refactor setup.py to extract stuff into build_rust.py

Signed-off-by: Bugen Zhao <i@bugenzhao.com>

.buildkite/hardware_tests/cpu.yaml

@@ -91,7 +91,7 @@ steps:
   - tests/quantization/test_cpu_wna16.py
   commands:
     - |
-      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 30m "
       pytest -x -v -s tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs
       pytest -x -v -s tests/quantization/test_cpu_wna16.py"
       

.buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml

@@ -1,68 +0,0 @@
-group: Intel
-steps:
-  - label: ":docker: Build XPU image"
-    soft_fail: true
-    optional: true
-    depends_on: []
-    key: image-build-xpu
-    commands:
-      - bash -lc '.buildkite/image_build/image_build_xpu.sh "public.ecr.aws/q9t5s3a7" "vllm-ci-test-repo" "$BUILDKITE_COMMIT"'
-    env:
-      DOCKER_BUILDKIT: "1"
-    retry:
-      automatic:
-        - exit_status: -1  # Agent was lost
-          limit: 2
-        - exit_status: -10  # Agent was lost
-          limit: 2
-  - label: "XPU example Test"
-    depends_on:
-      - image-build-xpu
-    timeout_in_minutes: 30
-    optional: true
-    device: intel_gpu
-    no_plugin: true
-    env:
-      REGISTRY: "public.ecr.aws/q9t5s3a7"
-      REPO: "vllm-ci-test-repo"
-    source_file_dependencies:
-      - .buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml
-      - .buildkite/scripts/hardware_ci/run-intel-ci-test.sh
-    commands:
-      - >-
-        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-        'bash .buildkite/scripts/hardware_ci/run-intel-ci-test.sh example'
-  - label: "XPU V1 test"
-    depends_on:
-      - image-build-xpu
-    timeout_in_minutes: 30
-    optional: true
-    device: intel_gpu
-    no_plugin: true
-    env:
-      REGISTRY: "public.ecr.aws/q9t5s3a7"
-      REPO: "vllm-ci-test-repo"
-    source_file_dependencies:
-      - .buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml
-      - .buildkite/scripts/hardware_ci/run-intel-ci-test.sh
-    commands:
-      - >-
-        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-        'bash .buildkite/scripts/hardware_ci/run-intel-ci-test.sh v1'
-  - label: "XPU server test"
-    depends_on:
-      - image-build-xpu
-    timeout_in_minutes: 30
-    optional: true
-    device: intel_gpu
-    no_plugin: true
-    env:
-      REGISTRY: "public.ecr.aws/q9t5s3a7"
-      REPO: "vllm-ci-test-repo"
-    source_file_dependencies:
-      - .buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml
-      - .buildkite/scripts/hardware_ci/run-intel-ci-test.sh
-    commands:
-      - >-
-        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-        'bash .buildkite/scripts/hardware_ci/run-intel-ci-test.sh server'

.buildkite/intel_jobs/misc_intel.yaml

@@ -57,16 +57,13 @@ steps:
   commands:
     - >-
       bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'pip install lm_eval[api]>=0.4.12 &&
-      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      'export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
       cd tests &&
       pytest -v -s v1/logits_processors --ignore=v1/logits_processors/test_custom_online.py --ignore=v1/logits_processors/test_custom_offline.py &&
       pytest -v -s v1/test_oracle.py &&
       pytest -v -s v1/test_request.py &&
       pytest -v -s v1/test_outputs.py &&
-      pytest -v -s v1/sample/test_topk_topp_sampler.py &&
-      pytest -v -s v1/sample/test_logprobs.py &&
-      pytest -v -s v1/sample/test_logprobs_e2e.py'
+      pytest -v -s v1/sample/test_topk_topp_sampler.py'
 
 - label: XPU CPU Offload
   timeout_in_minutes: 60

.buildkite/intel_jobs/model_runner_v2_intel.yaml

@@ -1,54 +0,0 @@
-group: Model Runner V2 Intel
-depends_on:
-  - image-build-xpu
-steps:
-- label: Model Runner V2 Core Tests (Intel)
-  timeout_in_minutes: 45
-  device: intel_gpu
-  no_plugin: true
-  working_dir: "."
-  env:
-    REGISTRY: "public.ecr.aws/q9t5s3a7"
-    REPO: "vllm-ci-test-repo"
-    VLLM_TEST_DEVICE: "xpu"
-  source_file_dependencies:
-    - vllm/v1/worker/gpu/
-    - vllm/v1/worker/gpu_worker.py
-    - vllm/v1/core/sched/
-    - vllm/v1/attention/
-    - tests/v1/engine/test_llm_engine.py
-    - tests/v1/e2e/
-  commands:
-    - >-
-      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
-      cd tests &&
-      pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics" &&
-      ENFORCE_EAGER=1 pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram" &&
-      pytest -v -s v1/e2e/general/test_min_tokens.py'
-
-- label: Model Runner V2 Examples (Intel)
-  timeout_in_minutes: 45
-  device: intel_gpu
-  no_plugin: true
-  working_dir: "."
-  env:
-    REGISTRY: "public.ecr.aws/q9t5s3a7"
-    REPO: "vllm-ci-test-repo"
-    VLLM_TEST_DEVICE: "xpu"
-  source_file_dependencies:
-    - vllm/v1/worker/gpu/
-    - vllm/v1/core/sched/
-    - vllm/v1/worker/gpu_worker.py
-    - examples/basic/offline_inference/
-    - examples/generate/multimodal/
-    - examples/features/
-  commands:
-    - >-
-      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
-      cd examples &&
-      python3 basic/offline_inference/chat.py &&
-      python3 basic/offline_inference/generate.py --model facebook/opt-125m &&
-      python3 generate/multimodal/vision_language_offline.py --seed 0 &&
-      python3 features/automatic_prefix_caching/prefix_caching_offline.py'

.buildkite/intel_jobs/test-intel.yaml

@@ -60,7 +60,6 @@ steps:
       - >-
         bash .buildkite/scripts/hardware_ci/run-intel-test.sh
         'cd tests &&
-        bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
         pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py &&
         pytest -v -s v1/engine --ignore=v1/engine/test_output_processor.py &&
         pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py -k "not test_topk_only and not test_topp_only and not test_topk_and_topp" &&
@@ -88,20 +87,3 @@ steps:
         cd tests &&
         pytest -v -s entrypoints/multimodal/openai/chat_completion/test_audio_in_video.py &&
         pytest -v -s benchmarks/test_serve_cli.py'
-  - label: "XPU quantization test"
-    depends_on:
-      - image-build-xpu
-    timeout_in_minutes: 30
-    device: intel_gpu
-    no_plugin: true
-    env:
-      REGISTRY: "public.ecr.aws/q9t5s3a7"
-      REPO: "vllm-ci-test-repo"
-    source_file_dependencies:
-      - vllm/
-      - .buildkite/intel_jobs/test-intel.yaml 
-    commands:
-      - >-
-        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-        'cd tests &&
-        pytest -v -s quantization/test_auto_round.py'

.buildkite/lm-eval-harness/configs/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8.yaml

@@ -6,7 +6,9 @@ tasks:
     value: 0.7142
   - name: "exact_match,flexible-extract"
     value: 0.4579
-moe_backend: "flashinfer_cutlass"
+env_vars:
+  VLLM_USE_FLASHINFER_MOE_FP8: "1"
+  VLLM_FLASHINFER_MOE_BACKEND: "throughput"
 limit: 1319
 num_fewshot: 5
 max_model_len: 262144

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

@@ -68,10 +68,6 @@ def launch_lm_eval(eval_config, tp_size):
     if current_platform.is_rocm() and "Nemotron-3" in eval_config["model_name"]:
         model_args += "attention_backend=TRITON_ATTN"
 
-    moe_backend = eval_config.get("moe_backend", None)
-    if moe_backend is not None:
-        model_args += f"moe_backend={moe_backend},"
-
     env_vars = eval_config.get("env_vars", None)
     with scoped_env_vars(env_vars):
         results = lm_eval.simple_evaluate(

.buildkite/release-pipeline.yaml

@@ -1,25 +1,12 @@
 # CUDA architecture lists — following PyTorch RELEASE.md
 # (https://github.com/pytorch/pytorch/blob/main/RELEASE.md)
 # SM86 included for broader Ampere coverage; SM89 for marlin fp8 support
-# These requested arches are filtered by CMake's CUDA_SUPPORTED_ARCHS before
-# per-kernel arch selection. Do not add +PTX here: top-level +PTX is stripped
-# during that filtering, so kernels that need PTX must request it locally.
 env:
-  # for CUDA >=13, sm_100+ targets have family specifiers (see CMakeLists.txt)
-  # so targets like 10.3 and 12.1 are automatically supported with this list
-  CUDA_ARCH_X86: "7.5 8.0 8.6 8.9 9.0 10.0 12.0"
-  # aarch64-only targets: Orin (8.7), Thor (11.0, CUDA 13+)
-  CUDA_ARCH_AARCH64: "8.0 8.7 8.9 9.0 10.0 11.0 12.0"
-
-  # for CUDA <13, we need to specify all needed targets
-  # some targets (10.3, 12.1) are skipped to limit the wheel size (< 500MB)
-  # please use CUDA 13 wheels or compile yourself on these new devices
+  CUDA_ARCH_X86: "7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX"
+  # aarch64 only architectures: 8.7 for Orin, 11.0 for Thor (since CUDA 13)
+  CUDA_ARCH_AARCH64: "8.0 8.7 8.9 9.0 10.0 11.0 12.0+PTX"
   CUDA_ARCH_X86_CU129: "7.5 8.0 8.6 8.9 9.0 10.0 12.0"
   CUDA_ARCH_AARCH64_CU129: "8.0 8.7 8.9 9.0 10.0 12.0"
-  
-  # pre-built mooncake wheels
-  # the manylinux_2_35 wheel has compatibility issue on Ubuntu 24.04
-  # so we use different wheels for the time being
   MOONCAKE_WHEEL_AARCH64_2_35: "https://vllm-wheels.s3.amazonaws.com/mooncake/mooncake_transfer_engine-0.3.10.post2-0da9dfea3-cp312-cp312-manylinux_2_35_aarch64.whl"
   MOONCAKE_WHEEL_AARCH64_2_39: "https://vllm-wheels.s3.amazonaws.com/mooncake/mooncake_transfer_engine-0.3.10.post2-0da9dfea3-cp312-cp312-manylinux_2_39_aarch64.whl"
   MOONCAKE_WHEEL_X86_64: "https://vllm-wheels.s3.amazonaws.com/mooncake/mooncake_transfer_engine-0.3.10.post2-0da9dfea3-cp312-cp312-manylinux_2_35_x86_64.whl"
@@ -859,6 +846,7 @@ steps:
         allow_failure: true
       - step: build-cpu-release-image-arm64
         allow_failure: true
+    if: build.env("NIGHTLY") != "1"
 
   - label: "Publish release images to DockerHub"
     depends_on:

.buildkite/scripts/ci-clean-log.sh

@@ -13,8 +13,5 @@ INPUT_FILE="$1"
 # Strip timestamps
 sed -i 's/^\[[0-9]\{4\}-[0-9]\{2\}-[0-9]\{2\}T[0-9]\{2\}:[0-9]\{2\}:[0-9]\{2\}Z\] //' "$INPUT_FILE"
 
-# Strip Buildkite inline timestamp markers (ESC _bk;t=<ms> BEL)
-sed -i 's/\x1B_bk;t=[0-9]*\x07//g' "$INPUT_FILE"
-
 # Strip colorization
 sed -i -r 's/\x1B\[[0-9;]*[mK]//g' "$INPUT_FILE"

.buildkite/scripts/ci-fetch-log.sh

@@ -1,178 +1,74 @@
 #!/bin/bash
-# Fetch vLLM Buildkite CI logs (public; no login required).
+# Usage: ./ci-fetch-log.sh <buildkite_job_url> [output_file]
+#        ./ci-fetch-log.sh <build_number> <job_uuid> [output_file]
 #
-# Usage:
-#   ci-fetch-log.sh [--soft|--all] --pr [<PR>]  failed jobs in the PR's latest
-#                                               build (current branch if omitted)
-#   ci-fetch-log.sh [--soft|--all] <build_url>  failed jobs in that build
-#   ci-fetch-log.sh <job_url> [output]          one job; both #<job_uuid> and
-#                                               ?sid=<id> URL forms work
-#   ci-fetch-log.sh <build> <job_uuid> [output]
+# Downloads the raw log for a Buildkite job from the public, unauthenticated
+# /organizations/<org>/pipelines/<pipeline>/builds/<n>/jobs/<uuid>/download
+# endpoint, then strips ANSI/timestamps via ci-clean-log.sh.
 #
-# --soft also fetches soft-failed jobs; --all fetches every finished job.
-# Saves each log as ci-<build>-<job-name>.log (ANSI/timestamps stripped) and
-# prints "<file>\t<job name>" per job. [output] is single-job only; "-"
-# streams to stdout. Existing files are kept; CI_FETCH_LOG_FORCE=1 refetches.
+# Find <build_number> and <job_uuid> via:
+#   gh pr checks <PR> --repo vllm-project/vllm
+# Each failing row's URL is .../builds/<build_number>#<job_uuid>.
+#
+# Default output path: ci-<build>-<uuid_first_13_chars>.log (e.g.
+# ci-68478-019e6b07-daae.log). Jobs in the same build share the UUID's
+# first 8 chars, so the second segment is needed for uniqueness when
+# fetching multiple jobs in parallel. The script refuses to overwrite an
+# existing output file; pass an explicit path or set CI_FETCH_LOG_FORCE=1
+# to override.
 
 set -euo pipefail
 
 ORG="vllm"
 PIPELINE="ci"
-UA="vllm-ci-fetch-log"
-UUID_RE='[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}'
 
 usage() {
-    sed -n '2,15p' "$0" | sed 's/^# \{0,1\}//'
+    echo "Usage: $0 <buildkite_job_url> [output_file]"
+    echo "       $0 <build_number> <job_uuid> [output_file]"
     exit 1
 }
 
-die() {
-    echo "$1" >&2
-    exit 1
-}
+if [ $# -lt 1 ]; then usage; fi
 
-BUILD="" JOB="" SID="" OUT=""
-SCOPE="failed"
-
-while :; do
-    case "${1:-}" in
-    --soft) SCOPE="soft" ;;
-    --all) SCOPE="all" ;;
-    *) break ;;
-    esac
-    shift
-done
-
-case "${1:-}" in
---pr)
-    PR="${2:-}"
-    # gh pr checks exits non-zero when checks are failing; that is the
-    # expected case here.
-    URL=$(gh pr checks ${PR:+"$PR"} --repo vllm-project/vllm 2>/dev/null |
-        grep -oE "https://buildkite.com/${ORG}/${PIPELINE}/builds/[0-9]+" |
-        sort -t/ -k7 -n | tail -1 || true)
-    [ -n "$URL" ] || die "No Buildkite build found via: gh pr checks ${PR:-<current branch>}"
-    BUILD="${URL##*/}"
-    ;;
-https://*)
+if [[ "$1" == https://* ]]; then
     BUILD=$(echo "$1" | sed -nE 's#.*/builds/([0-9]+).*#\1#p')
-    JOB=$(echo "$1" | grep -oE "#${UUID_RE}" | head -n 1 | cut -c2- || true)
-    SID=$(echo "$1" | grep -oE "[?&]sid=${UUID_RE}" | head -n 1 | sed 's/.*sid=//' || true)
+    JOB=$(echo "$1" | grep -oE '[0-9a-f]{8}-[0-9a-f-]+' | head -n 1)
     OUT="${2:-}"
-    [ -n "$BUILD" ] || die "Could not parse build number from: $1"
-    ;;
-[0-9]*)
-    [ $# -ge 2 ] || usage
+else
+    if [ $# -lt 2 ]; then usage; fi
     BUILD="$1"
     JOB="$2"
     OUT="${3:-}"
-    ;;
-*)
+fi
+
+if [ -z "$BUILD" ] || [ -z "$JOB" ]; then
+    echo "Could not parse build number or job UUID from: $1" >&2
     usage
-    ;;
-esac
+fi
+
+# Jobs in the same build share the UUID's first segment, so include the
+# second segment (chars 9-13, e.g. "019e6b07-daae") to keep default filenames
+# unique when fetching multiple jobs from one build in parallel.
+if [ -z "$OUT" ]; then
+    OUT="ci-${BUILD}-${JOB:0:13}.log"
+fi
+
+if [ -e "$OUT" ] && [ -z "${CI_FETCH_LOG_FORCE:-}" ]; then
+    echo "Refusing to overwrite existing $OUT (set CI_FETCH_LOG_FORCE=1 or pass an explicit output path)." >&2
+    exit 1
+fi
 
 COOKIES=$(mktemp)
-JOBS_TSV=$(mktemp)
-trap 'rm -f "$COOKIES" "$JOBS_TSV"' EXIT
+trap 'rm -f "$COOKIES"' EXIT
 
-# Buildkite issues a session cookie on first hit; later requests need it.
-curl -fsSL -c "$COOKIES" -A "$UA" \
+# Buildkite issues a session cookie on first hit; subsequent /download needs it.
+curl -fsSL -c "$COOKIES" -A "vllm-ci-fetch-log" \
     "https://buildkite.com/${ORG}/${PIPELINE}/builds/${BUILD}" -o /dev/null
 
-# The build's job list (id, step uuid, state, name) is served as JSON from
-# the user-facing /data/jobs endpoint. Flatten it to TSV for easy filtering:
-#   job_id  step_uuid  failed  soft_failed  finished  slug  name
-curl -fsSL -b "$COOKIES" -A "$UA" \
-    "https://buildkite.com/${ORG}/${PIPELINE}/builds/${BUILD}/data/jobs" |
-    python3 -c '
-import json, re, sys
+curl -fsSL -b "$COOKIES" -A "vllm-ci-fetch-log" \
+    "https://buildkite.com/organizations/${ORG}/pipelines/${PIPELINE}/builds/${BUILD}/jobs/${JOB}/download" \
+    -o "$OUT"
 
-data = json.load(sys.stdin)
-if data.get("has_next_page"):
-    print("warning: job list is paginated; some jobs not shown", file=sys.stderr)
-for r in data["records"]:
-    if r.get("type") != "script":
-        continue
-    name = (r.get("name") or "").replace("\t", " ").replace("\n", " ")
-    slug = re.sub(r"[^a-z0-9]+", "-", name.lower()).strip("-")[:60]
-    print("\t".join([
-        r["id"],
-        r.get("step_uuid") or "",
-        str(r.get("passed") is False),
-        str(bool(r.get("soft_failed"))),
-        str(bool(r.get("finished_at"))),
-        slug,
-        name,
-    ]))
-' >"$JOBS_TSV" || die "Could not list jobs for build ${BUILD}"
+bash "$(dirname "$0")/ci-clean-log.sh" "$OUT"
 
-if [ -n "$SID" ] && [ -z "$JOB" ]; then
-    # The ?sid= in builds/<N>/list URLs is the *step* uuid, not the job uuid.
-    JOB=$(awk -F'\t' -v s="$SID" '$1 == s || $2 == s {print $1; exit}' "$JOBS_TSV")
-    [ -n "$JOB" ] || die "No job matching sid=${SID} in build ${BUILD}"
-fi
-
-fetch_job() { # <job_uuid> <output_file>
-    curl -fsSL -b "$COOKIES" -A "$UA" \
-        "https://buildkite.com/organizations/${ORG}/pipelines/${PIPELINE}/builds/${BUILD}/jobs/$1/download" \
-        -o "$2"
-    bash "$(dirname "$0")/ci-clean-log.sh" "$2"
-}
-
-if [ -n "$JOB" ]; then
-    # Single-job mode.
-    NAME=$(awk -F'\t' -v j="$JOB" '$1 == j {print $7; exit}' "$JOBS_TSV")
-    SLUG=$(awk -F'\t' -v j="$JOB" '$1 == j {print $6; exit}' "$JOBS_TSV")
-    [ -n "$OUT" ] || OUT="ci-${BUILD}-${SLUG:-${JOB:0:13}}.log"
-    if [ "$OUT" = "-" ]; then
-        TMP=$(mktemp)
-        fetch_job "$JOB" "$TMP"
-        cat "$TMP"
-        rm -f "$TMP"
-        exit 0
-    fi
-    if [ -e "$OUT" ] && [ -z "${CI_FETCH_LOG_FORCE:-}" ]; then
-        die "Refusing to overwrite existing ${OUT} (set CI_FETCH_LOG_FORCE=1 or pass an output path)."
-    fi
-    fetch_job "$JOB" "$OUT"
-    printf '%s\t%s\n' "$OUT" "${NAME:-$JOB}"
-    exit 0
-fi
-
-# Build-wide mode: fetch finished jobs matching $SCOPE.
-[ -z "$OUT" ] || die "[output_file] is only valid when fetching a single job."
-
-case "$SCOPE" in
-failed) FILTER='$3 == "True" && $4 == "False" && $5 == "True"' ;;
-soft) FILTER='$3 == "True" && $5 == "True"' ;;
-all) FILTER='$5 == "True"' ;;
-esac
-
-if [ "$SCOPE" = "failed" ]; then
-    SOFT=$(awk -F'\t' '$3 == "True" && $4 == "True"' "$JOBS_TSV" | wc -l)
-    [ "$SOFT" -eq 0 ] || echo "Skipping ${SOFT} soft-failed job(s); use --soft to include them." >&2
-fi
-
-FOUND=0
-EMITTED=" "
-while IFS=$'\t' read -r job_id _ _ _ _ slug name; do
-    FOUND=$((FOUND + 1))
-    out="ci-${BUILD}-${slug:-${job_id:0:13}}.log"
-    # Retries share a name with the original job; disambiguate by uuid.
-    case "$EMITTED" in
-    *" $out "*) out="ci-${BUILD}-${slug:-job}-${job_id:0:13}.log" ;;
-    esac
-    EMITTED="${EMITTED}${out} "
-    if [ -e "$out" ] && [ -z "${CI_FETCH_LOG_FORCE:-}" ]; then
-        echo "Keeping existing ${out} (set CI_FETCH_LOG_FORCE=1 to refetch)." >&2
-    elif ! fetch_job "$job_id" "$out"; then
-        echo "Failed to download log for job ${job_id} (${name})." >&2
-        continue
-    fi
-    printf '%s\t%s\n' "$out" "$name"
-done < <(awk -F'\t' "$FILTER" "$JOBS_TSV")
-
-if [ "$FOUND" -eq 0 ]; then
-    echo "No matching jobs in build ${BUILD} (scope: ${SCOPE})." >&2
-fi
+echo "$OUT"

.buildkite/scripts/hardware_ci/run-intel-ci-test.sh

@@ -1,51 +0,0 @@
-#!/bin/bash
-
-set -euo pipefail
-
-test_suite="${1:-}"
-
-if [[ -z "${test_suite}" ]]; then
-  echo "Usage: $0 <example|v1|server>" >&2
-  exit 1
-fi
-
-case "${test_suite}" in
-  example)
-    pip install tblib==3.1.0
-
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 -O3 -cc.cudagraph_mode=NONE
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --attention-backend=TRITON_ATTN
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --quantization fp8
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --kv-cache-dtype fp8
-    python3 examples/basic/offline_inference/generate.py --model nvidia/Llama-3.1-8B-Instruct-FP8 --block-size 64 --enforce-eager --quantization modelopt --kv-cache-dtype fp8 --attention-backend TRITON_ATTN --max-model-len 4096
-    python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --block-size 64 --enforce-eager --max-model-len 8192
-    python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2
-    python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
-    python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --max-model-len 8192
-    ;;
-  v1)
-    cd tests
-
-    pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py
-    pytest -v -s v1/engine --ignore=v1/engine/test_output_processor.py
-    pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py -k "not test_topk_only and not test_topp_only and not test_topk_and_topp"
-    pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.py
-    pytest -v -s v1/structured_output
-    pytest -v -s v1/test_serial_utils.py
-    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py
-    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py --ignore=v1/kv_connector/unit/test_hf3fs_client.py --ignore=v1/kv_connector/unit/test_hf3fs_connector.py --ignore=v1/kv_connector/unit/test_hf3fs_metadata_server.py --ignore=v1/kv_connector/unit/test_offloading_connector.py
-    ;;
-  server)
-    pip install av
-    cd tests
-
-    pytest -v -s entrypoints/multimodal/openai/chat_completion/test_audio_in_video.py
-    pytest -v -s benchmarks/test_serve_cli.py
-    ;;
-  *)
-    echo "Unknown Intel test suite: ${test_suite}" >&2
-    exit 1
-    ;;
-esac

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

@@ -243,10 +243,8 @@ container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head
 
 # ---- Command source selection ----
 commands=""
-commands_source=""
 if [[ -n "${VLLM_TEST_COMMANDS:-}" ]]; then
   commands="${VLLM_TEST_COMMANDS}"
-  commands_source="env"
   echo "Commands sourced from VLLM_TEST_COMMANDS (quoting preserved)"
 elif [[ $# -gt 0 ]]; then
   all_yaml=true
@@ -305,12 +303,8 @@ if [[ -z "$commands" ]]; then
 fi
 
 echo "Raw commands: $commands"
-if [[ "$commands_source" != "env" ]]; then
-  commands=$(re_quote_pytest_markers "$commands")
-  echo "After re-quoting: $commands"
-else
-  echo "Skipping re-quoting for VLLM_TEST_COMMANDS input"
-fi
+commands=$(re_quote_pytest_markers "$commands")
+echo "After re-quoting: $commands"
 commands=$(apply_intel_test_overrides "$commands")
 echo "Final commands: $commands"
 

.buildkite/test-amd.yaml

@@ -398,7 +398,7 @@ steps:
   - tests/kernels/helion/
   - vllm/platforms/rocm.py
   commands:
-  - pip install helion==1.1.0
+  - pip install helion==1.0.0
   - pytest -v -s kernels/helion/
 
 - label: Kernels Mamba Test # TBD
@@ -415,6 +415,22 @@ steps:
   commands:
   - pytest -v -s kernels/mamba
 
+#-----------------------------------------------------------  mi250 · lora  ------------------------------------------------------------#
+
+- label: LoRA %N # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
+  agent_pool: mi250_1
+  parallelism: 4
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/lora
+  - tests/lora
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_qwen3_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py
+
 #------------------------------------------------------  mi250 · models / basic  -------------------------------------------------------#
 
 - label: Basic Models Test (Other CPU) # TBD
@@ -592,11 +608,6 @@ steps:
   - pytest -v -s plugins_tests/test_bge_m3_sparse_io_processor_plugins.py
   - pip uninstall bge_m3_sparse_plugin -y
   # END: `bge_m3_sparse io_processor` test
-  # BEGIN: `colbert_query io_processor` test
-  - pip install -e ./plugins/colbert_query_plugin
-  - pytest -v -s plugins_tests/test_colbert_query_io_processor_plugins.py
-  - pip uninstall colbert_query_plugin -y
-  # END: `colbert_query io_processor` test
   # BEGIN: `stat_logger` plugins test
   - pip install -e ./plugins/vllm_add_dummy_stat_logger
   - pytest -v -s plugins_tests/test_stats_logger_plugins.py
@@ -1683,20 +1694,6 @@ steps:
 
 #-----------------------------------------------------------  mi300 · lora  ------------------------------------------------------------#
 
-- label: LoRA %N # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  parallelism: 4
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/lora
-  - tests/lora
-  - vllm/platforms/rocm.py
-  commands:
-  - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_qwen3_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py
-
 - label: LoRA TP (Distributed) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
@@ -1781,9 +1778,10 @@ steps:
   - tests/models/multimodal/generation
   - tests/models/multimodal/test_mapping.py
   commands:
-  - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-  - pytest -v -s models/multimodal/generation -m 'not core_model' --ignore models/multimodal/generation/test_common.py
-  - pytest -v -s models/multimodal/test_mapping.py
+  - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@rocm-7.0-v2.3.0'
+  - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
+  - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+
 
 - label: Multi-Modal Models (Extended Generation 2) # TBD
   timeout_in_minutes: 180
@@ -1795,8 +1793,9 @@ steps:
   - vllm/
   - tests/models/multimodal/generation
   commands:
-  - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-  - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
+  - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@rocm-7.0-v2.3.0'
+  - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
+  - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
 
 - label: Multi-Modal Models (Extended Generation 3) # TBD
@@ -2755,7 +2754,7 @@ steps:
   - vllm/_aiter_ops.py
   - vllm/platforms/rocm.py
   commands:
-  - python3 benchmarks/attention_benchmarks/benchmark.py --backends ROCM_ATTN ROCM_AITER_FA ROCM_AITER_UNIFIED_ATTN --batch-specs "8q1s1k"
+  - python3 benchmarks/attention_benchmarks/benchmark.py --backends ROCM_ATTN ROCM_AITER_FA ROCM_AITER_UNIFIED_ATTN --batch-specs "8q1s1k" --repeats 1 --warmup-iters 1
 
 #--------------------------------------------------------  mi355 · distributed  --------------------------------------------------------#
 
@@ -2947,6 +2946,7 @@ steps:
   - vllm/model_executor/models/qwen3_5_mtp.py
   - vllm/transformers_utils/configs/qwen3_5.py
   - vllm/transformers_utils/configs/qwen3_5_moe.py
+  - vllm/model_executor/models/qwen.py
   - vllm/model_executor/models/qwen2.py
   - vllm/model_executor/models/qwen3.py
   - vllm/model_executor/models/qwen3_next.py
@@ -3184,6 +3184,7 @@ steps:
   - vllm/model_executor/models/qwen3_5_mtp.py
   - vllm/transformers_utils/configs/qwen3_5.py
   - vllm/transformers_utils/configs/qwen3_5_moe.py
+  - vllm/model_executor/models/qwen.py
   - vllm/model_executor/models/qwen2.py
   - vllm/model_executor/models/qwen3.py
   - vllm/model_executor/models/qwen3_next.py

.buildkite/test_areas/attention.yaml

@@ -15,7 +15,7 @@ steps:
     - pytest -v -s v1/attention
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       timeout_in_minutes: 70
       depends_on:
       - image-build-amd

.buildkite/test_areas/benchmarks.yaml

@@ -23,4 +23,4 @@ steps:
   - benchmarks/attention_benchmarks/
   - vllm/v1/attention/
   commands:
-  - python3 benchmarks/attention_benchmarks/benchmark.py --backends flash flashinfer --batch-specs "8q1s1k"
+  - python3 benchmarks/attention_benchmarks/benchmark.py --backends flash flashinfer --batch-specs "8q1s1k" --repeats 1 --warmup-iters 1

.buildkite/test_areas/disaggregated.yaml

@@ -61,20 +61,6 @@ steps:
     - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
-- label: Hybrid SSM NixlConnector PD prefix cache test (2 GPUs)
-  key: hybrid-ssm-nixlconnector-pd-prefix-cache-2-gpus
-  timeout_in_minutes: 25
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
-    - vllm/v1/core/sched/
-    - vllm/v1/core/kv_cache_coordinator.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
-    - bash v1/kv_connector/nixl_integration/run_mamba_prefix_cache_test.sh
-
 - label: MultiConnector (Nixl+Offloading) PD accuracy (2 GPUs)
   key: multiconnector-nixl-offloading-pd-accuracy-2-gpus
   timeout_in_minutes: 30

.buildkite/test_areas/engine.yaml

@@ -28,7 +28,7 @@ steps:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py test_jit_monitor.py
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       timeout_in_minutes: 60
       depends_on:
       - image-build-amd
@@ -44,7 +44,7 @@ steps:
     - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       timeout_in_minutes: 40
       depends_on:
       - image-build-amd

.buildkite/test_areas/entrypoints.yaml

@@ -46,7 +46,7 @@ steps:
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/serve/dev/rpc
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       depends_on:
       - image-build-amd
 
@@ -63,7 +63,7 @@ steps:
   - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/correctness
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       timeout_in_minutes: 80
       depends_on:
       - image-build-amd
@@ -82,7 +82,7 @@ steps:
   - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       timeout_in_minutes: 80
       depends_on:
       - image-build-amd
@@ -104,7 +104,7 @@ steps:
   - pytest -v -s entrypoints/anthropic
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       timeout_in_minutes: 60
       depends_on:
       - image-build-amd
@@ -165,7 +165,7 @@ steps:
   - pytest -s entrypoints/openai/correctness/
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       depends_on:
       - image-build-amd
       source_file_dependencies:

.buildkite/test_areas/kernels.yaml

@@ -75,19 +75,6 @@ steps:
     - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 2
 
-- label: Kernels Attention DiffKV Test (H100)
-  key: kernels-attention-diffkv-test-h100
-  timeout_in_minutes: 20
-  device: h100
-  num_devices: 1
-  source_file_dependencies:
-  - vllm/v1/attention/ops/triton_unified_attention_diffkv.py
-  - vllm/v1/attention/backends/triton_attn_diffkv.py
-  - vllm/v1/attention/backends/flash_attn_diffkv.py
-  - tests/kernels/attention/test_triton_unified_attention_diffkv.py
-  commands:
-    - pytest -v -s kernels/attention/test_triton_unified_attention_diffkv.py
-
 - label: Kernels Quantization Test %N
   key: kernels-quantization-test
   timeout_in_minutes: 90
@@ -100,7 +87,7 @@ steps:
   parallelism: 2
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       source_file_dependencies:
       - csrc/quantization/
       - vllm/model_executor/layers/quantization
@@ -237,7 +224,7 @@ steps:
   - vllm/utils/import_utils.py
   - tests/kernels/helion/
   commands:
-    - pip install helion==1.1.0
+    - pip install helion==1.0.0
     - pytest -v -s kernels/helion/
 
  

.buildkite/test_areas/lm_eval.yaml

@@ -14,7 +14,7 @@ steps:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       timeout_in_minutes: 55
       depends_on:
       - image-build-amd

.buildkite/test_areas/lora.yaml

@@ -12,17 +12,6 @@ steps:
   commands:
     - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_qwen3_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py 
   parallelism: 4
-  mirror:
-    amd:
-      device: mi325_1
-      working_dir: "/vllm-workspace/tests"
-      timeout_in_minutes: 60
-      source_file_dependencies:
-      - vllm/lora
-      - tests/lora
-      - vllm/platforms/rocm.py
-      depends_on:
-      - image-build-amd
 
 
 - label: LoRA TP (Distributed)

.buildkite/test_areas/misc.yaml

@@ -21,12 +21,6 @@ steps:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     # TODO: create another `optional` test group for slow tests
     - pytest -v -s -m 'not slow_test' v1/spec_decode
-  mirror:
-    amd:
-      device: mi300_1
-      timeout_in_minutes: 65
-      depends_on:
-      - image-build-amd
 
 - label: V1 Sample + Logits
   key: v1-sample-logits
@@ -144,26 +138,11 @@ steps:
     - vllm/v1/spec_decode/extract_hidden_states.py
     - vllm/model_executor/models/extract_hidden_states.py
     - vllm/transformers_utils/configs/extract_hidden_states.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/example_hidden_states_connector.py
     - tests/v1/kv_connector/extract_hidden_states_integration
   commands:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - pytest -v -s v1/kv_connector/extract_hidden_states_integration
 
-- label: Extract Hidden States Integration (2 GPUs)
-  key: extract-hidden-states-integration-2-gpus
-  timeout_in_minutes: 20
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/v1/spec_decode/extract_hidden_states.py
-    - vllm/model_executor/models/extract_hidden_states.py
-    - vllm/transformers_utils/configs/extract_hidden_states.py
-    - vllm/distributed/kv_transfer/kv_connector/v1/example_hidden_states_connector.py
-    - tests/v1/kv_connector/extract_hidden_states_integration
-  commands:
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -v -s -m 'distributed' v1/kv_connector/extract_hidden_states_integration
-
 - label: Regression
   key: regression
   timeout_in_minutes: 20
@@ -314,7 +293,6 @@ steps:
   - vllm/transformers_utils/
   - vllm/utils/
   - vllm/v1/
-  - tests/test_envs.py
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/test_pooling_params.py
@@ -331,7 +309,6 @@ steps:
   device: cpu-small
   commands:
   - python3 standalone_tests/lazy_imports.py
-  - pytest -v -s test_envs.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s test_pooling_params.py

.buildkite/test_areas/models_multimodal.yaml

@@ -15,7 +15,7 @@ steps:
     - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       depends_on:
       - image-build-amd
 
@@ -30,9 +30,10 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
     - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
+    - pytest -v -s models/multimodal/generation/test_vit_cudagraph.py -m core_model
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       depends_on:
       - image-build-amd
 
@@ -49,7 +50,7 @@ steps:
     - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       depends_on:
       - image-build-amd
 
@@ -62,16 +63,9 @@ steps:
   - tests/models/multimodal
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py  --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/generation/test_vit_cudagraph.py --ignore models/multimodal/processing
-    - pytest -v -s models/multimodal/generation/test_vit_cudagraph.py -m core_model
+    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py  --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/processing
     - pytest models/multimodal/generation/test_memory_leak.py -m core_model
     - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
-  mirror:
-    amd:
-      soft_fail: true
-      device: mi325_1
-      depends_on:
-      - image-build-amd
 
 - label: Multi-Modal Processor (CPU)
   key: multi-modal-processor-cpu
@@ -161,7 +155,7 @@ steps:
     - pytest -v -s models/multimodal/pooling -m 'not core_model'
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       timeout_in_minutes: 60
       depends_on:
       - image-build-amd

.buildkite/test_areas/plugins.yaml

@@ -27,10 +27,6 @@ steps:
   - 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
-  # test colbert_query io_processor plugin
-  - pip install -e ./plugins/colbert_query_plugin
-  - pytest -v -s plugins_tests/test_colbert_query_io_processor_plugins.py
-  - pip uninstall colbert_query_plugin -y
   # end io_processor plugins test
   # begin stat_logger plugins test
   - pip install -e ./plugins/vllm_add_dummy_stat_logger
@@ -44,17 +40,3 @@ steps:
   - pytest -v -s plugins_tests/test_oot_registration_online.py # it needs a clean process
   - pytest -v -s plugins_tests/test_oot_registration_offline.py # it needs a clean process
   - pytest -v -s plugins_tests/lora_resolvers # unit tests for in-tree lora resolver plugins
-
-
-- label: GGUF Plugin
-  key: gguf-plugin
-  device: h200_18gb
-  timeout_in_minutes: 30
-  soft_fail: true
-  optional: true
-  source_file_dependencies:
-  - vllm/model_executor/layers/quantization
-  - tests/plugins_tests/test_gguf_plugin.py
-  commands:
-    - pip install "vllm-gguf-plugin >= 0.0.2"
-    - pytest -v -s plugins_tests/gguf

.buildkite/test_areas/quantization.yaml

@@ -21,18 +21,6 @@ steps:
   - uv pip install --system conch-triton-kernels
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
-- label: Quantized Fusions
-  key: quantized-fusions
-  timeout_in_minutes: 30
-  source_file_dependencies:
-  - tests/fusion
-  - vllm/model_executor/layers/fusion
-  - vllm/model_executor/kernels/linear
-  - vllm/model_executor/layers/quantization/compressed_tensors
-  - vllm/model_executor/layers/quantization/modelopt.py
-  commands:
-    - pytest -v -s fusion/
-
 - label: Quantized MoE Test (B200)
   key: quantized-moe-test-b200
   timeout_in_minutes: 60

.buildkite/test_areas/rust_frontend.yaml

@@ -99,13 +99,9 @@ steps:
   - vllm/v1/engine/
   - vllm/v1/worker/
   - tests/utils.py
-  - tests/v1/distributed/test_external_lb_dp.py
-  - tests/v1/distributed/test_hybrid_lb_dp.py
   - tests/v1/distributed/test_internal_lb_dp.py
   commands:
   - export VLLM_USE_RUST_FRONTEND=1
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - export NCCL_CUMEM_HOST_ENABLE=0
   - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_internal_lb_dp.py -k "not 4 and not server_info"
-  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py -k "not 4 and not server_info"
-  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py -k "not 4 and not server_info"

.buildkite/test_areas/spec_decode.yaml

@@ -39,7 +39,7 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       timeout_in_minutes: 65
       depends_on:
       - image-build-amd
@@ -78,7 +78,7 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       timeout_in_minutes: 65
       depends_on:
       - image-build-amd
@@ -103,7 +103,7 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
   mirror:
     amd:
-      device: mi325_1
+      device: mi300_1
       timeout_in_minutes: 50
       depends_on:
       - image-build-amd

.github/CODEOWNERS

@@ -80,7 +80,7 @@
 /vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery @NickLucche
 
 # Model runner V2
-/vllm/v1/worker/gpu @WoosukKwon @njhill @yewentao256
+/vllm/v1/worker/gpu @WoosukKwon @njhill
 /vllm/v1/worker/gpu/kv_connector.py @orozery
 
 # CI & building
@@ -120,6 +120,16 @@
 /vllm/model_executor/models/transformers @hmellor
 /tests/models/test_transformers.py @hmellor
 
+# Observability
+/vllm/config/observability.py @markmc
+/vllm/v1/metrics @markmc
+/tests/v1/metrics @markmc
+/vllm/tracing.py @markmc
+/tests/v1/tracing/test_tracing.py @markmc
+/vllm/config/kv_events.py @markmc
+/vllm/distributed/kv_events.py @markmc
+/tests/distributed/test_events.py @markmc
+
 # Docs
 /docs/mkdocs @hmellor
 /docs/**/*.yml @hmellor

.github/actionlint.yaml

@@ -1,5 +0,0 @@
-# Custom self-hosted runner labels (e.g. the autoscaling vllm-runners pool) so
-# actionlint doesn't flag them as unknown in `runs-on`.
-self-hosted-runner:
-  labels:
-    - vllm-runners

.github/dependabot.yml

@@ -21,6 +21,7 @@ updates:
       - dependency-name: "torchvision"
       - dependency-name: "xformers"
       - dependency-name: "lm-format-enforcer"
+      - dependency-name: "gguf"
       - dependency-name: "compressed-tensors"
       - dependency-name: "ray[cgraph]" # Ray Compiled Graph
       - dependency-name: "lm-eval"

.github/mergify.yml

@@ -144,12 +144,12 @@ pull_request_rules:
     - label != stale
     - or:
       - files~=^examples/.*mistral.*\.py
-      - files~=^tests/.*(?:mistral|voxtral|mixtral|pixtral).*\.py
-      - files~=^vllm/model_executor/models/.*(?:mistral|voxtral|mixtral|pixtral).*\.py
+      - files~=^tests/.*mistral.*\.py
+      - files~=^vllm/model_executor/models/.*mistral.*\.py
       - files~=^vllm/reasoning/.*mistral.*\.py
       - files~=^vllm/tool_parsers/.*mistral.*\.py
-      - files~=^vllm/transformers_utils/.*(?:mistral|voxtral|pixtral).*\.py
-      - title~=(?i)(?:mistral|ministral|voxtral|mixtral|pixtral)
+      - files~=^vllm/transformers_utils/.*mistral.*\.py
+      - title~=(?i)Mistral
   actions:
     label:
       add:
@@ -388,13 +388,9 @@ pull_request_rules:
     - or:
       - files~=^tests/tool_use/
       - files~=^tests/tool_parsers/
-      - files~=^tests/parser/
-      - files~=^tests/reasoning/
       - files~=^tests/entrypoints/openai/.*tool.*
       - files~=^tests/entrypoints/anthropic/.*tool.*
       - files~=^vllm/tool_parsers/
-      - files~=^vllm/parser/
-      - files~=^vllm/reasoning/
       - files=docs/features/tool_calling.md
       - files~=^examples/tool_calling/
   actions:

.github/workflows/pre-commit.yml

@@ -46,16 +46,12 @@ jobs:
   pre-commit:
     needs: pre-run-check
     if: always() && (needs.pre-run-check.result == 'success' || needs.pre-run-check.result == 'skipped')
-    runs-on: [self-hosted, linux, x64, vllm-runners]
+    runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
     - uses: actions/setup-python@83679a892e2d95755f2dac6acb0bfd1e9ac5d548 # v6.1.0
       with:
         python-version: "3.12"
-    # Provide shellcheck on PATH so tools/pre_commit/shellcheck.sh skips its
-    # wget + tar -xJ self-download, which the self-hosted runner image lacks
-    # (no wget/xz). Pinned to shellcheck 0.10.0 to match the script's "stable".
-    - run: python -m pip install shellcheck-py==0.10.0.1
     - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"
     - run: echo "::add-matcher::.github/workflows/matchers/markdownlint.json"
     - run: echo "::add-matcher::.github/workflows/matchers/mypy.json"

.github/workflows/scripts/build.sh

@@ -9,7 +9,7 @@ PATH=${cuda_home}/bin:$PATH
 LD_LIBRARY_PATH=${cuda_home}/lib64:$LD_LIBRARY_PATH
 
 # Install requirements
-if [ "$(echo "$2" | cut -d. -f1)" = "12" ]; then
+if [ "$(echo $2 | cut -d. -f1)" = "12" ]; then
     sed -i 's/^nvidia-cutlass-dsl\[cu13\]>=/nvidia-cutlass-dsl>=/' requirements/cuda.txt
 fi
 $python_executable -m pip install -r requirements/build/cuda.txt -r requirements/cuda.txt
@@ -17,10 +17,7 @@ $python_executable -m pip install -r requirements/build/cuda.txt -r requirements
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1
 # Make sure release wheels are built for the following architectures
-# Do not add +PTX here: vLLM filters torch's top-level PTX flag when it
-# converts global gencode flags into per-kernel arch lists. If a specific
-# kernel needs PTX, add +PTX to that kernel's CMake arch list instead.
-export TORCH_CUDA_ARCH_LIST="7.5 8.0 8.6 8.9 9.0 10.0 12.0"
+export TORCH_CUDA_ARCH_LIST="7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX"
 
 bash tools/check_repo.sh
 

.gitignore

@@ -15,9 +15,6 @@ vllm/third_party/flashmla/flash_mla_interface.py
 # DeepGEMM vendored package built from source
 vllm/third_party/deep_gemm/
 
-# fmha_sm100 vendored package built from source
-vllm/third_party/fmha_sm100/
-
 # triton jit
 .triton
 
@@ -236,7 +233,7 @@ actionlint
 shellcheck*/
 
 # Ignore moe/marlin_moe gen code
-csrc/libtorch_stable/moe/marlin_moe_wna16/kernel_*
+csrc/moe/marlin_moe_wna16/kernel_*
 
 # Ignore ep_kernels_workspace folder
 ep_kernels_workspace/

.pre-commit-config.yaml

@@ -21,7 +21,7 @@ repos:
   rev: v21.1.2
   hooks:
   - id: clang-format
-    exclude: 'csrc/libtorch_stable/moe/topk_softmax_kernels.cu|vllm/third_party/.*'
+    exclude: 'csrc/(moe/topk_softmax_kernels.cu|libtorch_stable/quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
     types_or: [c++, cuda]
     args: [--style=file, --verbose]
 - repo: https://github.com/DavidAnson/markdownlint-cli2

AGENTS.md

@@ -105,26 +105,6 @@ The line length limit for Python code is 88 characters. If you are not sure, use
 
 Use [Google-style docstrings](https://google.github.io/styleguide/pyguide.html#38-comments-and-docstrings) (`Args:`/`Returns:`/`Raises:` sections), not reStructuredText/Sphinx fields (`:param:`, `:return:`, `:rtype:`).
 
