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

docs/design/fused_moe_modular_kernel.md

@@ -15,7 +15,7 @@ Based on the format of the input activations, FusedMoE implementations are broad
 The input activation format completely depends on the All2All Dispatch being used.
 
 * In the Contiguous variant, the All2All Dispatch returns the activations as a contiguous tensor of shape (M, K) along with TopK Ids and TopK weights of shape (M, num_topk). Look at `DeepEPHTPrepareAndFinalize` for an example.
-* In the Batched variant, the All2All Dispatch returns the activations as a tensor of shape (num_experts, max_tokens, K). Here, the activations/tokens that subscribe to the same expert are batched together. Note that not all entries of the tensor are valid. The activations tensor is typically accompanied by an `expert_num_tokens` tensor of size `num_experts`, where `expert_num_tokens[i]` indicates the number of valid tokens that subscribe to the ith expert. Look at `PplxPrepareAndFinalize` or `DeepEPLLPrepareAndFinalize` for an example.
+* In the Batched variant, the All2All Dispatch returns the activations as a tensor of shape (num_experts, max_tokens, K). Here, the activations/tokens that subscribe to the same expert are batched together. Note that not all entries of the tensor are valid. The activations tensor is typically accompanied by an `expert_num_tokens` tensor of size `num_experts`, where `expert_num_tokens[i]` indicates the number of valid tokens that subscribe to the ith expert. Look at `DeepEPLLPrepareAndFinalize` for an example.
 
 The FusedMoE operation is generally made of multiple operations, in both the Contiguous and Batched variants, as described in the diagrams below
 
@@ -37,31 +37,31 @@ The rest of the document will focus on the Contiguous / Non-Batched case. Extrap
 FusedMoEModularKernel splits the FusedMoE operation into 3 parts,
 
 1. TopKWeightAndReduce
-2. FusedMoEPrepareAndFinalize
-3. FusedMoEPermuteExpertsUnpermute
+2. FusedMoEPrepareAndFinalizeModular
+3. FusedMoEExpertsModular
 
 ### TopKWeightAndReduce
 
-The TopK Weight Application and Reduction components happen right after the Unpermute operation and before the All2All Combine. Note that the `FusedMoEPermuteExpertsUnpermute` is responsible for the Unpermute and `FusedMoEPrepareAndFinalize` is responsible for the All2All Combine. There is value in doing the TopK Weight Application and Reduction in the `FusedMoEPermuteExpertsUnpermute`. But some implementations choose to do it `FusedMoEPrepareAndFinalize`. In order to enable this flexibility, we have a TopKWeightAndReduce abstract class.
+The TopK Weight Application and Reduction components happen right after the Unpermute operation and before the All2All Combine. Note that the `FusedMoEExpertsModular` is responsible for the Unpermute and `FusedMoEPrepareAndFinalizeModular` is responsible for the All2All Combine. There is value in doing the TopK Weight Application and Reduction in the `FusedMoEExpertsModular`. But some implementations choose to do it `FusedMoEPrepareAndFinalizeModular`. In order to enable this flexibility, we have a TopKWeightAndReduce abstract class.
 
 Please find the implementations of TopKWeightAndReduce [here](../../vllm/model_executor/layers/fused_moe/topk_weight_and_reduce.py).
 
-`FusedMoEPrepareAndFinalize::finalize()` method accepts a `TopKWeightAndReduce` argument that is invoked inside the method.
-The `FusedMoEModularKernel` acts as a bridge between the `FusedMoEPermuteExpertsUnpermute` and `FusedMoEPerpareAndFinalize` implementations to determine where the TopK Weight Application and Reduction happens.
+`FusedMoEPrepareAndFinalizeModular::finalize()` method accepts a `TopKWeightAndReduce` argument that is invoked inside the method.
+The `FusedMoEModularKernel` acts as a bridge between the `FusedMoEExpertsModular` and `FusedMoEPerpareAndFinalize` implementations to determine where the TopK Weight Application and Reduction happens.
 
-* `FusedMoEPermuteExpertsUnpermute::finalize_weight_and_reduce_impl` method returns `TopKWeightAndReduceNoOp` if the `FusedMoEPermuteExpertsUnpermute` implementation does the weight application and reduction itself.
-* `FusedMoEPermuteExpertsUnpermute::finalize_weight_and_reduce_impl` method returns `TopKWeightAndReduceContiguous` / `TopKWeightAndReduceNaiveBatched` / `TopKWeightAndReduceDelegate` if the `FusedMoEPermuteExpertsUnpermute` implementation needs the `FusedMoEPrepareAndFinalize::finalize()` to do the weight application and reduction.
+* `FusedMoEExpertsModular::finalize_weight_and_reduce_impl` method returns `TopKWeightAndReduceNoOp` if the `FusedMoEExpertsModular` implementation does the weight application and reduction itself.
+* `FusedMoEExpertsModular::finalize_weight_and_reduce_impl` method returns `TopKWeightAndReduceContiguous` / `TopKWeightAndReduceNaiveBatched` / `TopKWeightAndReduceDelegate` if the `FusedMoEExpertsModular` implementation needs the `FusedMoEPrepareAndFinalizeModular::finalize()` to do the weight application and reduction.
 
