[NemotronH] Small fix reasoning parser (#36635)

Signed-off-by: Roi Koren <roik@nvidia.com>
(cherry picked from commit e661b9ee83)

.buildkite/hardware_tests/amd.yaml

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

.buildkite/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_cpu.sh

@@ -25,7 +25,9 @@ fi
 docker build --file docker/Dockerfile.cpu \
   --build-arg max_jobs=16 \
   --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
-  --build-arg VLLM_CPU_X86=true \
+  --build-arg VLLM_CPU_AVX512BF16=true \
+  --build-arg VLLM_CPU_AVX512VNNI=true \
+  --build-arg VLLM_CPU_AMXBF16=true \
   --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu \
   --target vllm-test \
   --progress plain .

.buildkite/lm-eval-harness/configs/models-large-rocm-fp8.txt

@@ -1 +0,0 @@
-Qwen3-235B-A22B-Instruct-2507-FP8.yaml

.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}"
             )
-
-            min_acceptable = ground_truth * (1 - rtol)
-            success = success and measured_value >= min_acceptable
+            success = success and np.isclose(ground_truth, measured_value, rtol=rtol)
 
     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

@@ -7,12 +7,12 @@ 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
+import regex as re
 
 pd.options.display.float_format = "{:.2f}".format
 plotly_found = util.find_spec("plotly.express") is not None
@@ -33,45 +33,6 @@ 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 +52,19 @@ 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 +75,12 @@ def compare_data_columns(
             "No common key columns found from info_cols across the input files."
         )
 
-    union_index = None
-    metas: list[pd.DataFrame] = []
-    staged: list[tuple[str, pd.Series, pd.Series | None]] = []
+    meta_added = False
 
     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
-        # 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:
+        if drop_column in df.columns:
             df = df.dropna(subset=[drop_column], ignore_index=True)
 
         for c in (
@@ -163,61 +105,35 @@ def compare_data_columns(
             meta = meta.groupby(level=key_cols, dropna=False).first()
 
         file_label = "/".join(file.split("/")[:-1]) or os.path.basename(file)
-
-        if data_column in df_idx.columns:
-            s = df_idx[data_column]
-            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 = df_idx[data_column]
+        if not s.index.is_unique:
+            s = s.groupby(level=key_cols, dropna=False).mean()
         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:
-            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)
+        frames.append(s)
         raw_data_cols.append(file_label)
-        metric_series_aligned.append(s_aligned)
+        compare_frames.append(s)
 
-        if debug and name_s is not None:
-            frames.append(name_s.reindex(union_index))
-
-        if len(metric_series_aligned) >= 2:
-            base = metric_series_aligned[0]
-            current = metric_series_aligned[-1]
-            if "P99" in str(data_column) or "Median" in str(data_column):
+        if len(compare_frames) >= 2:
+            base = compare_frames[0]
+            current = compare_frames[-1]
+            if "P99" in data_column or "Median" in 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 +204,24 @@ 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,
-    threshold: float,
-    slack_pct: float = 0.0,
+    df: pd.DataFrame, threshold: float
 ) -> pd.io.formats.style.Styler:
     conc_col = _find_concurrency_col(df)
     key_cols = [
@@ -304,24 +234,12 @@ def _highlight_threshold(
     ]
     conf_cols = [c for c in conf_cols if pd.api.types.is_numeric_dtype(df[c])]
 
-    try:
-        slack_pct = float(slack_pct or 0.0)
-    except Exception:
-        slack_pct = 0.0
-    slack_limit = threshold * (1.0 + slack_pct / 100.0)
-
-    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)
+    return df.style.map(
+        lambda v: "background-color:#e6ffe6;font-weight:bold;"
+        if pd.notna(v) and v <= threshold
+        else "",
+        subset=conf_cols,
+    )
 
 
 def highlight_ratio_columns(styler: pd.io.formats.style.Styler):
@@ -368,30 +286,11 @@ def _sanitize_sheet_name(name: str) -> str:
       - 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 = re.sub(r"[:\\/?*\[\]]", "_", name)
     name = name.strip().strip("'")
-    name = " ".join(name.split())
-
+    name = re.sub(r"\s+", " ", name)
     if not name:
         name = "sheet"
     return name[:31]
@@ -399,57 +298,30 @@ def _sanitize_sheet_name(name: str) -> str:
 
 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).
+    model = d.get("Model", "model")
+    model_short = str(model).split("/")[-1]
     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] = []
+    startrow = 0
     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)
+        pd.DataFrame([[title]]).to_excel(
+            writer, sheet_name=sheet, index=False, header=False, startrow=startrow
+        )
+        startrow += 1
+        df.to_excel(writer, sheet_name=sheet, index=False, startrow=startrow)
+        startrow += len(df) + 3
 
 
 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
+    s = re.sub(r"[^\w\-.]+", "_", str(s).strip())
+    return s[:180] if len(s) > 180 else s
 
 
 # -----------------------------
@@ -556,11 +428,7 @@ def _config_value_columns(df: pd.DataFrame, conc_col: str) -> list[str]:
 
 
 def _max_concurrency_ok(
-    df: pd.DataFrame,
-    conc_col: str,
-    cfg_col: str,
-    threshold: float,
-    slack_pct: float = 0.0,
+    df: pd.DataFrame, conc_col: str, cfg_col: str, threshold: float
 ):
     if df is None or conc_col not in df.columns or cfg_col not in df.columns:
         return pd.NA
@@ -573,14 +441,7 @@ def _max_concurrency_ok(
     if d.empty:
         return pd.NA
 
-    # Accept values up to (1 + slack_pct%) above the SLA.
-    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]
+    ok = d[d[cfg_col] <= threshold]
     if ok.empty:
         return pd.NA
 
@@ -646,25 +507,15 @@ 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(
-                ttft_group_df, conc_col, cfg, args.ttft_max_ms, args.ttft_slack_pct
-            )
+            _max_concurrency_ok(ttft_group_df, conc_col, cfg, args.ttft_max_ms)
             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
-            )
+            _max_concurrency_ok(tpot_group_df, conc_col, cfg, args.tpot_max_ms)
             if tpot_group_df is not None
             else pd.NA
         )
@@ -693,8 +544,8 @@ def build_valid_max_concurrency_summary_html(
         rows.append(
             {
                 "Configuration": cfg,
-                f"Max {conc_col} (TTFT ≤ {ttft_range})": ttft_max,
-                f"Max {conc_col} (TPOT ≤ {tpot_range})": tpot_max,
+                f"Max {conc_col} (TTFT ≤ {args.ttft_max_ms:g} ms)": ttft_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,
@@ -769,24 +620,15 @@ def build_valid_max_concurrency_summary_df(
     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
-            )
+            _max_concurrency_ok(ttft_group_df, conc_col, cfg, args.ttft_max_ms)
             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
-            )
+            _max_concurrency_ok(tpot_group_df, conc_col, cfg, args.tpot_max_ms)
             if tpot_group_df is not None
             else pd.NA
         )
@@ -815,8 +657,8 @@ def build_valid_max_concurrency_summary_df(
         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} (TTFT ≤ {args.ttft_max_ms:g} ms)": ttft_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,
@@ -909,21 +751,7 @@ def build_parser() -> argparse.ArgumentParser:
         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 ----
+    # ---- NEW: export options ----
     parser.add_argument(
         "--excel-out",
         type=str,
@@ -1015,13 +843,9 @@ def render_metric_table_html(
 
     metric_name = metric_label.lower()
     if "ttft" in metric_name:
-        styler = _highlight_threshold(
-            display_group, args.ttft_max_ms, args.ttft_slack_pct
-        )
+        styler = _highlight_threshold(display_group, args.ttft_max_ms)
     elif ("tpot" in metric_name) or ("median" in metric_name) or ("p99" in metric_name):
-        styler = _highlight_threshold(
-            display_group, args.tpot_max_ms, args.tpot_slack_pct
-        )
+        styler = _highlight_threshold(display_group, args.tpot_max_ms)
     else:
         styler = display_group.style
 
@@ -1138,46 +962,22 @@ def write_report_group_first(
         csv_dir.mkdir(parents=True, exist_ok=True)
 
     excel_path = args.excel_out or "perf_comparison.xlsx"
-    disable_excel = os.getenv("VLLM_COMPARE_DISABLE_EXCEL", "0") == "1"
-
-    # Prefer xlsxwriter for speed; fallback to openpyxl if unavailable.
-    excel_engine = (
-        os.getenv("VLLM_COMPARE_EXCEL_ENGINE", "xlsxwriter").strip() or "xlsxwriter"
-    )
-    if excel_engine == "xlsxwriter" and util.find_spec("xlsxwriter") is None:
-        excel_engine = "openpyxl"
-
-    excel_engine_kwargs = {}
-    if excel_engine == "xlsxwriter":
-        # Reduce memory pressure & usually faster writes.
-        excel_engine_kwargs = {"options": {"constant_memory": True}}
-
-    xw_ctx = (
-        nullcontext(None)
-        if disable_excel
-        else pd.ExcelWriter(
-            excel_path, engine=excel_engine, engine_kwargs=excel_engine_kwargs
-        )
-    )
-    with xw_ctx as xw:
-        used_sheets: set[str] = set()
+    with pd.ExcelWriter(excel_path, engine="openpyxl") as xw:
         # ---- Environment sheet (first) ----
         env_sheet = _sanitize_sheet_name("Environment")
         env_df = _load_env_df_for_inputs(args, files)
-        if xw is not None:
-            if env_df is None or env_df.empty:
-                pd.DataFrame(
-                    [
-                        {
-                            "Section": "Environment",
-                            "Key": "vllm_env.txt",
-                            "Value": "NOT FOUND (or empty)",
-                        }
-                    ]
-                ).to_excel(xw, sheet_name=env_sheet, index=False)
-            else:
-                env_df.to_excel(xw, sheet_name=env_sheet, index=False)
-            used_sheets.add(env_sheet)
+        if env_df is None or env_df.empty:
+            pd.DataFrame(
+                [
+                    {
+                        "Section": "Environment",
+                        "Key": "vllm_env.txt",
+                        "Value": "NOT FOUND (or empty)",
+                    }
+                ]
+            ).to_excel(xw, sheet_name=env_sheet, index=False)
+        else:
+            env_df.to_excel(xw, sheet_name=env_sheet, index=False)
         with open("perf_comparison.html", "w", encoding="utf-8") as main_fh:
             main_fh.write('<meta charset="utf-8">\n')
             for gkey in group_keys:
@@ -1193,19 +993,12 @@ def write_report_group_first(
 
                 main_fh.write(group_header)
 
-                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)
+                dedup_i = 1
+                while sheet in xw.sheets:
+                    dedup_i += 1
+                    sheet = _sanitize_sheet_name(f"{sheet_base}_{dedup_i}")
 
                 excel_blocks: list[tuple[str, pd.DataFrame]] = []
 
@@ -1266,7 +1059,7 @@ def write_report_group_first(
                         )
 