-### Coding style guidelines
-
-Follow these rules for all code changes in this repository:
-
-- Try to match existing code style.
-- Code should be self-documenting and self-explanatory.
-- Keep comments and docstrings minimal and concise.
-- Assume the reader is familiar with vLLM.
-
-### Diagnosing CI failures
-
-Buildkite logs are public; no login needed. Details: [docs/contributing/ci/failures.md](docs/contributing/ci/failures.md).
-
-```bash
-# All failed-job logs for a PR's latest build (current branch's PR if omitted):
-.buildkite/scripts/ci-fetch-log.sh --pr <PR>
-# Any Buildkite build or job URL also works:
-.buildkite/scripts/ci-fetch-log.sh "<buildkite_url>"
-```
-
 ### Commit messages
 
 Add attribution using commit trailers such as `Co-authored-by:` (other projects use `Assisted-by:` or `Generated-by:`). For example:

SECURITY.md

@@ -34,15 +34,6 @@ Vulnerabilities that cause denial of service or partial disruption, but do not a
 
 Minor issues such as informational disclosures, logging errors, non-exploitable flaws, or weaknesses that require local or high-privilege access and offer negligible impact. Examples include side channel attacks or hash collisions. These issues often have CVSS scores less than 4.0
 
-## Fix disclosure policy
-
-When a security report is accepted, the fix process depends on the severity:
-
-* **CRITICAL and HIGH severity**: Fixes are developed in a private security fork and coordinated with the prenotification group before public disclosure.
-* **MODERATE and LOW severity**: Fixes are developed and submitted as public pull requests. These issues do not require embargo since they do not enable arbitrary code execution or significant data breach, and public visibility accelerates community review and adoption of the fix.
-
-The vulnerability management team reserves the right to adjust the disclosure approach on a case-by-case basis, taking into account factors such as active exploitation, unusual attack surface, or coordination requirements with downstream vendors.
-
 ## Prenotification policy
 
 For certain security issues of CRITICAL, HIGH, or MODERATE severity level, we may prenotify certain organizations or vendors that ship vLLM. The purpose of this prenotification is to allow for a coordinated release of fixes for severe issues.

