Patch protobuf for CVE-2026-0994 (#34253)

Signed-off-by: Seiji Eicher <seiji@anyscale.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
(cherry picked from commit 5045d5c983)

.buildkite/hardware_tests/amd.yaml

@@ -1,7 +1,6 @@
-group: Hardware - AMD Build 
+group: Hardware
 steps:
   - label: "AMD: :docker: build image"
-    key: image-build-amd
     depends_on: []
     device: amd_cpu
     no_plugin: true
@@ -10,7 +9,7 @@ steps:
       docker build
       --build-arg max_jobs=16
       --build-arg REMOTE_VLLM=1
-      --build-arg ARG_PYTORCH_ROCM_ARCH='gfx90a;gfx942;gfx950'
+      --build-arg ARG_PYTORCH_ROCM_ARCH='gfx90a;gfx942'
       --build-arg VLLM_BRANCH=$BUILDKITE_COMMIT
       --tag "rocm/vllm-ci:${BUILDKITE_COMMIT}"
       -f docker/Dockerfile.rocm

.buildkite/hardware_tests/cpu.yaml

@@ -21,20 +21,6 @@ steps:
       pytest -x -v -s tests/kernels/moe/test_cpu_fused_moe.py
       pytest -x -v -s tests/kernels/test_onednn.py"
 
-- label: CPU-Compatibility Tests
-  depends_on: []
-  soft_fail: true
-  device: intel_cpu
-  no_plugin: true
-  source_file_dependencies:
-  - cmake/cpu_extension.cmake
-  - setup.py
-  - vllm/platforms/cpu.py
-  commands:
-    - |
-      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 20m "
-      bash .buildkite/scripts/hardware_ci/run-cpu-compatibility-test.sh"
-
 - label: CPU-Language Generation and Pooling Model Tests
   depends_on: []
   soft_fail: true

.buildkite/image_build/image_build.sh

@@ -8,7 +8,7 @@ clean_docker_tag() {
 }
 
 print_usage_and_exit() {
-    echo "Usage: $0 <registry> <repo> <commit> <branch> <image_tag> [<image_tag_latest>]"
+    echo "Usage: $0 <registry> <repo> <commit> <branch> <vllm_use_precompiled> <vllm_merge_base_commit> <cache_from> <cache_to>"
     exit 1
 }
 
@@ -142,16 +142,11 @@ resolve_parent_commit() {
 
 print_bake_config() {
     echo "--- :page_facing_up: Resolved bake configuration"
-    # Write to a temp directory to avoid polluting the repo root (which is the
+    BAKE_CONFIG_FILE="bake-config-build-${BUILDKITE_BUILD_NUMBER:-local}.json"
-    # Docker build context). Files left in the repo root get COPY'd into the
-    # image and can cause duplicate artifact uploads from downstream steps.
-    local bake_tmp
-    bake_tmp="$(mktemp -d)"
-    BAKE_CONFIG_FILE="${bake_tmp}/bake-config-build-${BUILDKITE_BUILD_NUMBER:-local}.json"
     docker buildx bake -f "${VLLM_BAKE_FILE_PATH}" -f "${CI_HCL_PATH}" --print "${TARGET}" | tee "${BAKE_CONFIG_FILE}" || true
     echo "Saved bake config to ${BAKE_CONFIG_FILE}"
     echo "--- :arrow_down: Uploading bake config to Buildkite"
-    (cd "$(dirname "${BAKE_CONFIG_FILE}")" && buildkite-agent artifact upload "$(basename "${BAKE_CONFIG_FILE}")")
+    buildkite-agent artifact upload "${BAKE_CONFIG_FILE}"
 }
 
 #################################
@@ -159,7 +154,7 @@ print_bake_config() {
 #################################
 print_instance_info
 
-if [[ $# -lt 5 ]]; then
+if [[ $# -lt 7 ]]; then
     print_usage_and_exit
 fi
 
@@ -168,8 +163,10 @@ REGISTRY=$1
 REPO=$2
 BUILDKITE_COMMIT=$3
 BRANCH=$4
-IMAGE_TAG=$5
+VLLM_USE_PRECOMPILED=$5
-IMAGE_TAG_LATEST=${6:-} # only used for main branch, optional
+VLLM_MERGE_BASE_COMMIT=$6
+IMAGE_TAG=$7
+IMAGE_TAG_LATEST=${8:-} # only used for main branch, optional
 
 # build config
 TARGET="test-ci"
@@ -196,6 +193,8 @@ export CACHE_FROM
 export CACHE_FROM_BASE_BRANCH
 export CACHE_FROM_MAIN
 export CACHE_TO
+export VLLM_USE_PRECOMPILED
+export VLLM_MERGE_BASE_COMMIT
 
 # print args
 echo "--- :mag: Arguments"
@@ -203,6 +202,8 @@ echo "REGISTRY: ${REGISTRY}"
 echo "REPO: ${REPO}"
 echo "BUILDKITE_COMMIT: ${BUILDKITE_COMMIT}"
 echo "BRANCH: ${BRANCH}"
+echo "VLLM_USE_PRECOMPILED: ${VLLM_USE_PRECOMPILED}"
+echo "VLLM_MERGE_BASE_COMMIT: ${VLLM_MERGE_BASE_COMMIT}"
 echo "IMAGE_TAG: ${IMAGE_TAG}"
 echo "IMAGE_TAG_LATEST: ${IMAGE_TAG_LATEST}"
 

.buildkite/image_build/image_build.yaml

@@ -5,7 +5,8 @@ steps:
     depends_on: []
     timeout_in_minutes: 600
     commands:
-    - if [[ "$BUILDKITE_BRANCH" == "main" ]]; then .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $IMAGE_TAG $IMAGE_TAG_LATEST; else .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $IMAGE_TAG; fi
+    - if [[ "$BUILDKITE_BRANCH" != "main" ]]; then .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $VLLM_USE_PRECOMPILED $VLLM_MERGE_BASE_COMMIT $IMAGE_TAG; fi
+    - if [[ "$BUILDKITE_BRANCH" == "main" ]]; then .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $VLLM_USE_PRECOMPILED $VLLM_MERGE_BASE_COMMIT $IMAGE_TAG $IMAGE_TAG_LATEST; fi
     retry:
       automatic:
         - exit_status: -1  # Agent was lost

.buildkite/image_build/image_build_cpu.sh

@@ -11,10 +11,10 @@ REPO=$2
 BUILDKITE_COMMIT=$3
 
 # authenticate with AWS ECR
-aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin $REGISTRY
 
 # skip build if image already exists
-if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu) ]]; then
+if [[ -z $(docker manifest inspect $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu) ]]; then
   echo "Image not found, proceeding with build..."
 else
   echo "Image found"
@@ -24,11 +24,13 @@ fi
 # build
 docker build --file docker/Dockerfile.cpu \
   --build-arg max_jobs=16 \
-  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
+  --build-arg buildkite_commit=$BUILDKITE_COMMIT \
-  --build-arg VLLM_CPU_X86=true \
+  --build-arg VLLM_CPU_AVX512BF16=true \
-  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu \
+  --build-arg VLLM_CPU_AVX512VNNI=true \
+  --build-arg VLLM_CPU_AMXBF16=true \
+  --tag $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu \
   --target vllm-test \
   --progress plain .
 
 # push
-docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu
+docker push $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu

.buildkite/image_build/image_build_cpu_arm64.sh

@@ -11,10 +11,10 @@ REPO=$2
 BUILDKITE_COMMIT=$3
 
 # authenticate with AWS ECR
-aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin $REGISTRY
 
 # skip build if image already exists
-if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu) ]]; then
+if [[ -z $(docker manifest inspect $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu) ]]; then
   echo "Image not found, proceeding with build..."
 else
   echo "Image found"
@@ -24,10 +24,10 @@ fi
 # build
 docker build --file docker/Dockerfile.cpu \
   --build-arg max_jobs=16 \
-  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
+  --build-arg buildkite_commit=$BUILDKITE_COMMIT \
-  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu \
+  --tag $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu \
   --target vllm-test \
   --progress plain .
 
 # push
-docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu
+docker push $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu

.buildkite/image_build/image_build_hpu.sh

@@ -11,10 +11,10 @@ REPO=$2
 BUILDKITE_COMMIT=$3
 
 # authenticate with AWS ECR
-aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin $REGISTRY
 
 # skip build if image already exists
-if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu) ]]; then
+if [[ -z $(docker manifest inspect $REGISTRY/$REPO:$BUILDKITE_COMMIT-hpu) ]]; then
   echo "Image not found, proceeding with build..."
 else
   echo "Image found"
@@ -25,10 +25,10 @@ fi
 docker build \
   --file tests/pytorch_ci_hud_benchmark/Dockerfile.hpu \
   --build-arg max_jobs=16 \
-  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
+  --build-arg buildkite_commit=$BUILDKITE_COMMIT \
-  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu \
+  --tag $REGISTRY/$REPO:$BUILDKITE_COMMIT-hpu \
   --progress plain \
   https://github.com/vllm-project/vllm-gaudi.git
 
 # push
-docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu
+docker push $REGISTRY/$REPO:$BUILDKITE_COMMIT-hpu

.buildkite/lm-eval-harness/run-lm-eval-chartqa-vllm-vlm-baseline.sh

@@ -2,7 +2,7 @@
 # We can use this script to compute baseline accuracy on chartqa for vllm.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install "lm-eval[api]>=0.4.11"
+#   pip install "lm-eval[api]>=0.4.9.2"
 
 usage() {
     echo``
@@ -41,4 +41,4 @@ lm_eval --model vllm-vlm \
   --tasks chartqa \
   --batch_size auto \
   --apply_chat_template \
-  --limit "$LIMIT"
+  --limit $LIMIT

.buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh

@@ -2,7 +2,7 @@
 # We can use this script to compute baseline accuracy on GSM for transformers.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install "lm-eval[api]>=0.4.11"
+#   pip install "lm-eval[api]>=0.4.9.2"
 
 usage() {
     echo``

.buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh

@@ -3,7 +3,7 @@
 # We use this for fp8, which HF does not support.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install "lm-eval[api]>=0.4.11"
+#   pip install "lm-eval[api]>=0.4.9.2"
 
 usage() {
     echo``

.buildkite/lm-eval-harness/run-lm-eval-mmlupro-vllm-baseline.sh

@@ -3,7 +3,7 @@
 # We use this for fp8, which HF does not support.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install "lm-eval[api]>=0.4.11"
+#   pip install "lm-eval[api]>=0.4.9.2"
 
 usage() {
     echo``
@@ -20,11 +20,14 @@ usage() {
     echo
 }
 
-while getopts "m:l:f:t:" OPT; do
+while getopts "m:b:l:f:t:" OPT; do
   case ${OPT} in
     m )
         MODEL="$OPTARG"
         ;;
+    b )
+        BATCH_SIZE="$OPTARG"
+        ;;
     l )
         LIMIT="$OPTARG"
         ;;

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

@@ -13,10 +13,9 @@ import os
 from contextlib import contextmanager
 
 import lm_eval
+import numpy as np
 import yaml
 
-from vllm.platforms import current_platform
-
 DEFAULT_RTOL = 0.08
 
 
@@ -64,9 +63,6 @@ def launch_lm_eval(eval_config, tp_size):
         "allow_deprecated_quantization=True,"
     )
 
-    if current_platform.is_rocm() and "Nemotron-3" in eval_config["model_name"]:
-        model_args += "attention_backend=TRITON_ATTN"
-
     env_vars = eval_config.get("env_vars", None)
     with scoped_env_vars(env_vars):
         results = lm_eval.simple_evaluate(
@@ -106,8 +102,6 @@ def test_lm_eval_correctness_param(config_filename, tp_size):
                 f"ground_truth={ground_truth:.3f} | "
                 f"measured={measured_value:.3f} | rtol={rtol}"
             )
-
+            success = success and np.isclose(ground_truth, measured_value, rtol=rtol)
-            min_acceptable = ground_truth * (1 - rtol)
-            success = success and measured_value >= min_acceptable
 
     assert success

.buildkite/performance-benchmarks/README.md

@@ -83,6 +83,7 @@ We test the throughput by using `vllm bench serve` with request rate = inf to co
         "server_parameters": {
             "model": "meta-llama/Meta-Llama-3-8B",
             "tensor_parallel_size": 1,
+            "swap_space": 16,
             "disable_log_stats": "",
             "load_format": "dummy"
         },

.buildkite/performance-benchmarks/scripts/compare-json-results.py

@@ -7,10 +7,8 @@ import argparse
 import html as _html
 import json
 import os
-from contextlib import nullcontext
 from dataclasses import dataclass
 from importlib import util
-from pathlib import Path
 
 import pandas as pd
 
@@ -33,45 +31,6 @@ pd.set_option("display.precision", 2)
 pd.set_option("display.float_format", lambda x: f"{x:.2f}")
 
 
-# -----------------------------
-# Concurrency normalization (NEW, small)
-# -----------------------------
-def _find_concurrency_col(df: pd.DataFrame) -> str:
-    for c in [
-        "# of max concurrency.",
-        "# of max concurrency",
-        "Max Concurrency",
-        "max_concurrency",
-        "Concurrency",
-    ]:
-        if c in df.columns:
-            return c
-
-    for c in df.columns:
-        if "concurr" in str(c).lower():
-            s = df[c]
-            if s.dtype.kind in "iu" and s.nunique() > 1 and s.min() >= 1:
-                return c
-
-    raise ValueError(
-        "Cannot infer concurrency column. "
-        "Please rename the column to one of the known names "
-        "or add an explicit override (e.g., --concurrency-col)."
-    )
-
-
-def _normalize_concurrency_in_df(
-    df: pd.DataFrame, canonical: str = "# of max concurrency."
-) -> pd.DataFrame:
-    if canonical in df.columns:
-        return df
-    detected = _find_concurrency_col(df)
-    if detected in df.columns and detected != canonical:
-        return df.rename(columns={detected: canonical})
-    df[canonical] = pd.NA
-    return df
-
-
 # -----------------------------
 # Core data compare
 # -----------------------------
@@ -91,25 +50,19 @@ def compare_data_columns(
     - Concat along axis=1 (indexes align), then reset_index so callers can
       group by columns.
     - If --debug, add a <file_label>_name column per file.
-
-    Minimal fix to support different max_concurrency lists across files:
-      - normalize concurrency column naming to "# of max concurrency."
-      - align on UNION of keys (missing points become NaN)
-      - BUGFIX: don't drop throughput rows based on P99/Median presence
     """
     print("\ncompare_data_column:", data_column)
 
     frames = []
     raw_data_cols: list[str] = []
+    compare_frames = []
 
-    # Determine key cols after normalizing concurrency
     cols_per_file: list[set] = []
     for f in files:
         try:
             df_tmp = pd.read_json(f, orient="records")
         except Exception as err:
             raise ValueError(f"Failed to read {f}") from err
-        df_tmp = _normalize_concurrency_in_df(df_tmp, canonical="# of max concurrency.")
         cols_per_file.append(set(df_tmp.columns))
 
     key_cols = [c for c in info_cols if all(c in cset for cset in cols_per_file)]
@@ -120,25 +73,12 @@ def compare_data_columns(
             "No common key columns found from info_cols across the input files."
         )
 
-    union_index = None
+    meta_added = False
-    metas: list[pd.DataFrame] = []
-    staged: list[tuple[str, pd.Series, pd.Series | None]] = []
 
     for file in files:
         df = pd.read_json(file, orient="records")
-        df = _normalize_concurrency_in_df(df, canonical="# of max concurrency.")
 
-        # BUGFIX: only drop rows for latency-like metrics; throughput rows may have
+        if drop_column in df.columns:
-        # NaN in P99/Median columns even if the column exists in the JSON.
-        metric_lc = str(data_column).lower()
-        is_latency_metric = (
-            "ttft" in metric_lc
-            or "tpot" in metric_lc
-            or "p99" in metric_lc
-            or "median" in metric_lc
-            or metric_lc.strip() in {"p99", "median"}
-        )
-        if is_latency_metric and drop_column in df.columns:
             df = df.dropna(subset=[drop_column], ignore_index=True)
 
         for c in (
@@ -163,61 +103,35 @@ def compare_data_columns(
             meta = meta.groupby(level=key_cols, dropna=False).first()
 
         file_label = "/".join(file.split("/")[:-1]) or os.path.basename(file)
-
+        s = df_idx[data_column]
-        if data_column in df_idx.columns:
+        if not s.index.is_unique:
-            s = df_idx[data_column]
+            s = s.groupby(level=key_cols, dropna=False).mean()
-            if not s.index.is_unique:
-                s = s.groupby(level=key_cols, dropna=False).mean()
-        else:
-            # keep NA series to preserve meta keys for union_index
-            s = pd.Series(pd.NA, index=meta.index)
         s.name = file_label
 
-        name_s = None
+        if not meta_added:
+            frames.append(meta)
+            meta_added = True
+
         if debug and name_column in df_idx.columns:
             name_s = df_idx[name_column]
             if not name_s.index.is_unique:
                 name_s = name_s.groupby(level=key_cols, dropna=False).first()
             name_s.name = f"{file_label}_name"
+            frames.append(name_s)
 
-        if union_index is None:
+        frames.append(s)
-            union_index = meta.index
-        else:
-            union_index = union_index.union(meta.index)
-        metas.append(meta)
-
-        staged.append((file_label, s, name_s))
-
-    if union_index is None:
-        raise ValueError("No data found after loading inputs.")
-
-    # meta first (union-aligned): build UNION meta across all files
-    if metas:
-        meta_union = pd.concat(metas, axis=0)
-        # Collapse duplicates on the MultiIndex; keep first non-null per column
-        meta_union = meta_union.groupby(level=key_cols, dropna=False).first()
-        frames.append(meta_union.reindex(union_index))
-
-    # values + ratios (union-aligned)
-    metric_series_aligned: list[pd.Series] = []
-    for file_label, s, name_s in staged:
-        s_aligned = s.reindex(union_index)
-        frames.append(s_aligned)
         raw_data_cols.append(file_label)
-        metric_series_aligned.append(s_aligned)
+        compare_frames.append(s)
 
-        if debug and name_s is not None:
+        if len(compare_frames) >= 2:
-            frames.append(name_s.reindex(union_index))
+            base = compare_frames[0]
-
+            current = compare_frames[-1]
-        if len(metric_series_aligned) >= 2:
+            if "P99" in data_column or "Median" in data_column:
-            base = metric_series_aligned[0]
-            current = metric_series_aligned[-1]
-            if "P99" in str(data_column) or "Median" in str(data_column):
                 ratio = base / current
             else:
                 ratio = current / base
             ratio = ratio.mask(base == 0)
-            ratio.name = f"Ratio 1 vs {len(metric_series_aligned)}"
+            ratio.name = f"Ratio 1 vs {len(compare_frames)}"
             frames.append(ratio)
 
     concat_df = pd.concat(frames, axis=1).reset_index(drop=True)
@@ -288,10 +202,24 @@ def split_json_by_tp_pp(
 # -----------------------------
 # Styling helpers
 # -----------------------------
+def _find_concurrency_col(df: pd.DataFrame) -> str:
+    for c in [
+        "# of max concurrency.",
+        "# of max concurrency",
+        "Max Concurrency",
+        "max_concurrency",
+        "Concurrency",
+    ]:
+        if c in df.columns:
+            return c
+    for c in df.columns:
+        if df[c].dtype.kind in "iu" and df[c].nunique() > 1 and df[c].min() >= 1:
+            return c
+    return "# of max concurrency."
+
+
 def _highlight_threshold(
-    df: pd.DataFrame,
+    df: pd.DataFrame, threshold: float
-    threshold: float,
-    slack_pct: float = 0.0,
 ) -> pd.io.formats.style.Styler:
     conc_col = _find_concurrency_col(df)
     key_cols = [
@@ -304,24 +232,12 @@ def _highlight_threshold(
     ]
     conf_cols = [c for c in conf_cols if pd.api.types.is_numeric_dtype(df[c])]
 
-    try:
+    return df.style.map(
-        slack_pct = float(slack_pct or 0.0)
+        lambda v: "background-color:#e6ffe6;font-weight:bold;"
-    except Exception:
+        if pd.notna(v) and v <= threshold
-        slack_pct = 0.0
+        else "",
-    slack_limit = threshold * (1.0 + slack_pct / 100.0)
+        subset=conf_cols,
-
+    )
-    def _cell(v):
-        if pd.isna(v):
-            return ""
-        if v <= threshold:
-            # Strict SLA
-            return "background-color:#e6ffe6;font-weight:bold;"
-        if v <= slack_limit:
-            # Within slack range
-            return "background-color:#ffe5cc;font-weight:bold;"
-        return ""
-
-    return df.style.map(_cell, subset=conf_cols)
 
 
 def highlight_ratio_columns(styler: pd.io.formats.style.Styler):
@@ -359,177 +275,6 @@ def _apply_two_decimals(
     return styler.format({c: "{:.2f}" for c in num_cols}, na_rep="")
 
 
-# -----------------------------
-# Export helpers (Excel + CSV)
-# -----------------------------
-def _sanitize_sheet_name(name: str) -> str:
-    """
-    Excel sheet constraints:
-      - max 31 chars
-      - cannot contain: : \ / ? * [ ]
-      - cannot be empty
-
-    NOTE: Use fast, non-regex operations here to avoid the third-party `regex`
-    module's compile overhead/edge-cases on some systems.
-    """
-    name = "sheet" if name is None else str(name)
-
-    # Replace illegal characters with underscore.
-    trans = str.maketrans(
-        {
-            ":": "_",
-            "\\": "_",
-            "/": "_",
-            "?": "_",
-            "*": "_",
-            "[": "_",
-            "]": "_",
-        }
-    )
-    name = name.translate(trans)
-
-    # Strip quotes/spaces and collapse whitespace.
-    name = name.strip().strip("'")
-    name = " ".join(name.split())
-
-    if not name:
-        name = "sheet"
-    return name[:31]
-
-
-def _group_to_sheet_base(group_cols: list[str], gkey_tuple) -> str:
-    d = dict(zip(group_cols, gkey_tuple))
-
-    # Always keep input/output lengths (these are important).
-    ilen = d.get("Input Len", "")
-    olen = d.get("Output Len", "")
-    lens = f"_{ilen}x{olen}" if ilen != "" and olen != "" else ""
-
-    # Shorten model name aggressively to make room for lens.
-    model = d.get("Model", "model")
-    leaf = str(model).split("/")[-1]
-
-    max_model_len = max(1, 31 - len(lens))
-    model_short = leaf[:max_model_len]
-
-    return _sanitize_sheet_name(f"{model_short}{lens}")
-
-
-def _write_tables_to_excel_sheet(
-    writer: pd.ExcelWriter, sheet: str, blocks: list[tuple[str, pd.DataFrame]]
-):
-    """Write all blocks to a sheet with a single to_excel() call.
-
-    Pandas+openpyxl can be extremely slow when called many times per sheet.
-    We flatten blocks into one table with a 'Section' column to keep structure
-    while making Excel generation fast and deterministic.
-    """
-    if not blocks:
-        pd.DataFrame().to_excel(writer, sheet_name=sheet, index=False)
-        return
-
-    combined_parts: list[pd.DataFrame] = []
-    for title, df in blocks:
-        df2 = df.copy()
-        # Put the section label as the first column for readability.
-        df2.insert(0, "Section", title)
-        combined_parts.append(df2)
-
-    combined = pd.concat(combined_parts, axis=0, ignore_index=True, sort=False)
-    combined.to_excel(writer, sheet_name=sheet, index=False)
-
-
-def _safe_filename(s: str) -> str:
-    # Fast path without the third-party `regex` module.
-    s = " ".join(str(s).strip().split())
-    allowed = []
-    for ch in s:
-        if ch.isalnum() or ch in "._-":
-            allowed.append(ch)
-        else:
-            allowed.append("_")
-    out = "".join(allowed)
-    return out[:180] if len(out) > 180 else out
-
-
-# -----------------------------
-# vLLM environment export helper
-# -----------------------------
-def _parse_vllm_env_txt(env_path: Path) -> pd.DataFrame:
-    """Parse vllm_env.txt into a flat table (Section, Key, Value).
-
-    Supports:
-      - section headers as standalone lines (no ':' or '=')
-      - key-value lines like 'OS: Ubuntu ...'
-      - env var lines like 'HF_HOME=/data/hf'
-    """
-    lines = env_path.read_text(encoding="utf-8", errors="replace").splitlines()
-    section = "General"
-    rows: list[dict] = []
-
-    def set_section(s: str):
-        nonlocal section
-        s = (s or "").strip()
-        if s:
-            section = s
-
-    for raw in lines:
-        stripped = raw.strip()
-        if not stripped:
-            continue
-        # divider lines like =====
-        if set(stripped) <= {"="}:
-            continue
-
-        # section header heuristic: short standalone line
-        if ":" not in stripped and "=" not in stripped and len(stripped) <= 64:
-            if stripped.lower().startswith("collecting environment information"):
-                continue
-            set_section(stripped)
-            continue
-
-        # env var style: KEY=VALUE (and not a URL with :)
-        if "=" in stripped and ":" not in stripped:
-            k, v = stripped.split("=", 1)
-            k = k.strip()
-            v = v.strip()
-            if k:
-                rows.append({"Section": section, "Key": k, "Value": v})
-            continue
-
-        # key: value
-        if ":" in stripped:
-            k, v = stripped.split(":", 1)
-            k = k.strip()
-            v = v.strip()
-            if k:
-                rows.append({"Section": section, "Key": k, "Value": v})
-            continue
-
-    return pd.DataFrame(rows, columns=["Section", "Key", "Value"])
-
-
-def _load_env_df_for_inputs(args, files: list[str]) -> pd.DataFrame | None:
-    """Load vllm_env.txt next to the *original* input JSON file.
-
-    Note: when only one -f is provided, the script may split JSON into ./splits/...,
-    but vllm_env.txt typically lives next to the original benchmark_results.json.
-    """
-    base_dir: Path | None = None
-    if getattr(args, "file", None):
-        base_dir = Path(args.file[0]).resolve().parent
-    elif files:
-        base_dir = Path(files[0]).resolve().parent
-    if base_dir is None:
-        return None
-
-    env_path = base_dir / "vllm_env.txt"
-    if not env_path.exists():
-        return None
-    df = _parse_vllm_env_txt(env_path)
-    return df
-
-
 # -----------------------------
 # Valid max concurrency summary helpers
 # -----------------------------
@@ -556,11 +301,7 @@ def _config_value_columns(df: pd.DataFrame, conc_col: str) -> list[str]:
 
 
 def _max_concurrency_ok(
-    df: pd.DataFrame,
+    df: pd.DataFrame, conc_col: str, cfg_col: str, threshold: float
-    conc_col: str,
-    cfg_col: str,
-    threshold: float,
-    slack_pct: float = 0.0,
 ):
     if df is None or conc_col not in df.columns or cfg_col not in df.columns:
         return pd.NA
@@ -573,14 +314,7 @@ def _max_concurrency_ok(
     if d.empty:
         return pd.NA
 
-    # Accept values up to (1 + slack_pct%) above the SLA.
+    ok = d[d[cfg_col] <= threshold]
-    try:
-        slack_pct = float(slack_pct or 0.0)
-    except Exception:
-        slack_pct = 0.0
-    effective_limit = float(threshold) * (1.0 + slack_pct / 100.0)
-
-    ok = d[d[cfg_col] <= effective_limit]
     if ok.empty:
         return pd.NA
 
@@ -646,25 +380,15 @@ def build_valid_max_concurrency_summary_html(
     if not cfg_cols:
         cfg_cols = sorted(set(ttft_cols) | set(tpot_cols) | set(tput_cols), key=str)
 
-    # Display SLA ranges in the table header (SLA .. SLA*(1+slack))
-    ttft_hi = args.ttft_max_ms * (1.0 + args.ttft_slack_pct / 100.0)
-    tpot_hi = args.tpot_max_ms * (1.0 + args.tpot_slack_pct / 100.0)
-    ttft_range = f"{args.ttft_max_ms:g}–{ttft_hi:g} ms (+{args.ttft_slack_pct:g}%)"
-    tpot_range = f"{args.tpot_max_ms:g}–{tpot_hi:g} ms (+{args.tpot_slack_pct:g}%)"
-
     rows = []
     for cfg in cfg_cols:
         ttft_max = (
-            _max_concurrency_ok(
+            _max_concurrency_ok(ttft_group_df, conc_col, cfg, args.ttft_max_ms)
-                ttft_group_df, conc_col, cfg, args.ttft_max_ms, args.ttft_slack_pct
-            )
             if ttft_group_df is not None
             else pd.NA
         )
         tpot_max = (
-            _max_concurrency_ok(
+            _max_concurrency_ok(tpot_group_df, conc_col, cfg, args.tpot_max_ms)
-                tpot_group_df, conc_col, cfg, args.tpot_max_ms, args.tpot_slack_pct
-            )
             if tpot_group_df is not None
             else pd.NA
         )
@@ -693,8 +417,8 @@ def build_valid_max_concurrency_summary_html(
         rows.append(
             {
                 "Configuration": cfg,
-                f"Max {conc_col} (TTFT ≤ {ttft_range})": ttft_max,
+                f"Max {conc_col} (TTFT ≤ {args.ttft_max_ms:g} ms)": ttft_max,
-                f"Max {conc_col} (TPOT ≤ {tpot_range})": tpot_max,
+                f"Max {conc_col} (TPOT ≤ {args.tpot_max_ms:g} ms)": tpot_max,
                 f"Max {conc_col} (Both)": both,
                 "Output Tput @ Both (tok/s)": tput_at_both,
                 "TTFT @ Both (ms)": ttft_at_both,
@@ -704,6 +428,7 @@ def build_valid_max_concurrency_summary_html(
 
     summary_df = pd.DataFrame(rows)
 
+    # --- Coerce numeric columns so Styler doesn't miss them due to object dtype ---
     for c in summary_df.columns:
         if c == "Configuration":
             continue
@@ -711,10 +436,12 @@ def build_valid_max_concurrency_summary_html(
 
     both_col = f"Max {conc_col} (Both)"
 
+    # --- Strict 2-decimal formatting for ALL non-Configuration columns ---
     formatters = {}
     for c in summary_df.columns:
         if c == "Configuration":
             continue
+        # default argument binds per-column formatter correctly
         formatters[c] = lambda v: "" if pd.isna(v) else f"{float(v):.2f}"
 
     styler = summary_df.style.format(formatters)
@@ -733,104 +460,6 @@ def build_valid_max_concurrency_summary_html(
     return title + styler.to_html(table_attributes='border="1" class="dataframe"')
 
 
-def build_valid_max_concurrency_summary_df(
-    tput_group_df: pd.DataFrame | None,
-    ttft_group_df: pd.DataFrame | None,
-    tpot_group_df: pd.DataFrame | None,
-    conc_col: str,
-    args,
-) -> pd.DataFrame | None:
-    if ttft_group_df is None and tpot_group_df is None:
-        return None
-
-    ttft_cols = (
-        _config_value_columns(ttft_group_df, conc_col)
-        if ttft_group_df is not None
-        else []
-    )
-    tpot_cols = (
-        _config_value_columns(tpot_group_df, conc_col)
-        if tpot_group_df is not None
-        else []
-    )
-    tput_cols = (
-        _config_value_columns(tput_group_df, conc_col)
-        if tput_group_df is not None
-        else []
-    )
-
-    if ttft_group_df is not None and tpot_group_df is not None:
-        cfg_cols = [c for c in ttft_cols if c in tpot_cols]
-        if tput_group_df is not None:
-            cfg_cols = [c for c in cfg_cols if c in tput_cols] or cfg_cols
-    else:
-        cfg_cols = ttft_cols or tpot_cols
-
-    if not cfg_cols:
-        cfg_cols = sorted(set(ttft_cols) | set(tpot_cols) | set(tput_cols), key=str)
-
-    ttft_hi = args.ttft_max_ms * (1.0 + args.ttft_slack_pct / 100.0)
-    tpot_hi = args.tpot_max_ms * (1.0 + args.tpot_slack_pct / 100.0)
-    ttft_range = f"{args.ttft_max_ms:g}–{ttft_hi:g} ms (+{args.ttft_slack_pct:g}%)"
-    tpot_range = f"{args.tpot_max_ms:g}–{tpot_hi:g} ms (+{args.tpot_slack_pct:g}%)"
-
-    rows = []
-    for cfg in cfg_cols:
-        ttft_max = (
-            _max_concurrency_ok(
-                ttft_group_df, conc_col, cfg, args.ttft_max_ms, args.ttft_slack_pct
-            )
-            if ttft_group_df is not None
-            else pd.NA
-        )
-        tpot_max = (
-            _max_concurrency_ok(
-                tpot_group_df, conc_col, cfg, args.tpot_max_ms, args.tpot_slack_pct
-            )
-            if tpot_group_df is not None
-            else pd.NA
-        )
-        both = (
-            pd.NA
-            if (pd.isna(ttft_max) or pd.isna(tpot_max))
-            else min(ttft_max, tpot_max)
-        )
-
-        tput_at_both = (
-            _value_at_concurrency(tput_group_df, conc_col, cfg, both)
-            if tput_group_df is not None
-            else pd.NA
-        )
-        ttft_at_both = (
-            _value_at_concurrency(ttft_group_df, conc_col, cfg, both)
-            if ttft_group_df is not None
-            else pd.NA
-        )
-        tpot_at_both = (
-            _value_at_concurrency(tpot_group_df, conc_col, cfg, both)
-            if tpot_group_df is not None
-            else pd.NA
-        )
-
-        rows.append(
-            {
-                "Configuration": cfg,
-                f"Max {conc_col} (TTFT ≤ {ttft_range})": ttft_max,
-                f"Max {conc_col} (TPOT ≤ {tpot_range})": tpot_max,
-                f"Max {conc_col} (Both)": both,
-                "Output Tput @ Both (tok/s)": tput_at_both,
-                "TTFT @ Both (ms)": ttft_at_both,
-                "TPOT @ Both (ms)": tpot_at_both,
-            }
-        )
-
-    df = pd.DataFrame(rows)
-    for c in df.columns:
-        if c != "Configuration":
-            df[c] = pd.to_numeric(df[c], errors="coerce")
-    return df
-
-
 # -----------------------------
 # Plot helper
 # -----------------------------
@@ -908,35 +537,6 @@ def build_parser() -> argparse.ArgumentParser:
         default=100.0,
         help="Reference limit for TPOT plots (ms)",
     )
-
-    # ---- SLA tolerance (slack) options ----
-    parser.add_argument(
-        "--ttft-slack-pct",
-        type=float,
-        default=5.0,
-        help="Allowed percentage above TTFT SLA (default: 5).",
-    )
-    parser.add_argument(
-        "--tpot-slack-pct",
-        type=float,
-        default=5.0,
-        help="Allowed percentage above TPOT SLA (default: 5).",
-    )
-
-    # ---- export options ----
-    parser.add_argument(
-        "--excel-out",
-        type=str,
-        default="perf_comparison.xlsx",
-        help="Write one sheet per (Model, Dataset, Input Len, Output Len).",
-    )
-    parser.add_argument(
-        "--csv-out-dir",
-        type=str,
-        default="",
-        help="If set, write per-group per-metric CSVs into this directory.",
-    )
-
     return parser
 