-### FusedMoEPrepareAndFinalize
+### FusedMoEPrepareAndFinalizeModular
 
-The `FusedMoEPrepareAndFinalize` abstract class exposes `prepare`, `prepare_no_receive`  and `finalize` functions.
-The `prepare` function is responsible for input activation Quantization and All2All Dispatch. If implemented, The `prepare_no_receive` is like `prepare` except it does not wait to receive results from other workers.  Instead it returns a "receiver" callback that must be invoked to wait for the final results of worker. It is not required that this method is supported by all `FusedMoEPrepareAndFinalize` classes, but if it is available, it can be used to interleave work with the initial all to all communication, e.g. interleaving shared experts with fused experts.  The `finalize` function is responsible for invoking the All2All Combine. Additionally the `finalize` function may or may not do the TopK weight application and reduction (Please refer to the TopKWeightAndReduce section)
+The `FusedMoEPrepareAndFinalizeModular` abstract class exposes `prepare`, `prepare_no_receive`  and `finalize` functions.
+The `prepare` function is responsible for input activation Quantization and All2All Dispatch. If implemented, The `prepare_no_receive` is like `prepare` except it does not wait to receive results from other workers.  Instead it returns a "receiver" callback that must be invoked to wait for the final results of worker. It is not required that this method is supported by all `FusedMoEPrepareAndFinalizeModular` classes, but if it is available, it can be used to interleave work with the initial all to all communication, e.g. interleaving shared experts with fused experts.  The `finalize` function is responsible for invoking the All2All Combine. Additionally the `finalize` function may or may not do the TopK weight application and reduction (Please refer to the TopKWeightAndReduce section)
 
-![FusedMoEPrepareAndFinalize Blocks](../assets/design/fused_moe_modular_kernel/prepare_and_finalize_blocks.png)
+![FusedMoEPrepareAndFinalizeModular Blocks](../assets/design/fused_moe_modular_kernel/prepare_and_finalize_blocks.png)
 
-### FusedMoEPermuteExpertsUnpermute
+### FusedMoEExpertsModular
 
-The `FusedMoEPermuteExpertsUnpermute` class is where the crux of the MoE operations happen. The `FusedMoEPermuteExpertsUnpermute` abstract class exposes a few important functions,
+The `FusedMoEExpertsModular` class is where the crux of the MoE operations happen. The `FusedMoEExpertsModular` abstract class exposes a few important functions,
 
 * apply()
 * workspace_shapes()
@@ -81,25 +81,25 @@ The `apply` method is where the implementations perform
 
 #### workspace_shapes()
 
-The core FusedMoE implementation performs a series of operations. It would be inefficient to create output memory for each of these operations separately. To that effect, implementations are required to declare 2 workspace shapes, the workspace datatype and the FusedMoE output shape as outputs of the workspace_shapes() method. This information is used to allocate the workspace tensors and the output tensor in `FusedMoEModularKernel::forward()` and passed on to the `FusedMoEPermuteExpertsUnpermute::apply()` method. The workspaces could then be used as intermediate buffers in the FusedMoE implementation.
+The core FusedMoE implementation performs a series of operations. It would be inefficient to create output memory for each of these operations separately. To that effect, implementations are required to declare 2 workspace shapes, the workspace datatype and the FusedMoE output shape as outputs of the workspace_shapes() method. This information is used to allocate the workspace tensors and the output tensor in `FusedMoEModularKernel::forward()` and passed on to the `FusedMoEExpertsModular::apply()` method. The workspaces could then be used as intermediate buffers in the FusedMoE implementation.
 
 #### finalize_weight_and_reduce_impl()
 
-It is sometimes efficient to perform TopK weight application and Reduction inside the `FusedMoEPermuteExpertsUnpermute::apply()`. Find an example [here](https://github.com/vllm-project/vllm/pull/20228). We have a `TopKWeightAndReduce` abstract class to facilitate such implementations. Please refer to the TopKWeightAndReduce section.
-`FusedMoEPermuteExpertsUnpermute::finalize_weight_and_reduce_impl()` returns the `TopKWeightAndReduce` object that the implementation wants the `FusedMoEPrepareAndFinalize::finalize()` to use.
+It is sometimes efficient to perform TopK weight application and Reduction inside the `FusedMoEExpertsModular::apply()`. Find an example [here](https://github.com/vllm-project/vllm/pull/20228). We have a `TopKWeightAndReduce` abstract class to facilitate such implementations. Please refer to the TopKWeightAndReduce section.
+`FusedMoEExpertsModular::finalize_weight_and_reduce_impl()` returns the `TopKWeightAndReduce` object that the implementation wants the `FusedMoEPrepareAndFinalizeModular::finalize()` to use.
 