benchmarks/attention_benchmarks/README.md

@@ -108,6 +108,7 @@ python benchmark.py \
     --backends flash triton flashinfer \
     --batch-specs "q2k" "8q1s1k" "2q2k_32q1s1k" \
     --num-layers 10 \
+    --repeats 5 \
     --output-csv results.csv
 ```
 
@@ -163,17 +164,14 @@ python benchmark.py \
 # Model configuration
 --num-layers N                      # Number of layers
 --head-dim N                        # Head dimension
---v-head-dim N                      # Value head dimension (defaults to --head-dim)
 --num-q-heads N                     # Query heads
 --num-kv-heads N                    # KV heads
 --block-size N                      # Block size
---kv-lora-rank N                    # MLA KV LoRA rank
---qk-nope-head-dim N                # MLA non-RoPE QK head dim
---qk-rope-head-dim N                # MLA RoPE QK head dim
 
 # Benchmark settings
 --device DEVICE                     # Device (default: cuda:0)
---warmup-ms N                       # Warmup window in ms for triton do_bench
+--repeats N                         # Repetitions
+--warmup-iters N                    # Warmup iterations
 --profile-memory                    # Profile memory usage
 
 # Parameter sweeps
@@ -213,6 +211,8 @@ config = BenchmarkConfig(
     num_kv_heads=1,
     block_size=128,
     device="cuda:0",
+    repeats=5,
+    warmup_iters=3,
 )
 
 # CUTLASS MLA with specific num_kv_splits
@@ -253,10 +253,14 @@ formatter.save_json(results, "output.json")
 
 ## Tips
 
-**1. Save results** - Always use `--output-csv` or `--output-json`
+**1. Warmup matters** - Use `--warmup-iters 10` for stable results
 
-**2. Test incrementally** - Start with `--num-layers 1`
+**2. Multiple repeats** - Use `--repeats 20` for low variance
 
-**3. Extended grammar** - Leverage spec decode, chunked prefill patterns
+**3. Save results** - Always use `--output-csv` or `--output-json`
 
-**4. Parameter sweeps** - Use `--sweep-param` and `--sweep-values` to find optimal values
+**4. Test incrementally** - Start with `--num-layers 1 --repeats 1`
+
+**5. Extended grammar** - Leverage spec decode, chunked prefill patterns
+
+**6. Parameter sweeps** - Use `--sweep-param` and `--sweep-values` to find optimal values

benchmarks/attention_benchmarks/benchmark.py

@@ -26,9 +26,6 @@ Examples:
 """
 
 import argparse
-import os
-import shutil
-import subprocess
 import sys
 from dataclasses import replace
 from pathlib import Path
@@ -53,16 +50,6 @@ from common import (
 from vllm.v1.worker.workspace import init_workspace_manager
 
 
-def _str2bool(v) -> bool:
-    if isinstance(v, bool):
-        return v
-    if v.lower() in ("true", "1", "yes", "t"):
-        return True
-    if v.lower() in ("false", "0", "no", "f"):
-        return False
-    raise argparse.ArgumentTypeError(f"expected a boolean, got {v!r}")
-
-
 def run_standard_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
     """Run standard attention benchmark (Flash/Triton/FlashInfer)."""
     from runner import run_attention_benchmark
@@ -96,15 +83,13 @@ def run_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
         else:
             return run_standard_attention_benchmark(config)
     except Exception as e:
-        error_msg = str(e) or repr(e)
         return BenchmarkResult(
             config=config,
             mean_time=float("inf"),
-            median_time=float("inf"),
             std_time=0,
             min_time=float("inf"),
             max_time=float("inf"),
-            error=error_msg,
+            error=str(e),
         )
 
 
@@ -130,12 +115,9 @@ def run_model_parameter_sweep(
     """
     all_results = []
 
-    sweep_desc = (
-        f"{sweep.param_name} = {sweep.values}"
-        if sweep.param_name
-        else f"{len(sweep.values)} configurations"
+    console.print(
+        f"[yellow]Model sweep mode: testing {sweep.param_name} = {sweep.values}[/]"
     )
-    console.print(f"[yellow]Model sweep mode: testing {sweep_desc}[/]")
 
     total = len(backends) * len(batch_specs) * len(sweep.values)
 
@@ -143,9 +125,9 @@ def run_model_parameter_sweep(
         for backend in backends:
             for spec in batch_specs:
                 for value in sweep.values:
-                    # Create config with modified model parameter(s)
+                    # Create config with modified model parameter
                     config_args = base_config_args.copy()
-                    sweep.apply(config_args, value)
+                    config_args[sweep.param_name] = value
 
                     # Create config with original backend for running
                     clean_config = BenchmarkConfig(
@@ -162,21 +144,13 @@ def run_model_parameter_sweep(
                     all_results.append(result)
 
                     if not result.success:
-                        err_label = (
-                            f"{sweep.param_name}={value}"
-                            if sweep.param_name
-                            else f"{value}"
-                        )
                         console.print(
-                            f"[red]Error {backend} {spec} {err_label}"
-                            f": {result.error}[/]"
+                            f"[red]Error {backend} {spec} {sweep.param_name}="
+                            f"{value}: {result.error}[/]"
                         )
 
                     pbar.update(1)
 
-    if base_config_args.get("ncu_profile"):
-        return all_results
-
     # Display sweep results - create separate table for each parameter value
     console.print("\n[bold green]Model Parameter Sweep Results:[/]")
     formatter = ResultsFormatter(console)
@@ -210,10 +184,7 @@ def run_model_parameter_sweep(
     )
 
     for param_value in sorted_param_values:
-        label = (
-            f"{sweep.param_name} = {param_value}" if sweep.param_name else param_value
-        )
-        console.print(f"\n[bold cyan]{label}[/]")
+        console.print(f"\n[bold cyan]{sweep.param_name} = {param_value}[/]")
         param_results = by_param_value[param_value]
 
         # Create modified results with original backend names
@@ -229,9 +200,8 @@ def run_model_parameter_sweep(
         formatter.print_table(modified_results, backends, compare_to_fastest=True)
 
     # Show optimal backend for each (param_value, batch_spec) combination
-    sweep_name = sweep.param_name or "config"
     console.print(
-        f"\n[bold cyan]Optimal backend for each ({sweep_name}, batch_spec):[/]"
+        f"\n[bold cyan]Optimal backend for each ({sweep.param_name}, batch_spec):[/]"
     )
 
     # Group by (param_value, batch_spec)
@@ -266,10 +236,7 @@ def run_model_parameter_sweep(
     for param_value, spec in sorted_keys:
         # Print header when param value changes
         if param_value != current_param_value:
-            header = (
-                f"{sweep.param_name}={param_value}" if sweep.param_name else param_value
-            )
-            console.print(f"\n  [bold]{header}:[/]")
+            console.print(f"\n  [bold]{sweep.param_name}={param_value}:[/]")
             current_param_value = param_value
 
         results = by_param_and_spec[(param_value, spec)]
@@ -355,9 +322,6 @@ def run_parameter_sweep(
 
                     pbar.update(1)
 
-    if base_config_args.get("ncu_profile"):
-        return all_results
-
     # Display sweep results
     console.print("\n[bold green]Sweep Results:[/]")
     backend_labels = [sweep.get_label(b, v) for b in backends for v in sweep_values]
@@ -495,20 +459,6 @@ def main():
         help="Prefill backends to compare (fa2, fa3, fa4). "
         "Uses the first decode backend for impl construction.",
     )
-    parser.add_argument(
-        "--fp8-output-scale",
-        type=float,
-        help="Static per-tensor scale enabling the MLA prefill FP8-output "
-        "comparison on FA4 (fused write vs standalone post-quant).",
-    )
-    parser.add_argument(
-        "--fuse-quant-op",
-        nargs="+",
-        type=_str2bool,
-        help="FP8-output write path(s) to run: false = bf16 attention + "
-        "standalone static-FP8 quant, true = FA4 writes FP8 directly. "
-        "Default: both.",
-    )
 
     # Batch specifications
     parser.add_argument(
@@ -524,35 +474,11 @@ def main():
     parser.add_argument("--num-q-heads", type=int, default=32, help="Query heads")
     parser.add_argument("--num-kv-heads", type=int, default=8, help="KV heads")
     parser.add_argument("--block-size", type=int, default=16, help="Block size")
-    parser.add_argument(
-        "--v-head-dim",
-        type=int,
-        default=None,
-        help="Value head dimension (defaults to --head-dim if unset)",
-    )
-
-    # MLA-specific model dimensions
-    parser.add_argument(
-        "--kv-lora-rank", type=int, default=None, help="MLA KV LoRA rank"
-    )
-    parser.add_argument(
-        "--qk-nope-head-dim", type=int, default=None, help="MLA non-RoPE QK head dim"
-    )
-    parser.add_argument(
-        "--qk-rope-head-dim", type=int, default=None, help="MLA RoPE QK head dim"
-    )
 
     # Benchmark settings
     parser.add_argument("--device", default="cuda:0", help="Device")
-    parser.add_argument(
-        "--warmup-ms",
-        type=int,
-        default=None,
-        help=(
-            "Warmup window in ms for triton's do_bench (default: triton's own). "
-            "Has no effect with CUDA graphs; pass --no-cuda-graphs to use it."
-        ),
-    )
+    parser.add_argument("--repeats", type=int, default=1, help="Repetitions")
+    parser.add_argument("--warmup-iters", type=int, default=3, help="Warmup iterations")
     parser.add_argument("--profile-memory", action="store_true", help="Profile memory")
     parser.add_argument(
         "--kv-cache-dtype",
@@ -565,33 +491,10 @@ def main():
         action=argparse.BooleanOptionalAction,
         default=True,
         help=(
-            "Use triton do_bench_cudagraph (True) or do_bench (False) "
-            "for timing. CUDA graphs eliminate CPU launch overhead "
-            "(default: True)"
+            "Launch kernels with CUDA graphs to eliminate CPU overhead"
+            "in measurements (default: True)"
         ),
     )
-    parser.add_argument(
-        "--num-splits",
-        type=int,
-        default=None,
-        help="FlashAttention split-K factor (0=auto heuristic, 1=disabled, >1=force N)",
-    )
-    parser.add_argument(
-        "--ncu-profile",
-        action="store_true",
-        default=False,
-        help=(
-            "Enable Nsight Compute profiling mode. Automatically wraps the "
-            "script with ncu, capturing a profile with source correlation. "
-            "Use --ncu-output to set the output file name."
-        ),
-    )
-    parser.add_argument(
-        "--ncu-output",
-        type=str,
-        default="profile",
-        help="Output file name for ncu profile (default: 'profile').",
-    )
 
     # Parameter sweep (use YAML config for advanced sweeps)
     parser.add_argument(
@@ -642,12 +545,6 @@ def main():
         # Prefill backends (e.g., ["fa3", "fa4"])
         args.prefill_backends = yaml_config.get("prefill_backends", None)
 
-        # FP8 output benchmark knobs; CLI wins.
-        if args.fp8_output_scale is None:
-            args.fp8_output_scale = yaml_config.get("fp8_output_scale", None)
-        if args.fuse_quant_op is None:
-            args.fuse_quant_op = yaml_config.get("fuse_quant_op", None)
-
         # Check for special modes
         args.mode = yaml_config.get("mode", None)
 
@@ -679,28 +576,23 @@ def main():
             model = yaml_config["model"]
             args.num_layers = model.get("num_layers", args.num_layers)
             args.head_dim = model.get("head_dim", args.head_dim)
-            args.v_head_dim = model.get("v_head_dim", args.v_head_dim)
             args.num_q_heads = model.get("num_q_heads", args.num_q_heads)
             args.num_kv_heads = model.get("num_kv_heads", args.num_kv_heads)
             args.block_size = model.get("block_size", args.block_size)
-            # MLA-specific dimensions
-            args.kv_lora_rank = model.get("kv_lora_rank", args.kv_lora_rank)
-            args.qk_nope_head_dim = model.get("qk_nope_head_dim", args.qk_nope_head_dim)
-            args.qk_rope_head_dim = model.get("qk_rope_head_dim", args.qk_rope_head_dim)
 
         # Benchmark settings (top-level keys)
         if "device" in yaml_config:
             args.device = yaml_config["device"]
-        if "warmup_ms" in yaml_config:
-            args.warmup_ms = yaml_config["warmup_ms"]
+        if "repeats" in yaml_config:
+            args.repeats = yaml_config["repeats"]
+        if "warmup_iters" in yaml_config:
+            args.warmup_iters = yaml_config["warmup_iters"]
         if "profile_memory" in yaml_config:
             args.profile_memory = yaml_config["profile_memory"]
         if "kv_cache_dtype" in yaml_config:
             args.kv_cache_dtype = yaml_config["kv_cache_dtype"]
         if "cuda_graphs" in yaml_config:
             args.cuda_graphs = yaml_config["cuda_graphs"]
-        if "ncu_profile" in yaml_config:
-            args.ncu_profile = yaml_config["ncu_profile"]
 
         # Parameter sweep configuration
         if "parameter_sweep" in yaml_config:
@@ -720,7 +612,7 @@ def main():
         if "model_parameter_sweep" in yaml_config:
             sweep_config = yaml_config["model_parameter_sweep"]
             args.model_parameter_sweep = ModelParameterSweep(
-                param_name=sweep_config.get("param_name"),
+                param_name=sweep_config["param_name"],
                 values=sweep_config["values"],
                 label_format=sweep_config.get(
                     "label_format", "{backend}_{param_name}_{value}"
@@ -739,32 +631,6 @@ def main():
 
         console.print()
 
-    # Re-exec under ncu if --ncu-profile and not already inside ncu. This runs
-    # after YAML processing so ncu_profile set via config file is honored.
-    if args.ncu_profile and "_NCU_INNER" not in os.environ:
-        ncu = shutil.which("ncu")
-        if ncu is None:
-            print("Error: 'ncu' not found in PATH", file=sys.stderr)
-            sys.exit(1)
-        cmd = [
-            ncu,
-            "--profile-from-start",
-            "off",
-            "--set",
-            "full",
-            "--import-source",
-            "yes",
-            "-o",
-            args.ncu_output,
-            sys.executable,
-            *sys.argv,
-        ]
-        env = os.environ.copy()
-        env["CUTE_DSL_LINEINFO"] = "1"
-        env["_NCU_INNER"] = "1"
-        print(f"Launching: {' '.join(cmd)}")
-        sys.exit(subprocess.call(cmd, env=env))
-
     # Handle CLI-based parameter sweep (if not from YAML)
     if (
         (not hasattr(args, "parameter_sweep") or args.parameter_sweep is None)
@@ -789,18 +655,6 @@ def main():
     console.print(f"Batch specs: {', '.join(args.batch_specs)}")
     console.print(f"KV cache dtype: {args.kv_cache_dtype}")
     console.print(f"CUDA graphs: {args.cuda_graphs}")
-    if args.warmup_ms is not None and args.cuda_graphs:
-        console.print(
-            "[yellow]Warning: --warmup-ms is ignored with CUDA graphs "
-            "(do_bench_cudagraph warms up internally). Pass --no-cuda-graphs "
-            "to use it.[/]"
-        )
-    if args.num_splits == 0 and args.cuda_graphs:
-        console.print(
-            "[yellow]Warning: --num-splits 0 (FA3 heuristic) is not CUDA-graph "
-            "compatible and may fail or fall back. Pass --no-cuda-graphs or use "
-            "--num-splits >=1.[/]"
-        )
     console.print()
 
     init_workspace_manager(args.device)
@@ -808,68 +662,8 @@ def main():
     # Run benchmarks
     all_results = []
 
-    # Under ncu profiling the kernels run only to be captured by the profiler;
-    # timings are placeholder zeros, so the result tables and saved metrics are
-    # skipped. The Nsight Compute report (--ncu-output) holds the real data.
-    if args.ncu_profile:
-        console.print(
-            "[dim]ncu profiling enabled: result tables and saved metrics are "
-            "skipped (timings are placeholder zeros).[/]"
-        )
-
-    # FA4 fused FP8 output vs standalone post-quant, on the same fa4 kernel:
-    # the delta is the post-quant kernel the fused path removes.
-    fp8_output_scale = getattr(args, "fp8_output_scale", None)
-    if fp8_output_scale is not None:
-        decode_backend = backends[0]
-        fuse_variants = args.fuse_quant_op or [False, True]
-        label_of = {False: "post_quant", True: "fused"}
-        console.print(
-            f"[yellow]FP8 output comparison @ scale={fp8_output_scale} "
-            f"(prefill=fa4, decode impl={decode_backend})[/]"
-        )
-        fp8_results = []
-        total = len(fuse_variants) * len(args.batch_specs)
-        with tqdm(total=total, desc="FP8 output benchmarking") as pbar:
-            for spec in args.batch_specs:
-                for fuse in fuse_variants:
-                    config = BenchmarkConfig(
-                        backend=decode_backend,
-                        batch_spec=spec,
-                        num_layers=args.num_layers,
-                        head_dim=args.head_dim,
-                        num_q_heads=args.num_q_heads,
-                        num_kv_heads=args.num_kv_heads,
-                        block_size=args.block_size,
-                        device=args.device,
-                        repeats=args.repeats,
-                        warmup_iters=args.warmup_iters,
-                        profile_memory=args.profile_memory,
-                        kv_cache_dtype=args.kv_cache_dtype,
-                        use_cuda_graphs=args.cuda_graphs,
-                        prefill_backend="fa4",
-                    )
-                    result = run_benchmark(
-                        config, output_scale=fp8_output_scale, fuse_quant_op=fuse
-                    )
-                    label = label_of[fuse]
-                    labeled_config = replace(result.config, backend=label)
-                    result = replace(result, config=labeled_config)
-                    fp8_results.append(result)
-
-                    if not result.success:
-                        console.print(f"[red]Error {label} {spec}: {result.error}[/]")
-
-                    pbar.update(1)
-
-        console.print("\n[bold green]FP8 Output Results:[/]")
-        formatter = ResultsFormatter(console)
-        labels = [label_of[f] for f in fuse_variants]
-        formatter.print_table(fp8_results, labels, compare_to_fastest=True)
-        all_results = fp8_results
-
     # Handle special mode: decode_vs_prefill comparison
-    elif hasattr(args, "mode") and args.mode == "decode_vs_prefill":
+    if hasattr(args, "mode") and args.mode == "decode_vs_prefill":
         console.print("[yellow]Mode: Decode vs Prefill pipeline comparison[/]")
         console.print(
             "[dim]For each query length, testing both decode and prefill pipelines[/]"
@@ -914,11 +708,11 @@ def main():
                         num_kv_heads=args.num_kv_heads,
                         block_size=args.block_size,
                         device=args.device,
+                        repeats=args.repeats,
+                        warmup_iters=args.warmup_iters,
                         profile_memory=args.profile_memory,
                         kv_cache_dtype=args.kv_cache_dtype,
                         use_cuda_graphs=args.cuda_graphs,
-                        ncu_profile=args.ncu_profile,
-                        warmup_ms=args.warmup_ms,
                     )
 
                     # Add decode pipeline config
@@ -955,7 +749,6 @@ def main():
                         result = BenchmarkResult(
                             config=config,
                             mean_time=timing["mean"],
-                            median_time=timing.get("median", timing["mean"]),
                             std_time=timing["std"],
                             min_time=timing["min"],
                             max_time=timing["max"],
@@ -977,7 +770,6 @@ def main():
                         result = BenchmarkResult(
                             config=config,
                             mean_time=float("inf"),
-                            median_time=float("inf"),
                             std_time=0,
                             min_time=float("inf"),
                             max_time=float("inf"),
@@ -987,9 +779,6 @@ def main():
 
                 pbar.update(1)
 
-        if args.ncu_profile:
-            return
-
         # Display decode vs prefill results
         console.print("\n[bold green]Decode vs Prefill Results:[/]")
 
@@ -1069,20 +858,15 @@ def main():
         base_config_args = {
             "num_layers": args.num_layers,
             "head_dim": args.head_dim,
-            "v_head_dim": args.v_head_dim,
             "num_q_heads": args.num_q_heads,
             "num_kv_heads": args.num_kv_heads,
             "block_size": args.block_size,
             "device": args.device,
+            "repeats": args.repeats,
+            "warmup_iters": args.warmup_iters,
             "profile_memory": args.profile_memory,
             "kv_cache_dtype": args.kv_cache_dtype,
             "use_cuda_graphs": args.cuda_graphs,
-            "ncu_profile": args.ncu_profile,
-            "warmup_ms": args.warmup_ms,
-            "num_splits": args.num_splits,
-            "kv_lora_rank": args.kv_lora_rank,
-            "qk_nope_head_dim": args.qk_nope_head_dim,
-            "qk_rope_head_dim": args.qk_rope_head_dim,
         }
         all_results = run_model_parameter_sweep(
             backends,
@@ -1098,17 +882,15 @@ def main():
         base_config_args = {
             "num_layers": args.num_layers,
             "head_dim": args.head_dim,
-            "v_head_dim": args.v_head_dim,
             "num_q_heads": args.num_q_heads,
             "num_kv_heads": args.num_kv_heads,
             "block_size": args.block_size,
             "device": args.device,
+            "repeats": args.repeats,
+            "warmup_iters": args.warmup_iters,
             "profile_memory": args.profile_memory,
             "kv_cache_dtype": args.kv_cache_dtype,
             "use_cuda_graphs": args.cuda_graphs,
-            "ncu_profile": args.ncu_profile,
-            "warmup_ms": args.warmup_ms,
-            "num_splits": args.num_splits,
         }
         all_results = run_parameter_sweep(
             backends, args.batch_specs, base_config_args, args.parameter_sweep, console
@@ -1132,17 +914,15 @@ def main():
                             batch_spec=spec,
                             num_layers=args.num_layers,
                             head_dim=args.head_dim,
-                            v_head_dim=getattr(args, "v_head_dim", None),
                             num_q_heads=args.num_q_heads,
                             num_kv_heads=args.num_kv_heads,
                             block_size=args.block_size,
                             device=args.device,
+                            repeats=args.repeats,
+                            warmup_iters=args.warmup_iters,
                             profile_memory=args.profile_memory,
                             kv_cache_dtype=args.kv_cache_dtype,
                             use_cuda_graphs=args.cuda_graphs,
-                            ncu_profile=args.ncu_profile,
-                            warmup_ms=args.warmup_ms,
-                            num_splits=args.num_splits,
                         )
 
                         result = run_benchmark(config)
@@ -1155,10 +935,9 @@ def main():
 
                         pbar.update(1)
 
-            if not args.ncu_profile:
-                console.print("\n[bold green]Results:[/]")
-                formatter = ResultsFormatter(console)
-                formatter.print_table(decode_results, backends)
+            console.print("\n[bold green]Results:[/]")
+            formatter = ResultsFormatter(console)
+            formatter.print_table(decode_results, backends)
 
         # Run prefill backend comparison
         if prefill_backends:
@@ -1183,8 +962,9 @@ def main():
                             num_kv_heads=args.num_kv_heads,
                             block_size=args.block_size,
                             device=args.device,
+                            repeats=args.repeats,
+                            warmup_iters=args.warmup_iters,
                             profile_memory=args.profile_memory,
-                            warmup_ms=args.warmup_ms,
                             prefill_backend=pb,
                         )
 
@@ -1200,17 +980,16 @@ def main():
 
                         pbar.update(1)
 
-            if not args.ncu_profile:
-                console.print("\n[bold green]Prefill Backend Results:[/]")
-                formatter = ResultsFormatter(console)
-                formatter.print_table(
-                    prefill_results, prefill_backends, compare_to_fastest=True
-                )
+            console.print("\n[bold green]Prefill Backend Results:[/]")
+            formatter = ResultsFormatter(console)
+            formatter.print_table(
+                prefill_results, prefill_backends, compare_to_fastest=True
+            )
 
         all_results = decode_results + prefill_results
 
-    # Save results (skip ncu profiling runs: timings are placeholder zeros)
-    if all_results and not args.ncu_profile:
+    # Save results
+    if all_results:
         formatter = ResultsFormatter(console)
         if args.output_csv:
             formatter.save_csv(all_results, args.output_csv)

benchmarks/attention_benchmarks/common.py

@@ -15,8 +15,6 @@ from batch_spec import get_batch_type, parse_batch_spec
 from rich.console import Console
 from rich.table import Table
 
-from vllm.triton_utils import triton
-
 
 def batch_spec_sort_key(spec: str) -> tuple[int, int, int]:
     """
@@ -36,30 +34,6 @@ def batch_spec_sort_key(spec: str) -> tuple[int, int, int]:
         return (0, 0, 0)
 
 
-def run_do_bench(
-    benchmark_fn,
-    use_cuda_graphs: bool,
-    warmup_ms: int | None = None,
-) -> list[float]:
-    kwargs: dict[str, Any] = {"return_mode": "all"}
-    if use_cuda_graphs:
-        result = triton.testing.do_bench_cudagraph(benchmark_fn, **kwargs)
-    else:
-        if warmup_ms is not None:
-            kwargs["warmup"] = warmup_ms
-        result = triton.testing.do_bench(benchmark_fn, **kwargs)
-    return result
-
-
-def run_ncu_profile(benchmark_fn) -> None:
-    benchmark_fn()
-    torch.accelerator.synchronize()
-    torch.cuda.cudart().cudaProfilerStart()
-    benchmark_fn()
-    torch.accelerator.synchronize()
-    torch.cuda.cudart().cudaProfilerStop()
-
-
 # Mock classes for vLLM attention infrastructure
 
 
@@ -208,37 +182,18 @@ class ParameterSweep:
 
 @dataclass
 class ModelParameterSweep:
-    """Configuration for sweeping model configuration parameter(s).
+    """Configuration for sweeping a model configuration parameter."""
 
-    Supports two modes:
-    - Single param: param_name="head_dim", values=[128, 256, 512]
-    - Multi param: values=[{head_dim: 192, v_head_dim: 128}, {head_dim: 256}]
-      When values are dicts, each dict's keys are applied as config overrides.
-    """
-
-    param_name: str | None = None
-    values: list[Any] | None = None
-    label_format: str = "{backend}_{param_name}_{value}"
+    param_name: str  # Name of the model config parameter to sweep (e.g., "num_q_heads")
+    values: list[Any]  # List of values to test
+    label_format: str = "{backend}_{param_name}_{value}"  # Result label template
 
     def get_label(self, backend: str, value: Any) -> str:
         """Generate a label for a specific parameter value."""
-        if isinstance(value, dict):
-            return self.label_format.format(
-                backend=backend, param_name=self.param_name, value=value, **value
-            )
         return self.label_format.format(
             backend=backend, param_name=self.param_name, value=value
         )
 
-    def apply(self, config_args: dict, value: Any) -> None:
-        """Apply a sweep value to config args."""
-        if isinstance(value, dict):
-            config_args.update(value)
-        elif self.param_name is not None:
-            config_args[self.param_name] = value
-        else:
-            raise ValueError("param_name must be set if sweep values are not dicts")
-
 
 @dataclass
 class BenchmarkConfig:
@@ -253,10 +208,10 @@ class BenchmarkConfig:
     block_size: int
     device: str
     dtype: torch.dtype = torch.float16
+    repeats: int = 1
+    warmup_iters: int = 3
     profile_memory: bool = False
     use_cuda_graphs: bool = False
-    ncu_profile: bool = False
-    warmup_ms: int | None = None
 
     # "auto" or "fp8"
     kv_cache_dtype: str = "auto"
@@ -271,7 +226,6 @@ class BenchmarkConfig:
     # Backend-specific tuning
     num_kv_splits: int | None = None  # CUTLASS MLA
     reorder_batch_threshold: int | None = None  # FlashAttn MLA, FlashMLA
-    num_splits: int | None = None  # FlashAttention split-K (0=auto, 1=disabled)
 
 
 @dataclass
@@ -280,7 +234,6 @@ class BenchmarkResult:
 
     config: BenchmarkConfig
     mean_time: float  # seconds
-    median_time: float  # seconds
     std_time: float  # seconds
     min_time: float  # seconds
     max_time: float  # seconds
@@ -299,7 +252,6 @@ class BenchmarkResult:
         return {
             "config": asdict(self.config),
             "mean_time": self.mean_time,
-            "median_time": self.median_time,
             "std_time": self.std_time,
             "min_time": self.min_time,
             "max_time": self.max_time,

benchmarks/attention_benchmarks/configs/mla_decode.yaml

@@ -56,6 +56,8 @@ backends:
   - TOKENSPEED_MLA  # Blackwell + R1 dims + FP8 KV (use --kv-cache-dtype fp8)
 
 device: "cuda:0"
+repeats: 100
+warmup_iters: 10
 profile_memory: true
 
 # Backend-specific tuning

benchmarks/attention_benchmarks/configs/mla_fa4_fp8_output.yaml

@@ -1,44 +0,0 @@
-# MLA prefill FP8-output microbenchmark (FA4).
-# Compares the fused FP8 write against bf16 attention + a standalone static-FP8
-# quant; the delta is the post-quant kernel the fused path removes.
-# DeepSeek-Coder-V2-Lite dims; FA4 needs SM100/110.
-#
-# Usage:
-#   python benchmark.py --config configs/mla_fa4_fp8_output.yaml
-
-description: "MLA prefill FA4 fused-FP8 output vs post-quant"
-
-model:
-  name: "deepseek-v2-lite"
-  num_layers: 27
-  num_q_heads: 16
-  num_kv_heads: 1
-  head_dim: 576
-  kv_lora_rank: 512
-  qk_nope_head_dim: 128
-  qk_rope_head_dim: 64
-  v_head_dim: 128
-  block_size: 128
-
-# Pure prefill (q_len == kv_len) so every token goes through forward_mha.
-batch_specs:
-  - "q512"
-  - "q1k"
-  - "q2k"
-  - "q4k"
-  - "q8k"
-  - "2q4k"
-  - "4q4k"
-  - "8q4k"
-
-# Only used to construct the MLA impl; the pure-prefill specs skip decode.
-decode_backends:
-  - CUTLASS_MLA
-
-# Sweep the two FP8 write paths (prefill backend is fixed to fa4).
-fp8_output_scale: 0.1
-fuse_quant_op: [false, true]
-
-device: "cuda:0"
-repeats: 50
-warmup_iters: 10

benchmarks/attention_benchmarks/configs/mla_mixed_batch.yaml

@@ -51,6 +51,8 @@ backends:
   - FLASHMLA         # Hopper only
 
 device: "cuda:0"
+repeats: 5
+warmup_iters: 3
 profile_memory: true
 
 # Analyze chunked prefill workspace size impact

benchmarks/attention_benchmarks/configs/mla_prefill.yaml

@@ -124,3 +124,5 @@ prefill_backends:
   - tokenspeed
 
 device: "cuda:0"
+repeats: 20
+warmup_iters: 5

benchmarks/attention_benchmarks/configs/mla_sparse_decode.yaml

@@ -53,4 +53,6 @@ backends:
   - FLASHINFER_MLA_SPARSE
 
 device: "cuda:0"
+repeats: 100
+warmup_iters: 10
 profile_memory: true

benchmarks/attention_benchmarks/configs/mla_sparse_prefill.yaml

@@ -57,4 +57,6 @@ backends:
   - FLASHINFER_MLA_SPARSE
 
 device: "cuda:0"
+repeats: 10
+warmup_iters: 3
 profile_memory: true

benchmarks/attention_benchmarks/configs/reorder_threshold.yaml

@@ -63,6 +63,8 @@ model:
 
 # Benchmark settings
 device: "cuda:0"
+repeats: 15          # More repeats for spec decode variance
+warmup_iters: 5
 profile_memory: false
 
 # Output

benchmarks/attention_benchmarks/configs/speculative_decode.yaml

@@ -49,6 +49,8 @@ backends:
 
 # Benchmark settings
 device: "cuda:0"
+repeats: 10  # More repeats for statistical significance
+warmup_iters: 5
 profile_memory: false
 
 # Test these threshold values for optimization

benchmarks/attention_benchmarks/configs/standard_attention.yaml

@@ -43,4 +43,6 @@ backends:
   - FLASHINFER
 
 device: "cuda:0"
+repeats: 5
+warmup_iters: 3
 profile_memory: false

benchmarks/attention_benchmarks/configs/standard_decode.yaml

@@ -1,142 +0,0 @@
-# Standard attention decode benchmark configuration
-# Sweeps num_q_heads and num_kv_heads to isolate effects of:
-#   1. GQA ratio (fixed num_q_heads=32, vary num_kv_heads)
-#   2. Absolute head count (fixed 4:1 ratio, vary scale)
-
-model:
-  num_layers: 32
-  num_q_heads: 32 # Base value, overridden by sweep
-  num_kv_heads: 8 # Base value, overridden by sweep
-  head_dim: 128
-  block_size: 16
-
-# Head count sweep: each entry overrides num_q_heads, num_kv_heads, and
-# head_dim where it differs from the base (128). Head counts are per-GPU
-# (i.e. after TP sharding).
-#
-# Group A — vary GQA ratio (fixed q=32, head_dim=128):
-#   32:32 (MHA), 32:8 (GQA 4:1), 32:4 (GQA 8:1), 32:1 (MQA)
-#
-# Groups B-E — real model configs at various TP degrees:
-#   Model              head_dim  Full    TP2      TP4      TP8
-#   Llama 3 8B           128    32:8   16:4      8:2      4:1
-#   Llama 3 70B          128    64:8   32:4     16:2      8:1
-#   GPT-OSS 120B          64    64:8   32:4     16:2      8:1
-#   Llama 3 405B         128   128:8   64:4     32:2     16:1
-model_parameter_sweep:
-  values:
-    # --- head_dim=128 (Llama 3 family) ---
-    - { num_q_heads: 32, num_kv_heads: 32, head_dim: 128 } # MHA 1:1
-    - { num_q_heads: 32, num_kv_heads: 1, head_dim: 128 } # MQA 32:1
-    - { num_q_heads: 4, num_kv_heads: 1, head_dim: 128 } # Llama 3 8B TP8
-    - { num_q_heads: 8, num_kv_heads: 2, head_dim: 128 } # Llama 3 8B TP4
-    - { num_q_heads: 16, num_kv_heads: 4, head_dim: 128 } # Llama 3 8B TP2
-    - { num_q_heads: 32, num_kv_heads: 8, head_dim: 128 } # Llama 3 8B TP1 / GQA 4:1
-    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 128 } # Llama 3 70B TP8
-    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 128 } # Llama 3 70B TP4
-    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 128 } # Llama 3 70B TP2 / GQA 8:1
-    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 128 } # Llama 3 70B TP1
-    - { num_q_heads: 16, num_kv_heads: 1, head_dim: 128 } # Llama 3 405B TP8
-    - { num_q_heads: 32, num_kv_heads: 2, head_dim: 128 } # Llama 3 405B TP4
-    - { num_q_heads: 64, num_kv_heads: 4, head_dim: 128 } # Llama 3 405B TP2
-    - { num_q_heads: 128, num_kv_heads: 8, head_dim: 128 } # Llama 3 405B TP1
-    # --- head_dim=64 (GPT-OSS 120B) ---
-    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 64 } # GPT-OSS 120B TP8
-    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 64 } # GPT-OSS 120B TP4
-    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 64 } # GPT-OSS 120B TP2
-    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 64 } # GPT-OSS 120B TP1
-  label_format: "{backend}_q{num_q_heads}kv{num_kv_heads}d{head_dim}"
-
-batch_specs:
-  # ---- batch_size x seq_len grid (decode: q_len=1) ----
-  # Small grid for quick iteration. Uncomment for full sweep.
-
-  # Batch size 1
-  - "q1s1k"
-  - "q1s512"
-  - "q1s2k"
-  - "q1s4k"
-  - "q1s8k"
-  - "q1s16k"
-  - "q1s32k"
-
-  # Batch size 2
-  - "2q1s512"
-  - "2q1s1k"
-  - "2q1s2k"
-  - "2q1s4k"
-  - "2q1s8k"
-  - "2q1s16k"
-  - "2q1s32k"
-
-  # Batch size 4
-  - "4q1s512"
-  - "4q1s1k"
-  - "4q1s2k"
-  - "4q1s4k"
-  - "4q1s8k"
-  - "4q1s16k"
-  - "4q1s32k"
-
-  # Batch size 8
-  - "8q1s1k"
-  - "8q1s512"
-  - "8q1s2k"
-  - "8q1s4k"
-  - "8q1s8k"
-  - "8q1s16k"
-  - "8q1s32k"
-
-  # Batch size 16
-  - "16q1s512"
-  - "16q1s1k"
-  - "16q1s2k"
-  - "16q1s4k"
-  - "16q1s8k"
-  - "16q1s16k"
-  - "16q1s32k"
-
-  # Batch size 32
-  - "32q1s512"
-  - "32q1s1k"
-  - "32q1s2k"
-  - "32q1s4k"
-  - "32q1s8k"
-  - "32q1s16k"
-  - "32q1s32k"
-
-  # Batch size 64
-  - "64q1s1k"
-  - "64q1s512"
-  - "64q1s2k"
-  - "64q1s4k"
-  - "64q1s8k"
-  - "64q1s16k"
-  - "64q1s32k"
-
-  # Batch size 128
-  - "128q1s512"
-  - "128q1s1k"
-  - "128q1s2k"
-  - "128q1s4k"
-  - "128q1s8k"
-  - "128q1s16k"
-  - "128q1s32k"
-
-  # Batch size 256
-  - "256q1s1k"
-  - "256q1s512"
-  - "256q1s2k"
-  - "256q1s4k"
-  - "256q1s8k"
-  - "256q1s16k"
-  - "256q1s32k"
-
-# Available backends: FLASH_ATTN, TRITON_ATTN, FLASHINFER
-backends:
-  - FLASH_ATTN
-  - TRITON_ATTN
-  - FLASHINFER
-
-device: "cuda:0"
-profile_memory: false

benchmarks/attention_benchmarks/configs/standard_prefill.yaml

@@ -1,108 +0,0 @@
-# Standard attention prefill benchmark configuration
-# Sweeps num_q_heads and num_kv_heads to isolate effects of:
-#   1. GQA ratio (fixed num_q_heads=32, vary num_kv_heads)
-#   2. Absolute head count (fixed 4:1 ratio, vary scale)
-
-model:
-  num_layers: 32
-  num_q_heads: 32 # Base value, overridden by sweep
-  num_kv_heads: 8 # Base value, overridden by sweep
-  head_dim: 128
-  block_size: 16
-
-# Head count sweep: each entry overrides num_q_heads, num_kv_heads, and
-# head_dim where it differs from the base (128). Head counts are per-GPU
-# (i.e. after TP sharding).
-#
-# Group A — vary GQA ratio (fixed q=32, head_dim=128):
-#   32:32 (MHA), 32:8 (GQA 4:1), 32:4 (GQA 8:1), 32:1 (MQA)
-#
-# Groups B-E — real model configs at various TP degrees:
-#   Model              head_dim  Full    TP2      TP4      TP8
-#   Llama 3 8B           128    32:8   16:4      8:2      4:1
-#   Llama 3 70B          128    64:8   32:4     16:2      8:1
-#   GPT-OSS 120B          64    64:8   32:4     16:2      8:1
-#   Llama 3 405B         128   128:8   64:4     32:2     16:1
-model_parameter_sweep:
-  values:
-    # --- head_dim=128 (Llama 3 family) ---
-    - { num_q_heads: 32, num_kv_heads: 32, head_dim: 128 } # MHA 1:1
-    - { num_q_heads: 32, num_kv_heads: 1, head_dim: 128 } # MQA 32:1
-    - { num_q_heads: 4, num_kv_heads: 1, head_dim: 128 } # Llama 3 8B TP8
-    - { num_q_heads: 8, num_kv_heads: 2, head_dim: 128 } # Llama 3 8B TP4
-    - { num_q_heads: 16, num_kv_heads: 4, head_dim: 128 } # Llama 3 8B TP2
-    - { num_q_heads: 32, num_kv_heads: 8, head_dim: 128 } # Llama 3 8B TP1 / GQA 4:1
-    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 128 } # Llama 3 70B TP8
-    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 128 } # Llama 3 70B TP4
-    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 128 } # Llama 3 70B TP2 / GQA 8:1
-    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 128 } # Llama 3 70B TP1
-    - { num_q_heads: 16, num_kv_heads: 1, head_dim: 128 } # Llama 3 405B TP8
-    - { num_q_heads: 32, num_kv_heads: 2, head_dim: 128 } # Llama 3 405B TP4
-    - { num_q_heads: 64, num_kv_heads: 4, head_dim: 128 } # Llama 3 405B TP2
-    - { num_q_heads: 128, num_kv_heads: 8, head_dim: 128 } # Llama 3 405B TP1
-    # --- head_dim=64 (GPT-OSS 120B) ---
-    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 64 } # GPT-OSS 120B TP8
-    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 64 } # GPT-OSS 120B TP4
-    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 64 } # GPT-OSS 120B TP2
-    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 64 } # GPT-OSS 120B TP1
-  label_format: "{backend}_q{num_q_heads}kv{num_kv_heads}d{head_dim}"
-
-batch_specs:
-  # ---- batch_size x prefill_len grid (prefill: q_len == seq_len) ----
-  # Total tokens = batch_size * prefill_len, and prefill compute scales with
-  # prefill_len^2, so the largest cells are expensive. Trim batch sizes or
-  # lengths for quick iteration.
-
-  # Batch size 1
-  - "q512"
-  - "q1k"
-  - "q2k"
-  - "q4k"
-  - "q8k"
-  - "q16k"
-  - "q32k"
-
-  # Batch size 2
-  - "2q512"
-  - "2q1k"
-  - "2q2k"
-  - "2q4k"
-  - "2q8k"
-  - "2q16k"
-  - "2q32k"
-
-  # Batch size 4
-  - "4q512"
-  - "4q1k"
-  - "4q2k"
-  - "4q4k"
-  - "4q8k"
-  - "4q16k"
-  - "4q32k"
-
-  # Batch size 8
-  - "8q512"
-  - "8q1k"
-  - "8q2k"
-  - "8q4k"
-  - "8q8k"
-  - "8q16k"
-  - "8q32k"
-
-  # Batch size 16
-  - "16q512"
-  - "16q1k"
-  - "16q2k"
-  - "16q4k"
-  - "16q8k"
-  - "16q16k"
-  - "16q32k"
-
-# Available backends: FLASH_ATTN, TRITON_ATTN, FLASHINFER
-backends:
-  - FLASH_ATTN
-  - TRITON_ATTN
-  - FLASHINFER
-
-device: "cuda:0"
-profile_memory: false

benchmarks/attention_benchmarks/mla_runner.py

@@ -8,8 +8,6 @@ This module provides helpers for running MLA backends without
 needing full VllmConfig integration.
 """
 
-import statistics
-
 import numpy as np
 import torch
 from batch_spec import parse_batch_spec
@@ -19,8 +17,6 @@ from common import (
     MockIndexer,
     MockKVBProj,
     MockLayer,
-    run_do_bench,
-    run_ncu_profile,
     setup_mla_dims,
 )
 
@@ -708,8 +704,6 @@ def _run_single_benchmark(
     device: torch.device,
     indexer=None,
     kv_cache_dtype: str | None = None,
-    output_scale: float | None = None,
-    fuse_quant_op: bool = False,
 ) -> BenchmarkResult:
     """
     Run a single benchmark iteration.
@@ -723,11 +717,6 @@ def _run_single_benchmark(
         mla_dims: MLA dimension configuration
         device: Target device
         indexer: Optional MockIndexer for sparse backends
-        output_scale: Static per-tensor FP8 scale for prefill output. None
-            keeps the plain bf16 output (no quantization).
-        fuse_quant_op: With output_scale set, True lets the prefill kernel write
-            FP8 directly; False runs bf16 attention then a standalone static-FP8
-            quant. The delta isolates the saved post-quant kernel.
 
     Returns:
         BenchmarkResult with timing statistics
@@ -831,86 +820,63 @@ def _run_single_benchmark(
             num_prefill, mla_dims, query_fmt, device, torch.bfloat16
         )
 
-    # Prefill FP8 output: fused (kernel writes e4m3) vs separate post-quant.
-    prefill_fp8_output = None
-    prefill_output_scale = None
-    prefill_quant_op = None
-    if has_prefill and output_scale is not None:
-        from vllm.platforms import current_platform
-
-        prefill_output_scale = torch.tensor(
-            [output_scale], device=device, dtype=torch.float32
-        )
-        if fuse_quant_op:
-            prefill_fp8_output = torch.empty_like(
-                prefill_inputs["output"], dtype=current_platform.fp8_dtype()
-            )
-        else:
-            from vllm.model_executor.layers.quantization.input_quant_fp8 import (
-                QuantFP8,
-            )
-            from vllm.model_executor.layers.quantization.utils.quant_utils import (
-                GroupShape,
-            )
-
-            prefill_quant_op = QuantFP8(static=True, group_shape=GroupShape.PER_TENSOR)
-
-    fused_output = output_scale is not None and fuse_quant_op
-
-    # Build forward function (runs a single decode/prefill pass)
+    # Build forward function
     def forward_fn():
         results = []
         if has_decode:
             results.append(impl.forward_mqa(decode_inputs, kv_cache, metadata, layer))
         if has_prefill:
-            out = impl.forward_mha(
-                prefill_inputs["q"],
-                prefill_inputs["k_c_normed"],
-                prefill_inputs["k_pe"],
-                kv_cache,
-                metadata,
-                prefill_inputs["k_scale"],
-                prefill_fp8_output if fused_output else prefill_inputs["output"],
-                prefill_output_scale if fused_output else None,
-            )
-            if fused_output:
-                out = prefill_fp8_output
-            elif prefill_quant_op is not None:
-                out, _ = prefill_quant_op(
-                    prefill_inputs["output"], prefill_output_scale
+            results.append(
+                impl.forward_mha(
+                    prefill_inputs["q"],
+                    prefill_inputs["k_c_normed"],
+                    prefill_inputs["k_pe"],
+                    kv_cache,
+                    metadata,
+                    prefill_inputs["k_scale"],
+                    prefill_inputs["output"],
                 )
-            results.append(out)
+            )
         return results[0] if len(results) == 1 else tuple(results)
 
-    def benchmark_fn():
-        for _ in range(config.num_layers):
+    # Warmup
+    for _ in range(config.warmup_iters):
+        forward_fn()
+    torch.accelerator.synchronize()
+
+    # Optionally capture a CUDA graph after warmup.
+    # Graph replay eliminates CPU launch overhead so timings reflect pure
+    # kernel time.
+    if config.use_cuda_graphs:
+        graph = torch.cuda.CUDAGraph()
+        with torch.cuda.graph(graph):
             forward_fn()
+        benchmark_fn = graph.replay
+    else:
+        benchmark_fn = forward_fn
 
-    if config.ncu_profile:
-        run_ncu_profile(benchmark_fn)
-        return BenchmarkResult(
-            config=config,
-            mean_time=0.0,
-            median_time=0.0,
-            std_time=0.0,
-            min_time=0.0,
-            max_time=0.0,
-            throughput_tokens_per_sec=0.0,
-        )
+    # Benchmark
+    times = []
+    for _ in range(config.repeats):
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
 
-    all_ms = run_do_bench(benchmark_fn, config.use_cuda_graphs, config.warmup_ms)
+        start.record()
+        for _ in range(config.num_layers):
+            benchmark_fn()
+        end.record()
 
-    # Convert ms to seconds per layer
-    times = [t / 1000.0 / config.num_layers for t in all_ms]
-    mean_time = statistics.mean(times)
+        torch.accelerator.synchronize()
+        elapsed_ms = start.elapsed_time(end)
+        times.append(elapsed_ms / 1000.0 / config.num_layers)
 
+    mean_time = float(np.mean(times))
     return BenchmarkResult(
         config=config,
         mean_time=mean_time,
-        median_time=statistics.median(times),
-        std_time=statistics.stdev(times) if len(times) > 1 else 0.0,
-        min_time=min(times),
-        max_time=max(times),
+        std_time=float(np.std(times)),
+        min_time=float(np.min(times)),
+        max_time=float(np.max(times)),
         throughput_tokens_per_sec=total_q / mean_time if mean_time > 0 else 0,
     )
 
@@ -920,8 +886,6 @@ def _run_mla_benchmark_batched(
     configs_with_params: list[tuple],  # [(config, threshold, num_splits), ...]
     index_topk: int = 2048,
     prefill_backend: str | None = None,
-    output_scale: float | None = None,
-    fuse_quant_op: bool = False,
 ) -> list[BenchmarkResult]:
     """
     Unified batched MLA benchmark runner for all backends.
@@ -1061,8 +1025,6 @@ def _run_mla_benchmark_batched(
                     device,
                     indexer=indexer,
                     kv_cache_dtype=kv_cache_dtype,
-                    output_scale=output_scale,
-                    fuse_quant_op=fuse_quant_op,
                 )
                 results.append(result)
 
@@ -1090,8 +1052,6 @@ def run_mla_benchmark(
     num_kv_splits: int | None = None,
     index_topk: int = 2048,
     prefill_backend: str | None = None,
-    output_scale: float | None = None,
-    fuse_quant_op: bool = False,
 ) -> BenchmarkResult | list[BenchmarkResult]:
     """
     Unified MLA benchmark runner for all backends.
@@ -1111,9 +1071,6 @@ def run_mla_benchmark(
         index_topk: Topk value for sparse MLA backends (default 2048)
         prefill_backend: Prefill backend name (e.g., "fa3", "fa4").
             When set, forces the specified FlashAttention version for prefill.
-        output_scale: Static per-tensor FP8 scale for prefill output (None = bf16).
-        fuse_quant_op: With output_scale set, fuse the FP8 write into the prefill
-            kernel vs a standalone post-quant kernel. See _run_single_benchmark.
 
     Returns:
         BenchmarkResult (single mode) or list of BenchmarkResult (batched mode)