                         excel_blocks.append(
-                            (metric_label, group_df.reset_index(drop=True))
+                            (metric_label, display_group.reset_index(drop=True))
                         )
                         if csv_dir:
                             fn = _safe_filename(
@@ -1274,7 +1067,7 @@ def write_report_group_first(
                                     "/", "_"
                                 )
                             )
-                            group_df.to_csv(csv_dir / f"{fn}.csv", index=False)
+                            display_group.to_csv(csv_dir / f"{fn}.csv", index=False)
 
                     summary_html = build_valid_max_concurrency_summary_html(
                         tput_group_df=tput_group_df,
@@ -1304,13 +1097,9 @@ def write_report_group_first(
                             )
                             summary_df.to_csv(csv_dir / f"{fn}.csv", index=False)
 
-                if do_excel:
-                    _write_tables_to_excel_sheet(xw, sheet, excel_blocks)
+                _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}")
+    print(f"Wrote Excel: {excel_path}")
     if csv_dir:
         print(f"Wrote CSVs under: {csv_dir}")
 

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

@@ -12,13 +12,6 @@ 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.
@@ -190,304 +183,6 @@ 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
   # $1: test type (latency or throughput)
@@ -652,48 +347,10 @@ run_serving_tests() {
     server_envs=$(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_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 "$server_params_no_model")
-
+    server_args=$(json2args "$server_params")
     server_envs=$(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 +382,14 @@ 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
@@ -778,14 +436,6 @@ run_serving_tests() {
       for max_concurrency in $max_concurrency_list; do
         new_test_name="${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 \
@@ -794,9 +444,8 @@ 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: $client_command"
@@ -818,11 +467,6 @@ run_serving_tests() {
         echo "$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
@@ -888,7 +532,6 @@ 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/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-text.json

@@ -149,39 +149,6 @@
         "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": {
@@ -221,45 +188,6 @@
         "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": {

.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 +139,14 @@
       }
     },
     {
-      "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
-      }
-    },
-    {
-      "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
+        "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,
@@ -88,6 +91,7 @@
         "server_parameters": {
             "model": "deepseek-ai/DeepSeek-R1",
             "tensor_parallel_size": 8,
+            "swap_space": 16,
             "disable_log_stats": "",
             "load_format": "dummy",
             "max-model-len": 2048,

.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/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/check-ray-compatibility.sh

@@ -16,23 +16,6 @@ RAY_BASE_URL="https://raw.githubusercontent.com/ray-project/ray/master/python"
 WORK_DIR=$(mktemp -d)
 trap 'rm -rf "$WORK_DIR"' EXIT
 
-# ── Detect PyTorch index URL ─────────────────────────────────────────────
-
-if python3 -c "import torch; assert torch.version.hip" 2>/dev/null; then
-    ROCM_VER=$(python3 -c "import torch; print(torch.version.hip.rsplit('.', 1)[0])")
-    CANDIDATE_URL="https://download.pytorch.org/whl/rocm${ROCM_VER}"
-    if curl -fsSL --head "${CANDIDATE_URL}/" >/dev/null 2>&1; then
-        TORCH_INDEX_URL="${CANDIDATE_URL}"
-    else
-        echo ">>> WARNING: ROCm ${ROCM_VER} wheel index not found at ${CANDIDATE_URL}"
-        echo ">>>          Falling back to default PyPI (resolution may be incomplete)"
-        TORCH_INDEX_URL=""
-    fi
-else
-    TORCH_INDEX_URL="https://download.pytorch.org/whl/cu129"
-fi
-echo ">>> Using PyTorch index: ${TORCH_INDEX_URL:-PyPI default}"
-
 # Fetch all Ray requirement files used in the LLM depset pipeline
 echo ">>> Fetching Ray requirement files"
 RAY_FILES=(
@@ -133,11 +116,6 @@ echo "============================================================"
 echo ">>> Resolving: Can Ray generate compatible lock files?"
 echo "============================================================"
 
-EXTRA_INDEX_ARGS=()
-if [[ -n "${TORCH_INDEX_URL}" ]]; then
-    EXTRA_INDEX_ARGS+=(--extra-index-url "${TORCH_INDEX_URL}")
-fi
-
 set +e
 uv pip compile \
     "${WORK_DIR}/requirements.txt" \
@@ -148,7 +126,7 @@ uv pip compile \
     -c "${WORK_DIR}/vllm-constraints.txt" \
     --python-version 3.12 \
     --python-platform x86_64-manylinux_2_31 \
-    "${EXTRA_INDEX_ARGS[@]}" \
+    --extra-index-url https://download.pytorch.org/whl/cu129 \
     --index-strategy unsafe-best-match \
     --unsafe-package setuptools \
     --unsafe-package ray \
@@ -188,19 +166,12 @@ See [issue #33599](https://github.com/vllm-project/vllm/issues/33599) for contex
 EOF
 fi
 
-# Notify Slack if webhook is configured and PR/branch are valid.
+# Notify Slack if webhook is configured.
 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=$(python3 -c '
+    echo ">>> Sending Slack notification"
+    # Single quotes are intentional: the f-string expressions are Python, not shell.
+    # shellcheck disable=SC2016
+    PAYLOAD=$(python3 -c '
 import json, os, sys
 pr = os.getenv("BUILDKITE_PULL_REQUEST", "N/A")
 branch = os.getenv("BUILDKITE_BRANCH", "unknown")
@@ -223,11 +194,10 @@ data = {
 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: $HTTP_CODE"
-    fi
+    HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" -X POST "$RAY_COMPAT_SLACK_WEBHOOK_URL" \
+        -H 'Content-type: application/json' \
+        -d "$PAYLOAD")
+    echo "    Slack webhook response: $HTTP_CODE"
 else
     echo ">>> Skipping Slack notification (RAY_COMPAT_SLACK_WEBHOOK_URL not set)"
 fi

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

@@ -205,13 +205,6 @@ re_quote_pytest_markers() {
       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}' "
@@ -249,11 +242,6 @@ re_quote_pytest_markers() {
 
   # 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
@@ -333,18 +321,15 @@ apply_rocm_test_overrides() {
   # --- 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_audio.py \
+    --ignore=entrypoints/openai/test_shutdown.py \
     --ignore=entrypoints/openai/test_completion.py \
-    --ignore=entrypoints/openai/models/test_models.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/completion/test_prompt_validation.py "}
-  fi
-
-  if [[ $cmds == *" entrypoints/serve"* ]]; then
-    cmds="${cmds} \
-    --ignore=entrypoints/serve/lora/test_lora_adapters.py"
+    --ignore=entrypoints/openai/test_root_path.py \
+    --ignore=entrypoints/openai/test_tokenization.py \
+    --ignore=entrypoints/openai/test_prompt_validation.py "}
   fi
 
   if [[ $cmds == *" entrypoints/llm "* ]]; then
@@ -507,8 +492,6 @@ else
     -e HF_TOKEN \
     -e AWS_ACCESS_KEY_ID \
     -e AWS_SECRET_ACCESS_KEY \
-    -e BUILDKITE_PARALLEL_JOB \
-    -e BUILDKITE_PARALLEL_JOB_COUNT \
     -v "${HF_CACHE}:${HF_MOUNT}" \
     -e "HF_HOME=${HF_MOUNT}" \
     -e "PYTHONPATH=${MYPYTHONPATH}" \

.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-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 "

.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

@@ -76,7 +76,7 @@ docker run --rm --runtime=habana --name="${container_name}" --network=host \
   -e PT_HPU_LAZY_MODE=1 \
   "${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-tpu-v1-test-part2.sh

@@ -127,7 +127,7 @@ run_and_track_test() {
 
 # --- Actual Test Execution ---
 run_and_track_test 1 "test_struct_output_generate.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
 run_and_track_test 2 "test_moe_pallas.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py"
 run_and_track_test 3 "test_lora.py" \

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

@@ -33,22 +33,23 @@ docker run \
     bash -c '
     set -e
     echo $ZE_AFFINITY_MASK
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 -O3 -cc.cudagraph_mode=NONE
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend 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/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --attention-backend=TRITON_ATTN
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --quantization fp8
-    python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4  --block-size 64 --enforce-eager --max-model-len 8192
-    python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2
-    python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
+    pip install tblib==3.1.0
+    python3 examples/offline_inference/basic/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 -O3 -cc.cudagraph_mode=NONE
+    python3 examples/offline_inference/basic/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 mp
+    python3 examples/offline_inference/basic/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 --quantization fp8
+    python3 examples/offline_inference/basic/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4  --block-size 64 --enforce-eager
+    python3 examples/offline_inference/basic/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2
+    python3 examples/offline_inference/basic/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
     cd tests
     pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py
     pytest -v -s v1/engine
     pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
-    pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.py
+    pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
     pytest -v -s v1/structured_output
     pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py
-    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py -k "not (test_register_kv_caches and FLASH_ATTN and True)"
+    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py
     pytest -v -s v1/test_serial_utils.py
 '

.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_prefetch_offload.sh

@@ -1,14 +1,11 @@
 #!/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]
-#
-# Environment variables:
-#   ATTENTION_BACKEND   - attention backend to use (e.g., FLASH_ATTN,
-#                         ROCM_ATTN, FLASHINFER). If unset, uses vllm default.
 THRESHOLD=${1:-0.25}
 NUM_Q=${2:-1319}
 PORT=${3:-8030}
@@ -25,14 +22,6 @@ wait_for_server() {
 
 MODEL="deepseek-ai/DeepSeek-V2-Lite"
 
-# ── Build optional vllm serve flags ─────────────────────────────────────
-
-EXTRA_ARGS=()
-if [[ -n "${ATTENTION_BACKEND:-}" ]]; then
-  echo "Using attention backend: ${ATTENTION_BACKEND}"
-  EXTRA_ARGS+=(--attention-backend "${ATTENTION_BACKEND}")
-fi
-
 cleanup() {
   if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
     kill "${SERVER_PID}" 2>/dev/null || true
@@ -51,8 +40,7 @@ vllm serve "$MODEL" \
   --offload-num-in-group 2 \
   --offload-prefetch-step 1 \
   --offload-params w13_weight w2_weight \
-  --port "$PORT" \
-  ${EXTRA_ARGS+"${EXTRA_ARGS[@]}"} &
+  --port "$PORT" &
 SERVER_PID=$!
 wait_for_server "$PORT"
 

.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)

.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/upload-nightly-wheels.sh

@@ -72,7 +72,7 @@ obj_json="objects.json"
 aws s3api list-objects-v2 --bucket "$BUCKET" --prefix "$SUBPATH/" --delimiter / --output json > "$obj_json"
 mkdir -p "$INDICES_OUTPUT_DIR"
 
-# call script to generate 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>/

.buildkite/test_areas/basic_correctness.yaml

@@ -14,3 +14,8 @@ steps:
   - pytest -v -s basic_correctness/test_cumem.py
   - pytest -v -s basic_correctness/test_basic_correctness.py
   - pytest -v -s basic_correctness/test_cpu_offload.py
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd

.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/"
@@ -59,7 +49,7 @@ steps:
   - export VLLM_TEST_CLEAN_GPU_MEMORY=1
   - pytest -s -v tests/compile/passes/distributed
 
-- label: Fusion and Compile Unit Tests (2xB200)
+- label: Fusion and Compile Unit Tests (B200)
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/"
   device: b200
@@ -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
-    - 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)"
+    # 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"
 