-![FusedMoEPermuteExpertsUnpermute Blocks](../assets/design/fused_moe_modular_kernel/fused_experts_blocks.png)
+![FusedMoEExpertsModular Blocks](../assets/design/fused_moe_modular_kernel/fused_experts_blocks.png)
 
 ### FusedMoEModularKernel
 
-`FusedMoEModularKernel` is composed of the `FusedMoEPrepareAndFinalize` and `FusedMoEPermuteExpertsUnpermute` objects.
+`FusedMoEModularKernel` is composed of the `FusedMoEPrepareAndFinalizeModular` and `FusedMoEExpertsModular` objects.
 `FusedMoEModularKernel` pseudocode/sketch,
 
 ```py
 class FusedMoEModularKernel:
     def __init__(self,
-                 prepare_finalize: FusedMoEPrepareAndFinalize,
-                 fused_experts: FusedMoEPermuteExpertsUnpermute):
+                 prepare_finalize: FusedMoEPrepareAndFinalizeModular,
+                 fused_experts: FusedMoEExpertsModular):
 
         self.prepare_finalize = prepare_finalize
         self.fused_experts = fused_experts
@@ -128,54 +128,53 @@ class FusedMoEModularKernel:
 
 ## How-To
 
-### How To Add a FusedMoEPrepareAndFinalize Type
+### How To Add a FusedMoEPrepareAndFinalizeModular Type
 
-Typically a FusedMoEPrepareAndFinalize type is backed by an All2All Dispatch & Combine implementation / kernel. For example,
+Typically a FusedMoEPrepareAndFinalizeModular type is backed by an All2All Dispatch & Combine implementation / kernel. For example,
 
-* PplxPrepareAndFinalize type is backed by Pplx All2All kernels,
 * DeepEPHTPrepareAndFinalize type is backed by DeepEP High-Throughput All2All kernels, and
 * DeepEPLLPrepareAndFinalize type is backed by DeepEP Low-Latency All2All kernels.
 
 #### Step 1: Add an All2All manager
 
-The purpose of the All2All Manager is to set up the All2All kernel implementations. The `FusedMoEPrepareAndFinalize` implementations typically fetch a kernel-implementation "handle" from the All2All Manager to invoke the Dispatch and Combine functions. Please look at the All2All Manager implementations [here](../../vllm/distributed/device_communicators/all2all.py).
+The purpose of the All2All Manager is to set up the All2All kernel implementations. The `FusedMoEPrepareAndFinalizeModular` implementations typically fetch a kernel-implementation "handle" from the All2All Manager to invoke the Dispatch and Combine functions. Please look at the All2All Manager implementations [here](../../vllm/distributed/device_communicators/all2all.py).
 
-#### Step 2: Add a FusedMoEPrepareAndFinalize Type
+#### Step 2: Add a FusedMoEPrepareAndFinalizeModular Type
 
-This section describes the significance of the various functions exposed by the `FusedMoEPrepareAndFinalize` abstract class.
+This section describes the significance of the various functions exposed by the `FusedMoEPrepareAndFinalizeModular` abstract class.
 
-`FusedMoEPrepareAndFinalize::prepare()`: The prepare method implements the Quantization and All2All Dispatch. Typically the Dispatch function from the relevant All2All Manager is invoked.
+`FusedMoEPrepareAndFinalizeModular::prepare()`: The prepare method implements the Quantization and All2All Dispatch. Typically the Dispatch function from the relevant All2All Manager is invoked.
 
-`FusedMoEPrepareAndFinalize::has_prepare_no_receive()`: Indicates whether or not this subclass implements `prepare_no_receive`. Defaults to False.
+`FusedMoEPrepareAndFinalizeModular::has_prepare_no_receive()`: Indicates whether or not this subclass implements `prepare_no_receive`. Defaults to False.
 
-`FusedMoEPrepareAndFinalize::prepare_no_receive()`: The prepare_no_receive method implements the Quantization and All2All Dispatch. It does not wait for the result of the dispatch operation but instead returns a thunk that can be invoked to wait for the final results. Typically the Dispatch function from the relevant All2All Manager is invoked.
+`FusedMoEPrepareAndFinalizeModular::prepare_no_receive()`: The prepare_no_receive method implements the Quantization and All2All Dispatch. It does not wait for the result of the dispatch operation but instead returns a thunk that can be invoked to wait for the final results. Typically the Dispatch function from the relevant All2All Manager is invoked.
 
-`FusedMoEPrepareAndFinalize::finalize()`: Maybe perform TopK Weight Application and Reduction and All2All Combine. Typically the Combine function from the relevant All2AllManager is invoked.
+`FusedMoEPrepareAndFinalizeModular::finalize()`: Maybe perform TopK Weight Application and Reduction and All2All Combine. Typically the Combine function from the relevant All2AllManager is invoked.
 