@@ -1138,12 +1095,7 @@ def run_mla_benchmark(
 
     # Use unified batched execution
     results = _run_mla_benchmark_batched(
-        backend,
-        configs_with_params,
-        index_topk,
-        prefill_backend=prefill_backend,
-        output_scale=output_scale,
-        fuse_quant_op=fuse_quant_op,
+        backend, configs_with_params, index_topk, prefill_backend=prefill_backend
     )
 
     # Return single result or list based on input

benchmarks/attention_benchmarks/runner.py

@@ -9,20 +9,13 @@ This module provides helpers for running standard attention backends
 """
 
 import logging
-import statistics
 import types
 from contextlib import contextmanager
 
+import numpy as np
 import torch
 from batch_spec import parse_batch_spec, reorder_for_flashinfer
-from common import (
-    BenchmarkConfig,
-    BenchmarkResult,
-    MockLayer,
-    get_attention_scale,
-    run_do_bench,
-    run_ncu_profile,
-)
+from common import BenchmarkConfig, BenchmarkResult, MockLayer, get_attention_scale
 
 from vllm.config import (
     CacheConfig,
@@ -215,13 +208,6 @@ def _create_backend_impl(
 
     scale = get_attention_scale(config.head_dim)
 
-    # Set v_head_dim for diff-headdim backends. Always reset (defaulting to
-    # head_dim) so a prior run's value doesn't leak into this one via the
-    # backend's class-level state.
-    if hasattr(backend_class, "set_head_size_v"):
-        v_dim = config.v_head_dim if config.v_head_dim is not None else config.head_dim
-        backend_class.set_head_size_v(v_dim)
-
     impl = backend_class.get_impl_cls()(
         num_heads=config.num_q_heads,
         head_size=config.head_dim,
@@ -314,7 +300,6 @@ def _create_input_tensors(
         from vllm.platforms import current_platform
 
         q_dtype = current_platform.fp8_dtype()
-    v_dim = config.v_head_dim if config.v_head_dim is not None else config.head_dim
     q_list = [
         torch.randn(
             total_q, config.num_q_heads, config.head_dim, device=device, dtype=dtype
@@ -328,7 +313,9 @@ def _create_input_tensors(
         for _ in range(config.num_layers)
     ]
     v_list = [
-        torch.randn(total_q, config.num_kv_heads, v_dim, device=device, dtype=dtype)
+        torch.randn(
+            total_q, config.num_kv_heads, config.head_dim, device=device, dtype=dtype
+        )
         for _ in range(config.num_layers)
     ]
     return q_list, k_list, v_list
@@ -402,17 +389,14 @@ def _run_single_benchmark(
     device: torch.device,
     dtype: torch.dtype,
 ) -> tuple:
-    """Run single benchmark using triton's do_bench_cudagraph/do_bench.
-
-    Returns:
-        (timing_stats, mem_stats) where timing_stats is a dict with
-        mean/std/min/max in seconds per layer.
-    """
+    """Run single benchmark iteration with warmup and timing loop."""
     total_q = q_list[0].shape[0]
-    v_dim = config.v_head_dim if config.v_head_dim is not None else config.head_dim
-    out = torch.empty(total_q, config.num_q_heads, v_dim, device=device, dtype=dtype)
+    out = torch.empty(
+        total_q, config.num_q_heads, config.head_dim, device=device, dtype=dtype
+    )
 
-    def benchmark_fn():
+    # Warmup
+    for _ in range(config.warmup_iters):
         for i in range(config.num_layers):
             impl.forward(
                 layer,
@@ -423,22 +407,52 @@ def _run_single_benchmark(
                 attn_metadata,
                 output=out,
             )
+    torch.accelerator.synchronize()
 
-    if config.ncu_profile:
-        run_ncu_profile(benchmark_fn)
-        timing_stats = dict.fromkeys(("mean", "median", "std", "min", "max"), 0.0)
+    # Optionally capture a CUDA graph after warmup.
+    # Graph replay eliminates CPU launch overhead so timings reflect pure
+    # kernel time.
+    if config.use_cuda_graphs:
+        graph = torch.cuda.CUDAGraph()
+        with torch.cuda.graph(graph):
+            for i in range(config.num_layers):
+                impl.forward(
+                    layer,
+                    q_list[i],
+                    k_list[i],
+                    v_list[i],
+                    cache_list[i],
+                    attn_metadata,
+                    output=out,
+                )
+        benchmark_fn = graph.replay
     else:
-        all_ms = run_do_bench(benchmark_fn, config.use_cuda_graphs, config.warmup_ms)
 
-        # Convert ms to seconds per layer
-        times = [t / 1000.0 / config.num_layers for t in all_ms]
-        timing_stats = {
-            "mean": statistics.mean(times),
-            "std": statistics.stdev(times) if len(times) > 1 else 0.0,
-            "min": min(times),
-            "max": max(times),
-            "median": statistics.median(times),
-        }
+        def benchmark_fn():
+            for i in range(config.num_layers):
+                impl.forward(
+                    layer,
+                    q_list[i],
+                    k_list[i],
+                    v_list[i],
+                    cache_list[i],
+                    attn_metadata,
+                    output=out,
+                )
+
+    # Benchmark
+    times = []
+    for _ in range(config.repeats):
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+
+        start.record()
+        benchmark_fn()
+        end.record()
+
+        torch.accelerator.synchronize()
+        elapsed_ms = start.elapsed_time(end)
+        times.append(elapsed_ms / 1000.0 / config.num_layers)  # seconds per layer
 
     mem_stats = {}
     if config.profile_memory:
@@ -447,7 +461,7 @@ def _run_single_benchmark(
             "reserved_mb": torch.accelerator.memory_reserved(device) / 1024**2,
         }
 
-    return timing_stats, mem_stats
+    return times, mem_stats
 
 
 # ============================================================================
@@ -527,12 +541,6 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 common_attn_metadata=common_metadata,
             )
 
-            # Override num_splits for split-K testing (FlashAttention only)
-            if config.num_splits is not None and hasattr(
-                attn_metadata, "max_num_splits"
-            ):
-                attn_metadata.max_num_splits = config.num_splits
-
             # Only quantize queries when the impl supports it
             quantize_query = config.kv_cache_dtype.startswith("fp8") and getattr(
                 impl, "supports_quant_query_input", False
@@ -545,7 +553,7 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 config, max_num_blocks, backend_class, device, dtype
             )
 
-            timing_stats, mem_stats = _run_single_benchmark(
+            times, mem_stats = _run_single_benchmark(
                 config,
                 impl,
                 layer,
@@ -558,16 +566,15 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 dtype,
             )
 
-    mean_time = timing_stats["mean"]
+    mean_time = np.mean(times)
     throughput = total_q / mean_time if mean_time > 0 else 0
 
     return BenchmarkResult(
         config=config,
         mean_time=mean_time,
-        median_time=timing_stats["median"],
-        std_time=timing_stats["std"],
-        min_time=timing_stats["min"],
-        max_time=timing_stats["max"],
+        std_time=np.std(times),
+        min_time=np.min(times),
+        max_time=np.max(times),
         throughput_tokens_per_sec=throughput,
         memory_allocated_mb=mem_stats.get("allocated_mb"),
         memory_reserved_mb=mem_stats.get("reserved_mb"),

benchmarks/benchmark_hidden_state_extraction.py

@@ -92,6 +92,7 @@ def run_baseline(
     llm = LLM(
         model=model,
         enable_prefix_caching=False,
+        enable_chunked_prefill=False,
         **extra_args,
     )
     sampling_params = SamplingParams(max_tokens=1)
@@ -193,6 +194,7 @@ async def _run_extraction_async(
     engine_args = AsyncEngineArgs(
         model=model,
         enable_prefix_caching=False,
+        enable_chunked_prefill=False,
         max_num_batched_tokens=40960,
         max_model_len=40960,
         speculative_config={

benchmarks/benchmark_pin_memory.py

@@ -1,358 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Benchmark and regression-test pinned (page-locked) CPU memory for vLLM.
-
-Verifies that enabling pinned memory does not regress throughput or latency
-compared to unpinned memory.  Each condition runs in an isolated ``spawn``
-subprocess so both start from a cold CUDA context, giving an unbiased
-comparison.
-
-Usage
------
-Run all tests with the default model::
-
-    python benchmarks/benchmark_pin_memory.py -v
-
-Override the model and optional max-model-len::
-
-    python benchmarks/benchmark_pin_memory.py --model unsloth/Qwen3-1.7B -v
-    python benchmarks/benchmark_pin_memory.py --model unsloth/Qwen3-1.7B \
-        --max-model-len 8192 -v
-
-Run only throughput or latency tests::
-
-    python benchmarks/benchmark_pin_memory.py -v -k test_throughput
-    python benchmarks/benchmark_pin_memory.py -v -k test_latency
-
-Run only the v1 or v2 runner variant::
-
-    python benchmarks/benchmark_pin_memory.py -v -k v1
-    python benchmarks/benchmark_pin_memory.py -v -k v2
-
-Note: on WSL2, v1 runner tests are skipped because pin memory is not available
-for the v1 runner without cpu_offload_gb.  Run on other platforms to exercise v1.
-"""
-
-import argparse
-import json
-import multiprocessing
-import sys
-import tempfile
-
-import pytest
-
-# Allow up to 2% degradation.  Both benchmark runs start from an identical
-# cold CUDA context (separate spawn subprocesses), so the measured difference
-# reflects the genuine pin_memory overhead rather than cold/warm ordering bias.
-_THROUGHPUT_TOLERANCE = 0.98
-_THROUGHPUT_NUM_REQUESTS = 200
-_THROUGHPUT_INPUT_LEN = 128
-_THROUGHPUT_OUTPUT_LEN = 512
-_THROUGHPUT_MAX_NUM_SEQS = 128
-
-# Latency benchmark constants — match latency.py defaults.
-_LATENCY_TOLERANCE = 1.02  # Allow up to 2% latency regression.
-_LATENCY_BATCH_SIZE = 64
-_LATENCY_INPUT_LEN = 32
-_LATENCY_OUTPUT_LEN = 128
-_LATENCY_WARMUP_ITERS = 5
-_LATENCY_BENCH_ITERS = 15
-
-_DEFAULT_MODEL = "unsloth/Qwen3-1.7B"
-_DEFAULT_MAX_MODEL_LEN = 16384
-
-
-def _benchmark_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(add_help=False)
-    parser.add_argument("--model", default=_DEFAULT_MODEL)
-    parser.add_argument("--max-model-len", type=int, default=_DEFAULT_MAX_MODEL_LEN)
-    args, _ = parser.parse_known_args()
-    return args
-
-
-@pytest.fixture
-def model() -> str:
-    return _benchmark_args().model
-
-
-@pytest.fixture
-def max_model_len() -> int:
-    return _benchmark_args().max_model_len
-
-
-def _skip_if_pin_memory_not_available(engine_args_kwargs: dict) -> None:
-    """Skip the current pytest test if pin_memory is unavailable for this config."""
-    import vllm.utils.platform_utils as pu
-    from vllm.config import set_current_vllm_config
-    from vllm.engine.arg_utils import EngineArgs
-
-    vllm_config = EngineArgs(**engine_args_kwargs).create_engine_config()
-    with set_current_vllm_config(vllm_config):
-        pu.is_pin_memory_available.cache_clear()
-        if not pu.is_pin_memory_available():
-            import os
-
-            runner = "v2" if os.environ.get("VLLM_USE_V2_MODEL_RUNNER") == "1" else "v1"
-            model = engine_args_kwargs.get("model", "unknown")
-            print(
-                f"\033[33mSKIP: pin_memory not available for "
-                f"{runner} runner, model={model}\033[0m"
-            )
-            pytest.skip("pin_memory not available for this configuration")
-
-
-def _throughput_worker(
-    pin: bool,
-    engine_args_kwargs: dict,
-    q: "multiprocessing.Queue[float]",
-    v2_mode: bool = False,
-) -> None:
-    """Run throughput benchmark in a fresh spawn subprocess.
-
-    Delegates to vllm/benchmarks/throughput.py main() using the random dataset,
-    so the methodology matches the official benchmark.  Results are written to a
-    temp JSON file and forwarded through the queue as tokens/s.
-
-    v2_mode: when True, monkeypatches is_uva_available() to always return True
-    so the v2 model runner's UVA buffers remain functional even when pin=False.
-    This isolates the non-UVA pin_memory paths in v2.
-    """
-    import vllm.utils.platform_utils as pu
-    from vllm.platforms import current_platform
-
-    pu.is_pin_memory_available.cache_clear()
-    pu.is_uva_available.cache_clear()
-    type(current_platform).is_pin_memory_available = classmethod(lambda cls: pin)
-    if v2_mode:
-        pu.is_uva_available = lambda: True
-
-    from vllm.benchmarks.throughput import add_cli_args
-    from vllm.benchmarks.throughput import main as throughput_main
-
-    parser = argparse.ArgumentParser()
-    add_cli_args(parser)
-    args = parser.parse_args([])
-
-    for key, val in engine_args_kwargs.items():
-        setattr(args, key, val)
-    args.max_num_seqs = _THROUGHPUT_MAX_NUM_SEQS
-    args.dataset_name = "random"
-    args.input_len = _THROUGHPUT_INPUT_LEN
-    args.output_len = _THROUGHPUT_OUTPUT_LEN
-    # Nullify defaults that conflict with explicit input/output_len.
-    args.random_input_len = None
-    args.random_output_len = None
-    args.random_prefix_len = None
-    args.num_prompts = _THROUGHPUT_NUM_REQUESTS
-    args.seed = 0
-    args.disable_detokenize = True
-
-    with tempfile.NamedTemporaryFile(mode="w", suffix=".json", delete=False) as f:
-        tmp_path = f.name
-    args.output_json = tmp_path
-
-    throughput_main(args)
-
-    with open(tmp_path) as f:
-        results = json.load(f)
-    q.put(results["tokens_per_second"])
-
-
-def _run_throughput_benchmark(
-    pin: bool,
-    engine_args_kwargs: dict,
-    v2_mode: bool = False,
-) -> float:
-    ctx = multiprocessing.get_context("spawn")
-    q = ctx.Queue()
-    p = ctx.Process(
-        target=_throughput_worker,
-        args=(pin, engine_args_kwargs, q, v2_mode),
-    )
-    p.start()
-    p.join()
-    if p.exitcode != 0:
-        raise RuntimeError(
-            f"Throughput benchmark subprocess (pin={pin}) exited with code {p.exitcode}"
-        )
-    return q.get()
-
-
-def _latency_worker(
-    pin: bool,
-    engine_args_kwargs: dict,
-    q: "multiprocessing.Queue[dict]",
-    v2_mode: bool = False,
-) -> None:
-    """Run latency benchmark in a fresh spawn subprocess.
-
-    Follows latency.py methodology: fixed batch of dummy token IDs, warmup
-    iterations to reach steady state, then timed iterations reduced to avg
-    and percentiles.  Results are written to a temp JSON file by latency_main
-    and forwarded through the queue.
-    """
-    import vllm.utils.platform_utils as pu
-    from vllm.platforms import current_platform
-
-    pu.is_pin_memory_available.cache_clear()
-    pu.is_uva_available.cache_clear()
-    type(current_platform).is_pin_memory_available = classmethod(lambda cls: pin)
-    if v2_mode:
-        pu.is_uva_available = lambda: True
-
-    from vllm.benchmarks.latency import add_cli_args
-    from vllm.benchmarks.latency import main as latency_main
-
-    parser = argparse.ArgumentParser()
-    add_cli_args(parser)
-    args = parser.parse_args([])
-
-    for key, val in engine_args_kwargs.items():
-        setattr(args, key, val)
-    args.input_len = _LATENCY_INPUT_LEN
-    args.output_len = _LATENCY_OUTPUT_LEN
-    args.batch_size = _LATENCY_BATCH_SIZE
-    args.num_iters_warmup = _LATENCY_WARMUP_ITERS
-    args.num_iters = _LATENCY_BENCH_ITERS
-    args.profile = False
-    args.disable_detokenize = True
-
-    with tempfile.NamedTemporaryFile(mode="w", suffix=".json", delete=False) as f:
-        tmp_path = f.name
-    args.output_json = tmp_path
-
-    latency_main(args)
-
-    with open(tmp_path) as f:
-        results = json.load(f)
-    q.put(results)
-
-
-def _run_latency_benchmark(
-    pin: bool,
-    engine_args_kwargs: dict,
-    v2_mode: bool = False,
-) -> dict:
-    ctx = multiprocessing.get_context("spawn")
-    q = ctx.Queue()
-    p = ctx.Process(
-        target=_latency_worker,
-        args=(pin, engine_args_kwargs, q, v2_mode),
-    )
-    p.start()
-    p.join()
-    if p.exitcode != 0:
-        raise RuntimeError(
-            f"Latency benchmark subprocess (pin={pin}) exited with code {p.exitcode}"
-        )
-    return q.get()
-
-
-@pytest.mark.parametrize(
-    "test_v2_runner",
-    [
-        pytest.param(False, id="v1"),
-        pytest.param(True, id="v2"),
-    ],
-)
-class TestPinnedMemory:
-    """Verify pinned memory yields >= throughput vs unpinned via real vLLM inference."""
-
-    def test_throughput(self, monkeypatch, test_v2_runner, model, max_model_len):
-        """Benchmark throughput with pin_memory forced on then off.
-
-        Delegates to vllm/benchmarks/throughput.py main() with the random
-        dataset.  Each condition runs in an isolated spawn subprocess so both
-        start from a cold CUDA context, giving an unbiased comparison.
-        """
-        monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
-        monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1" if test_v2_runner else "0")
-
-        engine_args_kwargs = dict(
-            model=model,
-            gpu_memory_utilization=0.88,
-            max_model_len=max_model_len,
-            enable_prefix_caching=False,
-        )
-
-        _skip_if_pin_memory_not_available(engine_args_kwargs)
-
-        unpinned_tps = _run_throughput_benchmark(
-            False, engine_args_kwargs, v2_mode=test_v2_runner
-        )
-        pinned_tps = _run_throughput_benchmark(
-            True, engine_args_kwargs, v2_mode=test_v2_runner
-        )
-
-        pct_diff = (pinned_tps - unpinned_tps) / unpinned_tps * 100
-        runner = "v2" if test_v2_runner else "v1"
-        print(
-            f"\n=== Throughput results ({runner} runner, {model}) ==="
-            f"\npin_memory=True:  {pinned_tps:.1f} tok/s"
-            f"\npin_memory=False: {unpinned_tps:.1f} tok/s"
-            f"\nDifference: {pct_diff:+.1f}% (pinned vs unpinned)"
-        )
-
-        assert pinned_tps >= unpinned_tps * _THROUGHPUT_TOLERANCE, (
-            f"Pinned throughput ({pinned_tps:.1f} tok/s) fell more than "
-            f"{(1.0 - _THROUGHPUT_TOLERANCE) * 100:.1f}% below "
-            f"unpinned ({unpinned_tps:.1f} tok/s)."
-        )
-
-    def test_latency(self, monkeypatch, test_v2_runner, model, max_model_len):
-        """Benchmark per-batch latency with pin_memory forced on then off.
-
-        Follows vllm/benchmarks/latency.py: fixed dummy-token batch, warmup
-        iterations to reach steady state, then timed iterations reduced to avg
-        and percentiles.  Subprocesses run serially so each gets a cold CUDA
-        context without GPU memory pressure from the other run.
-        """
-        monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
-        monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1" if test_v2_runner else "0")
-
-        engine_args_kwargs = dict(
-            model=model,
-            gpu_memory_utilization=0.88,
-            max_model_len=max_model_len,
-            enable_prefix_caching=False,
-        )
-
-        _skip_if_pin_memory_not_available(engine_args_kwargs)
-
-        unpinned = _run_latency_benchmark(
-            False, engine_args_kwargs, v2_mode=test_v2_runner
-        )
-        pinned = _run_latency_benchmark(
-            True, engine_args_kwargs, v2_mode=test_v2_runner
-        )
-
-        pct_diff = (
-            (pinned["avg_latency"] - unpinned["avg_latency"])
-            / unpinned["avg_latency"]
-            * 100
-        )
-        runner = "v2" if test_v2_runner else "v1"
-        print(
-            f"\n=== Latency results ({runner} runner, {model}) ==="
-            f"\npin_memory=True:  avg={pinned['avg_latency']:.3f}s"
-            f"  p50={pinned['percentiles']['50']:.3f}s"
-            f"  p99={pinned['percentiles']['99']:.3f}s"
-            f"\npin_memory=False: avg={unpinned['avg_latency']:.3f}s"
-            f"  p50={unpinned['percentiles']['50']:.3f}s"
-            f"  p99={unpinned['percentiles']['99']:.3f}s"
-            f"\nDifference: {pct_diff:+.1f}% (pinned vs unpinned)"
-        )
-
-        assert pinned["avg_latency"] <= unpinned["avg_latency"] * _LATENCY_TOLERANCE, (
-            f"Pinned avg latency ({pinned['avg_latency']:.3f}s) exceeded "
-            f"unpinned ({unpinned['avg_latency']:.3f}s) by more than "
-            f"{(_LATENCY_TOLERANCE - 1.0) * 100:.1f}%."
-        )
-
-
-if __name__ == "__main__":
-    _parser = argparse.ArgumentParser(add_help=False)
-    _parser.add_argument("--model", default=_DEFAULT_MODEL)
-    _parser.add_argument("--max-model-len", type=int, default=_DEFAULT_MAX_MODEL_LEN)
-    _, _remaining = _parser.parse_known_args()
-    sys.exit(pytest.main([__file__] + _remaining))

benchmarks/kernels/benchmark_flydsl_moe_w4a16.py

@@ -1,277 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-# Copyright (c) 2025 FlyDSL Project Contributors
-
-import json
-import os
-
-import torch
-from aiter.test_common import run_perftest
-
-from vllm.model_executor.layers.fused_moe import fused_experts
-from vllm.model_executor.layers.fused_moe.activation import MoEActivation
-from vllm.model_executor.layers.fused_moe.config import (
-    int4_w4a16_moe_quant_config,
-)
-from vllm.model_executor.layers.fused_moe.fused_flydsl_moe import fused_flydsl_moe
-from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe import (  # noqa: E501
-    compressed_tensors_moe_w4a16_flydsl,
-)
-from vllm.platforms import current_platform
-
-RoutingBuffers = tuple[
-    torch.Tensor,  # sorted_token_ids
-    torch.Tensor,  # sorted_weights
-    torch.Tensor,  # sorted_expert_ids
-    torch.Tensor,  # num_valid_ids (shape [1], i32)
-    int,  # sorted_size
-    int,  # blocks
-]
-
-MODEL_PARAMS_TO_TUNE = [
-    # (num_experts, inter_dim, hidden_size, topk)
-    (384, 256, 7168, 8),  # Kimi K2.5 TP=8
-    (384, 512, 7168, 8),  # Kimi K2.5 TP=4
-]
-
-NUM_TOKENS_TO_TUNE = [
-    1,
-    2,
-    4,
-    8,
-    16,
-    24,
-    32,
-    48,
-    64,
-    128,
-    256,
-    512,
-    1024,
-    2048,
-    4096,
-    8192,
-]
-
-TILE_M_SEARCH_SPACE = [16, 32, 64, 128, 256]
-TILE_N_SEARCH_SPACE = [16, 32, 64, 128, 256]
-TILE_K_SEARCH_SPACE = [16, 32, 64, 128, 256, 512]
-TILE_N2_SEARCH_SPACE = [16, 32, 64, 128, 256]
-TILE_K2_SEARCH_SPACE = [16, 32, 64, 128, 256, 512]
-
-TILE_CONFIGS = []
-for tile_m in TILE_M_SEARCH_SPACE:
-    for tile_n in TILE_N_SEARCH_SPACE:
-        for tile_k in TILE_K_SEARCH_SPACE:
-            for tile_n2 in TILE_N2_SEARCH_SPACE:
-                for tile_k2 in TILE_K2_SEARCH_SPACE:
-                    TILE_CONFIGS.append(
-                        {
-                            "tile_m": tile_m,
-                            "tile_n": tile_n,
-                            "tile_k": tile_k,
-                            "tile_n2": tile_n2,
-                            "tile_k2": tile_k2,
-                        }
-                    )
-
-
-def tune_flydsl_moe_w4a16(
-    device: str = "cuda", num_iters: int = 100, num_warmup: int = 10
-):
-    packed_factor = 8
-    w13_num_shards = 2
-    params_dtype = torch.bfloat16
-    group_size = 32
-    scale_factor = 0.01
-
-    for model_params in MODEL_PARAMS_TO_TUNE:
-        num_experts = model_params[0]
-        inter_dim = model_params[1]
-        hidden_size = model_params[2]
-        topk = model_params[3]
-        print(
-            f"\nTuning: num_experts={num_experts}, inter_dim={inter_dim}, "
-            f"hidden_size={hidden_size}, topk={topk}...\n"
-        )
-
-        w2_scales_size = inter_dim
-        num_groups_w2 = w2_scales_size // group_size
-        num_groups_w13 = hidden_size // group_size
-
-        w13_weight = torch.randint(
-            0,
-            255,
-            (num_experts, hidden_size // packed_factor, w13_num_shards * inter_dim),
-            dtype=torch.int32,
-            device=device,
-        )
-
-        w2_weight = torch.randint(
-            0,
-            255,
-            (num_experts, inter_dim // packed_factor, hidden_size),
-            dtype=torch.int32,
-            device=device,
-        )
-        w13_scale = scale_factor * torch.randn(
-            num_experts,
-            num_groups_w13,
-            w13_num_shards * inter_dim,
-            dtype=params_dtype,
-            device=device,
-        )
-        w2_scale = scale_factor * torch.randn(
-            num_experts, num_groups_w2, hidden_size, dtype=params_dtype, device=device
-        )
-
-        w13 = w13_weight
-        w13 = compressed_tensors_moe_w4a16_flydsl._gptq_int32_to_flydsl_packed(w13)
-        w13 = w13.view(-1).contiguous()
-
-        w2 = w2_weight
-        w2 = compressed_tensors_moe_w4a16_flydsl._gptq_int32_to_flydsl_packed(w2)
-        w2 = w2.view(-1).contiguous()
-
-        w13_scale_flydsl = w13_scale
-        w2_scale_flydsl = w2_scale
-
-        if group_size > 0 and w13_scale.dim() == 3 and w13_scale.shape[1] > 1:
-            E, G, N = w13_scale.shape
-            w13_scale_flydsl = (
-                w13_scale_flydsl.view(E, G // 2, 2, N)
-                .permute(0, 1, 3, 2)
-                .contiguous()
-                .view(-1)
-                .contiguous()
-            )
-        elif w13_scale.dim() == 3 and w13_scale.shape[1] == 1:
-            w13_scale_flydsl = w13_scale_flydsl.squeeze(1)
-
-        if group_size > 0 and w2_scale.dim() == 3 and w2_scale.shape[1] > 1:
-            E, G, N = w2_scale.shape
-            w2_scale_flydsl = (
-                w2_scale_flydsl.view(E, G // 2, 2, N)
-                .permute(0, 1, 3, 2)
-                .contiguous()
-                .view(-1)
-                .contiguous()
-            )
-        elif w2_scale.dim() == 3 and w2_scale.shape[1] == 1:
-            w2_scale_flydsl = w2_scale_flydsl.squeeze(1)
-
-        w13_scale_flydsl = w13_scale_flydsl.contiguous()
-        w2_scale_flydsl = w2_scale_flydsl.contiguous()
-
-        w13.is_shuffled = True
-        w2.is_shuffled = True
-
-        w13_weight_scale = w13_scale.transpose(1, 2).contiguous()
-        w2_weight_scale = w2_scale.transpose(1, 2).contiguous()
-        w13_weight_packed = w13_weight.transpose(1, 2).contiguous().view(torch.uint8)
-        w2_weight_packed = w2_weight.transpose(1, 2).contiguous().view(torch.uint8)
-
-        moe_quant_config = int4_w4a16_moe_quant_config(
-            w1_scale=w13_weight_scale,
-            w2_scale=w2_weight_scale,
-            w1_zp=None,
-            w2_zp=None,
-            block_shape=[0, group_size],
-        )
-
-        tuned_config = {}
-
-        for num_tokens in NUM_TOKENS_TO_TUNE:
-            score = torch.rand(
-                (num_tokens, num_experts), device=device, dtype=torch.float32
-            )
-            topk_vals, topk_ids = torch.topk(score, k=topk, dim=1)
-            topk_weights = torch.softmax(topk_vals, dim=1).to(torch.float32)
-            x = torch.randn(
-                (num_tokens, hidden_size), dtype=torch.bfloat16, device=device
-            )
-            us_best = float("inf")
-            for tile_config in TILE_CONFIGS:
-                try:
-                    tile_m = tile_config["tile_m"]
-                    tile_n = tile_config["tile_n"]
-                    tile_k = tile_config["tile_k"]
-                    tile_n2 = tile_config["tile_n2"]
-                    tile_k2 = tile_config["tile_k2"]
-
-                    model_dim = x.shape[1]
-                    assert model_dim % 64 == 0
-                    assert model_dim % tile_k == 0
-                    assert inter_dim % tile_n == 0
-                    assert model_dim % tile_n2 == 0
-                    assert inter_dim % tile_k2 == 0
-                    assert ((tile_m * tile_k2) % 256) == 0
-                    bytes_per_thread_x = (tile_m * tile_k2) // 256
-                    assert (bytes_per_thread_x % 4) == 0
-
-                    out, _us = run_perftest(
-                        fused_flydsl_moe,
-                        x,
-                        w13,
-                        w2,
-                        num_experts,
-                        inter_dim,
-                        topk_weights,
-                        topk_ids,
-                        num_iters=num_iters,
-                        num_warmup=num_warmup,
-                        w1_scale=w13_scale_flydsl,
-                        w2_scale=w2_scale_flydsl,
-                        topk=topk_weights.shape[-1],
-                        group_size=group_size,
-                        doweight_stage1=False,
-                        scale_is_bf16=True,
-                        config=tile_config,
-                    )
-                    torch.accelerator.synchronize()
-                except Exception:
-                    torch.accelerator.synchronize()
-                    continue
-                else:
-                    us = _us.item()
-                    if us < us_best:
-                        out_ref = fused_experts(
-                            x,
-                            w13_weight_packed,
-                            w2_weight_packed,
-                            topk_weights=topk_weights,
-                            topk_ids=topk_ids,
-                            activation=MoEActivation.SILU,
-                            apply_router_weight_on_input=False,
-                            global_num_experts=num_experts,
-                            expert_map=None,
-                            quant_config=moe_quant_config,
-                        )
-                        try:
-                            assert torch.allclose(out, out_ref, atol=0.5, rtol=0.1)
-                        except Exception:
-                            continue
-                        else:
-                            print(
-                                f"For [num_tokens={num_tokens}, num_experts={num_experts}, "  # noqa: E501
-                                f"inter_dim={inter_dim}] found new best "  # noqa: E501
-                                f"config={tile_config}, us={us:0.3f}"
-                            )
-                            us_best = us
-                            tuned_config[str(num_tokens)] = tile_config
-            device_name = current_platform.get_device_name().replace(" ", "_")
-            tuned_config_file_name = (
-                f"E={num_experts},N={inter_dim},device_name={device_name},"
-                f"dtype=int4_w4a16,backend=flydsl.json"
-            )
-            tuner_dir_path = os.path.dirname(os.path.realpath(__file__))
-            store_path = os.path.join(tuner_dir_path, tuned_config_file_name)
-            with open(store_path, "w") as f:
-                json.dump(tuned_config, f, indent=4)
-            print(
-                f"\nTuned config for num_tokens={num_tokens} was stored at {store_path}\n"  # noqa: E501
-            )
-
-
-if __name__ == "__main__":
-    tune_flydsl_moe_w4a16(device="cuda")

benchmarks/kernels/benchmark_moe.py

@@ -792,12 +792,6 @@ def get_model_params(config):
         topk = text_config.num_experts_per_tok
         intermediate_size = text_config.moe_intermediate_size
         hidden_size = text_config.hidden_size
-    elif architecture == "DiffusionGemmaForBlockDiffusion":
-        text_config = config.get_text_config()
-        E = text_config.num_experts
-        topk = text_config.top_k_experts
-        intermediate_size = text_config.moe_intermediate_size
-        hidden_size = text_config.hidden_size
     elif architecture == "HunYuanMoEV1ForCausalLM":
         E = config.num_experts
         topk = config.moe_topk[0]

benchmarks/kv_cache_watermark.sh

@@ -1,248 +0,0 @@
-#!/bin/bash
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-#
-# Reproducible demonstration of the KV cache watermark (`--watermark`) for
-# reducing preemption thrashing.
-#
-# The watermark is the fraction of total KV cache blocks the scheduler keeps
-# free when admitting a waiting/preempted request into the running queue.
-#
-# Why this workload triggers thrashing:
-#   Requests are admitted based on the KV cache they need *at admission time*.
-#   With `--scheduler-reserve-full-isl` (default) the input length is reserved up
-#   front, but the *output* length is unknown and unreserved. A decode-heavy
-#   workload (output >> input) at high concurrency therefore over-admits while
-#   requests are short, then runs out of KV cache as they all grow during decode
-#   -> the scheduler preempts (recompute) recently-admitted requests, re-prefills
-#   them later, and repeats. The watermark keeps a block of KV cache free so
-#   running requests can grow into it instead of triggering this churn.
-#
-# This script launches `vllm serve` under a deliberately KV-constrained config
-# and a decode-heavy workload, sweeping the watermark across several values, and
-# reports the preemption count (scraped from /metrics), throughput, and latency
-# percentiles for each. It then plots the results.
-#
-# Default workload: concurrency 200, input ~300 tokens, output ~4000 tokens
-# (+/- 20% variance), sized to run each config for ~5 minutes.
-#
-# Usage:
-#   benchmarks/kv_cache_watermark.sh
-#   MODEL=Qwen/Qwen2.5-14B-Instruct TP=2 benchmarks/kv_cache_watermark.sh
-#
-# Run inside the vLLM virtualenv (so `vllm` and `python` resolve to it).
-set -euo pipefail
-
-# ---- Config (override via environment) -------------------------------------
-MODEL=${MODEL:-Qwen/Qwen2.5-7B-Instruct}
-TP=${TP:-1}
-PORT=${PORT:-8000}
-URL="http://127.0.0.1:${PORT}"
-# Constrain the KV cache to a *near-critical* size: large enough that the engine
-# can run stably, but small enough that greedy over-admission tips it into
-# preemption thrashing. (Independent of GPU size, so the demo is reproducible.)
-# At the default workload this fits ~1.5x the mean concurrent KV demand.
-KV_CACHE_MEMORY_GB=${KV_CACHE_MEMORY_GB:-16}
-MAX_MODEL_LEN=${MAX_MODEL_LEN:-8192}
-MAX_NUM_SEQS=${MAX_NUM_SEQS:-256}
-# Optional weight loader (e.g. fastsafetensors on the GCP cluster).
-LOAD_FORMAT=${LOAD_FORMAT:-auto}
-# Decode-heavy workload: moderate input, long output, with length variance. The
-# long output means preempted requests have generated a lot before eviction, so
-# resuming them re-prefills a long sequence (high recomputation cost).
-INPUT_LEN=${INPUT_LEN:-1000}
-OUTPUT_LEN=${OUTPUT_LEN:-5000}
-RANGE_RATIO=${RANGE_RATIO:-0.2}
-CONCURRENCY=${CONCURRENCY:-128}
-# Enough prompts to keep each config saturated for ~5+ minutes.
-NUM_PROMPTS=${NUM_PROMPTS:-450}
-OUTDIR=${OUTDIR:-./watermark_bench_results}
-# Watermark fractions compared. "label value" per line; value=0 disables it.
-CONFIGS=${CONFIGS:-"off    0
-w0.02  0.02
-w0.05  0.05
-w0.10  0.10
-w0.15  0.15"}
-
-KV_CACHE_MEMORY_BYTES=$((KV_CACHE_MEMORY_GB * 1024 * 1024 * 1024))
-mkdir -p "$OUTDIR"
-
-SERVER_PID=""
-cleanup() { [[ -n "$SERVER_PID" ]] && kill "$SERVER_PID" 2>/dev/null || true; }
-trap cleanup EXIT
-
-scrape_preemptions() {
-  # Sum the vllm:num_preemptions_total counter across engines.
-  python - "${URL}/metrics" <<'PY'
-import sys, urllib.request
-total = 0.0
-try:
-    body = urllib.request.urlopen(sys.argv[1], timeout=10).read().decode("utf-8", "replace")
-    for line in body.splitlines():
-        if line.startswith("vllm:num_preemptions_total"):
-            total += float(line.rsplit(" ", 1)[-1])
-except Exception as e:  # noqa: BLE001
-    print(f"scrape error: {e}", file=sys.stderr)
-print(int(total))
-PY
-}
-
-wait_for_server() {
-  for _ in $(seq 1 300); do
-    if curl -s "${URL}/health" >/dev/null 2>&1; then return 0; fi
-    if ! kill -0 "$SERVER_PID" 2>/dev/null; then
-      echo "ERROR: server process exited during startup" >&2; return 1
-    fi
-    sleep 5
-  done
-  echo "ERROR: server did not become ready" >&2; return 1
-}
-
-run_one() {
-  local label=$1 watermark=$2
-  echo
-  echo "==================== watermark: ${label} (${watermark}) ===================="
-  vllm serve "$MODEL" \
-    --tensor-parallel-size "$TP" \
-    --load-format "$LOAD_FORMAT" \
-    --kv-cache-memory-bytes "$KV_CACHE_MEMORY_BYTES" \
-    --max-model-len "$MAX_MODEL_LEN" \
-    --max-num-seqs "$MAX_NUM_SEQS" \
-    --no-enable-prefix-caching \
-    --watermark "$watermark" \
-    --port "$PORT" >"${OUTDIR}/serve_${label}.log" 2>&1 &
-  SERVER_PID=$!
-  wait_for_server
-  sleep 5
-
-  local pre post
-  pre=$(scrape_preemptions)
-  vllm bench serve \
-    --backend vllm \
-    --base-url "$URL" \
-    --model "$MODEL" \
-    --dataset-name random \
-    --random-input-len "$INPUT_LEN" \
-    --random-output-len "$OUTPUT_LEN" \
-    --random-range-ratio "$RANGE_RATIO" \
-    --ignore-eos \
-    --num-prompts "$NUM_PROMPTS" \
-    --max-concurrency "$CONCURRENCY" \
-    --percentile-metrics "ttft,tpot,itl,e2el" \
-    --metric-percentiles "50,90,99" \
-    --save-result \
-    --result-dir "$OUTDIR" \
-    --result-filename "bench_${label}.json"
-  post=$(scrape_preemptions)
-  echo "${label} ${watermark} $((post - pre))" >>"${OUTDIR}/preemptions.txt"
-
-  kill "$SERVER_PID" 2>/dev/null || true
-  for _ in $(seq 1 60); do curl -s "${URL}/health" >/dev/null 2>&1 || break; sleep 2; done
-  SERVER_PID=""
-  sleep 10
-}
-
-: >"${OUTDIR}/preemptions.txt"
-while read -r label watermark; do
-  [[ -z "${label:-}" ]] && continue
-  run_one "$label" "$watermark"
-done <<<"$CONFIGS"
-
-echo
-echo "==================== summary ===================="
-python - "$OUTDIR" <<'PY'
-import json, os, sys
-outdir = sys.argv[1]
-pre = {}
-order = []
-for line in open(os.path.join(outdir, "preemptions.txt")):
-    label, watermark, n = line.split()
-    pre[label] = (float(watermark), int(n))
-    order.append(label)
-
-def g(d, *names):
-    for n in names:
-        if d.get(n) is not None:
-            return d[n]
-    return float("nan")
-
-cols = ["watermark", "frac", "preempt", "out_tok/s", "req/s",
-        "TTFT_p50", "TTFT_p99", "ITL_p99", "E2EL_p50"]
-print("  ".join(f"{c:>10}" for c in cols))
-rows = []
-for label in order:
-    watermark, n = pre[label]
-    d = json.load(open(os.path.join(outdir, f"bench_{label}.json")))
-    rows.append(dict(
-        label=label, watermark=watermark, preempt=n,
-        out_tok_s=g(d, "output_throughput"),
-        req_s=g(d, "request_throughput"),
-        ttft_p50=g(d, "p50_ttft_ms", "median_ttft_ms"),
-        ttft_p99=g(d, "p99_ttft_ms"),
-        itl_p99=g(d, "p99_itl_ms"),
-        e2el_p50=g(d, "p50_e2el_ms", "median_e2el_ms"),
-    ))
-    print("  ".join(f"{str(v):>10}" for v in [
-        label, watermark, n,
-        f"{rows[-1]['out_tok_s']:.0f}",
-        f"{rows[-1]['req_s']:.3f}",
-        f"{rows[-1]['ttft_p50']/1000:.2f}",
-        f"{rows[-1]['ttft_p99']/1000:.2f}",
-        f"{rows[-1]['itl_p99']:.2f}",
-        f"{rows[-1]['e2el_p50']/1000:.1f}",
-    ]))
-print("\n(TTFT/E2EL in seconds; ITL in ms. Lower preempt is better.)")
-
-# ---- Plot -------------------------------------------------------------------
-try:
-    import matplotlib
-    matplotlib.use("Agg")
-    import matplotlib.pyplot as plt
-except Exception as e:  # noqa: BLE001
-    print(f"\n(skip plot: matplotlib unavailable: {e})")
-    sys.exit(0)
-
-x = [r["watermark"] for r in rows]
-xt = [f"{r['watermark']:g}\n({r['label']})" for r in rows]
-idx = list(range(len(rows)))
-
-fig, axes = plt.subplots(2, 2, figsize=(12, 8))
-fig.suptitle(
-    f"KV cache watermark sweep — {os.path.basename(os.path.abspath(outdir))}",
-    fontsize=12,
-)
-
-ax = axes[0][0]
-ax.bar(idx, [r["preempt"] for r in rows], color="tab:red")
-ax.set_title("Preemptions (lower is better)")
-ax.set_ylabel("preemptions")
-ax.set_xticks(idx); ax.set_xticklabels(xt)
-
-ax = axes[0][1]
-ax.plot(idx, [r["out_tok_s"] for r in rows], "o-", color="tab:green")
-ax.set_title("Output throughput (higher is better)")
-ax.set_ylabel("tokens/s")
-ax.set_xticks(idx); ax.set_xticklabels(xt)
-
-ax = axes[1][0]
-ax.plot(idx, [r["itl_p99"] for r in rows], "o-", color="tab:blue")
-ax.set_title("Inter-token latency p99 (lower is better)")
-ax.set_ylabel("ITL p99 (ms)")
-ax.set_xlabel("watermark fraction")
-ax.set_xticks(idx); ax.set_xticklabels(xt)
-
-ax = axes[1][1]
-ax.plot(idx, [r["ttft_p50"] / 1000 for r in rows], "o-", label="TTFT p50")
-ax.plot(idx, [r["ttft_p99"] / 1000 for r in rows], "o-", label="TTFT p99")
-ax.plot(idx, [r["e2el_p50"] / 1000 for r in rows], "o-", label="E2EL p50")
-ax.set_title("Latency (lower is better)")
-ax.set_ylabel("seconds")
-ax.set_xlabel("watermark fraction")
-ax.set_xticks(idx); ax.set_xticklabels(xt)
-ax.legend()
-
-fig.tight_layout(rect=(0, 0, 1, 0.95))
-out_png = os.path.join(outdir, "watermark_results.png")
-fig.savefig(out_png, dpi=120)
-print(f"\nWrote plot: {out_png}")
-PY

cmake/cpu_extension.cmake

@@ -166,10 +166,6 @@ elseif (S390_FOUND)
         "-mtune=native")
 elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
     message(STATUS "RISC-V detected")
-    if(DEFINED VLLM_RVV_VLEN AND NOT VLLM_RVV_VLEN GREATER 0)
-        message(FATAL_ERROR
-            "VLLM_RVV_VLEN must be a positive integer; got '${VLLM_RVV_VLEN}'")
-    endif()
     # VLLM_RVV_VLEN selects the target VLEN. Auto-detected from /proc/cpuinfo
     # by default; override with -DVLLM_RVV_VLEN=128 or -DVLLM_RVV_VLEN=256.
     if(NOT DEFINED VLLM_RVV_VLEN)
@@ -193,7 +189,8 @@ elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
                 "RISC-V RVV is available but VLEN could not be auto-detected. "
                 "Please specify VLEN explicitly:\n"
                 "  -DVLLM_RVV_VLEN=128   (for VLEN=128 hardware)\n"
-                "  -DVLLM_RVV_VLEN=256   (for VLEN=256 hardware, e.g. Spacemit X100)")
+                "  -DVLLM_RVV_VLEN=256   (for VLEN=256 hardware, e.g. Spacemit X100)\n"
+                "  -DVLLM_RVV_VLEN=0     (force scalar, no RVV)")
         endif()
     endif()
     if(VLLM_RVV_VLEN AND VLLM_RVV_VLEN GREATER 0)
@@ -222,7 +219,7 @@ endif()
 
 
 # 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 OR RVV_FP16_FOUND OR RVV_BF16_FOUND)
+if (ENABLE_X86_ISA OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
     # Fetch and build Arm Compute Library (ACL) as oneDNN's backend for AArch64
     # TODO [fadara01]: remove this once ACL can be fetched and built automatically as a dependency of oneDNN
     set(ONEDNN_AARCH64_USE_ACL OFF CACHE BOOL "")
@@ -438,12 +435,6 @@ if(USE_ONEDNN)
         ${VLLM_EXT_SRC})
 endif()
 
-if (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
-    set(VLLM_EXT_SRC
-        "csrc/cpu/sgl-kernels/gemm_int4.cpp"
-        ${VLLM_EXT_SRC})
-endif()
-
 if (ENABLE_X86_ISA)
     set(VLLM_EXT_SRC_SGL
         "csrc/cpu/sgl-kernels/conv.cpp"

cmake/external_projects/fmha_sm100.cmake

@@ -1,48 +0,0 @@
-include(FetchContent)
-
-# If FMHA_SM100_SRC_DIR is set, fmha_sm100 is installed from that directory
-# instead of downloading. This is useful for local MSA development.
-if(DEFINED ENV{FMHA_SM100_SRC_DIR})
-  set(FMHA_SM100_SRC_DIR $ENV{FMHA_SM100_SRC_DIR})
-endif()
-
-if(FMHA_SM100_SRC_DIR)
-  FetchContent_Declare(
-    fmha_sm100
-    SOURCE_DIR ${FMHA_SM100_SRC_DIR}
-    CONFIGURE_COMMAND ""
-    BUILD_COMMAND ""
-  )
-else()
-  FetchContent_Declare(
-    fmha_sm100
-    GIT_REPOSITORY https://github.com/vllm-project/MSA.git
-    GIT_TAG 544eee5e09ae2dfa774d5b06739013f9b7402c57
-    GIT_PROGRESS TRUE
-    CONFIGURE_COMMAND ""
-    BUILD_COMMAND ""
-  )
-endif()
-
-FetchContent_GetProperties(fmha_sm100)
-if(NOT fmha_sm100_POPULATED)
-  FetchContent_Populate(fmha_sm100)
-endif()
-message(STATUS "fmha_sm100 is available at ${fmha_sm100_SOURCE_DIR}")
-
-add_custom_target(fmha_sm100)
-
-set(FMHA_SM100_PY_ROOT "${fmha_sm100_SOURCE_DIR}/python/fmha_sm100")
-
-install(FILES
-  "${FMHA_SM100_PY_ROOT}/__init__.py"
-  "${FMHA_SM100_PY_ROOT}/sparse.py"
-  DESTINATION vllm/third_party/fmha_sm100
-  COMPONENT fmha_sm100)
-
-install(DIRECTORY "${FMHA_SM100_PY_ROOT}/cute/"
-  DESTINATION vllm/third_party/fmha_sm100/cute
-  COMPONENT fmha_sm100
-  PATTERN "__pycache__" EXCLUDE
-  PATTERN "*.pyc" EXCLUDE
-  PATTERN ".git*" EXCLUDE)

cmake/external_projects/qutlass.cmake

@@ -32,33 +32,21 @@ endif()
 message(STATUS "[QUTLASS] QuTLASS is available at ${qutlass_SOURCE_DIR}")
 
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-  cuda_archs_loose_intersection(QUTLASS_SM120_ARCHS "12.0f" "${CUDA_ARCHS}")
-  cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(QUTLASS_ARCHS "10.0f;12.0f" "${CUDA_ARCHS}")
 else()
-  cuda_archs_loose_intersection(QUTLASS_SM120_ARCHS "12.0a;12.1a" "${CUDA_ARCHS}")
-  cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0a;10.3a" "${CUDA_ARCHS}")
-endif()
-
-# QUTLASS uses TARGET_CUDA_ARCH as a single preprocessor selector for all its
-# sources. Do not compile a mixed SM100/SM120 arch list with one selector; prefer
-# SM100 when both families are requested because that is the primary deployed
-# target for this extension today.
-if(QUTLASS_SM100_ARCHS)
-  set(QUTLASS_ARCHS "${QUTLASS_SM100_ARCHS}")
-  set(QUTLASS_TARGET_CC 100)
-  if(QUTLASS_SM120_ARCHS)
-    message(WARNING
-      "[QUTLASS] Both SM100 and SM120 archs were requested; selecting SM100 "
-      "because TARGET_CUDA_ARCH is a single compile-time selector.")
-  endif()
-elseif(QUTLASS_SM120_ARCHS)
-  set(QUTLASS_ARCHS "${QUTLASS_SM120_ARCHS}")
-  set(QUTLASS_TARGET_CC 120)
-else()
-  set(QUTLASS_ARCHS)
+  cuda_archs_loose_intersection(QUTLASS_ARCHS "12.0a;12.1a;10.0a;10.3a" "${CUDA_ARCHS}")
 endif()
 
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
+
+  if(QUTLASS_ARCHS MATCHES "10\\.(0a|3a|0f)")
+    set(QUTLASS_TARGET_CC 100)
+  elseif(QUTLASS_ARCHS MATCHES "12\\.[01][af]?")
+    set(QUTLASS_TARGET_CC 120)
+  else()
+    message(FATAL_ERROR "[QUTLASS] internal error parsing CUDA_ARCHS='${QUTLASS_ARCHS}'.")
+  endif()
+
   set(QUTLASS_SOURCES
     ${qutlass_SOURCE_DIR}/qutlass/csrc/bindings.cpp
     ${qutlass_SOURCE_DIR}/qutlass/csrc/gemm.cu

cmake/external_projects/vllm_flash_attn.cmake

@@ -39,7 +39,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 803020a8fa15407871341d41eba4919ade2ee1ee
+          GIT_TAG dd62dac706b1cf7895bd99b18c6cb7e7e117ee25
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/utils.cmake

@@ -487,9 +487,9 @@ endfunction()
 
 function(cuda_archs_sm90plus OUT_CUDA_ARCHS TGT_CUDA_ARCHS)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(_archs "9.0a;10.0f;11.0f;12.0f" "${TGT_CUDA_ARCHS}")
+    cuda_archs_loose_intersection(_archs "9.0a;10.0f;11.0f" "${TGT_CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(_archs "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${TGT_CUDA_ARCHS}")
+    cuda_archs_loose_intersection(_archs "9.0a;10.0a;10.1a;10.3a" "${TGT_CUDA_ARCHS}")
   endif()
   set(${OUT_CUDA_ARCHS} ${_archs} PARENT_SCOPE)
 endfunction()

csrc/cpu/cpu_attn.cpp

@@ -11,25 +11,13 @@ static inline cpu_attention::Fp8KVCacheDataType parse_fp8_kv_dtype(
   return cpu_attention::Fp8KVCacheDataType::kAuto;
 }
 
-bool cpu_attn_has_isa(const std::string& isa) {
-  if (isa == "rvv") {
-#if defined(__riscv) && defined(__riscv_v_min_vlen) && __riscv_v_min_vlen == 128
-    return true;
-#else
-    return false;
-#endif
-  }
-  return false;
-}
-
 torch::Tensor get_scheduler_metadata(
     const int64_t num_req, const int64_t num_heads_q,
     const int64_t num_heads_kv, const int64_t head_dim,
     const torch::Tensor& seq_lens, at::ScalarType dtype,
-    const torch::Tensor& query_start_loc, const bool causal,
+    const torch::Tensor& query_start_loc, const bool casual,
     const int64_t window_size, const std::string& isa_hint,
-    const bool enable_kv_split,
-    const std::optional<torch::Tensor>& dynamic_causal) {
+    const bool enable_kv_split) {
   cpu_attention::ISA isa;
   if (isa_hint == "amx") {
     isa = cpu_attention::ISA::AMX;
@@ -56,13 +44,24 @@ torch::Tensor get_scheduler_metadata(
   input.head_dim = head_dim;
   input.query_start_loc = query_start_loc.data_ptr<int32_t>();
   input.seq_lens = seq_lens.data_ptr<int32_t>();
-
-  input.sliding_window_size = window_size;
-  input.causal = causal;
+  if (window_size != -1) {
+    input.left_sliding_window_size = window_size - 1;
+    if (casual) {
+      input.right_sliding_window_size = 0;
+    } else {