 - label: Fusion E2E Config Sweep (H100)
   timeout_in_minutes: 30
@@ -132,9 +122,9 @@ steps:
   commands:
     - nvidia-smi
     # Run all models but only FLASHINFER, Inductor partition and native custom ops
-    # 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
     # Run just llama3 (fp8 & fp4) for all config combinations (only inductor partition)
-    - 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)"
+    - 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 llama-3)"
 
 - label: Fusion E2E TP2 Quick (H100)
   timeout_in_minutes: 20
@@ -150,8 +140,8 @@ 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_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_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"
 
 - label: Fusion E2E TP2 AR-RMS Config Sweep (H100)
   timeout_in_minutes: 40
@@ -205,7 +195,7 @@ steps:
   commands:
     - nvidia-smi
     # Run all models but only FLASHINFER, Inductor partition and native custom ops
-    # include qwen/deepseek with +quant_fp8 as -quant_fp8 rms+quant fusion is not supported
+    # include qwen 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_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_ar_rms.py -k "(FLASHINFER and inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and qwen3)) or 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)"

.buildkite/test_areas/distributed.yaml

@@ -15,29 +15,8 @@ steps:
   - pytest -v -s distributed/test_shm_buffer.py
   - pytest -v -s distributed/test_shm_storage.py
 
-- label: Distributed DP Tests (2 GPUs)
-  timeout_in_minutes: 20
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-  - vllm/distributed/
-  - vllm/engine/
-  - vllm/executor/
-  - vllm/worker/worker_base.py
-  - vllm/v1/engine/
-  - vllm/v1/worker/
-  - tests/v1/distributed
-  - tests/entrypoints/openai/test_multi_api_servers.py
-  commands:
-  # 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
-  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
-  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
-  - DP_SIZE=2 pytest -v -s entrypoints/openai/test_multi_api_servers.py
-
-- label: Distributed Compile + RPC Tests (2 GPUs)
-  timeout_in_minutes: 20
+- label: Distributed (2 GPUs)
+  timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   source_file_dependencies:
@@ -50,80 +29,61 @@ steps:
   - vllm/v1/worker/
   - tests/compile/fullgraph/test_basic_correctness.py
   - tests/compile/test_wrapper.py
-  - tests/entrypoints/llm/test_collective_rpc.py
-  commands:
-  # https://github.com/NVIDIA/nccl/issues/1838
-  - export NCCL_CUMEM_HOST_ENABLE=0
-  - pytest -v -s entrypoints/llm/test_collective_rpc.py
-  - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
-  - pytest -v -s ./compile/test_wrapper.py
-
-- label: Distributed Torchrun + Shutdown Tests (2 GPUs)
-  timeout_in_minutes: 20
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-  - vllm/distributed/
-  - vllm/engine/
-  - vllm/executor/
-  - vllm/worker/worker_base.py
-  - vllm/v1/engine/
-  - vllm/v1/worker/
   - tests/distributed/
+  - tests/entrypoints/llm/test_collective_rpc.py
+  - tests/v1/distributed
+  - tests/v1/entrypoints/openai/test_multi_api_servers.py
   - tests/v1/shutdown
   - tests/v1/worker/test_worker_memory_snapshot.py
   commands:
   # 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
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
+  - DP_SIZE=2 pytest -v -s v1/entrypoints/openai/test_multi_api_servers.py
+  - pytest -v -s entrypoints/llm/test_collective_rpc.py
+  - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
+  - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
   - VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
   - 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)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace"
-  num_devices: 4
-  source_file_dependencies:
-  - vllm/distributed/
-  - tests/distributed/test_torchrun_example.py
-  - tests/distributed/test_torchrun_example_moe.py
-  - examples/offline_inference/rlhf_colocate.py
-  - examples/rl/
-  - tests/examples/offline_inference/data_parallel.py
-  commands:
-  # https://github.com/NVIDIA/nccl/issues/1838
-  - export NCCL_CUMEM_HOST_ENABLE=0
-  # test with torchrun tp=2 and external_dp=2
-  - torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example.py
-  # test with torchrun tp=2 and pp=2
-  - PP_SIZE=2 torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example.py
-  # test with torchrun tp=4 and dp=1
-  - TP_SIZE=4 torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example_moe.py
-  # test with torchrun tp=2, pp=2 and dp=1
-  - PP_SIZE=2 TP_SIZE=2 torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example_moe.py
-  # test with torchrun tp=1 and dp=4 with ep
-  - DP_SIZE=4 ENABLE_EP=1 torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example_moe.py
-  # test with torchrun tp=2 and dp=2 with ep
-  - TP_SIZE=2 DP_SIZE=2 ENABLE_EP=1 torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example_moe.py
-  # test with internal dp
-  - python3 examples/offline_inference/data_parallel.py --enforce-eager
-  # rlhf examples
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/rl/rlhf_nccl.py
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/rl/rlhf_ipc.py
-
-- label: Distributed DP Tests (4 GPUs)
-  timeout_in_minutes: 30
+- label: Distributed Tests (4 GPUs)
+  timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:
   - vllm/distributed/
+  - tests/distributed/test_utils
+  - 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_utils
+  - tests/distributed/test_symm_mem_allreduce.py
   commands:
   # https://github.com/NVIDIA/nccl/issues/1838
   - export NCCL_CUMEM_HOST_ENABLE=0
+  # test with torchrun tp=2 and external_dp=2
+  - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  # test with torchrun tp=2 and pp=2
+  - PP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  # test with torchrun tp=4 and dp=1
+  - TP_SIZE=4 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
+  # test with torchrun tp=2, pp=2 and dp=1
+  - PP_SIZE=2 TP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
+  # test with torchrun tp=1 and dp=4 with ep
+  - DP_SIZE=4 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
+  # test with torchrun tp=2 and dp=2 with ep
+  - 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
   - 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
@@ -131,27 +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)
-  - pytest -v -s distributed/test_multiproc_executor.py::test_multiproc_executor_multi_node
+  # TODO: create a dedicated test section for multi-GPU example tests
+  # when we have multiple distributed example tests
+  # OLD rlhf examples
+  - 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 Tests (8 GPUs)(H100)
   timeout_in_minutes: 10
@@ -193,7 +146,7 @@ steps:
   num_devices: 2
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
-    - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/rl/rlhf_async_new_apis.py
+    # - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py --- failing, need to re-enable
     - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
     - pytest -v -s tests/v1/distributed/test_dbo.py
 
@@ -257,19 +210,6 @@ steps:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
     - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
-- label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
-  timeout_in_minutes: 30
-  device: a100
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 2
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - vllm/v1/worker/kv_connector_model_runner_mixin.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
-
 - label: Pipeline + Context Parallelism (4 GPUs)
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"

.buildkite/test_areas/engine.yaml

@@ -1,5 +1,5 @@
 group: Engine
-depends_on:
+depends_on: 
   - image-build
 steps:
 - label: Engine
@@ -14,30 +14,28 @@ steps:
   commands:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
 
-- label: Engine (1 GPU)
-  timeout_in_minutes: 30
+- label: V1 e2e + engine (1 GPU)
+  timeout_in_minutes: 45
   source_file_dependencies:
-    - vllm/v1/engine/
-    - tests/v1/engine/
+    - vllm/
+    - 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
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+      commands:
+      - pytest -v -s v1/e2e
+      - pytest -v -s v1/engine
 
 - label: V1 e2e (2 GPUs)
   timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
@@ -48,7 +46,7 @@ steps:
     - 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"
+    - pytest -v -s v1/e2e/test_spec_decode.py -k "tensor_parallelism"
   mirror:
     amd:
       device: mi325_2
@@ -64,21 +62,9 @@ steps:
     - 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"
+    - pytest -v -s v1/e2e/test_spec_decode.py -k "eagle_correctness_heavy"
   mirror:
     amd:
       device: mi325_4
       depends_on:
       - image-build-amd
-
-- label: V1 e2e (4xH100)
-  timeout_in_minutes: 60
-  device: h100
-  num_devices: 4
-  optional: true
-  source_file_dependencies:
-    - vllm/v1/attention/backends/utils.py
-    - vllm/v1/worker/gpu_model_runner.py
-    - tests/v1/e2e/test_hybrid_chunked_prefill.py
-  commands:
-    - pytest -v -s v1/e2e/test_hybrid_chunked_prefill.py

.buildkite/test_areas/entrypoints.yaml

@@ -10,7 +10,7 @@ steps:
   - tests/entrypoints/
   commands:
   - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/serve/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
+  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
 
 - label: Entrypoints Integration (LLM)
   timeout_in_minutes: 40
@@ -24,6 +24,11 @@ steps:
   - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Entrypoints Integration (API Server 1)
   timeout_in_minutes: 130
@@ -34,13 +39,8 @@ 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 --ignore=entrypoints/openai/test_multi_api_servers.py
+  - 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
@@ -48,11 +48,11 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/rpc
-  - tests/entrypoints/serve/instrumentator
+  - tests/entrypoints/instrumentator
   - tests/tool_use
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/serve/instrumentator
+  - pytest -v -s entrypoints/instrumentator
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
   - pytest -v -s tool_use
 
@@ -65,6 +65,11 @@ steps:
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/pooling
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Entrypoints Integration (Responses API)
   timeout_in_minutes: 50
@@ -75,6 +80,14 @@ steps:
   commands:
   - pytest -v -s entrypoints/openai/responses
 
+- label: Entrypoints V1
+  timeout_in_minutes: 50
+  source_file_dependencies:
+    - vllm/
+    - tests/v1
+  commands:
+    - pytest -v -s v1/entrypoints
+
 - label: OpenAI API Correctness
   timeout_in_minutes: 30
   source_file_dependencies:

.buildkite/test_areas/expert_parallelism.yaml

@@ -24,7 +24,8 @@ steps:
 
 - label: Elastic EP Scaling Test
   timeout_in_minutes: 20
-  device: h100
+  device: b200
+  optional: true
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:

.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
@@ -35,7 +34,7 @@ steps:
   parallelism: 2
 
 - label: Kernels MoE Test %N
-  timeout_in_minutes: 25
+  timeout_in_minutes: 60
   source_file_dependencies:
   - csrc/quantization/cutlass_w8a8/moe/
   - csrc/moe/
@@ -47,7 +46,7 @@ steps:
   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/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
-  parallelism: 5
+  parallelism: 2
 
 - label: Kernels Mamba Test
   timeout_in_minutes: 45
@@ -97,7 +96,7 @@ 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

.buildkite/test_areas/lm_eval.yaml

@@ -45,22 +45,6 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt
 
-- label: LM Eval Qwen3.5 Models (B200)
-  timeout_in_minutes: 120
-  device: b200
-  optional: true
-  num_devices: 2
-  source_file_dependencies:
-  - vllm/model_executor/models/qwen3_5.py
-  - vllm/model_executor/models/qwen3_5_mtp.py
-  - vllm/transformers_utils/configs/qwen3_5.py
-  - vllm/transformers_utils/configs/qwen3_5_moe.py
-  - vllm/model_executor/models/qwen3_next.py
-  - vllm/model_executor/models/qwen3_next_mtp.py
-  - vllm/model_executor/layers/fla/ops/
-  commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-qwen35-blackwell.txt
-
 - label: LM Eval Large Models (H200)
   timeout_in_minutes: 60
   device: h200