 
@@ -1015,13 +615,9 @@ def render_metric_table_html(
 
     metric_name = metric_label.lower()
     if "ttft" in metric_name:
-        styler = _highlight_threshold(
+        styler = _highlight_threshold(display_group, args.ttft_max_ms)
-            display_group, args.ttft_max_ms, args.ttft_slack_pct
-        )
     elif ("tpot" in metric_name) or ("median" in metric_name) or ("p99" in metric_name):
-        styler = _highlight_threshold(
+        styler = _highlight_threshold(display_group, args.tpot_max_ms)
-            display_group, args.tpot_max_ms, args.tpot_slack_pct
-        )
     else:
         styler = display_group.style
 
@@ -1061,6 +657,7 @@ def maybe_write_plot(
         markers=True,
     )
 
+    # Ensure plot hover + y tick labels are also 2 decimals.
     fig.update_traces(hovertemplate="%{y:.2f}<extra></extra>")
     fig.update_yaxes(tickformat=".2f")
 
@@ -1133,186 +730,87 @@ def write_report_group_first(
         for metric_label, (df, _) in metric_cache.items()
     }
 
-    csv_dir = Path(args.csv_out_dir) if args.csv_out_dir else None
+    with open("perf_comparison.html", "w", encoding="utf-8") as main_fh:
-    if csv_dir:
+        main_fh.write('<meta charset="utf-8">\n')
-        csv_dir.mkdir(parents=True, exist_ok=True)
+        for gkey in group_keys:
+            gkey_tuple = normalize_group_key(gkey)
+            suffix = build_group_suffix(group_cols_canonical, gkey_tuple)
+            sub_path = group_filename(gkey_tuple)
+            group_header = (
+                '<div style="font-size: 1.4em; font-weight: 700; '
+                'margin: 18px 0 10px 0;">'
+                f"{_html.escape(suffix)}"
+                "</div>\n"
+            )
 
-    excel_path = args.excel_out or "perf_comparison.xlsx"
+            main_fh.write(group_header)
-    disable_excel = os.getenv("VLLM_COMPARE_DISABLE_EXCEL", "0") == "1"
+            with open(sub_path, "w", encoding="utf-8") as sub_fh:
+                sub_fh.write('<meta charset="utf-8">\n')
+                sub_fh.write(group_header)
+                tput_group_df = None
+                ttft_group_df = None
+                tpot_group_df = None
+                conc_col = args.xaxis
 
-    # Prefer xlsxwriter for speed; fallback to openpyxl if unavailable.
+                for metric_label in plan.data_cols:
-    excel_engine = (
+                    gb = metric_groupbys[metric_label]
-        os.getenv("VLLM_COMPARE_EXCEL_ENGINE", "xlsxwriter").strip() or "xlsxwriter"
+                    df_sorted, raw_data_cols = metric_cache[metric_label]
-    )
-    if excel_engine == "xlsxwriter" and util.find_spec("xlsxwriter") is None:
-        excel_engine = "openpyxl"
 
-    excel_engine_kwargs = {}
+                    try:
-    if excel_engine == "xlsxwriter":
+                        group_df = gb.get_group(gkey)
-        # Reduce memory pressure & usually faster writes.
+                    except KeyError:
-        excel_engine_kwargs = {"options": {"constant_memory": True}}
+                        missing = (
+                            '<div style="font-size: 1.1em; font-weight: 600; '
+                            'margin: 10px 0;">'
+                            f"{_html.escape(metric_label)} — missing for this group"
+                            "</div>\n"
+                        )
 
-    xw_ctx = (
+                        main_fh.write(missing)
-        nullcontext(None)
+                        sub_fh.write(missing)
-        if disable_excel
+                        continue
-        else pd.ExcelWriter(
+
-            excel_path, engine=excel_engine, engine_kwargs=excel_engine_kwargs
+                    if conc_col not in group_df.columns:
-        )
+                        conc_col = _find_concurrency_col(group_df)
-    )
+
-    with xw_ctx as xw:
+                    mn = metric_label.lower().strip()
-        used_sheets: set[str] = set()
+                    if "tok/s" in mn:
-        # ---- Environment sheet (first) ----
+                        tput_group_df = group_df
-        env_sheet = _sanitize_sheet_name("Environment")
+                    elif "ttft" in mn:
-        env_df = _load_env_df_for_inputs(args, files)
+                        ttft_group_df = group_df
-        if xw is not None:
+                    elif mn in ("p99", "median") or "tpot" in mn:
-            if env_df is None or env_df.empty:
+                        tpot_group_df = group_df
-                pd.DataFrame(
+
-                    [
+                    display_group = group_df.drop(
-                        {
+                        columns=group_cols_canonical, errors="ignore"
-                            "Section": "Environment",
+                    )
-                            "Key": "vllm_env.txt",
+
-                            "Value": "NOT FOUND (or empty)",
+                    html = render_metric_table_html(
-                        }
+                        display_group, metric_label, suffix, args
-                    ]
+                    )
-                ).to_excel(xw, sheet_name=env_sheet, index=False)
+                    main_fh.write(html)
-            else:
+                    sub_fh.write(html)
-                env_df.to_excel(xw, sheet_name=env_sheet, index=False)
+
-            used_sheets.add(env_sheet)
+                    maybe_write_plot(
-        with open("perf_comparison.html", "w", encoding="utf-8") as main_fh:
+                        main_fh,
-            main_fh.write('<meta charset="utf-8">\n')
+                        sub_fh,
-            for gkey in group_keys:
+                        group_df=group_df,
-                gkey_tuple = normalize_group_key(gkey)
+                        raw_data_cols=raw_data_cols,
-                suffix = build_group_suffix(group_cols_canonical, gkey_tuple)
+                        metric_label=metric_label,
-                sub_path = group_filename(gkey_tuple)
+                        y_axis_col=y_axis_col,
-                group_header = (
+                        args=args,
-                    '<div style="font-size: 1.4em; font-weight: 700; '
+                    )
-                    'margin: 18px 0 10px 0;">'
+
-                    f"{_html.escape(suffix)}"
+                summary_html = build_valid_max_concurrency_summary_html(
-                    "</div>\n"
+                    tput_group_df=tput_group_df,
+                    ttft_group_df=ttft_group_df,
+                    tpot_group_df=tpot_group_df,
+                    conc_col=conc_col,
+                    args=args,
                 )
-
+                if summary_html:
-                main_fh.write(group_header)
+                    main_fh.write(summary_html)
-
+                    sub_fh.write(summary_html)
-                do_excel = xw is not None
-                sheet = _group_to_sheet_base(group_cols_canonical, gkey_tuple)
-                sheet_base = sheet
-                if do_excel:
-                    dedup_i = 1
-                    while sheet in used_sheets:
-                        dedup_i += 1
-                        suffix = f"_{dedup_i}"
-                        # Ensure uniqueness even when sheet names are truncated.
-                        base = str(sheet_base)
-                        keep = max(1, 31 - len(suffix))
-                        sheet = _sanitize_sheet_name(base[:keep] + suffix)
-                    used_sheets.add(sheet)
-
-                excel_blocks: list[tuple[str, pd.DataFrame]] = []
-
-                with open(sub_path, "w", encoding="utf-8") as sub_fh:
-                    sub_fh.write('<meta charset="utf-8">\n')
-                    sub_fh.write(group_header)
-                    tput_group_df = None
-                    ttft_group_df = None
-                    tpot_group_df = None
-                    conc_col = args.xaxis
-
-                    for metric_label in plan.data_cols:
-                        gb = metric_groupbys[metric_label]
-                        df_sorted, raw_data_cols = metric_cache[metric_label]
-
-                        try:
-                            group_df = gb.get_group(gkey)
-                        except KeyError:
-                            missing = (
-                                '<div style="font-size: 1.1em; font-weight: 600; '
-                                'margin: 10px 0;">'
-                                f"{_html.escape(metric_label)} — missing for this group"
-                                "</div>\n"
-                            )
-                            main_fh.write(missing)
-                            sub_fh.write(missing)
-                            continue
-
-                        if conc_col not in group_df.columns:
-                            conc_col = _find_concurrency_col(group_df)
-
-                        mn = metric_label.lower().strip()
-                        if "tok/s" in mn:
-                            tput_group_df = group_df
-                        elif "ttft" in mn:
-                            ttft_group_df = group_df
-                        elif mn in ("p99", "median") or "tpot" in mn:
-                            tpot_group_df = group_df
-
-                        display_group = group_df.drop(
-                            columns=group_cols_canonical, errors="ignore"
-                        )
-
-                        html = render_metric_table_html(
-                            display_group, metric_label, suffix, args
-                        )
-                        main_fh.write(html)
-                        sub_fh.write(html)
-
-                        maybe_write_plot(
-                            main_fh,
-                            sub_fh,
-                            group_df=group_df,
-                            raw_data_cols=raw_data_cols,
-                            metric_label=metric_label,
-                            y_axis_col=y_axis_col,
-                            args=args,
-                        )
-
-                        excel_blocks.append(
-                            (metric_label, group_df.reset_index(drop=True))
-                        )
-                        if csv_dir:
-                            fn = _safe_filename(
-                                f"{sheet}__{metric_label}".replace(" ", "_").replace(
-                                    "/", "_"
-                                )
-                            )
-                            group_df.to_csv(csv_dir / f"{fn}.csv", index=False)
-
-                    summary_html = build_valid_max_concurrency_summary_html(
-                        tput_group_df=tput_group_df,
-                        ttft_group_df=ttft_group_df,
-                        tpot_group_df=tpot_group_df,
-                        conc_col=conc_col,
-                        args=args,
-                    )
-                    if summary_html:
-                        main_fh.write(summary_html)
-                        sub_fh.write(summary_html)
-
-                    summary_df = build_valid_max_concurrency_summary_df(
-                        tput_group_df=tput_group_df,
-                        ttft_group_df=ttft_group_df,
-                        tpot_group_df=tpot_group_df,
-                        conc_col=conc_col,
-                        args=args,
-                    )
-                    if summary_df is not None:
-                        excel_blocks.append(
-                            ("Valid Max Concurrency Summary", summary_df)
-                        )
-                        if csv_dir:
-                            fn = _safe_filename(
-                                f"{sheet}__Valid_Max_Concurrency_Summary"
-                            )
-                            summary_df.to_csv(csv_dir / f"{fn}.csv", index=False)
-
-                if do_excel:
-                    _write_tables_to_excel_sheet(xw, sheet, excel_blocks)
-
-    if disable_excel:
-        print("Skipped Excel generation (VLLM_COMPARE_DISABLE_EXCEL=1).")
-    else:
-        print(f"Wrote Excel: {excel_path}")
-    if csv_dir:
-        print(f"Wrote CSVs under: {csv_dir}")
 
 
 def main():

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

@@ -1,4 +1,6 @@
 #!/bin/bash
+
+# This script should be run inside the CI process
 # This script assumes that we are already inside the vllm/ directory
 # Benchmarking results will be available inside vllm/benchmarks/results/
 
@@ -7,26 +9,14 @@
 set -x
 set -o pipefail
 
-# Environment-driven debug controls (like ON_CPU=1)
-DRY_RUN="${DRY_RUN:-0}"
-MODEL_FILTER="${MODEL_FILTER:-}"
-DTYPE_FILTER="${DTYPE_FILTER:-}"
-
-# Adaptive search controls
-ENABLE_ADAPTIVE_CONCURRENCY="${ENABLE_ADAPTIVE_CONCURRENCY:-0}"
-SLA_TTFT_MS="${SLA_TTFT_MS:-3000}"
-SLA_TPOT_MS="${SLA_TPOT_MS:-100}"
-ADAPTIVE_MAX_PROBES="${ADAPTIVE_MAX_PROBES:-8}"
-ADAPTIVE_MAX_CONCURRENCY="${ADAPTIVE_MAX_CONCURRENCY:-1024}"
-
 check_gpus() {
   if command -v nvidia-smi; then
     # check the number of GPUs and GPU type.
-    declare -g gpu_count=$(nvidia-smi --list-gpus | grep -c . || true)
+    declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
   elif command -v amd-smi; then
-    declare -g gpu_count=$(amd-smi list | grep -c 'GPU' || true)
+    declare -g gpu_count=$(amd-smi list | grep 'GPU' | wc -l)
   elif command -v hl-smi; then
-    declare -g gpu_count=$(hl-smi --list | grep -ci "Module ID" || true)
+    declare -g gpu_count=$(hl-smi --list | grep -i "Module ID" | wc -l)
   fi
 
   if [[ $gpu_count -gt 0 ]]; then
@@ -54,7 +44,7 @@ check_cpus() {
   declare -g numa_count=$(lscpu | grep "NUMA node(s):" | awk '{print $3}')
   if [[ $numa_count -gt 0 ]]; then
     echo "NUMA found."
-    echo "$numa_count"
+    echo $numa_count
   else
     echo "Need at least 1 NUMA to run benchmarking."
     exit 1
@@ -122,12 +112,13 @@ json2envs() {
 }
 
 wait_for_server() {
+  # wait for vllm server to start
+  # return 1 if vllm server crashes
   local timeout_val="1200"
   timeout "$timeout_val" bash -c '
-    until curl -sf http://localhost:8000/v1/models >/dev/null; do
+    until curl -X POST localhost:8000/v1/completions; do
       sleep 1
-    done
+    done' && return 0 || return 1
-  '
 }
 
 kill_processes_launched_by_current_bash() {
@@ -190,304 +181,6 @@ upload_to_buildkite() {
   $BUILDKITE_AGENT_COMMAND artifact upload "$RESULTS_FOLDER/*"
 }
 
-# -------------------------------
-# Adaptive concurrency helpers
-# -------------------------------
-result_json_path_for_serving() {
-  local test_name=$1
-  local qps=$2
-  local max_concurrency=$3
-  echo "$RESULTS_FOLDER/${test_name}_qps_${qps}_concurrency_${max_concurrency}.json"
-}
-
-extract_metric_ms() {
-  local metric_name=$1
-  local json_file=$2
-
-  [[ -f "$json_file" ]] || return 0
-
-  if [[ "$metric_name" == "ttft" ]]; then
-    jq -r '
-      [
-        .ttft_ms.p99?,
-        .metrics.ttft_ms.p99?,
-        .ttft.p99?,
-        .metrics.ttft.p99?,
-        .p99_ttft_ms?,
-        .ttft_ms.mean?,
-        .metrics.ttft_ms.mean?,
-        .ttft.mean?,
-        .metrics.ttft.mean?,
-        .mean_ttft_ms?
-      ] | map(select(. != null)) | .[0] // empty
-    ' "$json_file"
-  else
-    jq -r '
-      [
-        .tpot_ms.p99?,
-        .metrics.tpot_ms.p99?,
-        .tpot.p99?,
-        .metrics.tpot.p99?,
-        .p99_tpot_ms?,
-        .itl_ms.p99?,
-        .metrics.itl_ms.p99?,
-        .inter_token_latency_ms.p99?,
-        .tpot_ms.mean?,
-        .metrics.tpot_ms.mean?,
-        .tpot.mean?,
-        .metrics.tpot.mean?,
-        .itl_ms.mean?,
-        .metrics.itl_ms.mean?,
-        .mean_tpot_ms?,
-        .mean_itl_ms?
-      ] | map(select(. != null)) | .[0] // empty
-    ' "$json_file"
-  fi
-}
-
-evaluate_sla_from_json() {
-  local json_file=$1
-  local ttft
-  local tpot
-  local pass
-
-  [[ -f "$json_file" ]] || return 2
-
-  ttft=$(extract_metric_ms ttft "$json_file")
-  tpot=$(extract_metric_ms tpot "$json_file")
-
-  [[ -n "$ttft" && -n "$tpot" ]] || return 2
-
-  pass=$(jq -n \
-    --argjson ttft "$ttft" \
-    --argjson tpot "$tpot" \
-    --argjson sla_ttft "$SLA_TTFT_MS" \
-    --argjson sla_tpot "$SLA_TPOT_MS" \
-    '($ttft <= $sla_ttft) and ($tpot <= $sla_tpot)')
-
-  [[ "$pass" == "true" ]]
-}
-
-write_adaptive_summary_json() {
-  local summary_file=$1
-  local test_name=$2
-  local qps=$3
-  local static_last_pass=$4
-  local static_first_fail=$5
-  local final_last_pass=$6
-  local final_first_fail=$7
-
-  jq -n \
-    --arg test_name "$test_name" \
-    --arg qps "$qps" \
-    --argjson sla_ttft "$SLA_TTFT_MS" \
-    --argjson sla_tpot "$SLA_TPOT_MS" \
-    --arg static_last_pass "${static_last_pass:-}" \
-    --arg static_first_fail "${static_first_fail:-}" \
-    --arg final_last_pass "${final_last_pass:-}" \
-    --arg final_first_fail "${final_first_fail:-}" \
-    '{
-      test_name: $test_name,
-      qps: $qps,
-      sla_ttft_ms: $sla_ttft,
-      sla_tpot_ms: $sla_tpot,
-      static_last_pass: (if $static_last_pass == "" then null else ($static_last_pass | tonumber) end),
-      static_first_fail: (if $static_first_fail == "" then null else ($static_first_fail | tonumber) end),
-      final_last_pass: (if $final_last_pass == "" then null else ($final_last_pass | tonumber) end),
-      final_first_fail: (if $final_first_fail == "" then null else ($final_first_fail | tonumber) end)
-    }' > "$summary_file"
-}
-
-run_single_serving_probe() {
-  local test_name=$1
-  local qps=$2
-  local max_concurrency=$3
-  local tp=$4
-  local compilation_config_mode=$5
-  local optimization_level=$6
-  local client_args_effective=$7
-  local client_remote_args=$8
-  local server_command=$9
-
-  local new_test_name="${test_name}_qps_${qps}_concurrency_${max_concurrency}"
-  local result_json
-  local num_prompts_arg=""
-  local client_command
-
-  result_json=$(result_json_path_for_serving "$test_name" "$qps" "$max_concurrency")
-
-  if [[ -f "$result_json" ]]; then
-    evaluate_sla_from_json "$result_json"
-    return $?
-  fi
-
-  if [[ -n "${PROMPTS_PER_CONCURRENCY}" ]]; then
-    num_prompts=$(( max_concurrency * PROMPTS_PER_CONCURRENCY ))
-    if (( num_prompts < MIN_NUM_PROMPTS )); then num_prompts=$MIN_NUM_PROMPTS; fi
-    if (( num_prompts > MAX_NUM_PROMPTS )); then num_prompts=$MAX_NUM_PROMPTS; fi
-    num_prompts_arg="--num-prompts $num_prompts"
-  fi
-
-  client_command="vllm bench serve \
-    --save-result \
-    --result-dir $RESULTS_FOLDER \
-    --result-filename ${new_test_name}.json \
-    --request-rate $qps \
-    --max-concurrency $max_concurrency \
-    $num_prompts_arg \
-    --metadata tensor_parallel_size=$tp compilation_config.mode=$compilation_config_mode optimization_level=$optimization_level adaptive_search=1 \
-    $client_args_effective $client_remote_args "
-
-  echo "Adaptive probe: $client_command"
-
-  if [[ "${DRY_RUN:-0}" != "1" ]]; then
-    bash -c "$client_command"
-  fi
-
-  jq_output=$(jq -n \
-    --arg server "$server_command" \
-    --arg client "$client_command" \
-    --arg gpu "$gpu_type" \
-    '{
-      server_command: $server,
-      client_command: $client,
-      gpu_type: $gpu,
-      adaptive_search: true
-    }')
-  echo "$jq_output" > "$RESULTS_FOLDER/${new_test_name}.commands"
-
-  evaluate_sla_from_json "$result_json"
-}
-
-adaptive_refine_from_static_results() {
-  local test_name=$1
-  local qps=$2
-  local max_concurrency_list_raw=$3
-  local tp=$4
-  local compilation_config_mode=$5
-  local optimization_level=$6
-  local client_args_effective=$7
-  local client_remote_args=$8
-  local server_command=$9
-
-  local sorted_points
-  local point
-  local rc
-  local static_last_pass=""
-  local static_first_fail=""
-  local largest_static=""
-  local step_hint=1
-  local previous_point=""
-  local low
-  local high
-  local mid
-  local probes=0
-  local summary_file="$RESULTS_FOLDER/${test_name}_qps_${qps}_sla_summary.json"
-
-  [[ "${ENABLE_ADAPTIVE_CONCURRENCY}" == "1" ]] || return 0
-  [[ "${DRY_RUN:-0}" != "1" ]] || return 0
-
-  sorted_points=$(for point in $max_concurrency_list_raw; do printf '%s\n' "$point"; done | tr -d "'" | awk '/^[0-9]+$/' | sort -n | uniq)
-  [[ -n "$sorted_points" ]] || return 0
-
-  while read -r point; do
-    [[ -z "$point" ]] && continue
-    largest_static="$point"
-    evaluate_sla_from_json "$(result_json_path_for_serving "$test_name" "$qps" "$point")"
-    rc=$?
-    if (( rc == 0 )); then
-      static_last_pass="$point"
-    elif (( rc == 1 )); then
-      if [[ -n "$static_last_pass" ]]; then
-        static_first_fail="$point"
-        break
-      fi
-    fi
-
-    if [[ -n "$previous_point" ]]; then
-      step_hint=$(( point - previous_point ))
-      if (( step_hint < 1 )); then step_hint=1; fi
-    fi
-    previous_point="$point"
-  done <<< "$sorted_points"
-
-  if [[ -z "$static_last_pass" ]]; then
-    write_adaptive_summary_json "$summary_file" "$test_name" "$qps" "" "$static_first_fail" "" "$static_first_fail"
-    return 0
-  fi
-
-  if [[ -n "$static_first_fail" ]]; then
-    low=$static_last_pass
-    high=$static_first_fail
-    while (( low + 1 < high )) && (( probes < ADAPTIVE_MAX_PROBES )); do
-      mid=$(( (low + high) / 2 ))
-      probes=$(( probes + 1 ))
-      run_single_serving_probe \
-        "$test_name" "$qps" "$mid" "$tp" \
-        "$compilation_config_mode" "$optimization_level" \
-        "$client_args_effective" "$client_remote_args" "$server_command"
-      rc=$?
-      if (( rc == 0 )); then
-        low=$mid
-      elif (( rc == 1 )); then
-        high=$mid
-      else
-        break
-      fi
-    done
-    write_adaptive_summary_json "$summary_file" "$test_name" "$qps" "$static_last_pass" "$static_first_fail" "$low" "$high"
-    return 0
-  fi
-
-  low=$largest_static
-  high=""
-  while (( probes < ADAPTIVE_MAX_PROBES )); do
-    point=$(( low + step_hint ))
-    if (( point > ADAPTIVE_MAX_CONCURRENCY )); then
-      point=$ADAPTIVE_MAX_CONCURRENCY
-    fi
-    (( point > low )) || break
-    probes=$(( probes + 1 ))
-    run_single_serving_probe \
-      "$test_name" "$qps" "$point" "$tp" \
-      "$compilation_config_mode" "$optimization_level" \
-      "$client_args_effective" "$client_remote_args" "$server_command"
-    rc=$?
-    if (( rc == 0 )); then
-      low=$point
-      (( point == ADAPTIVE_MAX_CONCURRENCY )) && break
-      step_hint=$(( step_hint * 2 ))
-      if (( step_hint < 1 )); then step_hint=1; fi
-    elif (( rc == 1 )); then
-      high=$point
-      break
-    else
-      break
-    fi
-  done
-
-  if [[ -n "$high" ]]; then
-    while (( low + 1 < high )) && (( probes < ADAPTIVE_MAX_PROBES )); do
-      mid=$(( (low + high) / 2 ))
-      probes=$(( probes + 1 ))
-      run_single_serving_probe \
-        "$test_name" "$qps" "$mid" "$tp" \
-        "$compilation_config_mode" "$optimization_level" \
-        "$client_args_effective" "$client_remote_args" "$server_command"
-      rc=$?
-      if (( rc == 0 )); then
-        low=$mid
-      elif (( rc == 1 )); then
-        high=$mid
-      else
-        break
-      fi
-    done
-  fi
-
-  write_adaptive_summary_json "$summary_file" "$test_name" "$qps" "$static_last_pass" "" "$low" "$high"
-}
-
 run_benchmark_tests() {
   # run benchmark tests using `vllm bench <test_type>` command
   # $1: test type (latency or throughput)
@@ -559,16 +252,37 @@ run_benchmark_tests() {
   done
 }
 
-run_latency_tests() { run_benchmark_tests "latency" "$1"; }
+run_latency_tests() {
-run_startup_tests() { run_benchmark_tests "startup" "$1"; }
+  run_benchmark_tests "latency" "$1"
-run_throughput_tests() { run_benchmark_tests "throughput" "$1"; }
+}
 
-merge_serving_tests_stream() {
+run_startup_tests() {
-  # Emit merged serving test objects, optionally filtered by MODEL_FILTER/DTYPE_FILTER in DRY_RUN mode.
+  run_benchmark_tests "startup" "$1"
-  # This helper does NOT modify JSON; it only filters the stream in dry-run mode.
+}
-  local serving_test_file="$1"
+
-  # shellcheck disable=SC2016
+run_throughput_tests() {
-  local merged='
+  run_benchmark_tests "throughput" "$1"
+}
+
+run_serving_tests() {
+  # run serving tests using `vllm bench serve` command
+  # $1: a json file specifying serving test cases
+  #
+  # Supported JSON formats:
+  # 1) Plain format: top-level array
+  #    [ { "test_name": "...", "server_parameters": {...}, ... }, ... ]
+  #
+  # 2) Default parameters field + plain format tests
+  #    {
+  #      "defaults": { ... },
+  #      "tests": [ { "test_name": "...", "server_parameters": {...}, ... }, ... ]
+  #    }
+
+  local serving_test_file
+  serving_test_file=$1
+
+  # Iterate over serving tests
+  jq -c '
     if type == "array" then
       # Plain format: test cases array
       .[]
@@ -590,50 +304,7 @@ merge_serving_tests_stream() {
     else
       error("Unsupported serving test file format: must be array or object with .tests")
     end
-  '
+  ' "$serving_test_file" | while read -r params; do
-
-  jq -c "$merged" "$serving_test_file" | \
-  if [[ "${DRY_RUN:-0}" == "1" && ( "${MODEL_FILTER}${DTYPE_FILTER}" != "" ) ]]; then
-    jq -c --arg model "$MODEL_FILTER" --arg dtype "$DTYPE_FILTER" '
-      select((($model|length)==0)
-             or ((.server_parameters.model // "") == $model)
-             or ((.client_parameters.model // "") == $model))
-      | select((($dtype|length)==0) or ((.server_parameters.dtype // "") == $dtype))
-    '
-  else
-    cat
-  fi
-}
-
-run_serving_tests() {
-  # run serving tests using `vllm bench serve` command
-  # $1: a json file specifying serving test cases
-  #
-  # Supported JSON formats:
-  # 1) Plain format: top-level array
-  #    [ { "test_name": "...", "server_parameters": {...}, ... }, ... ]
-  #
-  # 2) Default parameters field + plain format tests
-  #    {
-  #      "defaults": { ... },
-  #      "tests": [ { "test_name": "...", "server_parameters": {...}, ... }, ... ]
-  #    }
-
-  local serving_test_file
-  serving_test_file=$1
-
-  # In dry-run mode, if filters are provided but no tests match, fail fast.
-  if [[ "${DRY_RUN:-0}" == "1" && ( "${MODEL_FILTER}${DTYPE_FILTER}" != "" ) ]]; then
-    local count
-    count=$(merge_serving_tests_stream "$serving_test_file" | wc -l | tr -d ' ')
-    if [[ "$count" -eq 0 ]]; then
-      echo "No matching serving tests found in $serving_test_file for model='$MODEL_FILTER' dtype='$DTYPE_FILTER'." >&2
-      return 0
-    fi
-  fi
-
-  # Iterate over serving tests (merged + optional filtered stream)
-  merge_serving_tests_stream "$serving_test_file" | while read -r params; do
     # get the test name, and append the GPU type back to it.
     test_name=$(echo "$params" | jq -r '.test_name')
     if [[ ! "$test_name" =~ ^serving_ ]]; then
@@ -652,48 +323,10 @@ run_serving_tests() {
     server_envs=$(echo "$params" | jq -r '.server_environment_variables')
     client_params=$(echo "$params" | jq -r '.client_parameters')
 
-    # vLLM serve CLI: model must be positional (no --model). Convert server_parameters accordingly.
+    server_args=$(json2args "$server_params")
-    server_model=$(echo "$server_params" | jq -r '.model // empty')
-    if [[ -z "$server_model" || "$server_model" == "null" ]]; then
-      echo "Error: serving test '$test_name' is missing server_parameters.model" >&2
-      exit 1
-    fi
-    server_params_no_model=$(echo "$server_params" | jq -c 'del(.model)')
-    server_args=$(json2args "$server_params_no_model")
-
     server_envs=$(json2envs "$server_envs")
     client_args=$(json2args "$client_params")
 
-    # ------------------------------------------------------------
-    # Option 1: Dynamic num-prompts scaling based on max_concurrency
-    #
-    # If PROMPTS_PER_CONCURRENCY is set, override JSON num_prompts with:
-    #   num_prompts = max_concurrency * PROMPTS_PER_CONCURRENCY
-    #
-    # If PROMPTS_PER_CONCURRENCY is NOT set, keep JSON num_prompts behavior
-    # unchanged (i.e., whatever is in serving-tests-*.json).
-    # ------------------------------------------------------------
-    PROMPTS_PER_CONCURRENCY="${PROMPTS_PER_CONCURRENCY-}"  # no default on purpose
-    MIN_NUM_PROMPTS="${MIN_NUM_PROMPTS:-1}"
-    MAX_NUM_PROMPTS="${MAX_NUM_PROMPTS:-1000000}"
-
-    if [[ -n "${PROMPTS_PER_CONCURRENCY}" ]]; then
-      # Remove any fixed --num-prompts from JSON-derived args (avoid duplicates)
-      # Remove any fixed --num-prompts from JSON-derived args (avoid duplicates)
-      # Handles: --num-prompts 123   and   --num-prompts=123
-      client_args_no_np="$(
-        printf ' %s ' "$client_args" \
-        | sed -E \
-          -e 's/[[:space:]]--num-prompts=([^[:space:]]+)([[:space:]]|$)/ /g' \
-          -e 's/[[:space:]]--num-prompts[[:space:]]+([^[:space:]]+)([[:space:]]|$)/ /g'
-      )"
-      # normalize whitespace
-      client_args_no_np="$(echo "$client_args_no_np" | tr -s ' ' | sed -E 's/^ //; s/ $//')"
-      client_args_no_np="$(echo "$client_args_no_np" | xargs)"
-      client_args_effective="$client_args_no_np"
-    else
-      client_args_effective="$client_args"
-    fi
     # qps_list
     qps_list=$(echo "$params" | jq -r '.qps_list')
     qps_list=$(echo "$qps_list" | jq -r '.[] | @sh')
@@ -725,13 +358,14 @@ run_serving_tests() {
     fi
 
     # check if server model and client model is aligned
+    server_model=$(echo "$server_params" | jq -r '.model')
     client_model=$(echo "$client_params" | jq -r '.model')
     if [[ $server_model != "$client_model" ]]; then
       echo "Server model and client model must be the same. Skip testcase $test_name."
       continue
     fi
 
-    server_command="$server_envs vllm serve $server_model \
+    server_command="$server_envs vllm serve \
       $server_args"
 
     # run the server
@@ -739,7 +373,7 @@ run_serving_tests() {
     echo "Server command: $server_command"
     # support remote vllm server
     client_remote_args=""
-    if [[ -z "${REMOTE_HOST}" && "${DRY_RUN:-0}" != "1" ]]; then
+    if [[ -z "${REMOTE_HOST}" ]]; then
       bash -c "$server_command" &
       server_pid=$!
       # wait until the server is alive
@@ -750,9 +384,6 @@ run_serving_tests() {
         echo ""
         echo "vLLM failed to start within the timeout period."
       fi
-    elif [[ "${DRY_RUN:-0}" == "1" ]]; then
-        # dry-run: don't start server
-        echo "Dry Run."
     else
       server_command="Using Remote Server $REMOTE_HOST $REMOTE_PORT"
       if [[ ${REMOTE_PORT} ]]; then
@@ -771,21 +402,15 @@ run_serving_tests() {
     for qps in $qps_list; do
       # remove the surrounding single quote from qps
       if [[ "$qps" == *"inf"* ]]; then
+        echo "qps was $qps"
         qps="inf"
+        echo "now qps is $qps"
       fi
 
       # iterate over different max_concurrency
       for max_concurrency in $max_concurrency_list; do
-        new_test_name="${test_name}_qps_${qps}_concurrency_${max_concurrency}"
+        new_test_name=$test_name"_qps_"$qps"_concurrency_"$max_concurrency
         echo " new test name $new_test_name"
-        # If PROMPTS_PER_CONCURRENCY is set, compute per-concurrency --num-prompts.
-        num_prompts_arg=""
-        if [[ -n "${PROMPTS_PER_CONCURRENCY}" ]]; then
-          num_prompts=$(( max_concurrency * PROMPTS_PER_CONCURRENCY ))
-          if (( num_prompts < MIN_NUM_PROMPTS )); then num_prompts=$MIN_NUM_PROMPTS; fi
-          if (( num_prompts > MAX_NUM_PROMPTS )); then num_prompts=$MAX_NUM_PROMPTS; fi
-          num_prompts_arg="--num-prompts $num_prompts"
-        fi
         # pass the tensor parallel size, the compilation mode, and the optimization
         # level to the client so that they can be used on the benchmark dashboard
         client_command="vllm bench serve \
@@ -794,16 +419,13 @@ run_serving_tests() {
           --result-filename ${new_test_name}.json \
           --request-rate $qps \
           --max-concurrency $max_concurrency \
-          $num_prompts_arg \
           --metadata tensor_parallel_size=$tp compilation_config.mode=$compilation_config_mode optimization_level=$optimization_level \
-          $client_args_effective $client_remote_args "
+          $client_args $client_remote_args "
 
         echo "Running test case $test_name with qps $qps"
         echo "Client command: $client_command"
 