+      input.right_sliding_window_size = window_size - 1;
+    }
+  } else {
+    input.left_sliding_window_size = -1;
+    if (casual) {
+      input.right_sliding_window_size = 0;
+    } else {
+      input.right_sliding_window_size = -1;
+    }
+  }
+  input.casual = casual;
   input.isa = isa;
   input.enable_kv_split = enable_kv_split;
-  input.dynamic_causal =
-      dynamic_causal.has_value() ? dynamic_causal->data_ptr<bool>() : nullptr;
 
   VLLM_DISPATCH_FLOATING_TYPES(dtype, "get_scheduler_metadata", [&]() {
     CPU_ATTN_DISPATCH(head_dim, isa, 0, [&]() {
@@ -176,11 +175,10 @@ void cpu_attention_with_kv_cache(
     const torch::Tensor& seq_lens,         // [num_tokens]
     const double scale, const bool causal,
     const std::optional<torch::Tensor>& alibi_slopes,  // [num_heads]
-    const int64_t sliding_window,
+    const int64_t sliding_window_left, const int64_t sliding_window_right,
     const torch::Tensor& block_table,  // [num_tokens, max_block_num]
     const double softcap, const torch::Tensor& scheduler_metadata,
-    const std::optional<torch::Tensor>& s_aux,           // [num_heads]
-    const std::optional<torch::Tensor>& dynamic_causal,  // [num_reqs]
+    const std::optional<torch::Tensor>& s_aux,  // [num_heads]
     const double k_scale = 1.0, const double v_scale = 1.0,
     const std::string& kv_cache_dtype = "auto") {
   TORCH_CHECK_EQ(query.dim(), 3);
@@ -222,11 +220,13 @@ void cpu_attention_with_kv_cache(
   input.alibi_slopes =
       alibi_slopes.has_value() ? alibi_slopes->data_ptr<float>() : nullptr;
   input.s_aux = s_aux.has_value() ? s_aux->data_ptr<c10::BFloat16>() : nullptr;
-  input.dynamic_causal =
-      dynamic_causal.has_value() ? dynamic_causal->data_ptr<bool>() : nullptr;
   input.scale = scale;
   input.causal = causal;
-  input.sliding_window_size = sliding_window;
+  input.sliding_window_left = sliding_window_left;
+  input.sliding_window_right = sliding_window_right;
+  if (input.causal) {
+    input.sliding_window_right = 0;
+  }
   input.softcap = static_cast<float>(softcap);
 
   if (is_fp8) {

csrc/cpu/cpu_attn_impl.hpp

@@ -388,13 +388,13 @@ class AttentionScheduler {
     int32_t head_dim;
     int32_t* query_start_loc;
     int32_t* seq_lens;
-    int32_t sliding_window_size;
-    bool causal;
+    int32_t left_sliding_window_size;
+    int32_t right_sliding_window_size;
+    bool casual;
     cpu_attention::ISA isa;
     int32_t max_num_q_per_iter;  // max Q head num can be hold in registers
     int32_t kv_block_alignment;  // context length alignment requirement
     bool enable_kv_split;
-    bool* dynamic_causal;
   };
 
   static constexpr int32_t MaxQTileIterNum = 128;
@@ -403,8 +403,7 @@ class AttentionScheduler {
       : available_cache_size_(cpu_utils::get_available_l2_size()) {}
 
   torch::Tensor schedule(const ScheduleInput& input) const {
-    const bool causal = input.causal;
-    const bool is_dynamic_causal = input.dynamic_causal != nullptr;
+    const bool casual = input.casual;
     const int32_t thread_num = omp_get_max_threads();
     const int64_t cache_size = cpu_utils::get_available_l2_size();
     const int32_t max_num_q_per_iter = input.max_num_q_per_iter;
@@ -435,7 +434,8 @@ class AttentionScheduler {
     const int32_t default_tile_token_num = default_tile_size / q_head_per_kv;
     const int32_t split_kv_q_token_num_threshold =
         input.enable_kv_split ? 1 : 0;
-    const int32_t sliding_window_size = input.sliding_window_size;
+    const int32_t left_sliding_window_size = input.left_sliding_window_size;
+    const int32_t right_sliding_window_size = input.right_sliding_window_size;
     TORCH_CHECK_LE(split_kv_q_token_num_threshold * q_head_per_kv, 16);
 
     // get total kv len
@@ -444,9 +444,7 @@ class AttentionScheduler {
       const int32_t seq_len = input.seq_lens[req_id];
       const int32_t q_token_num =
           input.query_start_loc[req_id + 1] - input.query_start_loc[req_id];
-      const bool req_causal =
-          is_dynamic_causal ? input.dynamic_causal[req_id] : causal;
-      const int32_t q_start_pos = seq_len - q_token_num;
+      const int32_t q_start_pos = (casual ? (seq_len - q_token_num) : 0);
       const int32_t kv_start_pos = 0;
       const int32_t kv_end_pos = seq_len;
 
@@ -458,7 +456,7 @@ class AttentionScheduler {
         const int32_t q_tile_pos_right = q_tile_pos_left + q_tile_token_num;
         const auto [kv_tile_pos_left, kv_tile_pos_right] = calcu_kv_tile_pos(
             kv_start_pos, kv_end_pos, q_tile_pos_left, q_tile_pos_right,
-            sliding_window_size, req_causal);
+            left_sliding_window_size, right_sliding_window_size);
         const auto [aligned_kv_tile_pos_left, aligned_kv_tile_pos_right] =
             align_kv_tile_pos(kv_tile_pos_left, kv_tile_pos_right,
                               kv_len_alignment);
@@ -486,9 +484,7 @@ class AttentionScheduler {
       const int32_t seq_len = input.seq_lens[req_id];
       const int32_t q_token_num =
           input.query_start_loc[req_id + 1] - input.query_start_loc[req_id];
-      const bool req_causal =
-          is_dynamic_causal ? input.dynamic_causal[req_id] : causal;
-      const int32_t q_start_pos = seq_len - q_token_num;
+      const int32_t q_start_pos = (casual ? (seq_len - q_token_num) : 0);
       const int32_t kv_start_pos = 0;
       const int32_t kv_end_pos = seq_len;
       int32_t local_split_id = 0;
@@ -502,7 +498,7 @@ class AttentionScheduler {
         const int32_t q_tile_pos_right = q_tile_pos_left + q_tile_token_num;
         const auto [kv_tile_pos_left, kv_tile_pos_right] = calcu_kv_tile_pos(
             kv_start_pos, kv_end_pos, q_tile_pos_left, q_tile_pos_right,
-            sliding_window_size, req_causal);
+            left_sliding_window_size, right_sliding_window_size);
         const auto [aligned_kv_tile_pos_left, aligned_kv_tile_pos_right] =
             align_kv_tile_pos(kv_tile_pos_left, kv_tile_pos_right,
                               kv_len_alignment);
@@ -712,41 +708,15 @@ class AttentionScheduler {
     return metadata_tensor;
   }
 
-  FORCE_INLINE static std::pair<int32_t, int32_t> calcu_sliding_window_size(
-      int32_t window_size, bool causal) {
-    int32_t left_sliding_window_size, right_sliding_window_size;
-    if (window_size != -1) {
-      left_sliding_window_size = window_size - 1;
-      if (causal) {
-        right_sliding_window_size = 0;
-      } else {
-        right_sliding_window_size = window_size - 1;
-      }
-    } else {
-      left_sliding_window_size = -1;
-      if (causal) {
-        right_sliding_window_size = 0;
-      } else {
-        right_sliding_window_size = -1;
-      }
-    }
-
-    return {left_sliding_window_size, right_sliding_window_size};
-  }
-
   FORCE_INLINE static std::pair<int32_t, int32_t> calcu_kv_tile_pos(
       int32_t kv_left_pos, int32_t kv_right_pos, int32_t q_left_pos,
-      int32_t q_right_pos, int32_t window_size, bool causal) {
-    auto [left_sliding_window_size, right_sliding_window_size] =
-        calcu_sliding_window_size(window_size, causal);
-
-    if (left_sliding_window_size != -1) {
-      kv_left_pos =
-          std::max(kv_left_pos, q_left_pos - left_sliding_window_size);
+      int32_t q_right_pos, int32_t sliding_window_left,
+      int32_t sliding_window_right) {
+    if (sliding_window_left != -1) {
+      kv_left_pos = std::max(kv_left_pos, q_left_pos - sliding_window_left);
     }
-    if (right_sliding_window_size != -1) {
-      kv_right_pos =
-          std::min(kv_right_pos, q_right_pos + right_sliding_window_size);
+    if (sliding_window_right != -1) {
+      kv_right_pos = std::min(kv_right_pos, q_right_pos + sliding_window_right);
     }
     return {kv_left_pos, kv_right_pos};
   }
@@ -835,10 +805,10 @@ struct AttentionInput {
   int32_t* block_table;
   float* alibi_slopes;
   c10::BFloat16* s_aux;
-  bool* dynamic_causal;
   float scale;
   bool causal;
-  int32_t sliding_window_size;
+  int32_t sliding_window_left;
+  int32_t sliding_window_right;
   float softcap;
   // FP8 KV cache scales (used by FP8 attention implementations)
   float k_scale_fp8 = 1.0f;
@@ -852,8 +822,8 @@ struct AttentionInput {
       logits_buffer_t *__restrict__ logits_buffer,                          \
       float *__restrict__ partial_q_buffer, float *__restrict__ max_buffer, \
       float *__restrict__ sum_buffer, int32_t *__restrict__ block_table,    \
-      const int32_t kv_end_pos, const int32_t kv_tile_start_pos,            \
-      const int32_t kv_tile_end_pos, const int32_t kv_tile_token_num,       \
+      const int32_t kv_tile_start_pos, const int32_t kv_tile_end_pos,       \
+      const int32_t kv_tile_token_num,                                      \
       const int64_t kv_cache_num_blocks_stride, const int32_t q_head_num,   \
       const int32_t q_token_num, const int32_t q_tile_start_pos,            \
       const int32_t q_heads_per_kv, const int32_t block_size,               \
@@ -864,7 +834,7 @@ struct AttentionInput {
 
 #define CPU_ATTENTION_PARAMS                                                  \
   q_heads_buffer, k_head_cache_ptr, v_head_cache_ptr, logits_buffer,          \
-      partial_q_buffer, max_buffer, sum_buffer, block_table, kv_end_pos,      \
+      partial_q_buffer, max_buffer, sum_buffer, block_table,                  \
       kv_tile_start_pos, kv_tile_end_pos, kv_tile_token_num,                  \
       kv_cache_num_blocks_stride, q_head_num, q_token_num, q_tile_start_pos,  \
       q_heads_per_kv, block_size, left_window_size, right_window_size, scale, \
@@ -947,7 +917,6 @@ class AttentionMainLoop {
     //  - max_buffer: [MaxQHeadNumPerIteration, 1], store max logits
     //  - sum_buffer: [MaxQHeadNumPerIteration, 1], store sum of exp
     //  - block_table
-    //  - kv_end_pos: un-aligned end position of KV cache
     //  - kv_tile_start_pos: start position of KV cache, aligned to
     //  BlockSizeAlignment
     //  - kv_tile_end_pos: end position of KV cache, aligned to
@@ -1074,7 +1043,7 @@ class AttentionMainLoop {
         }
 
         apply_mask(logits_buffer, kv_tile_token_num, q_tile_start_pos,
-                   kv_end_pos, kv_tile_start_pos, kv_tile_end_pos, q_token_num,
+                   kv_tile_start_pos, kv_tile_end_pos, q_token_num,
                    q_heads_per_kv, left_window_size, right_window_size);
 
         // if (debug_info){
@@ -1157,7 +1126,7 @@ class AttentionMainLoop {
 
     void apply_mask(logits_buffer_t* __restrict__ logits_buffer,
                     const int64_t logits_buffer_stride,
-                    const int32_t q_tile_start_pos, const int32_t kv_end_pos,
+                    const int32_t q_tile_start_pos,
                     const int32_t kv_tile_start_pos,
                     const int32_t kv_tile_end_pos, const int32_t q_token_num,
                     const int32_t q_heads_per_kv,
@@ -1185,7 +1154,7 @@ class AttentionMainLoop {
                            std::max(kv_tile_start_pos,
                                     curr_token_pos + sliding_window_right + 1));
           }
-          return std::min(pos, kv_end_pos);
+          return pos;
         }();
 
         int32_t left_invalid_token_num = left_kv_pos - kv_tile_start_pos;
@@ -1472,16 +1441,15 @@ class AttentionMainLoop {
       const int64_t q_head_num_stride = input->query_num_heads_stride;
       const int64_t kv_cache_head_num_stride = input->cache_num_kv_heads_stride;
       const int64_t kv_cache_block_num_stride = input->cache_num_blocks_stride;
-      const int32_t sliding_window_size = input->sliding_window_size;
+      const int32_t sliding_window_left = input->sliding_window_left;
+      const int32_t sliding_window_right = input->sliding_window_right;
       const int32_t block_size = input->block_size;
       const float scale = input->scale;
       const float softcap_scale = input->softcap;
       const float* alibi_slopes = input->alibi_slopes;
       const c10::BFloat16* s_aux = input->s_aux;
-      const bool* dynamic_causal = input->dynamic_causal;
-      const bool is_dynamic_causal = dynamic_causal != nullptr;
 
-      const bool causal = input->causal;
+      const bool casual = input->causal;
       int32_t* const block_table = input->block_table;
       const int64_t block_table_stride = input->blt_num_tokens_stride;
 
@@ -1564,11 +1532,6 @@ class AttentionMainLoop {
                 &curr_workitem_groups[workitem_group_idx];
 
             const int32_t current_group_idx = current_workitem_group->req_id;
-            const int32_t current_group_causal =
-                is_dynamic_causal ? dynamic_causal[current_group_idx] : causal;
-            auto [sliding_window_left, sliding_window_right] =
-                AttentionScheduler::calcu_sliding_window_size(
-                    sliding_window_size, current_group_causal);
             const int32_t kv_start_pos =
                 current_workitem_group->kv_split_pos_start;
             const int32_t kv_end_pos = current_workitem_group->kv_split_pos_end;
@@ -1596,7 +1559,8 @@ class AttentionMainLoop {
             const int32_t q_end = input->query_start_loc[current_group_idx + 1];
             const int32_t q_start = input->query_start_loc[current_group_idx];
             const int32_t seq_len = input->seq_lens[current_group_idx];
-            const int32_t q_start_pos = seq_len - (q_end - q_start);
+            const int32_t q_start_pos =
+                (casual ? seq_len - (q_end - q_start) : 0);
             const int32_t block_num = (seq_len + block_size - 1) / block_size;
             // Only apply sink for the first KV split
             bool use_sink = (s_aux != nullptr &&
@@ -1646,8 +1610,8 @@ class AttentionMainLoop {
               const auto [kv_tile_start_pos, kv_tile_end_pos] =
                   AttentionScheduler::calcu_kv_tile_pos(
                       kv_start_pos, kv_end_pos, q_tile_start_pos,
-                      q_tile_end_pos, sliding_window_size,
-                      current_group_causal);
+                      q_tile_end_pos, sliding_window_left,
+                      sliding_window_right);
               const auto [rounded_kv_tile_start_pos, rounded_kv_tile_end_pos] =
                   AttentionScheduler::align_kv_tile_pos(
                       kv_tile_start_pos, kv_tile_end_pos, blocksize_alignment);
@@ -1760,8 +1724,8 @@ class AttentionMainLoop {
                               actual_kv_tile_pos_right] =
                       AttentionScheduler::calcu_kv_tile_pos(
                           kv_tile_pos_left, kv_tile_pos_right, q_tile_pos_left,
-                          q_tile_pos_right, sliding_window_size,
-                          current_group_causal);
+                          q_tile_pos_right, sliding_window_left,
+                          sliding_window_right);
                   const int32_t q_iter_idx =
                       q_head_tile_token_offset / curr_max_q_token_num_per_iter;
 
@@ -1825,7 +1789,7 @@ class AttentionMainLoop {
                   attn_impl.template execute_attention<Attention>(
                       curr_q_heads_buffer, curr_k_cache, curr_v_cache,
                       logits_buffer, curr_partial_q_buffer, curr_max_buffer,
-                      curr_sum_buffer, curr_block_table, kv_end_pos,
+                      curr_sum_buffer, curr_block_table,
                       aligned_actual_kv_tile_pos_left,
                       aligned_actual_kv_tile_pos_right, actual_kv_token_num,
                       kv_cache_block_num_stride, q_tile_head_num,

csrc/cpu/cpu_fused_moe.cpp

@@ -1,5 +1,3 @@
-#include <sleef.h>
-
 #include "cpu/cpu_types.hpp"
 #include "cpu/utils.hpp"
 #include "cpu/micro_gemm/cpu_micro_gemm_vec.hpp"
@@ -165,6 +163,7 @@ void gelu_tanh_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
   vec_op::FP32Vec16 w1_vec(0.7978845608028654);
   vec_op::FP32Vec16 w2_vec(0.5);
   vec_op::FP32Vec16 w3_vec(0.044715);
+  alignas(64) float temp[16];
 
   for (int32_t m = 0; m < m_size; ++m) {
     for (int32_t n = 0; n < dim; n += 16) {
@@ -172,9 +171,12 @@ void gelu_tanh_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
       vec_op::FP32Vec16 up_vec(up + n);
       auto gate_pow3_vec = gate_vec * gate_vec * gate_vec;
       auto inner_vec = w1_vec * (gate_vec + w3_vec * gate_pow3_vec);
-      // Note: can't use fast_exp form because diffusiongemma will generate
-      // wrong results
-      vec_op::FP32Vec16 tanh_vec(Sleef_tanhf16_u10(inner_vec.reg));
+
+      inner_vec.save(temp);
+      for (int32_t i = 0; i < 16; ++i) {
+        temp[i] = std::tanh(temp[i]);
+      }
+      vec_op::FP32Vec16 tanh_vec(temp);
       auto gelu_tanh = gate_vec * w2_vec * (one_vec + tanh_vec);
       auto gated_output_fp32 = up_vec * gelu_tanh;
       scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);

csrc/cpu/cpu_types_riscv_defs.hpp

@@ -57,10 +57,6 @@ typedef RVVTYPE(vfloat32, LMUL_512, _t) fixed_fp32x16_t
 typedef RVVTYPE(vfloat32, LMUL_1024, _t) fixed_fp32x32_t
     __attribute__((riscv_rvv_vector_bits(1024)));
 
-// int8
-typedef RVVTYPE(vint8, LMUL_128, _t) fixed_i8x16_t
-    __attribute__((riscv_rvv_vector_bits(128)));
-
 // int32
 typedef RVVTYPE(vint32, LMUL_256, _t) fixed_i32x8_t
     __attribute__((riscv_rvv_vector_bits(256)));

csrc/cpu/cpu_types_riscv_impl.hpp

@@ -9,14 +9,10 @@
 
 #include <algorithm>
 #include <cmath>
-#include <cstdint>
 #include <cstring>
 #include <iostream>
 #include <limits>
 #include <torch/all.h>
-
-#include "float_convert.hpp"
-
 namespace vec_op {
 
 // FP8 KV cache is not supported on RISC-V. These tag types and the
@@ -249,7 +245,8 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
     float tmp[8];
     for (int i = 0; i < 8; ++i) {
-      tmp[i] = bf16_to_float(u16[i]);
+      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
+      std::memcpy(&tmp[i], &v, 4);
     }
     reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_256)(tmp, 8);
   }
@@ -259,7 +256,9 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     RVVI(__riscv_vse32_v_f32, LMUL_256)(tmp, reg_fp32, 8);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < 8; ++i) {
-      u16[i] = float_to_bf16(tmp[i]);
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
     }
   }
   void save(void* ptr, int elem_num) const {
@@ -267,7 +266,9 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     RVVI(__riscv_vse32_v_f32, LMUL_256)(tmp, reg_fp32, 8);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < elem_num; ++i) {
-      u16[i] = float_to_bf16(tmp[i]);
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
     }
   }
   void save_strided(void* ptr, ptrdiff_t stride) const {
@@ -276,8 +277,10 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     uint8_t* u8 = static_cast<uint8_t*>(ptr);
     ptrdiff_t byte_stride = stride * sizeof(uint16_t);
     for (int i = 0; i < 8; ++i) {
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) =
-          float_to_bf16(tmp[i]);
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      uint16_t val = static_cast<uint16_t>(v >> 16);
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
     }
   }
 };
@@ -289,7 +292,8 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
     float tmp[16];
     for (int i = 0; i < 16; ++i) {
-      tmp[i] = bf16_to_float(u16[i]);
+      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
+      std::memcpy(&tmp[i], &v, 4);
     }
     reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_512)(tmp, 16);
   }
@@ -302,7 +306,9 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     RVVI(__riscv_vse32_v_f32, LMUL_512)(tmp, reg_fp32, 16);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < 16; ++i) {
-      u16[i] = float_to_bf16(tmp[i]);
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
     }
   }
   void save(void* ptr, int elem_num) const {
@@ -310,7 +316,9 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     RVVI(__riscv_vse32_v_f32, LMUL_512)(tmp, reg_fp32, 16);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < elem_num; ++i) {
-      u16[i] = float_to_bf16(tmp[i]);
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
     }
   }
   void save_strided(void* ptr, ptrdiff_t stride) const {
@@ -319,8 +327,10 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     uint8_t* u8 = static_cast<uint8_t*>(ptr);
     ptrdiff_t byte_stride = stride * sizeof(uint16_t);
     for (int i = 0; i < 16; ++i) {
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) =
-          float_to_bf16(tmp[i]);
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      uint16_t val = static_cast<uint16_t>(v >> 16);
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
     }
   }
 };
@@ -333,7 +343,8 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
     float tmp[32];
     for (int i = 0; i < 32; ++i) {
-      tmp[i] = bf16_to_float(u16[i]);
+      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
+      std::memcpy(&tmp[i], &v, 4);
     }
     reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_1024)(tmp, 32);
   }
@@ -360,7 +371,9 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     RVVI(__riscv_vse32_v_f32, LMUL_1024)(tmp, reg_fp32, 32);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < 32; ++i) {
-      u16[i] = float_to_bf16(tmp[i]);
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
     }
   }
 
@@ -369,7 +382,9 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     RVVI(__riscv_vse32_v_f32, LMUL_1024)(tmp, reg_fp32, 32);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < elem_num; ++i) {
-      u16[i] = float_to_bf16(tmp[i]);
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
     }
   }
 
@@ -379,8 +394,10 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     uint8_t* u8 = static_cast<uint8_t*>(ptr);
     ptrdiff_t byte_stride = stride * sizeof(uint16_t);
     for (int i = 0; i < 32; ++i) {
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) =
-          float_to_bf16(tmp[i]);
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      uint16_t val = static_cast<uint16_t>(v >> 16);
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
     }
   }
 };
@@ -717,18 +734,10 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return FP32Vec16(
         RVVI(__riscv_vfmax_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
   }
-  FP32Vec16 max(const FP32Vec16& b, const int elem_num) const {
-    return FP32Vec16(
-        RVVI(__riscv_vfmax_vv_f32, LMUL_512)(reg, b.reg, elem_num));
-  }
   FP32Vec16 min(const FP32Vec16& b) const {
     return FP32Vec16(
         RVVI(__riscv_vfmin_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
   }
-  FP32Vec16 min(const FP32Vec16& b, const int elem_num) const {
-    return FP32Vec16(
-        RVVI(__riscv_vfmin_vv_f32, LMUL_512)(reg, b.reg, elem_num));
-  }
   FP32Vec16 abs() const {
     return FP32Vec16(RVVI(__riscv_vfabs_v_f32, LMUL_512)(reg, VEC_ELEM_NUM));
   }
@@ -858,27 +867,6 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   }
 };
 
-struct INT8Vec16 : public Vec<INT8Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  fixed_i8x16_t reg;
-
-  explicit INT8Vec16(const FP32Vec16& vec) {
-    auto i32_vec =
-        RVVI(__riscv_vfcvt_x_f_v_i32, LMUL_512)(vec.reg, VEC_ELEM_NUM);
-    auto i16_vec = RVVI(__riscv_vnclip_wx_i16, LMUL_256)(
-        i32_vec, 0, __RISCV_VXRM_RNU, VEC_ELEM_NUM);
-    reg = RVVI(__riscv_vnclip_wx_i8, LMUL_128)(i16_vec, 0, __RISCV_VXRM_RNU,
-                                               VEC_ELEM_NUM);
-  }
-
-  void save(int8_t* ptr) const {
-    RVVI(__riscv_vse8_v_i8, LMUL_128)(ptr, reg, VEC_ELEM_NUM);
-  }
-  void save(int8_t* ptr, int elem_num) const {
-    RVVI(__riscv_vse8_v_i8, LMUL_128)(ptr, reg, elem_num);
-  }
-};
-
 // ============================================================================
 // Type Traits & Global Helpers
 // ============================================================================
@@ -968,7 +956,9 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16& v)
 #else
 template <>
 inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
-  *reinterpret_cast<uint16_t*>(ptr) = float_to_bf16(v);
+  uint32_t val;
+  std::memcpy(&val, &v, 4);
+  *reinterpret_cast<uint16_t*>(ptr) = static_cast<uint16_t>(val >> 16);
 }
 inline BF16Vec8::BF16Vec8(const FP32Vec8& v) : reg_fp32(v.reg) {}
 inline BF16Vec16::BF16Vec16(const FP32Vec16& v) : reg_fp32(v.reg) {}

csrc/cpu/cpu_types_vxe.hpp

@@ -3,9 +3,7 @@
 #define CPU_TYPES_VXE_HPP
 
 #include <vecintrin.h>
-#include <bit>
 #include <cmath>
-#include <cstdint>
 #include <limits>
 #include <torch/all.h>
 namespace vec_op {
@@ -819,7 +817,8 @@ inline void storeFP32<::c10::Half>(float v, ::c10::Half* ptr) {
   // intrinsics for FP32 to FP16 conversion does not use IEEE rounding and can
   // produce incorrect results for some inputs. Process each of the 4 vectors
   // separately.
-  uint32_t in = std::bit_cast<uint32_t>(v);
+  uint32_t in;
+  std::memcpy(&in, &v, sizeof(in));
 
   uint32_t s = (in & 0x80000000) >> 16;  // Sign
   uint32_t e = (in & 0x7F800000) >> 23;  // Exponent

csrc/cpu/float_convert.hpp

@@ -1,15 +1,14 @@
-#pragma once
 
-#include <bit>
-#include <cstdint>
-
-inline float bf16_to_float(uint16_t bf16) {
+static float bf16_to_float(uint16_t bf16) {
   uint32_t bits = static_cast<uint32_t>(bf16) << 16;
-  return std::bit_cast<float>(bits);
+  float fp32;
+  std::memcpy(&fp32, &bits, sizeof(fp32));
+  return fp32;
 }
 
-inline uint16_t float_to_bf16(float fp32) {
-  uint32_t bits = std::bit_cast<uint32_t>(fp32);
+static uint16_t float_to_bf16(float fp32) {
+  uint32_t bits;
+  std::memcpy(&bits, &fp32, sizeof(fp32));
   return static_cast<uint16_t>(bits >> 16);
 }
 
@@ -19,13 +18,14 @@ inline uint16_t float_to_bf16(float fp32) {
  * Codes below copied from
  * https://github.com/PrincetonVision/marvin/tree/master/tools/tensorIO_matlab
  *************************************************/
-inline uint16_t float_to_fp16(float fp32) {
+static uint16_t float_to_fp16(float fp32) {
   uint16_t fp16;
 
+  unsigned x;
   unsigned u, remainder, shift, lsb, lsb_s1, lsb_m1;
   unsigned sign, exponent, mantissa;
 
-  uint32_t x = std::bit_cast<uint32_t>(fp32);
+  std::memcpy(&x, &fp32, sizeof(fp32));
   u = (x & 0x7fffffff);
 
   // Get rid of +NaN/-NaN case first.
@@ -77,11 +77,12 @@ inline uint16_t float_to_fp16(float fp32) {
   return fp16;
 }
 
-inline float fp16_to_float(uint16_t fp16) {
+static float fp16_to_float(uint16_t fp16) {
   unsigned sign = ((fp16 >> 15) & 1);
   unsigned exponent = ((fp16 >> 10) & 0x1f);
   unsigned mantissa = ((fp16 & 0x3ff) << 13);
-  uint32_t temp;
+  int temp;
+  float fp32;
   if (exponent == 0x1f) { /* NaN or Inf */
     mantissa = (mantissa ? (sign = 0, 0x7fffff) : 0);
     exponent = 0xff;
@@ -100,5 +101,6 @@ inline float fp16_to_float(uint16_t fp16) {
     exponent += 0x70;
   }
   temp = ((sign << 31) | (exponent << 23) | mantissa);
-  return std::bit_cast<float>(temp);
+  std::memcpy(&fp32, &temp, sizeof(temp));
+  return fp32;
 }

csrc/cpu/generate_cpu_attn_dispatch.py

@@ -11,7 +11,7 @@ import os
 HEAD_DIMS_32 = [32, 64, 96, 128, 160, 192, 224, 256, 512]
 
 # Head dimensions divisible by 16 but not 32 (VEC16 only)
-HEAD_DIMS_16 = [48, 80, 112]
+HEAD_DIMS_16 = [80, 112]
 
 # ISA types
 ISA_TYPES = {

csrc/cpu/layernorm.cpp

@@ -4,9 +4,8 @@ namespace {
 template <typename scalar_t>
 void rms_norm_impl(scalar_t* __restrict__ out,
                    const scalar_t* __restrict__ input,
-                   const scalar_t* __restrict__ weight, const bool has_weight,
-                   const float epsilon, const int num_tokens,
-                   const int hidden_size) {
+                   const scalar_t* __restrict__ weight, const float epsilon,
+                   const int num_tokens, const int hidden_size) {
   using scalar_vec_t = vec_op::vec_t<scalar_t>;
   constexpr int VEC_ELEM_NUM = scalar_vec_t::get_elem_num();
   TORCH_CHECK(hidden_size % VEC_ELEM_NUM == 0);
@@ -28,15 +27,12 @@ void rms_norm_impl(scalar_t* __restrict__ out,
 
     for (int j = 0; j < hidden_size; j += VEC_ELEM_NUM) {
       scalar_vec_t x(input_p + j);
+      scalar_vec_t w(weight + j);
+
       vec_op::FP32Vec8 fp32_x(x);
-      vec_op::FP32Vec8 fp32_out;
-      if (has_weight) {
-        scalar_vec_t w(weight + j);
-        vec_op::FP32Vec8 fp32_w(w);
-        fp32_out = fp32_x * fp32_s_variance * fp32_w;
-      } else {
-        fp32_out = fp32_x * fp32_s_variance;
-      }
+      vec_op::FP32Vec8 fp32_w(w);
+
+      vec_op::FP32Vec8 fp32_out = fp32_x * fp32_s_variance * fp32_w;
 
       scalar_vec_t out(fp32_out);
       out.save(output_p + j);
@@ -48,8 +44,8 @@ template <typename scalar_t>
 void fused_add_rms_norm_impl(scalar_t* __restrict__ input,
                              scalar_t* __restrict__ residual,
                              const scalar_t* __restrict__ weight,
-                             const bool has_weight, const float epsilon,
-                             const int num_tokens, const int hidden_size) {
+                             const float epsilon, const int num_tokens,
+                             const int hidden_size) {
   using scalar_vec_t = vec_op::vec_t<scalar_t>;
   constexpr int VEC_ELEM_NUM = scalar_vec_t::get_elem_num();
   TORCH_CHECK(hidden_size % VEC_ELEM_NUM == 0);
@@ -76,18 +72,13 @@ void fused_add_rms_norm_impl(scalar_t* __restrict__ input,
     vec_op::FP32Vec8 fp32_s_variance(s_variance);
 
     for (int j = 0; j < hidden_size; j += VEC_ELEM_NUM) {
-      vec_op::FP32Vec8 fp32_out;
-      if (has_weight) {
-        scalar_vec_t w(weight + j);
-        scalar_vec_t res(residual_p + j);
-        vec_op::FP32Vec8 fp32_w(w);
-        vec_op::FP32Vec8 fp32_res(res);
-        fp32_out = fp32_res * fp32_s_variance * fp32_w;
-      } else {
-        scalar_vec_t res(residual_p + j);
-        vec_op::FP32Vec8 fp32_res(res);
-        fp32_out = fp32_res * fp32_s_variance;
-      }
+      scalar_vec_t w(weight + j);
+      scalar_vec_t res(residual_p + j);
+
+      vec_op::FP32Vec8 fp32_w(w);
+      vec_op::FP32Vec8 fp32_res(res);
+
+      vec_op::FP32Vec8 fp32_out = fp32_res * fp32_s_variance * fp32_w;
 
       scalar_vec_t out(fp32_out);
       out.save(input_p + j);
@@ -96,41 +87,31 @@ void fused_add_rms_norm_impl(scalar_t* __restrict__ input,
 }
 }  // namespace
 
-void rms_norm(torch::Tensor& out, torch::Tensor& input,
-              std::optional<torch::Tensor> weight, double epsilon) {
+void rms_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
+              double epsilon) {
   int hidden_size = input.size(-1);
   int num_tokens = input.numel() / hidden_size;
-  const bool has_weight = weight.has_value();
-  if (has_weight) {
-    TORCH_CHECK(weight->is_contiguous());
-  }
 
   VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "rms_norm_impl", [&] {
     CPU_KERNEL_GUARD_IN(rms_norm_impl)
     rms_norm_impl(out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),
-                  has_weight ? weight->data_ptr<scalar_t>() : nullptr,
-                  has_weight, epsilon, num_tokens, hidden_size);
+                  weight.data_ptr<scalar_t>(), epsilon, num_tokens,
+                  hidden_size);
     CPU_KERNEL_GUARD_OUT(rms_norm_impl)
   });
 }
 
 void fused_add_rms_norm(torch::Tensor& input, torch::Tensor& residual,
-                        std::optional<torch::Tensor> weight, double epsilon) {
+                        torch::Tensor& weight, double epsilon) {
   int hidden_size = input.size(-1);
   int num_tokens = input.numel() / hidden_size;
-  const bool has_weight = weight.has_value();
-  if (has_weight) {
-    TORCH_CHECK(weight->scalar_type() == input.scalar_type());
-    TORCH_CHECK(weight->is_contiguous());
-  }
 
   VLLM_DISPATCH_FLOATING_TYPES(
       input.scalar_type(), "fused_add_rms_norm_impl", [&] {
         CPU_KERNEL_GUARD_IN(fused_add_rms_norm_impl)
         fused_add_rms_norm_impl(
             input.data_ptr<scalar_t>(), residual.data_ptr<scalar_t>(),
-            has_weight ? weight->data_ptr<scalar_t>() : nullptr, has_weight,
-            epsilon, num_tokens, hidden_size);
+            weight.data_ptr<scalar_t>(), epsilon, num_tokens, hidden_size);
         CPU_KERNEL_GUARD_OUT(fused_add_rms_norm_impl)
       });
 }

csrc/cpu/sgl-kernels/gemm_int4.cpp

@@ -268,23 +268,6 @@ void _dequant_gemm_accum_small_M(
   _dequant_gemm_accum_small_M<M, N, ldb, sym_quant_act>(C, A, scales_a, qzeros_a, B, scales_b, qzeros_b, K, lda, ldc);
 #endif
 
-template <int64_t N, int64_t ldb>
-inline int32_t load_uint4_vnni(const uint8_t* __restrict__ B, int64_t k, int64_t n) {
-  // B is packed as [_block_k / 4, N / 2, 4] for VNNI4. Each byte stores two
-  // columns from adjacent 8-column groups for one K lane.
-  constexpr int64_t n_group_size = 8;
-  constexpr int64_t vnni_size = 4;
-  static_assert(N % (2 * n_group_size) == 0);
-
-  int64_t n_group = n / n_group_size;
-  int64_t ni = n % n_group_size;
-  int64_t ki = k % vnni_size;
-  int64_t k_base = k - ki;
-  int64_t packed_n = (n_group / 2) * n_group_size + ni;
-  uint8_t packed = B[k_base * ldb + packed_n * vnni_size + ki];
-  return (n_group % 2 == 0) ? (packed & 0x0f) : ((packed >> 4) & 0x0f);
-}
-
 template <int64_t N, int64_t ldb, bool sym_quant_act>
 void _dequant_gemm_accum(
     float* C,
@@ -338,24 +321,7 @@ void _dequant_gemm_accum(
   } else
 #endif
   {
-    for (int64_t m = 0; m < M; ++m) {
-      for (int64_t n = 0; n < N; ++n) {
-        int32_t acc = 0;
-        for (int64_t k = 0; k < K; ++k) {
-          int32_t b = load_uint4_vnni<N, ldb>(B, k, n) - qzeros_b[n];
-          if constexpr (sym_quant_act) {
-            const int8_t* A_s8 = reinterpret_cast<const int8_t*>(A);
-            acc += static_cast<int32_t>(A_s8[m * lda + k]) * b;
-          } else {
-            acc += static_cast<int32_t>(A[m * lda + k]) * b;
-          }
-        }
-        if constexpr (!sym_quant_act) {
-          acc -= qzeros_a[m] * compensation[n];
-        }
-        C[m * ldc + n] += static_cast<float>(acc) * scales_a[m] * scales_b[n];
-      }
-    }
+    TORCH_CHECK(false, "tinygemm_kernel: scalar path not implemented!");
   }
 }
 
@@ -530,11 +496,9 @@ void _da8w4_linear_impl(
         store_out<out_dtype, BLOCK_N>(C_tmp, output + mci * block_m * N + nc * BLOCK_N, m_size, N /*lda*/);
       }
     }
-#if defined(CPU_CAPABILITY_AVX512)
     if (use_brgemm) {
       at::native::cpublas::brgemm_release();
     }
-#endif
   });
 }
 

csrc/cpu/sgl-kernels/vec.h

@@ -245,7 +245,7 @@ quantize_row_int8(uint8_t* __restrict__ Aq, float& As, const scalar_t* __restric
 
   for (int64_t k = 0; k < K; ++k) {
     const float val = static_cast<float>(A[k]) * inv_scale;
-    Aq[k] = static_cast<uint8_t>(static_cast<int32_t>(std::round(val)) + 128);
+    Aq[k] = (uint8_t)(std::round(val)) + 128;
   }
   As = scale;
 }

csrc/cpu/torch_bindings.cpp

@@ -146,16 +146,13 @@ at::Tensor causal_conv1d_update_cpu(
 void activation_lut_bf16(torch::Tensor& out, torch::Tensor& input,
                          const std::string& activation);
 
-bool cpu_attn_has_isa(const std::string& isa);
-
 torch::Tensor get_scheduler_metadata(
     const int64_t num_req, const int64_t num_heads_q,
     const int64_t num_heads_kv, const int64_t head_dim,
     const torch::Tensor& seq_lens, at::ScalarType dtype,
     const torch::Tensor& query_start_loc, const bool casual,
     const int64_t window_size, const std::string& isa_hint,
-    const bool enable_kv_split,
-    const std::optional<torch::Tensor>& dynamic_causal);
+    const bool enable_kv_split);
 
 void cpu_attn_reshape_and_cache(const torch::Tensor& key,
                                 const torch::Tensor& value,
@@ -172,10 +169,10 @@ void cpu_attention_with_kv_cache(
     const torch::Tensor& query_start_loc, const torch::Tensor& seq_lens,
     const double scale, const bool causal,
     const std::optional<torch::Tensor>& alibi_slopes,
-    const int64_t sliding_window_left, const torch::Tensor& block_table,
-    const double softcap, const torch::Tensor& scheduler_metadata,
-    const std::optional<torch::Tensor>& s_aux,
-    const std::optional<torch::Tensor>& dynamic_causal, const double k_scale,
+    const int64_t sliding_window_left, const int64_t sliding_window_right,
+    const torch::Tensor& block_table, const double softcap,
+    const torch::Tensor& scheduler_metadata,
+    const std::optional<torch::Tensor>& s_aux, const double k_scale,
     const double v_scale, const std::string& kv_cache_dtype);
 
 // Note: just for avoiding importing errors
@@ -312,13 +309,13 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // Layernorm
   // Apply Root Mean Square (RMS) Normalization to the input tensor.
   ops.def(
-      "rms_norm(Tensor! out, Tensor input, Tensor? weight, float epsilon) -> "
+      "rms_norm(Tensor! out, Tensor input, Tensor weight, float epsilon) -> "
       "()");
   ops.impl("rms_norm", torch::kCPU, &rms_norm);
 
   // In-place fused Add and RMS Normalization.
   ops.def(
-      "fused_add_rms_norm(Tensor! input, Tensor! residual, Tensor? weight, "
+      "fused_add_rms_norm(Tensor! input, Tensor! residual, Tensor weight, "
       "float epsilon) -> ()");
   ops.impl("fused_add_rms_norm", torch::kCPU, &fused_add_rms_norm);
 
@@ -332,9 +329,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("rotary_embedding", torch::kCPU, &rotary_embedding);
 
   // Quantization
-#if defined(__AVX512F__) || defined(__AVX2__) ||                               \
-    (defined(__aarch64__) && !defined(__APPLE__)) || defined(__powerpc64__) || \
-    defined(__riscv_v)
+#if defined(__AVX512F__) || defined(__AVX2__) || \
+    (defined(__aarch64__) && !defined(__APPLE__)) || defined(__powerpc64__)
   // Helper function to release oneDNN handlers
   ops.def("release_dnnl_matmul_handler(int handler) -> ()",
           &release_dnnl_matmul_handler);
@@ -432,6 +428,19 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("int8_scaled_mm_with_quant", torch::kCPU,
            &int8_scaled_mm_with_quant);
 
+  // Adapted from sglang: INT4 W4A8 kernels
+  ops.def(
+      "convert_weight_packed_scale_zp(Tensor weight, Tensor qzeros, Tensor "
+      "scales, int quant_method_4bit) -> (Tensor, "
+      "Tensor, Tensor)");
+  ops.impl("convert_weight_packed_scale_zp", torch::kCPU,
+           &convert_weight_packed_scale_zp);
+
+  ops.def(
+      "int4_scaled_mm_cpu(Tensor(a0!) x, Tensor(a1!) w, Tensor(a2!) w_zeros, "
+      "Tensor(a3!) w_scales, Tensor? bias) -> Tensor");
+  ops.impl("int4_scaled_mm_cpu", torch::kCPU, &int4_scaled_mm_cpu);
+
   // Adapted from sglang: FP8 W8A16 kernel
   ops.def(
       "fp8_scaled_mm_cpu(Tensor(a0!) mat1, Tensor(a1!) mat2, Tensor(a2!) "
@@ -458,23 +467,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("causal_conv1d_update_cpu", torch::kCPU, &causal_conv1d_update_cpu);
 #endif
 
-#if (defined(__AVX512BF16__) && defined(__AVX512F__) && \
-     defined(__AVX512VNNI__)) ||                        \
-    defined(__riscv)
-  // Adapted from sglang: INT4 W4A8 kernels
-  ops.def(
-      "convert_weight_packed_scale_zp(Tensor weight, Tensor qzeros, Tensor "
-      "scales, int quant_method_4bit) -> (Tensor, "
-      "Tensor, Tensor)");
-  ops.impl("convert_weight_packed_scale_zp", torch::kCPU,
-           &convert_weight_packed_scale_zp);
-
-  ops.def(
-      "int4_scaled_mm_cpu(Tensor(a0!) x, Tensor(a1!) w, Tensor(a2!) w_zeros, "
-      "Tensor(a3!) w_scales, Tensor? bias) -> Tensor");
-  ops.impl("int4_scaled_mm_cpu", torch::kCPU, &int4_scaled_mm_cpu);
-#endif
-
   // Adapted from sglang: GDN kernels
   ops.def(
       "chunk_gated_delta_rule_cpu(Tensor query, Tensor key, Tensor value, "
@@ -499,12 +491,11 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("fused_gdn_gating_cpu", torch::kCPU, &fused_gdn_gating_cpu);
 
   // CPU attention kernels
-  ops.def("cpu_attn_has_isa(str isa) -> bool", &cpu_attn_has_isa);
   ops.def(
       "get_scheduler_metadata(int num_req, int num_heads_q, int num_heads_kv, "
       "int head_dim, Tensor seq_lens, ScalarType dtype, Tensor "
       "query_start_loc, bool casual, int window_size, str isa_hint, bool "
-      "enable_kv_split, Tensor? dynamic_causal) -> Tensor",
+      "enable_kv_split) -> Tensor",
       &get_scheduler_metadata);
   ops.def(
       "cpu_attn_reshape_and_cache(Tensor key, Tensor value, Tensor(a2!) "
@@ -516,9 +507,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "cpu_attention_with_kv_cache(Tensor query, Tensor key_cache, Tensor "
       "value_cache, Tensor(a3!) output, Tensor query_start_loc, Tensor "
       "seq_lens, float scale, bool causal, Tensor? alibi_slopes, SymInt "
-      "sliding_window_size, Tensor block_table, "
-      "float softcap, Tensor scheduler_metadata, Tensor? s_aux, Tensor? "
-      "dynamic_causal, "
+      "sliding_window_left, SymInt sliding_window_right, Tensor block_table, "
+      "float softcap, Tensor scheduler_metadata, Tensor? s_aux, "
       "float k_scale=1.0, float v_scale=1.0, str kv_cache_dtype=\"auto\") -> "
       "()",
       &cpu_attention_with_kv_cache);

csrc/cutlass_extensions/vllm_cutlass_library_extension.py

@@ -57,13 +57,13 @@ VLLMDataTypeVLLMScalarTypeTag: dict[VLLMDataType | DataType, str] = {
 }
 
 VLLMDataTypeTorchDataTypeTag: dict[VLLMDataType | DataType, str] = {
-    DataType.u8: "torch::headeronly::ScalarType::Byte",
-    DataType.s8: "torch::headeronly::ScalarType::Char",
-    DataType.e4m3: "torch::headeronly::ScalarType::Float8_e4m3fn",
-    DataType.s32: "torch::headeronly::ScalarType::Int",
-    DataType.f16: "torch::headeronly::ScalarType::Half",
-    DataType.bf16: "torch::headeronly::ScalarType::BFloat16",
-    DataType.f32: "torch::headeronly::ScalarType::Float",
+    DataType.u8: "at::ScalarType::Byte",
+    DataType.s8: "at::ScalarType::Char",
+    DataType.e4m3: "at::ScalarType::Float8_e4m3fn",
+    DataType.s32: "at::ScalarType::Int",
+    DataType.f16: "at::ScalarType::Half",
+    DataType.bf16: "at::ScalarType::BFloat16",
+    DataType.f32: "at::ScalarType::Float",
 }
 
 VLLMKernelScheduleTag: dict[MixedInputKernelScheduleType | KernelScheduleType, str] = {

csrc/libtorch_stable/activation_kernels.cu

@@ -10,20 +10,11 @@
 
 namespace vllm {
 
-// `alpha` and `beta` are applied to opposite operands:
-//   - alpha lives INSIDE the activation (the activated half): the gated
-//     activation computes act_half * sigmoid(alpha * act_half).
-//   - beta is added to the OTHER (non-activated) half before the multiply.
-// So the result is always ACT(act_half, alpha) * (other_half + beta).
-// Which half is which depends on `act_first` (see below). Defaults
-// alpha=1.0, beta=0.0 reproduce the plain SwiGLU/GeGLU behavior.
-template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&, const float),
+template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
           bool act_first, bool HAS_CLAMP>
 __device__ __forceinline__ scalar_t compute(const scalar_t& x,
                                             const scalar_t& y,
-                                            const float limit,
-                                            const float alpha,
-                                            const float beta) {
+                                            const float limit) {
   if constexpr (act_first) {
     scalar_t gate = x;
     scalar_t up = y;
@@ -31,9 +22,7 @@ __device__ __forceinline__ scalar_t compute(const scalar_t& x,
       gate = (scalar_t)fminf((float)gate, limit);
       up = (scalar_t)fmaxf(fminf((float)up, limit), -limit);
     }
-    // act_first: gate is the activated half -> alpha applies to gate;
-    // beta is added to up (the non-activated half).
-    return (scalar_t)(ACT_FN(gate, alpha) * ((float)up + beta));
+    return ACT_FN(gate) * up;
   } else {
     scalar_t gate = x;
     scalar_t up = y;
@@ -41,68 +30,55 @@ __device__ __forceinline__ scalar_t compute(const scalar_t& x,
       gate = (scalar_t)fmaxf(fminf((float)gate, limit), -limit);
       up = (scalar_t)fminf((float)up, limit);
     }
-    // !act_first: up is the activated half -> alpha applies to up;
-    // beta is added to gate (the non-activated half).
-    return (scalar_t)(((float)gate + beta) * ACT_FN(up, alpha));
+    return gate * ACT_FN(up);
   }
 }
 
-template <typename packed_t,
-          packed_t (*PACKED_ACT_FN)(const packed_t&, const float),
+template <typename packed_t, packed_t (*PACKED_ACT_FN)(const packed_t&),
           bool act_first, bool HAS_CLAMP>
 __device__ __forceinline__ packed_t packed_compute(const packed_t& x,
                                                    const packed_t& y,
-                                                   const float limit,
-                                                   const float alpha,
-                                                   const float beta) {
+                                                   const float limit) {
   if constexpr (act_first) {
     packed_t gate = x;
     packed_t up = y;
-    float2 u = cast_to_float2(up);
     if constexpr (HAS_CLAMP) {
       float2 g = cast_to_float2(gate);
+      float2 u = cast_to_float2(up);
       g.x = fminf(g.x, limit);
       g.y = fminf(g.y, limit);
       u.x = fmaxf(fminf(u.x, limit), -limit);
       u.y = fmaxf(fminf(u.y, limit), -limit);
       gate = cast_to_packed<packed_t>(g);
+      up = cast_to_packed<packed_t>(u);
     }
-    // act_first: gate is the activated half -> alpha applies to gate;
-    // beta is added to up (the non-activated half).
-    float2 activated = cast_to_float2(PACKED_ACT_FN(gate, alpha));
-    activated.x *= u.x + beta;
-    activated.y *= u.y + beta;
-    return cast_to_packed<packed_t>(activated);
+    return packed_mul(PACKED_ACT_FN(gate), up);
   } else {
     packed_t gate = x;
     packed_t up = y;
-    float2 g = cast_to_float2(gate);
     if constexpr (HAS_CLAMP) {
+      float2 g = cast_to_float2(gate);
       float2 u = cast_to_float2(up);
       g.x = fmaxf(fminf(g.x, limit), -limit);
       g.y = fmaxf(fminf(g.y, limit), -limit);
       u.x = fminf(u.x, limit);
       u.y = fminf(u.y, limit);
+      gate = cast_to_packed<packed_t>(g);
       up = cast_to_packed<packed_t>(u);
     }
-    // !act_first: up is the activated half -> alpha applies to up;
-    // beta is added to gate (the non-activated half).
-    float2 activated = cast_to_float2(PACKED_ACT_FN(up, alpha));
-    activated.x *= g.x + beta;
-    activated.y *= g.y + beta;
-    return cast_to_packed<packed_t>(activated);
+    return packed_mul(gate, PACKED_ACT_FN(up));
   }
 }
 
 // Activation and gating kernel template.
 template <typename scalar_t, typename packed_t,
-          scalar_t (*ACT_FN)(const scalar_t&, const float),
-          packed_t (*PACKED_ACT_FN)(const packed_t&, const float),
-          bool act_first, bool use_vec, bool HAS_CLAMP, bool use_256b = false>
+          scalar_t (*ACT_FN)(const scalar_t&),
+          packed_t (*PACKED_ACT_FN)(const packed_t&), bool act_first,
+          bool use_vec, bool HAS_CLAMP, bool use_256b = false>
 __global__ void act_and_mul_kernel(
     scalar_t* __restrict__ out,          // [..., d]
     const scalar_t* __restrict__ input,  // [..., 2, d]
-    const int d, const float limit, const float alpha, const float beta) {
+    const int d, const float limit) {
   const scalar_t* x_ptr = input + blockIdx.x * 2 * d;
   const scalar_t* y_ptr = x_ptr + d;
   scalar_t* out_ptr = out + blockIdx.x * d;
@@ -129,7 +105,7 @@ __global__ void act_and_mul_kernel(
       for (int j = 0; j < pvec_t::NUM_ELTS; j++) {
         x.elts[j] =
             packed_compute<packed_t, PACKED_ACT_FN, act_first, HAS_CLAMP>(
-                x.elts[j], y.elts[j], limit, alpha, beta);
+                x.elts[j], y.elts[j], limit);
       }
       if constexpr (use_256b) {
         st256(x, &out_vec[i]);
@@ -142,34 +118,29 @@ __global__ void act_and_mul_kernel(
     for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
       const scalar_t x = VLLM_LDG(&x_ptr[idx]);
       const scalar_t y = VLLM_LDG(&y_ptr[idx]);
-      out_ptr[idx] = compute<scalar_t, ACT_FN, act_first, HAS_CLAMP>(
-          x, y, limit, alpha, beta);
+      out_ptr[idx] =
+          compute<scalar_t, ACT_FN, act_first, HAS_CLAMP>(x, y, limit);
     }
   }
 }
 