.buildkite/test_areas/lora.yaml

@@ -8,7 +8,7 @@ steps:
   - vllm/lora
   - tests/lora
   commands:
-    - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_llm_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemoel_lora.py 
+    - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_llm_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py
   parallelism: 4
 
 
@@ -30,5 +30,4 @@ steps:
     - pytest -v -s -x lora/test_llama_tp.py
     - pytest -v -s -x lora/test_llm_with_multi_loras.py
     - pytest -v -s -x lora/test_olmoe_tp.py
-    - pytest -v -s -x lora/test_gptoss_tp.py
-    - pytest -v -s -x lora/test_qwen35_densemoel_lora.py
+    - pytest -v -s -x lora/test_gptoss_tp.py

.buildkite/test_areas/misc.yaml

@@ -67,13 +67,12 @@ steps:
   - examples/
   commands:
     - pip install tensorizer # for tensorizer test
-     # for basic
-    - python3 basic/offline_inference/chat.py
-    - 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
-    - python3 basic/offline_inference/classify.py
-    - python3 basic/offline_inference/embed.py
-    - python3 basic/offline_inference/score.py
+    - python3 offline_inference/basic/chat.py # for basic
+    - python3 offline_inference/basic/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 offline_inference/basic/classify.py
+    - python3 offline_inference/basic/embed.py
+    - python3 offline_inference/basic/score.py
     # for multi-modal models
     - python3 offline_inference/audio_language.py --seed 0
     - python3 offline_inference/vision_language.py --seed 0

.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/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 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

@@ -65,7 +65,7 @@ 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_multimodal.yaml

@@ -2,67 +2,18 @@ group: Models - Multimodal
 depends_on: 
   - image-build
 steps:
-- label: "Multi-Modal Models (Standard) 1: qwen2"
-  timeout_in_minutes: 45
+- label: Multi-Modal Models (Standard) # 60min
+  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"
-    - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
-  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
+    - pip freeze | grep -E 'torch'
+    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
     - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
 
-- label: Multi-Modal Processor (CPU)
+- label: Multi-Modal Processor Test (CPU)
   depends_on: 
   - image-build-cpu
   timeout_in_minutes: 60
@@ -95,44 +46,29 @@ steps:
   commands:
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
 
-- label: Multi-Modal Models (Extended Generation 1)
+- label: Multi-Modal Models (Extended) 1
   optional: true
   source_file_dependencies:
   - vllm/
-  - tests/models/multimodal/generation
-  - tests/models/multimodal/test_mapping.py
+  - tests/models/multimodal
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/multimodal/generation -m 'not core_model' --ignore models/multimodal/generation/test_common.py
-    - pytest -v -s models/multimodal/test_mapping.py
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
+    - pytest -v -s models/multimodal -m 'not core_model' --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/processing
 
-- label: Multi-Modal Models (Extended Generation 2)
+- label: Multi-Modal Models (Extended) 2
   optional: true
   source_file_dependencies:
   - vllm/
-  - tests/models/multimodal/generation
+  - 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 'split(group=0) and not core_model'
 
-- label: Multi-Modal Models (Extended Generation 3)
+- label: Multi-Modal Models (Extended) 3
   optional: true
   source_file_dependencies:
   - vllm/
-  - tests/models/multimodal/generation
+  - 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 'split(group=1) and not core_model'
-
-- label: Multi-Modal Models (Extended Pooling)
-  optional: true
-  source_file_dependencies:
-  - vllm/
-  - tests/models/multimodal/pooling
-  commands:
-    - pytest -v -s models/multimodal/pooling -m 'not core_model'

.buildkite/test_areas/plugins.yaml

@@ -15,12 +15,9 @@ 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
-  # test generic io_processor plugins functions
-  - pytest -v -s ./plugins_tests/test_io_processor_plugins.py
-  # test Terratorch io_processor plugins
+  # begin io_processor plugins test, all the code in between uses the prithvi_io_processor plugin
   - 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
@@ -36,6 +33,6 @@ 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/pytorch.yaml

@@ -35,7 +35,7 @@ steps:
   # as it is a heavy test that is covered in other steps.
   # Use `find` to launch multiple instances of pytest so that
   # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
-  - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
+  - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -exec pytest -s -v {} \\;"
 
 - label: PyTorch Fullgraph
   timeout_in_minutes: 30

.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"

.github/CODEOWNERS

@@ -75,7 +75,7 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/multimodal @DarkLight1337 @ywang96 @NickLucche
 /tests/quantization @mgoin @robertgshaw2-redhat @yewentao256 @pavanimajety
 /tests/test_inputs.py @DarkLight1337 @ywang96
-/tests/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
+/tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm
 /tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC @orozery
 /tests/weight_loading @mgoin @youkaichao @yewentao256
@@ -171,7 +171,6 @@ mkdocs.yaml @hmellor
 
 # Pooling models
 /examples/pooling @noooop
-/docs/models/pooling_models @noooop
 /tests/models/*/pooling* @noooop
 /tests/entrypoints/pooling @noooop
 /vllm/config/pooler.py @noooop

.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>
 
@@ -260,7 +261,7 @@ pull_request_rules:
       - files=examples/offline_inference/structured_outputs.py
       - files=examples/online_serving/structured_outputs/structured_outputs.py
       - files~=^tests/v1/structured_output/
-      - files=tests/entrypoints/llm/test_struct_output_generate.py
+      - files=tests/v1/entrypoints/llm/test_struct_output_generate.py
       - files~=^vllm/v1/structured_output/
   actions:
     label:
@@ -333,10 +334,9 @@ pull_request_rules:
     - label != stale
     - or:
       - files~=^tests/tool_use/
-      - files~=^tests/tool_parsers/
-      - files~=^tests/entrypoints/openai/.*tool.*
-      - files~=^tests/entrypoints/anthropic/.*tool.*
-      - files~=^vllm/tool_parsers/
+      - files~=^tests/entrypoints/openai/tool_parsers/
+      - 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_*
       - files=examples/offline_inference/chat_with_tools.py
@@ -382,7 +382,7 @@ 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/scripts/cleanup_pr_body.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+
+set -eu
+
+# ensure 1 argument is passed
+if [ "$#" -ne 1 ]; then
+    echo "Usage: $0 <pr_number>"
+    exit 1
+fi
+
+PR_NUMBER=$1
+OLD=/tmp/orig_pr_body.txt
+NEW=/tmp/new_pr_body.txt
+
+gh pr view --json body --template "{{.body}}" "${PR_NUMBER}" > "${OLD}"
+cp "${OLD}" "${NEW}"
+
+# Remove markdown comments (like the <!-- markdownlint-disable --> at the start)
+sed -i '/<!--.*-->$/d' "${NEW}"
+
+# Remove "PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED."
+sed -i '/PLEASE FILL IN THE PR DESCRIPTION HERE.*$/d' "${NEW}"
+
+# Remove all lines after and including "**BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE**"
+sed -i '/\*\*BEFORE SUBMITTING, PLEASE READ.*\*\*/,$d' "${NEW}"
+
+# Remove HTML <details> section that includes <summary> text of "PR Checklist (Click to Expand)"
+python3 - <<EOF
+import regex as re
+
+with open("${NEW}", "r") as file:
+    content = file.read()
+
+pattern = re.compile(r'(---\n\n)?<details>.*?<summary>.*?PR Checklist \(Click to Expand\).*?</summary>.*?</details>', re.DOTALL)
+content = re.sub(pattern, '', content)
+
+with open("${NEW}", "w") as file:
+    file.write(content)
+EOF
+
+# Run this only if ${NEW} is different than ${OLD}
+if ! cmp -s "${OLD}" "${NEW}"; then
+    gh pr edit --body-file "${NEW}" "${PR_NUMBER}"
+    echo
+    echo "Updated PR body:"
+    echo
+    cat "${NEW}"
+else
+    echo "No changes needed"
+fi

.github/workflows/cleanup_pr_body.yml

@@ -0,0 +1,32 @@
+name: Cleanup PR Body
+
+on:
+  pull_request_target:
+    types: [opened, reopened, edited]
+
+permissions:
+  pull-requests: write
+
+jobs:
+  update-description:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
+
+      - name: Set up Python
+        uses: actions/setup-python@83679a892e2d95755f2dac6acb0bfd1e9ac5d548 # v6.1.0
+        with:
+          python-version: '3.12'
+          cache: 'pip'
+
+      - name: Install Python dependencies
+        run: |
+          python3 -m pip install --upgrade pip
+          python3 -m pip install regex
+
+      - name: Update PR description
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: bash .github/scripts/cleanup_pr_body.sh "${{ github.event.number }}"

.github/workflows/issue_autolabel.yml

@@ -383,107 +383,4 @@ jobs:
                   core.notice(`All users for label "${label}" already mentioned, skipping comment`);
                 }
               }
-            }
-
-      - name: Request missing ROCm info from issue author
-        if: contains(steps.label-step.outputs.labels_added, 'rocm') && contains(toJSON(github.event.issue.labels.*.name), 'bug')
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd  # v8.0.0
-        with:
-          script: |
-            const body = (context.payload.issue.body || '').toLowerCase();
-
-            // Check for existing bot comments to avoid duplicate requests
-            const comments = await github.rest.issues.listComments({
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              issue_number: context.issue.number,
-            });
-            const botAlreadyAsked = comments.data.some(
-              c => c.user.type === 'Bot' && c.body.includes('<!-- rocm-info-request -->')
-            );
-            if (botAlreadyAsked) {
-              core.notice('ROCm info request already posted, skipping');
-              return;
-            }
-
-            // Define required information and detection patterns
-            const requiredInfo = [
-              {
-                name: 'Reproducer',
-                patterns: [
-                  /reproduc/i, /minimal.?example/i, /repro\b/i, /steps to reproduce/i,
-                  /code.?snippet/i, /sample.?code/i,
-                  /```python[\s\S]*?```/, /```bash[\s\S]*?```/, /```sh[\s\S]*?```/,
-                ],
-                ask: 'A minimal reproducer (code snippet or script that triggers the issue)',
-              },
-              {
-                name: 'Error message',
-                patterns: [
-                  /error/i, /traceback/i, /exception/i, /fault/i, /crash/i,
-                  /failed/i, /abort/i, /panic/i,
-                ],
-                ask: 'The full error message or traceback',
-              },
-              {
-                name: 'Installation method',
-                patterns: [
-                  /docker/i, /rocm\/pytorch/i, /dockerfile/i, /from source/i,
-                  /pip install/i, /build.?from/i, /container/i, /image/i,
-                  /wheel/i, /\.whl/i, /nightly/i,
-                ],
-                ask: 'How you installed vLLM (Docker image name, pip install, or build from source steps)',
-              },
-              {
-                name: 'Command',
-                patterns: [
-                  /vllm serve/i, /python\s+\S+\.py/i, /```bash[\s\S]*?```/,
-                  /```sh[\s\S]*?```/, /command/i, /launch/i, /run\s/i,
-                  /--model/i, /--tensor-parallel/i, /--gpu-memory/i,
-                ],
-                ask: 'The command you used to launch vLLM (e.g., `vllm serve ...` or the Python script)',
-              },
-              {
-                name: 'GFX architecture',
-                patterns: [
-                  /gfx\d{3,4}/i, /mi\d{3}/i, /mi\d{2}\b/i, /radeon/i,
-                  /gpu.?arch/i, /rocm-smi/i, /rocminfo/i, /navi/i,
-                  /instinct/i,
-                ],
-                ask: 'Your GPU model and GFX architecture (e.g., MI300X / gfx942) — run `rocminfo | grep gfx`',
-              },
-            ];
-
-            const issueBody = context.payload.issue.body || '';
-            const missing = requiredInfo.filter(info =>
-              !info.patterns.some(p => p.test(issueBody))
-            );
-
-            if (missing.length === 0) {
-              core.notice('All required ROCm info appears to be present');
-              return;
-            }
-
-            const author = context.payload.issue.user.login;
-            const checklist = requiredInfo.map(info => {
-              const found = !missing.includes(info);
-              return `- [${found ? 'x' : ' '}] ${info.ask}`;
-            }).join('\n');
-            const message = [
-              '<!-- rocm-info-request -->',
-              `Hi @${author}, thanks for reporting this ROCm issue!`,
-              '',
-              'To help us investigate, please make sure the following information is included:',
-              '',
-              checklist,
-              '',
-              'Please provide any unchecked items above. This will help us reproduce and resolve the issue faster. Thank you!',
-            ].join('\n');
-
-            await github.rest.issues.createComment({
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              issue_number: context.issue.number,
-              body: message,
-            });
-            core.notice(`Requested missing ROCm info from @${author}: ${missing.map(m => m.name).join(', ')}`);
+            }