-        if [[ "${DRY_RUN:-0}" != "1" ]]; then
+        bash -c "$client_command"
-          bash -c "$client_command"
-        fi
 
         # record the benchmarking commands
         jq_output=$(jq -n \
@@ -818,23 +440,15 @@ run_serving_tests() {
         echo "$jq_output" >"$RESULTS_FOLDER/${new_test_name}.commands"
 
       done
-
-      adaptive_refine_from_static_results \
-        "$test_name" "$qps" "$max_concurrency_list" "$tp" \
-        "$compilation_config_mode" "$optimization_level" \
-        "$client_args_effective" "$client_remote_args" "$server_command"
     done
 
     # clean up
-    if [[ "${DRY_RUN:-0}" != "1" ]]; then
+    kill -9 $server_pid
-      kill -9 "$server_pid"
+    kill_gpu_processes
-      kill_gpu_processes
-    fi
   done
 }
 
 main() {
-
   local ARCH
   ARCH=''
   if [[ "$ON_CPU" == "1" ]]; then
@@ -844,13 +458,7 @@ main() {
      check_gpus
      ARCH="$arch_suffix"
   fi
-
+  check_hf_token
-  # DRY_RUN does not execute vLLM; do not require HF_TOKEN.
-  if [[ "${DRY_RUN:-0}" != "1" ]]; then
-    check_hf_token
-  else
-    echo "DRY_RUN=1 -> skip HF_TOKEN validation"
-  fi
 
   # dependencies
   (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
@@ -871,16 +479,11 @@ main() {
 
   # dump vllm info via vllm collect-env
   env_output=$(vllm collect-env)
+
   echo "$env_output" >"$RESULTS_FOLDER/vllm_env.txt"
 
   # benchmarking
-  run_serving_tests $QUICK_BENCHMARK_ROOT/tests/"${SERVING_JSON:-serving-tests$ARCH.json}" || exit $?
+  run_serving_tests $QUICK_BENCHMARK_ROOT/tests/"${SERVING_JSON:-serving-tests$ARCH.json}"
-
-  if [[ "${DRY_RUN:-0}" == "1" ]]; then
-    echo "DRY_RUN=1 -> skip latency/startup/throughput suites"
-    exit 0
-  fi
-
   run_latency_tests $QUICK_BENCHMARK_ROOT/tests/"${LATENCY_JSON:-latency-tests$ARCH.json}"
   run_startup_tests $QUICK_BENCHMARK_ROOT/tests/"${STARTUP_JSON:-startup-tests$ARCH.json}"
   run_throughput_tests $QUICK_BENCHMARK_ROOT/tests/"${THROUGHPUT_JSON:-throughput-tests$ARCH.json}"
@@ -888,7 +491,6 @@ main() {
   # postprocess benchmarking results
   pip install tabulate pandas
   python3 $QUICK_BENCHMARK_ROOT/scripts/convert-results-json-to-markdown.py
-  python3 $QUICK_BENCHMARK_ROOT/scripts/compare-json-results.py -f $RESULTS_FOLDER/benchmark_results.json
 
   upload_to_buildkite
 }

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

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

.buildkite/performance-benchmarks/tests/serving-tests-cpu-asr.json

@@ -1,37 +0,0 @@
-{
-  "defaults": {
-    "qps_list": [
-      "inf"
-    ],
-    "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-    "server_environment_variables": {
-      "VLLM_RPC_TIMEOUT": 100000,
-      "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120
-    },
-    "server_parameters": {
-      "dtype": "bfloat16",
-      "model": "openai/whisper-large-v3-turbo"
-    },
-    "client_parameters": {
-      "model": "openai/whisper-large-v3-turbo",
-      "backend": "openai-audio",
-      "endpoint": "/v1/audio/transcriptions",
-      "dataset_name": "hf",
-      "dataset_path": "openslr/librispeech_asr",
-      "hf_subset": "clean",
-      "hf_split": "test",
-      "no_stream": "",
-      "no_oversample": "",
-      "num_prompts": 200
-    }
-  },
-  "tests": [
-    {
-      "test_name": "serving_whisper_large_v3_turbo_librispeech_clean_tp1",
-      "server_parameters": {
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {}
-    }
-  ]
-}

.buildkite/performance-benchmarks/tests/serving-tests-cpu-embed.json

@@ -1,41 +0,0 @@
-{
-  "defaults": {
-    "qps_list": [
-      "inf"
-    ],
-    "max_concurrency_list": [
-      32,
-      64,
-      128
-    ],
-    "server_environment_variables": {
-      "VLLM_RPC_TIMEOUT": 100000,
-      "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-      "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-      "VLLM_CPU_SGL_KERNEL": 1,
-      "VLLM_CPU_KVCACHE_SPACE": 40
-    },
-    "server_parameters": {
-      "dtype": "bfloat16",
-      "model": "jinaai/jina-embeddings-v3",
-      "trust_remote_code": ""
-    },
-    "client_parameters": {
-      "model": "jinaai/jina-embeddings-v3",
-      "backend": "openai-embeddings",
-      "endpoint": "/v1/embeddings",
-      "dataset_name": "sharegpt",
-      "dataset_path": "ShareGPT_V3_unfiltered_cleaned_split.json",
-      "num_prompts": 200
-    }
-  },
-  "tests": [
-    {
-      "test_name": "serving_jina_embed_v3_tp1_sharegpt",
-      "server_parameters": {
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {}
-    }
-  ]
-}

.buildkite/performance-benchmarks/tests/serving-tests-cpu-text.json

@@ -1,355 +0,0 @@
-{
-  "defaults": {
-    "qps_list": [
-      "inf"
-    ],
-    "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
-    "server_environment_variables": {
-      "VLLM_RPC_TIMEOUT": 100000,
-      "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
-      "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
-      "VLLM_CPU_SGL_KERNEL": 1,
-      "VLLM_CPU_KVCACHE_SPACE": 40
-    },
-    "server_parameters": {
-      "model": "meta-llama/Llama-3.1-8B-Instruct",
-      "tensor_parallel_size": 1,
-      "dtype": "bfloat16",
-      "distributed_executor_backend": "mp",
-      "block_size": 128,
-      "trust_remote_code": "",
-      "disable_log_stats": "",
-      "max_num_batched_tokens": 2048,
-      "max_num_seqs": 256
-    },
-    "client_parameters": {
-      "model": "meta-llama/Llama-3.1-8B-Instruct",
-      "backend": "vllm",
-      "ignore-eos": "",
-      "num_prompts": 200
-    }
-  },
-  "tests": [
-    {
-      "test_name": "serving_llama8B_tp1_sharegpt",
-      "server_parameters": {
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "dataset_name": "sharegpt",
-        "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp2_sharegpt",
-      "server_parameters": {
-        "tensor_parallel_size": 2
-      },
-      "client_parameters": {
-        "dataset_name": "sharegpt",
-        "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp1_random_128_128",
-      "server_parameters": {
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp2_random_128_128",
-      "server_parameters": {
-        "tensor_parallel_size": 2
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp4_random_128_128",
-      "server_parameters": {
-        "tensor_parallel_size": 4
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp1_random_128_2048",
-      "server_parameters": {
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 2048
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp2_random_128_2048",
-      "server_parameters": {
-        "tensor_parallel_size": 2
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 2048
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp4_random_128_2048",
-      "server_parameters": {
-        "tensor_parallel_size": 4
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 2048
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp1_random_2048_128",
-      "server_parameters": {
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 2048,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp2_random_2048_128",
-      "server_parameters": {
-        "tensor_parallel_size": 2
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 2048,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp4_random_2048_128",
-      "server_parameters": {
-        "tensor_parallel_size": 4
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 2048,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp1_random_2048_2048",
-      "server_parameters": {
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 2048,
-        "random-output-len": 2048
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp2_random_2048_2048",
-      "server_parameters": {
-        "tensor_parallel_size": 2
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 2048,
-        "random-output-len": 2048
-      }
-    },
-    {
-      "test_name": "serving_llama8B_tp4_random_2048_2048",
-      "server_parameters": {
-        "tensor_parallel_size": 4
-      },
-      "client_parameters": {
-        "dataset_name": "random",
-        "random-input-len": 2048,
-        "random-output-len": 2048
-      }
-    },
-    {
-      "test_name": "serving_llama8B_int4_tp1_random_128_128",
-      "server_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_int4_tp2_random_128_128",
-      "server_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "tensor_parallel_size": 2
-      },
-      "client_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_int4_tp4_random_128_128",
-      "server_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "tensor_parallel_size": 4
-      },
-      "client_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_int8_tp1_random_128_128",
-      "server_parameters": {
-        "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_int8_tp2_random_128_128",
-      "server_parameters": {
-        "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-        "tensor_parallel_size": 2
-      },
-      "client_parameters": {
-        "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_int8_tp4_random_128_128",
-      "server_parameters": {
-        "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-        "tensor_parallel_size": 4
-      },
-      "client_parameters": {
-        "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama3B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "meta-llama/Llama-3.2-3B-Instruct",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "meta-llama/Llama-3.2-3B-Instruct",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_granite2B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "ibm-granite/granite-3.2-2b-instruct",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "ibm-granite/granite-3.2-2b-instruct",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_qwen1.7B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "Qwen/Qwen3-1.7B",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "Qwen/Qwen3-1.7B",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_qwen4B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "Qwen/Qwen3-4B",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "Qwen/Qwen3-4B",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_qwen8B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "Qwen/Qwen3-8B",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "Qwen/Qwen3-8B",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_glm9B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "zai-org/glm-4-9b-hf",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "zai-org/glm-4-9b-hf",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_gemma7B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "google/gemma-7b",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "google/gemma-7b",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    }
-  ]
-}

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

@@ -72,6 +72,17 @@
         "random-output-len": 128
       }
     },
+    {
+      "test_name": "serving_llama8B_tp4_random_128_128",
+      "server_parameters": {
+        "tensor_parallel_size": 4
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
     {
       "test_name": "serving_llama8B_tp1_random_128_2048",
       "server_parameters": {
@@ -94,6 +105,17 @@
         "random-output-len": 2048
       }
     },
+    {
+      "test_name": "serving_llama8B_tp4_random_128_2048",
+      "server_parameters": {
+        "tensor_parallel_size": 4
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 2048
+      }
+    },
     {
       "test_name": "serving_llama8B_tp1_random_2048_128",
       "server_parameters": {
@@ -117,25 +139,144 @@
       }
     },
     {
-      "test_name": "serving_llama8B_tp1_random_2048_2048",
+      "test_name": "serving_llama8B_tp4_random_2048_128",
       "server_parameters": {
-        "tensor_parallel_size": 1
+        "tensor_parallel_size": 4
       },
       "client_parameters": {
         "dataset_name": "random",
         "random-input-len": 2048,
-        "random-output-len": 2048
+        "random-output-len": 128
       }
     },
     {
-      "test_name": "serving_llama8B_tp2_random_2048_2048",
+      "test_name": "serving_llama8B_int4_tp1_random_128_128",
       "server_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_int4_tp2_random_128_128",
+      "server_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
         "tensor_parallel_size": 2
       },
       "client_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
         "dataset_name": "random",
-        "random-input-len": 2048,
+        "random-input-len": 128,
-        "random-output-len": 2048
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_int4_tp4_random_128_128",
+      "server_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "tensor_parallel_size": 4
+      },
+      "client_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama3B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "meta-llama/Llama-3.2-3B-Instruct",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "meta-llama/Llama-3.2-3B-Instruct",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_granite2B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "ibm-granite/granite-3.2-2b-instruct",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "ibm-granite/granite-3.2-2b-instruct",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_qwen1.7B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "Qwen/Qwen3-1.7B",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "Qwen/Qwen3-1.7B",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_qwen4B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "Qwen/Qwen3-4B",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "Qwen/Qwen3-4B",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_qwen8B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "Qwen/Qwen3-8B",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "Qwen/Qwen3-8B",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_glm9B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "zai-org/glm-4-9b-hf",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "zai-org/glm-4-9b-hf",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_gemma7B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "google/gemma-7b",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "google/gemma-7b",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
       }
     }
   ]

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

@@ -10,6 +10,7 @@
         "server_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 1,
+            "swap_space": 16,
             "disable_log_stats": "",
             "load_format": "dummy",
             "max-model-len": 2048,
@@ -36,6 +37,7 @@
         "server_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
             "tensor_parallel_size": 4,
+            "swap_space": 16,
             "disable_log_stats": "",
             "load_format": "dummy",
             "max-model-len": 2048,
@@ -62,6 +64,7 @@
         "server_parameters": {
             "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
             "tensor_parallel_size": 2,
+            "swap_space": 16,
             "disable_log_stats": "",
             "load_format": "dummy",
             "max-model-len": 2048,
@@ -75,83 +78,5 @@
             "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
             "num_prompts": 200
         }
-    },
-    {
-        "test_name": "serving_deepseek_r1",
-        "qps_list": [1, 4, 16, "inf"],
-        "server_environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "server_parameters": {
-            "model": "deepseek-ai/DeepSeek-R1",
-            "tensor_parallel_size": 8,
-            "disable_log_stats": "",
-            "load_format": "dummy",
-            "max-model-len": 2048,
-            "max-num-seqs": 200,
-            "async-scheduling": "",
-            "dtype": "bfloat16"
-        },
-        "client_parameters": {
-            "model": "deepseek-ai/DeepSeek-R1",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_llama4_maverick_17b128e_instruct_fp8",
-        "qps_list": [1, 4, 16, "inf"],
-        "server_environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "server_parameters": {
-            "model": "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
-            "tensor_parallel_size": 8,
-            "disable_log_stats": "",
-            "max-model-len": 2048,
-            "max-num-seqs": 128,
-            "async-scheduling": "",
-            "enable_expert_parallel": "",
-            "max-num-batched-tokens": 4096
-        },
-        "client_parameters": {
-            "model": "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
-    },
-    {
-        "test_name": "serving_qwen3_8b",
-        "qps_list": [1, 4, 10, "inf"],
-        "server_environment_variables": {
-            "PT_HPU_LAZY_MODE": 1,
-            "PT_HPU_ENABLE_LAZY_COLLECTIVES": 1,
-            "VLLM_CONTIGUOUS_PA": 1,
-            "VLLM_DEFRAG": 1
-        },
-        "server_parameters": {
-            "model": "Qwen/Qwen-3-8B",
-            "tensor_parallel_size": 1,
-            "dtype": "bfloat16",
-            "disable_log_stats": "",
-            "async-scheduling": ""
-        },
-        "client_parameters": {
-            "model": "Qwen/Qwen-3-8B",
-            "backend": "vllm",
-            "dataset_name": "sharegpt",
-            "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
-            "num_prompts": 200
-        }
     }
 ]

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

@@ -5,6 +5,7 @@
         "server_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 1,
+            "swap_space": 16,
             "disable_log_stats": "",
             "load_format": "dummy"
         },
@@ -22,6 +23,7 @@
         "server_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
             "tensor_parallel_size": 4,
+            "swap_space": 16,
             "disable_log_stats": "",
             "load_format": "dummy"
         },
@@ -39,6 +41,7 @@
         "server_parameters": {
             "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
             "tensor_parallel_size": 2,
+            "swap_space": 16,
             "disable_log_stats": "",
             "load_format": "dummy"
         },
@@ -56,6 +59,7 @@
         "server_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct", 
             "tensor_parallel_size": 4,
+            "swap_space": 16,
             "speculative_config": {
                 "model": "turboderp/Qwama-0.5B-Instruct",
                 "num_speculative_tokens": 4,

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

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

.buildkite/release-pipeline.yaml

@@ -83,7 +83,7 @@ steps:
         agents:
           queue: cpu_queue_postmerge
         commands:
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_X86=true --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_AMXBF16=true --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
           - "mkdir artifacts"
           - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
           - "bash .buildkite/scripts/upload-nightly-wheels.sh manylinux_2_35"
@@ -152,7 +152,7 @@ steps:
           queue: cpu_queue_postmerge
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_X86=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_AMXBF16=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
           - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest"
           - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
         env:

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

@@ -25,7 +25,7 @@ S3_REGION="${AWS_DEFAULT_REGION:-us-west-2}"
 S3_URL="http://${S3_BUCKET}.s3-website-${S3_REGION}.amazonaws.com"
 
 # Format ROCm version for path (e.g., "7.1" -> "rocm710")
-ROCM_VERSION_PATH="rocm$(echo "${ROCM_VERSION}" | tr -d '.')"
+ROCM_VERSION_PATH="rocm$(echo ${ROCM_VERSION} | tr -d '.')"
 ROCM_PATH="rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}"
 buildkite-agent annotate --style 'success' --context 'rocm-release-workflow' << EOF
 ## ROCm Wheel and Docker Image Releases
@@ -68,7 +68,7 @@ aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/triton
 aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/torchvision-*.whl .
 aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/torchaudio-*.whl .
 aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/amdsmi-*.whl .
-aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/amd_aiter-*.whl .
+aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/aiter-*.whl .
 aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/flash-attn-*.whl .
 \`\`\`
 
@@ -80,7 +80,7 @@ aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/flash-
 - **torchvision**: TorchVision for ROCm PyTorch
 - **torchaudio**: Torchaudio for ROCm PyTorch
 - **amdsmi**: AMD SMI Python bindings
-- **amd_aiter**: Aiter for ROCm
+- **aiter**: Aiter for ROCm
 - **flash-attn**: Flash Attention for ROCm
 
 ### :warning: Notes

.buildkite/scripts/cache-rocm-base-wheels.sh

@@ -83,7 +83,7 @@ case "${1:-}" in
             exit 1
         fi
 
-        WHEEL_COUNT=$(find artifacts/rocm-base-wheels -maxdepth 1 -name '*.whl' 2>/dev/null | wc -l)
+        WHEEL_COUNT=$(ls artifacts/rocm-base-wheels/*.whl 2>/dev/null | wc -l)
         if [[ "$WHEEL_COUNT" -eq 0 ]]; then
             echo "ERROR: No wheels found in artifacts/rocm-base-wheels/" >&2
             exit 1
@@ -110,9 +110,9 @@ case "${1:-}" in
 
         echo ""
         echo "Downloaded wheels:"
-        find artifacts/rocm-base-wheels -maxdepth 1 -name '*.whl' -exec ls -lh {} \;
+        ls -lh artifacts/rocm-base-wheels/
 
-        WHEEL_COUNT=$(find artifacts/rocm-base-wheels -maxdepth 1 -name '*.whl' 2>/dev/null | wc -l)
+        WHEEL_COUNT=$(ls artifacts/rocm-base-wheels/*.whl 2>/dev/null | wc -l)
         echo ""
         echo "Total: $WHEEL_COUNT wheels"
         echo "========================================"

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

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

.buildkite/scripts/cherry-pick-from-milestone.sh

@@ -134,7 +134,7 @@ log_info "Fetching merged PRs from milestone '${MILESTONE}'..."
 
 # Store PR data in a temp file
 PR_DATA=$(mktemp)
-trap 'rm -f "$PR_DATA"' EXIT
+trap "rm -f $PR_DATA" EXIT
 
 if ! gh pr list --state merged --search "milestone:${MILESTONE}" \
     --limit 1000 \

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

@@ -1,57 +1,25 @@
 #!/bin/bash
 
-# This script runs tests inside the corresponding ROCm docker container.
+# This script runs test inside the corresponding ROCm docker container.
-# It handles both single-node and multi-node test configurations.
-#
-# Multi-node detection: Instead of matching on fragile group names, we detect
-# multi-node jobs structurally by looking for the bracket command syntax
-# "[node0_cmds] && [node1_cmds]" or via the NUM_NODES environment variable.
-#
-###############################################################################
-# QUOTING / COMMAND PASSING
-#
-# Passing commands as positional arguments ($*) is fragile when the command
-# string itself contains double quotes, e.g.:
-#
-#   bash run-amd-test.sh "export FLAGS="value" && pytest -m "not slow""
-#
-# The outer shell resolves the nested quotes *before* this script runs, so
-# the script receives mangled input it cannot fully recover.
-#
-# Preferred: pass commands via the VLLM_TEST_COMMANDS environment variable:
-#
-#   export VLLM_TEST_COMMANDS='export FLAGS="value" && pytest -m "not slow"'
-#   bash run-amd-test.sh
-#
-# Single-quoted assignment preserves all inner double quotes verbatim.
-# The $* path is kept for backward compatibility but callers should migrate.
-###############################################################################
 set -o pipefail
 
 # Export Python path
 export PYTHONPATH=".."
 
-###############################################################################
+# Print ROCm version
-# Helper Functions
+echo "--- Confirming Clean Initial State"
-###############################################################################
+while true; do
+        sleep 3
+        if grep -q clean /opt/amdgpu/etc/gpu_state; then
+                echo "GPUs state is \"clean\""
+                break
+        fi
+done
 
-wait_for_clean_gpus() {
+echo "--- ROCm info"
-  local timeout=${1:-300}
+rocminfo
-  local start=$SECONDS
-  echo "--- Waiting for clean GPU state (timeout: ${timeout}s)"
-  while true; do
-    if grep -q clean /opt/amdgpu/etc/gpu_state; then
-      echo "GPUs state is \"clean\""
-      return
-    fi
-    if (( SECONDS - start >= timeout )); then
-      echo "Error: GPUs did not reach clean state within ${timeout}s" >&2
-      exit 1
-    fi
-    sleep 3
-  done
-}
 
+# cleanup older docker images
 cleanup_docker() {
   # Get Docker's root directory
   docker_root=$(docker info -f '{{.DockerRootDir}}')
@@ -60,12 +28,15 @@ cleanup_docker() {
     exit 1
   fi
   echo "Docker root directory: $docker_root"
-
+  # Check disk usage of the filesystem where Docker's root directory is located
   disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
+  # Define the threshold
   threshold=70
   if [ "$disk_usage" -gt "$threshold" ]; then
     echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
+    # Remove dangling images (those that are not tagged and not used by any container)
     docker image prune -f
+    # Remove unused volumes / force the system prune for old images as well.
     docker volume prune -f && docker system prune --force --filter "until=72h" --all
     echo "Docker images and volumes cleanup completed."
   else
@@ -74,445 +45,193 @@ cleanup_docker() {
 }
 
 cleanup_network() {
-  local max_nodes=${NUM_NODES:-2}
+  for node in $(seq 0 $((NUM_NODES-1))); do
-  for node in $(seq 0 $((max_nodes - 1))); do
+    if docker pr -a -q -f name="node${node}" | grep -q .; then
-    if docker ps -a -q -f name="node${node}" | grep -q .; then
+      docker stop "node${node}"
-      docker stop "node${node}" || true
     fi
   done
-  if docker network ls | grep -q docker-net; then
+  if docker network ls | grep docker-net; then
-    docker network rm docker-net || true
+    docker network rm docker-net
   fi
 }
 
-is_multi_node() {
+# Call the cleanup docker function
-  local cmds="$1"
-  # Primary signal: NUM_NODES environment variable set by the pipeline
-  if [[ "${NUM_NODES:-1}" -gt 1 ]]; then
-    return 0
-  fi
-  # Fallback: detect the bracket syntax structurally
-  # Pattern: [...] && [...] (per-node command arrays)
-  if [[ "$cmds" =~ \[.*\].*\&\&.*\[.*\] ]]; then
-    return 0
-  fi
-  return 1
-}
-
-handle_pytest_exit() {
-  local exit_code=$1
-  if [ "$exit_code" -eq 5 ]; then
-    echo "Pytest exit code 5 (no tests collected) - treating as success."
-    exit 0
-  fi
-  exit "$exit_code"
-}
-
-###############################################################################
-# Pytest marker/keyword re-quoting
-#
-# When commands are passed through Buildkite -> shell -> $* -> bash -c,
-# quotes around multi-word pytest -m/-k expressions get stripped:
-#   pytest -v -s -m 'not cpu_test' v1/core
-# becomes:
-#   pytest -v -s -m not cpu_test v1/core
-#
-# pytest then interprets "cpu_test" as a file path, not part of the marker.
-#
-# This function detects unquoted expressions after -m/-k and re-quotes them
-# by collecting tokens until a recognizable boundary is reached:
-#   - test path (contains '/')
-#   - test file (ends with '.py')
-#   - another pytest flag (--xxx or -x single-char flags)
-#   - command separator (&& || ; |)
-#   - environment variable assignment (FOO=bar)
-#
-# Single-word markers (e.g. -m cpu_test, -m hybrid_model) pass through
-# unquoted since they have no spaces and work fine.
-#
-# Already-quoted expressions (containing literal single quotes) are passed
-# through untouched to avoid double-quoting values injected by
-# apply_rocm_test_overrides.
-#
-# NOTE: This ONLY fixes -m/-k flags. It cannot recover arbitrary inner
-# double-quotes stripped by the calling shell (see header comment).
-# Use VLLM_TEST_COMMANDS to avoid the problem entirely.
-###############################################################################
-re_quote_pytest_markers() {
-  local input="$1"
-  local output=""
-  local collecting=false
-  local marker_buf=""
-
-  # Strip backslash-newline continuations, then flatten remaining newlines
-  local flat="${input//$'\\\n'/ }"
-  flat="${flat//$'\n'/ }"
-
-  # Disable globbing to prevent *.py etc. from expanding during read -ra
-  local restore_glob
-  restore_glob="$(shopt -p -o noglob 2>/dev/null || true)"
-  set -o noglob
-  local -a words
-  read -ra words <<< "$flat"
-  eval "$restore_glob"
-
-  for word in "${words[@]}"; do
-    if $collecting; then
-      # If the token we're about to collect already contains a literal
-      # single quote, the expression was already quoted upstream.
-      # Flush and stop collecting.
-      if [[ "$word" == *"'"* ]]; then
-        if [[ -n "$marker_buf" ]]; then
-          # Should not normally happen (partial buf + quote), flush raw
-          output+="${marker_buf} "
-          marker_buf=""
-        fi
-        output+="${word} "
-        collecting=false
-        continue
-      fi
-
-      local is_boundary=false
-      case "$word" in
-        # Line-continuation artifact
-        "\\")
-          is_boundary=true ;;
-        # Command separators
-        "&&"|"||"|";"|"|")
-          is_boundary=true ;;
-        # Long flags (--ignore, --shard-id, etc.)
-        --*)
-          is_boundary=true ;;
-        # Short flags (-v, -s, -x, etc.) but NOT negative marker tokens
-        # like "not" which don't start with "-". Also skip -k/-m which
-        # would start a new marker (handled below).
-        -[a-zA-Z])
-          is_boundary=true ;;
-        # Test path (contains /)
-        */*)
-          is_boundary=true ;;
-        # Test file (ends with .py, possibly with ::method)
-        *.py|*.py::*)
-          is_boundary=true ;;
-        # Environment variable assignment preceding a command (FOO=bar)
-        *=*)
-          # Only treat as boundary if it looks like VAR=value, not
-          # pytest filter expressions like num_gpus=2 inside markers
-          if [[ "$word" =~ ^[A-Z_][A-Z0-9_]*= ]]; then
-            is_boundary=true
-          fi
-          ;;
-      esac
-
-      if $is_boundary; then
-        # Strip surrounding double quotes if present (from upstream
-        # single-to-double conversion); without this, wrapping below
-        # would produce '"expr"' with literal double-quote characters.
-        if [[ "$marker_buf" == '"'*'"' ]]; then
-          marker_buf="${marker_buf#\"}"
-          marker_buf="${marker_buf%\"}"
-        fi
-        # Flush the collected marker expression
-        if [[ "$marker_buf" == *" "* || "$marker_buf" == *"("* ]]; then
-          output+="'${marker_buf}' "
-        else
-          output+="${marker_buf} "
-        fi
-        collecting=false
-        marker_buf=""
-        # Check if this boundary word itself starts a new -m/-k
-        if [[ "$word" == "-m" || "$word" == "-k" ]]; then
-          output+="${word} "
-          collecting=true
-        # Drop stray backslash tokens silently
-        elif [[ "$word" == "\\" ]]; then
-          :
-        else
-          output+="${word} "
-        fi
-      else
-        # Accumulate into marker buffer
-        if [[ -n "$marker_buf" ]]; then
-          marker_buf+=" ${word}"
-        else
-          marker_buf="${word}"
-        fi
-      fi
-    elif [[ "$word" == "-m" || "$word" == "-k" ]]; then
-      output+="${word} "
-      collecting=true
-      marker_buf=""
-    else
-      output+="${word} "
-    fi
-  done
-
-  # Flush any trailing marker expression (marker at end of command)
-  if $collecting && [[ -n "$marker_buf" ]]; then
-    # Strip surrounding double quotes (see mid-stream flush comment)
-    if [[ "$marker_buf" == '"'*'"' ]]; then
-      marker_buf="${marker_buf#\"}"
-      marker_buf="${marker_buf%\"}"
-    fi
-    if [[ "$marker_buf" == *" "* || "$marker_buf" == *"("* ]]; then
-      output+="'${marker_buf}'"
-    else
-      output+="${marker_buf}"
-    fi
-  fi
-
-  echo "${output% }"
-}
-
-###############################################################################
-# ROCm-specific pytest command rewrites
-#
-# These apply ignore flags and environment overrides for tests that are not
-# yet supported or behave differently on ROCm hardware. Kept as a single
-# function so new exclusions are easy to add in one place.
-###############################################################################
-
-apply_rocm_test_overrides() {
-  local cmds="$1"
-
-  # --- Model registry filter ---
-  if [[ $cmds == *"pytest -v -s models/test_registry.py"* ]]; then
-    cmds=${cmds//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
-  fi
-
-  # --- LoRA: disable custom paged attention ---
-  if [[ $cmds == *"pytest -v -s lora"* ]]; then
-    cmds=${cmds//"pytest -v -s lora"/"VLLM_ROCM_CUSTOM_PAGED_ATTN=0 pytest -v -s lora"}
-  fi
-
-  # --- Kernel ignores ---
-  if [[ $cmds == *" kernels/core"* ]]; then
-    cmds="${cmds} \
-    --ignore=kernels/core/test_fused_quant_layernorm.py \
-    --ignore=kernels/core/test_permute_cols.py"
-  fi
-
-  if [[ $cmds == *" kernels/attention"* ]]; then
-    cmds="${cmds} \
-    --ignore=kernels/attention/test_attention_selector.py \
-    --ignore=kernels/attention/test_encoder_decoder_attn.py \
-    --ignore=kernels/attention/test_flash_attn.py \
-    --ignore=kernels/attention/test_flashinfer.py \
-    --ignore=kernels/attention/test_prefix_prefill.py \
-    --ignore=kernels/attention/test_cascade_flash_attn.py \
-    --ignore=kernels/attention/test_mha_attn.py \
-    --ignore=kernels/attention/test_lightning_attn.py \
-    --ignore=kernels/attention/test_attention.py"
-  fi
-
-  if [[ $cmds == *" kernels/quantization"* ]]; then
-    cmds="${cmds} \
-    --ignore=kernels/quantization/test_int8_quant.py \
-    --ignore=kernels/quantization/test_machete_mm.py \
-    --ignore=kernels/quantization/test_block_fp8.py \
-    --ignore=kernels/quantization/test_block_int8.py \
-    --ignore=kernels/quantization/test_marlin_gemm.py \
-    --ignore=kernels/quantization/test_cutlass_scaled_mm.py \
-    --ignore=kernels/quantization/test_int8_kernel.py"
-  fi
-
-  if [[ $cmds == *" kernels/mamba"* ]]; then
-    cmds="${cmds} \
-    --ignore=kernels/mamba/test_mamba_mixer2.py \
-    --ignore=kernels/mamba/test_causal_conv1d.py \
-    --ignore=kernels/mamba/test_mamba_ssm_ssd.py"
-  fi
-
-  if [[ $cmds == *" kernels/moe"* ]]; then
-    cmds="${cmds} \
-    --ignore=kernels/moe/test_moe.py \
-    --ignore=kernels/moe/test_cutlass_moe.py \
-    --ignore=kernels/moe/test_triton_moe_ptpc_fp8.py"
-  fi
-
-  # --- Entrypoint ignores ---
-  if [[ $cmds == *" entrypoints/openai "* ]]; then
-    cmds=${cmds//" entrypoints/openai "/" entrypoints/openai \
-    --ignore=entrypoints/openai/chat_completion/test_audio.py \
-    --ignore=entrypoints/openai/completion/test_shutdown.py \
-    --ignore=entrypoints/openai/test_completion.py \
-    --ignore=entrypoints/openai/test_models.py \
-    --ignore=entrypoints/openai/test_lora_adapters.py \
-    --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
-    --ignore=entrypoints/openai/chat_completion/test_root_path.py \
-    --ignore=entrypoints/openai/test_tokenization.py \
-    --ignore=entrypoints/openai/completion/test_prompt_validation.py "}
-  fi
-
-  if [[ $cmds == *" entrypoints/llm "* ]]; then
-    cmds=${cmds//" entrypoints/llm "/" entrypoints/llm \
-    --ignore=entrypoints/llm/test_chat.py \
-    --ignore=entrypoints/llm/test_accuracy.py \
-    --ignore=entrypoints/llm/test_init.py \
-    --ignore=entrypoints/llm/test_prompt_validation.py "}
-  fi
-
-  # Clean up escaped newlines from --ignore appends
-  cmds=$(echo "$cmds" | sed 's/ \\ / /g')
-
-  echo "$cmds"
-}
-
-###############################################################################
-# Main
-###############################################################################
-
-# --- GPU initialization ---
-echo "--- Confirming Clean Initial State"
-wait_for_clean_gpus
-
-echo "--- ROCm info"
-rocminfo
-
-# --- Docker housekeeping ---
 cleanup_docker
 
 echo "--- Resetting GPUs"
-echo "reset" > /opt/amdgpu/etc/gpu_state
-wait_for_clean_gpus
 