-// Gated activations take an `alpha` argument that scales the sigmoid input
-// (`x * sigmoid(alpha * x)`). alpha defaults to 1.0 at all call sites, which
-// is exactly SiLU; only the clamp path (silu_and_mul_with_clamp) passes a
-// non-default alpha. Activations that do not use alpha simply ignore it.
 template <typename T>
-__device__ __forceinline__ T silu_kernel(const T& x, const float alpha) {
-  // x * sigmoid(alpha * x)
-  return (T)(((float)x) / (1.0f + expf((float)-x * alpha)));
+__device__ __forceinline__ T silu_kernel(const T& x) {
+  // x * sigmoid(x)
+  return (T)(((float)x) / (1.0f + expf((float)-x)));
 }
 
 template <typename packed_t>
-__device__ __forceinline__ packed_t packed_silu_kernel(const packed_t& val,
-                                                       const float alpha) {
-  // x * sigmoid(alpha * x)
+__device__ __forceinline__ packed_t packed_silu_kernel(const packed_t& val) {
+  // x * sigmoid(x)
   float2 fval = cast_to_float2(val);
-  fval.x = fval.x / (1.0f + expf(-fval.x * alpha));
-  fval.y = fval.y / (1.0f + expf(-fval.y * alpha));
+  fval.x = fval.x / (1.0f + expf(-fval.x));
+  fval.y = fval.y / (1.0f + expf(-fval.y));
   return cast_to_packed<packed_t>(fval);
 }
 
 template <typename T>
-__device__ __forceinline__ T gelu_kernel(const T& x, const float /*alpha*/) {
+__device__ __forceinline__ T gelu_kernel(const T& x) {
   // Equivalent to PyTorch GELU with 'none' approximation.
   // Refer to:
   // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L36-L38
@@ -179,8 +150,7 @@ __device__ __forceinline__ T gelu_kernel(const T& x, const float /*alpha*/) {
 }
 
 template <typename packed_t>
-__device__ __forceinline__ packed_t packed_gelu_kernel(const packed_t& val,
-                                                       const float /*alpha*/) {
+__device__ __forceinline__ packed_t packed_gelu_kernel(const packed_t& val) {
   // Equivalent to PyTorch GELU with 'none' approximation.
   // Refer to:
   // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L36-L38
@@ -192,8 +162,7 @@ __device__ __forceinline__ packed_t packed_gelu_kernel(const packed_t& val,
 }
 
 template <typename T>
-__device__ __forceinline__ T gelu_tanh_kernel(const T& x,
-                                              const float /*alpha*/) {
+__device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
   // Equivalent to PyTorch GELU with 'tanh' approximation.
   // Refer to:
   // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L25-L30
@@ -207,7 +176,7 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x,
 
 template <typename packed_t>
 __device__ __forceinline__ packed_t
-packed_gelu_tanh_kernel(const packed_t& val, const float /*alpha*/) {
+packed_gelu_tanh_kernel(const packed_t& val) {
   // Equivalent to PyTorch GELU with 'tanh' approximation.
   // Refer to:
   // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L25-L30
@@ -233,7 +202,7 @@ packed_gelu_tanh_kernel(const packed_t& val, const float /*alpha*/) {
 // clamped (max only) and up input is clamped (both sides) before the
 // activation function is applied.
 #define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, PACKED_KERNEL, ACT_FIRST,        \
-                                      HAS_CLAMP, LIMIT, ALPHA, BETA)           \
+                                      HAS_CLAMP, LIMIT)                        \
   auto dtype = input.scalar_type();                                            \
   int d = input.size(-1) / 2;                                                  \
   int64_t num_tokens = input.numel() / input.size(-1);                         \
@@ -261,7 +230,7 @@ packed_gelu_tanh_kernel(const packed_t& val, const float /*alpha*/) {
             PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
             ACT_FIRST, true, HAS_CLAMP, true><<<grid, block, 0, stream>>>(     \
             out.mutable_data_ptr<scalar_t>(),                                  \
-            input.const_data_ptr<scalar_t>(), d, LIMIT, ALPHA, BETA);          \
+            input.const_data_ptr<scalar_t>(), d, LIMIT);                       \
       });                                                                      \
     } else {                                                                   \
       VLLM_STABLE_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {   \
@@ -271,7 +240,7 @@ packed_gelu_tanh_kernel(const packed_t& val, const float /*alpha*/) {
             PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
             ACT_FIRST, true, HAS_CLAMP, false><<<grid, block, 0, stream>>>(    \
             out.mutable_data_ptr<scalar_t>(),                                  \
-            input.const_data_ptr<scalar_t>(), d, LIMIT, ALPHA, BETA);          \
+            input.const_data_ptr<scalar_t>(), d, LIMIT);                       \
       });                                                                      \
     }                                                                          \
   } else {                                                                     \
@@ -283,7 +252,7 @@ packed_gelu_tanh_kernel(const packed_t& val, const float /*alpha*/) {
           PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>,   \
           ACT_FIRST, false, HAS_CLAMP><<<grid, block, 0, stream>>>(            \
           out.mutable_data_ptr<scalar_t>(), input.const_data_ptr<scalar_t>(),  \
-          d, LIMIT, ALPHA, BETA);                                              \
+          d, LIMIT);                                                           \
     });                                                                        \
   }
 
@@ -291,18 +260,14 @@ void silu_and_mul(torch::stable::Tensor& out,    // [..., d]
                   torch::stable::Tensor& input)  // [..., 2 * d]
 {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
-                                true, false, 0.0f, 1.0f, 0.0f);
+                                true, false, 0.0f);
 }
 
 void silu_and_mul_clamp(torch::stable::Tensor& out,    // [..., d]
                         torch::stable::Tensor& input,  // [..., 2 * d]
-                        double limit, double alpha, double beta) {
-  // out = (gate.clamp(max=limit) * sigmoid(alpha * gate.clamp(max=limit)))
-  //       * (up.clamp(+-limit) + beta)
-  // alpha=1.0, beta=0.0 reduce this to silu(gate) * up.
+                        double limit) {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
-                                true, true, (float)limit, (float)alpha,
-                                (float)beta);
+                                true, true, (float)limit);
 }
 
 void mul_and_silu(torch::stable::Tensor& out,    // [..., d]
@@ -311,22 +276,21 @@ void mul_and_silu(torch::stable::Tensor& out,    // [..., d]
   // The difference between mul_and_silu and silu_and_mul is that mul_and_silu
   // applies the silu to the latter half of the input.
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
-                                false, false, 0.0f, 1.0f, 0.0f);
+                                false, false, 0.0f);
 }
 
 void gelu_and_mul(torch::stable::Tensor& out,    // [..., d]
                   torch::stable::Tensor& input)  // [..., 2 * d]
 {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel, vllm::packed_gelu_kernel,
-                                true, false, 0.0f, 1.0f, 0.0f);
+                                true, false, 0.0f);
 }
 
 void gelu_tanh_and_mul(torch::stable::Tensor& out,    // [..., d]
                        torch::stable::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel,
-                                vllm::packed_gelu_tanh_kernel, true, false,
-                                0.0f, 1.0f, 0.0f);
+  LAUNCH_ACTIVATION_GATE_KERNEL(
+      vllm::gelu_tanh_kernel, vllm::packed_gelu_tanh_kernel, true, false, 0.0f);
 }
 
 namespace vllm {

csrc/libtorch_stable/cuda_vec_utils.cuh

@@ -21,7 +21,7 @@
 // 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(CUDART_VERSION) && CUDART_VERSION >= 12090
+    defined(CUDA_VERSION) && CUDA_VERSION >= 12090
   #define VLLM_256B_PTX_ENABLED 1
 #else
   #define VLLM_256B_PTX_ENABLED 0

csrc/libtorch_stable/dispatch_utils.h

@@ -30,28 +30,6 @@
   THO_DISPATCH_SWITCH(TYPE, NAME,                            \
                       VLLM_STABLE_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
 
-#define VLLM_STABLE_DISPATCH_CASE_INTEGRAL_TYPES(...)                  \
-  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Byte, __VA_ARGS__)  \
-  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Char, __VA_ARGS__)  \
-  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Short, __VA_ARGS__) \
-  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Int, __VA_ARGS__)   \
-  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Long, __VA_ARGS__)
-
-#define VLLM_STABLE_DISPATCH_CASE_INTEGRAL_AND_UNSIGNED_TYPES(...)      \
-  VLLM_STABLE_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__)                 \
-  THO_DISPATCH_CASE(torch::headeronly::ScalarType::UInt16, __VA_ARGS__) \
-  THO_DISPATCH_CASE(torch::headeronly::ScalarType::UInt32, __VA_ARGS__) \
-  THO_DISPATCH_CASE(torch::headeronly::ScalarType::UInt64, __VA_ARGS__)
-
-#define VLLM_STABLE_DISPATCH_INTEGRAL_TYPES(TYPE, NAME, ...) \
-  THO_DISPATCH_SWITCH(TYPE, NAME,                            \
-                      VLLM_STABLE_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__))
-
-#define VLLM_STABLE_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(TYPE, NAME, ...) \
-  THO_DISPATCH_SWITCH(                                                    \
-      TYPE, NAME,                                                         \
-      VLLM_STABLE_DISPATCH_CASE_INTEGRAL_AND_UNSIGNED_TYPES(__VA_ARGS__))
-
 // FP8 type dispatch - ROCm uses FNUZ format, CUDA uses OCP format
 #ifdef USE_ROCM
   #define VLLM_STABLE_DISPATCH_CASE_FP8_TYPES(...)                 \

csrc/libtorch_stable/fp32_router_gemm.cu

@@ -175,52 +175,49 @@ void invokeFp32RouterGemm(float* output, InputT const* mat_a,
 }
 
 // ---------------------------------------------------------------------------
-// Explicit instantiations: M=1..32, for both input types, for the supported
-// (E, H) pairs:  (256, 3072) [MiniMax-M2/M2.5]  and  (128, 6144) [MiniMax-M3].
+// Explicit instantiations: M=1..32, E=256, H=3072, for both input types
 // ---------------------------------------------------------------------------
 
-#define INSTANTIATE(T, M, E, H)                                    \
-  template void invokeFp32RouterGemm<T, M, E, H>(float*, T const*, \
-                                                 float const*, cudaStream_t);
+#define INSTANTIATE(T, M)                              \
+  template void invokeFp32RouterGemm<T, M, 256, 3072>( \
+      float*, T const*, float const*, cudaStream_t);
 
-#define INSTANTIATE_ALL(T, E, H) \
-  INSTANTIATE(T, 1, E, H)        \
-  INSTANTIATE(T, 2, E, H)        \
-  INSTANTIATE(T, 3, E, H)        \
-  INSTANTIATE(T, 4, E, H)        \
-  INSTANTIATE(T, 5, E, H)        \
-  INSTANTIATE(T, 6, E, H)        \
-  INSTANTIATE(T, 7, E, H)        \
-  INSTANTIATE(T, 8, E, H)        \
-  INSTANTIATE(T, 9, E, H)        \
-  INSTANTIATE(T, 10, E, H)       \
-  INSTANTIATE(T, 11, E, H)       \
-  INSTANTIATE(T, 12, E, H)       \
-  INSTANTIATE(T, 13, E, H)       \
-  INSTANTIATE(T, 14, E, H)       \
-  INSTANTIATE(T, 15, E, H)       \
-  INSTANTIATE(T, 16, E, H)       \
-  INSTANTIATE(T, 17, E, H)       \
-  INSTANTIATE(T, 18, E, H)       \
-  INSTANTIATE(T, 19, E, H)       \
-  INSTANTIATE(T, 20, E, H)       \
-  INSTANTIATE(T, 21, E, H)       \
-  INSTANTIATE(T, 22, E, H)       \
-  INSTANTIATE(T, 23, E, H)       \
-  INSTANTIATE(T, 24, E, H)       \
-  INSTANTIATE(T, 25, E, H)       \
-  INSTANTIATE(T, 26, E, H)       \
-  INSTANTIATE(T, 27, E, H)       \
-  INSTANTIATE(T, 28, E, H)       \
-  INSTANTIATE(T, 29, E, H)       \
-  INSTANTIATE(T, 30, E, H)       \
-  INSTANTIATE(T, 31, E, H)       \
-  INSTANTIATE(T, 32, E, H)
+#define INSTANTIATE_ALL(T) \
+  INSTANTIATE(T, 1)        \
+  INSTANTIATE(T, 2)        \
+  INSTANTIATE(T, 3)        \
+  INSTANTIATE(T, 4)        \
+  INSTANTIATE(T, 5)        \
+  INSTANTIATE(T, 6)        \
+  INSTANTIATE(T, 7)        \
+  INSTANTIATE(T, 8)        \
+  INSTANTIATE(T, 9)        \
+  INSTANTIATE(T, 10)       \
+  INSTANTIATE(T, 11)       \
+  INSTANTIATE(T, 12)       \
+  INSTANTIATE(T, 13)       \
+  INSTANTIATE(T, 14)       \
+  INSTANTIATE(T, 15)       \
+  INSTANTIATE(T, 16)       \
+  INSTANTIATE(T, 17)       \
+  INSTANTIATE(T, 18)       \
+  INSTANTIATE(T, 19)       \
+  INSTANTIATE(T, 20)       \
+  INSTANTIATE(T, 21)       \
+  INSTANTIATE(T, 22)       \
+  INSTANTIATE(T, 23)       \
+  INSTANTIATE(T, 24)       \
+  INSTANTIATE(T, 25)       \
+  INSTANTIATE(T, 26)       \
+  INSTANTIATE(T, 27)       \
+  INSTANTIATE(T, 28)       \
+  INSTANTIATE(T, 29)       \
+  INSTANTIATE(T, 30)       \
+  INSTANTIATE(T, 31)       \
+  INSTANTIATE(T, 32)
 
-INSTANTIATE_ALL(float, 256, 3072)
-INSTANTIATE_ALL(__nv_bfloat16, 256, 3072)
-INSTANTIATE_ALL(float, 128, 6144)
-INSTANTIATE_ALL(__nv_bfloat16, 128, 6144)
+INSTANTIATE_ALL(float)
+INSTANTIATE_ALL(__nv_bfloat16)
 
 #undef INSTANTIATE_ALL
 #undef INSTANTIATE

csrc/libtorch_stable/fp32_router_gemm_entry.cu

@@ -22,42 +22,36 @@ inline int getSMVersion() {
 
 }  // namespace
 
+static constexpr int FP32_NUM_EXPERTS = 256;
+static constexpr int FP32_HIDDEN_DIM = 3072;
 static constexpr int FP32_MAX_TOKENS = 32;
 
-// Supported (hidden_dim, num_experts) pairs (must match the instantiations in
-// fp32_router_gemm.cu): (3072, 256) for MiniMax-M2/M2.5, (6144, 128) for M3.
-static inline bool fp32_router_gemm_supported(int hidden_dim, int num_experts) {
-  return (hidden_dim == 3072 && num_experts == 256) ||
-         (hidden_dim == 6144 && num_experts == 128);
-}
-
 // Forward declarations — 4 template params must match fp32_router_gemm.cu
 template <typename InputT, int kNumTokens, int kNumExperts, int kHiddenDim>
 void invokeFp32RouterGemm(float* output, InputT const* mat_a,
                           float const* mat_b, cudaStream_t stream);
 
-// LoopUnroller templated on InputT, kNumExperts and kHiddenDim
-template <typename InputT, int kNumExperts, int kHiddenDim, int kBegin,
-          int kEnd>
+// LoopUnroller templated on InputT
+template <typename InputT, int kBegin, int kEnd>
 struct Fp32LoopUnroller {
   static void unroll(int num_tokens, float* output, InputT const* mat_a,
                      float const* mat_b, cudaStream_t stream) {
     if (num_tokens == kBegin) {
-      invokeFp32RouterGemm<InputT, kBegin, kNumExperts, kHiddenDim>(
+      invokeFp32RouterGemm<InputT, kBegin, FP32_NUM_EXPERTS, FP32_HIDDEN_DIM>(
           output, mat_a, mat_b, stream);
     } else {
-      Fp32LoopUnroller<InputT, kNumExperts, kHiddenDim, kBegin + 1,
-                       kEnd>::unroll(num_tokens, output, mat_a, mat_b, stream);
+      Fp32LoopUnroller<InputT, kBegin + 1, kEnd>::unroll(num_tokens, output,
+                                                         mat_a, mat_b, stream);
     }
   }
 };
 
-template <typename InputT, int kNumExperts, int kHiddenDim, int kEnd>
-struct Fp32LoopUnroller<InputT, kNumExperts, kHiddenDim, kEnd, kEnd> {
+template <typename InputT, int kEnd>
+struct Fp32LoopUnroller<InputT, kEnd, kEnd> {
   static void unroll(int num_tokens, float* output, InputT const* mat_a,
                      float const* mat_b, cudaStream_t stream) {
     if (num_tokens == kEnd) {
-      invokeFp32RouterGemm<InputT, kEnd, kNumExperts, kHiddenDim>(
+      invokeFp32RouterGemm<InputT, kEnd, FP32_NUM_EXPERTS, FP32_HIDDEN_DIM>(
           output, mat_a, mat_b, stream);
     } else {
       throw std::invalid_argument(
@@ -66,23 +60,6 @@ struct Fp32LoopUnroller<InputT, kNumExperts, kHiddenDim, kEnd, kEnd> {
   }
 };
 
-// Dispatch over the supported (num_experts, hidden_dim) pairs.
-template <typename InputT>
-void dispatchFp32RouterGemm(int num_experts, int hidden_dim, int num_tokens,
-                            float* output, InputT const* mat_a,
-                            float const* mat_b, cudaStream_t stream) {
-  if (num_experts == 256 && hidden_dim == 3072) {
-    Fp32LoopUnroller<InputT, 256, 3072, 1, FP32_MAX_TOKENS>::unroll(
-        num_tokens, output, mat_a, mat_b, stream);
-  } else if (num_experts == 128 && hidden_dim == 6144) {
-    Fp32LoopUnroller<InputT, 128, 6144, 1, FP32_MAX_TOKENS>::unroll(
-        num_tokens, output, mat_a, mat_b, stream);
-  } else {
-    throw std::invalid_argument(
-        "fp32_router_gemm: unsupported (hidden_dim, num_experts) pair");
-  }
-}
-
 void fp32_router_gemm(
     torch::stable::Tensor& output,       // [num_tokens, num_experts]
     torch::stable::Tensor const& mat_a,  // [num_tokens, hidden_dim]
@@ -108,10 +85,10 @@ void fp32_router_gemm(
   STD_TORCH_CHECK(
       mat_a.size(1) == mat_b.size(1),
       "fp32_router_gemm: mat_a and mat_b must have the same hidden_dim");
-  STD_TORCH_CHECK(
-      fp32_router_gemm_supported(hidden_dim, num_experts),
-      "fp32_router_gemm: supported (hidden_dim, num_experts) pairs are "
-      "(3072, 256) and (6144, 128)");
+  STD_TORCH_CHECK(hidden_dim == FP32_HIDDEN_DIM,
+                  "fp32_router_gemm: expected hidden_dim=3072");
+  STD_TORCH_CHECK(num_experts == FP32_NUM_EXPERTS,
+                  "fp32_router_gemm: expected num_experts=256");
   STD_TORCH_CHECK(num_tokens <= FP32_MAX_TOKENS,
                   "fp32_router_gemm: num_tokens must be in [0, 32]");
   STD_TORCH_CHECK(
@@ -136,13 +113,12 @@ void fp32_router_gemm(
   if (mat_a.scalar_type() == torch::headeronly::ScalarType::BFloat16) {
     auto const* mat_a_ptr =
         reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr());
-    dispatchFp32RouterGemm<__nv_bfloat16>(num_experts, hidden_dim, num_tokens,
-                                          out_ptr, mat_a_ptr, mat_b_ptr,
-                                          stream);
+    Fp32LoopUnroller<__nv_bfloat16, 1, FP32_MAX_TOKENS>::unroll(
+        num_tokens, out_ptr, mat_a_ptr, mat_b_ptr, stream);
   } else {
     auto const* mat_a_ptr = reinterpret_cast<float const*>(mat_a.data_ptr());
-    dispatchFp32RouterGemm<float>(num_experts, hidden_dim, num_tokens, out_ptr,
-                                  mat_a_ptr, mat_b_ptr, stream);
+    Fp32LoopUnroller<float, 1, FP32_MAX_TOKENS>::unroll(
+        num_tokens, out_ptr, mat_a_ptr, mat_b_ptr, stream);
   }
 }
 

csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu

@@ -18,7 +18,7 @@
  *   ROPE_DIM  = 64   (RoPE applied to dims [NOPE_DIM, HEAD_DIM))
  *   NOPE_DIM  = 448
  *   QUANT_BLOCK = 64 (UE8M0 FP8 quant block)
- *   FP8_MAX   = 224.0f on ROCm FNUZ / 448.0f on OCP
+ *   FP8_MAX   = 448.0f
  *   is_neox=false (GPT-J interleaved pairs)
  *   cos_sin_cache layout [max_pos, rope_dim] = cos || sin (cos first, sin
  *     second along last dim; each half is rope_dim/2 = 32 values)
@@ -61,11 +61,10 @@
 #ifdef USE_ROCM
 // ROCm-compatible FP8 conversion helpers
 __device__ __forceinline__ uint8_t rocm_cvt_float_to_fp8_e4m3(float val) {
-  // gfx942 uses FNUZ FP8; other ROCm targets use OCP E4M3.
-  #if defined(__gfx942__)
-  __hip_fp8_e4m3_fnuz fp8_val(val);
-  #else
+  #if defined(HIP_FP8_TYPE_OCP)
   __hip_fp8_e4m3 fp8_val(val);
+  #else
+  __hip_fp8_e4m3_fnuz fp8_val(val);
   #endif
   return reinterpret_cast<uint8_t&>(fp8_val);
 }
@@ -91,13 +90,7 @@ constexpr int kQuantBlock = 64;
 constexpr int kNumQuantBlocks = kNopeDim / kQuantBlock;   // 7
 constexpr int kScaleBytesPerToken = kNumQuantBlocks + 1;  // 8 (7 real + 1 pad)
 constexpr int kTokenDataBytes = kNopeDim + kRopeDim * 2;  // 448 + 128 = 576
-// FNUZ on gfx942 / OCP elsewhere. FNUZ uses 224.0 (not the dtype's raw
-// 240.0) to match the rest of vLLM's FNUZ pipeline.
-#if defined(USE_ROCM) && defined(__gfx942__)
-constexpr float kFp8Max = 224.0f;
-#else
 constexpr float kFp8Max = 448.0f;
-#endif
 
 #ifndef USE_ROCM
 // When num_tokens is less than this threshold,

csrc/libtorch_stable/fused_minimax_m3_qknorm_rope_kv_insert_kernel.cu

@@ -1,675 +0,0 @@
-/*
- * SPDX-License-Identifier: Apache-2.0
- * SPDX-FileCopyrightText: Copyright contributors to the vLLM project
- *
- * Horizontally-fused MiniMax-M3 attention pre-processing kernel.
- *
- * Replaces the per-token Python sequence in
- * ``MiniMaxM3SparseAttention.forward`` / ``MiniMaxM3Attention.forward``:
- *
- *     q  = q_norm(q);  k = k_norm(k);  q, k = rotary_emb(pos, q, k)
- *     index_q = index_q_norm(index_q);  index_k = index_k_norm(index_k)
- *     index_q, index_k = rotary_emb(pos, index_q, index_k)
- *     _insert_kv(k, v, index_k)
- *
- * All branches share head_dim=128 and the *same* partial-NeoX RoPE table
- * (``rotary_dim`` rotated, the trailing dims pass through).  The four norms
- * are Gemma-style RMSNorm (``x * rsqrt(mean(x^2)+eps) * (1 + weight)``) with
- * independent weights.
- *
- * Everything lives in a single fused ``qkv`` tensor.  The sparse layer's
- * fused projection (MinimaxM3QKVParallelLinearWithIndexer) emits, per token::
- *
- *     [ q | k | v | index_q | index_k ]   (the "5 results")
- *
- * while the dense layer emits just ``[ q | k | v ]``.  The kernel reads the
- * index branch straight out of that packed row -- no separate index tensors.
- *
- * One kernel, one grid; each warp owns one (token, head-slot) pair.  Slot
- * enumeration per token:
- *     [0, nq)                         Q  heads   -> norm(q_w)  + RoPE, write
- * qkv [nq, nq+nkv)                    K  heads   -> norm(k_w)  + RoPE, write
- * qkv
- *                                                  (+ insert into key cache)
- *     [nq+nkv, nq+2*nkv)              V  heads   -> insert into value cache
- *     IQ heads (niq)                             -> norm(iq_w) + RoPE, write iq
- *     IK       (1)                               -> norm(ik_w) + RoPE
- *                                                  (+ insert into index cache)
- *
- * The IQ/IK warps address the index_q/index_k sub-blocks *inside* qkv at the
- * fixed physical offsets (nq+2*nkv)*128 and (nq+2*nkv+niq)*128.
- *
- * Dense vs sparse is a compile-time choice via the ``kIsSparse``/``kInsertKV``
- * template bools (3 instantiations: dense <false,false>, sparse-profiling
- * <true,false>, sparse-serving <true,true>), so the index slots, the V slots
- * and the cache inserts fold away entirely on paths that don't use them. The
- * dense layer passes no caches/index: norm+RoPE happens in place and the
- * generic ``Attention`` layer owns the cache write.
- *
- * Q/K and (sparse) index_q/index_k are all rewritten in place inside the fused
- * ``qkv`` tensor.  Caches (bf16) are scatter-written by slot.
- */
-
-#include <cmath>
-#include <cuda_runtime.h>
-#include <type_traits>
-
-#include "torch_utils.h"
-
-#include "../cuda_compat.h"
-#include "../type_convert.cuh"
-#include "../attention/dtype_fp8.cuh"
-#include "dispatch_utils.h"
-
-#ifdef USE_ROCM
-  #include "../quantization/w8a8/fp8/amd/quant_utils.cuh"
-#else
-  #include "../quantization/w8a8/fp8/nvidia/quant_utils.cuh"
-#endif
-
-#ifndef FINAL_MASK
-  #ifdef USE_ROCM
-    #define FINAL_MASK 0xffffffffffffffffULL
-  #else
-    #define FINAL_MASK 0xffffffffu
-  #endif
-#endif
-
-namespace vllm {
-namespace minimax_m3_fused_ops {
-
-namespace {
-inline int getSMVersion() {
-  auto* props = get_device_prop();
-  return props->major * 10 + props->minor;
-}
-}  // namespace
-
-// ────────────────────────────────────────────────────────────────────────────
-// Constants (hard-coded for MiniMax-M3-preview).
-// ────────────────────────────────────────────────────────────────────────────
-constexpr int kHeadDim = 128;
-constexpr int kNumLanes = 32;
-constexpr int kElemsPerLane = kHeadDim / kNumLanes;  // 4
-
-// ────────────────────────────────────────────────────────────────────────────
-// Helpers
-// ────────────────────────────────────────────────────────────────────────────
-__device__ __forceinline__ float warpReduceSum(float val) {
-#pragma unroll
-  for (int mask = 16; mask > 0; mask >>= 1) {
-    val += __shfl_xor_sync(FINAL_MASK, val, mask, 32);
-  }
-  return val;
-}
-
-// Gemma RMSNorm over the full head (no-op when ``weight == nullptr``), rounded
-// back to scalar_t like the materialized unfused norm output, followed by
-// partial NeoX RoPE on the leading ``rotary_dim`` dims. Each lane owns
-// ``kElemsPerLane`` contiguous dims [laneId*4, laneId*4+4).
-template <typename scalar_t>
-__device__ __forceinline__ void normAndRope(
-    float (&elems)[kElemsPerLane], int const laneId, float const eps,
-    scalar_t const* __restrict__ weight,  // [kHeadDim] or nullptr (no norm)
-    bool const do_rope, int const rotary_dim,
-    scalar_t const* __restrict__ cos_ptr,  // cos_sin_cache + pos*rotary_dim
-    bool const apply_norm) {
-  // ── Gemma RMSNorm: x * rsqrt(mean(x^2)+eps) * (1 + w) ──────────────────
-  if (apply_norm) {
-    float sumsq = 0.0f;
-#pragma unroll
-    for (int i = 0; i < kElemsPerLane; i++) sumsq += elems[i] * elems[i];
-    sumsq = warpReduceSum(sumsq);
-    float const rms_rcp = rsqrtf(sumsq / static_cast<float>(kHeadDim) + eps);
-#pragma unroll
-    for (int i = 0; i < kElemsPerLane; i++) {
-      int const dim = laneId * kElemsPerLane + i;
-      float const w = 1.0f + static_cast<float>(weight[dim]);
-      elems[i] = elems[i] * rms_rcp * w;
-    }
-  }
-
-  // ── Partial NeoX RoPE on dims [0, rotary_dim) ──────────────────────────
-  // half = rotary_dim/2.  Pair (i, i+half) for i in [0, half).  Lane L owns
-  // dims [4L, 4L+4); since half is a multiple of 4, a lane lies wholly in the
-  // first half (own=x[i]) or second half (own=x[i+half]); its partner lives
-  // ``half/4`` lanes away (XOR with that distance).
-  if (do_rope) {
-    int const half = rotary_dim / 2;
-    int const dim0 = laneId * kElemsPerLane;
-    bool const in_rope = dim0 < rotary_dim;
-    int const lane_xor = half / kElemsPerLane;  // partner-lane distance
-
-    float partner[kElemsPerLane];
-#pragma unroll
-    for (int i = 0; i < kElemsPerLane; i++) {
-      partner[i] = __shfl_xor_sync(FINAL_MASK, elems[i], lane_xor, 32);
-    }
-    if (in_rope) {
-      bool const first_half = dim0 < half;
-      int const i_base = first_half ? dim0 : (dim0 - half);  // cos/sin index
-      scalar_t const* sin_ptr = cos_ptr + half;
-#pragma unroll
-      for (int i = 0; i < kElemsPerLane; i++) {
-        float const c = static_cast<float>(cos_ptr[i_base + i]);
-        float const s = static_cast<float>(sin_ptr[i_base + i]);
-        if (first_half) {
-          elems[i] = elems[i] * c - partner[i] * s;
-        } else {
-          elems[i] = elems[i] * c + partner[i] * s;
-        }
-      }
-    }
-  }
-}
-
-// Load 4 contiguous bf16 -> 4 fp32 registers.
-template <typename scalar_t>
-__device__ __forceinline__ void loadElems(scalar_t const* __restrict__ src,
-                                          float (&elems)[kElemsPerLane]) {
-  using Converter = vllm::_typeConvert<scalar_t>;
-  uint2 v = *reinterpret_cast<uint2 const*>(src);
-  auto const* p =
-      reinterpret_cast<typename Converter::packed_hip_type const*>(&v);
-#pragma unroll
-  for (int i = 0; i < kElemsPerLane / 2; i++) {
-    float2 f2 = Converter::convert(p[i]);
-    elems[2 * i] = f2.x;
-    elems[2 * i + 1] = f2.y;
-  }
-}
-
-// Store 4 fp32 registers -> 4 contiguous bf16.
-template <typename scalar_t>
-__device__ __forceinline__ void storeElems(
-    scalar_t* __restrict__ dst, float const (&elems)[kElemsPerLane]) {
-  using Converter = vllm::_typeConvert<scalar_t>;
-  uint2 v;
-  auto* p = reinterpret_cast<typename Converter::packed_hip_type*>(&v);
-#pragma unroll
-  for (int i = 0; i < kElemsPerLane / 2; i++) {
-    p[i] = Converter::convert(make_float2(elems[2 * i], elems[2 * i + 1]));
-  }
-  *reinterpret_cast<uint2*>(dst) = v;
-}
-
-template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
-__device__ __forceinline__ void storeCacheElems(
-    cache_t* __restrict__ dst, float const (&elems)[kElemsPerLane]) {
-  if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
-    // kAuto means unquantized KV cache here: cache_t == scalar_t, so store the
-    // model dtype directly. FP8 cache dtypes use the conversion path below.
-    storeElems<scalar_t>(reinterpret_cast<scalar_t*>(dst), elems);
-  } else {
-#pragma unroll
-    for (int i = 0; i < kElemsPerLane; i++) {
-      dst[i] = fp8::scaled_convert<cache_t, float, kv_dt>(elems[i], 1.0f);
-    }
-  }
-}
-
-// ────────────────────────────────────────────────────────────────────────────
-// Kernel
-// ────────────────────────────────────────────────────────────────────────────
-// Grid: 1D, ceil(num_tokens * slots_per_token / warps_per_block).
-// Each warp = one (token, slot).
-//
-// `kIsSparse` and `kInsertKV` are compile-time template bools, so all the
-// branch decisions that distinguish the dense layer from the sparse layer
-// (index slots, KV/index inserts, V slots) fold away per instantiation.
-// Three instantiations are built: dense <false,false>, sparse-profiling
-// <true,false> and sparse-serving <true,true>.  Slots per token:
-//     Q : nq                            (always — norm+RoPE)
-//     K : nkv                           (always — norm+RoPE; +K-cache insert)
-//     V : nkv  only if kInsertKV        (V-cache insert; no warps in dense)
-//     IQ: niq  only if kIsSparse        (norm+RoPE)
-//     IK: 1    only if kIsSparse        (norm+RoPE; +index-cache insert)
-template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt,
-          bool kIsSparse, bool kInsertKV>
-__global__ void fusedMiniMaxM3QNormRopeKVInsertKernel(
-    scalar_t* __restrict__ qkv,  // [N, qkv_row] in/out (packs index if sparse)
-    scalar_t* __restrict__ q_out,        // [N, nq*128] contiguous, or nullptr
-    scalar_t* __restrict__ index_q_out,  // [N, niq*128] contiguous, or nullptr
-    scalar_t const* __restrict__ q_norm_w,
-    scalar_t const* __restrict__ k_norm_w,
-    scalar_t const* __restrict__ iq_norm_w,
-    scalar_t const* __restrict__ ik_norm_w,
-    scalar_t const* __restrict__ cos_sin_cache,  // [max_pos, rotary_dim]
-    int64_t const* __restrict__ positions,       // [N] i64
-    int64_t const* __restrict__ slot_mapping,    // main K/V slots or nullptr
-    int64_t const* __restrict__ index_slot_mapping,  // index K slots/nullptr
-    cache_t* __restrict__ kv_cache,      // [nb,2,bs,nkv,128] or nullptr
-    scalar_t* __restrict__ index_cache,  // [nb*bs, 128] or nullptr
-    float const eps, int const rotary_dim, int const num_tokens, int const nq,
-    int const nkv, int const niq, int const block_size,
-    // kv_cache strides (in elements) for logical shape [nb, 2, bs, nkv, 128].
-    // The head_dim (last) dim is always innermost-contiguous (stride 1), so the
-    // NHD/HND layout choice is fully captured by these four strides: NHD keeps
-    // s_token < s_head, HND swaps them. dim_base addresses head_dim directly.
-    int64_t const kv_s_block, int64_t const kv_s_kv, int64_t const kv_s_token,
-    int64_t const kv_s_head) {
-#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
-  // _typeConvert<BFloat16> is unavailable on pre-Ampere; the M3 kernel only
-  // runs with bf16/fp16 inputs in practice.  Discard the bf16 body there.
-  if constexpr (std::is_same_v<scalar_t, c10::BFloat16>) {
-    return;
-  } else {
-#endif
-    int const warpsPerBlock = blockDim.x / 32;
-    int const laneId = threadIdx.x % 32;
-    int const globalWarpIdx = blockIdx.x * warpsPerBlock + (threadIdx.x / 32);
-
-    // Slot layout (compile-time gated: dense has neither V nor index slots).
-    int const v_slots = kInsertKV ? nkv : 0;
-    int const idx_slots = kIsSparse ? niq + 1 : 0;
-    int const slots_per_token = nq + nkv + v_slots + idx_slots;
-
-    int const tokenIdx = globalWarpIdx / slots_per_token;
-    int const slot = globalWarpIdx % slots_per_token;
-    if (tokenIdx >= num_tokens) return;
-
-    // Slot boundaries.
-    int const k_begin = nq;
-    int const v_begin = nq + nkv;             // valid only when kInsertKV
-    int const iq_begin = nq + nkv + v_slots;  // index block start
-    int const ik_slot = iq_begin + niq;       // valid only when kIsSparse
-
-    bool const isQ = slot < k_begin;
-    bool const isK = slot >= k_begin && slot < v_begin;
-    bool isV = false;
-    if constexpr (kInsertKV) isV = slot >= v_begin && slot < v_begin + nkv;
-    bool isIQ = false, isIK = false;
-    if constexpr (kIsSparse) {
-      isIQ = slot >= iq_begin && slot < ik_slot;
-      isIK = slot == ik_slot;
-    }
-
-    int const dim_base = laneId * kElemsPerLane;
-    // Physical row width of qkv: the dense layer packs [q|k|v]; the sparse
-    // layer additionally packs [index_q (niq heads) | index_k (1 head)].
-    int const qkv_row = (nq + 2 * nkv + (kIsSparse ? (niq + 1) : 0)) * kHeadDim;
-
-    // ── Resolve source pointer + per-branch parameters. ────────────────────
-    scalar_t* row_ptr = nullptr;       // in-place output location
-    scalar_t const* norm_w = nullptr;  // nullptr -> skip norm (V)
-    bool do_rope = true;
-    int head = 0;  // kv head index for inserts
-
-    if (isQ) {
-      row_ptr =
-          qkv + static_cast<int64_t>(tokenIdx) * qkv_row + slot * kHeadDim;
-      norm_w = q_norm_w;
-    } else if (isK) {
-      head = slot - k_begin;
-      row_ptr =
-          qkv + static_cast<int64_t>(tokenIdx) * qkv_row + slot * kHeadDim;
-      norm_w = k_norm_w;
-    } else if (isV) {
-      // qkv V section starts at slot index (nq + nkv): slot * kHeadDim is the
-      // correct in-tensor offset.
-      head = slot - v_begin;
-      row_ptr =
-          qkv + static_cast<int64_t>(tokenIdx) * qkv_row + slot * kHeadDim;
-      norm_w = nullptr;  // V: no norm, no rope
-      do_rope = false;
-    } else if (isIQ) {
-      // index_q sub-block lives at physical offset (nq+2*nkv)*128 in qkv.
-      int const ih = slot - iq_begin;
-      row_ptr = qkv + static_cast<int64_t>(tokenIdx) * qkv_row +
-                (nq + 2 * nkv + ih) * kHeadDim;
-      norm_w = iq_norm_w;
-    } else {  // isIK -- single shared index key at (nq+2*nkv+niq)*128.
-      row_ptr = qkv + static_cast<int64_t>(tokenIdx) * qkv_row +
-                (nq + 2 * nkv + niq) * kHeadDim;
-      norm_w = ik_norm_w;
-    }
-
-    // Store destination.  Q and index_q are gathered into dedicated contiguous
-    // output buffers (when provided) so the downstream SM100 sparse kernel's
-    // flat TMA descriptor can address them as [tokens*heads, head_dim]; this
-    // folds the de-interleaving into the store the kernel already does, instead
-    // of a separate q.contiguous() copy.  Everything else stays in place.
-    scalar_t* store_ptr = row_ptr;
-    if (isQ && q_out != nullptr) {
-      store_ptr = q_out + static_cast<int64_t>(tokenIdx) * nq * kHeadDim +
-                  slot * kHeadDim;
-    } else if (isIQ && index_q_out != nullptr) {
-      store_ptr = index_q_out +
-                  static_cast<int64_t>(tokenIdx) * niq * kHeadDim +
-                  (slot - iq_begin) * kHeadDim;
-    }
-
-    // PDL: wait for the predecessor kernel (the qkv-projection GEMM that
-    // produces ``qkv``) to finish before touching any global memory.  No-op
-    // when PDL is not enabled on the launch.  The CUDA runtime wrapper emits
-    // the griddepcontrol.wait PTX with the required memory clobber internally.
-#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
-    cudaGridDependencySynchronize();
-#endif
-
-    // ── Load -> norm+rope (fp32) -> store back in place. ───────────────────
-    float elems[kElemsPerLane];
-    loadElems<scalar_t>(row_ptr + dim_base, elems);
-
-    if (!isV) {
-      int64_t const pos = positions[tokenIdx];
-      scalar_t const* cos_ptr = cos_sin_cache + pos * rotary_dim;
-      normAndRope<scalar_t>(elems, laneId, eps, norm_w, do_rope, rotary_dim,
-                            cos_ptr, /*apply_norm=*/norm_w != nullptr);
-      storeElems<scalar_t>(store_ptr + dim_base, elems);
-    }
-
-    // ── Cache inserts (sparse serving only). ───────────────────────────────
-    if constexpr (kInsertKV) {
-      // Guard (not early-return) so every thread reaches the PDL trigger below.
-      int64_t const sm = (isK || isV)
-                             ? slot_mapping[tokenIdx]
-                             : (isIK ? index_slot_mapping[tokenIdx] : -1);
-      if (sm >= 0) {  // skip padded / unscheduled tokens
-        if (isIK) {
-          scalar_t* dst = index_cache + sm * kHeadDim + dim_base;
-          storeElems<scalar_t>(dst, elems);
-        } else if (isK || isV) {
-          // kv_cache logical shape [num_blocks, 2, block_size, nkv, head_dim].
-          // Paging is logical (block = sm/block_size, token = sm%block_size);
-          // the physical NHD/HND layout is honoured via the passed strides.
-          int64_t const b = sm / block_size;
-          int64_t const t = sm % block_size;
-          int const kv = isK ? 0 : 1;
-          int64_t const off =
-              b * kv_s_block + kv * kv_s_kv + t * kv_s_token + head * kv_s_head;
-          storeCacheElems<scalar_t, cache_t, kv_dt>(kv_cache + off + dim_base,
-                                                    elems);
-        }
-      }
-    }
-
-    // PDL: signal that this kernel is done so a dependent successor may launch
-    // early.  No-op when PDL is not enabled on the launch.
-#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
-    cudaTriggerProgrammaticLaunchCompletion();
-#endif
-#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
-  }
-#endif
-}
-
-// ────────────────────────────────────────────────────────────────────────────
-// Launch wrapper
-// ────────────────────────────────────────────────────────────────────────────
-template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
-void launchFusedMiniMaxM3(scalar_t* qkv, scalar_t* q_out, scalar_t* index_q_out,
-                          scalar_t const* q_norm_w, scalar_t const* k_norm_w,
-                          scalar_t const* iq_norm_w, scalar_t const* ik_norm_w,
-                          scalar_t const* cos_sin_cache,
-                          int64_t const* positions, int64_t const* slot_mapping,
-                          int64_t const* index_slot_mapping, cache_t* kv_cache,
-                          scalar_t* index_cache, float const eps,
-                          int const rotary_dim, int const num_tokens,
-                          int const nq, int const nkv, int const niq,
-                          int const block_size, int64_t const kv_s_block,
-                          int64_t const kv_s_kv, int64_t const kv_s_token,
-                          int64_t const kv_s_head, bool const has_index,
-                          bool const insert_kv, cudaStream_t stream) {
-  // Slot count must match the kernel's compile-time gating.
-  int const v_slots = insert_kv ? nkv : 0;
-  int const idx_slots = has_index ? niq + 1 : 0;
-  int const slots_per_token = nq + nkv + v_slots + idx_slots;
-
-  constexpr int kBlockSize = 256;
-  constexpr int kWarpsPerBlock = kBlockSize / 32;
-  int64_t const total_warps =
-      static_cast<int64_t>(num_tokens) * slots_per_token;
-  int const grid =
-      static_cast<int>((total_warps + kWarpsPerBlock - 1) / kWarpsPerBlock);
-  if (grid == 0) return;
-
-#ifndef USE_ROCM
-  // PDL: enable programmatic stream serialization whenever the hardware
-  // supports it (SM90+).  On pre-Hopper GPUs the attribute is unavailable, so
-  // leave numAttrs = 0 and launch as a regular kernel via cudaLaunchKernelEx.
-  static int const sm_version = getSMVersion();
-  cudaLaunchConfig_t config;
-  config.gridDim = dim3(grid);
-  config.blockDim = dim3(kBlockSize);
-  config.dynamicSmemBytes = 0;
-  config.stream = stream;
-  cudaLaunchAttribute attrs[1];
-  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
-  attrs[0].val.programmaticStreamSerializationAllowed = 1;
-  config.attrs = attrs;
-  config.numAttrs = (sm_version >= 90) ? 1 : 0;
-
-  #define LAUNCH(IS_SPARSE, INSERT)                                           \
-    cudaLaunchKernelEx(                                                       \
-        &config,                                                              \
-        fusedMiniMaxM3QNormRopeKVInsertKernel<scalar_t, cache_t, kv_dt,       \
-                                              IS_SPARSE, INSERT>,             \
-        qkv, q_out, index_q_out, q_norm_w, k_norm_w, iq_norm_w, ik_norm_w,    \
-        cos_sin_cache, positions, slot_mapping, index_slot_mapping, kv_cache, \
-        index_cache, eps, rotary_dim, num_tokens, nq, nkv, niq, block_size,   \
-        kv_s_block, kv_s_kv, kv_s_token, kv_s_head)
-#else
-  // ROCm: standard kernel launch syntax (no PDL/stream serialization).
-  // clang-format off
-  #define LAUNCH(IS_SPARSE, INSERT)                                          \
-    fusedMiniMaxM3QNormRopeKVInsertKernel<scalar_t, cache_t, kv_dt,          \
-                                          IS_SPARSE, INSERT>                 \
-        <<<grid, kBlockSize, 0, stream>>>(                                   \
-            qkv, q_out, index_q_out, q_norm_w, k_norm_w, iq_norm_w,          \
-            ik_norm_w, cos_sin_cache, positions, slot_mapping,               \
-            index_slot_mapping, kv_cache, index_cache, eps, rotary_dim,      \
-            num_tokens, nq, nkv, niq, block_size, kv_s_block, kv_s_kv,       \
-            kv_s_token, kv_s_head)
-  // clang-format on
-#endif
-
-  if (has_index) {
-    if (insert_kv) {
-      LAUNCH(true, true);  // sparse serving
-    } else {
-      LAUNCH(true, false);  // sparse profiling
-    }
-  } else {
-    // Dense layer: never has an index branch and never inserts here (the
-    // generic Attention layer owns the KV insert).
-    LAUNCH(false, false);
-  }
-#undef LAUNCH
-}
-
-}  // namespace minimax_m3_fused_ops
-}  // namespace vllm
-
-#define CALL_FUSED_MINIMAX_M3(_RAW_T, CACHE_T, KV_DTYPE)                       \
-  vllm::minimax_m3_fused_ops::launchFusedMiniMaxM3<st, CACHE_T, KV_DTYPE>(     \
-      reinterpret_cast<st*>(qkv.data_ptr()),                                   \
-      q_out.has_value() ? reinterpret_cast<st*>(q_out->data_ptr()) : nullptr,  \
-      index_q_out.has_value() ? reinterpret_cast<st*>(index_q_out->data_ptr()) \
-                              : nullptr,                                       \
-      reinterpret_cast<st const*>(q_norm_weight.data_ptr()),                   \
-      reinterpret_cast<st const*>(k_norm_weight.data_ptr()),                   \
-      has_index ? reinterpret_cast<st const*>(index_q_norm_weight->data_ptr()) \
-                : nullptr,                                                     \
-      has_index ? reinterpret_cast<st const*>(index_k_norm_weight->data_ptr()) \
-                : nullptr,                                                     \
-      reinterpret_cast<st const*>(cos_sin_cache.data_ptr()),                   \
-      reinterpret_cast<int64_t const*>(positions.data_ptr()),                  \
-      insert_kv ? reinterpret_cast<int64_t const*>(slot_mapping->data_ptr())   \
-                : nullptr,                                                     \
-      insert_kv ? reinterpret_cast<int64_t const*>(                            \
-                      effective_index_slot_mapping->data_ptr())                \
-                : nullptr,                                                     \
-      insert_kv ? reinterpret_cast<CACHE_T*>(kv_cache->data_ptr()) : nullptr,  \
-      (insert_kv && has_index)                                                 \
-          ? reinterpret_cast<st*>(index_cache->data_ptr())                     \
-          : nullptr,                                                           \
-      static_cast<float>(eps), static_cast<int>(rotary_dim), num_tokens, nq,   \
-      nkv, niq, static_cast<int>(block_size), kv_s_block, kv_s_kv, kv_s_token, \
-      kv_s_head, has_index, insert_kv, stream)
-
-// ────────────────────────────────────────────────────────────────────────────
-// Torch op wrapper
-// ────────────────────────────────────────────────────────────────────────────
-void fused_minimax_m3_qknorm_rope_kv_insert(
-    torch::stable::Tensor& qkv,  // [N, qkv_row] (packs index if sparse)
-    torch::stable::Tensor const& q_norm_weight,  // [128]
-    torch::stable::Tensor const& k_norm_weight,  // [128]
-    torch::stable::Tensor const& cos_sin_cache,  // [max_pos, rotary_dim]
-    torch::stable::Tensor const& positions,      // [N] i64
-    int64_t num_heads, int64_t num_kv_heads, int64_t rotary_dim, double eps,
-    std::optional<torch::stable::Tensor> index_q_norm_weight,  // [128]
-    std::optional<torch::stable::Tensor> index_k_norm_weight,  // [128]
-    int64_t num_index_heads,                                  // niq; 0 => dense
-    std::optional<torch::stable::Tensor> slot_mapping,        // [N] i64
-    std::optional<torch::stable::Tensor> index_slot_mapping,  // [N] i64
-    std::optional<torch::stable::Tensor> kv_cache,     // [nb,2,bs,nkv,128]
-    std::optional<torch::stable::Tensor> index_cache,  // [nb,bs,128]
-    int64_t block_size,
-    std::optional<torch::stable::Tensor> q_out,  // [N, nq*128] contiguous
-    std::optional<torch::stable::Tensor>
-        index_q_out,  // [N, niq*128] contiguous
-    const std::string& kv_cache_dtype) {
-  STD_TORCH_CHECK(qkv.is_cuda() && qkv.is_contiguous(),
-                  "qkv must be contiguous CUDA");
-  STD_TORCH_CHECK(
-      qkv.scalar_type() == torch::headeronly::ScalarType::Half ||
-          qkv.scalar_type() == torch::headeronly::ScalarType::BFloat16,
-      "qkv must be float16 or bfloat16");
-  STD_TORCH_CHECK(
-      positions.is_cuda() &&
-          positions.scalar_type() == torch::headeronly::ScalarType::Long,
-      "positions must be int64 CUDA");
-  STD_TORCH_CHECK(cos_sin_cache.is_cuda() && cos_sin_cache.is_contiguous(),
-                  "cos_sin_cache must be contiguous CUDA");
-  STD_TORCH_CHECK(cos_sin_cache.scalar_type() == qkv.scalar_type(),
-                  "cos_sin_cache dtype must match qkv");
-  STD_TORCH_CHECK(
-      cos_sin_cache.dim() == 2 && cos_sin_cache.size(1) == rotary_dim,
-      "cos_sin_cache shape [max_pos, rotary_dim]");
-
-  STD_TORCH_CHECK(q_norm_weight.scalar_type() == qkv.scalar_type() &&
-                      k_norm_weight.scalar_type() == qkv.scalar_type(),
-                  "q/k norm weight dtype must match qkv");
-  STD_TORCH_CHECK(
-      q_norm_weight.numel() == vllm::minimax_m3_fused_ops::kHeadDim &&
-          k_norm_weight.numel() == vllm::minimax_m3_fused_ops::kHeadDim,
-      "q/k norm weight must have 128 elements");
-  STD_TORCH_CHECK(rotary_dim > 0 && rotary_dim % 8 == 0 &&
-                      rotary_dim <= vllm::minimax_m3_fused_ops::kHeadDim,
-                  "rotary_dim must be a positive multiple of 8 and <= 128");
-
-  int const num_tokens = static_cast<int>(qkv.size(0));
-  int const nq = static_cast<int>(num_heads);
-  int const nkv = static_cast<int>(num_kv_heads);
-  int const niq = static_cast<int>(num_index_heads);
-
-  // The sparse layer packs the index branch ([index_q (niq heads) | index_k
-  // (1 head)]) right after [q|k|v] in the same row; the dense layer does not.
-  bool const has_index = niq > 0;
-  bool const insert_kv = kv_cache.has_value();
-  vllm::Fp8KVCacheDataType const kv_dt =
-      vllm::get_fp8_kv_cache_data_type(kv_cache_dtype);
-  int const kHeadDim = vllm::minimax_m3_fused_ops::kHeadDim;
-  int const expected_row =
-      (nq + 2 * nkv + (has_index ? niq + 1 : 0)) * kHeadDim;
-  STD_TORCH_CHECK(qkv.size(1) == expected_row,
-                  "qkv last dim must be (num_heads + 2*num_kv_heads"
-                  " + num_index_heads + 1) * 128 for sparse, "
-                  "(num_heads + 2*num_kv_heads) * 128 for dense");
-
-  // Only the sparse layer inserts here (dense lets the generic Attention layer
-  // own the KV write); there is no dense+insert kernel instantiation.
-  STD_TORCH_CHECK(
-      !insert_kv || has_index,
-      "insert mode (kv_cache) requires the index branch (sparse layer)");
-  if (has_index) {
-    STD_TORCH_CHECK(