-`FusedMoEPrepareAndFinalize::activation_format()`: Return `FusedMoEActivationFormat.BatchedExperts` if the output of the prepare method (i.e. the All2All dispatch) is Batched. Return `FusedMoEActivationFormat.Standard` otherwise.
+`FusedMoEPrepareAndFinalizeModular::activation_format()`: Return `FusedMoEActivationFormat.BatchedExperts` if the output of the prepare method (i.e. the All2All dispatch) is Batched. Return `FusedMoEActivationFormat.Standard` otherwise.
 
-`FusedMoEPrepareAndFinalize::topk_indices_dtype()`: Data type of the TopK ids. Some All2All kernels have strict requirements pertaining to the data type of the TopK ids. This requirement is passed on to the `FusedMoe::select_experts` function so it could be respected. If there are no strict requirements return None.
+`FusedMoEPrepareAndFinalizeModular::topk_indices_dtype()`: Data type of the TopK ids. Some All2All kernels have strict requirements pertaining to the data type of the TopK ids. This requirement is passed on to the `FusedMoe::select_experts` function so it could be respected. If there are no strict requirements return None.
 
-`FusedMoEPrepareAndFinalize::max_num_tokens_per_rank()`: This is the maximum number of tokens that would be submitted to the All2All Dispatch at once.
+`FusedMoEPrepareAndFinalizeModular::max_num_tokens_per_rank()`: This is the maximum number of tokens that would be submitted to the All2All Dispatch at once.
 
-`FusedMoEPrepareAndFinalize::num_dispatchers()`: Total number of dispatching units. This value determines the size of the Dispatch output. The Dispatch output is of shape (num_local_experts, max_num_tokens, K). Here max_num_tokens = num_dispatchers() * max_num_tokens_per_rank().
+`FusedMoEPrepareAndFinalizeModular::num_dispatchers()`: Total number of dispatching units. This value determines the size of the Dispatch output. The Dispatch output is of shape (num_local_experts, max_num_tokens, K). Here max_num_tokens = num_dispatchers() * max_num_tokens_per_rank().
 
-We suggest picking an already existing `FusedMoEPrepareAndFinalize` implementation that matches your All2All implementation closely and using it as a reference.
+We suggest picking an already existing `FusedMoEPrepareAndFinalizeModular` implementation that matches your All2All implementation closely and using it as a reference.
 
-### How To Add a FusedMoEPermuteExpertsUnpermute Type
+### How To Add a FusedMoEExpertsModular Type
 
-FusedMoEPermuteExpertsUnpermute performs the core of the FusedMoE operations. The various functions exposed by the abstract class and their significance is as follows,
+FusedMoEExpertsModular performs the core of the FusedMoE operations. The various functions exposed by the abstract class and their significance is as follows,
 
-`FusedMoEPermuteExpertsUnpermute::activation_formats()`: Return the supported Input and Output activation formats. i.e. Contiguous / Batched format.
+`FusedMoEExpertsModular::activation_formats()`: Return the supported Input and Output activation formats. i.e. Contiguous / Batched format.
 
-`FusedMoEPermuteExpertsUnpermute::supports_chunking()`: Return True if the implementation supports chunking. Typically
+`FusedMoEExpertsModular::supports_chunking()`: Return True if the implementation supports chunking. Typically
 implementations that input `FusedMoEActivationFormat.Standard` support chunking and `FusedMoEActivationFormat.BatchedExperts` do not.
 
-`FusedMoEPermuteExpertsUnpermute::supports_expert_map()`: Return True if the implementation supports expert map.
+`FusedMoEExpertsModular::supports_expert_map()`: Return True if the implementation supports expert map.
 
-`FusedMoEPermuteExpertsUnpermute::workspace_shapes()` /
-`FusedMoEPermuteExpertsUnpermute::finalize_weight_and_reduce_impl` /
-`FusedMoEPermuteExpertsUnpermute::apply`: Refer to `FusedMoEPermuteExpertsUnpermute` section above.
+`FusedMoEExpertsModular::workspace_shapes()` /
+`FusedMoEExpertsModular::finalize_weight_and_reduce_impl` /
+`FusedMoEExpertsModular::apply`: Refer to `FusedMoEExpertsModular` section above.
 
 ### FusedMoEModularKernel Initialization
 
@@ -187,14 +186,14 @@ implementations that input `FusedMoEActivationFormat.Standard` support chunking
 
 #### maybe_make_prepare_finalize
 
-The `maybe_make_prepare_finalize` method is responsible for constructing an instance of `FusedMoEPrepareAndFinalize` when appropriate based on the current all2all backend, e.g. when EP + DP is enabled.  The base class method currently constructs all the `FusedMoEPrepareAndFinalize` objects for the EP+DP case.  Derived classes can override this method to construct prepare/finalize objects for different scenarios, e.g. `ModelOptNvFp4FusedMoE` can construct a `FlashInferCutlassMoEPrepareAndFinalize` for the EP+TP case.