-# --- Pull test image ---
+echo "reset" > /opt/amdgpu/etc/gpu_state
+
+while true; do
+        sleep 3
+        if grep -q clean /opt/amdgpu/etc/gpu_state; then
+                echo "GPUs state is \"clean\""
+                break
+        fi
+done
+
 echo "--- Pulling container"
 image_name="rocm/vllm-ci:${BUILDKITE_COMMIT}"
 container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 docker pull "${image_name}"
 
 remove_docker_container() {
-  docker rm -f "${container_name}" || docker image rm -f "${image_name}" || true
+   docker rm -f "${container_name}" || docker image rm -f "${image_name}" || true
 }
 trap remove_docker_container EXIT
 
-# --- Prepare commands ---
 echo "--- Running container"
 
 HF_CACHE="$(realpath ~)/huggingface"
 mkdir -p "${HF_CACHE}"
 HF_MOUNT="/root/.cache/huggingface"
 
-# ---- Command source selection ----
+commands=$@
-# Prefer VLLM_TEST_COMMANDS (preserves all inner quoting intact).
-# Fall back to $* for backward compatibility, but warn that inner
-# double-quotes will have been stripped by the calling shell.
-if [[ -n "${VLLM_TEST_COMMANDS:-}" ]]; then
-  commands="${VLLM_TEST_COMMANDS}"
-  echo "Commands sourced from VLLM_TEST_COMMANDS (quoting preserved)"
-else
-  commands="$*"
-  if [[ -z "$commands" ]]; then
-    echo "Error: No test commands provided." >&2
-    echo "Usage:" >&2
-    echo "  Preferred:  VLLM_TEST_COMMANDS='...' bash $0" >&2
-    echo "  Legacy:     bash $0 \"commands here\"" >&2
-    exit 1
-  fi
-  echo "Commands sourced from positional args (legacy mode)"
-  echo "WARNING: Inner double-quotes in the command string may have been"
-  echo "  stripped by the calling shell. If you see syntax errors, switch to:"
-  echo "  export VLLM_TEST_COMMANDS='your commands here'"
-  echo "  bash $0"
-fi
-
 echo "Raw commands: $commands"
 
-# Fix quoting before ROCm overrides (so overrides see correct structure)
+commands=${commands//"pytest -v -s basic_correctness/test_basic_correctness.py"/"pytest -v -s basic_correctness/test_basic_correctness.py"}
-commands=$(re_quote_pytest_markers "$commands")
-echo "After re-quoting: $commands"
 
-commands=$(apply_rocm_test_overrides "$commands")
+if [[ $commands == *"pytest -v -s models/test_registry.py"* ]]; then
+  commands=${commands//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
+fi
+
+commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"pytest -v -s compile/test_basic_correctness.py"}
+
+if [[ $commands == *"pytest -v -s lora"* ]]; then
+  commands=${commands//"pytest -v -s lora"/"VLLM_ROCM_CUSTOM_PAGED_ATTN=0 pytest -v -s lora"}
+fi
+
+#ignore certain kernels tests
+if [[ $commands == *" kernels/core"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/core/test_fused_quant_layernorm.py \
+  --ignore=kernels/core/test_permute_cols.py"
+fi
+
+if [[ $commands == *" kernels/attention"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/attention/test_attention_selector.py \
+  --ignore=kernels/attention/test_encoder_decoder_attn.py \
+  --ignore=kernels/attention/test_flash_attn.py \
+  --ignore=kernels/attention/test_flashinfer.py \
+  --ignore=kernels/attention/test_prefix_prefill.py \
+  --ignore=kernels/attention/test_cascade_flash_attn.py \
+  --ignore=kernels/attention/test_mha_attn.py \
+  --ignore=kernels/attention/test_lightning_attn.py \
+  --ignore=kernels/attention/test_attention.py"
+fi
+
+if [[ $commands == *" kernels/quantization"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/quantization/test_int8_quant.py \
+  --ignore=kernels/quantization/test_machete_mm.py \
+  --ignore=kernels/quantization/test_block_fp8.py \
+  --ignore=kernels/quantization/test_block_int8.py \
+  --ignore=kernels/quantization/test_marlin_gemm.py \
+  --ignore=kernels/quantization/test_cutlass_scaled_mm.py \
+  --ignore=kernels/quantization/test_int8_kernel.py"
+fi
+
+if [[ $commands == *" kernels/mamba"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/mamba/test_mamba_mixer2.py \
+  --ignore=kernels/mamba/test_causal_conv1d.py \
+  --ignore=kernels/mamba/test_mamba_ssm_ssd.py"
+fi
+
+if [[ $commands == *" kernels/moe"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/moe/test_moe.py \
+  --ignore=kernels/moe/test_cutlass_moe.py \
+  --ignore=kernels/moe/test_triton_moe_ptpc_fp8.py"
+fi
+
+#ignore certain Entrypoints/openai tests
+if [[ $commands == *" entrypoints/openai "* ]]; then
+  commands=${commands//" entrypoints/openai "/" entrypoints/openai \
+  --ignore=entrypoints/openai/test_audio.py \
+  --ignore=entrypoints/openai/test_shutdown.py \
+  --ignore=entrypoints/openai/test_completion.py \
+  --ignore=entrypoints/openai/test_models.py \
+  --ignore=entrypoints/openai/test_lora_adapters.py \
+  --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
+  --ignore=entrypoints/openai/test_root_path.py \
+  --ignore=entrypoints/openai/test_tokenization.py \
+  --ignore=entrypoints/openai/test_prompt_validation.py "}
+fi
+
+#ignore certain Entrypoints/llm tests
+if [[ $commands == *" entrypoints/llm "* ]]; then
+  commands=${commands//" entrypoints/llm "/" entrypoints/llm \
+  --ignore=entrypoints/llm/test_chat.py \
+  --ignore=entrypoints/llm/test_accuracy.py \
+  --ignore=entrypoints/llm/test_init.py \
+  --ignore=entrypoints/llm/test_prompt_validation.py "}
+fi
+
+commands=$(echo "$commands" | sed 's/ \\ / /g')
 echo "Final commands: $commands"
 
+# --ignore=entrypoints/openai/test_encoder_decoder.py \
+# --ignore=entrypoints/openai/test_embedding.py \
+# --ignore=entrypoints/openai/test_oot_registration.py
+# --ignore=entrypoints/openai/test_accuracy.py \
+# --ignore=entrypoints/openai/test_models.py <= Fails on MI250 but passes on MI300 as of 2025-03-13
+
+
 MYPYTHONPATH=".."
 
-# Verify GPU access
+# Test that we're launching on the machine that has
+# proper access to GPUs
 render_gid=$(getent group render | cut -d: -f3)
 if [[ -z "$render_gid" ]]; then
   echo "Error: 'render' group not found. This is required for GPU access." >&2
   exit 1
 fi
 
-# --- RDMA device passthrough (conditional) ---
+if [[ $commands == *"VLLM_TEST_GROUP_NAME=mi325_4-2-node-tests-4-gpus-in-total"* ]]; then
-# If the host has RDMA devices, pass them through so tests like
-# test_moriio_connector can access ibverbs. On hosts without RDMA
-# hardware the tests will gracefully skip via _rdma_available().
-RDMA_FLAGS=""
-if [ -d /dev/infiniband ]; then
-  echo "RDMA devices detected on host, enabling passthrough"
-  RDMA_FLAGS="--device /dev/infiniband --cap-add=IPC_LOCK"
-else
-  echo "No RDMA devices found on host, RDMA tests will be skipped"
-fi
 
-# --- Route: multi-node vs single-node ---
-if is_multi_node "$commands"; then
-  echo "--- Multi-node job detected"
   export DCKR_VER=$(docker --version | sed 's/Docker version \(.*\), build .*/\1/')
 
-  # Parse the bracket syntax:  prefix ; [node0_cmds] && [node1_cmds]
+  if [[ "$commands" =~ ^(.*)"["(.*)"] && ["(.*)"]"$ ]]; then
-  #   BASH_REMATCH[1] = prefix (everything before first bracket)
+      prefix=$( echo "${BASH_REMATCH[1]}" | sed 's/;//g')
-  #   BASH_REMATCH[2] = comma-separated node0 commands
+      echo "PREFIX: ${prefix}"
-  #   BASH_REMATCH[3] = comma-separated node1 commands
+      export composite_command="(command rocm-smi || true)"
-  if [[ "$commands" =~ ^(.*)\[(.*)"] && ["(.*)\]$ ]]; then
+      myIFS=$IFS
-    prefix=$(echo "${BASH_REMATCH[1]}" | sed 's/;//g')
+      IFS=','
-    echo "PREFIX: ${prefix}"
+      read -ra node0 <<< ${BASH_REMATCH[2]}
+      read -ra node1 <<< ${BASH_REMATCH[3]}
+      IFS=$myIFS
+      for i in "${!node0[@]}";do 
+        command_node_0=$(echo ${node0[i]} | sed 's/\"//g')
+        command_node_1=$(echo ${node1[i]} | sed 's/\"//g')
         
-    export composite_command="(command rocm-smi || true)"
+        export commands="./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 ${image_name} '${command_node_0}' '${command_node_1}'"
-    saved_IFS=$IFS
+        echo "COMMANDS: ${commands}"
-    IFS=','
+        composite_command=$(echo "${composite_command} && ${commands}")
-    read -ra node0 <<< "${BASH_REMATCH[2]}"
+      done
-    read -ra node1 <<< "${BASH_REMATCH[3]}"
+      /bin/bash -c "${composite_command}"
-    IFS=$saved_IFS
+      cleanup_network
-
-    if [[ ${#node0[@]} -ne ${#node1[@]} ]]; then
-      echo "Warning: node0 has ${#node0[@]} commands, node1 has ${#node1[@]}. They will be paired by index."
-    fi
-
-    for i in "${!node0[@]}"; do
-      command_node_0=$(echo "${node0[i]}" | sed 's/\"//g')
-      command_node_1=$(echo "${node1[i]}" | sed 's/\"//g')
-
-      step_cmd="./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 ${image_name} '${command_node_0}' '${command_node_1}'"
-      echo "COMMANDS: ${step_cmd}"
-      composite_command="${composite_command} && ${step_cmd}"
-    done
-
-    /bin/bash -c "${composite_command}"
-    exit_code=$?
-    cleanup_network
-    handle_pytest_exit "$exit_code"
   else
-    echo "Multi-node job detected but failed to parse bracket command syntax."
+      echo "Failed to parse node commands! Exiting."
-    echo "Expected format: prefix ; [node0_cmd1, node0_cmd2] && [node1_cmd1, node1_cmd2]"
+      cleanup_network
-    echo "Got: $commands"
+      exit 111
-    cleanup_network
-    exit 111
   fi
 else
-  echo "--- Single-node job"
   echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
   docker run \
-    --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+          --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
-    $RDMA_FLAGS \
+          --network=host \
-    --network=host \
+          --shm-size=16gb \
-    --shm-size=16gb \
+          --group-add "$render_gid" \
-    --group-add "$render_gid" \
+          --rm \
-    --rm \
+          -e HF_TOKEN \
-    -e HF_TOKEN \
+          -e AWS_ACCESS_KEY_ID \
-    -e AWS_ACCESS_KEY_ID \
+          -e AWS_SECRET_ACCESS_KEY \
-    -e AWS_SECRET_ACCESS_KEY \
+          -v "${HF_CACHE}:${HF_MOUNT}" \
-    -e BUILDKITE_PARALLEL_JOB \
+          -e "HF_HOME=${HF_MOUNT}" \
-    -e BUILDKITE_PARALLEL_JOB_COUNT \
+          -e "PYTHONPATH=${MYPYTHONPATH}" \
-    -v "${HF_CACHE}:${HF_MOUNT}" \
+          --name "${container_name}" \
-    -e "HF_HOME=${HF_MOUNT}" \
+          "${image_name}" \
-    -e "PYTHONPATH=${MYPYTHONPATH}" \
+          /bin/bash -c "${commands}"
-    --name "${container_name}" \
-    "${image_name}" \
-    /bin/bash -c "${commands}"
-
-  exit_code=$?
-  handle_pytest_exit "$exit_code"
 fi

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

@@ -1,65 +0,0 @@
-#!/bin/bash
-set -euox pipefail
-
-export VLLM_CPU_KVCACHE_SPACE=1 
-export VLLM_CPU_CI_ENV=1
-# Reduce sub-processes for acceleration
-export TORCH_COMPILE_DISABLE=1 
-export VLLM_ENABLE_V1_MULTIPROCESSING=0
-
-SDE_ARCHIVE="sde-external-10.7.0-2026-02-18-lin.tar.xz"
-SDE_CHECKSUM="CA3D4086DE4ACB3FAEDF9F57B541C6936B7D5E19AE2BF763B6EA933573A0A217"
-wget "https://downloadmirror.intel.com/913594/${SDE_ARCHIVE}"
-echo "${SDE_CHECKSUM}  ${SDE_ARCHIVE}" | sha256sum --check
-mkdir -p sde
-tar -xvf "./${SDE_ARCHIVE}" --strip-components=1 -C ./sde/
-
-wait_for_pid_and_check_log() {
-    local pid="$1"
-    local log_file="$2"
-    local exit_status
-
-    if [ -z "$pid" ] || [ -z "$log_file" ]; then
-        echo "Usage: wait_for_pid_and_check_log <PID> <LOG_FILE>"
-        return 1
-    fi
-
-    echo "Waiting for process $pid to finish..."
-    
-    # Use the 'wait' command to pause the script until the specific PID exits.
-    # The 'wait' command's own exit status will be that of the waited-for process.
-    if wait "$pid"; then
-        exit_status=$?
-        echo "Process $pid finished with exit status $exit_status (Success)."
-    else
-        exit_status=$?
-        echo "Process $pid finished with exit status $exit_status (Failure)."
-    fi
-
-    if [ "$exit_status" -ne 0 ]; then
-        echo "Process exited with a non-zero status."
-        echo "--- Last few lines of log file: $log_file ---"
-        tail -n 50 "$log_file"
-        echo "---------------------------------------------"
-        return 1 # Indicate failure based on exit status
-    fi
-
-    echo "No errors detected in log file and process exited successfully."
-    return 0
-}
-
-# Test Sky Lake (AVX512F)
-./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_0.log 2>&1 &
-PID_TEST_0=$!
-
-# Test Cascade Lake (AVX512F + VNNI)
-./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_1.log 2>&1 &
-PID_TEST_1=$!
-
-# Test Cooper Lake (AVX512F + VNNI + BF16)
-./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_2.log 2>&1 &
-PID_TEST_2=$!
-
-wait_for_pid_and_check_log $PID_TEST_0 test_0.log
-wait_for_pid_and_check_log $PID_TEST_1 test_1.log
-wait_for_pid_and_check_log $PID_TEST_2 test_2.log

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

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

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

@@ -34,7 +34,7 @@ function cpu_tests() {
   # offline inference
   docker exec cpu-test bash -c "
     set -e
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m"
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
 
   # Run model tests
   docker exec cpu-test bash -c "

.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh

@@ -27,7 +27,7 @@ function cpu_tests() {
   podman exec -it "$container_id" bash -c "
     export TORCH_COMPILE_DISABLE=1
     set -xve
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m" >> "$HOME"/test_basic.log
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m" >> $HOME/test_basic.log
 
   # Run basic model test
   podman exec -it "$container_id" bash -c "
@@ -43,7 +43,7 @@ function cpu_tests() {
     pytest -v -s tests/models/language/generation/test_common.py::test_models[False-False-5-32-google/gemma-1.1-2b-it]
     pytest -v -s tests/models/language/pooling/test_classification.py::test_models[float-jason9693/Qwen2.5-1.5B-apeach]
     # TODO: Below test case tests/models/language/pooling/test_embedding.py::test_models[True-ssmits/Qwen2-7B-Instruct-embed-base] fails on ppc64le. Disabling it for time being.
-    # pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model" >> "$HOME"/test_rest.log
+    # pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model" >> $HOME/test_rest.log
 }
 
 # All of CPU tests are expected to be finished less than 40 mins.

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

@@ -16,5 +16,5 @@ echo "--- :docker: Building Docker image"
 docker build --progress plain --tag "$IMAGE_NAME" --target vllm-test -f docker/Dockerfile.cpu .
 
 # Run the image, setting --shm-size=4g for tensor parallel.
-docker run --rm --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN -e VLLM_CPU_KVCACHE_SPACE=16 -e VLLM_CPU_CI_ENV=1 -e VLLM_CPU_SIM_MULTI_NUMA=1 --shm-size=4g "$IMAGE_NAME" \
+docker run --rm --cpuset-cpus=$CORE_RANGE --cpuset-mems=$NUMA_NODE -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN -e VLLM_CPU_KVCACHE_SPACE=16 -e VLLM_CPU_CI_ENV=1 -e VLLM_CPU_SIM_MULTI_NUMA=1 --shm-size=4g $IMAGE_NAME \
-        timeout "$TIMEOUT_VAL" bash -c "set -euox pipefail; echo \"--- Print packages\"; pip list; echo \"--- Running tests\"; ${TEST_COMMAND}"
+        timeout $TIMEOUT_VAL bash -c "set -euox pipefail; echo \"--- Print packages\"; pip list; echo \"--- Running tests\"; ${TEST_COMMAND}"

.buildkite/scripts/hardware_ci/run-gh200-test.sh

@@ -25,5 +25,5 @@ remove_docker_container
 
 # Run the image and test offline inference
 docker run -e HF_TOKEN -e VLLM_WORKER_MULTIPROC_METHOD=spawn -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
-    python3 examples/basic/offline_inference/generate.py --model meta-llama/Llama-3.2-1B
+    python3 examples/offline_inference/basic/generate.py --model meta-llama/Llama-3.2-1B
 '

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

@@ -1,42 +1,17 @@
 #!/bin/bash
 
-# This script builds the HPU docker image and runs the offline inference inside the container.
+# This script build the CPU docker image and run the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
-#
-# vllm-gaudi compatibility pinning:
-#   The vllm-gaudi plugin is installed on top of the vllm upstream checkout used by this CI job.
-#   When upstream vllm changes its API, the plugin may break before it has been updated.
-#   To handle this, the vllm-gaudi repository maintains a file:
-#     vllm/last-good-commit-for-vllm-gaudi/VLLM_COMMUNITY_COMMIT
-#   The first line of that file controls what version of vllm is used inside the Docker image:
-#     - "latest"        : no checkout override; the current Buildkite CI commit is used as-is.
-#     - "<commit SHA>"  : vllm is checked out to that specific commit before building, pinning
-#                         the test to a known-compatible baseline.
-#   To unpin (resume testing against the live vllm tip), set the file content back to "latest".
 set -exuo pipefail
 
-# Fetch the vllm community commit reference from vllm-gaudi (first line only).
-VLLM_COMMUNITY_COMMIT=$(curl -s \
-  https://raw.githubusercontent.com/vllm-project/vllm-gaudi/vllm/last-good-commit-for-vllm-gaudi/VLLM_COMMUNITY_COMMIT \
-  | head -1 | tr -d '\n')
-
-echo "Using vllm community commit: ${VLLM_COMMUNITY_COMMIT}"
-
 # Try building the docker image
 image_name="hpu/upstream-vllm-ci:${BUILDKITE_COMMIT}"
 container_name="hpu-upstream-vllm-ci-${BUILDKITE_COMMIT}-container"
-cat <<EOF | docker build -t "${image_name}" -f - .
+cat <<EOF | docker build -t ${image_name} -f - .
 FROM gaudi-base-image:latest
 
 COPY ./ /workspace/vllm
 
-# If VLLM_COMMUNITY_COMMIT is a specific commit (not "latest"), check it out to pin vllm
-# to the version known to be compatible with vllm-gaudi. When the value is "latest",
-# the current checkout (the Buildkite CI commit) is used unchanged.
-RUN if [ "${VLLM_COMMUNITY_COMMIT}" != "latest" ]; then \
-      cd /workspace/vllm && git fetch --unshallow 2>/dev/null || true && git checkout ${VLLM_COMMUNITY_COMMIT}; \
-    fi
-
 WORKDIR /workspace/vllm
 
 ENV no_proxy=localhost,127.0.0.1
@@ -64,19 +39,19 @@ EOF
 # functions, while other platforms only need one remove_docker_container
 # function.
 EXITCODE=1
-remove_docker_containers() { docker rm -f "${container_name}" || true; }
+remove_docker_containers() { docker rm -f ${container_name} || true; }
 trap 'remove_docker_containers; exit $EXITCODE;' EXIT
 remove_docker_containers
 
 echo "Running HPU plugin v1 test"
-docker run --rm --runtime=habana --name="${container_name}" --network=host \
+docker run --rm --runtime=habana --name=${container_name} --network=host \
   -e HABANA_VISIBLE_DEVICES=all \
   -e VLLM_SKIP_WARMUP=true \
   -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
   -e PT_HPU_LAZY_MODE=1 \
   "${image_name}" \
   /bin/bash -c '
-  cd vllm; timeout 120s python -u examples/basic/offline_inference/generate.py --model facebook/opt-125m
+  cd vllm; timeout 120s python -u examples/offline_inference/basic/generate.py --model facebook/opt-125m
 '
 
 EXITCODE=$?

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

@@ -41,7 +41,6 @@ get_config() {
         echo "Error: file '${TEST_RUN_CONFIG_FILE}' does not exist in the warehouse" >&2
         exit 1
     fi
-    # shellcheck source=/dev/null
     source "${TEST_RUN_CONFIG_FILE}"
     echo "Base docker image name that get from configuration: ${BASE_IMAGE_NAME}"
     return 0
@@ -49,8 +48,9 @@ get_config() {
 
 # get test running configuration.
 fetch_vllm_test_cfg
+get_config
 # Check if the function call was successful. If not, exit the script.
-if ! get_config; then
+if [ $? -ne 0 ]; then
   exit 1
 fi
 
@@ -62,14 +62,14 @@ agent_idx=$(echo "${BUILDKITE_AGENT_NAME}" | awk -F'-' '{print $(NF-1)}')
 echo "agent_idx: ${agent_idx}"
 builder_name="cachebuilder${agent_idx}"
 builder_cache_dir="/mnt/docker-cache${agent_idx}"
-mkdir -p "${builder_cache_dir}"
+mkdir -p ${builder_cache_dir}
 
 # Try building the docker image
 cat <<EOF | DOCKER_BUILDKIT=1 docker build \
-    --add-host cache-service-vllm.nginx-pypi-cache.svc.cluster.local:"${PYPI_CACHE_HOST}" \
+    --add-host cache-service-vllm.nginx-pypi-cache.svc.cluster.local:${PYPI_CACHE_HOST} \
-    --builder "${builder_name}" --cache-from type=local,src="${builder_cache_dir}" \
+    --builder ${builder_name} --cache-from type=local,src=${builder_cache_dir} \
-                           --cache-to type=local,dest="${builder_cache_dir}",mode=max \
+                           --cache-to type=local,dest=${builder_cache_dir},mode=max \
-    --progress=plain --load -t "${image_name}" -f - .
+    --progress=plain --load -t ${image_name} -f - .
 FROM ${BASE_IMAGE_NAME}
 
 # Define environments
@@ -116,7 +116,7 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
     source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
     source /usr/local/Ascend/nnal/atb/set_env.sh && \
-    export LD_LIBRARY_PATH=\$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/$(uname -i)-linux/devlib && \
+    export LD_LIBRARY_PATH=\$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/`uname -i`-linux/devlib && \
     python3 -m pip install -v -e /workspace/vllm-ascend/ --extra-index https://download.pytorch.org/whl/cpu/
 
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
@@ -139,7 +139,7 @@ trap remove_docker_container EXIT
 # Generate corresponding --device args based on BUILDKITE_AGENT_NAME
 # Ascend NPU BUILDKITE_AGENT_NAME format is {hostname}-{agent_idx}-{npu_card_num}cards, and agent_idx starts from 1.
 #   e.g. atlas-a2-001-1-2cards means this is the 1-th agent on atlas-a2-001 host, and it has 2 NPU cards.
-#   returns one argument per line: --device, /dev/davinciX, ...
+#   returns --device /dev/davinci0 --device /dev/davinci1
 parse_and_gen_devices() {
     local input="$1"
     local index cards_num
@@ -151,24 +151,29 @@ parse_and_gen_devices() {
         return 1
     fi
 
+    local devices=""
     local i=0
     while (( i < cards_num )); do
         local dev_idx=$(((index - 1)*cards_num + i ))
-        printf '%s\n' "--device"
+        devices="$devices --device /dev/davinci${dev_idx}"
-        printf '%s\n' "/dev/davinci${dev_idx}"
         ((i++))
     done
+
+    # trim leading space
+    devices="${devices#"${devices%%[![:space:]]*}"}"
+    # Output devices: assigned to the caller variable
+    printf '%s' "$devices"
 }
 
-mapfile -t device_args < <(parse_and_gen_devices "${BUILDKITE_AGENT_NAME}") || exit 1
+devices=$(parse_and_gen_devices "${BUILDKITE_AGENT_NAME}") || exit 1
 
 # Run the image and execute the Out-Of-Tree (OOT) platform interface test case on Ascend NPU hardware.
 # This test checks whether the OOT platform interface is functioning properly in conjunction with
 # the hardware plugin vllm-ascend.
 model_cache_dir=/mnt/modelscope${agent_idx}
-mkdir -p "${model_cache_dir}"
+mkdir -p ${model_cache_dir}
 docker run \
-    "${device_args[@]}" \
+    ${devices} \
     --device /dev/davinci_manager \
     --device /dev/devmm_svm \
     --device /dev/hisi_hdc \
@@ -177,7 +182,7 @@ docker run \
     -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
     -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
     -v /etc/ascend_install.info:/etc/ascend_install.info \
-    -v "${model_cache_dir}":/root/.cache/modelscope \
+    -v ${model_cache_dir}:/root/.cache/modelscope \
     --entrypoint="" \
     --name "${container_name}" \
     "${image_name}" \

.buildkite/scripts/hardware_ci/run-tpu-v1-test-part2.sh

@@ -61,7 +61,7 @@ echo "Results will be stored in: $RESULTS_DIR"
 echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
     && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
-    && python3 -m pip install --progress-bar off "lm-eval[api]>=0.4.11" \
+    && python3 -m pip install --progress-bar off "lm-eval[api]>=0.4.9.2" \
     && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
 echo "--- Python dependencies installed ---"
 

.buildkite/scripts/hardware_ci/run-tpu-v1-test.sh

@@ -61,7 +61,7 @@ echo "Results will be stored in: $RESULTS_DIR"
 echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
     && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
-    && python3 -m pip install --progress-bar off "lm-eval[api]>=0.4.11" \
+    && python3 -m pip install --progress-bar off "lm-eval[api]>=0.4.9.2" \
     && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
 echo "--- Python dependencies installed ---"
 

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

@@ -8,7 +8,7 @@ image_name="xpu/vllm-ci:${BUILDKITE_COMMIT}"
 container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 
 # Try building the docker image
-docker build -t "${image_name}" -f docker/Dockerfile.xpu .
+docker build -t ${image_name} -f docker/Dockerfile.xpu .
 
 # Setup cleanup
 remove_docker_container() {
@@ -34,17 +34,17 @@ docker run \
     set -e
     echo $ZE_AFFINITY_MASK
     pip install tblib==3.1.0
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    python3 examples/offline_inference/basic/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/offline_inference/basic/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 ray
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
+    python3 examples/offline_inference/basic/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/offline_inference/basic/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/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --quantization fp8
-    python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4  --block-size 64 --enforce-eager
+    python3 examples/offline_inference/basic/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4  --block-size 64 --enforce-eager
-    python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2
+    python3 examples/offline_inference/basic/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2
-    python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
+    python3 examples/offline_inference/basic/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
     cd tests
-    pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py
+    pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py
     pytest -v -s v1/engine
     pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
     pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py

.buildkite/scripts/push-nightly-builds.sh

@@ -21,16 +21,16 @@ echo "Pushing original tag $ORIG_TAG_NAME$ORIG_TAG_SUFFIX to new nightly tag nam
 
 # pull original arch-dependent images from AWS ECR Public
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-x86_64"$ORIG_TAG_SUFFIX"
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-x86_64$ORIG_TAG_SUFFIX
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-aarch64"$ORIG_TAG_SUFFIX"
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-aarch64$ORIG_TAG_SUFFIX
 # tag arch-dependent images
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-x86_64"$ORIG_TAG_SUFFIX" vllm/vllm-openai:"$TAG_NAME"-x86_64
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-x86_64$ORIG_TAG_SUFFIX vllm/vllm-openai:$TAG_NAME-x86_64
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-aarch64"$ORIG_TAG_SUFFIX" vllm/vllm-openai:"$TAG_NAME"-aarch64
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-aarch64$ORIG_TAG_SUFFIX vllm/vllm-openai:$TAG_NAME-aarch64
 # push arch-dependent images to DockerHub
-docker push vllm/vllm-openai:"$TAG_NAME"-x86_64
+docker push vllm/vllm-openai:$TAG_NAME-x86_64
-docker push vllm/vllm-openai:"$TAG_NAME"-aarch64
+docker push vllm/vllm-openai:$TAG_NAME-aarch64
 # push arch-independent manifest to DockerHub
-docker manifest create vllm/vllm-openai:"$TAG_NAME" vllm/vllm-openai:"$TAG_NAME"-x86_64 vllm/vllm-openai:"$TAG_NAME"-aarch64 --amend
+docker manifest create vllm/vllm-openai:$TAG_NAME vllm/vllm-openai:$TAG_NAME-x86_64 vllm/vllm-openai:$TAG_NAME-aarch64 --amend
-docker manifest create vllm/vllm-openai:"$TAG_NAME"-"$BUILDKITE_COMMIT" vllm/vllm-openai:"$TAG_NAME"-x86_64 vllm/vllm-openai:"$TAG_NAME"-aarch64 --amend
+docker manifest create vllm/vllm-openai:$TAG_NAME-$BUILDKITE_COMMIT vllm/vllm-openai:$TAG_NAME-x86_64 vllm/vllm-openai:$TAG_NAME-aarch64 --amend
-docker manifest push vllm/vllm-openai:"$TAG_NAME"
+docker manifest push vllm/vllm-openai:$TAG_NAME
-docker manifest push vllm/vllm-openai:"$TAG_NAME"-"$BUILDKITE_COMMIT"
+docker manifest push vllm/vllm-openai:$TAG_NAME-$BUILDKITE_COMMIT

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

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

.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh

@@ -51,14 +51,14 @@ for BACK in "${BACKENDS[@]}"; do
     --enable-eplb \
     --trust-remote-code \
     --max-model-len 2048 \
-    --all2all-backend "$BACK" \
+    --all2all-backend $BACK \
-    --port "$PORT" &
+    --port $PORT &
   SERVER_PID=$!
-  wait_for_server "$PORT"
+  wait_for_server $PORT
 
   TAG=$(echo "$MODEL" | tr '/: \\n' '_____')
   OUT="${OUT_DIR}/${TAG}_${BACK}.json"
-  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port "$PORT" --num-questions "${NUM_Q}" --save-results "${OUT}"
+  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port $PORT --num-questions ${NUM_Q} --save-results ${OUT}
   python3 - <<PY
 import json; acc=json.load(open('${OUT}'))['accuracy']
 print(f"${MODEL} ${BACK}: accuracy {acc:.3f}")

.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_prefetch_offload.sh

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

.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh

@@ -47,20 +47,20 @@ for BACK in "${BACKENDS[@]}"; do
   vllm serve "$MODEL" \
     --enforce-eager \
     --enable-eplb \
-    --all2all-backend "$BACK" \
+    --all2all-backend $BACK \
     --eplb-config '{"window_size":10, "step_interval":100, "num_redundant_experts":0, "log_balancedness":true}' \
-    --tensor-parallel-size "${TENSOR_PARALLEL_SIZE}" \
+    --tensor-parallel-size ${TENSOR_PARALLEL_SIZE} \
-    --data-parallel-size "${DATA_PARALLEL_SIZE}" \
+    --data-parallel-size ${DATA_PARALLEL_SIZE} \
     --enable-expert-parallel \
     --trust-remote-code \
     --max-model-len 2048 \
-    --port "$PORT" &
+    --port $PORT &
   SERVER_PID=$!
-  wait_for_server "$PORT"
+  wait_for_server $PORT
 
   TAG=$(echo "$MODEL" | tr '/: \\n' '_____')
   OUT="${OUT_DIR}/${TAG}_${BACK}.json"
-  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port "$PORT" --num-questions "${NUM_Q}" --save-results "${OUT}"
+  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port $PORT --num-questions ${NUM_Q} --save-results ${OUT}
   python3 - <<PY
 import json; acc=json.load(open('${OUT}'))['accuracy']
 print(f"${MODEL} ${BACK}: accuracy {acc:.3f}")

.buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh

@@ -24,7 +24,7 @@ if command -v rocm-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:
   BACKENDS=("allgather_reducescatter")
   # Disable MOE padding for ROCm since it is causing eplb to fail
   export VLLM_ROCM_MOE_PADDING=0
-  PLATFORM_ARGS=("--no-async-scheduling" "--attention-backend=TRITON_ATTN")
+  PLATFORM_ARGS=("--no-async-scheduling")
   echo "Disabled async scheduling for ROCm platform due to issues with spec decode."
 else
   # Non-ROCm platform (CUDA/other)
@@ -51,20 +51,20 @@ for BACK in "${BACKENDS[@]}"; do
     --tensor-parallel-size 4 \
     --enable-expert-parallel \
     --enable-eplb \
-    --all2all-backend "$BACK" \
+    --all2all-backend $BACK \
     --eplb-config '{"window_size":200,"step_interval":600,"use_async":true}' \
     --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":1}' \
     --trust-remote-code \
     --max-model-len 2048 \
     --gpu-memory-utilization 0.9 \
     "${PLATFORM_ARGS[@]}" \
-    --port "$PORT" &
+    --port $PORT &
   SERVER_PID=$!
-  wait_for_server "$PORT"
+  wait_for_server $PORT
 
   TAG=$(echo "$MODEL" | tr '/: \\n' '_____')
   OUT="${OUT_DIR}/${TAG}_${BACK}.json"
-  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port "$PORT" --num-questions "${NUM_Q}" --save-results "${OUT}"
+  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port $PORT --num-questions ${NUM_Q} --save-results ${OUT}
   python3 - <<PY
 import json; acc=json.load(open('${OUT}'))['accuracy']
 print(f"${MODEL} ${BACK}: accuracy {acc:.3f}")