-        index_q_norm_weight.has_value() && index_k_norm_weight.has_value(),
-        "index branch requires both index norm weights");
-    STD_TORCH_CHECK(index_q_norm_weight->scalar_type() == qkv.scalar_type() &&
-                        index_k_norm_weight->scalar_type() == qkv.scalar_type(),
-                    "index norm weights dtype must match qkv");
-    STD_TORCH_CHECK(index_q_norm_weight->numel() == kHeadDim &&
-                        index_k_norm_weight->numel() == kHeadDim,
-                    "index norm weights must have 128 elements");
-  }
-  // kv_cache strides (logical shape [nb, 2, bs, nkv, head_dim]). Read straight
-  // off the tensor so the kernel honours whatever physical layout the attention
-  // backend allocated (NHD: stride order (0,1,2,3,4); HND: (0,1,3,2,4)). No new
-  // op argument is needed -- the strides ride along with the tensor itself.
-  int64_t kv_s_block = 0, kv_s_kv = 0, kv_s_token = 0, kv_s_head = 0;
-  torch::stable::Tensor const* effective_index_slot_mapping = nullptr;
-  if (insert_kv) {
-    STD_TORCH_CHECK(
-        slot_mapping.has_value() && slot_mapping->is_cuda() &&
-            slot_mapping->scalar_type() == torch::headeronly::ScalarType::Long,
-        "insert mode requires int64 CUDA slot_mapping");
-    STD_TORCH_CHECK(
-        !index_slot_mapping.has_value() ||
-            (index_slot_mapping->is_cuda() &&
-             index_slot_mapping->scalar_type() ==
-                 torch::headeronly::ScalarType::Long &&
-             index_slot_mapping->numel() == slot_mapping->numel()),
-        "index_slot_mapping must be int64 CUDA with slot_mapping length");
-    if (kv_dt == vllm::Fp8KVCacheDataType::kAuto) {
-      STD_TORCH_CHECK(kv_cache->scalar_type() == qkv.scalar_type(),
-                      "auto kv_cache dtype must match qkv");
-    } else {
-      STD_TORCH_CHECK(
-          kv_cache->scalar_type() == torch::headeronly::ScalarType::Byte,
-          "fp8 kv_cache must use uint8 storage");
-    }
-    STD_TORCH_CHECK(index_cache.has_value() &&
-                        index_cache->scalar_type() == qkv.scalar_type(),
-                    "insert mode requires matching index_cache");
-    STD_TORCH_CHECK(kv_cache->dim() == 5 && kv_cache->stride(4) == 1,
-                    "kv_cache must be [nb,2,bs,nkv,head_dim] with contiguous "
-                    "head_dim (stride(4)==1)");
-    kv_s_block = kv_cache->stride(0);
-    kv_s_kv = kv_cache->stride(1);
-    kv_s_token = kv_cache->stride(2);
-    kv_s_head = kv_cache->stride(3);
-    effective_index_slot_mapping = index_slot_mapping.has_value()
-                                       ? &index_slot_mapping.value()
-                                       : &slot_mapping.value();
-  }
-  // Optional contiguous gather targets: when given, the normed/roped q (and
-  // index_q) are written here instead of in place, so callers avoid a separate
-  // .contiguous() copy.  index_q_out only makes sense on the sparse path.
-  if (q_out.has_value()) {
-    STD_TORCH_CHECK(
-        q_out->is_cuda() && q_out->is_contiguous() &&
-            q_out->scalar_type() == qkv.scalar_type(),
-        "q_out must be a contiguous CUDA tensor matching qkv dtype");
-    STD_TORCH_CHECK(
-        q_out->numel() == static_cast<int64_t>(num_tokens) * nq * kHeadDim,
-        "q_out must have num_tokens * num_heads * 128 elements");
-  }
-  if (index_q_out.has_value()) {
-    STD_TORCH_CHECK(
-        has_index,
-        "index_q_out requires the index branch (num_index_heads > 0)");
-    STD_TORCH_CHECK(
-        index_q_out->is_cuda() && index_q_out->is_contiguous() &&
-            index_q_out->scalar_type() == qkv.scalar_type(),
-        "index_q_out must be a contiguous CUDA tensor matching qkv dtype");
-    STD_TORCH_CHECK(index_q_out->numel() ==
-                        static_cast<int64_t>(num_tokens) * niq * kHeadDim,
-                    "index_q_out must have num_tokens * num_index_heads * 128 "
-                    "elements");
-  }
-
-  const torch::stable::accelerator::DeviceGuard device_guard(
-      qkv.get_device_index());
-  auto stream = get_current_cuda_stream(qkv.get_device_index());
-
-  VLLM_STABLE_DISPATCH_HALF_TYPES(
-      qkv.scalar_type(), "fused_minimax_m3_qknorm_rope_kv_insert", [&] {
-        using st = scalar_t;
-        DISPATCH_BY_KV_CACHE_DTYPE(qkv.scalar_type(), kv_cache_dtype,
-                                   CALL_FUSED_MINIMAX_M3);
-      });
-}
-
-#undef CALL_FUSED_MINIMAX_M3

csrc/libtorch_stable/layernorm_kernels.cu

@@ -11,7 +11,7 @@
 namespace vllm {
 
 // TODO(woosuk): Further optimize this kernel.
-template <typename scalar_t, int VEC_SIZE, int NUM_DIMS, bool HasWeight>
+template <typename scalar_t, int VEC_SIZE, int NUM_DIMS>
 __global__ void rms_norm_kernel(
     scalar_t* __restrict__ out,           // [..., hidden_size]
     const scalar_t* __restrict__ input,   // [..., hidden_size]
@@ -20,7 +20,7 @@ __global__ void rms_norm_kernel(
     const int64_t input_stride_d4,        // input.stride(-4)
     const int64_t input_shape_d2,         // input.size(-2)
     const int64_t input_shape_d3,         // input.size(-3)
-    const scalar_t* __restrict__ weight,  // [hidden_size], null if !HasWeight
+    const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
   __shared__ float s_variance;
   float variance = 0.0f;
@@ -74,19 +74,11 @@ __global__ void rms_norm_kernel(
   for (int i = threadIdx.x; i < hidden_size / VEC_SIZE; i += blockDim.x) {
     vec_n_t<scalar_t, VEC_SIZE> dst;
     vec_n_t<scalar_t, VEC_SIZE> src1 = v_in[i];
-    vec_n_t<scalar_t, VEC_SIZE> src2;
-    if constexpr (HasWeight) {
-      src2 = v_w[i];
-    }
+    vec_n_t<scalar_t, VEC_SIZE> src2 = v_w[i];
 #pragma unroll
     for (int j = 0; j < VEC_SIZE; j++) {
       float x = static_cast<float>(src1.val[j]);
-      scalar_t normalized = static_cast<scalar_t>(x * s_variance);
-      if constexpr (HasWeight) {
-        dst.val[j] = normalized * src2.val[j];
-      } else {
-        dst.val[j] = normalized;
-      }
+      dst.val[j] = static_cast<scalar_t>(x * s_variance) * src2.val[j];
     }
     v_out[i] = dst;
   }
@@ -96,13 +88,13 @@ __global__ void rms_norm_kernel(
    Additional optimizations we can make in this case are
    packed and vectorized operations, which help with the
    memory latency bottleneck. */
-template <typename scalar_t, int width, bool HasWeight>
+template <typename scalar_t, int width>
 __global__ std::enable_if_t<(width > 0) && _typeConvert<scalar_t>::exists>
 fused_add_rms_norm_kernel(
     scalar_t* __restrict__ input,  // [..., hidden_size]
     const int64_t input_stride,
     scalar_t* __restrict__ residual,      // [..., hidden_size]
-    const scalar_t* __restrict__ weight,  // [hidden_size], null if !HasWeight
+    const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
   // Sanity checks on our vector struct and type-punned pointer arithmetic
   static_assert(std::is_pod_v<_f16Vec<scalar_t, width>>);
@@ -144,21 +136,13 @@ fused_add_rms_norm_kernel(
     int id = blockIdx.x * vec_hidden_size + idx;
     int64_t strided_id = blockIdx.x * vec_input_stride + idx;
     _f16Vec<scalar_t, width> res = residual_v[id];
+    _f16Vec<scalar_t, width> w = weight_v[idx];
     _f16Vec<scalar_t, width> out;
     using Converter = _typeConvert<scalar_t>;
-    if constexpr (HasWeight) {
-      _f16Vec<scalar_t, width> w = weight_v[idx];
 #pragma unroll
-      for (int j = 0; j < width; ++j) {
-        float x = Converter::convert(res.data[j]);
-        out.data[j] = Converter::convert(x * s_variance) * w.data[j];
-      }
-    } else {
-#pragma unroll
-      for (int j = 0; j < width; ++j) {
-        float x = Converter::convert(res.data[j]);
-        out.data[j] = Converter::convert(x * s_variance);
-      }
+    for (int j = 0; j < width; ++j) {
+      float x = Converter::convert(res.data[j]);
+      out.data[j] = Converter::convert(x * s_variance) * w.data[j];
     }
     input_v[strided_id] = out;
   }
@@ -167,13 +151,13 @@ fused_add_rms_norm_kernel(
 /* Generic fused_add_rms_norm_kernel
    The width field is not used here but necessary for other specializations.
  */
-template <typename scalar_t, int width, bool HasWeight>
+template <typename scalar_t, int width>
 __global__ std::enable_if_t<(width == 0) || !_typeConvert<scalar_t>::exists>
 fused_add_rms_norm_kernel(
     scalar_t* __restrict__ input,  // [..., hidden_size]
     const int64_t input_stride,
     scalar_t* __restrict__ residual,      // [..., hidden_size]
-    const scalar_t* __restrict__ weight,  // [hidden_size], null if !HasWeight
+    const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
   __shared__ float s_variance;
   float variance = 0.0f;
@@ -197,29 +181,23 @@ fused_add_rms_norm_kernel(
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
     float x = (float)residual[blockIdx.x * hidden_size + idx];
-    if constexpr (HasWeight) {
-      input[blockIdx.x * input_stride + idx] =
-          (scalar_t)(x * s_variance) * weight[idx];
-    } else {
-      input[blockIdx.x * input_stride + idx] = (scalar_t)(x * s_variance);
-    }
+    input[blockIdx.x * input_stride + idx] =
+        (scalar_t)(x * s_variance) * weight[idx];
   }
 }
 
 }  // namespace vllm
 
-void rms_norm(torch::stable::Tensor& out,    // [..., hidden_size]
-              torch::stable::Tensor& input,  // [..., hidden_size]
-              std::optional<torch::stable::Tensor> weight,  // [hidden_size]
+void rms_norm(torch::stable::Tensor& out,     // [..., hidden_size]
+              torch::stable::Tensor& input,   // [..., hidden_size]
+              torch::stable::Tensor& weight,  // [hidden_size]
               double epsilon) {
   STD_TORCH_CHECK(out.is_contiguous());
   if (input.stride(-1) != 1) {
     input = torch::stable::contiguous(input);
   }
   STD_TORCH_CHECK(input.stride(-1) == 1);
-  if (weight.has_value()) {
-    STD_TORCH_CHECK(weight->is_contiguous());
-  }
+  STD_TORCH_CHECK(weight.is_contiguous());
 
   int hidden_size = input.size(-1);
 
@@ -237,69 +215,46 @@ void rms_norm(torch::stable::Tensor& out,    // [..., hidden_size]
   const torch::stable::accelerator::DeviceGuard device_guard(
       input.get_device_index());
   const cudaStream_t stream = get_current_cuda_stream();
-  const bool has_weight = weight.has_value();
   VLLM_STABLE_DISPATCH_RANK234(num_dims, [&] {
     VLLM_STABLE_DISPATCH_FLOATING_TYPES(
         input.scalar_type(), "rms_norm_kernel", [&] {
-          const scalar_t* weight_ptr =
-              has_weight ? weight->const_data_ptr<scalar_t>() : nullptr;
           const int calculated_vec_size =
               std::gcd(16 / sizeof(scalar_t), hidden_size);
           const int block_size =
               std::min(hidden_size / calculated_vec_size, max_block_size);
           dim3 block(block_size);
           VLLM_STABLE_DISPATCH_VEC_SIZE(calculated_vec_size, [&] {
-            if (has_weight) {
-              vllm::rms_norm_kernel<scalar_t, vec_size, tensor_rank, true>
-                  <<<grid, block, 0, stream>>>(
-                      out.mutable_data_ptr<scalar_t>(),
-                      input.const_data_ptr<scalar_t>(), input_stride_d2,
-                      input_stride_d3, input_stride_d4, input_shape_d2,
-                      input_shape_d3, weight_ptr, epsilon, num_tokens,
-                      hidden_size);
-            } else {
-              vllm::rms_norm_kernel<scalar_t, vec_size, tensor_rank, false>
-                  <<<grid, block, 0, stream>>>(
-                      out.mutable_data_ptr<scalar_t>(),
-                      input.const_data_ptr<scalar_t>(), input_stride_d2,
-                      input_stride_d3, input_stride_d4, input_shape_d2,
-                      input_shape_d3, weight_ptr, epsilon, num_tokens,
-                      hidden_size);
-            }
+            vllm::rms_norm_kernel<scalar_t, vec_size, tensor_rank>
+                <<<grid, block, 0, stream>>>(
+                    out.mutable_data_ptr<scalar_t>(),
+                    input.const_data_ptr<scalar_t>(), input_stride_d2,
+                    input_stride_d3, input_stride_d4, input_shape_d2,
+                    input_shape_d3, weight.const_data_ptr<scalar_t>(), epsilon,
+                    num_tokens, hidden_size);
           });
         });
   });
 }
 
-#define LAUNCH_FUSED_ADD_RMS_NORM(width, has_weight)                       \
-  VLLM_STABLE_DISPATCH_FLOATING_TYPES(                                     \
-      input.scalar_type(), "fused_add_rms_norm_kernel", [&] {              \
-        if (has_weight) {                                                  \
-          vllm::fused_add_rms_norm_kernel<scalar_t, width, true>           \
-              <<<grid, block, 0, stream>>>(                                \
-                  input.mutable_data_ptr<scalar_t>(), input_stride,        \
-                  residual.mutable_data_ptr<scalar_t>(),                   \
-                  weight->const_data_ptr<scalar_t>(), epsilon, num_tokens, \
-                  hidden_size);                                            \
-        } else {                                                           \
-          vllm::fused_add_rms_norm_kernel<scalar_t, width, false>          \
-              <<<grid, block, 0, stream>>>(                                \
-                  input.mutable_data_ptr<scalar_t>(), input_stride,        \
-                  residual.mutable_data_ptr<scalar_t>(), nullptr, epsilon, \
-                  num_tokens, hidden_size);                                \
-        }                                                                  \
+#define LAUNCH_FUSED_ADD_RMS_NORM(width)                                \
+  VLLM_STABLE_DISPATCH_FLOATING_TYPES(                                  \
+      input.scalar_type(), "fused_add_rms_norm_kernel", [&] {           \
+        vllm::fused_add_rms_norm_kernel<scalar_t, width>                \
+            <<<grid, block, 0, stream>>>(                               \
+                input.mutable_data_ptr<scalar_t>(), input_stride,       \
+                residual.mutable_data_ptr<scalar_t>(),                  \
+                weight.const_data_ptr<scalar_t>(), epsilon, num_tokens, \
+                hidden_size);                                           \
       });
 
 void fused_add_rms_norm(torch::stable::Tensor& input,     // [..., hidden_size]
                         torch::stable::Tensor& residual,  // [..., hidden_size]
-                        std::optional<torch::stable::Tensor> weight,
+                        torch::stable::Tensor& weight,    // [hidden_size]
                         double epsilon) {
+  STD_TORCH_CHECK(weight.scalar_type() == input.scalar_type());
   STD_TORCH_CHECK(input.scalar_type() == residual.scalar_type());
   STD_TORCH_CHECK(residual.is_contiguous());
-  if (weight.has_value()) {
-    STD_TORCH_CHECK(weight->scalar_type() == input.scalar_type());
-    STD_TORCH_CHECK(weight->is_contiguous());
-  }
+  STD_TORCH_CHECK(weight.is_contiguous());
   int hidden_size = input.size(-1);
   int64_t input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
@@ -314,33 +269,30 @@ void fused_add_rms_norm(torch::stable::Tensor& input,     // [..., hidden_size]
   const torch::stable::accelerator::DeviceGuard device_guard(
       input.get_device_index());
   const cudaStream_t stream = get_current_cuda_stream();
-  constexpr int vector_width = 8;
-  constexpr int req_alignment_bytes = vector_width * 2;
+  /*If the tensor types are FP16/BF16, try to use the optimized kernel
+    with packed + vectorized ops.
+    Max optimization is achieved with a width-8 vector of FP16/BF16s
+    since we can load at most 128 bits at once in a global memory op.
+    However, this requires each tensor's data to be aligned to 16
+    bytes.
+   */
   auto inp_ptr = reinterpret_cast<std::uintptr_t>(input.data_ptr());
   auto res_ptr = reinterpret_cast<std::uintptr_t>(residual.data_ptr());
+  auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight.data_ptr());
+  constexpr int vector_width = 8;
+  constexpr int req_alignment_bytes =
+      vector_width * 2;  // vector_width * sizeof(bfloat16 or float16) (float32
+                         // falls back to non-vectorized version anyway)
+  bool ptrs_are_aligned = inp_ptr % req_alignment_bytes == 0 &&
+                          res_ptr % req_alignment_bytes == 0 &&
+                          wt_ptr % req_alignment_bytes == 0;
   bool offsets_are_multiple_of_vector_width =
       hidden_size % vector_width == 0 && input_stride % vector_width == 0;
   bool batch_invariant_launch = vllm::vllm_is_batch_invariant();
-  const bool has_weight = weight.has_value();
-  if (has_weight) {
-    auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight->data_ptr());
-    bool ptrs_are_aligned = inp_ptr % req_alignment_bytes == 0 &&
-                            res_ptr % req_alignment_bytes == 0 &&
-                            wt_ptr % req_alignment_bytes == 0;
-    if (ptrs_are_aligned && offsets_are_multiple_of_vector_width &&
-        !batch_invariant_launch) {
-      LAUNCH_FUSED_ADD_RMS_NORM(8, true);
-    } else {
-      LAUNCH_FUSED_ADD_RMS_NORM(0, true);
-    }
+  if (ptrs_are_aligned && offsets_are_multiple_of_vector_width &&
+      !batch_invariant_launch) {
+    LAUNCH_FUSED_ADD_RMS_NORM(8);
   } else {
-    bool ptrs_are_aligned = inp_ptr % req_alignment_bytes == 0 &&
-                            res_ptr % req_alignment_bytes == 0;
-    if (ptrs_are_aligned && offsets_are_multiple_of_vector_width &&
-        !batch_invariant_launch) {
-      LAUNCH_FUSED_ADD_RMS_NORM(8, false);
-    } else {
-      LAUNCH_FUSED_ADD_RMS_NORM(0, false);
-    }
+    LAUNCH_FUSED_ADD_RMS_NORM(0);
   }
 }

csrc/libtorch_stable/moe/moe_ops.h

@@ -1,87 +0,0 @@
-#pragma once
-
-#include <torch/csrc/stable/tensor.h>
-
-#include <optional>
-#include <tuple>
-
-void topk_softmax(torch::stable::Tensor& topk_weights,
-                  torch::stable::Tensor& topk_indices,
-                  torch::stable::Tensor& token_expert_indices,
-                  torch::stable::Tensor& gating_output, bool renormalize,
-                  std::optional<torch::stable::Tensor> bias);
-
-void topk_sigmoid(torch::stable::Tensor& topk_weights,
-                  torch::stable::Tensor& topk_indices,
-                  torch::stable::Tensor& token_expert_indices,
-                  torch::stable::Tensor& gating_output, bool renormalize,
-                  std::optional<torch::stable::Tensor> bias);
-
-void topk_softplus_sqrt(
-    torch::stable::Tensor& topk_weights, torch::stable::Tensor& topk_indices,
-    torch::stable::Tensor& token_expert_indices,
-    torch::stable::Tensor& gating_output, bool renormalize,
-    double routed_scaling_factor,
-    const std::optional<torch::stable::Tensor>& correction_bias,
-    const std::optional<torch::stable::Tensor>& input_ids,
-    const std::optional<torch::stable::Tensor>& tid2eid);
-
-void moe_sum(torch::stable::Tensor& input, torch::stable::Tensor& output);
-
-void moe_align_block_size(
-    torch::stable::Tensor topk_ids, int64_t num_experts, int64_t block_size,
-    torch::stable::Tensor sorted_token_ids, torch::stable::Tensor experts_ids,
-    torch::stable::Tensor num_tokens_post_pad,
-    std::optional<torch::stable::Tensor> maybe_expert_map);
-
-void batched_moe_align_block_size(
-    int64_t max_tokens_per_batch, int64_t block_size,
-    const torch::stable::Tensor& expert_num_tokens,
-    torch::stable::Tensor sorted_ids, torch::stable::Tensor expert_ids,
-    torch::stable::Tensor num_tokens_post_pad);
-
-void moe_lora_align_block_size(
-    torch::stable::Tensor topk_ids, torch::stable::Tensor token_lora_mapping,
-    int64_t num_experts, int64_t block_size, int64_t max_loras,
-    int64_t max_num_tokens_padded, int64_t max_num_m_blocks,
-    torch::stable::Tensor sorted_token_ids, torch::stable::Tensor expert_ids,
-    torch::stable::Tensor num_tokens_post_pad,
-    torch::stable::Tensor adapter_enabled, torch::stable::Tensor lora_ids,
-    std::optional<torch::stable::Tensor> maybe_expert_map);
-#ifndef USE_ROCM
-torch::stable::Tensor moe_wna16_gemm(
-    torch::stable::Tensor input, torch::stable::Tensor output,
-    torch::stable::Tensor b_qweight, torch::stable::Tensor b_scales,
-    std::optional<torch::stable::Tensor> b_qzeros,
-    std::optional<torch::stable::Tensor> topk_weights,
-    torch::stable::Tensor sorted_token_ids, torch::stable::Tensor expert_ids,
-    torch::stable::Tensor num_tokens_post_pad, int64_t top_k,
-    int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N, int64_t BLOCK_SIZE_K,
-    int64_t bit);
-
-std::tuple<torch::stable::Tensor, torch::stable::Tensor> grouped_topk(
-    const torch::stable::Tensor& scores, int64_t n_group, int64_t topk_group,
-    int64_t topk, bool renormalize, double routed_scaling_factor,
-    const torch::stable::Tensor& bias, int64_t scoring_func);
-#endif
-
-bool moe_permute_unpermute_supported();
-
-int64_t moe_permute_sort_workspace_size(int64_t num_expanded_rows,
-                                        int64_t num_expert);
-
-void shuffle_rows(const torch::stable::Tensor& input_tensor,
-                  const torch::stable::Tensor& dst2src_map,
-                  torch::stable::Tensor& output_tensor);
-
-#ifndef USE_ROCM
-// DeepSeek V3 optimized router GEMM kernel for SM90+
-// Computes output = mat_a @ mat_b.T where:
-//   mat_a: [num_tokens, hidden_dim] in bf16
-//   mat_b: [num_experts, hidden_dim] in bf16
-//   output: [num_tokens, num_experts] in bf16 or fp32
-// Supports num_tokens in [1, 16], num_experts in {256, 384}, hidden_dim = 7168
-void dsv3_router_gemm(torch::stable::Tensor& output,
-                      const torch::stable::Tensor& mat_a,
-                      const torch::stable::Tensor& mat_b);
-#endif

csrc/libtorch_stable/moe/moe_permute_unpermute_op.cu

@@ -1,319 +0,0 @@
-#include <cuda.h>
-#include <cuda_runtime.h>
-#include <torch/csrc/stable/accelerator.h>
-#include <torch/csrc/stable/ops.h>
-#include <torch/csrc/stable/tensor.h>
-#include <torch/headeronly/core/ScalarType.h>
-#include <torch/headeronly/util/Exception.h>
-
-#include "core/registration.h"
-#include "libtorch_stable/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.h"
-#include "libtorch_stable/torch_utils.h"
-
-#include <torch/csrc/stable/library.h>
-
-// moe_permute kernels require at least CUDA 12.0
-#if defined(CUDA_VERSION) && (CUDA_VERSION >= 12000)
-
-namespace {
-
-int64_t product_integers(torch::headeronly::IntHeaderOnlyArrayRef sizes) {
-  int64_t numel = 1;
-  for (int64_t s : sizes) {
-    numel *= s;
-  }
-  return numel;
-}
-
-torch::stable::Tensor maybe_allocate_tensor(
-    const std::optional<torch::stable::Tensor>& maybe_tensor,
-    torch::headeronly::IntHeaderOnlyArrayRef expected_sizes,
-    torch::headeronly::ScalarType dtype, torch::stable::Device device,
-    char const* name) {
-  auto expected_numel = product_integers(expected_sizes);
-  if (maybe_tensor.has_value()) {
-    auto tensor = maybe_tensor.value();
-    STD_TORCH_CHECK(tensor.device() == device, name,
-                    " must be on the same device");
-    STD_TORCH_CHECK(tensor.scalar_type() == dtype, name,
-                    " has incorrect dtype");
-    STD_TORCH_CHECK(tensor.is_contiguous(), name, " must be contiguous");
-    STD_TORCH_CHECK(tensor.numel() >= expected_numel, name,
-                    " is too small for the requested shape");
-    auto flat_tensor = torch::stable::view(tensor, {tensor.numel()});
-    return torch::stable::view(
-        torch::stable::narrow(flat_tensor, 0, 0, expected_numel),
-        expected_sizes);
-  }
-  return torch::stable::empty(expected_sizes, dtype, std::nullopt, device);
-}
-
-}  // namespace
-
-int64_t moe_permute_sort_workspace_size(int64_t num_expanded_rows,
-                                        int64_t n_expert) {
-  return static_cast<int64_t>(
-      CubKeyValueSorter::getWorkspaceSize(num_expanded_rows, n_expert));
-}
-
-void moe_permute_impl(
-    const torch::stable::Tensor& input,                 // [n_token, hidden]
-    const torch::stable::Tensor& topk_ids,              // [n_token, topk]
-    const torch::stable::Tensor& token_expert_indices,  // [n_token, topk]
-    const std::optional<torch::stable::Tensor>& expert_map,  // [n_expert]
-    int64_t n_expert, int64_t n_local_expert, int64_t topk,
-    torch::stable::Tensor& permuted_input,  // [permuted_size, hidden]
-    torch::stable::Tensor& expert_first_token_offset,  // [n_local_expert + 1]
-    torch::stable::Tensor& inv_permuted_idx,           // [n_token, topk]
-    torch::stable::Tensor& permuted_idx,               // [permute_size]
-    const std::optional<torch::stable::Tensor>& maybe_sort_workspace,
-    const std::optional<torch::stable::Tensor>& maybe_permuted_experts_id,
-    const std::optional<torch::stable::Tensor>& maybe_sorted_row_idx,
-    const std::optional<torch::stable::Tensor>& maybe_topk_ids_for_sort) {
-  STD_TORCH_CHECK(expert_first_token_offset.scalar_type() ==
-                      torch::headeronly::ScalarType::Long,
-                  "expert_first_token_offset must be int64");
-  STD_TORCH_CHECK(topk_ids.scalar_type() == torch::headeronly::ScalarType::Int,
-                  "topk_ids must be int32");
-  STD_TORCH_CHECK(
-      token_expert_indices.scalar_type() == torch::headeronly::ScalarType::Int,
-      "token_expert_indices must be int32");
-  STD_TORCH_CHECK(
-      inv_permuted_idx.scalar_type() == torch::headeronly::ScalarType::Int,
-      "inv_permuted_idx must be int32");
-  STD_TORCH_CHECK(expert_first_token_offset.size(0) == n_local_expert + 1,
-                  "expert_first_token_offset shape != n_local_expert+1");
-  STD_TORCH_CHECK(
-      inv_permuted_idx.sizes().equals(token_expert_indices.sizes()),
-      "token_expert_indices shape must be same as inv_permuted_idx");
-
-  auto device = input.device();
-  auto n_token = input.sizes()[0];
-  auto n_hidden = input.sizes()[1];
-  auto expanded_rows = n_token * topk;
-  auto stream = get_current_cuda_stream(input.get_device_index());
-
-  auto sorter_size = moe_permute_sort_workspace_size(expanded_rows, n_expert);
-  auto sort_workspace = maybe_allocate_tensor(
-      maybe_sort_workspace, {sorter_size}, torch::headeronly::ScalarType::Char,
-      device, "sort_workspace");
-  auto permuted_experts_id = maybe_allocate_tensor(
-      maybe_permuted_experts_id, topk_ids.sizes(),
-      torch::headeronly::ScalarType::Int, device, "permuted_experts_id");
-  auto sorted_row_idx = maybe_allocate_tensor(
-      maybe_sorted_row_idx, inv_permuted_idx.sizes(),
-      torch::headeronly::ScalarType::Int, device, "sorted_row_idx");
-
-  CubKeyValueSorter sorter{};
-  int64_t* valid_num_ptr = nullptr;
-  torch::stable::Tensor topk_ids_for_sort = topk_ids;
-
-  if (expert_map.has_value()) {
-    const int* expert_map_ptr = get_ptr<int>(expert_map.value());
-    valid_num_ptr =
-        get_ptr<int64_t>(expert_first_token_offset) + n_local_expert;
-    topk_ids_for_sort = maybe_allocate_tensor(
-        maybe_topk_ids_for_sort, topk_ids.sizes(),
-        torch::headeronly::ScalarType::Int, device, "topk_ids_for_sort");
-    torch::stable::copy_(topk_ids_for_sort, topk_ids);
-    preprocessTopkIdLauncher(get_ptr<int>(topk_ids_for_sort), n_token * topk,
-                             expert_map_ptr, n_expert, stream);
-  }
-
-  sortAndScanExpert(
-      get_ptr<const int>(topk_ids_for_sort), get_ptr<int>(token_expert_indices),
-      get_ptr<int>(permuted_experts_id), get_ptr<int>(sorted_row_idx),
-      get_ptr<int64_t>(expert_first_token_offset), n_token, n_expert,
-      n_local_expert, topk, sorter, get_ptr<int>(sort_workspace), stream);
-
-  MOE_DISPATCH(input.scalar_type(), [&] {
-    expandInputRowsKernelLauncher<scalar_t>(
-        get_ptr<scalar_t>(input), get_ptr<scalar_t>(permuted_input),
-        get_ptr<int>(sorted_row_idx), get_ptr<int>(inv_permuted_idx),
-        get_ptr<int>(permuted_idx), get_ptr<int64_t>(expert_first_token_offset),
-        n_token, valid_num_ptr, n_hidden, topk, n_local_expert, stream);
-  });
-}
-
-void moe_permute(
-    const torch::stable::Tensor& input,                 // [n_token, hidden]
-    const torch::stable::Tensor& topk_ids,              // [n_token, topk]
-    const torch::stable::Tensor& token_expert_indices,  // [n_token, topk]
-    const std::optional<torch::stable::Tensor>& expert_map,  // [n_expert]
-    int64_t n_expert, int64_t n_local_expert, int64_t topk,
-    torch::stable::Tensor& permuted_input,  // [permuted_size, hidden]
-    torch::stable::Tensor& expert_first_token_offset,  // [n_local_expert + 1]
-    torch::stable::Tensor& inv_permuted_idx,           // [n_token, topk]
-    torch::stable::Tensor& permuted_idx) {             // [permute_size]
-  moe_permute_impl(input, topk_ids, token_expert_indices, expert_map, n_expert,
-                   n_local_expert, topk, permuted_input,
-                   expert_first_token_offset, inv_permuted_idx, permuted_idx,
-                   std::nullopt, std::nullopt, std::nullopt, std::nullopt);
-}
-
-void moe_permute_with_scratch(
-    const torch::stable::Tensor& input, const torch::stable::Tensor& topk_ids,
-    const torch::stable::Tensor& token_expert_indices,
-    const std::optional<torch::stable::Tensor>& expert_map, int64_t n_expert,
-    int64_t n_local_expert, int64_t topk, torch::stable::Tensor& permuted_input,
-    torch::stable::Tensor& expert_first_token_offset,
-    torch::stable::Tensor& inv_permuted_idx,
-    torch::stable::Tensor& permuted_idx, torch::stable::Tensor& sort_workspace,
-    torch::stable::Tensor& permuted_experts_id,
-    torch::stable::Tensor& sorted_row_idx,
-    torch::stable::Tensor& topk_ids_for_sort) {
-  moe_permute_impl(input, topk_ids, token_expert_indices, expert_map, n_expert,
-                   n_local_expert, topk, permuted_input,
-                   expert_first_token_offset, inv_permuted_idx, permuted_idx,
-                   sort_workspace, permuted_experts_id, sorted_row_idx,
-                   topk_ids_for_sort);
-}
-
-void moe_unpermute(
-    const torch::stable::Tensor&
-        permuted_hidden_states,                     // [n_token * topk, hidden]
-    const torch::stable::Tensor& topk_weights,      // [n_token, topk]
-    const torch::stable::Tensor& inv_permuted_idx,  // [n_token, topk]
-    const std::optional<torch::stable::Tensor>&
-        expert_first_token_offset,  // [n_local_expert+1]
-    int64_t topk,
-    torch::stable::Tensor& hidden_states) {  // [n_token, hidden]
-  STD_TORCH_CHECK(
-      permuted_hidden_states.scalar_type() == hidden_states.scalar_type(),
-      "permuted_hidden_states dtype must be same as hidden_states");
-
-  auto n_token = hidden_states.size(0);
-  auto n_hidden = hidden_states.size(1);
-  auto stream = get_current_cuda_stream(hidden_states.get_device_index());
-
-  int64_t const* valid_ptr = nullptr;
-  if (expert_first_token_offset.has_value()) {
-    int n_local_expert = expert_first_token_offset.value().size(0) - 1;
-    valid_ptr =
-        get_ptr<int64_t>(expert_first_token_offset.value()) + n_local_expert;
-  }
-
-  MOE_DISPATCH(hidden_states.scalar_type(), [&] {
-    finalizeMoeRoutingKernelLauncher<scalar_t, scalar_t>(
-        get_ptr<scalar_t>(permuted_hidden_states),
-        get_ptr<scalar_t>(hidden_states), get_ptr<float>(topk_weights),
-        get_ptr<int>(inv_permuted_idx), n_token, n_hidden, topk, valid_ptr,
-        stream);
-  });
-}
-
-template <typename T>
-__global__ void shuffleInputRowsKernel(const T* input,
-                                       const int32_t* dst2src_map, T* output,
-                                       int64_t num_src_rows,
-                                       int64_t num_dst_rows, int64_t num_cols) {
-  int64_t dest_row_idx = blockIdx.x;
-  int64_t const source_row_idx = dst2src_map[dest_row_idx];
-
-  if (blockIdx.x < num_dst_rows) {
-    // Load 128-bits per thread
-    constexpr int64_t ELEM_PER_THREAD = 128 / sizeof(T) / 8;
-    using DataElem = cutlass::Array<T, ELEM_PER_THREAD>;
-
-    // Duplicate and permute rows
-    auto const* source_row_ptr =
-        reinterpret_cast<DataElem const*>(input + source_row_idx * num_cols);
-    auto* dest_row_ptr =
-        reinterpret_cast<DataElem*>(output + dest_row_idx * num_cols);
-
-    int64_t const start_offset = threadIdx.x;
-    int64_t const stride = blockDim.x;
-    int64_t const num_elems_in_col = num_cols / ELEM_PER_THREAD;
-
-    for (int elem_index = start_offset; elem_index < num_elems_in_col;
-         elem_index += stride) {
-      dest_row_ptr[elem_index] = source_row_ptr[elem_index];
-    }
-  }
-}
-
-void shuffle_rows(const torch::stable::Tensor& input_tensor,
-                  const torch::stable::Tensor& dst2src_map,
-                  torch::stable::Tensor& output_tensor) {
-  STD_TORCH_CHECK(input_tensor.scalar_type() == output_tensor.scalar_type(),
-                  "Input and output tensors must have the same data type");
-
-  auto stream = get_current_cuda_stream(output_tensor.get_device_index());
-  const int64_t blocks = output_tensor.size(0);
-  const int64_t threads = 256;
-  const int64_t num_dest_rows = output_tensor.size(0);
-  const int64_t num_src_rows = input_tensor.size(0);
-  const int64_t num_cols = input_tensor.size(1);
-
-  STD_TORCH_CHECK(!(num_cols % (128 / input_tensor.element_size() / 8)),
-                  "num_cols must be divisible by 128 / "
-                  "input_tensor.element_size() / 8");
-
-  MOE_DISPATCH(input_tensor.scalar_type(), [&] {
-    shuffleInputRowsKernel<scalar_t><<<blocks, threads, 0, stream>>>(
-        reinterpret_cast<const scalar_t*>(input_tensor.const_data_ptr()),
-        reinterpret_cast<const int32_t*>(dst2src_map.const_data_ptr()),
-        reinterpret_cast<scalar_t*>(output_tensor.mutable_data_ptr()),
-        num_src_rows, num_dest_rows, num_cols);
-  });
-}
-
-#else
-
-int64_t moe_permute_sort_workspace_size(int64_t num_expanded_rows,
-                                        int64_t n_expert) {
-  STD_TORCH_CHECK(
-      false, "moe_permute_sort_workspace_size is not supported on CUDA < 12.0");
-}
-
-void moe_permute(const torch::stable::Tensor& input,
-                 const torch::stable::Tensor& topk_ids,
-                 const torch::stable::Tensor& token_expert_indices,
-                 const std::optional<torch::stable::Tensor>& expert_map,
-                 int64_t n_expert, int64_t n_local_expert, int64_t topk,
-                 torch::stable::Tensor& permuted_input,
-                 torch::stable::Tensor& expert_first_token_offset,
-                 torch::stable::Tensor& inv_permuted_idx,
-                 torch::stable::Tensor& permuted_idx) {
-  STD_TORCH_CHECK(false, "moe_permute is not supported on CUDA < 12.0");
-}
-
-void moe_permute_with_scratch(
-    const torch::stable::Tensor& input, const torch::stable::Tensor& topk_ids,
-    const torch::stable::Tensor& token_expert_indices,
-    const std::optional<torch::stable::Tensor>& expert_map, int64_t n_expert,
-    int64_t n_local_expert, int64_t topk, torch::stable::Tensor& permuted_input,
-    torch::stable::Tensor& expert_first_token_offset,
-    torch::stable::Tensor& inv_permuted_idx,
-    torch::stable::Tensor& permuted_idx, torch::stable::Tensor& sort_workspace,
-    torch::stable::Tensor& permuted_experts_id,
-    torch::stable::Tensor& sorted_row_idx,
-    torch::stable::Tensor& topk_ids_for_sort) {
-  STD_TORCH_CHECK(false,
-                  "moe_permute_with_scratch is not supported on CUDA < 12.0");
-}
-
-void moe_unpermute(
-    const torch::stable::Tensor& permuted_hidden_states,
-    const torch::stable::Tensor& topk_weights,
-    const torch::stable::Tensor& inv_permuted_idx,
-    const std::optional<torch::stable::Tensor>& expert_first_token_offset,
-    int64_t topk, torch::stable::Tensor& hidden_states) {
-  STD_TORCH_CHECK(false, "moe_unpermute is not supported on CUDA < 12.0");
-}
-
-#endif
-
-bool moe_permute_unpermute_supported() {
-#if defined(CUDA_VERSION) && (CUDA_VERSION >= 12000)
-  return true;
-#else
-  return false;
-#endif
-}
-
-STABLE_TORCH_LIBRARY_IMPL(_moe_C, CUDA, m) {
-  m.impl("moe_permute", TORCH_BOX(&moe_permute));
-  m.impl("moe_permute_with_scratch", TORCH_BOX(&moe_permute_with_scratch));
-  m.impl("moe_unpermute", TORCH_BOX(&moe_unpermute));
-}

csrc/libtorch_stable/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm.cu

@@ -0,0 +1,69 @@
+// 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/csrc/stable/library.h>
+#include <torch/csrc/stable/tensor.h>
+#include "libtorch_stable/torch_utils.h"
+
+#include "cutlass_mxfp8_grouped_mm_launcher.cuh"
+
+void cutlass_mxfp8_grouped_mm(const torch::stable::Tensor& a,
+                              const torch::stable::Tensor& b,
+                              const torch::stable::Tensor& sfa,
+                              const torch::stable::Tensor& sfb,
+                              torch::stable::Tensor& d,
+                              const torch::stable::Tensor& problem_sizes,
+                              const torch::stable::Tensor& expert_offsets,
+                              const torch::stable::Tensor& blockscale_offsets) {
+#if defined(CUTLASS_ARCH_MMA_SM100_SUPPORTED)
+  STD_TORCH_CHECK(problem_sizes.dim() == 2, "problem_sizes must be 2D tensor");
+  STD_TORCH_CHECK(problem_sizes.size(1) == 3,
+                  "problem_sizes must have shape (num_experts, 3)");
+  STD_TORCH_CHECK(
+      problem_sizes.size(0) == expert_offsets.size(0),
+      "Number of experts in problem_sizes must match expert_offsets");
+  STD_TORCH_CHECK(
+      problem_sizes.scalar_type() == torch::headeronly::ScalarType::Int,
+      "problem_sizes must be int32");
+  STD_TORCH_CHECK(
+      expert_offsets.scalar_type() == torch::headeronly::ScalarType::Int,
+      "expert_offsets must be int32");
+  STD_TORCH_CHECK(
+      blockscale_offsets.scalar_type() == torch::headeronly::ScalarType::Int,
+      "blockscale_offsets must be int32");
+  STD_TORCH_CHECK(a.dim() == 2,
+                  "a must be a 2D tensor of shape (num_tokens, k)");
+  STD_TORCH_CHECK(b.dim() == 3,
+                  "b must be a 3D tensor of shape (num_experts, k, n)");
+  STD_TORCH_CHECK(a.size(1) == b.size(1) && a.size(1) % 128 == 0,
+                  "k should align 128");
+  STD_TORCH_CHECK(b.size(2) % 128 == 0, "n should align 128");
+  STD_TORCH_CHECK(a.stride(1) == 1, "a must be row major");
+  STD_TORCH_CHECK(b.stride(1) == 1, "b must be column major");
+
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      a.get_device_index());
+  auto stream = get_current_cuda_stream(a.get_device_index());
+  if (d.scalar_type() == torch::headeronly::ScalarType::BFloat16) {
+    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.scalar_type() == torch::headeronly::ScalarType::Half) {
+    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 {
+    STD_TORCH_CHECK(false, "dtype must be kFloat16 or kBFloat16");
+  }
+#else
+  STD_TORCH_CHECK(false,
+                  "No implemented cutlass_mxfp8_grouped_mm for "
+                  "current device");
+#endif
+}
+
+STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, m) {
+  m.impl("cutlass_mxfp8_grouped_mm", TORCH_BOX(&cutlass_mxfp8_grouped_mm));
+}

csrc/libtorch_stable/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/libtorch_stable/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm_launcher.cuh

@@ -0,0 +1,198 @@
+// 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 <torch/csrc/stable/tensor.h>
+#include <torch/headeronly/util/Exception.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"
+#include "libtorch_stable/torch_utils.h"
+
+namespace expert_specialization {
+
+template <typename GemmTraits>
+void cutlass_mxfp8_grouped_mm_pre_compute(
+    torch::stable::Tensor& a_ptrs, torch::stable::Tensor& b_ptrs,
+    torch::stable::Tensor& sfa_ptrs, torch::stable::Tensor& sfb_ptrs,
+    torch::stable::Tensor& d_ptrs, torch::stable::Tensor& stride_a,
+    torch::stable::Tensor& stride_b, torch::stable::Tensor& stride_d,
+    torch::stable::Tensor& layout_sfa, torch::stable::Tensor& layout_sfb,
+    const torch::stable::Tensor& a, const torch::stable::Tensor& b,
+    const torch::stable::Tensor& sfa, const torch::stable::Tensor& sfb,
+    const torch::stable::Tensor& d, const torch::stable::Tensor& problem_sizes,
+    const torch::stable::Tensor& expert_offsets,
+    const torch::stable::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 = static_cast<int>(expert_offsets.size(0));
+  STD_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::stable::Tensor& a_ptrs,
+                              const torch::stable::Tensor& b_ptrs,
+                              const torch::stable::Tensor& sfa_ptrs,
+                              const torch::stable::Tensor& sfb_ptrs,
+                              const torch::stable::Tensor& d_ptrs,
+                              const torch::stable::Tensor& stride_a,
+                              const torch::stable::Tensor& stride_b,
+                              const torch::stable::Tensor& stride_d,
+                              const torch::stable::Tensor& layout_sfa,
+                              const torch::stable::Tensor& layout_sfb,
+                              const torch::stable::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 = d_ptrs.get_device_index();
+  hw_info.sm_count = get_device_prop()->multiProcessorCount;
+  hw_info.cluster_shape = GemmTraits::MMAConfig::preferred_cluster;
+  hw_info.cluster_shape_fallback = GemmTraits::MMAConfig::fallback_cluster;
+
+  int num_experts = static_cast<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);
+  STD_TORCH_CHECK(can_implement_status == cutlass::Status::kSuccess,
+                  "Failed to implement GEMM");
+
+  size_t workspace_size = gemm.get_workspace_size(arguments);
+  torch::stable::Tensor workspace = torch::stable::empty(
+      {static_cast<int64_t>(workspace_size)},
+      torch::headeronly::ScalarType::Byte, std::nullopt, d_ptrs.device());
+
+  auto status = gemm.initialize(arguments, workspace.data_ptr(), stream);
+  STD_TORCH_CHECK(status == cutlass::Status::kSuccess,
+                  "Failed to initialize GEMM");
+
+  status = gemm.run(stream, nullptr, true);  // Enable PDL
+  STD_TORCH_CHECK(status == cutlass::Status::kSuccess, "Failed to run GEMM");
+}
+
+template <typename OutType>
+void cutlass_mxfp8_grouped_mm_dispatch_out_dtype(
+    const torch::stable::Tensor& a, const torch::stable::Tensor& b,
+    const torch::stable::Tensor& sfa, const torch::stable::Tensor& sfb,
+    torch::stable::Tensor& d, const torch::stable::Tensor& problem_sizes,
+    const torch::stable::Tensor& expert_offsets,
+    const torch::stable::Tensor& blockscale_offsets, cudaStream_t stream) {
+  int num_experts = static_cast<int>(problem_sizes.size(0));
+  auto device = a.device();
+
+  torch::stable::Tensor a_ptrs = torch::stable::empty(
+      num_experts, torch::headeronly::ScalarType::Long, std::nullopt, device);
+  torch::stable::Tensor b_ptrs = torch::stable::empty(
+      num_experts, torch::headeronly::ScalarType::Long, std::nullopt, device);
+  torch::stable::Tensor sfa_ptrs = torch::stable::empty(
+      num_experts, torch::headeronly::ScalarType::Long, std::nullopt, device);
+  torch::stable::Tensor sfb_ptrs = torch::stable::empty(
+      num_experts, torch::headeronly::ScalarType::Long, std::nullopt, device);
+  torch::stable::Tensor d_ptrs = torch::stable::empty(
+      num_experts, torch::headeronly::ScalarType::Long, std::nullopt, device);
+
+  torch::stable::Tensor stride_a = torch::stable::empty(
+      num_experts, torch::headeronly::ScalarType::Long, std::nullopt, device);
+  torch::stable::Tensor stride_b = torch::stable::empty(
+      num_experts, torch::headeronly::ScalarType::Long, std::nullopt, device);
+  torch::stable::Tensor stride_d = torch::stable::empty(
+      num_experts, torch::headeronly::ScalarType::Long, std::nullopt, device);
+  torch::stable::Tensor layout_sfa =
+      torch::stable::empty({num_experts, 5}, torch::headeronly::ScalarType::Int,
+                           std::nullopt, device);
+  torch::stable::Tensor layout_sfb =
+      torch::stable::empty({num_experts, 5}, torch::headeronly::ScalarType::Int,
+                           std::nullopt, device);
+
+  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/libtorch_stable/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/libtorch_stable/moe/mxfp8_moe/mxfp8_experts_quant.cu

@@ -0,0 +1,66 @@
+// 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/csrc/stable/library.h>
+#include <torch/csrc/stable/tensor.h>
+#include "libtorch_stable/torch_utils.h"
+
+#include "mxfp8_experts_quant.cuh"
+
+void mxfp8_experts_quant(const torch::stable::Tensor& input,
+                         const torch::stable::Tensor& problem_sizes,
+                         const torch::stable::Tensor& expert_offsets,
+                         const torch::stable::Tensor& blockscale_offsets,
+                         torch::stable::Tensor& quant_output,
+                         torch::stable::Tensor& scale_factor) {
+#if defined(CUTLASS_ARCH_MMA_SM100_SUPPORTED)
+  STD_TORCH_CHECK(input.dim() == 2, "input must be 2D tensor");
+  STD_TORCH_CHECK(input.size(1) % 128 == 0, "k must align to 128");
+  STD_TORCH_CHECK(input.stride(1) == 1, "input must be row major");
+  STD_TORCH_CHECK(problem_sizes.dim() == 2, "problem_sizes must be 2D tensor");
+  STD_TORCH_CHECK(
+      problem_sizes.scalar_type() == torch::headeronly::ScalarType::Int,
+      "problem_sizes must be int32");
+  STD_TORCH_CHECK(
+      expert_offsets.scalar_type() == torch::headeronly::ScalarType::Int,
+      "expert_offsets must be int32");
+  STD_TORCH_CHECK(
+      blockscale_offsets.scalar_type() == torch::headeronly::ScalarType::Int,
+      "blockscale_offsets must be int32");
+
+  auto groups = problem_sizes.size(0);
+  STD_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");
+  STD_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");
+
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      input.get_device_index());
+  if (input.scalar_type() == torch::headeronly::ScalarType::BFloat16) {
+    expert_specialization::launch_mxfp8_experts_quant<__nv_bfloat16>(
+        input, problem_sizes, expert_offsets, blockscale_offsets, quant_output,
+        scale_factor);
+  } else if (input.scalar_type() == torch::headeronly::ScalarType::Half) {
+    expert_specialization::launch_mxfp8_experts_quant<__half>(
+        input, problem_sizes, expert_offsets, blockscale_offsets, quant_output,
+        scale_factor);
+  } else {
+    STD_TORCH_CHECK(false, "dtype must be kFloat16 or kBFloat16");
+  }
+#else
+  STD_TORCH_CHECK(false,
+                  "No implemented mxfp8_experts_quant for "
+                  "current device");
+#endif
+}
+
+// Registered here (not torch_bindings.cpp) because ENABLE_ES_MXFP8_GROUPED_MM
+// is applied only under COMPILE_LANGUAGE:CUDA.
+STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, m) {
+  m.impl("mxfp8_experts_quant", TORCH_BOX(&mxfp8_experts_quant));
+}

csrc/libtorch_stable/moe/mxfp8_moe/mxfp8_experts_quant.cuh

@@ -0,0 +1,416 @@
+// 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 <cuda.h>
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+
+#include <torch/csrc/inductor/aoti_torch/c/shim.h>
+#include <torch/csrc/stable/macros.h>
+#include <torch/csrc/stable/tensor.h>
+#include <torch/headeronly/util/Exception.h>
+
+#include <cuda/ptx>
+
+#include "cute/tensor.hpp"
+#include "libtorch_stable/torch_utils.h"
+
+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::stable::Tensor& input,
+                                const torch::stable::Tensor& problem_sizes,
+                                const torch::stable::Tensor& expert_offsets,
+                                const torch::stable::Tensor& blockscale_offsets,
+                                torch::stable::Tensor& quant_output,
+                                torch::stable::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;
+  STD_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(get_device_prop()->multiProcessorCount * max_active_blocks_per_sm,