+The `maybe_make_prepare_finalize` method is responsible for constructing an instance of `FusedMoEPrepareAndFinalizeModular` when appropriate based on the current all2all backend, e.g. when EP + DP is enabled.  The base class method currently constructs all the `FusedMoEPrepareAndFinalizeModular` objects for the EP+DP case.  Derived classes can override this method to construct prepare/finalize objects for different scenarios, e.g. `ModelOptNvFp4FusedMoE` can construct a `FlashInferCutlassMoEPrepareAndFinalize` for the EP+TP case.
 Please refer to the implementations in,
 
 * `ModelOptNvFp4FusedMoE`
 
 #### select_gemm_impl
 
-The `select_gemm_impl` method is undefined in the base class. It is the responsibility of the derived class to implement a method that constructs a valid/appropriate `FusedMoEPermuteExpertsUnpermute` object.
+The `select_gemm_impl` method is undefined in the base class. It is the responsibility of the derived class to implement a method that constructs a valid/appropriate `FusedMoEExpertsModular` object.
 Please refer to the implementations in,
 
 * `UnquantizedFusedMoEMethod`
@@ -206,7 +205,7 @@ derived classes.
 
 #### init_prepare_finalize
 
-Based on the input and env settings, the `init_prepare_finalize` method creates the appropriate `FusedMoEPrepareAndFinalize` object. The method then queries `select_gemm_impl` for the appropriate `FusedMoEPermuteExpertsUnpermute` object and builds the `FusedMoEModularKernel` object
+Based on the input and env settings, the `init_prepare_finalize` method creates the appropriate `FusedMoEPrepareAndFinalizeModular` object. The method then queries `select_gemm_impl` for the appropriate `FusedMoEExpertsModular` object and builds the `FusedMoEModularKernel` object
 
 Please take a look at [init_prepare_finalize](https://github.com/vllm-project/vllm/blob/1cbf951ba272c230823b947631065b826409fa62/vllm/model_executor/layers/fused_moe/layer.py#L188).
 **Important**: The `FusedMoEMethodBase` derived classes use the `FusedMoEMethodBase::fused_experts` object in their `apply` methods. When settings permit the construction of a valid `FusedMoEModularKernel` object, we override `FusedMoEMethodBase::fused_experts` with it. This essentially makes the derived classes agnostic to what FusedMoE implementation is used.
@@ -215,9 +214,9 @@ Please take a look at [init_prepare_finalize](https://github.com/vllm-project/vl
 
 We have `FusedMoEModularKernel` unit tests at [test_modular_kernel_combinations.py](../../tests/kernels/moe/test_modular_kernel_combinations.py).
 
-The unit test iterates through all combinations of `FusedMoEPrepareAndFinalize` and `FusedMoEPremuteExpertsUnpermute` types and if they are
+The unit test iterates through all combinations of `FusedMoEPrepareAndFinalizeModular` and `FusedMoEPremuteExpertsUnpermute` types and if they are
 compatible, runs some correctness tests.
-If you are adding some `FusedMoEPrepareAndFinalize` / `FusedMoEPermuteExpertsUnpermute` implementations,
+If you are adding some `FusedMoEPrepareAndFinalizeModular` / `FusedMoEExpertsModular` implementations,
 
 1. Add the implementation type to `MK_ALL_PREPARE_FINALIZE_TYPES` and `MK_FUSED_EXPERT_TYPES` in [mk_objects.py](../../tests/kernels/moe/modular_kernel_tools/mk_objects.py) respectively.
 2. Update `Config::is_batched_prepare_finalize()`, `Config::is_batched_fused_experts()`, `Config::is_standard_fused_experts()`,
@@ -226,24 +225,24 @@ If you are adding some `FusedMoEPrepareAndFinalize` / `FusedMoEPermuteExpertsUnp
 
 Doing this will add the new implementation to the test suite.
 
-### How To Check `FusedMoEPrepareAndFinalize` & `FusedMoEPermuteExpertsUnpermute` Compatibility
+### How To Check `FusedMoEPrepareAndFinalizeModular` & `FusedMoEExpertsModular` Compatibility
 
 The unit test file [test_modular_kernel_combinations.py](../../tests/kernels/moe/test_modular_kernel_combinations.py) can also be executed as a standalone script.
-Example: `python3 -m tests.kernels.moe.test_modular_kernel_combinations --pf-type PplxPrepareAndFinalize --experts-type BatchedTritonExperts`
-As a side effect, this script can be used to test `FusedMoEPrepareAndFinalize` & `FusedMoEPermuteExpertsUnpermute` compatibility. When invoked
+Example: `python3 -m tests.kernels.moe.test_modular_kernel_combinations --pf-type DeepEPLLPrepareAndFinalize --experts-type BatchedTritonExperts`