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

@@ -1,14 +1,11 @@
 name: macOS Apple Silicon Smoke Test
 
 on:
-  schedule:
-    # Daily at 2:30 AM UTC
-    - cron: '30 2 * * *'
+  push:
+    branches:
+      - main
   workflow_dispatch:  # Manual trigger
 
-permissions:
-  contents: read
-
 jobs:
   macos-m1-smoke-test:
     runs-on: macos-latest

.github/workflows/new_pr_bot.yml

@@ -1,96 +0,0 @@
-name: New PR Bot
-
-on:
-  pull_request_target:
-    types: [opened]
-
-permissions:
-  pull-requests: write
-
-jobs:
-  update-description:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Update PR description
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
-        with:
-          script: |
-            const { owner, repo } = context.repo;
-            const pr_number = context.issue.number;
-
-            const { data: pr } = await github.rest.pulls.get({
-              owner,
-              repo,
-              pull_number: pr_number,
-            });
-
-            let body = pr.body || '';
-            const original = body;
-
-            // Remove markdown comments (<!-- ... -->)
-            body = body.replace(/^<!--.*-->$/gm, '');
-
-            // Remove "PLEASE FILL IN THE PR DESCRIPTION HERE ..."
-            body = body.replace(/^PLEASE FILL IN THE PR DESCRIPTION HERE.*$/gm, '');
-
-            // Remove all lines after and including "**BEFORE SUBMITTING, PLEASE READ ..."
-            body = body.replace(/\*\*BEFORE SUBMITTING, PLEASE READ.*\*\*[\s\S]*$/, '');
-
-            // Remove <details> section containing "PR Checklist (Click to Expand)"
-            body = body.replace(/(---\n\n)?<details>[\s\S]*?<summary>[\s\S]*?PR Checklist \(Click to Expand\)[\s\S]*?<\/summary>[\s\S]*?<\/details>/g, '');
-
-            if (body !== original) {
-              await github.rest.pulls.update({
-                owner,
-                repo,
-                pull_number: pr_number,
-                body,
-              });
-              console.log('Updated PR body');
-            } else {
-              console.log('No changes needed');
-            }
-
-  reminder-comment:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Post welcome comment for first-time contributors
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
-        with:
-          script: |
-            const { owner, repo } = context.repo;
-            const prAuthor = context.payload.pull_request.user.login;
-
-            const { data: searchResults } = await github.rest.search.issuesAndPullRequests({
-              q: `repo:${owner}/${repo} type:pr author:${prAuthor}`,
-              per_page: 1,
-            });
-
-            const authorPRCount = searchResults.total_count;
-            console.log(`Found ${authorPRCount} PRs by ${prAuthor}`);
-
-            if (authorPRCount === 1) {
-              console.log(`Posting welcome comment for first-time contributor: ${prAuthor}`);
-              await github.rest.issues.createComment({
-                owner,
-                repo,
-                issue_number: context.issue.number,
-                body: [
-                  '\u{1f44b} Hi! Thank you for contributing to the vLLM project.',
-                  '',
-                  '\u{1f4ac} Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.',
-                  '',
-                  'Just a reminder: PRs would not trigger full CI run by default.',
-                  '',
-                  'Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.',
-                  '',
-                  'To run CI, PR reviewers can either: Add `ready` label to the PR or enable auto-merge.',
-                  '',
-                  'If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.',
-                  '',
-                  '\u{1f680}',
-                ].join('\n'),
-              });
-            } else {
-              console.log(`Skipping comment for ${prAuthor} - not their first PR (${authorPRCount} PRs found)`);
-            }

.github/workflows/pre-commit.yml

@@ -11,39 +11,9 @@ concurrency:
 
 permissions:
   contents: read
-  pull-requests: read
 
 jobs:
-  pre-run-check:
-    if: github.event_name == 'pull_request'
-    runs-on: ubuntu-latest
-    steps:
-    - name: Check PR label and author merge count
-      uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
-      with:
-        script: |
-          const { data: pr } = await github.rest.pulls.get({
-            ...context.repo,
-            pull_number: context.payload.pull_request.number,
-          });
-
-          const hasReadyLabel = pr.labels.some(l => l.name === 'ready');
-
-          const { data: mergedPRs } = await github.rest.search.issuesAndPullRequests({
-            q: `repo:${context.repo.owner}/${context.repo.repo} is:pr is:merged author:${pr.user.login}`,
-            per_page: 4,
-          });
-          const mergedCount = mergedPRs.total_count;
-
-          if (hasReadyLabel || mergedCount >= 4) {
-            core.info(`Check passed: ready label=${hasReadyLabel}, 4+ merged PRs=${mergedCount >= 4}`);
-          } else {
-            core.setFailed(`PR must have the 'ready' label or the author must have at least 4 merged PRs (found ${mergedCount}).`);
-          }
-
   pre-commit:
-    needs: pre-run-check
-    if: always() && (needs.pre-run-check.result == 'success' || needs.pre-run-check.result == 'skipped')
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1

.github/workflows/reminder_comment.yml

@@ -0,0 +1,54 @@
+name: PR Reminder Comment Bot
+permissions:
+  pull-requests: write
+on:
+  pull_request_target:
+    types: [opened]
+jobs:
+  pr_reminder:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Remind to run full CI on PR
+        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+        with:
+          script: |
+            try {
+              // Get the PR author
+              const prAuthor = context.payload.pull_request.user.login;
+              
+              // Check if this is the author's first PR in this repository
+              // Use GitHub's search API to find all PRs by this author
+              const { data: searchResults } = await github.rest.search.issuesAndPullRequests({
+                q: `repo:${context.repo.owner}/${context.repo.repo} type:pr author:${prAuthor}`,
+                per_page: 100  
+              });
+              
+              const authorPRCount = searchResults.total_count;
+              
+              console.log(`Found ${authorPRCount} PRs by ${prAuthor}`);
+              
+              // Only post comment if this is the first PR (only one PR by this author)
+              if (authorPRCount === 1) {
+                console.log(`Posting welcome comment for first-time contributor: ${prAuthor}`);
+                await github.rest.issues.createComment({
+                owner: context.repo.owner,
+                repo: context.repo.repo,
+                issue_number: context.issue.number,
+                body: '👋 Hi! Thank you for contributing to the vLLM project.\n\n' +
+                  '💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.\n\n' +
+                  'Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run `fastcheck` CI which starts running only a small and essential subset of CI tests to quickly catch errors. \n\n' +
+                  'You ask your reviewers to trigger select CI tests on top of `fastcheck` CI. \n\n' +
+                  'Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.\n\n' +
+                  'To run CI, PR reviewers can either: Add `ready` label to the PR or enable auto-merge.\n\n' +
+                  'If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.\n\n' +
+                  '🚀'
+                });
+              } else {
+                console.log(`Skipping comment for ${prAuthor} - not their first PR (${authorPRCount} PRs found)`);
+              }
+            } catch (error) {
+              console.error('Error checking PR history or posting comment:', error);
+              // Don't fail the workflow, just log the error
+            }
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.gitignore

@@ -189,9 +189,11 @@ cython_debug/
 .vscode/
 
 # Claude
+CLAUDE.md
 .claude/
 
 # Codex
+AGENTS.md
 .codex/
 
 # Cursor

.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
-  rev: v0.21.0
+- repo: https://github.com/igorshubovych/markdownlint-cli
+  rev: v0.45.0
   hooks:
-  - id: markdownlint-cli2
-    language_version: lts
-    args: [--fix]
-    exclude: ^CLAUDE\.md$
+  - id: markdownlint
+    exclude: '.*\.inc\.md'
+    stages: [manual] # Only run in CI
 - 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

.readthedocs.yaml

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

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.
@@ -340,6 +340,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   list(APPEND VLLM_EXT_SRC
     "csrc/quantization/awq/gemm_kernels.cu"
+    "csrc/permute_cols.cu"
     "csrc/quantization/w8a8/cutlass/scaled_mm_entry.cu"
     "csrc/quantization/fp4/nvfp4_quant_entry.cu"
     "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
@@ -985,48 +986,6 @@ define_extension_target(
 # Setting this variable sidesteps the issue by calling the driver directly.
 target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
 
-# add OR VLLM_GPU_LANG STREQUAL "HIP" here once
-# https://github.com/vllm-project/vllm/issues/35163 is resolved
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  #
-  # _C_stable_libtorch extension (ops registered via STABLE_TORCH_LIBRARY)
-  #
-  set(VLLM_STABLE_EXT_SRC
-    "csrc/libtorch_stable/torch_bindings.cpp")
-
-  if(VLLM_GPU_LANG STREQUAL "CUDA")
-    list(APPEND VLLM_STABLE_EXT_SRC "csrc/libtorch_stable/permute_cols.cu")
-  endif()
-
-  if(VLLM_GPU_LANG STREQUAL "CUDA")
-    set_gencode_flags_for_srcs(
-      SRCS "${VLLM_STABLE_EXT_SRC}"
-      CUDA_ARCHS "${CUDA_ARCHS}")
-  endif()
-
-  message(STATUS "Enabling C_stable extension.")
-  define_extension_target(
-    _C_stable_libtorch
-    DESTINATION vllm
-    LANGUAGE ${VLLM_GPU_LANG}
-    SOURCES ${VLLM_STABLE_EXT_SRC}
-    COMPILE_FLAGS ${VLLM_GPU_FLAGS}
-    ARCHITECTURES ${VLLM_GPU_ARCHES}
-    USE_SABI 3
-    WITH_SOABI)
-
-  # Set TORCH_TARGET_VERSION for stable ABI compatibility.
-  # This ensures we only use C-shim APIs available in PyTorch 2.10.
-  # _C_stable_libtorch is abi compatible with PyTorch >= TORCH_TARGET_VERSION
-  # which is currently set to 2.10.
-  target_compile_definitions(_C_stable_libtorch PRIVATE
-    TORCH_TARGET_VERSION=0x020A000000000000ULL)
-
-  # Needed to use cuda APIs from C-shim
-  target_compile_definitions(_C_stable_libtorch PRIVATE
-    USE_CUDA)
-endif()
-
 #
 # _moe_C extension
 #
@@ -1040,7 +999,6 @@ 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/gpt_oss_router_gemm.cu"
     "csrc/moe/router_gemm.cu")
 endif()
 