+            1, 1);
+  dim3 block(THREAD_BLOCK_SIZE, 1, 1);
+  int num_experts = static_cast<int>(problem_sizes.size(0));
+  auto stream = get_current_cuda_stream(input.get_device_index());
+  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/libtorch_stable/moe/permute_unpermute_kernels/dispatch.h

@@ -1,60 +0,0 @@
-#pragma once
-
-#include <cuda_fp8.h>
-#include <torch/headeronly/core/ScalarType.h>
-#include <torch/headeronly/util/Exception.h>
-
-#define MOE_SWITCH(TYPE, ...)                                         \
-  const auto _st = (TYPE);                                            \
-  switch (_st) {                                                      \
-    __VA_ARGS__                                                       \
-    default:                                                          \
-      STD_TORCH_CHECK(false, "[moe permute]data type dispatch fail!") \
-  }
-
-#define MOE_DISPATCH_CASE(enum_type, ...)                  \
-  case enum_type: {                                        \
-    using scalar_t = ScalarType2CudaType<enum_type>::type; \
-    __VA_ARGS__();                                         \
-    break;                                                 \
-  }
-
-#define MOE_DISPATCH_FLOAT_CASE(...)                                           \
-  MOE_DISPATCH_CASE(torch::headeronly::ScalarType::Float, __VA_ARGS__)         \
-  MOE_DISPATCH_CASE(torch::headeronly::ScalarType::Half, __VA_ARGS__)          \
-  MOE_DISPATCH_CASE(torch::headeronly::ScalarType::BFloat16, __VA_ARGS__)      \
-  MOE_DISPATCH_CASE(torch::headeronly::ScalarType::Float8_e5m2, __VA_ARGS__)   \
-  MOE_DISPATCH_CASE(torch::headeronly::ScalarType::Float8_e4m3fn, __VA_ARGS__) \
-  MOE_DISPATCH_CASE(torch::headeronly::ScalarType::Byte, __VA_ARGS__)
-
-#define MOE_DISPATCH(TYPE, ...) \
-  MOE_SWITCH(TYPE, MOE_DISPATCH_FLOAT_CASE(__VA_ARGS__))
-
-template <torch::headeronly::ScalarType type>
-struct ScalarType2CudaType;
-
-template <>
-struct ScalarType2CudaType<torch::headeronly::ScalarType::Float> {
-  using type = float;
-};
-template <>
-struct ScalarType2CudaType<torch::headeronly::ScalarType::Half> {
-  using type = half;
-};
-template <>
-struct ScalarType2CudaType<torch::headeronly::ScalarType::BFloat16> {
-  using type = __nv_bfloat16;
-};
-// uint8 for packed fp4
-template <>
-struct ScalarType2CudaType<torch::headeronly::ScalarType::Byte> {
-  using type = uint8_t;
-};
-template <>
-struct ScalarType2CudaType<torch::headeronly::ScalarType::Float8_e5m2> {
-  using type = __nv_fp8_e5m2;
-};
-template <>
-struct ScalarType2CudaType<torch::headeronly::ScalarType::Float8_e4m3fn> {
-  using type = __nv_fp8_e4m3;
-};

csrc/libtorch_stable/ops.h

@@ -3,9 +3,6 @@
 #include <torch/csrc/stable/library.h>
 #include <torch/csrc/stable/tensor.h>
 
-#include <optional>
-#include <string>
-
 void per_token_group_quant_fp8(const torch::stable::Tensor& input,
                                torch::stable::Tensor& output_q,
                                torch::stable::Tensor& output_s,
@@ -188,12 +185,11 @@ torch::stable::Tensor hadacore_transform(torch::stable::Tensor& x,
 
 // Layernorm kernels (shared CUDA/ROCm)
 void rms_norm(torch::stable::Tensor& out, torch::stable::Tensor& input,
-              std::optional<torch::stable::Tensor> weight, double epsilon);
+              torch::stable::Tensor& weight, double epsilon);
 
 void fused_add_rms_norm(torch::stable::Tensor& input,
                         torch::stable::Tensor& residual,
-                        std::optional<torch::stable::Tensor> weight,
-                        double epsilon);
+                        torch::stable::Tensor& weight, double epsilon);
 
 // Layernorm-quant kernels (shared CUDA/ROCm)
 void rms_norm_static_fp8_quant(torch::stable::Tensor& out,
@@ -285,25 +281,6 @@ minimax_allreduce_rms_qk(torch::stable::Tensor qkv,
                          int64_t const nranks, double const eps);
 #endif
 
-// Horizontally-fused MiniMax-M3 QK-norm + partial NeoX RoPE (+ optional KV /
-// index-cache insert). Dense layer: norm+RoPE only; sparse layer: also packs
-// the index branch and scatters k/v/index_k into their paged caches.
-void fused_minimax_m3_qknorm_rope_kv_insert(
-    torch::stable::Tensor& qkv, torch::stable::Tensor const& q_norm_weight,
-    torch::stable::Tensor const& k_norm_weight,
-    torch::stable::Tensor const& cos_sin_cache,
-    torch::stable::Tensor const& positions, int64_t num_heads,
-    int64_t num_kv_heads, int64_t rotary_dim, double eps,
-    std::optional<torch::stable::Tensor> index_q_norm_weight,
-    std::optional<torch::stable::Tensor> index_k_norm_weight,
-    int64_t num_index_heads, std::optional<torch::stable::Tensor> slot_mapping,
-    std::optional<torch::stable::Tensor> index_slot_mapping,
-    std::optional<torch::stable::Tensor> kv_cache,
-    std::optional<torch::stable::Tensor> index_cache, int64_t block_size,
-    std::optional<torch::stable::Tensor> q_out,
-    std::optional<torch::stable::Tensor> index_q_out,
-    const std::string& kv_cache_dtype);
-
 // Sampler kernels (shared CUDA/ROCm)
 void apply_repetition_penalties_(
     torch::stable::Tensor& logits, const torch::stable::Tensor& prompt_mask,
@@ -369,8 +346,7 @@ void free_shared_buffer(int64_t buffer);
 // Activation kernels (shared CUDA/ROCm)
 void silu_and_mul(torch::stable::Tensor& out, torch::stable::Tensor& input);
 void silu_and_mul_clamp(torch::stable::Tensor& out,
-                        torch::stable::Tensor& input, double limit,
-                        double alpha = 1.0, double beta = 0.0);
+                        torch::stable::Tensor& input, double limit);
 void mul_and_silu(torch::stable::Tensor& out, torch::stable::Tensor& input);
 void gelu_and_mul(torch::stable::Tensor& out, torch::stable::Tensor& input);
 void gelu_tanh_and_mul(torch::stable::Tensor& out,
@@ -421,6 +397,35 @@ torch::stable::Tensor gptq_gemm(torch::stable::Tensor a,
 void gptq_shuffle(torch::stable::Tensor q_weight, torch::stable::Tensor q_perm,
                   int64_t bit);
 
+// GGML kernels (shared CUDA/ROCm)
+torch::stable::Tensor ggml_dequantize(
+    torch::stable::Tensor W, int64_t type, int64_t m, int64_t n,
+    std::optional<torch::headeronly::ScalarType> const& dtype);
+
+torch::stable::Tensor ggml_mul_mat_vec_a8(torch::stable::Tensor W,
+                                          torch::stable::Tensor X, int64_t type,
+                                          int64_t row);
+
+torch::stable::Tensor ggml_mul_mat_a8(torch::stable::Tensor W,
+                                      torch::stable::Tensor X, int64_t type,
+                                      int64_t row);
+
+torch::stable::Tensor ggml_moe_a8(torch::stable::Tensor X,
+                                  torch::stable::Tensor W,
+                                  torch::stable::Tensor sorted_token_ids,
+                                  torch::stable::Tensor expert_ids,
+                                  torch::stable::Tensor num_tokens_post_padded,
+                                  int64_t type, int64_t row, int64_t top_k,
+                                  int64_t tokens);
+
+torch::stable::Tensor ggml_moe_a8_vec(torch::stable::Tensor X,
+                                      torch::stable::Tensor W,
+                                      torch::stable::Tensor topk_ids,
+                                      int64_t top_k, int64_t type, int64_t row,
+                                      int64_t tokens);
+
+int64_t ggml_moe_get_block_size(int64_t type);
+
 void paged_attention_v1(
     torch::stable::Tensor& out, torch::stable::Tensor& query,
     torch::stable::Tensor& key_cache, torch::stable::Tensor& value_cache,

csrc/libtorch_stable/quantization/fp4/mxfp4_experts_quant.cu

@@ -27,24 +27,15 @@
 #include <torch/csrc/stable/tensor.h>
 #include "libtorch_stable/torch_utils.h"
 #include "libtorch_stable/dispatch_utils.h"
-#include "libtorch_stable/cutlass_extensions/common.hpp"
 #include "../../cuda_vec_utils.cuh"
 #include "cuda_utils.h"
 
 #include "nvfp4_utils.cuh"
-
-#if defined(CUDART_VERSION) && CUDART_VERSION >= 12090
-  #define VLLM_MXFP4_EXPERTS_QUANT_SUPPORTED 1
 static_assert(CVT_FP4_ELTS_PER_THREAD == 16,
               "MXFP4 experts quant requires PACK16 mode (CUDA >= 12.9)");
-#else
-  #define VLLM_MXFP4_EXPERTS_QUANT_SUPPORTED 0
-#endif
 
 #include "libtorch_stable/launch_bounds_utils.h"
 
-#if VLLM_MXFP4_EXPERTS_QUANT_SUPPORTED
-
 namespace vllm {
 
 // MXFP4 block size constants
@@ -113,7 +104,7 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
                 &input_offset_by_experts[chunk_start + 12]));
         local_offsets[16] = __ldca(&input_offset_by_experts[chunk_start + 16]);
 
-  #pragma unroll
+#pragma unroll
         for (int i = 0; i < 16; i++) {
           if (rowIdx >= local_offsets[i] && rowIdx < local_offsets[i + 1]) {
             rowIdx_in_expert = rowIdx - local_offsets[i];
@@ -318,14 +309,14 @@ void mxfp4_quant_impl(void* output, void* output_scale, void* input,
 
 }  // namespace vllm
 
-  /*Quantization entry for mxfp4 experts quantization*/
-  #define CHECK_TH_CUDA(x, m) \
-    STD_TORCH_CHECK(x.is_cuda(), m, "must be a CUDA tensor")
-  #define CHECK_CONTIGUOUS(x, m) \
-    STD_TORCH_CHECK(x.is_contiguous(), m, "must be contiguous")
-  #define CHECK_INPUT(x, m) \
-    CHECK_TH_CUDA(x, m);    \
-    CHECK_CONTIGUOUS(x, m);
+/*Quantization entry for mxfp4 experts quantization*/
+#define CHECK_TH_CUDA(x, m) \
+  STD_TORCH_CHECK(x.is_cuda(), m, "must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x, m) \
+  STD_TORCH_CHECK(x.is_contiguous(), m, "must be contiguous")
+#define CHECK_INPUT(x, m) \
+  CHECK_TH_CUDA(x, m);    \
+  CHECK_CONTIGUOUS(x, m);
 
 constexpr auto HALF = torch::headeronly::ScalarType::Half;
 constexpr auto BF16 = torch::headeronly::ScalarType::BFloat16;
@@ -373,28 +364,12 @@ static void validate_mxfp4_experts_quant_inputs(
   STD_TORCH_CHECK(output_scale.size(1) * 4 == padded_k);
 }
 
-#endif  // VLLM_MXFP4_EXPERTS_QUANT_SUPPORTED
-
-static bool mxfp4_experts_quant_sm_supported(int64_t cuda_device_capability) {
-#if VLLM_MXFP4_EXPERTS_QUANT_SUPPORTED
-  return cuda_device_capability >= 100 && cuda_device_capability < 120;
-#else
-  return false;
-#endif
-}
-
 void mxfp4_experts_quant(
     torch::stable::Tensor& output, torch::stable::Tensor& output_scale,
     torch::stable::Tensor const& input,
     torch::stable::Tensor const& input_offset_by_experts,
     torch::stable::Tensor const& output_scale_offset_by_experts,
     int64_t n_experts) {
-#if VLLM_MXFP4_EXPERTS_QUANT_SUPPORTED
-  int32_t sm = get_sm_version_num();
-  STD_TORCH_CHECK(mxfp4_experts_quant_sm_supported(sm),
-                  "No compiled MXFP4 experts quant kernel for SM ", sm,
-                  ". Recompile with SM10x/11x FP4 support and CUDA >= 12.9.");
-
   auto m_topk = input.size(0);
   auto k = input.size(1);
 
@@ -415,10 +390,6 @@ void mxfp4_experts_quant(
             output_scale_offset_by_experts.data_ptr(), m_topk, k, n_experts,
             stream);
       });
-#else
-  STD_TORCH_CHECK_NOT_IMPLEMENTED(false,
-                                  "MXFP4 experts quant requires CUDA >= 12.9.");
-#endif
 }
 
 void silu_and_mul_mxfp4_experts_quant(
@@ -427,12 +398,6 @@ void silu_and_mul_mxfp4_experts_quant(
     torch::stable::Tensor const& input_offset_by_experts,
     torch::stable::Tensor const& output_scale_offset_by_experts,
     int64_t n_experts) {
-#if VLLM_MXFP4_EXPERTS_QUANT_SUPPORTED
-  int32_t sm = get_sm_version_num();
-  STD_TORCH_CHECK(mxfp4_experts_quant_sm_supported(sm),
-                  "No compiled SiLU+Mul MXFP4 experts quant kernel for SM ", sm,
-                  ". Recompile with SM10x/11x FP4 support and CUDA >= 12.9.");
-
   auto m_topk = input.size(0);
   auto k_times_2 = input.size(1);
   STD_TORCH_CHECK(k_times_2 % 2 == 0, "input width must be even (gate || up)");
@@ -455,29 +420,13 @@ void silu_and_mul_mxfp4_experts_quant(
             output_scale_offset_by_experts.data_ptr(), m_topk, k, n_experts,
             stream);
       });
-#else
-  STD_TORCH_CHECK_NOT_IMPLEMENTED(
-      false, "SiLU+Mul MXFP4 experts quant requires CUDA >= 12.9.");
-#endif
 }
 
-bool mxfp4_experts_quant_supported(int64_t cuda_device_capability) {
-  return mxfp4_experts_quant_sm_supported(cuda_device_capability);
-}
-
-STABLE_TORCH_LIBRARY_FRAGMENT(_C, m) {
-  m.def("mxfp4_experts_quant_supported(int cuda_device_capability) -> bool");
-}
-
-// Registered here so the CUDA 12.8 stub and CUDA 12.9+ implementation stay
-// tied to the same translation unit.
+// Registered here (not torch_bindings.cpp) because VLLM_GPU_FLAGS is applied
+// only under COMPILE_LANGUAGE:CUDA, so ENABLE_NVFP4_SM100 is invisible to
+// .cpp files and cannot gate the registration from there.
 STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, m) {
   m.impl("mxfp4_experts_quant", TORCH_BOX(&mxfp4_experts_quant));
   m.impl("silu_and_mul_mxfp4_experts_quant",
          TORCH_BOX(&silu_and_mul_mxfp4_experts_quant));
 }
-
-STABLE_TORCH_LIBRARY_IMPL(_C, CompositeExplicitAutograd, m) {
-  m.impl("mxfp4_experts_quant_supported",
-         TORCH_BOX(&mxfp4_experts_quant_supported));
-}

csrc/libtorch_stable/quantization/fp4/nvfp4_utils.cuh

@@ -22,15 +22,15 @@
 
 #include "../../cuda_vec_utils.cuh"
 
-#if defined(NVFP4_ENABLE_ELTS16) && defined(CUDART_VERSION) && \
-    CUDART_VERSION >= 12090
+#if defined(NVFP4_ENABLE_ELTS16) && defined(CUDA_VERSION) && \
+    CUDA_VERSION >= 12090
   #define ELTS_PER_THREAD 16
-  #define CVT_FP4_PACK16 1
 constexpr int CVT_FP4_ELTS_PER_THREAD = 16;
+constexpr bool CVT_FP4_PACK16 = true;
 #else
   #define ELTS_PER_THREAD 8
-  #define CVT_FP4_PACK16 0
 constexpr int CVT_FP4_ELTS_PER_THREAD = 8;
+constexpr bool CVT_FP4_PACK16 = false;
 #endif
 
 constexpr int CVT_FP4_SF_VEC_SIZE = 16;
@@ -237,30 +237,21 @@ __device__ __forceinline__ fp4_packed_t cvt_warp_fp16_to_fp4(
   // Get the final absolute maximum values.
   float vecMax = float(__hmax(localMax.x, localMax.y));
 
+  // Get the SF (max value of the vector / max value of e2m1).
+  // maximum value of e2m1 = 6.0.
+  // TODO: use half as compute data type.
+  float SFValue = SFScaleVal * (vecMax * reciprocal_approximate_ftz(6.0f));
   // 8 bits representation of the SF.
-  float SFValue;
   uint8_t fp8SFVal;
-
+  // Write the SF to global memory (STG.8).
   if constexpr (UE8M0_SF) {
-    // OCP MX spec E8M0 scale computation (MXFP4 path):
-    // scale_exp = biased_exponent(round_up(vecMax)) - 2
-    //   -2 because max E2M1 value is 6.0 ≈ 2^2.58; we use 2^2=4 as the
-    //   safe divisor so that max_val / scale <= 6.0 for values near 2^n.
-    uint32_t max_bits = __float_as_uint(vecMax);
-    // Add rounding bias at mantissa bit 21 (equivalent to bf16 val_to_add=32
-    // at bit 5). Threshold: values with mantissa >= 0.75 (i.e. >= 1.75*2^n)
-    // round up to the next power of 2.
-    uint32_t rounded_bits = (max_bits + (1u << 21)) & 0xFF800000u;
-    uint32_t biased_exp = (rounded_bits >> 23) & 0xFFu;
-    uint32_t scale_exp = (biased_exp > 2u) ? (biased_exp - 2u) : 0u;
-    scale_exp = min(scale_exp, 254u);
-    fp8SFVal = static_cast<uint8_t>(scale_exp);
-    // Reconstruct scale as float32: scale = 2^(scale_exp - 127)
-    uint32_t sf_bits = scale_exp << 23;
-    SFValue = __uint_as_float(sf_bits);
+    // Extract the 8 exponent bits from float32.
+    // float 32bits = 1 sign bit + 8 exponent bits + 23 mantissa bits.
+    uint32_t tmp = reinterpret_cast<uint32_t&>(SFValue) >> 23;
+    fp8SFVal = tmp & 0xff;
+    // Convert back to fp32.
+    reinterpret_cast<uint32_t&>(SFValue) = tmp << 23;
   } else {
-    // NVFP4 path: scale = max / 6.0, stored as E4M3.
-    SFValue = SFScaleVal * (vecMax * reciprocal_approximate_ftz(6.0f));
     // Here SFValue is always positive, so E4M3 is the same as UE4M3.
     __nv_fp8_e4m3 tmp = __nv_fp8_e4m3(SFValue);
     reinterpret_cast<__nv_fp8_e4m3&>(fp8SFVal) = tmp;
@@ -271,21 +262,13 @@ __device__ __forceinline__ fp4_packed_t cvt_warp_fp16_to_fp4(
   // Write the SF to global memory (STG.8).
   if (SFout) *SFout = fp8SFVal;
 
-  // Get the output scale (= 1 / SFValue for the MXFP4/UE8M0 path where
-  // SFScaleVal=1).  Use exact division for UE8M0 to ensure bit-exact scaling
-  // that matches the reference QDQ implementation (dividing by a power-of-2
-  // scale is exact in IEEE 754).
-  float outputScale;
-  if constexpr (UE8M0_SF) {
-    // SFValue is always a power of 2 for UE8M0, so 1/SFValue is exact.
-    outputScale = SFValue != 0.0f ? (1.0f / SFValue) : 0.0f;
-  } else {
-    // NVFP4 path: use fast approximate reciprocal (original behavior).
-    outputScale = SFValue != 0.0f
-                      ? reciprocal_approximate_ftz(
+  // Get the output scale.
+  // Recipe: final_scale = reciprocal(fp32(fp8(SFValue * SFScaleVal))) *
+  //                       reciprocal(SFScaleVal))
+  float outputScale =
+      SFValue != 0.0f ? reciprocal_approximate_ftz(
                             SFValue * reciprocal_approximate_ftz(SFScaleVal))
                       : 0.0f;
-  }
 
   // Convert the input to float.
   float2 fp2Vals[CVT_FP4_ELTS_PER_THREAD / 2];

csrc/libtorch_stable/quantization/gguf/dequantize.cuh

@@ -0,0 +1,571 @@
+// copied and adapted from https://github.com/ggerganov/llama.cpp/blob/b2899/ggml-cuda/convert.cu
+// Dequant functions
+static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
+    const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    const int vui = x[ib].qs[iqs];
+
+    v.x = __int2half_rn(vui & 0xF);
+    v.y = __int2half_rn(vui >> 4);
+
+    v = __hsub2(v, __floats2half2_rn(8.0f, 8.0f));
+    v = __hmul2(v, {d, d});
+}
+
+static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
+    const block_q4_1 * x = (const block_q4_1 *) vx;
+
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);
+
+    const int vui = x[ib].qs[iqs];
+
+    v.x = __int2half_rn(vui & 0xF);
+    v.y = __int2half_rn(vui >> 4);
+
+    v = __hmul2(v, {d, d});
+    v = __hadd2(v, {m, m});
+}
+
+static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
+    const block_q5_0 * x = (const block_q5_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    uint32_t qh;
+    memcpy(&qh, x[ib].qh, sizeof(qh));
+
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+
+    v.x = __int2half_rn((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y = __int2half_rn((x[ib].qs[iqs] >>  4) | xh_1);
+
+    v = __hsub2(v, __floats2half2_rn(16.0f, 16.0f));
+    v = __hmul2(v, {d, d});
+}
+
+static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
+    const block_q5_1 * x = (const block_q5_1 *) vx;
+
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);
+
+    uint32_t qh;
+    memcpy(&qh, x[ib].qh, sizeof(qh));
+
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+
+    v.x = __int2half_rn((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y = __int2half_rn((x[ib].qs[iqs] >>  4) | xh_1);
+
+    v = __hmul2(v, {d, d});
+    v = __hadd2(v, {m, m});
+}
+
+static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
+    const block_q8_0 * x = (const block_q8_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    v.x = __int2half_rn(x[ib].qs[iqs + 0]);
+    v.y = __int2half_rn(x[ib].qs[iqs + 1]);
+
+    v = __hmul2(v, {d, d});
+}
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) {
+    const int i = 2*(blockDim.x*blockIdx.x + threadIdx.x);
+
+    if (i >= k) {
+        return;
+    }
+
+    const int ib = i/qk; // block index
+    const int iqs = (i%qk)/qr; // quant index
+    const int iybs = i - i%qk; // y block start index
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    dfloat2 v;
+    dequantize_kernel(vx, ib, iqs, v);
+
+    y[iybs + iqs + 0]        = convert_from_half<dst_t>(v.x);
+    y[iybs + iqs + y_offset] = convert_from_half<dst_t>(v.y);
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const auto i   = blockIdx.x;
+    const block_q2_K * x = (const block_q2_K *) vx;
+
+    const auto tid = threadIdx.x;
+    const int n   = tid/32;
+    const int l   = tid - 32*n;
+    const int is  = 8*n + l/16;
+
+    const uint8_t q = x[i].qs[32*n + l];
+    dst_t * y = yy + i*QK_K + 128*n;
+
+    half dall = __low2half(x[i].dm);
+    half dmin = __high2half(x[i].dm);
+    y[l+ 0] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+0] & 0xF) * ((q >> 0) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+0] >> 4))));
+    y[l+32] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+2] & 0xF) * ((q >> 2) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+2] >> 4))));
+    y[l+64] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+4] & 0xF) * ((q >> 4) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+4] >> 4))));
+    y[l+96] = convert_from_half<dst_t>(__hsub(__hmul(dall, __int2half_rn((x[i].scales[is+6] & 0xF) * ((q >> 6) & 3))), __hmul(dmin,  __int2half_rn(x[i].scales[is+6] >> 4))));
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const auto i = blockIdx.x;
+    const block_q3_K * x = (const block_q3_K *) vx;
+
+    const auto r = threadIdx.x/4;
+    const int tid = r/2;
+    const int is0 = r%2;
+    const int l0 = 16*is0 + 4*(threadIdx.x%4);
+    const int n = tid / 4;
+    const int j = tid - 4*n;
+
+    uint8_t m = 1 << (4*n + j);
+    int is = 8*n + 2*j + is0;
+    int shift = 2*j;
+
+    int8_t us = is <  4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
+                is <  8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
+                is < 12 ? (x[i].scales[is-8] >>  4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
+                          (x[i].scales[is-8] >>  4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
+    half d_all = x[i].d;
+    half dl = __hmul(d_all,  __int2half_rn(us - 32));
+
+    dst_t * y = yy + i*QK_K + 128*n + 32*j;
+    const uint8_t * q = x[i].qs + 32*n;
+    const uint8_t * hm = x[i].hmask;
+
+    for (int l = l0; l < l0+4; ++l) {
+        y[l] = convert_from_half<dst_t>(__hmul(dl,  __int2half_rn((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4))));
+    }
+}
+
+static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
+    if (j < 4) {
+        d = q[j] & 63; m = q[j + 4] & 63;
+    } else {
+        d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
+        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const block_q4_K * x = (const block_q4_K *) vx;
+
+    const auto i = blockIdx.x;
+
+    // assume 32 threads
+    const auto tid = threadIdx.x;
+    const int il  = tid/8;
+    const int ir  = tid%8;
+    const int is  = 2*il;
+    const int n   = 4;
+
+    dst_t * y = yy + i*QK_K + 64*il + n*ir;
+
+    const half dall = __low2half(x[i].dm);
+    const half dmin = __high2half(x[i].dm);
+
+    const uint8_t * q = x[i].qs + 32*il + n*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, sc, m);
+    const half d1 = __hmul(dall, __int2half_rn(sc));
+    const half m1 = __hmul(dmin,  __int2half_rn(m));
+    get_scale_min_k4(is + 1, x[i].scales, sc, m);
+    const half d2 = __hmul(dall, __int2half_rn(sc));
+    const half m2 = __hmul(dmin, __int2half_rn(m));
+    for (int l = 0; l < n; ++l) {
+        y[l + 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn(q[l] & 0xF)), m1));
+        y[l +32] = convert_from_half<dst_t>(__hsub(__hmul(d2,  __int2half_rn(q[l] >> 4)), m2));
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const block_q5_K * x = (const block_q5_K *) vx;
+
+    const auto i = blockIdx.x;
+
+    // assume 64 threads - this is very slightly better than the one below
+    const auto tid = threadIdx.x;
+    const int il  = tid/16;   // il is in 0...3
+    const int ir  = tid%16;   // ir is in 0...15
+    const int is  = 2*il;     // is is in 0...6
+
+    dst_t * y = yy + i*QK_K + 64*il + 2*ir;
+
+    const half dall = __low2half(x[i].dm);
+    const half dmin = __high2half(x[i].dm);
+
+    const uint8_t * ql = x[i].qs + 32*il + 2*ir;
+    const uint8_t * qh = x[i].qh + 2*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, sc, m);
+    const half d1 = __hmul(dall, __int2half_rn(sc)); const half m1 = __hmul(dmin, __int2half_rn(m));
+    get_scale_min_k4(is + 1, x[i].scales, sc, m);
+    const half d2 = __hmul(dall, __int2half_rn(sc)); const half m2 = __hmul(dmin, __int2half_rn(m));
+
+    uint8_t   hm  = 1 << (2*il);
+    y[ 0] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0))), m1));
+    y[ 1] = convert_from_half<dst_t>(__hsub(__hmul(d1, __int2half_rn((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0))), m1));
+    hm <<= 1;
+    y[32] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[0] >>  4) + (qh[0] & hm ? 16 : 0))), m2));
+    y[33] = convert_from_half<dst_t>(__hsub(__hmul(d2, __int2half_rn((ql[1] >>  4) + (qh[1] & hm ? 16 : 0))), m2));
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const block_q6_K * x = (const block_q6_K *) vx;
+
+    const auto i = blockIdx.x;
+
+    // assume 64 threads - this is very slightly better than the one below
+    const auto tid = threadIdx.x;
+    const int ip  = tid/32;   // ip is 0 or 1
+    const int il  = tid - 32*ip; // 0...32
+    const int is  = 8*ip + il/16;
+
+    dst_t * y = yy + i*QK_K + 128*ip + il;
+
+    const half d = x[i].d;
+
+    const uint8_t * ql = x[i].ql + 64*ip + il;
+    const uint8_t   qh = x[i].qh[32*ip + il];
+    const int8_t  * sc = x[i].scales + is;
+
+    y[ 0] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32))));
+    y[32] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32))));
+    y[64] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32))));
+    y[96] = convert_from_half<dst_t>(__hmul(d, __int2half_rn(sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32))));
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const auto i   = blockIdx.x;
+    const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
+
+    const auto tid = threadIdx.x;
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const uint8_t  * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]);
+    const uint32_t aux32 = q2[2] | (q2[3] << 16);
+    const float d = __half2float(x[i].d) * (0.5f + (aux32 >> 28)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const auto i   = blockIdx.x;
+    const block_iq2_xs * x = (const block_iq2_xs *) vx;
+
+    const auto tid = threadIdx.x;
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
+    const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const auto i   = blockIdx.x;
+    const block_iq2_s * x = (const block_iq2_s *) vx;
+
+    const auto tid = threadIdx.x;
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
+    const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const auto i   = blockIdx.x;
+    const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
+
+    const auto tid = threadIdx.x;
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t  * q3 = x[i].qs + 8*ib;
+    const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib;
+    const uint8_t  * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
+    const uint8_t  * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float d = __half2float(x[i].d) * (0.5f + (aux32 >> 28)) * 0.5f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const auto i   = blockIdx.x;
+    const block_iq3_s * x = (const block_iq3_s *) vx;
+
+    const auto tid = threadIdx.x;
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * qs = x[i].qs + 8*ib;
+    const uint8_t * grid1 = (const uint8_t *)(iq3xs_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)));
+    const uint8_t * grid2 = (const uint8_t *)(iq3xs_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)));
+    const float d = __half2float(x[i].d) * (0.5f + ((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)) * 0.5f;
+    const uint8_t signs = x[i].signs[4*ib + il];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq1_s * x = (const block_iq1_s  *) vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
+    const float d = __half2float(x[i].d) * (2*((x[i].qh[ib] >> 12) & 7) + 1);
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq1_m * x = (const block_iq1_m  *) vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * sc = (const uint16_t *)x[i].scales;
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const int64_t ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4);
+    const float d = __half2float(scale.f16) * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1);
+    const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA;
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const auto i   = blockIdx.x;
+    const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
+
+    const auto tid = threadIdx.x;
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ib].qs + 4*il;
+    const float d = __half2float(x[ib].d);
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const auto i   = blockIdx.x;
+    const block_iq4_xs * x = (const block_iq4_xs *)vx;
+
+    const auto tid = threadIdx.x;
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[i].qs + 16*ib + 4*il;
+    const float d = __half2float(x[i].d) * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+}
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
+    dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+}
+
+template<typename dst_t>
+static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_xxs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq3_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq1_m<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+    dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static to_cuda_ggml_t<dst_t> ggml_get_to_cuda(int64_t type) {
+    switch (type) {
+        case 2:
+            return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
+        case 3:
+            return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
+        case 6:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case 7:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case 8:
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+        case 10:
+            return dequantize_row_q2_K_cuda;
+        case 11:
+            return dequantize_row_q3_K_cuda;
+        case 12:
+            return dequantize_row_q4_K_cuda;
+        case 13:
+            return dequantize_row_q5_K_cuda;
+        case 14:
+            return dequantize_row_q6_K_cuda;
+        case 16:
+            return dequantize_row_iq2_xxs_cuda;
+        case 17:
+            return dequantize_row_iq2_xs_cuda;
+        case 18:
+            return dequantize_row_iq3_xxs_cuda;
+        case 19:
+            return dequantize_row_iq1_s_cuda;
+        case 20:
+            return dequantize_row_iq4_nl_cuda;
+        case 21:
+            return dequantize_row_iq3_s_cuda;
+        case 22:
+            return dequantize_row_iq2_s_cuda;
+        case 23:
+            return dequantize_row_iq4_xs_cuda;
+        case 29:
+            return dequantize_row_iq1_m_cuda;
+        default:
+            return nullptr;
+    }
+}

csrc/libtorch_stable/quantization/gguf/gguf_kernel.cu

@@ -0,0 +1,557 @@
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+
+#include "../../../cuda_compat.h"
+#include "../../dispatch_utils.h"
+#include "../../torch_utils.h"
+
+#include <torch/csrc/stable/ops.h>
+
+#include "ggml-common.h"
+#include "vecdotq.cuh"
+#include "dequantize.cuh"
+#include "mmvq.cuh"
+#include "mmq.cuh"
+#include "moe.cuh"
+#include "moe_vec.cuh"
+
+// Q8 gemv
+template <typename scalar_t>
+static __global__ void quantize_q8_1(const scalar_t* __restrict__ x,
+                                     void* __restrict__ vy, const int kx,
+                                     const int kx_padded) {
+  const auto ix = blockDim.x * blockIdx.x + threadIdx.x;
+  if (ix >= kx_padded) {
+    return;
+  }
+  const auto iy = blockDim.y * blockIdx.y + threadIdx.y;
+  const int i_padded = iy * kx_padded + ix;
+
+  block_q8_1* y = (block_q8_1*)vy;
+
+  const int ib = i_padded / QK8_1;   // block index
+  const int iqs = i_padded % QK8_1;  // quant index
+
+  const float xi = ix < kx ? static_cast<float>(x[iy * kx + ix]) : 0.0f;
+  float amax = fabsf(xi);
+  float sum = xi;
+
+#pragma unroll
+  for (int mask = 16; mask > 0; mask >>= 1) {
+    amax = fmaxf(amax, VLLM_SHFL_XOR_SYNC_WIDTH(amax, mask, 32));
+    sum += VLLM_SHFL_XOR_SYNC_WIDTH(sum, mask, 32);
+  }
+
+  const float d = amax / 127;
+  const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
+
+  y[ib].qs[iqs] = q;
+
+  if (iqs > 0) {
+    return;
+  }
+
+  y[ib].ds.x = __float2half(d);
+  y[ib].ds.y = __float2half(sum);
+}
+
+template <typename scalar_t>
+static void quantize_row_q8_1_cuda(const scalar_t* x, void* vy, const int kx,
+                                   const int ky, cudaStream_t stream) {
+  const int64_t kx_padded = (kx + 512 - 1) / 512 * 512;
+  const int block_num_x =
+      (kx_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
+  constexpr int MAX_BLOCK_SIZE = 65535;
+  for (int off = 0; off < ky; off += MAX_BLOCK_SIZE) {
+    const int num_blocks_y = std::min(ky, off + MAX_BLOCK_SIZE) - off;
+    const dim3 num_blocks(block_num_x, num_blocks_y, 1);
+    const dim3 block_size(CUDA_DEQUANTIZE_BLOCK_SIZE, 1, 1);
+    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(
+        &x[off * kx], (int32_t*)vy + off * (kx_padded / 32 * 9), kx, kx_padded);
+  }
+}
+
+torch::stable::Tensor ggml_dequantize(
+    torch::stable::Tensor W,  // quant weight
+    int64_t type, int64_t m, int64_t n,
+    std::optional<torch::headeronly::ScalarType> const& dtype) {
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      W.get_device_index());
+  auto dtype_ = dtype.value_or(torch::headeronly::ScalarType::Half);
+  auto DW = torch::stable::empty({m, n}, dtype_, std::nullopt, W.device());
+  cudaStream_t stream = get_current_cuda_stream();
+
+  VLLM_STABLE_DISPATCH_FLOATING_TYPES(DW.scalar_type(), "ggml_dequantize", [&] {
+    auto to_cuda = ggml_get_to_cuda<scalar_t>(type);
+    to_cuda((void*)W.data_ptr(), (scalar_t*)DW.data_ptr(), m * n, stream);
+  });
+
+  return DW;
+}
+
+torch::stable::Tensor ggml_mul_mat_vec_a8(
+    torch::stable::Tensor W,  // quant weight
+    torch::stable::Tensor X,  // input
+    int64_t type, int64_t row) {
+  int col = X.sizes()[1];
+  int vecs = X.sizes()[0];
+  const int padded = (col + 512 - 1) / 512 * 512;
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      X.get_device_index());
+  auto Y = torch::stable::empty({vecs, row}, X.scalar_type(), std::nullopt,
+                                W.device());
+  cudaStream_t stream = get_current_cuda_stream();
+  auto quant_X = torch::stable::empty({vecs, padded / 32 * 9},
+                                      torch::headeronly::ScalarType::Int,
+                                      std::nullopt, W.device());
+  VLLM_STABLE_DISPATCH_FLOATING_TYPES(
+      X.scalar_type(), "ggml_mul_mat_vec_a8", [&] {
+        quantize_row_q8_1_cuda<scalar_t>((scalar_t*)X.data_ptr(),
+                                         (void*)quant_X.data_ptr(), col, vecs,
+                                         stream);
+        switch (type) {
+          case 2:
+            mul_mat_vec_q4_0_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 3:
+            mul_mat_vec_q4_1_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 6:
+            mul_mat_vec_q5_0_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 7:
+            mul_mat_vec_q5_1_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 8:
+            mul_mat_vec_q8_0_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 10:
+            mul_mat_vec_q2_K_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 11:
+            mul_mat_vec_q3_K_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 12:
+            mul_mat_vec_q4_K_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 13:
+            mul_mat_vec_q5_K_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 14:
+            mul_mat_vec_q6_K_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 16:
+            mul_mat_vec_iq2_xxs_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 17:
+            mul_mat_vec_iq2_xs_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 18:
+            mul_mat_vec_iq3_xxs_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 19:
+            mul_mat_vec_iq1_s_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 20:
+            mul_mat_vec_iq4_nl_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 21:
+            mul_mat_vec_iq3_s_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 22:
+            mul_mat_vec_iq2_s_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 23:
+            mul_mat_vec_iq4_xs_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+          case 29:
+            mul_mat_vec_iq1_m_q8_1_cuda<scalar_t>(
+                (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+                (scalar_t*)Y.data_ptr(), col, row, vecs, stream);
+            break;
+        }
+      });
+  return Y;
+}
+
+torch::stable::Tensor ggml_mul_mat_a8(torch::stable::Tensor W,  // quant weight
+                                      torch::stable::Tensor X,  // input
+                                      int64_t type, int64_t row) {
+  int col = X.sizes()[1];
+  int padded = (col + 512 - 1) / 512 * 512;
+  int batch = X.sizes()[0];
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      X.get_device_index());
+  auto Y = torch::stable::empty({batch, row}, X.scalar_type(), std::nullopt,
+                                W.device());
+  cudaStream_t stream = get_current_cuda_stream();
+  auto quant_X = torch::stable::empty({batch, padded / 32 * 9},
+                                      torch::headeronly::ScalarType::Int,
+                                      std::nullopt, W.device());
+  VLLM_STABLE_DISPATCH_FLOATING_TYPES(X.scalar_type(), "ggml_mul_mat_a8", [&] {
+    quantize_row_q8_1_cuda((scalar_t*)X.data_ptr(), (void*)quant_X.data_ptr(),
+                           col, batch, stream);
+
+    switch (type) {
+      case 2:
+        ggml_mul_mat_q4_0_q8_1_cuda(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), col, row, batch, padded, row, stream);
+        break;
+      case 3:
+        ggml_mul_mat_q4_1_q8_1_cuda(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), col, row, batch, padded, row, stream);
+        break;
+      case 6:
+        ggml_mul_mat_q5_0_q8_1_cuda(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), col, row, batch, padded, row, stream);
+        break;
+      case 7:
+        ggml_mul_mat_q5_1_q8_1_cuda(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), col, row, batch, padded, row, stream);
+        break;
+      case 8:
+        ggml_mul_mat_q8_0_q8_1_cuda(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), col, row, batch, padded, row, stream);
+        break;
+      case 10:
+        ggml_mul_mat_q2_K_q8_1_cuda(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), col, row, batch, padded, row, stream);
+        break;
+      case 11:
+        ggml_mul_mat_q3_K_q8_1_cuda(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), col, row, batch, padded, row, stream);
+        break;
+      case 12:
+        ggml_mul_mat_q4_K_q8_1_cuda(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), col, row, batch, padded, row, stream);
+        break;
+      case 13:
+        ggml_mul_mat_q5_K_q8_1_cuda(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), col, row, batch, padded, row, stream);
+        break;
+      case 14:
+        ggml_mul_mat_q6_K_q8_1_cuda(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), col, row, batch, padded, row, stream);
+        break;
+    }
+  });
+  return Y;
+}
+
+torch::stable::Tensor ggml_moe_a8(torch::stable::Tensor X,  // input
+                                  torch::stable::Tensor W,  // expert weights
+                                  torch::stable::Tensor sorted_token_ids,
+                                  torch::stable::Tensor expert_ids,
+                                  torch::stable::Tensor num_tokens_post_padded,
+                                  int64_t type, int64_t row, int64_t top_k,
+                                  int64_t tokens) {
+  int col = X.sizes()[1];
+  int padded = (col + 512 - 1) / 512 * 512;
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      X.get_device_index());
+  auto Y = torch::stable::empty({tokens * top_k, row}, X.scalar_type(),
+                                std::nullopt, W.device());
+  cudaStream_t stream = get_current_cuda_stream();
+  auto quant_X = torch::stable::empty({tokens, padded / 32 * 9},
+                                      torch::headeronly::ScalarType::Int,
+                                      std::nullopt, W.device());
+  VLLM_STABLE_DISPATCH_FLOATING_TYPES(X.scalar_type(), "ggml_moe_a8", [&] {
+    quantize_row_q8_1_cuda((scalar_t*)X.data_ptr(), (void*)quant_X.data_ptr(),
+                           col, tokens, stream);
+    switch (type) {
+      case 2:
+        ggml_moe_q4_0_q8_1_cuda(
+            (void*)quant_X.data_ptr(), (void*)W.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)sorted_token_ids.data_ptr(),
+            (int*)expert_ids.data_ptr(),
+            (int*)num_tokens_post_padded.data_ptr(), W.stride(0), col, row,
+            tokens, padded, row, top_k, sorted_token_ids.sizes()[0], stream);
+        break;
+      case 3:
+        ggml_moe_q4_1_q8_1_cuda(
+            (void*)quant_X.data_ptr(), (void*)W.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)sorted_token_ids.data_ptr(),
+            (int*)expert_ids.data_ptr(),
+            (int*)num_tokens_post_padded.data_ptr(), W.stride(0), col, row,
+            tokens, padded, row, top_k, sorted_token_ids.sizes()[0], stream);
+        break;
+      case 6:
+        ggml_moe_q5_0_q8_1_cuda(
+            (void*)quant_X.data_ptr(), (void*)W.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)sorted_token_ids.data_ptr(),
+            (int*)expert_ids.data_ptr(),
+            (int*)num_tokens_post_padded.data_ptr(), W.stride(0), col, row,
+            tokens, padded, row, top_k, sorted_token_ids.sizes()[0], stream);
+        break;
+      case 7:
+        ggml_moe_q5_1_q8_1_cuda(
+            (void*)quant_X.data_ptr(), (void*)W.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)sorted_token_ids.data_ptr(),
+            (int*)expert_ids.data_ptr(),
+            (int*)num_tokens_post_padded.data_ptr(), W.stride(0), col, row,
+            tokens, padded, row, top_k, sorted_token_ids.sizes()[0], stream);
+        break;
+      case 8:
+        ggml_moe_q8_0_q8_1_cuda(
+            (void*)quant_X.data_ptr(), (void*)W.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)sorted_token_ids.data_ptr(),
+            (int*)expert_ids.data_ptr(),
+            (int*)num_tokens_post_padded.data_ptr(), W.stride(0), col, row,
+            tokens, padded, row, top_k, sorted_token_ids.sizes()[0], stream);
+        break;
+      case 10:
+        ggml_moe_q2_K_q8_1_cuda(
+            (void*)quant_X.data_ptr(), (void*)W.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)sorted_token_ids.data_ptr(),
+            (int*)expert_ids.data_ptr(),
+            (int*)num_tokens_post_padded.data_ptr(), W.stride(0), col, row,
+            tokens, padded, row, top_k, sorted_token_ids.sizes()[0], stream);
+        break;
+      case 11:
+        ggml_moe_q3_K_q8_1_cuda(
+            (void*)quant_X.data_ptr(), (void*)W.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)sorted_token_ids.data_ptr(),
+            (int*)expert_ids.data_ptr(),
+            (int*)num_tokens_post_padded.data_ptr(), W.stride(0), col, row,
+            tokens, padded, row, top_k, sorted_token_ids.sizes()[0], stream);
+        break;
+      case 12:
+        ggml_moe_q4_K_q8_1_cuda(
+            (void*)quant_X.data_ptr(), (void*)W.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)sorted_token_ids.data_ptr(),
+            (int*)expert_ids.data_ptr(),
+            (int*)num_tokens_post_padded.data_ptr(), W.stride(0), col, row,
+            tokens, padded, row, top_k, sorted_token_ids.sizes()[0], stream);
+        break;
+      case 13:
+        ggml_moe_q5_K_q8_1_cuda(
+            (void*)quant_X.data_ptr(), (void*)W.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)sorted_token_ids.data_ptr(),
+            (int*)expert_ids.data_ptr(),
+            (int*)num_tokens_post_padded.data_ptr(), W.stride(0), col, row,
+            tokens, padded, row, top_k, sorted_token_ids.sizes()[0], stream);
+        break;
+      case 14:
+        ggml_moe_q6_K_q8_1_cuda(
+            (void*)quant_X.data_ptr(), (void*)W.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)sorted_token_ids.data_ptr(),
+            (int*)expert_ids.data_ptr(),
+            (int*)num_tokens_post_padded.data_ptr(), W.stride(0), col, row,
+            tokens, padded, row, top_k, sorted_token_ids.sizes()[0], stream);
+        break;
+    }
+  });
+  return Y;
+}
+
+torch::stable::Tensor ggml_moe_a8_vec(
+    torch::stable::Tensor X,  // input
+    torch::stable::Tensor W,  // expert weights
+    torch::stable::Tensor topk_ids, int64_t top_k, int64_t type, int64_t row,
+    int64_t tokens) {
+  int col = X.sizes()[1];
+  const int padded = (col + 512 - 1) / 512 * 512;
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      X.get_device_index());
+  auto Y = torch::stable::empty({tokens * top_k, row}, X.scalar_type(),
+                                std::nullopt, W.device());
+  torch::stable::fill_(Y, 0.0);
+  cudaStream_t stream = get_current_cuda_stream();
+  auto quant_X = torch::stable::empty({tokens, padded / 32 * 9},
+                                      torch::headeronly::ScalarType::Int,
+                                      std::nullopt, W.device());
+  VLLM_STABLE_DISPATCH_FLOATING_TYPES(X.scalar_type(), "ggml_moe_vec_a8", [&] {
+    quantize_row_q8_1_cuda<scalar_t>((scalar_t*)X.data_ptr(),
+                                     (void*)quant_X.data_ptr(), col, tokens,
+                                     stream);
+    switch (type) {
+      case 2:
+        moe_vec_q4_0_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 3:
+        moe_vec_q4_1_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 6:
+        moe_vec_q5_0_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 7:
+        moe_vec_q5_1_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 8:
+        moe_vec_q8_0_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 10:
+        moe_vec_q2_K_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 11:
+        moe_vec_q3_K_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 12:
+        moe_vec_q4_K_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 13:
+        moe_vec_q5_K_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 14:
+        moe_vec_q6_K_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 16:
+        moe_vec_iq2_xxs_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 17:
+        moe_vec_iq2_xs_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 18:
+        moe_vec_iq3_xxs_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 19:
+        moe_vec_iq1_s_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 20:
+        moe_vec_iq4_nl_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 21:
+        moe_vec_iq3_s_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 22:
+        moe_vec_iq2_s_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 23:
+        moe_vec_iq4_xs_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+      case 29:
+        moe_vec_iq1_m_q8_1_cuda<scalar_t>(
+            (void*)W.data_ptr(), (void*)quant_X.data_ptr(),
+            (scalar_t*)Y.data_ptr(), (int*)topk_ids.data_ptr(), top_k, tokens,
+            col, row, quant_X.stride(0), stream);
+        break;
+    }
+  });
+  return Y;
+}
+
+int64_t ggml_moe_get_block_size(int64_t type) {
+  switch (type) {
+    case 2:
+      return MOE_X_Q4_0;
+    case 3:
+      return MOE_X_Q4_1;
+    case 6:
+      return MOE_X_Q5_0;
+    case 7:
+      return MOE_X_Q5_1;
+    case 8:
+      return MOE_X_Q8_0;
+    case 10:
+      return MOE_X_Q2_K;
+    case 11:
+      return MOE_X_Q3_K;
+    case 12:
+      return MOE_X_Q4_K;
+    case 13:
+      return MOE_X_Q5_K;
+    case 14:
+      return MOE_X_Q6_K;
+  }
+  return 0;
+}