.buildkite/scripts/tool_call/run-bfcl-eval.sh

@@ -1,248 +0,0 @@
-#!/bin/bash
-# Run BFCL (Berkeley Function Call Leaderboard) tool-calling correctness
-# evaluation against a local vLLM server.
-#
-# Usage:
-#   # Run with defaults (gpt-oss-20b, multi_turn)
-#   bash .buildkite/scripts/tool_call/run-bfcl-eval.sh
-#
-#   # Run with gpt-oss-120b and multiple test categories
-#   BFCL_MODEL="openai/gpt-oss-120b" BFCL_TP_SIZE=4 \
-#     BFCL_TEST_CATEGORY="live_simple, multiple, parallel_multiple" \
-#     bash .buildkite/scripts/tool_call/run-bfcl-eval.sh
-#
-#   # Chain both API types (use BFCL_OUTPUT_DIR to avoid overwriting results)
-#   BFCL_OUTPUT_DIR=./bfcl-chat-completions BFCL_API_TYPE=chat_completions \
-#     bash .buildkite/scripts/tool_call/run-bfcl-eval.sh && \
-#   BFCL_OUTPUT_DIR=./bfcl-responses BFCL_API_TYPE=responses \
-#     bash .buildkite/scripts/tool_call/run-bfcl-eval.sh
-#
-# Environment variables (all optional, with defaults):
-#   BFCL_MODEL          - HF model name (default: openai/gpt-oss-20b)
-#   BFCL_API_TYPE       - API type: "chat_completions" or "responses" (default: chat_completions)
-#   BFCL_OUTPUT_DIR     - Directory for BFCL results (default: current working directory)
-#   BFCL_TEST_CATEGORY  - BFCL test categories (default: multi_turn)
-#   BFCL_TOOL_CALL_PARSER - Tool call parser name (default: openai)
-#   BFCL_NUM_THREADS    - Threads for BFCL generate (default: 8)
-#   BFCL_TP_SIZE        - Tensor parallel size (default: 1)
-#   BFCL_MAX_MODEL_LEN  - Max model length (default: 4096)
-#   BFCL_PORT           - Server port (default: 8000)
-#   BFCL_REASONING_PARSER - Reasoning parser name (default: disabled)
-#   BFCL_EXTRA_ARGS     - Additional vLLM server args
-
-set -euo pipefail
-
-# ---- Configuration ----
-MODEL="${BFCL_MODEL:-openai/gpt-oss-20b}"
-API_TYPE="${BFCL_API_TYPE:-chat_completions}"
-OUTPUT_DIR="${BFCL_OUTPUT_DIR:-}"
-TEST_CATEGORY="${BFCL_TEST_CATEGORY:-multi_turn}"
-TOOL_CALL_PARSER="${BFCL_TOOL_CALL_PARSER:-openai}"
-NUM_THREADS="${BFCL_NUM_THREADS:-8}"
-TP_SIZE="${BFCL_TP_SIZE:-1}"
-MAX_MODEL_LEN="${BFCL_MAX_MODEL_LEN:-4096}"
-PORT="${BFCL_PORT:-8000}"
-REASONING_PARSER="${BFCL_REASONING_PARSER:-}"
-EXTRA_ARGS="${BFCL_EXTRA_ARGS:-}"
-
-# Set up output directory
-if [ -n "$OUTPUT_DIR" ]; then
-    mkdir -p "$OUTPUT_DIR"
-    OUTPUT_DIR="$(cd "$OUTPUT_DIR" && pwd)"
-fi
-
-echo "============================================"
-echo "BFCL Tool Call Correctness Evaluation"
-echo "============================================"
-echo "Model:          $MODEL"
-echo "Tool parser:    $TOOL_CALL_PARSER"
-echo "API type:       $API_TYPE"
-echo "Output dir:     ${OUTPUT_DIR:-<cwd>}"
-echo "Test category:  $TEST_CATEGORY"
-echo "TP size:        $TP_SIZE"
-echo "Max model len:  $MAX_MODEL_LEN"
-echo "Port:           $PORT"
-echo "Num threads:    $NUM_THREADS"
-echo "============================================"
-
-# ---- Install bfcl-eval if missing ----
-if ! python3 -c "import bfcl_eval" 2>/dev/null; then
-    echo "Installing bfcl-eval..."
-    pip install "bfcl-eval>=2025.10.20.1,<2026"
-fi
-
-# ---- Cleanup handler ----
-SERVER_PID=""
-cleanup() {
-    if [ -n "$SERVER_PID" ]; then
-        echo "Stopping vLLM server (pid=$SERVER_PID)..."
-        kill "$SERVER_PID" 2>/dev/null || true
-        wait "$SERVER_PID" 2>/dev/null || true
-    fi
-    # Remove BFCL lock files (created by filelock for thread-safe writes)
-    rm -rf .file_locks/
-    if [ -n "${OUTPUT_DIR:-}" ]; then
-        rm -rf "$OUTPUT_DIR/.file_locks/"
-    fi
-}
-trap cleanup EXIT
-
-# ---- Start vLLM server ----
-echo "Starting vLLM server..."
-
-SERVE_ARGS=(
-    "$MODEL"
-    --port "$PORT"
-    --enable-auto-tool-choice
-    --tool-call-parser "$TOOL_CALL_PARSER"
-    --tensor-parallel-size "$TP_SIZE"
-    --max-model-len "$MAX_MODEL_LEN"
-    --enforce-eager
-    --no-enable-prefix-caching
-)
-
-# Append reasoning parser if specified
-if [ -n "$REASONING_PARSER" ]; then
-    SERVE_ARGS+=(--reasoning-parser "$REASONING_PARSER")
-fi
-
-# Append any extra args
-if [ -n "$EXTRA_ARGS" ]; then
-    read -ra EXTRA_ARGS_ARRAY <<< "$EXTRA_ARGS"
-    SERVE_ARGS+=("${EXTRA_ARGS_ARRAY[@]}")
-fi
-
-echo "Command: vllm serve ${SERVE_ARGS[*]}"
-vllm serve "${SERVE_ARGS[@]}" &
-SERVER_PID=$!
-
-# ---- Wait for server to be ready ----
-echo "Waiting for vLLM server to start (timeout: 600s)..."
-SECONDS_WAITED=0
-until curl -sf "http://localhost:${PORT}/health" > /dev/null 2>&1; do
-    if [ $SECONDS_WAITED -ge 600 ]; then
-        echo ""
-        echo "ERROR: vLLM server failed to start within 600s"
-        exit 1
-    fi
-    if (( SECONDS_WAITED % 30 == 0 && SECONDS_WAITED > 0 )); then
-        echo "  Still waiting... (${SECONDS_WAITED}s elapsed)"
-    fi
-    sleep 2
-    SECONDS_WAITED=$((SECONDS_WAITED + 2))
-done
-echo "vLLM server is ready. (started in ${SECONDS_WAITED}s)"
-
-# ---- Run BFCL evaluation ----
-# bfcl-eval has no CLI entry point; generate() and evaluate() are Typer
-# functions that must be called from Python. The MODEL_CONFIG_MAPPING must
-# be patched in-process so BFCL knows to use the OpenAI-compatible handler
-# against our local vLLM server.
-bfcl_exit_code=0
-python3 - "$MODEL" "$TEST_CATEGORY" "$NUM_THREADS" "$PORT" "$API_TYPE" "$OUTPUT_DIR" << 'PYEOF' || bfcl_exit_code=$?
-import os
-import sys
-
-model = sys.argv[1]
-test_category = sys.argv[2]
-num_threads = int(sys.argv[3])
-port = sys.argv[4]
-api_type = sys.argv[5]
-output_dir = sys.argv[6] if len(sys.argv) > 6 and sys.argv[6] else os.getcwd()
-
-os.environ["OPENAI_BASE_URL"] = f"http://localhost:{port}/v1"
-os.environ["OPENAI_API_KEY"] = "dummy"
-os.environ["BFCL_PROJECT_ROOT"] = output_dir
-
-import bfcl_eval.constants.model_config as bfcl_model_config
-from bfcl_eval.constants.model_config import ModelConfig
-from bfcl_eval.model_handler.api_inference.openai_completion import (
-    OpenAICompletionsHandler,
-)
-from bfcl_eval.model_handler.api_inference.openai_response import (
-    OpenAIResponsesHandler,
-)
-
-if api_type == "responses":
-    handler = OpenAIResponsesHandler
-else:
-    handler = OpenAICompletionsHandler
-
-bfcl_model_config.MODEL_CONFIG_MAPPING[model] = ModelConfig(
-    model_name=model,
-    display_name=f"{model} (FC) (vLLM)",
-    url=f"https://huggingface.co/{model}",
-    org="",
-    license="apache-2.0",
-    model_handler=handler,
-    input_price=None,
-    output_price=None,
-    is_fc_model=True,
-    underscore_to_dot=True,
-)
-
-from bfcl_eval.__main__ import evaluate, generate
-import inspect
-import typer
-
-
-def _get_default_kwargs(function):
-    kwargs = {}
-    for k, v in inspect.signature(function).parameters.items():
-        if v.default is not inspect.Parameter.empty:
-            default = v.default
-            if isinstance(default, typer.models.OptionInfo):
-                default = default.default
-            kwargs[k] = default
-    return kwargs
-
-
-# ---- generate ----
-print(f"=== BFCL generate: model={model} test_category={test_category} ===")
-gen_kwargs = _get_default_kwargs(generate)
-gen_kwargs["model"] = [model]
-gen_kwargs["test_category"] = [c.strip() for c in test_category.split(",")]
-gen_kwargs["skip_server_setup"] = True
-gen_kwargs["num_threads"] = num_threads
-generate(**gen_kwargs)
-
-# ---- evaluate ----
-print(f"=== BFCL evaluate: model={model} test_category={test_category} ===")
-eval_kwargs = _get_default_kwargs(evaluate)
-eval_kwargs["model"] = [model]
-eval_kwargs["test_category"] = [c.strip() for c in test_category.split(",")]
-evaluate(**eval_kwargs)
-
-print("=== BFCL evaluation completed successfully ===")
-PYEOF
-
-# ---- Upload results to buildkite ----
-if command -v buildkite-agent &>/dev/null; then
-    if [ $bfcl_exit_code -eq 0 ]; then
-        STYLE="success"
-        STATUS="PASSED"
-    else
-        STYLE="error"
-        STATUS="FAILED"
-    fi
-
-    buildkite-agent annotate --style "$STYLE" --context "bfcl-results" <<EOF
-### BFCL Tool Call Correctness - ${STATUS}
-- **Model:** \`${MODEL}\`
-- **Parser:** \`${TOOL_CALL_PARSER}\`
-- **API type:** \`${API_TYPE}\`
-- **Test category:** \`${TEST_CATEGORY}\`
-EOF
-
-    # BFCL writes results to $BFCL_PROJECT_ROOT/result/ and scores to
-    # $BFCL_PROJECT_ROOT/score/
-    RESULTS_ROOT="${OUTPUT_DIR:-.}"
-    if [ -d "$RESULTS_ROOT/result" ]; then
-        buildkite-agent artifact upload "$RESULTS_ROOT/result/**/*"
-    fi
-    if [ -d "$RESULTS_ROOT/score" ]; then
-        buildkite-agent artifact upload "$RESULTS_ROOT/score/**/*"
-    fi
-fi
-
-exit $bfcl_exit_code

.buildkite/scripts/tpu/docker_run_bm.sh

@@ -9,11 +9,10 @@ ENV_FILE=$1
 
 # For testing on local vm, use `set -a` to export all variables
 source /etc/environment
-# shellcheck source=/dev/null
+source $ENV_FILE
-source "$ENV_FILE"
 
 remove_docker_container() { 
-    docker rm -f "$CONTAINER_NAME" || true;
+    docker rm -f $CONTAINER_NAME || true;
 }
 
 trap remove_docker_container EXIT
@@ -42,13 +41,13 @@ echo
 echo "starting docker...$CONTAINER_NAME"
 echo    
 docker run \
- -v "$DOWNLOAD_DIR":"$DOWNLOAD_DIR" \
+ -v $DOWNLOAD_DIR:$DOWNLOAD_DIR \
- --env-file "$ENV_FILE" \
+ --env-file $ENV_FILE \
  -e HF_TOKEN="$HF_TOKEN" \
- -e TARGET_COMMIT="$BUILDKITE_COMMIT" \
+ -e TARGET_COMMIT=$BUILDKITE_COMMIT \
- -e MODEL="$MODEL" \
+ -e MODEL=$MODEL \
  -e WORKSPACE=/workspace \
- --name "$CONTAINER_NAME" \
+ --name $CONTAINER_NAME \
  -d \
  --privileged \
  --network host \

.buildkite/scripts/tpu/run_bm.sh

@@ -42,21 +42,21 @@ echo "lanching vllm..."
 echo "logging to $VLLM_LOG"
 echo
 
-vllm serve "$MODEL" \
+vllm serve $MODEL \
  --seed 42 \
- --max-num-seqs "$MAX_NUM_SEQS" \
+ --max-num-seqs $MAX_NUM_SEQS \
- --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS" \
+ --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
- --tensor-parallel-size "$TENSOR_PARALLEL_SIZE" \
+ --tensor-parallel-size $TENSOR_PARALLEL_SIZE \
  --no-enable-prefix-caching \
- --download_dir "$DOWNLOAD_DIR" \
+ --download_dir $DOWNLOAD_DIR \
- --max-model-len "$MAX_MODEL_LEN" > "$VLLM_LOG" 2>&1 &
+ --max-model-len $MAX_MODEL_LEN > "$VLLM_LOG" 2>&1 &
 
 
 echo "wait for 20 minutes.."
 echo
 # sleep 1200
 # wait for 10 minutes...
-for _ in {1..120}; do
+for i in {1..120}; do
     # TODO: detect other type of errors.
     if grep -Fq "raise RuntimeError" "$VLLM_LOG"; then
         echo "Detected RuntimeError, exiting."
@@ -78,11 +78,11 @@ echo "logging to $BM_LOG"
 echo
 vllm bench serve \
     --backend vllm \
-    --model "$MODEL"  \
+    --model $MODEL  \
     --dataset-name sonnet \
     --dataset-path benchmarks/sonnet_4x.txt \
-    --sonnet-input-len "$INPUT_LEN" \
+    --sonnet-input-len $INPUT_LEN \
-    --sonnet-output-len "$OUTPUT_LEN" \
+    --sonnet-output-len $OUTPUT_LEN \
     --ignore-eos > "$BM_LOG"
 
 echo "completed..."

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

@@ -72,19 +72,20 @@ obj_json="objects.json"
 aws s3api list-objects-v2 --bucket "$BUCKET" --prefix "$SUBPATH/" --delimiter / --output json > "$obj_json"
 mkdir -p "$INDICES_OUTPUT_DIR"
 
-# call script to generate indices for all existing wheels
+# call script to generate indicies for all existing wheels
 # this indices have relative paths that could work as long as it is next to the wheel directory in s3
 # i.e., the wheels are always in s3://vllm-wheels/<commit>/
 # and indices can be placed in /<commit>/, or /nightly/, or /<version>/
-alias_args=()
+if [[ ! -z "$DEFAULT_VARIANT_ALIAS" ]]; then
-if [[ -n "$DEFAULT_VARIANT_ALIAS" ]]; then
+    alias_arg="--alias-to-default $DEFAULT_VARIANT_ALIAS"
-    alias_args=(--alias-to-default "$DEFAULT_VARIANT_ALIAS")
+else
+    alias_arg=""
 fi
 
 # HACK: we do not need regex module here, but it is required by pre-commit hook
 # To avoid any external dependency, we simply replace it back to the stdlib re module
 sed -i 's/import regex as re/import re/g' .buildkite/scripts/generate-nightly-index.py
-$PYTHON .buildkite/scripts/generate-nightly-index.py --version "$SUBPATH" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" --comment "commit $BUILDKITE_COMMIT" "${alias_args[@]}"
+$PYTHON .buildkite/scripts/generate-nightly-index.py --version "$SUBPATH" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" --comment "commit $BUILDKITE_COMMIT" $alias_arg
 
 # copy indices to /<commit>/ unconditionally
 echo "Uploading indices to $S3_COMMIT_PREFIX"
@@ -99,9 +100,9 @@ fi
 # re-generate and copy to /<pure_version>/ only if it does not have "dev" in the version
 if [[ "$version" != *"dev"* ]]; then
     echo "Re-generating indices for /$pure_version/"
-    rm -rf "${INDICES_OUTPUT_DIR:?}/*"
+    rm -rf "$INDICES_OUTPUT_DIR/*"
     mkdir -p "$INDICES_OUTPUT_DIR"
     # wheel-dir is overridden to be the commit directory, so that the indices point to the correct wheel path
-    $PYTHON .buildkite/scripts/generate-nightly-index.py --version "$pure_version" --wheel-dir "$SUBPATH" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" --comment "version $pure_version" "${alias_args[@]}"
+    $PYTHON .buildkite/scripts/generate-nightly-index.py --version "$pure_version" --wheel-dir "$SUBPATH" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" --comment "version $pure_version" $alias_arg
     aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/$pure_version/"
 fi

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

@@ -7,7 +7,7 @@ SUBPATH=$BUILDKITE_COMMIT
 S3_COMMIT_PREFIX="s3://$BUCKET/$SUBPATH/"
 
 RELEASE_VERSION=$(buildkite-agent meta-data get release-version)
-GIT_VERSION=$(git describe --exact-match --tags "$BUILDKITE_COMMIT" 2>/dev/null)
+GIT_VERSION=$(git describe --exact-match --tags $BUILDKITE_COMMIT 2>/dev/null)
 
 echo "Release version from Buildkite: $RELEASE_VERSION"
 
@@ -54,13 +54,10 @@ mkdir -p $DIST_DIR
 # include only wheels for the release version, ignore all files with "dev" or "rc" in the name (without excluding 'aarch64')
 aws s3 cp --recursive --exclude "*" --include "vllm-${PURE_VERSION}*.whl" --exclude "*dev*" --exclude "*rc[0-9]*" "$S3_COMMIT_PREFIX" $DIST_DIR
 echo "Wheels copied to local directory"
-# generate source distribution using setup.py
+# generate source tarball
-python setup.py sdist --dist-dir=$DIST_DIR
+git archive --format=tar.gz --output="$DIST_DIR/vllm-${PURE_VERSION}.tar.gz" $BUILDKITE_COMMIT
 ls -la $DIST_DIR
 
-SDIST_FILE=$(find $DIST_DIR -name "vllm*.tar.gz")
-echo "Found sdist: $SDIST_FILE"
-
 # upload wheels to PyPI (only default variant, i.e. files without '+' in the name)
 PYPI_WHEEL_FILES=$(find $DIST_DIR -name "vllm-${PURE_VERSION}*.whl" -not -name "*+*")
 if [[ -z "$PYPI_WHEEL_FILES" ]]; then
@@ -68,6 +65,6 @@ if [[ -z "$PYPI_WHEEL_FILES" ]]; then
   exit 1
 fi
 
-python3 -m twine check "$PYPI_WHEEL_FILES" "$SDIST_FILE"
+python3 -m twine check $PYPI_WHEEL_FILES
-python3 -m twine upload --non-interactive --verbose "$PYPI_WHEEL_FILES" "$SDIST_FILE"
+python3 -m twine upload --non-interactive --verbose $PYPI_WHEEL_FILES
-echo "Wheels and source distribution uploaded to PyPI"
+echo "Wheels uploaded to PyPI"

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

@@ -55,7 +55,7 @@ mkdir -p all-rocm-wheels
 cp artifacts/rocm-base-wheels/*.whl all-rocm-wheels/ 2>/dev/null || true
 cp artifacts/rocm-vllm-wheel/*.whl all-rocm-wheels/ 2>/dev/null || true
 
-WHEEL_COUNT=$(find all-rocm-wheels -maxdepth 1 -name '*.whl' 2>/dev/null | wc -l)
+WHEEL_COUNT=$(ls all-rocm-wheels/*.whl 2>/dev/null | wc -l)
 echo "Total wheels to upload: $WHEEL_COUNT"
 
 if [ "$WHEEL_COUNT" -eq 0 ]; then
@@ -115,7 +115,7 @@ if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]] |
 fi
 
 # Extract version from vLLM wheel and update version-specific index
-VLLM_WHEEL=$(find all-rocm-wheels -maxdepth 1 -name 'vllm*.whl' 2>/dev/null | head -1)
+VLLM_WHEEL=$(ls all-rocm-wheels/vllm*.whl 2>/dev/null | head -1)
 if [ -n "$VLLM_WHEEL" ]; then
     VERSION=$(unzip -p "$VLLM_WHEEL" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
     echo "Version in wheel: $VERSION"

.buildkite/test_areas/benchmarks.yaml

@@ -17,15 +17,3 @@ steps:
   - tests/benchmarks/
   commands:
   - pytest -v -s benchmarks/
-
-- label: Attention Benchmarks Smoke Test (B200)
-  device: b200
-  num_gpus: 2
-  optional: true
-  working_dir: "/vllm-workspace/"
-  timeout_in_minutes: 10
-  source_file_dependencies:
-  - benchmarks/attention_benchmarks/
-  - vllm/v1/attention/
-  commands:
-  - python3 benchmarks/attention_benchmarks/benchmark.py --backends flash flashinfer --batch-specs "8q1s1k" --repeats 1 --warmup-iters 1

.buildkite/test_areas/compile.yaml

@@ -36,16 +36,6 @@ steps:
   - export VLLM_TEST_CLEAN_GPU_MEMORY=1
   - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py
 
-- label: AsyncTP Correctness Tests (B200)
-  timeout_in_minutes: 50
-  working_dir: "/vllm-workspace/"
-  device: b200
-  optional: true
-  num_devices: 2
-  commands:
-  - export VLLM_TEST_CLEAN_GPU_MEMORY=1
-  - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py
-
 - label: Distributed Compile Unit Tests (2xH100)
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/"
@@ -101,8 +91,8 @@ steps:
     - nvidia-smi
     # Run all models and attn backends but only Inductor partition and native custom ops
     - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and not +rms_norm and not +quant_fp8"
-    # Qwen/Deepseek requires +quant_fp8 as -quant_fp8 rms+quant fusion is not supported
+    # Qwen requires +quant_fp8 as -quant_fp8 rms+quant fusion is not supported
-    - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and not +rms_norm and +quant_fp8 and (qwen3 or deepseek)"
+    - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and not +rms_norm and +quant_fp8 and qwen3"
 
 - label: Fusion E2E Config Sweep (H100)
   timeout_in_minutes: 30
@@ -131,10 +121,13 @@ steps:
   optional: true
   commands:
     - nvidia-smi
-    # Run all models but only FLASHINFER, Inductor partition and native custom ops
+    # Run all models and attn backends but only Inductor partition and native custom ops
-    # Qwen/Deepseek requires +quant_fp8 as -quant_fp8 rms+quant fusion is not supported
+    # -k "inductor_partition and not +rms_norm and not +quant_fp8"
-    # Run just llama3 (fp8 & fp4) for all config combinations (only inductor partition)
+    # Qwen requires +quant_fp8 as -quant_fp8 rms+quant fusion is not supported
-    - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and (FLASHINFER and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek)) or llama-3)"
+    # -k "inductor_partition and not +rms_norm and +quant_fp8 and qwen3"
+    # Run just llama3 (fp8 & fp4) for all config combinations
+    # -k "llama-3"
+    - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and not +rms_norm and not +quant_fp8" -k "inductor_partition and not +rms_norm and +quant_fp8 and qwen3" -k "llama-3"
 
 - label: Fusion E2E TP2 Quick (H100)
   timeout_in_minutes: 20
@@ -150,8 +143,8 @@ steps:
   commands:
     - nvidia-smi
     # Run all models and attn backends but only Inductor partition and native custom ops
-    - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek))"
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "inductor_partition and not +rms_norm and not +quant_fp8"
-    - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek))"
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "inductor_partition and not +rms_norm and not +quant_fp8"
 
 - label: Fusion E2E TP2 AR-RMS Config Sweep (H100)
   timeout_in_minutes: 40
@@ -169,7 +162,7 @@ steps:
     - tests/compile/fusions_e2e/
   commands:
     - nvidia-smi
-    # Run just llama3 (fp8 & bf16) for all config combinations
+    # Run just llama3 (fp4 & fp8 & bf16) for all config combinations
     - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "llama-3"
 
 - label: Fusion E2E TP2 AsyncTP Config Sweep (H100)
@@ -204,8 +197,7 @@ steps:
     - tests/compile/fusions_e2e/
   commands:
     - nvidia-smi
-    # Run all models but only FLASHINFER, Inductor partition and native custom ops
+    # Run all models and attn backends but only Inductor partition and native custom ops
-    # include qwen/deepseek with +quant_fp8 as -quant_fp8 rms+quant fusion is not supported
     # for ar-rms-quant-fp4, also sweep llama3
-    - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "(FLASHINFER and inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek))) or Llama-3.1-8B-Instruct-FP4"
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "inductor_partition and not +rms_norm and not +quant_fp8" -k "Llama-3.1-8B-Instruct-FP4"
-    - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "FLASHINFER and inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and (qwen3 or deepseek))"
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "inductor_partition and not +rms_norm and not +quant_fp8"

.buildkite/test_areas/distributed.yaml

@@ -50,18 +50,23 @@ steps:
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
   - pytest -v -s v1/worker/test_worker_memory_snapshot.py
 
-- label: Distributed Torchrun + Examples (4 GPUs)
+- label: Distributed Tests (4 GPUs)
-  timeout_in_minutes: 30
+  timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:
   - vllm/distributed/
-  - tests/distributed/test_torchrun_example.py
+  - tests/distributed/test_utils
-  - tests/distributed/test_torchrun_example_moe.py
+  - tests/distributed/test_pynccl
+  - tests/distributed/test_events
+  - tests/compile/fullgraph/test_basic_correctness.py
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - examples/offline_inference/new_weight_syncing/
   - tests/examples/offline_inference/data_parallel.py
+  - tests/v1/distributed
+  - tests/v1/engine/test_engine_core_client.py
+  - tests/distributed/test_symm_mem_allreduce.py
   commands:
   # https://github.com/NVIDIA/nccl/issues/1838
   - export NCCL_CUMEM_HOST_ENABLE=0
@@ -79,27 +84,6 @@ steps:
   - TP_SIZE=2 DP_SIZE=2 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
   # test with internal dp
   - python3 ../examples/offline_inference/data_parallel.py --enforce-eager
-  # OLD rlhf examples
-  - cd ../examples/offline_inference
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
-  # NEW rlhf examples
-  - cd new_weight_syncing
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_nccl.py
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_ipc.py
-
-- label: Distributed DP Tests (4 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-  - vllm/distributed/
-  - tests/v1/distributed
-  - tests/v1/engine/test_engine_core_client.py
-  - tests/distributed/test_utils
-  commands:
-  # https://github.com/NVIDIA/nccl/issues/1838
-  - export NCCL_CUMEM_HOST_ENABLE=0
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
@@ -107,27 +91,20 @@ steps:
   - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py
   - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
   - pytest -v -s distributed/test_utils.py
-
-- label: Distributed Compile + Comm (4 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-  - vllm/distributed/
-  - tests/distributed/test_pynccl
-  - tests/distributed/test_events
-  - tests/compile/fullgraph/test_basic_correctness.py
-  - tests/distributed/test_symm_mem_allreduce.py
-  - tests/distributed/test_multiproc_executor.py
-  commands:
-  # https://github.com/NVIDIA/nccl/issues/1838
-  - export NCCL_CUMEM_HOST_ENABLE=0
   - pytest -v -s compile/fullgraph/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s distributed/test_events.py
   - pytest -v -s distributed/test_symm_mem_allreduce.py
-  # test multi-node TP with multiproc executor (simulated on single node)
+  # TODO: create a dedicated test section for multi-GPU example tests
-  - pytest -v -s distributed/test_multiproc_executor.py::test_multiproc_executor_multi_node
+  # when we have multiple distributed example tests
+  # OLD rlhf examples
+  - cd ../examples/offline_inference
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
+  # NEW rlhf examples
+  - cd new_weight_syncing
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_async_new_apis.py
 
 - label: Distributed Tests (8 GPUs)(H100)
   timeout_in_minutes: 10
@@ -169,7 +146,6 @@ steps:
   num_devices: 2
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
-    - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py
     - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
     - pytest -v -s tests/v1/distributed/test_dbo.py
 
@@ -189,7 +165,6 @@ steps:
   num_devices: 2
   num_nodes: 2
   no_plugin: true
-  optional: true # TODO: revert once infra issue solved
   source_file_dependencies:
   - vllm/distributed/
   - vllm/engine/
@@ -222,31 +197,7 @@ steps:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
     - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
-- label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs)
+- label: Pipeline + Context Parallelism (4 GPUs))
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-
-- label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
-  timeout_in_minutes: 30
-  device: a100
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - vllm/v1/worker/kv_connector_model_runner_mixin.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
-
-- label: Pipeline + Context Parallelism (4 GPUs)
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
   num_devices: 4

.buildkite/test_areas/e2e_integration.yaml

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

.buildkite/test_areas/engine.yaml

@@ -14,59 +14,17 @@ steps:
   commands:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
 
-- label: Engine (1 GPU)
+- label: V1 e2e + engine
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
   source_file_dependencies:
-    - vllm/v1/engine/
+    - vllm/
-    - tests/v1/engine/
+    - tests/v1
   commands:
+    # TODO: accuracy does not match, whether setting
+    # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
+    - pytest -v -s v1/e2e
+    # Run this test standalone for now;
+    # need to untangle use (implicit) use of spawn/fork across the tests.
     - pytest -v -s v1/engine/test_preprocess_error_handling.py
+    # Run the rest of v1/engine tests
     - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
-
-- label: e2e Scheduling (1 GPU)
-  timeout_in_minutes: 30
-  source_file_dependencies:
-    - vllm/v1/
-    - tests/v1/e2e/general/
-  commands:
-    - pytest -v -s v1/e2e/general/test_async_scheduling.py
-
-- label: e2e Core (1 GPU)
-  timeout_in_minutes: 30
-  source_file_dependencies:
-    - vllm/v1/
-    - tests/v1/e2e/general/
-  commands:
-    - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
-
-- label: V1 e2e (2 GPUs)
-  timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
-  optional: true
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/
-    - tests/v1/e2e
-  commands:
-    # Only run tests that need exactly 2 GPUs
-    - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"
-  mirror:
-    amd:
-      device: mi325_2
-      depends_on:
-      - image-build-amd
-
-- label: V1 e2e (4 GPUs)
-  timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
-  optional: true
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/
-    - tests/v1/e2e
-  commands:
-    # Only run tests that need 4 GPUs
-    - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "eagle_correctness_heavy"
-  mirror:
-    amd:
-      device: mi325_4
-      depends_on:
-      - image-build-amd

.buildkite/test_areas/entrypoints.yaml

@@ -34,26 +34,23 @@ steps:
   - tests/entrypoints/test_chat_utils
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/chat_completion/test_oot_registration.py --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/  --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/  --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses
   - pytest -v -s entrypoints/test_chat_utils.py
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
 
 - label: Entrypoints Integration (API Server 2)
   timeout_in_minutes: 130
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
-  - tests/entrypoints/rpc
-  - tests/entrypoints/instrumentator
   - tests/tool_use
+  - tests/entrypoints/sleep
+  - tests/entrypoints/instrumentator
+  - tests/entrypoints/rpc
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/instrumentator
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
+  - pytest -v -s entrypoints/instrumentator
+  - pytest -v -s entrypoints/sleep
   - pytest -v -s tool_use
 
 - label: Entrypoints Integration (Pooling)
@@ -82,11 +79,6 @@ steps:
     - tests/v1
   commands:
     - pytest -v -s v1/entrypoints
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
 
 - label: OpenAI API Correctness
   timeout_in_minutes: 30

.buildkite/test_areas/expert_parallelism.yaml

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

.buildkite/test_areas/kernels.yaml

@@ -8,9 +8,8 @@ steps:
   - csrc/
   - tests/kernels/core
   - tests/kernels/test_top_k_per_row.py
-  - tests/kernels/test_concat_mla_q.py
   commands:
-    - pytest -v -s kernels/core kernels/test_top_k_per_row.py kernels/test_concat_mla_q.py
+    - pytest -v -s kernels/core kernels/test_top_k_per_row.py
 
 - label: Kernels Attention Test %N
   timeout_in_minutes: 35
@@ -45,8 +44,7 @@ steps:
   - vllm/envs.py
   - vllm/config
   commands:
-    - pytest -v -s kernels/moe --ignore=kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+    - pytest -v -s kernels/moe --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
-    - pytest -v -s kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 2
 
 - label: Kernels Mamba Test
@@ -72,7 +70,7 @@ steps:
   - tests/kernels/moe/test_batched_deepgemm.py
   - tests/kernels/attention/test_deepgemm_attention.py
   commands:
-    - pytest -v -s kernels/quantization/test_block_fp8.py
+    - pytest -v -s kernels/quantization/test_block_fp8.py -k deep_gemm
     - pytest -v -s kernels/moe/test_deepgemm.py
     - pytest -v -s kernels/moe/test_batched_deepgemm.py
     - pytest -v -s kernels/attention/test_deepgemm_attention.py
@@ -97,7 +95,7 @@ steps:
   - vllm/platforms/cuda.py
   commands:
     - nvidia-smi
-    - python3 examples/basic/offline_inference/chat.py
+    - python3 examples/offline_inference/basic/chat.py
     # Attention
     # num_heads2 broken by https://github.com/flashinfer-ai/flashinfer/issues/1353
     - pytest -v -s tests/kernels/attention/test_attention_selector.py
@@ -117,7 +115,6 @@ steps:
     - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
     - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
-    - pytest -v -s tests/kernels/moe/test_flashinfer_moe.py
     - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
     # e2e
     - pytest -v -s tests/models/quantization/test_nvfp4.py
@@ -157,6 +154,8 @@ steps:
   commands:
     - pytest -v -s kernels/moe/test_deepep_deepgemm_moe.py
     - pytest -v -s kernels/moe/test_deepep_moe.py
+    - pytest -v -s kernels/moe/test_pplx_cutlass_moe.py
+    # - pytest -v -s kernels/moe/test_pplx_moe.py - failing on main
   