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/benchmark.py

@@ -47,8 +47,6 @@ from common import (
     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 +59,7 @@ 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(
-        config.backend, config, prefill_backend=config.prefill_backend, **kwargs
-    )
+    return run_mla(config.backend, config, **kwargs)
 
 
 def run_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
@@ -444,27 +440,20 @@ 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 +469,6 @@ 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(
@@ -528,7 +502,7 @@ def main():
 
         # Override args with YAML values, but CLI args take precedence
         # Check if CLI provided backends (they would be non-None and not default)
-        cli_backends_provided = args.backend is not None or args.backends is not None
+        cli_backends_provided = args.backends is not None or args.backend is not None
 
         # Backend(s) - only use YAML if CLI didn't specify
         if not cli_backends_provided:
@@ -538,12 +512,6 @@ def main():
             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 +521,21 @@ 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.
-        cli_batch_specs_provided = args.batch_specs is not None
-        if not cli_batch_specs_provided:
-            if "batch_spec_ranges" in yaml_config:
-                # Generate batch specs from ranges
-                generated_specs = generate_batch_specs_from_ranges(
-                    yaml_config["batch_spec_ranges"]
-                )
-                # Combine with any explicit batch_specs
-                if "batch_specs" in yaml_config:
-                    args.batch_specs = yaml_config["batch_specs"] + generated_specs
-                else:
-                    args.batch_specs = generated_specs
-                console.print(
-                    f"[dim]Generated {len(generated_specs)} batch specs from ranges[/]"
-                )
-            elif "batch_specs" in yaml_config:
-                args.batch_specs = yaml_config["batch_specs"]
+        if "batch_spec_ranges" in yaml_config:
+            # Generate batch specs from ranges
+            generated_specs = generate_batch_specs_from_ranges(
+                yaml_config["batch_spec_ranges"]
+            )
+            # Combine with any explicit batch_specs
+            if "batch_specs" in yaml_config:
+                args.batch_specs = yaml_config["batch_specs"] + generated_specs
+            else:
+                args.batch_specs = generated_specs
+            console.print(
+                f"[dim]Generated {len(generated_specs)} batch specs from ranges[/]"
+            )
+        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"]
@@ -595,10 +560,6 @@ def main():
             args.warmup_iters = yaml_config["warmup_iters"]
         if "profile_memory" in yaml_config:
             args.profile_memory = yaml_config["profile_memory"]
-        if "kv_cache_dtype" in yaml_config:
-            args.kv_cache_dtype = yaml_config["kv_cache_dtype"]
-        if "cuda_graphs" in yaml_config:
-            args.cuda_graphs = yaml_config["cuda_graphs"]
 
         # Parameter sweep configuration
         if "parameter_sweep" in yaml_config:
@@ -652,19 +613,10 @@ 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 +669,6 @@ 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 +821,6 @@ 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 +843,6 @@ 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 +850,37 @@ def main():
 
     else:
         # Normal mode: compare backends
-        decode_results = []
-        prefill_results = []
+        total = len(backends) * len(args.batch_specs)
 
-        # Run decode backend comparison
-        if not prefill_backends:
-            # No prefill backends specified: compare decode backends as before
-            total = len(backends) * len(args.batch_specs)
+        with tqdm(total=total, desc="Benchmarking") as pbar:
+            for spec in args.batch_specs:
+                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,
+                    )
 
-            with tqdm(total=total, desc="Benchmarking") as pbar:
-                for spec in args.batch_specs:
-                    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)
+                    all_results.append(result)
 
-                        result = run_benchmark(config)
-                        decode_results.append(result)
+                    if not result.success:
+                        console.print(f"[red]Error {backend} {spec}: {result.error}[/]")
 
-                        if not result.success:
-                            console.print(
-                                f"[red]Error {backend} {spec}: {result.error}[/]"
-                            )
+                    pbar.update(1)
 
-                        pbar.update(1)
-
-            console.print("\n[bold green]Results:[/]")
-            formatter = ResultsFormatter(console)
-            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
+        # Display results
+        console.print("\n[bold green]Results:[/]")
+        formatter = ResultsFormatter(console)
+        formatter.print_table(all_results, backends)
 
     # Save results
     if all_results:

benchmarks/attention_benchmarks/common.py

@@ -30,7 +30,7 @@ def batch_spec_sort_key(spec: str) -> tuple[int, int, int]:
         max_kv_len = max(r.kv_len for r in requests) if requests else 0
         return (batch_size, max_q_len, max_kv_len)
     except Exception:
-        # Fallback for unparsable specs
+        # Fallback for unparseable specs
         return (0, 0, 0)
 
 
@@ -77,7 +77,6 @@ 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]:
         """
@@ -213,11 +212,7 @@ 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
@@ -372,7 +367,6 @@ class ResultsFormatter:
                     "backend",
                     "batch_spec",
                     "num_layers",
-                    "kv_cache_dtype",
                     "mean_time",
                     "std_time",
                     "throughput",
@@ -386,7 +380,6 @@ 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,

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
-  - "4q2ks4k_32q1s2k"       # 4 extend + 32 decode
-  - "2q1ks8k_32q1s2k"       # 2 large extend + 32 decode
+  - "2q1kkv2k_16q1s1k"       # 2 extend + 16 decode
+  - "4q2kkv4k_32q1s2k"       # 4 extend + 32 decode
+  - "2q1kkv8k_32q1s2k"       # 2 large extend + 32 decode
 
   # Explicitly chunked prefill
   - "q8k"           # 8k prefill with chunking hint

benchmarks/attention_benchmarks/configs/mla_prefill.yaml

@@ -1,19 +1,4 @@
-# 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"
+# MLA prefill-only benchmark configuration for sparse backends
 
 model:
   name: "deepseek-v3"
@@ -27,25 +12,20 @@ model:
   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
+# 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
-  - "q512"
-  - "q1k"
-  - "q2k"
-  - "q4k"
-  - "q8k"
+  - "1q512"
+  - "1q1k"
+  - "1q2k"
+  - "1q4k"
+  - "1q8k"
 
   # Batched pure prefill
   - "2q512"
@@ -64,63 +44,19 @@ batch_specs:
   - "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"
+  # Extend
+  - "1q512s4k"
+  - "1q512s8k"
+  - "1q1ks8k"
+  - "1q2ks8k"
+  - "1q2ks16k"
+  - "1q4ks16k"
 
-  # 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
+backends:
+  - FLASHMLA_SPARSE
+  - FLASHINFER_MLA_SPARSE
 
 device: "cuda:0"
-repeats: 20
-warmup_iters: 5
+repeats: 10
+warmup_iters: 3
+profile_memory: true

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/mla_runner.py

@@ -60,11 +60,8 @@ 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.
@@ -78,9 +75,6 @@ def create_minimal_vllm_config(
                   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
@@ -151,13 +145,14 @@ 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 +164,7 @@ 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,
@@ -177,84 +172,9 @@ 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
-# ============================================================================
-
-# Maps prefill backend names to attention config overrides.
-# FA backends set flash_attn_version and disable non-FA paths.
-# Non-FA backends enable their specific path and disable others.
-_PREFILL_BACKEND_CONFIG: dict[str, dict] = {
-    "fa2": {
-        "flash_attn_version": 2,
-        "disable_flashinfer_prefill": True,
-        "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,
-    },
-}
-
-
-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 Configuration
 # ============================================================================
 
 
@@ -284,7 +204,6 @@ def _get_backend_config(backend: str) -> dict:
     Returns:
         Dict with backend configuration
     """
-    from vllm.v1.attention.backend import MultipleOf
     from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
     try:
@@ -301,8 +220,8 @@ def _get_backend_config(backend: str) -> dict:
     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
+    if hasattr(block_size, "value"):
+        # Handle MultipleOf enum
         block_size = None
 
     # Check if sparse via class method if available
@@ -537,7 +456,6 @@ def _create_backend_impl(
     device: torch.device,
     max_num_tokens: int = 8192,
     index_topk: int | None = None,
-    kv_cache_dtype: str = "auto",
 ):
     """
     Create backend implementation instance.
@@ -586,7 +504,7 @@ def _create_backend_impl(
         "num_kv_heads": mla_dims["num_kv_heads"],
         "alibi_slopes": None,
         "sliding_window": None,
-        "kv_cache_dtype": kv_cache_dtype,
+        "kv_cache_dtype": "auto",
         "logits_soft_cap": None,
         "attn_type": "decoder",
         "kv_sharing_target_layer_name": None,
@@ -704,7 +622,6 @@ def _run_single_benchmark(
     mla_dims: dict,
     device: torch.device,
     indexer=None,
-    kv_cache_dtype: str | None = None,
 ) -> BenchmarkResult:
     """
     Run a single benchmark iteration.
@@ -738,123 +655,53 @@ def _run_single_benchmark(
     )
 
     # Create KV cache
-    if kv_cache_dtype is None:
-        kv_cache_dtype = getattr(config, "kv_cache_dtype", "auto")
-    head_size = mla_dims["kv_lora_rank"] + mla_dims["qk_rope_head_dim"]
-    if kv_cache_dtype == "fp8_ds_mla":
-        # FlashMLA sparse custom format: 656 bytes per token, stored as uint8.
-        # Layout: kv_lora_rank fp8 bytes + 4 float32 tile scales
-        #         + 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(
+        num_blocks,
+        block_size,
+        mla_dims["kv_lora_rank"] + mla_dims["qk_rope_head_dim"],
+        device=device,
+        dtype=torch.bfloat16,
+    )
 
-        kv_cache = torch.zeros(
-            num_blocks,
-            block_size,
-            head_size,
-            device=device,
-            dtype=torch.uint8,
-        ).view(current_platform.fp8_dtype())
-    else:
-        kv_cache = torch.zeros(
-            num_blocks,
-            block_size,
-            head_size,
-            device=device,
-            dtype=torch.bfloat16,
-        )
+    # Create input tensors for both decode and prefill modes
+    decode_inputs, prefill_inputs = _create_input_tensors(
+        total_q,
+        mla_dims,
+        backend_cfg["query_format"],
+        device,
+        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:
+    # Determine which forward method to use
+    if is_sparse:
+        # Sparse backends use forward_mqa
+        forward_fn = lambda: impl.forward_mqa(decode_inputs, kv_cache, metadata, layer)
+    elif 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:
         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()
-
-    # 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
+    torch.cuda.synchronize()
 