+As a side effect, this script can be used to test `FusedMoEPrepareAndFinalizeModular` & `FusedMoEExpertsModular` compatibility. When invoked
 with incompatible types, the script will error.
 
 ### How To Profile
 
 Please take a look at [profile_modular_kernel.py](../../tests/kernels/moe/modular_kernel_tools/profile_modular_kernel.py)
 The script can be used to generate Torch traces for a single `FusedMoEModularKernel::forward()` call for any compatible
-`FusedMoEPrepareAndFinalize` and `FusedMoEPermuteExpertsUnpermute` types.
-Example: `python3 -m tests.kernels.moe.modular_kernel_tools.profile_modular_kernel --pf-type PplxPrepareAndFinalize --experts-type BatchedTritonExperts`
+`FusedMoEPrepareAndFinalizeModular` and `FusedMoEExpertsModular` types.
+Example: `python3 -m tests.kernels.moe.modular_kernel_tools.profile_modular_kernel --pf-type DeepEPLLPrepareAndFinalize --experts-type BatchedTritonExperts`
 
-## FusedMoEPrepareAndFinalize Implementations
+## FusedMoEPrepareAndFinalizeModular Implementations
 
 See [Fused MoE Kernel features](./moe_kernel_features.md#fused-moe-modular-all2all-backends) for a list of all the available modular prepare and finalize subclasses.
 
-## FusedMoEPermuteExpertsUnpermute
+## FusedMoEExpertsModular
 
 See [Fused MoE Kernel features](./moe_kernel_features.md#fused-moe-experts-kernels) for a list of all the available modular experts.

docs/design/io_processor_plugins.md

@@ -13,12 +13,13 @@ IOProcessorInput = TypeVar("IOProcessorInput")
 IOProcessorOutput = TypeVar("IOProcessorOutput")
 
 class IOProcessor(ABC, Generic[IOProcessorInput, IOProcessorOutput]):
-    def __init__(self, vllm_config: VllmConfig):
+    """Abstract interface for pre/post-processing of engine I/O."""
+
+    def __init__(self, vllm_config: VllmConfig, renderer: BaseRenderer):
         super().__init__()
 
         self.vllm_config = vllm_config
 
-    @abstractmethod
     def parse_data(self, data: object) -> IOProcessorInput:
         raise NotImplementedError
 
@@ -32,7 +33,7 @@ class IOProcessor(ABC, Generic[IOProcessorInput, IOProcessorOutput]):
         self,
         params: PoolingParams | None = None,
     ) -> PoolingParams:
-        return params or PoolingParams()
+        return params or PoolingParams(task="plugin")
 
     @abstractmethod
     def pre_process(

docs/design/metrics.md

@@ -656,7 +656,7 @@ vLLM has support for OpenTelemetry tracing:
 - Added by <https://github.com/vllm-project/vllm/pull/4687> and reinstated by <https://github.com/vllm-project/vllm/pull/20372>
 - Configured with `--oltp-traces-endpoint` and `--collect-detailed-traces`
 - [OpenTelemetry blog post](https://opentelemetry.io/blog/2024/llm-observability/)
-- [User-facing docs](../examples/online_serving/opentelemetry.md)
+- [User-facing docs](../../examples/online_serving/opentelemetry/README.md)
 - [Blog post](https://medium.com/@ronen.schaffer/follow-the-trail-supercharging-vllm-with-opentelemetry-distributed-tracing-aa655229b46f)
 - [IBM product docs](https://www.ibm.com/docs/en/instana-observability/current?topic=mgaa-monitoring-large-language-models-llms-vllm-public-preview)
 

docs/design/model_runner_v2.md

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

docs/design/moe_kernel_features.md

@@ -4,17 +4,17 @@ The purpose of this document is to provide an overview of the various MoE kernel
 
 ## Fused MoE Modular All2All backends
 
-There are a number of all2all communication backends that are used to implement expert parallelism (EP) for the `FusedMoE` layer. The different `FusedMoEPrepareAndFinalize` subclasses provide an interface for each all2all backend.
+There are a number of all2all communication backends that are used to implement expert parallelism (EP) for the `FusedMoE` layer. The different `FusedMoEPrepareAndFinalizeModular` subclasses provide an interface for each all2all backend.
 
 The following table describes the relevant features of each backend, i.e. activation format, supported quantization schemes and async support.
 
-The output activation format (standard or batched) corresponds to the output of the prepare step of the `FusedMoEPrepareAndFinalize` subclass, and the finalize step requires the same format. All the backend `prepare` methods expect activations in the standard format and all the `finalize` methods return activations in standard format. More details on the formats can be found in the [Fused MoE Modular Kernel](./fused_moe_modular_kernel.md) document.