 - label: Kernels Fp4 MoE Test (B200)
   timeout_in_minutes: 60

.buildkite/test_areas/lm_eval.yaml

@@ -11,17 +11,17 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
 
-# - label: LM Eval Large Models (4 GPUs)(A100)
+- label: LM Eval Large Models (4 GPUs)(A100)
-#   device: a100
+  device: a100
-#   optional: true
+  optional: true
-#   num_devices: 4
+  num_devices: 4
-#   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-#   source_file_dependencies:
+  source_file_dependencies:
-#   - csrc/
+  - csrc/
-#   - vllm/model_executor/layers/quantization
+  - vllm/model_executor/layers/quantization
-#   commands:
+  commands:
-#   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-#   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
+  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
 - label: LM Eval Large Models (4 GPUs)(H100)
   device: h100
@@ -73,29 +73,3 @@ steps:
   num_devices: 2
   commands:
     - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor-dp-ep/config-b200.txt
-
-- label: GPQA Eval (GPT-OSS) (H100)
-  timeout_in_minutes: 120
-  device: h100
-  optional: true
-  num_devices: 2
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  - tests/evals/gpt_oss/
-  commands:
-    - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-h100.txt
-
-- label: GPQA Eval (GPT-OSS) (B200)
-  timeout_in_minutes: 120
-  device: b200
-  optional: true
-  num_devices: 2
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  - tests/evals/gpt_oss/
-  commands:
-    - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-b200.txt

.buildkite/test_areas/misc.yaml

@@ -9,7 +9,6 @@ steps:
     - tests/v1
   commands:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     # split the test to avoid interference
     - pytest -v -s -m 'not cpu_test' v1/core
     - pytest -v -s v1/executor
@@ -17,7 +16,6 @@ steps:
     - pytest -v -s v1/sample
     - pytest -v -s v1/logits_processors
     - pytest -v -s v1/worker
-    # TODO: create another `optional` test group for slow tests
     - pytest -v -s -m 'not slow_test' v1/spec_decode
     - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
     - pytest -v -s -m 'not cpu_test' v1/metrics
@@ -27,11 +25,6 @@ steps:
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
 
 - label: V1 Others (CPU)
   depends_on:
@@ -67,13 +60,12 @@ steps:
   - examples/
   commands:
     - pip install tensorizer # for tensorizer test
-     # for basic
+    - python3 offline_inference/basic/chat.py # for basic
-    - python3 basic/offline_inference/chat.py
+    - python3 offline_inference/basic/generate.py --model facebook/opt-125m
-    - python3 basic/offline_inference/generate.py --model facebook/opt-125m
+    - python3 offline_inference/basic/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10
-    - python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10
+    - python3 offline_inference/basic/classify.py
-    - python3 basic/offline_inference/classify.py
+    - python3 offline_inference/basic/embed.py
-    - python3 basic/offline_inference/embed.py
+    - python3 offline_inference/basic/score.py
-    - python3 basic/offline_inference/score.py
     # for multi-modal models
     - python3 offline_inference/audio_language.py --seed 0
     - python3 offline_inference/vision_language.py --seed 0
@@ -116,11 +108,9 @@ steps:
   timeout_in_minutes: 50
   source_file_dependencies:
   - vllm/
-  - tests/detokenizer
   - tests/multimodal
   - tests/utils_
   commands:
-  - pytest -v -s detokenizer
   - pytest -v -s -m 'not cpu_test' multimodal
   - pytest -v -s utils_
 
@@ -133,7 +123,6 @@ steps:
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/test_pooling_params.py
-  - tests/test_ray_env.py
   - tests/multimodal
   - tests/renderers
   - tests/standalone_tests/lazy_imports.py
@@ -147,7 +136,6 @@ steps:
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s test_pooling_params.py
-  - pytest -v -s test_ray_env.py
   - pytest -v -s -m 'cpu_test' multimodal
   - pytest -v -s renderers
   - pytest -v -s tokenizers_
@@ -155,6 +143,20 @@ steps:
   - pytest -v -s transformers_utils
   - pytest -v -s config
 
+- label: GPT-OSS Eval (B200)
+  timeout_in_minutes: 60
+  working_dir: "/vllm-workspace/"
+  device: b200
+  optional: true
+  source_file_dependencies:
+  - tests/evals/gpt_oss
+  - vllm/model_executor/models/gpt_oss.py
+  - vllm/model_executor/layers/quantization/mxfp4.py
+  - vllm/v1/attention/backends/flashinfer.py
+  commands:
+    - uv pip install --system 'gpt-oss[eval]==0.0.5'
+    - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
+
 - label: Batch Invariance (H100)
   timeout_in_minutes: 25
   device: h100

.buildkite/test_areas/model_executor.yaml

@@ -9,9 +9,9 @@ steps:
   - vllm/config/model.py
   - vllm/model_executor
   - tests/model_executor
-  - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
+  - tests/entrypoints/openai/test_tensorizer_entrypoint.py
   commands:
     - apt-get update && apt-get install -y curl libsodium23
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - pytest -v -s model_executor
-    - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py
+    - pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py

.buildkite/test_areas/model_runner_v2.yaml

@@ -1,110 +0,0 @@
-group: Model Runner V2
-depends_on:
-  - image-build
-steps:
-- label: Model Runner V2 Core Tests
-  timeout_in_minutes: 45
-  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/
-  - tests/v1/entrypoints/llm/test_struct_output_generate.py
-  commands:
-  - set -x
-  - export VLLM_USE_V2_MODEL_RUNNER=1
-  - pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics"
-  # This requires eager until we sort out CG correctness issues.
-  # TODO: remove ENFORCE_EAGER here after https://github.com/vllm-project/vllm/pull/32936 is merged.
-  - ENFORCE_EAGER=1 pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram"
-  - pytest -v -s v1/e2e/general/test_context_length.py
-  - pytest -v -s v1/e2e/general/test_min_tokens.py
-  # Temporary hack filter to exclude ngram spec decoding based tests.
-  - pytest -v -s v1/entrypoints/llm/test_struct_output_generate.py -k "xgrammar and not speculative_config6 and not speculative_config7 and not speculative_config8 and not speculative_config0"
-
-- label: Model Runner V2 Examples
-  timeout_in_minutes: 45
-  working_dir: "/vllm-workspace/examples"
-  source_file_dependencies:
-    - vllm/v1/worker/gpu/
-    - vllm/v1/core/sched/
-    - vllm/v1/worker/gpu_worker.py
-    - examples/offline_inference/
-    - examples/basic/offline_inference/
-    - examples/pooling/embed/vision_embedding_offline.py
-    - examples/others/tensorize_vllm_model.py
-  commands:
-    - set -x
-    - export VLLM_USE_V2_MODEL_RUNNER=1
-    - pip install tensorizer # for tensorizer test
-    - python3 basic/offline_inference/chat.py # for basic
-    - python3 basic/offline_inference/generate.py --model facebook/opt-125m
-    #- python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10  # TODO
-    #- python3 basic/offline_inference/embed.py   # TODO
-    # for multi-modal models
-    - python3 offline_inference/audio_language.py --seed 0
-    - python3 offline_inference/vision_language.py --seed 0
-    - python3 offline_inference/vision_language_multi_image.py --seed 0
-    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
-    # for pooling models
-    - python3 pooling/embed/vision_embedding_offline.py --seed 0
-    # for features demo
-    - python3 offline_inference/prefix_caching.py
-    - python3 offline_inference/llm_engine_example.py
-    - python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
-    - python3 offline_inference/spec_decode.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
-    # https://github.com/vllm-project/vllm/pull/26682 uses slightly more memory in PyTorch 2.9+ causing this test to OOM in 1xL4 GPU
-    - python3 offline_inference/spec_decode.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
-
-- label: Model Runner V2 Distributed (2 GPUs)
-  timeout_in_minutes: 45
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/v1/worker/gpu/
-    - vllm/v1/worker/gpu_worker.py
-    - tests/basic_correctness/test_basic_correctness.py
-    - tests/v1/distributed/test_async_llm_dp.py
-    - tests/v1/distributed/test_eagle_dp.py
-  commands:
-    - set -x
-    - export VLLM_USE_V2_MODEL_RUNNER=1
-    # The "and not True" here is a hacky way to exclude the prompt_embeds cases which aren't yet supported.
-    - TARGET_TEST_SUITE=L4 pytest -v -s basic_correctness/test_basic_correctness.py -m 'distributed(num_gpus=2)' -k "not ray and not True"
-    # https://github.com/NVIDIA/nccl/issues/1838
-    - export NCCL_CUMEM_HOST_ENABLE=0
-    - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py -k "not ray"
-    - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
-
-# These require fix https://github.com/vllm-project/vllm/pull/36280
-- label: Model Runner V2 Pipeline Parallelism (4 GPUs)
-  timeout_in_minutes: 60
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/v1/worker/gpu/
-    - vllm/v1/worker/gpu_worker.py
-    - tests/distributed/test_pipeline_parallel.py
-    #- tests/distributed/test_pp_cudagraph.py
-  commands:
-    - set -x
-    - export VLLM_USE_V2_MODEL_RUNNER=1
-    - pytest -v -s distributed/test_pipeline_parallel.py -k "not ray and not Jamba"
-    # TODO: Uncomment once https://github.com/vllm-project/vllm/pull/35162 is merged.
-    #- pytest -v -s distributed/test_pp_cudagraph.py -k "not ray"
-
-- label: Model Runner V2 Spec Decode
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/v1/worker/gpu/
-  - vllm/v1/worker/gpu_worker.py
-  - tests/v1/spec_decode/test_max_len.py
-  - tests/v1/e2e/spec_decode/test_spec_decode.py
-  commands:
-  - set -x
-  - export VLLM_USE_V2_MODEL_RUNNER=1
-  - pytest -v -s v1/spec_decode/test_max_len.py -k "eagle or mtp"
-  - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "eagle or mtp"

.buildkite/test_areas/models_basic.yaml

@@ -4,6 +4,7 @@ depends_on:
 steps:
 - label: Basic Models Tests (Initialization)
   timeout_in_minutes: 45
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -15,6 +16,7 @@ steps:
 
 - label: Basic Models Tests (Extra Initialization) %N
   timeout_in_minutes: 45
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/model_executor/models/
@@ -36,12 +38,6 @@ steps:
   - tests/models/test_registry.py
   commands:
     - pytest -v -s models/test_terratorch.py models/test_transformers.py models/test_registry.py
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
-    
 
 - label: Basic Models Test (Other CPU) # 5min
   depends_on: 
@@ -65,7 +61,7 @@ steps:
     - pytest -v -s tests/models/test_transformers.py
     - pytest -v -s tests/models/multimodal/processing/
     - pytest -v -s tests/models/multimodal/test_mapping.py
-    - python3 examples/basic/offline_inference/chat.py
+    - python3 examples/offline_inference/basic/chat.py
     - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
     # Whisper needs spawn method to avoid deadlock
     - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper

.buildkite/test_areas/models_language.yaml

@@ -4,6 +4,7 @@ depends_on:
 steps:
 - label: Language Models Tests (Standard)
   timeout_in_minutes: 25
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -15,6 +16,7 @@ steps:
 
 - label: Language Models Tests (Extra Standard) %N
   timeout_in_minutes: 45
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/model_executor/models/
@@ -30,6 +32,7 @@ steps:
 
 - label: Language Models Tests (Hybrid) %N
   timeout_in_minutes: 75
+  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -37,7 +40,7 @@ steps:
   commands:
     # Install fast path packages for testing against transformers
     # Note: also needed to run plamo2 model in vLLM
-    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
+    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.2.5'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
     # Shard hybrid language model tests
     - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
@@ -45,6 +48,7 @@ steps:
 
 - label: Language Models Test (Extended Generation) # 80min
   timeout_in_minutes: 110
+  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/
@@ -52,21 +56,13 @@ steps:
   commands:
     # Install fast path packages for testing against transformers
     # Note: also needed to run plamo2 model in vLLM
-    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
+    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.2.5'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
     - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
-      commands:
-      - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
-      - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
-      - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
 - label: Language Models Test (PPL)
   timeout_in_minutes: 110
+  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/
@@ -76,20 +72,17 @@ steps:
 
 - label: Language Models Test (Extended Pooling)  # 36min
   timeout_in_minutes: 50
+  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/
   - tests/models/language/pooling
   commands:
     - pytest -v -s models/language/pooling -m 'not core_model'
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
 
 - label: Language Models Test (MTEB)
   timeout_in_minutes: 110
+  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/

.buildkite/test_areas/models_multimodal.yaml

@@ -2,65 +2,16 @@ group: Models - Multimodal
 depends_on: 
   - image-build
 steps:
-- label: "Multi-Modal Models (Standard) 1: qwen2"
+- label: Multi-Modal Models (Standard) # 60min
-  timeout_in_minutes: 45
+  timeout_in_minutes: 80
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen2"
+    - pip freeze | grep -E 'torch'
-    - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
+    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
-
-- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
-  timeout_in_minutes: 45
-  source_file_dependencies:
-  - vllm/
-  - tests/models/multimodal
-  commands:
-    - 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
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
-
-- label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
-  timeout_in_minutes: 45
-  source_file_dependencies:
-  - vllm/
-  - tests/models/multimodal
-  commands:
-    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma"
-    - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
-
-- label: "Multi-Modal Models (Standard) 4: other + whisper"
-  timeout_in_minutes: 45
-  source_file_dependencies:
-  - vllm/
-  - 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/processing
     - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
 
 - label: Multi-Modal Processor Test (CPU)
   depends_on: 
@@ -69,7 +20,6 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
-  - tests/models/registry.py
   device: cpu
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
@@ -80,7 +30,6 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
-  - tests/models/registry.py
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/processing/test_tensor_schema.py
@@ -103,11 +52,6 @@ steps:
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal -m 'not core_model' --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/processing
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
 
 - label: Multi-Modal Models (Extended) 2
   optional: true
@@ -126,3 +70,12 @@ steps:
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
+
+# This test is used only in PR development phase to test individual models and should never run on main
+- label: Custom Models
+  optional: true
+  commands:
+    - echo 'Testing custom models...'
+    # PR authors can temporarily add commands below to test individual models
+    # e.g. pytest -v -s models/encoder_decoder/vision_language/test_mllama.py
+    # *To avoid merge conflicts, remember to REMOVE (not just comment out) them before merging the PR*

.buildkite/test_areas/plugins.yaml

@@ -15,17 +15,10 @@ steps:
   - pytest -v -s plugins_tests/test_platform_plugins.py
   - pip uninstall vllm_add_dummy_platform -y
   # end platform plugin tests
-  # begin io_processor plugins test
+  # begin io_processor plugins test, all the code in between uses the prithvi_io_processor plugin
-  # test generic io_processor plugins functions
-  - pytest -v -s ./plugins_tests/test_io_processor_plugins.py
-  # test Terratorch io_processor plugins
   - pip install -e ./plugins/prithvi_io_processor_plugin
-  - pytest -v -s plugins_tests/test_terratorch_io_processor_plugins.py
+  - pytest -v -s plugins_tests/test_io_processor_plugins.py
   - pip uninstall prithvi_io_processor_plugin -y
-  # test bge_m3_sparse io_processor plugin
-  - pip install -e ./plugins/bge_m3_sparse_plugin
-  - pytest -v -s plugins_tests/test_bge_m3_sparse_io_processor_plugins.py
-  - pip uninstall bge_m3_sparse_plugin -y
   # end io_processor plugins test
   # begin stat_logger plugins test
   - pip install -e ./plugins/vllm_add_dummy_stat_logger
@@ -36,6 +29,6 @@ steps:
   - pytest -v -s plugins_tests/test_scheduler_plugins.py
   - pip install -e ./plugins/vllm_add_dummy_model
   - pytest -v -s distributed/test_distributed_oot.py
-  - pytest -v -s entrypoints/openai/chat_completion/test_oot_registration.py # it needs a clean process
+  - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
   - pytest -v -s models/test_oot_registration.py # it needs a clean process
   - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins

.buildkite/test_areas/ray_compat.yaml

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

.buildkite/test_areas/samplers.yaml

@@ -12,10 +12,3 @@ steps:
   commands:
     - pytest -v -s samplers
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
-      commands:
-      - pytest -v -s samplers

.buildkite/test_areas/spec_decode.yaml

@@ -1,40 +0,0 @@
-group: Spec Decode
-depends_on:
-  - image-build
-steps:
-- label: Spec Decode Eagle
-  timeout_in_minutes: 30
-  source_file_dependencies:
-    - vllm/v1/spec_decode/
-    - vllm/v1/worker/gpu/spec_decode/
-    - tests/v1/e2e/spec_decode/
-  commands:
-    - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
-
-- label: Spec Decode Speculators + MTP
-  timeout_in_minutes: 30
-  source_file_dependencies:
-    - vllm/v1/spec_decode/
-    - vllm/v1/worker/gpu/spec_decode/
-    - vllm/transformers_utils/configs/speculators/
-    - tests/v1/e2e/spec_decode/
-  commands:
-    - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
-
-- label: Spec Decode Ngram + Suffix
-  timeout_in_minutes: 30
-  source_file_dependencies:
-    - vllm/v1/spec_decode/
-    - vllm/v1/worker/gpu/spec_decode/
-    - tests/v1/e2e/spec_decode/
-  commands:
-    - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
-
-- label: Spec Decode Draft Model
-  timeout_in_minutes: 30
-  source_file_dependencies:
-    - vllm/v1/spec_decode/
-    - vllm/v1/worker/gpu/spec_decode/
-    - tests/v1/e2e/spec_decode/
-  commands:
-    - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"

.buildkite/test_areas/weight_loading.yaml

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

.github/.bc-linter.yml

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

.github/CODEOWNERS

@@ -2,66 +2,45 @@
 # for more info about CODEOWNERS file
 
 # This lists cover the "core" components of vLLM that require careful review
-/vllm/compilation @zou3519 @youkaichao @ProExpertProg @BoyuanFeng
+/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @njhill @22quinn
-/vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery
+/vllm/model_executor/layers/attention @LucasWilkinson
-/vllm/lora @jeejeelee
-/vllm/model_executor/layers/attention @LucasWilkinson @MatthewBonanni
 /vllm/model_executor/layers/fused_moe @mgoin @pavanimajety
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety
 /vllm/model_executor/layers/mamba @tdoublep
 /vllm/model_executor/model_loader @22quinn
 /vllm/model_executor/layers/batch_invariant.py @yewentao256 
 /vllm/multimodal @DarkLight1337 @ywang96 @NickLucche @tjtanaa
-/vllm/vllm_flash_attn @LucasWilkinson @MatthewBonanni
+/vllm/vllm_flash_attn @LucasWilkinson
+/vllm/lora @jeejeelee
+/vllm/reasoning @aarnphm @chaunceyjiang
+/vllm/entrypoints @aarnphm @chaunceyjiang
+/vllm/tool_parsers @aarnphm @chaunceyjiang
+/vllm/compilation @zou3519 @youkaichao @ProExpertProg
+/vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery
 CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
 # so spam a lot of people
 /vllm/config @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
-/vllm/config/cache.py @heheda12345
+/vllm/config/cache.py @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg @heheda12345
-
-# Entrypoints
-/vllm/entrypoints/anthropic @mgoin @DarkLight1337
-/vllm/entrypoints/cli @hmellor @mgoin @DarkLight1337 @russellb
-/vllm/entrypoints/mcp @heheda12345
-/vllm/entrypoints/openai @aarnphm @chaunceyjiang @DarkLight1337 @russellb
-/vllm/entrypoints/openai/realtime @njhill
-/vllm/entrypoints/openai/speech_to_text @NickLucche
-/vllm/entrypoints/pooling @noooop
-/vllm/entrypoints/sagemaker @DarkLight1337
-/vllm/entrypoints/serve @njhill
-/vllm/entrypoints/*.py @njhill
-/vllm/entrypoints/chat_utils.py @DarkLight1337
-/vllm/entrypoints/llm.py @DarkLight1337
-
-# Input/Output Processing
-/vllm/sampling_params.py @njhill @NickLucche
-/vllm/pooling_params.py @noooop @DarkLight1337
-/vllm/tokenizers @DarkLight1337 @njhill
-/vllm/renderers @DarkLight1337 @njhill
-/vllm/reasoning @aarnphm @chaunceyjiang
-/vllm/tool_parsers @aarnphm @chaunceyjiang
 
 # vLLM V1
-/vllm/v1/attention @LucasWilkinson @MatthewBonanni
+/vllm/v1/attention @LucasWilkinson
 /vllm/v1/attention/backend.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @njhill
 /vllm/v1/attention/backends/mla @pavanimajety
 /vllm/v1/attention/backends/flashinfer.py @mgoin @pavanimajety
 /vllm/v1/attention/backends/triton_attn.py @tdoublep
 /vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC @orozery
 /vllm/v1/sample @22quinn @houseroad @njhill
-/vllm/v1/spec_decode @benchislett @luccafong @MatthewBonanni
+/vllm/v1/spec_decode @benchislett @luccafong
 /vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
 /vllm/v1/kv_cache_interface.py @heheda12345
 /vllm/v1/kv_offload @ApostaC @orozery
-/vllm/v1/engine @njhill
+/vllm/v1/worker/gpu/kv_connector.py @orozery
-/vllm/v1/executor @njhill
+/vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery
-/vllm/v1/worker @njhill
-/vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery @NickLucche
 
 # Model runner V2
-/vllm/v1/worker/gpu @WoosukKwon @njhill
+/vllm/v1/worker/gpu @WoosukKwon
-/vllm/v1/worker/gpu/kv_connector.py @orozery
 
 # Test ownership
 /.buildkite/lm-eval-harness @mgoin 
@@ -136,8 +115,8 @@ mkdocs.yaml @hmellor
 /vllm/model_executor/models/mixtral*.py @patrickvonplaten
 /vllm/model_executor/models/voxtral*.py @patrickvonplaten
 /vllm/model_executor/models/pixtral*.py @patrickvonplaten
-/vllm/tokenizers/mistral.py @patrickvonplaten
 /vllm/transformers_utils/configs/mistral.py @patrickvonplaten
+/vllm/transformers_utils/tokenizers/mistral.py @patrickvonplaten
 
 # Kernels
 /vllm/v1/attention/ops/chunked_prefill_paged_decode.py @tdoublep
@@ -173,7 +152,9 @@ mkdocs.yaml @hmellor
 /examples/pooling @noooop
 /tests/models/*/pooling* @noooop
 /tests/entrypoints/pooling @noooop
+/vllm/entrypoints/pooling @noooop
 /vllm/config/pooler.py @noooop
+/vllm/pooling_params.py @noooop
 /vllm/model_executor/layers/pooler @noooop
 
 # Security guide and policies

.github/mergify.yml

@@ -3,7 +3,6 @@ pull_request_rules:
   description: Automatically apply documentation label
   conditions:
     - label != stale
-    - -closed
     - or:
       - files~=^[^/]+\.md$
       - files~=^docs/
@@ -27,7 +26,7 @@ pull_request_rules:
         Hi @{{author}}, the pre-commit checks have failed. Please run:
 
         ```bash 
-        uv pip install pre-commit>=4.5.1
+        uv pip install pre-commit
         pre-commit install
         pre-commit run --all-files
         ```
@@ -38,13 +37,15 @@ pull_request_rules:
 
         > [!TIP]
         > <details>
-        > <summary>Is <code>mypy</code> failing?</summary>
+        > <summary>Is <code>mypy</code> or <code>markdownlint</code> failing?</summary>
         > <br/>
-        > <code>mypy</code> is run differently in CI. If the failure is related to this check, please use the following command to run it locally:
+        > <code>mypy</code> and <code>markdownlint</code> are run differently in CI. If the failure is related to either of these checks, please use the following commands to run them locally:
         >
         > ```bash
         > # For mypy (substitute "3.10" with the failing version if needed)
         > pre-commit run --hook-stage manual mypy-3.10
+        > # For markdownlint
+        > pre-commit run --hook-stage manual markdownlint
         > ```
         > </details>
 
@@ -258,7 +259,8 @@ pull_request_rules:
       - files=benchmarks/run_structured_output_benchmark.sh
       - files=docs/features/structured_outputs.md
       - files=examples/offline_inference/structured_outputs.py
-      - files=examples/online_serving/structured_outputs/structured_outputs.py
+      - files=examples/online_serving/openai_chat_completion_structured_outputs.py
+      - files=examples/online_serving/openai_chat_completion_structured_outputs_with_reasoning.py
       - files~=^tests/v1/structured_output/
       - files=tests/v1/entrypoints/llm/test_struct_output_generate.py
       - files~=^vllm/v1/structured_output/
@@ -334,7 +336,7 @@ pull_request_rules:
     - or:
       - files~=^tests/tool_use/
       - files~=^tests/entrypoints/openai/tool_parsers/
-      - files=tests/entrypoints/openai/chat_completion/test_chat_with_tool_reasoning.py
+      - files=tests/entrypoints/openai/test_chat_with_tool_reasoning.py
       - files~=^vllm/entrypoints/openai/tool_parsers/
       - files=docs/features/tool_calling.md
       - files~=^examples/tool_chat_*
@@ -381,7 +383,7 @@ pull_request_rules:
     - or:
       - files~=^vllm/model_executor/model_loader/tensorizer.py
       - files~=^vllm/model_executor/model_loader/tensorizer_loader.py
-      - files~=^tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
+      - files~=^tests/entrypoints/openai/test_tensorizer_entrypoint.py
       - files~=^tests/model_executor/model_loader/tensorizer_loader/
   actions:
     assign:

.github/workflows/bc-lint.yml

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

.github/workflows/cleanup_pr_body.yml

@@ -19,7 +19,6 @@ jobs:
         uses: actions/setup-python@83679a892e2d95755f2dac6acb0bfd1e9ac5d548 # v6.1.0
         with:
           python-version: '3.12'
-          cache: 'pip'
 
       - name: Install Python dependencies
         run: |

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

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

.gitignore

@@ -3,8 +3,6 @@
 
 # vllm-flash-attn built from source
 vllm/vllm_flash_attn/*
-!vllm/vllm_flash_attn/__init__.py
-!vllm/vllm_flash_attn/flash_attn_interface.py
 
 # OpenAI triton kernels copied from source
 vllm/third_party/triton_kernels/*
@@ -189,9 +187,11 @@ cython_debug/
 .vscode/
 
 # Claude
+CLAUDE.md
 .claude/
 
 # Codex
+AGENTS.md
 .codex/
 
 # Cursor
@@ -238,6 +238,3 @@ ep_kernels_workspace/
 vllm/grpc/vllm_engine_pb2.py
 vllm/grpc/vllm_engine_pb2_grpc.py
 vllm/grpc/vllm_engine_pb2.pyi
-
-# Ignore generated cpu headers 
-csrc/cpu/cpu_attn_dispatch_generated.h

.pre-commit-config.yaml

@@ -13,7 +13,7 @@ repos:
     args: [--output-format, github, --fix]
   - id: ruff-format
 - repo: https://github.com/crate-ci/typos
-  rev: v1.43.5
+  rev: v1.38.1
   hooks:
   - id: typos
     args: [--force-exclude]
@@ -24,13 +24,12 @@ repos:
     exclude: 'csrc/(moe/topk_softmax_kernels.cu|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
+- repo: https://github.com/igorshubovych/markdownlint-cli
-  rev: v0.21.0
+  rev: v0.45.0
   hooks:
-  - id: markdownlint-cli2
+  - id: markdownlint
-    language_version: lts
+    exclude: '.*\.inc\.md'
-    args: [--fix]
+    stages: [manual] # Only run in CI
-    exclude: ^CLAUDE\.md$
 - repo: https://github.com/rhysd/actionlint
   rev: v1.7.7
   hooks:
@@ -56,7 +55,7 @@ repos:
       language: python
       types_or: [python, pyi]
       require_serial: true
-      additional_dependencies: ["mypy[faster-cache]==1.19.1", regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
+      additional_dependencies: [mypy==1.11.1, regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
   - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.10
     entry: python tools/pre_commit/mypy.py 1 "3.10"
@@ -128,13 +127,6 @@ repos:
     language: python
     types: [python]
     additional_dependencies: [regex]
-  # prevent use torch.cuda APIs
-  - id: check-torch-cuda-call
-    name: "Prevent new 'torch.cuda' APIs call"
-    entry: python tools/pre_commit/check_torch_cuda.py
-    language: python
-    types: [python]
-    additional_dependencies: [regex]
   - id: validate-config
     name: Validate configuration has default values and that each field has a docstring
     entry: python tools/pre_commit/validate_config.py
@@ -151,11 +143,6 @@ repos:
     name: Check attention backend documentation is up to date
     entry: python tools/pre_commit/generate_attention_backend_docs.py --check
     language: python
-  - id: check-boolean-context-manager
-    name: Check for boolean ops in with-statements
-    entry: python tools/pre_commit/check_boolean_context_manager.py
-    language: python
-    types: [python]
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

.readthedocs.yaml

@@ -9,15 +9,13 @@ build:
     python: "3.12"
   jobs:
     post_checkout:
-      # - bash docs/maybe_skip_pr_build.sh
+      - git fetch --unshallow || true
-      - git fetch origin main --unshallow --no-tags --filter=blob:none || true
-    pre_create_environment:
-      - pip install uv
-    create_environment:
-      - uv venv $READTHEDOCS_VIRTUALENV_PATH
-    install:
-      - uv pip install --python $READTHEDOCS_VIRTUALENV_PATH/bin/python --no-cache-dir -r requirements/docs.txt 
 
 mkdocs:
   configuration: mkdocs.yaml
   fail_on_warning: true
+
+# Optionally declare the Python requirements required to build your docs
+python:
+  install:
+    - requirements: requirements/docs.txt

AGENTS.md

@@ -1,113 +0,0 @@
-# Agent Instructions for vLLM
-
-> These instructions apply to **all** AI-assisted contributions to `vllm-project/vllm`.
-> Breaching these guidelines can result in automatic banning.
-
-## 1. Contribution Policy (Mandatory)
-
-### Duplicate-work checks
-
-Before proposing a PR, run these checks:
-
-```bash
-gh issue view <issue_number> --repo vllm-project/vllm --comments
-gh pr list --repo vllm-project/vllm --state open --search "<issue_number> in:body"
-gh pr list --repo vllm-project/vllm --state open --search "<short area keywords>"
-```
-
-- If an open PR already addresses the same fix, do not open another.
-- If your approach is materially different, explain the difference in the issue.
-
-### No low-value busywork PRs
-
-Do not open one-off PRs for tiny edits (single typo, isolated style change, one mutable default, etc.). Mechanical cleanups are acceptable only when bundled with substantive work.
-
-### Accountability
-
-- Pure code-agent PRs are **not allowed**. A human submitter must understand and defend the change end-to-end.
-- The submitting human must review every changed line and run relevant tests.
-- PR descriptions for AI-assisted work **must** include:
-    - Why this is not duplicating an existing PR.
-    - Test commands run and results.
-    - Clear statement that AI assistance was used.
-
-### Fail-closed behavior
-
-If work is duplicate/trivial busywork, **do not proceed**. Return a short explanation of what is missing.
-
----
-
-## 2. Development Workflow
-
-### Environment setup
-
-```bash
-# Install `uv` if you don't have it already:
-curl -LsSf https://astral.sh/uv/install.sh | sh
-
-# Always use `uv` for Python environment management:
-uv venv --python 3.12
-source .venv/bin/activate
-
-# Always make sure `pre-commit` and its hooks are installed:
-uv pip install -r requirements/lint.txt
-pre-commit install
-```
-
-### Installing dependencies
-
-```bash
-# If you are only making Python changes:
-VLLM_USE_PRECOMPILED=1 uv pip install -e .
-
-# If you are also making C/C++ changes:
-uv pip install -e .
-```
-
-### Running tests
-
-Tests require extra dependencies.
-All versions for test dependencies should be read from `requirements/test.txt`
-
-```bash
-# Install bare minimum test dependencies:
-uv pip install pytest pytest-asyncio tblib
-
-# Install additional test dependencies as needed, or install them all as follows:
-uv pip install -r requirements/test.txt
-
-# Run specific test from specific test file
-pytest tests/path/to/test.py -v -s -k test_name
-
-# Run all tests in directory
-pytest tests/path/to/dir -v -s
-```
-
-### Running linters
-
-```bash
-# Run all pre-commit hooks on staged files:
-pre-commit run
-
-# Run on all files:
-pre-commit run --all-files
-
-# Run a specific hook:
-pre-commit run ruff-check --all-files
-
-# Run mypy as it is in CI:
-pre-commit run mypy-3.10 --all-files --hook-stage manual
-```
-
-### Commit messages
-
-Add attribution using commit trailers such as `Co-authored-by:` (other projects use `Assisted-by:` or `Generated-by:`). For example:
-
-```text
-Your commit message here
-
-Co-authored-by: GitHub Copilot
-Co-authored-by: Claude
-Co-authored-by: gemini-code-assist
-Signed-off-by: Your Name <your.email@example.com>
-```

CLAUDE.md

@@ -1 +0,0 @@
-@AGENTS.md

CMakeLists.txt

@@ -37,7 +37,7 @@ install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
 set(PYTHON_SUPPORTED_VERSIONS "3.10" "3.11" "3.12" "3.13")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151")
 
 # ROCm installation prefix. Default to /opt/rocm but allow override via
 # -DROCM_PATH=/your/rocm/path when invoking cmake.
@@ -293,7 +293,6 @@ set(VLLM_EXT_SRC
   "csrc/fused_qknorm_rope_kernel.cu"
   "csrc/layernorm_quant_kernels.cu"
   "csrc/sampler.cu"
-  "csrc/topk.cu"
   "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/w8a8/int8/scaled_quant.cu"
@@ -725,7 +724,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # CUTLASS MoE kernels
 
   # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and ONLY works
-  # on Hopper). get_cutlass_(batched_)moe_mm_data should only be compiled
+  # on Hopper). get_cutlass_(pplx_)moe_mm_data should only be compiled
   # if it's possible to compile MoE kernels that use its output.
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
@@ -771,51 +770,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  # Expert-specialization MXFP8 blockscaled grouped kernels (SM100+).
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(ES_MXFP8_GROUPED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(ES_MXFP8_GROUPED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
-  endif()
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND ES_MXFP8_GROUPED_MM_ARCHS)
-    set(SRCS
-      "csrc/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm.cu"
-      "csrc/moe/mxfp8_moe/mxfp8_experts_quant.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${ES_MXFP8_GROUPED_MM_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_ES_MXFP8_GROUPED_MM_SM100=1")
-    message(STATUS "Building ES MXFP8 grouped kernels for archs: ${ES_MXFP8_GROUPED_MM_ARCHS}")
-  else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8
-        AND ES_MXFP8_GROUPED_MM_ARCHS)
-      message(STATUS "Not building ES MXFP8 grouped kernels as CUDA Compiler version is "
-                     "not >= 12.8.")
-    else()
-      message(STATUS "Not building ES MXFP8 grouped kernels as no compatible archs found "
-                     "in CUDA target architectures.")
-    endif()
-  endif()
-
-  # DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper and later)
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
-  endif()
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND DSV3_FUSED_A_GEMM_ARCHS)
-    set(DSV3_FUSED_A_GEMM_SRC "csrc/dsv3_fused_a_gemm.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${DSV3_FUSED_A_GEMM_SRC}"
-      CUDA_ARCHS "${DSV3_FUSED_A_GEMM_ARCHS}")
-    list(APPEND VLLM_EXT_SRC ${DSV3_FUSED_A_GEMM_SRC})
-    message(STATUS "Building dsv3_fused_a_gemm for archs: ${DSV3_FUSED_A_GEMM_ARCHS}")
-  else()
-    message(STATUS "Not building dsv3_fused_a_gemm as no compatible archs found "
-                   "in CUDA target architectures.")
-  endif()
-
   # moe_data.cu is used by all CUTLASS MoE kernels.
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
     cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
@@ -998,8 +952,7 @@ set(VLLM_MOE_EXT_SRC
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_MOE_EXT_SRC
     "csrc/moe/moe_wna16.cu"
-    "csrc/moe/grouped_topk_kernels.cu"
+    "csrc/moe/grouped_topk_kernels.cu")
-    "csrc/moe/router_gemm.cu")
 endif()
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
@@ -1128,27 +1081,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
                    " in CUDA target architectures")
   endif()
-
-  # DeepSeek V3 router GEMM kernel - requires SM90+
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(DSV3_ROUTER_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
-  endif()
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND DSV3_ROUTER_GEMM_ARCHS)
-    set(DSV3_ROUTER_GEMM_SRC
-      "csrc/moe/dsv3_router_gemm_entry.cu"
-      "csrc/moe/dsv3_router_gemm_float_out.cu"
-      "csrc/moe/dsv3_router_gemm_bf16_out.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${DSV3_ROUTER_GEMM_SRC}"
-      CUDA_ARCHS "${DSV3_ROUTER_GEMM_ARCHS}")
-    list(APPEND VLLM_MOE_EXT_SRC "${DSV3_ROUTER_GEMM_SRC}")
-    message(STATUS "Building DSV3 router GEMM kernel for archs: ${DSV3_ROUTER_GEMM_ARCHS}")
-  else()
-    message(STATUS "Not building DSV3 router GEMM kernel as no compatible archs found"
-                   " (requires SM90+ and CUDA >= 12.0)")
-  endif()
 endif()
 
 message(STATUS "Enabling moe extension.")

benchmarks/attention_benchmarks/README.md

@@ -187,7 +187,7 @@ python benchmark.py \
 ## Hardware Requirements
 
 | Backend | Hardware |
-| ------- | -------- |
+|---------|----------|
 | Flash/Triton/FlashInfer | Any CUDA GPU |
 | CUTLASS MLA | Blackwell (SM100+) |
 | FlashAttn MLA | Hopper (SM90+) |

benchmarks/attention_benchmarks/__init__.py

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

benchmarks/attention_benchmarks/batch_spec.py

@@ -229,40 +229,3 @@ def get_batch_stats(requests: list[BatchRequest]) -> dict:
             sum(r.kv_len for r in requests) / len(requests) if requests else 0
         ),
     }
-
-
-def get_batch_type(batch_spec: str, spec_decode_threshold: int = 8) -> str:
-    """
-    Classify a batch spec into a type string.
-
-    Args:
-        batch_spec: Batch specification string (e.g., "q2k", "8q1s1k", "2q2k_8q1s1k")
-        spec_decode_threshold: Max q_len to be considered spec-decode vs extend
-
-    Returns:
-        Type string: "prefill", "decode", "spec-decode", "extend", or "mixed (types...)"
-    """
-    requests = parse_batch_spec(batch_spec)
-
-    # Classify each request
-    types_present = set()
-    for req in requests:
-        if req.is_decode:
-            types_present.add("decode")
-        elif req.is_prefill:
-            types_present.add("prefill")
-        elif req.is_extend:
-            # Distinguish spec-decode (small q_len) from extend (chunked prefill)
-            if req.q_len <= spec_decode_threshold:
-                types_present.add("spec-decode")
-            else:
-                types_present.add("extend")
-
-    if len(types_present) == 1:
-        return types_present.pop()
-    elif len(types_present) > 1:
-        # Sort for consistent output
-        sorted_types = sorted(types_present)
-        return f"mixed ({'+'.join(sorted_types)})"
-    else:
-        return "unknown"

benchmarks/attention_benchmarks/benchmark.py

@@ -43,12 +43,9 @@ from common import (
     ModelParameterSweep,
     ParameterSweep,
     ResultsFormatter,
-    batch_spec_sort_key,
     is_mla_backend,
 )
 
-from vllm.v1.worker.workspace import init_workspace_manager
-
 
 def run_standard_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
     """Run standard attention benchmark (Flash/Triton/FlashInfer)."""
@@ -61,9 +58,7 @@ def run_mla_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
     """Run MLA benchmark with appropriate backend."""
     from mla_runner import run_mla_benchmark as run_mla
 
-    return run_mla(
+    return run_mla(config.backend, config, **kwargs)
-        config.backend, config, prefill_backend=config.prefill_backend, **kwargs
-    )
 
 
 def run_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
@@ -223,13 +218,10 @@ def run_model_parameter_sweep(
                         by_param_and_spec[key].append(r)
                         break
 
-    # Sort by param value then spec (batch_size, q_len, kv_len)
+    # Sort by param value then spec
     sorted_keys = sorted(
         by_param_and_spec.keys(),
-        key=lambda x: (
+        key=lambda x: (int(x[0]) if x[0].isdigit() else x[0], x[1]),
-            int(x[0]) if x[0].isdigit() else x[0],
-            batch_spec_sort_key(x[1]),
-        ),
     )
 
     current_param_value = None
@@ -338,7 +330,7 @@ def run_parameter_sweep(
                 by_spec[spec] = []
             by_spec[spec].append(r)
 
-    for spec in sorted(by_spec.keys(), key=batch_spec_sort_key):
+    for spec in sorted(by_spec.keys()):
         results = by_spec[spec]
         best = min(results, key=lambda r: r.mean_time)
         console.print(
@@ -444,27 +436,20 @@ def main():
     # Backend selection
     parser.add_argument(
         "--backends",
-        "--decode-backends",
         nargs="+",
-        help="Decode backends to benchmark (flash, triton, flashinfer, cutlass_mla, "
+        help="Backends to benchmark (flash, triton, flashinfer, cutlass_mla, "
         "flashinfer_mla, flashattn_mla, flashmla)",
     )
     parser.add_argument(
         "--backend",
         help="Single backend (alternative to --backends)",
     )
-    parser.add_argument(
-        "--prefill-backends",
-        nargs="+",
-        help="Prefill backends to compare (fa2, fa3, fa4). "
-        "Uses the first decode backend for impl construction.",
-    )
 
     # Batch specifications
     parser.add_argument(
         "--batch-specs",
         nargs="+",
-        default=None,
+        default=["q2k", "8q1s1k"],
         help="Batch specifications using extended grammar",
     )
 
@@ -480,21 +465,6 @@ def main():
     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",
-        default="auto",
-        choices=["auto", "fp8"],
-        help="KV cache dtype: auto or fp8",
-    )
-    parser.add_argument(
-        "--cuda-graphs",
-        action=argparse.BooleanOptionalAction,
-        default=True,
-        help=(
-            "Launch kernels with CUDA graphs to eliminate CPU overhead"
-            "in measurements (default: True)"
-        ),
-    )
 
     # Parameter sweep (use YAML config for advanced sweeps)
     parser.add_argument(
@@ -526,24 +496,15 @@ def main():
         if "description" in yaml_config:
             console.print(f"[dim]{yaml_config['description']}[/]")
 
-        # Override args with YAML values, but CLI args take precedence
+        # Override args with YAML values
-        # Check if CLI provided backends (they would be non-None and not default)
+        # (YAML takes precedence unless CLI arg was explicitly set)
-        cli_backends_provided = args.backend is not None or args.backends is not None
+        # Backend(s)
-
+        if "backend" in yaml_config:
-        # Backend(s) - only use YAML if CLI didn't specify
+            args.backend = yaml_config["backend"]
-        if not cli_backends_provided:
+            args.backends = None
-            if "backend" in yaml_config:
+        elif "backends" in yaml_config:
-                args.backend = yaml_config["backend"]
+            args.backends = yaml_config["backends"]
-                args.backends = None
+            args.backend = None
-            elif "backends" in yaml_config:
-                args.backends = yaml_config["backends"]
-                args.backend = None
-            elif "decode_backends" in yaml_config:
-                args.backends = yaml_config["decode_backends"]
-                args.backend = None
-
-        # Prefill backends (e.g., ["fa3", "fa4"])
-        args.prefill_backends = yaml_config.get("prefill_backends", None)
 
         # Check for special modes
         if "mode" in yaml_config:
@@ -553,24 +514,21 @@ def main():
 
         # Batch specs and sizes
         # Support both explicit batch_specs and generated batch_spec_ranges
-        # CLI --batch-specs takes precedence over YAML when provided.
+        if "batch_spec_ranges" in yaml_config:
-        cli_batch_specs_provided = args.batch_specs is not None
+            # Generate batch specs from ranges
-        if not cli_batch_specs_provided:
+            generated_specs = generate_batch_specs_from_ranges(
-            if "batch_spec_ranges" in yaml_config:
+                yaml_config["batch_spec_ranges"]
-                # Generate batch specs from ranges
+            )
-                generated_specs = generate_batch_specs_from_ranges(
+            # Combine with any explicit batch_specs
-                    yaml_config["batch_spec_ranges"]
+            if "batch_specs" in yaml_config:
-                )
+                args.batch_specs = yaml_config["batch_specs"] + generated_specs
-                # Combine with any explicit batch_specs
+            else:
-                if "batch_specs" in yaml_config:
+                args.batch_specs = generated_specs
-                    args.batch_specs = yaml_config["batch_specs"] + generated_specs
+            console.print(
-                else:
+                f"[dim]Generated {len(generated_specs)} batch specs from ranges[/]"
-                    args.batch_specs = generated_specs
+            )
-                console.print(
+        elif "batch_specs" in yaml_config:
-                    f"[dim]Generated {len(generated_specs)} batch specs from ranges[/]"
+            args.batch_specs = yaml_config["batch_specs"]
-                )
-            elif "batch_specs" in yaml_config:
-                args.batch_specs = yaml_config["batch_specs"]
 
         if "batch_sizes" in yaml_config:
             args.batch_sizes = yaml_config["batch_sizes"]
@@ -586,19 +544,13 @@ def main():
             args.num_kv_heads = model.get("num_kv_heads", args.num_kv_heads)
             args.block_size = model.get("block_size", args.block_size)
 
-        # Benchmark settings (top-level keys)
+        # Benchmark settings
-        if "device" in yaml_config:
+        if "benchmark" in yaml_config:
-            args.device = yaml_config["device"]
+            bench = yaml_config["benchmark"]
-        if "repeats" in yaml_config:
+            args.device = bench.get("device", args.device)
-            args.repeats = yaml_config["repeats"]
+            args.repeats = bench.get("repeats", args.repeats)
-        if "warmup_iters" in yaml_config:
+            args.warmup_iters = bench.get("warmup_iters", args.warmup_iters)
-            args.warmup_iters = yaml_config["warmup_iters"]
+            args.profile_memory = bench.get("profile_memory", args.profile_memory)
-        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"]
 
         # Parameter sweep configuration
         if "parameter_sweep" in yaml_config:
@@ -652,19 +604,10 @@ def main():
 
     # Determine backends
     backends = args.backends or ([args.backend] if args.backend else ["flash"])
-    prefill_backends = getattr(args, "prefill_backends", None)
-    if not args.batch_specs:
-        args.batch_specs = ["q2k", "8q1s1k"]
     console.print(f"Backends: {', '.join(backends)}")
-    if prefill_backends:
-        console.print(f"Prefill backends: {', '.join(prefill_backends)}")
     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}")
     console.print()
 
-    init_workspace_manager(args.device)
-
     # Run benchmarks
     all_results = []
 
@@ -717,8 +660,6 @@ def main():
                         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,
                     )
 
                     # Add decode pipeline config
@@ -871,8 +812,6 @@ def main():
             "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,
         }
         all_results = run_model_parameter_sweep(
             backends,
@@ -895,8 +834,6 @@ def main():
             "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,
         }
         all_results = run_parameter_sweep(
             backends, args.batch_specs, base_config_args, args.parameter_sweep, console
@@ -904,95 +841,37 @@ def main():
 
     else:
         # Normal mode: compare backends
-        decode_results = []
+        total = len(backends) * len(args.batch_specs)
-        prefill_results = []
 
-        # Run decode backend comparison
+        with tqdm(total=total, desc="Benchmarking") as pbar:
-        if not prefill_backends:
+            for spec in args.batch_specs:
-            # No prefill backends specified: compare decode backends as before
+                for backend in backends:
-            total = len(backends) * len(args.batch_specs)
+                    config = BenchmarkConfig(
+                        backend=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,
+                    )
 
-            with tqdm(total=total, desc="Benchmarking") as pbar:
+                    result = run_benchmark(config)
-                for spec in args.batch_specs:
+                    all_results.append(result)
-                    for backend in backends:
-                        config = BenchmarkConfig(
-                            backend=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,
-                        )
 
-                        result = run_benchmark(config)
+                    if not result.success:
-                        decode_results.append(result)
+                        console.print(f"[red]Error {backend} {spec}: {result.error}[/]")
 
-                        if not result.success:
+                    pbar.update(1)
-                            console.print(
-                                f"[red]Error {backend} {spec}: {result.error}[/]"
-                            )
 
-                        pbar.update(1)
+        # Display results
-
+        console.print("\n[bold green]Results:[/]")
-            console.print("\n[bold green]Results:[/]")
+        formatter = ResultsFormatter(console)
-            formatter = ResultsFormatter(console)
+        formatter.print_table(all_results, backends)
-            formatter.print_table(decode_results, backends)
-
-        # Run prefill backend comparison
-        if prefill_backends:
-            # Use first decode backend for impl construction
-            decode_backend = backends[0]
-            total = len(prefill_backends) * len(args.batch_specs)
-
-            console.print(
-                f"[yellow]Prefill comparison mode: "
-                f"using {decode_backend} for decode impl[/]"
-            )
-
-            with tqdm(total=total, desc="Prefill benchmarking") as pbar:
-                for spec in args.batch_specs:
-                    for pb in prefill_backends:
-                        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,
-                            prefill_backend=pb,
-                        )
-
-                        result = run_benchmark(config)
-
-                        # Label result with prefill backend name for display
-                        labeled_config = replace(result.config, backend=pb)
-                        result = replace(result, config=labeled_config)
-                        prefill_results.append(result)
-
-                        if not result.success:
-                            console.print(f"[red]Error {pb} {spec}: {result.error}[/]")
-
-                        pbar.update(1)
-
-            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
     if all_results:

benchmarks/attention_benchmarks/common.py

@@ -10,37 +10,18 @@ from dataclasses import asdict, dataclass
 from pathlib import Path
 from typing import Any
 
+import numpy as np
 import torch
-from batch_spec import get_batch_type, parse_batch_spec
 from rich.console import Console
 from rich.table import Table
 
-
-def batch_spec_sort_key(spec: str) -> tuple[int, int, int]:
-    """
-    Extract sorting key from batch spec: (batch_size, max_q_len, max_kv_len).
-
-    This ensures results are sorted by batch size first, then query length,
-    then sequence length, rather than alphabetically.
-    """
-    try:
-        requests = parse_batch_spec(spec)
-        batch_size = len(requests)
-        max_q_len = max(r.q_len for r in requests) if requests else 0
-        max_kv_len = max(r.kv_len for r in requests) if requests else 0
-        return (batch_size, max_q_len, max_kv_len)
-    except Exception:
-        # Fallback for unparsable specs
-        return (0, 0, 0)
-
-
 # Mock classes for vLLM attention infrastructure
 
 
 class MockHfConfig:
     """Mock HuggingFace config that satisfies vLLM's requirements."""
 
-    def __init__(self, mla_dims: dict, index_topk: int | None = None):
+    def __init__(self, mla_dims: dict):
         self.num_attention_heads = mla_dims["num_q_heads"]
         self.num_key_value_heads = mla_dims["num_kv_heads"]
         self.hidden_size = mla_dims["head_dim"] * mla_dims["num_q_heads"]
@@ -51,8 +32,6 @@ class MockHfConfig:
         self.qk_rope_head_dim = mla_dims["qk_rope_head_dim"]
         self.v_head_dim = mla_dims["v_head_dim"]
         self.qk_head_dim = mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"]
-        if index_topk is not None:
-            self.index_topk = index_topk
 
     def get_text_config(self):
         return self
@@ -61,7 +40,10 @@ class MockHfConfig:
 # Import AttentionLayerBase at module level to avoid circular dependencies
 try:
     from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
+
+    _HAS_ATTENTION_LAYER_BASE = True
 except ImportError:
+    _HAS_ATTENTION_LAYER_BASE = False
     AttentionLayerBase = object  # Fallback
 
 
@@ -77,7 +59,6 @@ class MockKVBProj:
         self.qk_nope_head_dim = qk_nope_head_dim
         self.v_head_dim = v_head_dim
         self.out_dim = qk_nope_head_dim + v_head_dim
-        self.weight = torch.empty(0, dtype=torch.bfloat16)
 
     def __call__(self, x: torch.Tensor) -> tuple[torch.Tensor]:
         """
@@ -101,38 +82,6 @@ class MockKVBProj:
         return (result,)  # Return as tuple to match ColumnParallelLinear API
 
 
-class MockIndexer:
-    """Mock Indexer for sparse MLA backends.
-
-    Provides topk_indices_buffer that sparse MLA backends use to determine
-    which KV cache slots to attend to for each token.
-    """
-
-    def __init__(
-        self,
-        max_num_tokens: int,
-        topk_tokens: int,
-        device: torch.device,
-    ):
-        self.topk_tokens = topk_tokens
-        self.topk_indices_buffer = torch.zeros(
-            (max_num_tokens, topk_tokens),
-            dtype=torch.int32,
-            device=device,
-        )
-
-    def fill_random_indices(self, num_tokens: int, max_kv_len: int):
-        """Fill topk_indices_buffer with random valid indices for benchmarking."""
-        indices = torch.randint(
-            0,
-            max_kv_len,
-            (num_tokens, self.topk_tokens),
-            dtype=torch.int32,
-            device=self.topk_indices_buffer.device,
-        )
-        self.topk_indices_buffer[:num_tokens] = indices
-
-
 class MockLayer(AttentionLayerBase):
     """Mock attention layer with scale parameters and impl.
 
@@ -164,6 +113,95 @@ class MockLayer(AttentionLayerBase):
         return self._kv_cache_spec
 
 
+class MockModelConfig:
+    """Mock model configuration."""
+
+    def __init__(
+        self,
+        num_q_heads: int,
+        num_kv_heads: int,
+        head_dim: int,
+        dtype: torch.dtype = torch.float16,
+        max_model_len: int = 32768,
+    ):
+        self._n_q = num_q_heads
+        self._n_kv = num_kv_heads
+        self._d = head_dim
+        self.dtype = dtype
+        self.max_model_len = max_model_len
+
+    def get_num_attention_heads(self, _=None) -> int:
+        return self._n_q
+
+    def get_num_kv_heads(self, _=None) -> int:
+        return self._n_kv
+
+    def get_head_size(self) -> int:
+        return self._d
+
+    def get_num_layers(self) -> int:
+        """Mock method for layer count queries."""
+        return 1
+
+    def get_sliding_window_for_layer(self, _layer_idx: int):
+        """Mock method for sliding window queries."""
+        return None
+
+    def get_logits_soft_cap_for_layer(self, _layer_idx: int):
+        """Mock method for logits soft cap queries."""
+        return None
+
+    def get_sm_scale_for_layer(self, _layer_idx: int) -> float:
+        """Mock method for SM scale queries."""
+        return 1.0 / (self.get_head_size() ** 0.5)
+
+
+class MockParallelConfig:
+    """Mock parallel configuration."""
+
+    pass
+
+
+class MockCompilationConfig:
+    """Mock compilation configuration."""
+
+    def __init__(self):
+        self.full_cuda_graph = False
+        self.static_forward_context = {}
+
+
+class MockVLLMConfig:
+    """Mock VLLM configuration."""
+
+    def __init__(self):
+        self.compilation_config = MockCompilationConfig()
+
+
+class MockRunner:
+    """Mock GPU runner for metadata builders."""
+
+    def __init__(
+        self,
+        seq_lens: np.ndarray,
+        query_start_locs: np.ndarray,
+        device: torch.device,
+        num_q_heads: int,
+        num_kv_heads: int,
+        head_dim: int,
+        dtype: torch.dtype,
+    ):
+        self.model_config = MockModelConfig(num_q_heads, num_kv_heads, head_dim, dtype)
+        self.parallel_config = MockParallelConfig()
+        self.vllm_config = MockVLLMConfig()
+        self.seq_lens_np = seq_lens
+        self.query_start_loc_np = query_start_locs
+        self.device = device
+        self.attention_chunk_size = None
+        self.num_query_heads = num_q_heads
+        self.num_kv_heads = num_kv_heads
+        self.dtype = dtype
+
+
 @dataclass
 class ParameterSweep:
     """Configuration for sweeping a backend parameter."""
@@ -213,11 +251,7 @@ class BenchmarkConfig:
     profile_memory: bool = False
     use_cuda_graphs: bool = False
 
-    # "auto" or "fp8"
-    kv_cache_dtype: str = "auto"
-
     # MLA-specific
-    prefill_backend: str | None = None
     kv_lora_rank: int | None = None
     qk_nope_head_dim: int | None = None
     qk_rope_head_dim: int | None = None
@@ -282,19 +316,14 @@ class ResultsFormatter:
             backends: List of backend names being compared
             compare_to_fastest: Show percentage comparison to fastest
         """
-        # Group by batch spec, preserving first-occurrence order
+        # Group by batch spec
         by_spec = {}
-        specs_order = []
         for r in results:
             spec = r.config.batch_spec
             if spec not in by_spec:
                 by_spec[spec] = {}
-                specs_order.append(spec)
             by_spec[spec][r.config.backend] = r
 
-        # Sort specs by (batch_size, q_len, kv_len) instead of alphabetically
-        specs_order = sorted(by_spec.keys(), key=batch_spec_sort_key)
-
         # Create shortened backend names for display
         def shorten_backend_name(name: str) -> str:
             """Shorten long backend names for table display."""
@@ -308,8 +337,6 @@ class ResultsFormatter:
 
         table = Table(title="Attention Benchmark Results")
         table.add_column("Batch\nSpec", no_wrap=True)
-        table.add_column("Type", no_wrap=True)
-        table.add_column("Batch\nSize", justify="right", no_wrap=True)
 
         multi = len(backends) > 1
         for backend in backends:
@@ -323,14 +350,12 @@ class ResultsFormatter:
                 table.add_column(col_rel, justify="right", no_wrap=False)
 
         # Add rows
-        for spec in specs_order:
+        for spec in sorted(by_spec.keys()):
             spec_results = by_spec[spec]
             times = {b: r.mean_time for b, r in spec_results.items() if r.success}
             best_time = min(times.values()) if times else 0.0
 
-            batch_type = get_batch_type(spec)
+            row = [spec]
-            batch_size = len(parse_batch_spec(spec))
-            row = [spec, batch_type, str(batch_size)]
             for backend in backends:
                 if backend in spec_results:
                     r = spec_results[backend]
@@ -372,7 +397,6 @@ class ResultsFormatter:
                     "backend",
                     "batch_spec",
                     "num_layers",
-                    "kv_cache_dtype",
                     "mean_time",
                     "std_time",
                     "throughput",
@@ -386,7 +410,6 @@ class ResultsFormatter:
                         "backend": r.config.backend,
                         "batch_spec": r.config.batch_spec,
                         "num_layers": r.config.num_layers,
-                        "kv_cache_dtype": r.config.kv_cache_dtype,
                         "mean_time": r.mean_time,
                         "std_time": r.std_time,
                         "throughput": r.throughput_tokens_per_sec or 0,
@@ -463,11 +486,10 @@ def get_attention_scale(head_dim: int) -> float:
 
 def is_mla_backend(backend: str) -> bool:
     """
-    Check if backend is an MLA backend using the AttentionBackendEnum.
+    Check if backend is an MLA backend using the backend's is_mla() property.
 
     Args:
-        backend: Backend name matching AttentionBackendEnum exactly
+        backend: Backend name (e.g., "CUTLASS_MLA", "FLASHINFER_MLA")
-        (e.g., "FLASHMLA_SPARSE")
 
     Returns:
         True if the backend is an MLA backend, False otherwise
@@ -475,8 +497,7 @@ def is_mla_backend(backend: str) -> bool:
     from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
     try:
-        backend_enum = AttentionBackendEnum[backend]
+        backend_class = AttentionBackendEnum[backend.upper()].get_class()
-        backend_class = backend_enum.get_class()
         return backend_class.is_mla()
-    except (KeyError, ValueError, ImportError, AttributeError):
+    except (KeyError, ValueError, ImportError):
         return False

benchmarks/attention_benchmarks/configs/mla_decode.yaml

@@ -3,7 +3,7 @@
 model:
   name: "deepseek-v3"
   num_layers: 60
-  num_q_heads: 128  # Base value, can be swept for TP simulation
+  num_q_heads: 128
   num_kv_heads: 1  # MLA uses single latent KV
   head_dim: 576
   kv_lora_rank: 512
@@ -12,13 +12,6 @@ model:
   v_head_dim: 128
   block_size: 128  # CUTLASS MLA and FlashAttn MLA use 128
 
-# Model parameter sweep: simulate tensor parallelism by varying num_q_heads
-# TP=1: 128 heads, TP=2: 64 heads, TP=4: 32 heads, TP=8: 16 heads
-model_parameter_sweep:
-  param_name: "num_q_heads"
-  values: [128, 64, 32, 16]
-  label_format: "{backend}_{value}h"
-
 batch_specs:
   # Small batches, varying sequence lengths
   - "16q1s512"     # 16 requests, 512 KV cache
@@ -41,30 +34,28 @@ batch_specs:
   # Very large batches
   - "128q1s1k"     # 128 requests, 1k KV cache
   - "128q1s2k"     # 128 requests, 2k KV cache
-  - "128q1s4k"     # 128 requests, 4k KV cache
-  - "128q1s8k"     # 128 requests, 8k KV cache
 
   # Long context
   - "32q1s16k"     # 32 requests, 16k KV cache
   - "32q1s32k"     # 32 requests, 32k KV cache
 
 backends:
-  - CUTLASS_MLA
+  - cutlass_mla
-  - FLASHINFER_MLA
+  - flashinfer_mla
-  - FLASH_ATTN_MLA  # Hopper only
+  - flashattn_mla  # Hopper only
-  - FLASHMLA        # Hopper only
+  - flashmla        # Hopper only
 
 device: "cuda:0"
-repeats: 100
+repeats: 5
-warmup_iters: 10
+warmup_iters: 3
 profile_memory: true
 
 # Backend-specific tuning
-CUTLASS_MLA:
+cutlass_mla:
   num_kv_splits: auto  # or specific value like 4, 8, 16
 
-FLASH_ATTN_MLA:
+flashattn_mla:
   reorder_batch_threshold: 512
 
-FLASHMLA:
+flashmla:
   reorder_batch_threshold: 1

benchmarks/attention_benchmarks/configs/mla_mixed_batch.yaml

@@ -30,9 +30,9 @@ batch_specs:
   - "2q16k_32q1s4k"         # 2 very large prefill + 32 decode
 
   # Context extension + decode
-  - "2q1ks2k_16q1s1k"       # 2 extend + 16 decode
+  - "2q1kkv2k_16q1s1k"       # 2 extend + 16 decode
-  - "4q2ks4k_32q1s2k"       # 4 extend + 32 decode
+  - "4q2kkv4k_32q1s2k"       # 4 extend + 32 decode
-  - "2q1ks8k_32q1s2k"       # 2 large extend + 32 decode
+  - "2q1kkv8k_32q1s2k"       # 2 large extend + 32 decode
 
   # Explicitly chunked prefill
   - "q8k"           # 8k prefill with chunking hint
@@ -45,10 +45,10 @@ batch_specs:
   - "4q4k_60q1s4k"          # 4 prefill + 60 decode
 
 backends:
-  - CUTLASS_MLA
+  - cutlass_mla
-  - FLASHINFER_MLA
+  - flashinfer_mla
-  - FLASH_ATTN_MLA   # Hopper only
+  - flashattn_mla   # Hopper only
-  - FLASHMLA         # Hopper only
+  - flashmla        # Hopper only
 
 device: "cuda:0"
 repeats: 5

benchmarks/attention_benchmarks/configs/mla_prefill.yaml

@@ -1,126 +0,0 @@
-# MLA prefill backend comparison
-#
-# Compares all available MLA prefill backends:
-#   FA backends:  fa2, fa3, fa4 (FlashAttention versions)
-#   Non-FA:       flashinfer, cudnn, trtllm (Blackwell-only, require flashinfer)
-#
-# Uses cutlass_mla as the decode backend for impl construction
-# (only the prefill path is exercised).
-#
-# Backends that aren't available on the current platform will report errors
-# in the results table (e.g., fa3 on Blackwell, cudnn without artifactory).
-#
-# Usage:
-#   python benchmark.py --config configs/mla_prefill.yaml
-
-description: "MLA prefill backend comparison"
-
-model:
-  name: "deepseek-v3"
-  num_layers: 60
-  num_q_heads: 128
-  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
-
-# 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
-
-batch_specs:
-  # Pure prefill
-  - "q512"
-  - "q1k"
-  - "q2k"
-  - "q4k"
-  - "q8k"
-
-  # Batched pure prefill
-  - "2q512"
-  - "2q1k"
-  - "2q2k"
-  - "2q4k"
-  - "2q8k"
-  - "4q512"
-  - "4q1k"
-  - "4q2k"
-  - "4q4k"
-  - "4q8k"
-  - "8q512"
-  - "8q1k"
-  - "8q2k"
-  - "8q4k"
-  - "8q8k"
-
-  # Chunked prefill / extend
-  # Short context
-  - "q128s1k"
-  - "q256s2k"
-  - "q512s4k"
-  - "q1ks4k"
-  - "q2ks8k"
-  - "2q128s1k"
-  - "2q256s2k"
-  - "2q512s4k"
-  - "2q1ks4k"
-  - "2q2ks8k"
-  - "4q128s1k"
-  - "4q256s2k"
-  - "4q512s4k"
-  - "4q1ks4k"
-  - "4q2ks8k"
-  - "8q128s1k"
-  - "8q256s2k"
-  - "8q512s4k"
-  - "8q1ks4k"
-
-  # Medium context
-  - "q128s16k"
-  - "q512s16k"
-  - "q1ks16k"
-  - "q2ks16k"
-  - "2q128s16k"
-  - "2q512s16k"
-  - "2q1ks16k"
-  - "2q2ks16k"
-  - "4q128s16k"
-  - "4q512s16k"
-  - "4q1ks16k"
-  - "4q2ks16k"
-
-  # Long context
-  - "q128s64k"
-  - "q512s64k"
-  - "q1ks64k"
-  - "q2ks64k"
-  - "2q128s64k"
-  - "2q512s64k"
-  - "2q1ks64k"
-  - "2q2ks64k"
-
-decode_backends:
-  - CUTLASS_MLA
-
-prefill_backends:
-  - fa2
-  - fa3
-  - fa4
-  - flashinfer
-  - cudnn
-  - trtllm
-
-device: "cuda:0"
-repeats: 20
-warmup_iters: 5

benchmarks/attention_benchmarks/configs/mla_sparse_decode.yaml

@@ -1,58 +0,0 @@
-# MLA decode-only benchmark configuration
-
-model:
-  name: "deepseek-v3"
-  num_layers: 60
-  num_q_heads: 128  # Base value, can be swept for TP simulation
-  num_kv_heads: 1  # MLA uses single latent KV
-  head_dim: 576
-  kv_lora_rank: 512
-  qk_nope_head_dim: 128
-  qk_rope_head_dim: 64
-  v_head_dim: 128
-  block_size: 128  # CUTLASS MLA and FlashAttn MLA use 128
-
-# Model parameter sweep: simulate tensor parallelism by varying num_q_heads
-# TP=1: 128 heads, TP=2: 64 heads, TP=4: 32 heads, TP=8: 16 heads
-model_parameter_sweep:
-  param_name: "num_q_heads"
-  values: [128, 64, 32, 16]
-  label_format: "{backend}_{value}h"
-
-batch_specs:
-  # Small batches, varying sequence lengths
-  - "16q1s512"     # 16 requests, 512 KV cache
-  - "16q1s1k"      # 16 requests, 1k KV cache
-  - "16q1s2k"      # 16 requests, 2k KV cache
-  - "16q1s4k"      # 16 requests, 4k KV cache
-
-  # Medium batches
-  - "32q1s1k"      # 32 requests, 1k KV cache
-  - "32q1s2k"      # 32 requests, 2k KV cache
-  - "32q1s4k"      # 32 requests, 4k KV cache
-  - "32q1s8k"      # 32 requests, 8k KV cache
-
-  # Large batches
-  - "64q1s1k"      # 64 requests, 1k KV cache
-  - "64q1s2k"      # 64 requests, 2k KV cache
-  - "64q1s4k"      # 64 requests, 4k KV cache
-  - "64q1s8k"      # 64 requests, 8k KV cache
-
-  # Very large batches
-  - "128q1s1k"     # 128 requests, 1k KV cache
-  - "128q1s2k"     # 128 requests, 2k KV cache
-  - "128q1s4k"     # 128 requests, 4k KV cache
-  - "128q1s8k"     # 128 requests, 8k KV cache
-
-  # Long context
-  - "32q1s16k"     # 32 requests, 16k KV cache
-  - "32q1s32k"     # 32 requests, 32k KV cache
-
-backends:
-  - FLASHMLA_SPARSE
-  - FLASHINFER_MLA_SPARSE
-
-device: "cuda:0"
-repeats: 100
-warmup_iters: 10
-profile_memory: true

benchmarks/attention_benchmarks/configs/mla_sparse_prefill.yaml

@@ -1,62 +0,0 @@
-# MLA prefill-only benchmark configuration for sparse backends
-
-model:
-  name: "deepseek-v3"
-  num_layers: 60
-  num_q_heads: 128
-  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
-
-# Model parameter sweep: simulate tensor parallelism by varying num_q_heads
-# TP=1: 128 heads, TP=2: 64 heads, TP=4: 32 heads, TP=8: 16 heads
-model_parameter_sweep:
-  param_name: "num_q_heads"
-  values: [128, 64, 32, 16]
-  label_format: "{backend}_{value}h"
-
-batch_specs:
-  # Pure prefill
-  - "1q512"
-  - "1q1k"
-  - "1q2k"
-  - "1q4k"
-  - "1q8k"
-
-  # Batched pure prefill
-  - "2q512"
-  - "2q1k"
-  - "2q2k"
-  - "2q4k"
-  - "2q8k"
-  - "4q512"
-  - "4q1k"
-  - "4q2k"
-  - "4q4k"
-  - "4q8k"
-  - "8q512"
-  - "8q1k"
-  - "8q2k"
-  - "8q4k"
-  - "8q8k"
-
-  # Extend
-  - "1q512s4k"
-  - "1q512s8k"
-  - "1q1ks8k"
-  - "1q2ks8k"
-  - "1q2ks16k"
-  - "1q4ks16k"
-
-backends:
-  - FLASHMLA_SPARSE
-  - FLASHINFER_MLA_SPARSE
-
-device: "cuda:0"
-repeats: 10
-warmup_iters: 3
-profile_memory: true

benchmarks/attention_benchmarks/configs/reorder_threshold.yaml

@@ -6,7 +6,7 @@
 description: "Decode vs Prefill pipeline crossover analysis"
 