     # Benchmark
     times = []
@@ -864,10 +711,10 @@ 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)
 
@@ -886,7 +733,6 @@ 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.
@@ -898,13 +744,11 @@ def _run_mla_benchmark_batched(
     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 +758,7 @@ 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
@@ -931,91 +775,26 @@ def _run_mla_benchmark_batched(
     # 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
-        # with the current VllmConfig. These are @functools.cache decorated and
-        # would otherwise return stale results from a previous backend's config.
-        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)
@@ -1040,7 +819,6 @@ def _run_mla_benchmark_batched(
                     mla_dims,
                     device,
                     indexer=indexer,
-                    kv_cache_dtype=kv_cache_dtype,
                 )
                 results.append(result)
 
@@ -1067,7 +845,6 @@ def run_mla_benchmark(
     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.
@@ -1085,8 +862,6 @@ def run_mla_benchmark(
                                  (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)
@@ -1110,9 +885,7 @@ def run_mla_benchmark(
         return_single = True
 
     # Use unified batched execution
-    results = _run_mla_benchmark_batched(
-        backend, configs_with_params, index_topk, prefill_backend=prefill_backend
-    )
+    results = _run_mla_benchmark_batched(backend, configs_with_params, index_topk)
 
     # Return single result or list based on input
     return results[0] if return_single else results

benchmarks/attention_benchmarks/runner.py

@@ -140,7 +140,8 @@ def _create_vllm_config(
 
     cache_config = CacheConfig(
         block_size=config.block_size,
-        cache_dtype=config.kv_cache_dtype,
+        cache_dtype="auto",
+        swap_space=0,
     )
     cache_config.num_gpu_blocks = max_num_blocks
     cache_config.num_cpu_blocks = 0
@@ -215,7 +216,7 @@ def _create_backend_impl(
         num_kv_heads=config.num_kv_heads,
         alibi_slopes=None,
         sliding_window=None,
-        kv_cache_dtype=config.kv_cache_dtype,
+        kv_cache_dtype="auto",
     )
 
     kv_cache_spec = FullAttentionSpec(
@@ -288,22 +289,12 @@ def _create_input_tensors(
     total_q: int,
     device: torch.device,
     dtype: torch.dtype,
-    quantize_query: bool = False,
 ) -> tuple:
-    """Create Q, K, V input tensors for all layers.
-
-    When quantize_query is True, queries are cast to fp8 to match backends
-    that require query/key/value dtype consistency.
-    """
-    q_dtype = dtype
-    if quantize_query:
-        from vllm.platforms import current_platform
-
-        q_dtype = current_platform.fp8_dtype()
+    """Create Q, K, V input tensors for all layers."""
     q_list = [
         torch.randn(
             total_q, config.num_q_heads, config.head_dim, device=device, dtype=dtype
-        ).to(q_dtype)
+        )
         for _ in range(config.num_layers)
     ]
     k_list = [
@@ -354,17 +345,10 @@ def _create_kv_cache(
     # Compute inverse permutation to get back to logical view
     inv_order = [stride_order.index(i) for i in range(len(stride_order))]
 
-    # Use fp8 dtype for cache when requested.
-    cache_dtype = dtype
-    if config.kv_cache_dtype == "fp8":
-        from vllm.platforms import current_platform
-
-        cache_dtype = current_platform.fp8_dtype()
-
     cache_list = []
     for _ in range(config.num_layers):
         # Allocate in physical layout order (contiguous in memory)
-        cache = torch.zeros(*physical_shape, device=device, dtype=cache_dtype)
+        cache = torch.zeros(*physical_shape, device=device, dtype=dtype)
         # Permute to logical view
         cache = cache.permute(*inv_order)
         cache_list.append(cache)
@@ -407,38 +391,7 @@ def _run_single_benchmark(
                 attn_metadata,
                 output=out,
             )
-    torch.accelerator.synchronize()
-
-    # Optionally capture a CUDA graph after warmup.
-    # Graph replay eliminates CPU launch overhead so timings reflect pure
-    # kernel time.
-    if config.use_cuda_graphs:
-        graph = torch.cuda.CUDAGraph()
-        with torch.cuda.graph(graph):
-            for i in range(config.num_layers):
-                impl.forward(
-                    layer,
-                    q_list[i],
-                    k_list[i],
-                    v_list[i],
-                    cache_list[i],
-                    attn_metadata,
-                    output=out,
-                )
-        benchmark_fn = graph.replay
-    else:
-
-        def benchmark_fn():
-            for i in range(config.num_layers):
-                impl.forward(
-                    layer,
-                    q_list[i],
-                    k_list[i],
-                    v_list[i],
-                    cache_list[i],
-                    attn_metadata,
-                    output=out,
-                )
+    torch.cuda.synchronize()
 
     # Benchmark
     times = []
@@ -447,18 +400,27 @@ def _run_single_benchmark(
         end = torch.cuda.Event(enable_timing=True)
 
         start.record()
-        benchmark_fn()
+        for i in range(config.num_layers):
+            impl.forward(
+                layer,
+                q_list[i],
+                k_list[i],
+                v_list[i],
+                cache_list[i],
+                attn_metadata,
+                output=out,
+            )
         end.record()
 
-        torch.accelerator.synchronize()
+        torch.cuda.synchronize()
         elapsed_ms = start.elapsed_time(end)
         times.append(elapsed_ms / 1000.0 / config.num_layers)  # seconds per layer
 
     mem_stats = {}
     if config.profile_memory:
         mem_stats = {
-            "allocated_mb": torch.accelerator.memory_allocated(device) / 1024**2,
-            "reserved_mb": torch.accelerator.memory_reserved(device) / 1024**2,
+            "allocated_mb": torch.cuda.memory_allocated(device) / 1024**2,
+            "reserved_mb": torch.cuda.memory_reserved(device) / 1024**2,
         }
 
     return times, mem_stats
@@ -482,7 +444,7 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
         BenchmarkResult with timing and memory statistics
     """
     device = torch.device(config.device)
-    torch.accelerator.set_device_index(device)
+    torch.cuda.set_device(device)
 
     backend_cfg = _get_backend_config(config.backend)
 
@@ -541,12 +503,8 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 common_attn_metadata=common_metadata,
             )
 
-            # Only quantize queries when the impl supports it
-            quantize_query = config.kv_cache_dtype.startswith("fp8") and getattr(
-                impl, "supports_quant_query_input", False
-            )
             q_list, k_list, v_list = _create_input_tensors(
-                config, total_q, device, dtype, quantize_query=quantize_query
+                config, total_q, device, dtype
             )
 
             cache_list = _create_kv_cache(

benchmarks/auto_tune/README.md

@@ -41,7 +41,7 @@ MODEL=meta-llama/Llama-3.3-70B-Instruct SYSTEM=TPU TP=8 DOWNLOAD_DIR='' INPUT_LE
 | --- | --- | --- |
 | `BASE` | **Required.** The absolute path to the parent directory of your vLLM repository directory. | `"$HOME"` |
 | `MODEL` | **Required.** The Hugging Face model identifier to be served by vllm. | `"meta-llama/Llama-3.1-8B-Instruct"` |
-| `SYSTEM` | **Required.** The hardware you are running on. Choices: `TPU` or `GPU`. (For other systems, it might not support saving profiles) | `"TPU"` |
+| `SYSTEM`| **Required.** The hardware you are running on. Choices: `TPU` or `GPU`. (For other systems, it might not support saving profiles) | `"TPU"` |
 | `TP` | **Required.** The tensor-parallelism size. | `1` |
 | `DOWNLOAD_DIR` | **Required.** Directory to download and load model weights from. | `""` (default download path) |
 | `INPUT_LEN` | **Required.** Request input length. | `4000` |

benchmarks/benchmark_long_document_qa_throughput.py

@@ -40,9 +40,9 @@ LLM engine. You can refer to the `vllm.engine.arg_utils.EngineArgs` for more
 details.
 """
 
+import dataclasses
 import random
 import time
-from dataclasses import fields
 
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
@@ -124,7 +124,7 @@ def main(args):
 
     # Create the LLM engine
     engine_args = EngineArgs.from_cli_args(args)
-    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
+    llm = LLM(**dataclasses.asdict(engine_args))
     sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
 
     print("------warm up------")

benchmarks/benchmark_prefix_caching.py

@@ -32,7 +32,6 @@ import dataclasses
 import json
 import random
 import time
-from dataclasses import fields
 
 from transformers import PreTrainedTokenizerBase
 
@@ -197,7 +196,7 @@ def main(args):
 
     engine_args = EngineArgs.from_cli_args(args)
 
-    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
+    llm = LLM(**dataclasses.asdict(engine_args))
 
     sampling_params = SamplingParams(
         temperature=0,

benchmarks/benchmark_prioritization.py

@@ -3,10 +3,10 @@
 """Benchmark offline prioritization."""
 
 import argparse
+import dataclasses
 import json
 import random
 import time
-from dataclasses import fields
 
 from transformers import AutoTokenizer, PreTrainedTokenizerBase
 
@@ -79,7 +79,7 @@ def run_vllm(
 ) -> float:
     from vllm import LLM, SamplingParams
 
-    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
+    llm = LLM(**dataclasses.asdict(engine_args))
 
     assert all(
         llm.llm_engine.model_config.max_model_len >= (request[1] + request[2])

benchmarks/benchmark_topk_topp.py

@@ -94,18 +94,15 @@ def create_logits(
 
 def measure_memory() -> tuple[int, int]:
     """Return (allocated, reserved) memory in bytes."""
-    torch.accelerator.synchronize()
-    return (
-        torch.accelerator.memory_allocated(),
-        torch.accelerator.max_memory_allocated(),
-    )
+    torch.cuda.synchronize()
+    return torch.cuda.memory_allocated(), torch.cuda.max_memory_allocated()
 
 
 def reset_memory_stats():
     """Reset peak memory statistics."""
     reset_buffer_cache()
-    torch.accelerator.reset_peak_memory_stats()
-    torch.accelerator.empty_cache()
+    torch.cuda.reset_peak_memory_stats()
+    torch.cuda.empty_cache()
     gc.collect()
 
 
@@ -126,7 +123,7 @@ def benchmark_function(
     for _ in range(warmup_iters):
         logits_copy = logits.clone()
         func(logits_copy, k, p)
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     # Reset memory stats before benchmark
     reset_memory_stats()
@@ -143,7 +140,7 @@ def benchmark_function(
         func(logits_copy, k, p)
         end_events[i].record()
 