+The output activation format (standard or batched) corresponds to the output of the prepare step of the `FusedMoEPrepareAndFinalizeModular` subclass, and the finalize step requires the same format. All the backend `prepare` methods expect activations in the standard format and all the `finalize` methods return activations in standard format. More details on the formats can be found in the [Fused MoE Modular Kernel](./fused_moe_modular_kernel.md) document.
 
-The quantization types and formats enumerate which quantization schemes are supported by each `FusedMoEPrepareAndFinalize` class. The quantization can happen before or after the dispatch based on the format the all2all backend supports, e.g. deepep_high_throughput supports only block-quantized fp8 format. Any other format will result in dispatching in higher precision and quantizing afterwards. The output of the prepare step for each backend is the quantized type. The finalize step generally requires the same input type as the original activations, e.g. if the original input is bfloat16 and the quantization scheme is fp8 with per-tensor scales, `prepare` will return fp8/per-tensor scale activations and `finalize` will take bfloat16 activations. See the diagrams in [Fused MoE Modular Kernel](./fused_moe_modular_kernel.md) for more details on the types and formats of activations at each step of the MoE process. If no quantization type is specified, the kernel operates on float16 and/or bfloat16.
+The quantization types and formats enumerate which quantization schemes are supported by each `FusedMoEPrepareAndFinalizeModular` class. The quantization can happen before or after the dispatch based on the format the all2all backend supports, e.g. deepep_high_throughput supports only block-quantized fp8 format. Any other format will result in dispatching in higher precision and quantizing afterwards. The output of the prepare step for each backend is the quantized type. The finalize step generally requires the same input type as the original activations, e.g. if the original input is bfloat16 and the quantization scheme is fp8 with per-tensor scales, `prepare` will return fp8/per-tensor scale activations and `finalize` will take bfloat16 activations. See the diagrams in [Fused MoE Modular Kernel](./fused_moe_modular_kernel.md) for more details on the types and formats of activations at each step of the MoE process. If no quantization type is specified, the kernel operates on float16 and/or bfloat16.
 
 Async backends support the use of DBO (Dual Batch Overlap) and shared expert overlap (where shared experts are computed during the combine step).
 
-Certain models require the topk weights to be applied to the input activations rather than the output activations when topk==1, e.g. Llama. For modular kernels, this feature is supported by the `FusedMoEPrepareAndFinalize` subclass. For non-modular kernels, it is up to the experts function to deal with this flag.
+Certain models require the topk weights to be applied to the input activations rather than the output activations when topk==1, e.g. Llama. For modular kernels, this feature is supported by the `FusedMoEPrepareAndFinalizeModular` subclass. For non-modular kernels, it is up to the experts function to deal with this flag.
 
 Unless otherwise specified, backends are controlled via the `--all2all-backend` command-line argument (or the `all2all_backend` parameter in `ParallelConfig`). All backends except `flashinfer` only work with EP+DP or EP+TP. `Flashinfer` can work with EP or DP without EP.
 
@@ -32,13 +32,10 @@ th {
 
 | Backend | Output act. format | Quant. types | Quant. format | Async | Apply Weight On Input | Subclass |
 |---------|--------------------|--------------|---------------|-------|-----------------------|-----------|
-| naive | standard | all<sup>1</sup> | G,A,T | N | <sup>6</sup> | [layer.py][vllm.model_executor.layers.fused_moe.layer.FusedMoE |
-| pplx | batched | fp8,int8 | G,A,T | Y | Y | [`PplxPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.pplx_prepare_finalize.PplxPrepareAndFinalize] |
-| deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize.DeepEPLLPrepareAndFinalize] |
-| deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize.DeepEPHTPrepareAndFinalize] |
+| naive | standard | all<sup>1</sup> | G,A,T | N | <sup>6</sup> | [layer.py][vllm.model_executor.layers.fused_moe.layer.FusedMoE] |
+| deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize.DeepEPHTPrepareAndFinalize] |
+| deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize.DeepEPLLPrepareAndFinalize] |
 | flashinfer_all2allv | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferA2APrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_a2a_prepare_finalize.FlashInferA2APrepareAndFinalize] |