 # Test FlashAttn MLA
-backend: FLASH_ATTN_MLA
+backend: flashattn_mla
 
 # Mode: decode_vs_prefill comparison (special sweep mode)
 # For each batch spec, we'll test both decode and prefill pipelines
@@ -62,10 +62,11 @@ model:
   block_size: 128
 
 # Benchmark settings
-device: "cuda:0"
+benchmark:
-repeats: 15          # More repeats for spec decode variance
+  device: "cuda:0"
-warmup_iters: 5
+  repeats: 15          # More repeats for spec decode variance
-profile_memory: false
+  warmup_iters: 5
+  profile_memory: false
 
 # Output
 output:

benchmarks/attention_benchmarks/configs/speculative_decode.yaml

@@ -41,17 +41,18 @@ batch_specs:
 
 # Backends that support query length > 1
 backends:
-  - FLASH_ATTN_MLA    # reorder_batch_threshold = 512
+  - flashattn_mla    # reorder_batch_threshold = 512
-  - FLASHMLA          # reorder_batch_threshold = 1 (tunable)
+  - flashmla          # reorder_batch_threshold = 1 (tunable)
 
 # FlashInfer-MLA also supports uniform spec-as-decode but with different mechanism
-# - FLASHINFER_MLA
+# - flashinfer_mla
 
 # Benchmark settings
-device: "cuda:0"
+benchmark:
-repeats: 10  # More repeats for statistical significance
+  device: "cuda:0"
-warmup_iters: 5
+  repeats: 10  # More repeats for statistical significance
-profile_memory: false
+  warmup_iters: 5
+  profile_memory: false
 
 # Test these threshold values for optimization
 parameter_sweep:

benchmarks/attention_benchmarks/configs/standard_attention.yaml

@@ -25,22 +25,14 @@ batch_specs:
   - "4q1k_16q1s2k"     # 4 prefill + 16 decode
   - "2q4k_32q1s1k"     # 2 large prefill + 32 decode
 
-  # Speculative decode (q <= 8)
+  # Context extension
-  - "16q2s1k"         # 16 requests, 2 spec tokens, 1k KV cache
+  - "q1ks2k"          # 1k query, 2k sequence (chunked prefill)
-  - "16q4s1k"         # 16 requests, 4 spec tokens, 1k KV cache
-  - "16q8s1k"         # 16 requests, 8 spec tokens, 1k KV cache
-  - "32q4s2k"         # 32 requests, 4 spec tokens, 2k KV cache
-  - "8q8s4k"          # 8 requests, 8 spec tokens, 4k KV cache
-
-  # Context extension (chunked prefill)
-  - "q1ks2k"          # 1k query, 2k sequence
   - "2q1ks4k"         # 2 requests: 1k query, 4k sequence
 
-# Available backends: FLASH_ATTN, TRITON_ATTN, FLASHINFER
 backends:
-  - FLASH_ATTN
+  - flash
-  - TRITON_ATTN
+  - triton
-  - FLASHINFER
+  - flashinfer
 
 device: "cuda:0"
 repeats: 5

benchmarks/attention_benchmarks/mla_runner.py

@@ -8,13 +8,14 @@ This module provides helpers for running MLA backends without
 needing full VllmConfig integration.
 """
 
+import importlib
+
 import numpy as np
 import torch
 from batch_spec import parse_batch_spec
 from common import (
     BenchmarkResult,
     MockHfConfig,
-    MockIndexer,
     MockKVBProj,
     MockLayer,
     setup_mla_dims,
@@ -60,11 +61,7 @@ def create_minimal_vllm_config(
     model_name: str = "deepseek-v3",
     block_size: int = 128,
     max_num_seqs: int = 256,
-    max_num_batched_tokens: int = 8192,
     mla_dims: dict | None = None,
-    index_topk: int | None = None,
-    prefill_backend: str | None = None,
-    kv_cache_dtype: str = "auto",
 ) -> VllmConfig:
     """
     Create minimal VllmConfig for MLA benchmarks.
@@ -76,11 +73,6 @@ def create_minimal_vllm_config(
         max_num_seqs: Maximum number of sequences
         mla_dims: Optional custom MLA dimensions dict. If not provided, uses
                   setup_mla_dims(model_name)
-        index_topk: Optional topk value for sparse MLA backends. If provided,
-                    the config will include index_topk for sparse attention.
-        prefill_backend: Prefill backend name (e.g., "fa3", "fa4", "flashinfer",
-                        "cudnn", "trtllm"). Configures the attention config to
-                        force the specified prefill backend.
 
     Returns:
         VllmConfig for benchmarking
@@ -90,7 +82,7 @@ def create_minimal_vllm_config(
         mla_dims = setup_mla_dims(model_name)
 
     # Create mock HF config first (avoids downloading from HuggingFace)
-    mock_hf_config = MockHfConfig(mla_dims, index_topk=index_topk)
+    mock_hf_config = MockHfConfig(mla_dims)
 
     # Create a temporary minimal config.json to avoid HF downloads
     # This ensures consistent ModelConfig construction without network access
@@ -128,12 +120,16 @@ def create_minimal_vllm_config(
             seed=0,
             max_model_len=32768,
             quantization=None,
+            quantization_param_path=None,
             enforce_eager=False,
+            max_context_len_to_capture=None,
+            max_seq_len_to_capture=8192,
             max_logprobs=20,
             disable_sliding_window=False,
             skip_tokenizer_init=True,
             served_model_name=None,
             limit_mm_per_prompt=None,
+            use_async_output_proc=True,
             config_format="auto",
         )
     finally:
@@ -151,13 +147,14 @@ def create_minimal_vllm_config(
     cache_config = CacheConfig(
         block_size=block_size,
         gpu_memory_utilization=0.9,
-        cache_dtype=kv_cache_dtype,
+        swap_space=0,
+        cache_dtype="auto",
         enable_prefix_caching=False,
     )
 
     scheduler_config = SchedulerConfig(
         max_num_seqs=max_num_seqs,
-        max_num_batched_tokens=max(max_num_batched_tokens, max_num_seqs),
+        max_num_batched_tokens=8192,
         max_model_len=32768,
         is_encoder_decoder=False,
         enable_chunked_prefill=True,
@@ -169,7 +166,7 @@ def create_minimal_vllm_config(
 
     compilation_config = CompilationConfig()
 
-    vllm_config = VllmConfig(
+    return VllmConfig(
         model_config=model_config,
         cache_config=cache_config,
         parallel_config=parallel_config,
@@ -177,147 +174,62 @@ def create_minimal_vllm_config(
         compilation_config=compilation_config,
     )
 
-    if prefill_backend is not None:
-        prefill_cfg = get_prefill_backend_config(prefill_backend)
-        if prefill_cfg["flash_attn_version"] is not None:
-            vllm_config.attention_config.flash_attn_version = prefill_cfg[
-                "flash_attn_version"
-            ]
-        vllm_config.attention_config.disable_flashinfer_prefill = prefill_cfg[
-            "disable_flashinfer_prefill"
-        ]
-        vllm_config.attention_config.use_cudnn_prefill = prefill_cfg[
-            "use_cudnn_prefill"
-        ]
-        vllm_config.attention_config.use_trtllm_ragged_deepseek_prefill = prefill_cfg[
-            "use_trtllm_ragged_deepseek_prefill"
-        ]
-
-    return vllm_config
-
 
 # ============================================================================
-# Prefill Backend Configuration
+# Backend Configuration
 # ============================================================================
 
-# Maps prefill backend names to attention config overrides.
+
-# FA backends set flash_attn_version and disable non-FA paths.
+# Backend name to class name prefix mapping
-# Non-FA backends enable their specific path and disable others.
+_BACKEND_NAME_MAP = {
-_PREFILL_BACKEND_CONFIG: dict[str, dict] = {
+    "flashattn_mla": "FlashAttnMLA",
-    "fa2": {
+    "flashmla": "FlashMLA",
-        "flash_attn_version": 2,
+    "flashinfer_mla": "FlashInferMLA",
-        "disable_flashinfer_prefill": True,
+    "cutlass_mla": "CutlassMLA",
-        "use_cudnn_prefill": False,
-        "use_trtllm_ragged_deepseek_prefill": False,
-    },
-    "fa3": {
-        "flash_attn_version": 3,
-        "disable_flashinfer_prefill": True,
-        "use_cudnn_prefill": False,
-        "use_trtllm_ragged_deepseek_prefill": False,
-    },
-    "fa4": {
-        "flash_attn_version": 4,
-        "disable_flashinfer_prefill": True,
-        "use_cudnn_prefill": False,
-        "use_trtllm_ragged_deepseek_prefill": False,
-    },
-    "flashinfer": {
-        "flash_attn_version": None,
-        "disable_flashinfer_prefill": False,
-        "use_cudnn_prefill": False,
-        "use_trtllm_ragged_deepseek_prefill": False,
-    },
-    "cudnn": {
-        "flash_attn_version": None,
-        "disable_flashinfer_prefill": True,
-        "use_cudnn_prefill": True,
-        "use_trtllm_ragged_deepseek_prefill": False,
-    },
-    "trtllm": {
-        "flash_attn_version": None,
-        "disable_flashinfer_prefill": True,
-        "use_cudnn_prefill": False,
-        "use_trtllm_ragged_deepseek_prefill": True,
-    },
 }
 
-
+# Special properties that differ from defaults
-def get_prefill_backend_config(prefill_backend: str) -> dict:
-    """Get attention config overrides for a prefill backend."""
-    if prefill_backend not in _PREFILL_BACKEND_CONFIG:
-        raise ValueError(
-            f"Unknown prefill backend: {prefill_backend!r}. "
-            f"Available: {list(_PREFILL_BACKEND_CONFIG.keys())}"
-        )
-    return _PREFILL_BACKEND_CONFIG[prefill_backend]
-
-
-# ============================================================================
-# Decode Backend Configuration
-# ============================================================================
-
-
-# Backend-specific properties that can't be inferred from the backend class
-# Keys are AttentionBackendEnum names (uppercase)
 _BACKEND_PROPERTIES = {
-    "FLASHMLA": {
+    "flashmla": {
         "query_format": "concat",  # Single concatenated tensor (vs tuple)
+        "block_size": 64,  # FlashMLA uses fixed block size
     },
-    "FLASHMLA_SPARSE": {
+    "flashinfer_mla": {
-        "query_format": "concat",  # Single concatenated tensor (vs tuple)
+        "block_size": 64,  # FlashInfer MLA only supports 32 or 64
     },
 }
 
 
 def _get_backend_config(backend: str) -> dict:
     """
-    Get backend configuration from AttentionBackendEnum.
+    Get backend configuration using naming conventions.
 
-    Uses the registry to get the backend class and extract configuration
+    All MLA backends follow the pattern:
-    from its methods (get_impl_cls, get_builder_cls, is_sparse, etc.).
+    - Module: vllm.v1.attention.backends.mla.{backend}
-
+    - Impl: {Name}Impl
-    Args:
+    - Metadata: {Name}Metadata (or MLACommonMetadata)
-        backend: Backend name matching AttentionBackendEnum exactly
+    - DecodeMetadata: {Name}DecodeMetadata (or MLACommonDecodeMetadata)
-        (e.g., "FLASHMLA_SPARSE")
+    - MetadataBuilder: {Name}MetadataBuilder
-
-    Returns:
-        Dict with backend configuration
     """
-    from vllm.v1.attention.backend import MultipleOf
+    if backend not in _BACKEND_NAME_MAP:
-    from vllm.v1.attention.backends.registry import AttentionBackendEnum
+        raise ValueError(f"Unknown backend: {backend}")
 
-    try:
+    name = _BACKEND_NAME_MAP[backend]
-        backend_enum = AttentionBackendEnum[backend]
-        backend_class = backend_enum.get_class()
-    except (KeyError, ValueError) as e:
-        valid_backends = [e.name for e in AttentionBackendEnum if e.name != "CUSTOM"]
-        raise ValueError(
-            f"Unknown backend: {backend}. "
-            f"Valid MLA backends: {[b for b in valid_backends if 'MLA' in b]}"
-        ) from e
-
-    # Get block size from backend class
-    block_sizes = backend_class.get_supported_kernel_block_sizes()
-    # Use first supported block size (backends typically support one for MLA)
-    block_size = block_sizes[0] if block_sizes else None
-    if isinstance(block_size, MultipleOf):
-        # No fixed block size; fall back to config value
-        block_size = None
-
-    # Check if sparse via class method if available
-    is_sparse = getattr(backend_class, "is_sparse", lambda: False)()
-
-    # Get properties that can't be inferred
     props = _BACKEND_PROPERTIES.get(backend, {})
 
+    # Check if backend uses common metadata (FlashInfer, CUTLASS)
+    uses_common = backend in ("flashinfer_mla", "cutlass_mla")
+
     return {
-        "backend_class": backend_class,
+        "module": f"vllm.v1.attention.backends.mla.{backend}",
-        "impl_class": backend_class.get_impl_cls(),
+        "impl_class": f"{name}Impl",
-        "builder_class": backend_class.get_builder_cls(),
+        "metadata_class": "MLACommonMetadata" if uses_common else f"{name}Metadata",
+        "decode_metadata_class": "MLACommonDecodeMetadata"
+        if uses_common
+        else f"{name}DecodeMetadata",
+        "builder_class": f"{name}MetadataBuilder",
         "query_format": props.get("query_format", "tuple"),
-        "block_size": block_size,
+        "block_size": props.get("block_size", None),
-        "is_sparse": is_sparse,
     }
 
 
@@ -535,27 +447,22 @@ def _create_backend_impl(
     mla_dims: dict,
     vllm_config: VllmConfig,
     device: torch.device,
-    max_num_tokens: int = 8192,
-    index_topk: int | None = None,
-    kv_cache_dtype: str = "auto",
 ):
     """
     Create backend implementation instance.
 
     Args:
-        backend_cfg: Backend configuration dict from _get_backend_config()
+        backend_cfg: Backend configuration dict
         mla_dims: MLA dimension configuration
         vllm_config: VllmConfig instance
         device: Target device
-        max_num_tokens: Maximum number of tokens for sparse indexer buffer
-        index_topk: Topk value for sparse MLA backends
 
     Returns:
-        Tuple of (impl, layer, builder_instance, indexer)
+        Tuple of (impl, layer, builder_instance)
     """
-    # Get classes from backend config (already resolved by _get_backend_config)
+    # Import backend classes
-    impl_class = backend_cfg["impl_class"]
+    backend_module = importlib.import_module(backend_cfg["module"])
-    builder_class = backend_cfg["builder_class"]
+    impl_class = getattr(backend_module, backend_cfg["impl_class"])
 
     # Calculate scale
     scale = 1.0 / np.sqrt(mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"])
@@ -567,44 +474,26 @@ def _create_backend_impl(
         v_head_dim=mla_dims["v_head_dim"],
     )
 
-    # Create indexer for sparse backends
-    indexer = None
-    if backend_cfg.get("is_sparse", False):
-        if index_topk is None:
-            index_topk = 2048  # Default topk for sparse MLA
-        indexer = MockIndexer(
-            max_num_tokens=max_num_tokens,
-            topk_tokens=index_topk,
-            device=device,
-        )
-
-    # Build impl kwargs
-    impl_kwargs = {
-        "num_heads": mla_dims["num_q_heads"],
-        "head_size": mla_dims["head_dim"],
-        "scale": scale,
-        "num_kv_heads": mla_dims["num_kv_heads"],
-        "alibi_slopes": None,
-        "sliding_window": None,
-        "kv_cache_dtype": kv_cache_dtype,
-        "logits_soft_cap": None,
-        "attn_type": "decoder",
-        "kv_sharing_target_layer_name": None,
-        "q_lora_rank": None,
-        "kv_lora_rank": mla_dims["kv_lora_rank"],
-        "qk_nope_head_dim": mla_dims["qk_nope_head_dim"],
-        "qk_rope_head_dim": mla_dims["qk_rope_head_dim"],
-        "qk_head_dim": mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"],
-        "v_head_dim": mla_dims["v_head_dim"],
-        "kv_b_proj": mock_kv_b_proj,
-    }
-
-    # Add indexer for sparse backends
-    if indexer is not None:
-        impl_kwargs["indexer"] = indexer
-
     # Create impl
-    impl = impl_class(**impl_kwargs)
+    impl = impl_class(
+        num_heads=mla_dims["num_q_heads"],
+        head_size=mla_dims["head_dim"],
+        scale=scale,
+        num_kv_heads=mla_dims["num_kv_heads"],
+        alibi_slopes=None,
+        sliding_window=None,
+        kv_cache_dtype="auto",
+        logits_soft_cap=None,
+        attn_type="decoder",
+        kv_sharing_target_layer_name=None,
+        q_lora_rank=None,
+        kv_lora_rank=mla_dims["kv_lora_rank"],
+        qk_nope_head_dim=mla_dims["qk_nope_head_dim"],
+        qk_rope_head_dim=mla_dims["qk_rope_head_dim"],
+        qk_head_dim=mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"],
+        v_head_dim=mla_dims["v_head_dim"],
+        kv_b_proj=mock_kv_b_proj,
+    )
 
     # Initialize DCP attributes
     if not hasattr(impl, "dcp_world_size") or impl.dcp_world_size in (None, -1):
@@ -626,7 +515,9 @@ def _create_backend_impl(
 
     # Create builder instance if needed
     builder_instance = None
-    if builder_class:
+    if backend_cfg["builder_class"]:
+        builder_class = getattr(backend_module, backend_cfg["builder_class"])
+
         # Populate static_forward_context so builder can find the layer
         # MockLayer inherits from AttentionLayerBase, so isinstance checks pass
         vllm_config.compilation_config.static_forward_context = {"placeholder": layer}
@@ -638,7 +529,7 @@ def _create_backend_impl(
             device=device,
         )
 
-    return impl, layer, builder_instance, indexer
+    return impl, layer, builder_instance
 
 
 # ============================================================================
@@ -703,8 +594,6 @@ def _run_single_benchmark(
     backend_cfg: dict,
     mla_dims: dict,
     device: torch.device,
-    indexer=None,
-    kv_cache_dtype: str | None = None,
 ) -> BenchmarkResult:
     """
     Run a single benchmark iteration.
@@ -717,7 +606,6 @@ def _run_single_benchmark(
         backend_cfg: Backend configuration dict
         mla_dims: MLA dimension configuration
         device: Target device
-        indexer: Optional MockIndexer for sparse backends
 
     Returns:
         BenchmarkResult with timing statistics
@@ -725,9 +613,7 @@ def _run_single_benchmark(
     # Parse batch spec
     requests = parse_batch_spec(config.batch_spec)
     q_lens = [r.q_len for r in requests]
-    kv_lens = [r.kv_len for r in requests]
     total_q = sum(q_lens)
-    max_kv_len = max(kv_lens)
 
     # Determine block size
     block_size = backend_cfg["block_size"] or config.block_size
@@ -738,123 +624,45 @@ def _run_single_benchmark(
     )
 
     # Create KV cache
-    if kv_cache_dtype is None:
+    kv_cache = torch.zeros(
-        kv_cache_dtype = getattr(config, "kv_cache_dtype", "auto")
+        num_blocks,
-    head_size = mla_dims["kv_lora_rank"] + mla_dims["qk_rope_head_dim"]
+        block_size,
-    if kv_cache_dtype == "fp8_ds_mla":
+        mla_dims["kv_lora_rank"] + mla_dims["qk_rope_head_dim"],
-        # FlashMLA sparse custom format: 656 bytes per token, stored as uint8.
+        device=device,
-        # Layout: kv_lora_rank fp8 bytes + 4 float32 tile scales
+        dtype=torch.bfloat16,
-        #         + 2*rope_dim bf16 bytes
+    )
-        # = 512 + 16 + 128 = 656 bytes for DeepSeek dims.
-        kv_cache = torch.zeros(
-            num_blocks,
-            block_size,
-            656,
-            device=device,
-            dtype=torch.uint8,
-        )
-    elif kv_cache_dtype == "fp8":
-        from vllm.platforms import current_platform
 
-        kv_cache = torch.zeros(
+    # Create input tensors for both decode and prefill modes
-            num_blocks,
+    decode_inputs, prefill_inputs = _create_input_tensors(
-            block_size,
+        total_q,
-            head_size,
+        mla_dims,
-            device=device,
+        backend_cfg["query_format"],
-            dtype=torch.uint8,
+        device,
-        ).view(current_platform.fp8_dtype())
+        torch.bfloat16,
+    )
+
+    # Determine which forward method to use based on metadata
+    if metadata.decode is not None:
+        forward_fn = lambda: impl._forward_decode(
+            decode_inputs, kv_cache, metadata, layer
+        )
+    elif metadata.prefill is not None:
+        forward_fn = lambda: impl._forward_prefill(
+            prefill_inputs["q"],
+            prefill_inputs["k_c_normed"],
+            prefill_inputs["k_pe"],
+            kv_cache,
+            metadata,
+            prefill_inputs["k_scale"],
+            prefill_inputs["output"],
+        )
     else:
-        kv_cache = torch.zeros(
-            num_blocks,
-            block_size,
-            head_size,
-            device=device,
-            dtype=torch.bfloat16,
-        )
-
-    # Fill indexer with random indices for sparse backends
-    is_sparse = backend_cfg.get("is_sparse", False)
-    if is_sparse and indexer is not None:
-        indexer.fill_random_indices(total_q, max_kv_len)
-
-    # Determine which forward methods to use based on metadata.
-    # Sparse MLA backends always use forward_mqa
-    has_decode = is_sparse or getattr(metadata, "decode", None) is not None
-    has_prefill = not is_sparse and getattr(metadata, "prefill", None) is not None
-    if not has_decode and not has_prefill:
         raise RuntimeError("Metadata has neither decode nor prefill metadata")
 
-    num_decode = (
-        metadata.num_decode_tokens
-        if (has_decode and has_prefill)
-        else total_q
-        if has_decode
-        else 0
-    )
-    num_prefill = total_q - num_decode
-
-    # Some backends requires fp8 queries when using fp8 KV cache.
-    is_fp8_kvcache = kv_cache_dtype.startswith("fp8")
-    quantize_query = is_fp8_kvcache and getattr(
-        impl, "supports_quant_query_input", False
-    )
-
-    # quantize_query forces concat format
-    query_fmt = "concat" if quantize_query else backend_cfg["query_format"]
-
-    # Create decode query tensors
-    if has_decode:
-        decode_inputs, _ = _create_input_tensors(
-            num_decode, mla_dims, query_fmt, device, torch.bfloat16
-        )
-        # Cast decode query to fp8 if the backend supports it
-        if quantize_query:
-            from vllm.platforms import current_platform
-
-            if isinstance(decode_inputs, tuple):
-                decode_inputs = torch.cat(list(decode_inputs), dim=-1)
-            decode_inputs = decode_inputs.to(current_platform.fp8_dtype())
-
-    # Create prefill input tensors
-    if has_prefill:
-        _, prefill_inputs = _create_input_tensors(
-            num_prefill, mla_dims, query_fmt, device, torch.bfloat16
-        )
-
-    # Build forward function
-    def forward_fn():
-        results = []
-        if has_decode:
-            results.append(impl.forward_mqa(decode_inputs, kv_cache, metadata, layer))
-        if has_prefill:
-            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"],
-                )
-            )
-        return results[0] if len(results) == 1 else tuple(results)
-
     # Warmup
     for _ in range(config.warmup_iters):
         forward_fn()
-    torch.accelerator.synchronize()
+    torch.cuda.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
 
     # Benchmark
     times = []
@@ -864,10 +672,10 @@ def _run_single_benchmark(
 
         start.record()
         for _ in range(config.num_layers):
-            benchmark_fn()
+            forward_fn()
         end.record()
 
-        torch.accelerator.synchronize()
+        torch.cuda.synchronize()
         elapsed_ms = start.elapsed_time(end)
         times.append(elapsed_ms / 1000.0 / config.num_layers)
 
@@ -885,26 +693,20 @@ def _run_single_benchmark(
 def _run_mla_benchmark_batched(
     backend: str,
     configs_with_params: list[tuple],  # [(config, threshold, num_splits), ...]
-    index_topk: int = 2048,
-    prefill_backend: str | None = None,
 ) -> list[BenchmarkResult]:
     """
     Unified batched MLA benchmark runner for all backends.
 
-    Works for: flashattn_mla, flashmla, flashinfer_mla, cutlass_mla,
+    Works for: flashattn_mla, flashmla, flashinfer_mla, cutlass_mla
-               flashinfer_mla_sparse, flashmla_sparse
 
     This function reuses backend initialization across multiple benchmarks
     to avoid setup/teardown overhead.
 
     Args:
-        backend: Backend name (decode backend used for impl construction)
+        backend: Backend name
         configs_with_params: List of (config, threshold, num_splits) tuples
             - threshold: reorder_batch_threshold (FlashAttn/FlashMLA only)
             - num_splits: num_kv_splits (CUTLASS only)
-        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.
 
     Returns:
         List of BenchmarkResult objects
@@ -914,7 +716,7 @@ def _run_mla_benchmark_batched(
 
     backend_cfg = _get_backend_config(backend)
     device = torch.device(configs_with_params[0][0].device)
-    torch.accelerator.set_device_index(device)
+    torch.cuda.set_device(device)
 
     # Determine block size
     config_block_size = configs_with_params[0][0].block_size
@@ -928,94 +730,21 @@ def _run_mla_benchmark_batched(
     if mla_dims is None:
         mla_dims = setup_mla_dims("deepseek-v3")
 
-    # Determine if this is a sparse backend
-    is_sparse = backend_cfg.get("is_sparse", False)
-
-    # Extract kv_cache_dtype from the first config
-    kv_cache_dtype = getattr(first_config, "kv_cache_dtype", "auto")
-
-    # FlashMLA sparse only supports "fp8_ds_mla" internally (not generic "fp8").
-    # Remap here so the user can pass --kv-cache-dtype fp8 regardless of backend.
-    if backend.upper() == "FLASHMLA_SPARSE" and kv_cache_dtype == "fp8":
-        kv_cache_dtype = "fp8_ds_mla"
-
-    # Compute max total_q across all configs so the metadata builder buffer
-    # and scheduler config are large enough for all batch specs.
-    max_total_q = max(
-        sum(r.q_len for r in parse_batch_spec(cfg.batch_spec))
-        for cfg, *_ in configs_with_params
-    )
-
     # Create and set vLLM config for MLA (reused across all benchmarks)
     vllm_config = create_minimal_vllm_config(
         model_name="deepseek-v3",  # Used only for model path
         block_size=block_size,
-        max_num_batched_tokens=max_total_q,
         mla_dims=mla_dims,  # Use custom dims from config or default
-        index_topk=index_topk if is_sparse else None,
-        prefill_backend=prefill_backend,
-        kv_cache_dtype=kv_cache_dtype,
     )
 
     results = []
 
     with set_current_vllm_config(vllm_config):
-        # Clear cached prefill backend detection functions so they re-evaluate
+        # Create backend impl, layer, and builder (reused across benchmarks)
-        # with the current VllmConfig. These are @functools.cache decorated and
+        impl, layer, builder_instance = _create_backend_impl(
-        # would otherwise return stale results from a previous backend's config.
+            backend_cfg, mla_dims, vllm_config, device
-        from vllm.model_executor.layers.attention.mla_attention import (
-            use_cudnn_prefill,
-            use_flashinfer_prefill,
-            use_trtllm_ragged_deepseek_prefill,
         )
 
-        use_flashinfer_prefill.cache_clear()
-        use_cudnn_prefill.cache_clear()
-        use_trtllm_ragged_deepseek_prefill.cache_clear()
-
-        # Create backend impl, layer, builder, and indexer (reused across benchmarks)
-        impl, layer, builder_instance, indexer = _create_backend_impl(
-            backend_cfg,
-            mla_dims,
-            vllm_config,
-            device,
-            max_num_tokens=max_total_q,
-            index_topk=index_topk if is_sparse else None,
-            kv_cache_dtype=kv_cache_dtype,
-        )
-
-        # Verify the actual prefill backend matches what was requested
-        if prefill_backend is not None:
-            prefill_cfg = get_prefill_backend_config(prefill_backend)
-            fa_version = prefill_cfg["flash_attn_version"]
-
-            if fa_version is not None:
-                # FA backend: verify the impl's FA version
-                actual_fa_version = getattr(impl, "vllm_flash_attn_version", None)
-                if actual_fa_version != fa_version:
-                    raise RuntimeError(
-                        f"Prefill backend '{prefill_backend}' requested FA "
-                        f"version {fa_version}, but the impl is using FA "
-                        f"version {actual_fa_version}. Check "
-                        f"vllm/v1/attention/backends/fa_utils.py."
-                    )
-            else:
-                # Non-FA backend: verify the builder picked the right path
-                expected_flags = {
-                    "flashinfer": "_use_fi_prefill",
-                    "cudnn": "_use_cudnn_prefill",
-                    "trtllm": "_use_trtllm_ragged_prefill",
-                }
-                flag_name = expected_flags.get(prefill_backend)
-                if flag_name and not getattr(builder_instance, flag_name, False):
-                    raise RuntimeError(
-                        f"Prefill backend '{prefill_backend}' was requested "
-                        f"but the metadata builder did not enable it. This "
-                        f"usually means a dependency is missing (e.g., "
-                        f"flashinfer not installed) or the platform doesn't "
-                        f"support it."
-                    )
-
         # Run each benchmark with the shared impl
         for config, threshold, num_splits in configs_with_params:
             # Set threshold for this benchmark (FlashAttn/FlashMLA only)
@@ -1039,8 +768,6 @@ def _run_mla_benchmark_batched(
                     backend_cfg,
                     mla_dims,
                     device,
-                    indexer=indexer,
-                    kv_cache_dtype=kv_cache_dtype,
                 )
                 results.append(result)
 
@@ -1066,27 +793,20 @@ def run_mla_benchmark(
     config,
     reorder_batch_threshold: int | None = None,
     num_kv_splits: int | None = None,
-    index_topk: int = 2048,
-    prefill_backend: str | None = None,
 ) -> BenchmarkResult | list[BenchmarkResult]:
     """
     Unified MLA benchmark runner for all backends.
 
-    Works for: flashattn_mla, flashmla, flashinfer_mla, cutlass_mla,
+    Works for: flashattn_mla, flashmla, flashinfer_mla, cutlass_mla
-               flashinfer_mla_sparse, flashmla_sparse
 
     Always uses batched execution internally for optimal performance.
 
     Args:
-        backend: Backend name (flashattn_mla, flashmla, flashinfer_mla, cutlass_mla,
+        backend: Backend name (flashattn_mla, flashmla, flashinfer_mla, cutlass_mla)
-                 flashinfer_mla_sparse, flashmla_sparse)
         config: BenchmarkConfig or list of (BenchmarkConfig, param) tuples
         reorder_batch_threshold: Threshold override for FlashAttn/FlashMLA
                                  (single config mode only)
         num_kv_splits: Number of KV splits for CUTLASS (single config mode only)
-        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.
 
     Returns:
         BenchmarkResult (single mode) or list of BenchmarkResult (batched mode)
@@ -1096,9 +816,9 @@ def run_mla_benchmark(
         # Already in batched format
         if len(config) > 0 and isinstance(config[0], tuple):
             # Format: [(cfg, param), ...] where param is threshold or num_splits
-            if backend in ("flashattn_mla", "flashmla", "flashmla_sparse"):
+            if backend in ("flashattn_mla", "flashmla"):
                 configs_with_params = [(cfg, param, None) for cfg, param in config]
-            else:  # cutlass_mla, flashinfer_mla, or sparse backends
+            else:  # cutlass_mla or flashinfer_mla
                 configs_with_params = [(cfg, None, param) for cfg, param in config]
         else:
             # Format: [cfg, ...] - just configs
@@ -1110,9 +830,7 @@ def run_mla_benchmark(
         return_single = True
 
     # Use unified batched execution
-    results = _run_mla_benchmark_batched(
+    results = _run_mla_benchmark_batched(backend, configs_with_params)
-        backend, configs_with_params, index_topk, prefill_backend=prefill_backend
-    )
 
     # Return single result or list based on input
     return results[0] if return_single else results