-| MoEPrepareAndFinalizeNoEP<sup>5</sup> | standard | fp8,int8 | G,A,T | N | Y | [`MoEPrepareAndFinalizeNoEP`][vllm.model_executor.layers.fused_moe.prepare_finalize.MoEPrepareAndFinalizeNoEP] |
-| BatchedPrepareAndFinalize<sup>5</sup> | batched | fp8,int8 | G,A,T | N | Y | [`BatchedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.fused_batched_moe.BatchedPrepareAndFinalize] |
 
 !!! info "Table key"
     1. All types: mxfp4, nvfp4, int4, int8, fp8
@@ -68,7 +65,7 @@ Modular kernels are supported by the following `FusedMoEMethodBase` classes.
 
 There are a number of MoE experts kernel implementations for different quantization types and architectures. Most follow the general API of the base Triton [`fused_experts`][vllm.model_executor.layers.fused_moe.fused_moe.fused_experts] function. Many have modular kernel adapters, so they can be used with compatible all2all backends. This table lists each experts kernel and its particular properties.
 
-Each kernel must be provided with one of the supported input activation formats. Some flavors of kernels support both standard and batched formats through different entry points, e.g. `TritonExperts` and `BatchedTritonExperts`. Batched format kernels are currently only needed for matching with certain all2all backends, e.g. `pplx` and `DeepEPLLPrepareAndFinalize`.
+Each kernel must be provided with one of the supported input activation formats. Some flavors of kernels support both standard and batched formats through different entry points, e.g. `TritonExperts` and `BatchedTritonExperts`. Batched format kernels are currently only needed for matching with certain all2all backends, e.g. `DeepEPLLPrepareAndFinalize`.
 
 Similar to the backend kernels, each experts kernel only supports certain quantization formats. For non-modular experts, the activations will be in the original type and quantized internally by the kernel. Modular experts will expect the activations to already be in the quantized format. Both types of experts will yield outputs in the original activation type.
 
@@ -76,9 +73,9 @@ Each experts kernel supports one or more activation functions, e.g. silu or gelu
 
 As with the backends, some experts support applying topk weights on the input activations. The entries in the column in this table only apply to the non-modular experts.
 
-Most experts flavors include an equivalent modular interface which will be a subclass of `FusedMoEPermuteExpertsUnpermute`.
+Most experts flavors include an equivalent modular interface which will be a subclass of `FusedMoEExpertsModular`.
 
-To be used with a particular `FusedMoEPrepareAndFinalize` subclass, MoE kernels must have compatible activation formats, quantization types and quantization formats.
+To be used with a particular `FusedMoEPrepareAndFinalizeModular` subclass, MoE kernels must have compatible activation formats, quantization types and quantization formats.
 
 | Kernel | Input act. format | Quant. types | Quant. format | Activation function | Apply Weight On Input | Modular | Source |
 |--------|-------------------|--------------|---------------|---------------------|-----------------------|---------|--------|
@@ -107,8 +104,8 @@ To be used with a particular `FusedMoEPrepareAndFinalize` subclass, MoE kernels
 
 The following table shows "families" of modular kernels that are intended to work together. There are some combinations which may work but have not yet been tested, e.g. flashinfer with other fp8 experts. Note that the "naive" backend will work with any non-modular experts.
 
-| backend | `FusedMoEPrepareAndFinalize` subclasses | `FusedMoEPermuteExpertsUnpermute` subclasses |
+| backend | `FusedMoEPrepareAndFinalizeModular` subclasses | `FusedMoEExpertsModular` subclasses |
 |---------|-----------------------------------------|----------------------------------------------|
 | deepep_high_throughput | `DeepEPHTPrepareAndFinalize` |  `DeepGemmExperts`,</br>`TritonExperts`,</br>`TritonOrDeepGemmExperts`,</br>`CutlassExpertsFp8`, </br>`MarlinExperts` |
-| deepep_low_latency,</br>pplx | `DeepEPLLPrepareAndFinalize`,</br>`PplxPrepareAndFinalize` |  `BatchedDeepGemmExperts`,</br>`BatchedTritonExperts`,</br>`CutlassBatchedExpertsFp8`,</br>`BatchedMarlinExperts` |
+| deepep_low_latency | `DeepEPLLPrepareAndFinalize` |  `BatchedDeepGemmExperts`,</br>`BatchedTritonExperts`,</br>`CutlassBatchedExpertsFp8`,</br>`BatchedMarlinExperts` |
 | flashinfer | `FlashInferCutlassMoEPrepareAndFinalize` | `FlashInferExperts` |