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     # Calculate timing
     times = [

benchmarks/kernels/bench_concat_mla_q.py

@@ -1,98 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import argparse
-
-import torch
-
-from vllm import _custom_ops as ops
-from vllm.triton_utils import triton
-
-# DeepSeek V3 dimensions
-NOPE_DIM = 512
-ROPE_DIM = 64
-NUM_HEADS = 128
-
-NUM_TOKENS = [8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192]
-
-
-def get_configs():
-    return NUM_TOKENS
-
-
-def make_inputs(num_tokens, dtype):
-    """Create inputs matching the real code path.
-
-    Args:
-        contiguous_nope: If False, simulate the transposed BMM output
-                         (non-contiguous nope with stride pattern from
-                         [N,B,L].transpose(0,1)).
-    """
-    # Simulate: bmm output [N, B, L].transpose(0, 1) -> [B, N, L]
-    raw = torch.randn(NUM_HEADS, num_tokens, NOPE_DIM, dtype=dtype, device="cuda")
-    ql_nope = raw.transpose(0, 1)
-
-    q_pe = torch.randn(num_tokens, NUM_HEADS, ROPE_DIM, dtype=dtype, device="cuda")
-    return ql_nope, q_pe
-
-
-# ---- Non-contiguous nope benchmark (real code path) ----
-@triton.testing.perf_report(
-    triton.testing.Benchmark(
-        x_names=["num_tokens"],
-        x_vals=get_configs(),
-        line_arg="provider",
-        line_vals=["torch_cat", "concat_mla_q"],
-        line_names=["torch.cat", "concat_mla_q (v8)"],
-        styles=[("blue", "--"), ("green", "-")],
-        ylabel="Latency (us)",
-        plot_name="concat_mla_q-transposed",
-        args={},
-    )
-)
-def bench_transposed(num_tokens, provider):
-    dtype = torch.bfloat16
-    ql_nope, q_pe = make_inputs(num_tokens, dtype)
-
-    q_out = torch.empty(
-        num_tokens, NUM_HEADS, NOPE_DIM + ROPE_DIM, dtype=dtype, device="cuda"
-    )
-
-    quantiles = [0.5, 0.2, 0.8]
-
-    if provider == "torch_cat":
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: torch.cat((ql_nope, q_pe), dim=-1), quantiles=quantiles, rep=500
-        )
-    else:
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: ops.concat_mla_q(ql_nope, q_pe, q_out), quantiles=quantiles, rep=500
-        )
-
-    return ms * 1000, max_ms * 1000, min_ms * 1000  # us
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Benchmark concat_mla_q vs torch.cat")
-    parser.add_argument(
-        "--save-path", type=str, default=None, help="Path to save benchmark results"
-    )
-    args = parser.parse_args()
-
-    print("\n" + "=" * 70)
-    print("CONCAT MLA Q KERNEL BENCHMARKS")
-    print("=" * 70)
-    print(f"Dimensions: nope={NOPE_DIM}, rope={ROPE_DIM}, heads={NUM_HEADS}")
-    print(
-        f"Per-head output: {NOPE_DIM + ROPE_DIM} bf16 = "
-        f"{(NOPE_DIM + ROPE_DIM) * 2} bytes"
-    )
-    print(f"num_tokens (decode=batch_size, prefill=chunk_size): {NUM_TOKENS}")
-    print("=" * 70)
-
-    print("\n--- Non-contiguous nope inputs (transposed BMM output) ---")
-    bench_transposed.run(print_data=True, save_path=args.save_path)
-
-    print("\n" + "=" * 70)
-    print("Benchmarking complete!")
-    print("=" * 70)

benchmarks/kernels/bench_cp_gather_fp8.py

@@ -1,153 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import argparse
-import math
-
-import torch
-
-from vllm import _custom_ops as ops
-from vllm.triton_utils import triton
-
-# DeepSeek V3 MLA dimensions
-NOPE_DIM = 512
-ROPE_DIM = 64
-HEAD_DIM = NOPE_DIM + ROPE_DIM  # 576 BF16 output elements per token
-ENTRY_BYTES = 656  # 512 FP8 + 16 scales + 128 BF16 RoPE
-BLOCK_SIZE = 64  # tokens per physical cache block - get_supported_kernel_block_sizes
-
-# Realistic prefill scenarios:
-#   - 1 long prefill: single request, 16K-96K tokens
-#   - 4 medium prefills: 4 requests, 4K-24K tokens each
-#   - 16 shorter prefills: 16 requests, 1K-6K tokens each
-SCENARIOS = [
-    # (label, num_reqs, total_tokens_list)
-    ("1-req", 1, [8192, 16384, 32768, 65536, 98304]),
-    ("4-reqs", 4, [8192, 16384, 32768, 65536, 98304]),
-    ("16-reqs", 16, [8192, 16384, 32768, 65536, 98304]),
-]
-
-
-def make_inputs(total_tokens, num_reqs, block_size):
-    """Create synthetic FP8 cache, block table, and output buffer.
-
-    Fills the cache with random bytes (we only measure throughput,
-    not correctness). Block table maps each request to contiguous
-    physical blocks.
-    """
-    # Divide tokens evenly across requests
-    base_len = total_tokens // num_reqs
-    remainder = total_tokens % num_reqs
-    seq_lens = [base_len + (1 if r < remainder else 0) for r in range(num_reqs)]
-
-    # workspace_starts: cumulative sum of seq_lens
-    workspace_starts = [0] * num_reqs
-    for r in range(1, num_reqs):
-        workspace_starts[r] = workspace_starts[r - 1] + seq_lens[r - 1]
-
-    # Physical blocks needed per request
-    blocks_per_req = [math.ceil(s / block_size) for s in seq_lens]
-    total_blocks = sum(blocks_per_req)
-    max_blocks = max(blocks_per_req)
-
-    # Allocate cache with random data (content doesn't matter for perf)
-    cache = torch.randint(
-        0,
-        256,
-        (total_blocks, block_size, ENTRY_BYTES),
-        dtype=torch.uint8,
-        device="cuda",
-    )
-
-    # Block table: contiguous block assignments
-    block_table = torch.zeros(num_reqs, max_blocks, dtype=torch.int32, device="cuda")
-    block_idx = 0
-    for r in range(num_reqs):
-        for b in range(blocks_per_req[r]):
-            block_table[r, b] = block_idx
-            block_idx += 1
-
-    # Output workspace
-    dst = torch.zeros(total_tokens, HEAD_DIM, dtype=torch.bfloat16, device="cuda")
-
-    seq_lens_t = torch.tensor(seq_lens, dtype=torch.int32, device="cuda")
-    workspace_starts_t = torch.tensor(
-        workspace_starts, dtype=torch.int32, device="cuda"
-    )
-
-    return cache, dst, block_table, seq_lens_t, workspace_starts_t
-
-
-def bench_scenario(label, num_reqs, total_tokens_list, save_path):
-    """Run benchmark for a specific (num_reqs, total_tokens) scenario."""
-
-    @triton.testing.perf_report(
-        triton.testing.Benchmark(
-            x_names=["total_tokens"],
-            x_vals=total_tokens_list,
-            line_arg="provider",
-            line_vals=["cuda_kernel"],
-            line_names=["cp_gather_fp8 (CUDA)"],
-            styles=[("green", "-")],
-            ylabel="Latency (us)",
-            plot_name=f"cp_gather_fp8-{label}-bs{BLOCK_SIZE}",
-            args={"num_reqs": num_reqs},
-        )
-    )
-    def bench_fn(total_tokens, provider, num_reqs):
-        cache, dst, block_table, seq_lens_t, ws_starts = make_inputs(
-            total_tokens, num_reqs, BLOCK_SIZE
-        )
-
-        quantiles = [0.5, 0.2, 0.8]
-
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: ops.cp_gather_and_upconvert_fp8_kv_cache(
-                cache, dst, block_table, seq_lens_t, ws_starts, num_reqs
-            ),
-            quantiles=quantiles,
-            rep=500,
-        )
-
-        return ms * 1000, max_ms * 1000, min_ms * 1000  # us
-
-    seq_len_per_req = total_tokens_list[0] // num_reqs
-    seq_len_per_req_max = total_tokens_list[-1] // num_reqs
-    print(
-        f"\n--- {label}: {num_reqs} request(s), "
-        f"~{seq_len_per_req}-{seq_len_per_req_max} tokens/req ---"
-    )
-    bench_fn.run(print_data=True, save_path=save_path)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
-        description="Benchmark cp_gather_and_upconvert_fp8_kv_cache"
-    )
-    parser.add_argument(
-        "--save-path",
-        type=str,
-        default=None,
-        help="Path to save benchmark results as CSV",
-    )
-    args = parser.parse_args()
-
-    # Print data volume info for bandwidth analysis
-    read_per_token = ENTRY_BYTES  # 656 bytes from cache
-    write_per_token = HEAD_DIM * 2  # 576 * 2 = 1152 bytes to workspace
-    total_per_token = read_per_token + write_per_token  # 1808 bytes
-
-    print("\n" + "=" * 70)
-    print("CP_GATHER_AND_UPCONVERT_FP8_KV_CACHE BENCHMARKS")
-    print("=" * 70)
-    print(f"Cache entry: {ENTRY_BYTES} bytes (512 FP8 + 16 scales + 128 RoPE)")
-    print(f"Output row:  {HEAD_DIM} BF16 = {HEAD_DIM * 2} bytes")
-    print(f"Per token:   {total_per_token} bytes (read + write)")
-    print(f"Block size:  {BLOCK_SIZE} tokens/block")
-    print("=" * 70)
-
-    for label, num_reqs, total_tokens_list in SCENARIOS:
-        bench_scenario(label, num_reqs, total_tokens_list, args.save_path)
-
-    print("\n" + "=" * 70)
-    print("Benchmarking complete!")
-    print("=" * 70)

benchmarks/kernels/benchmark_2d_silu_mul_fp8_quant.py

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

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -64,7 +64,7 @@ def bench_run(
     per_out_ch: bool,
     mkn: tuple[int, int, int],
 ):
-    init_workspace_manager(torch.accelerator.current_device_index())
+    init_workspace_manager(torch.cuda.current_device())
     (m, k, n) = mkn
 
     dtype = torch.half
@@ -171,7 +171,7 @@ def bench_run(
                 activation=MoEActivation.SILU,
                 global_num_experts=num_experts,
             )
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     # Create CUDA graphs for Triton (match benchmark_moe.py pattern exactly)
     triton_stream = torch.cuda.Stream()
@@ -187,14 +187,14 @@ def bench_run(
                 topk_ids,
                 quant_config=quant_config,
             )
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     def bench_cuda_graph(graph, num_warmup=5, num_iters=100):
         """Benchmark CUDA graph using events like benchmark_moe.py"""
         # Warmup
         for _ in range(num_warmup):
             graph.replay()
-        torch.accelerator.synchronize()
+        torch.cuda.synchronize()
 
         # Timing
         start_event = torch.Event(enable_timing=True)
@@ -202,7 +202,7 @@ def bench_run(
 
         latencies = []
         for _ in range(num_iters):
-            torch.accelerator.synchronize()
+            torch.cuda.synchronize()
             start_event.record()
             graph.replay()
             end_event.record()

benchmarks/kernels/benchmark_cutlass_moe_nvfp4.py

@@ -307,7 +307,7 @@ def bench_run(
     def replay_graph(graph, num_repeats):
         for _ in range(num_repeats):
             graph.replay()
-        torch.accelerator.synchronize()
+        torch.cuda.synchronize()
 
     cutlass_stream = torch.cuda.Stream()
     cutlass_graph = torch.cuda.CUDAGraph()
@@ -330,7 +330,7 @@ def bench_run(
             e=num_experts,
             device=device,
         )
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     triton_stream = torch.cuda.Stream()
     triton_graph = torch.cuda.CUDAGraph()
@@ -345,7 +345,7 @@ def bench_run(
             w2_fp8scale,
             a_fp8_scale,
         )
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     min_run_time = 5
     num_warmup = 5

benchmarks/kernels/benchmark_device_communicators.py

@@ -342,7 +342,7 @@ class CommunicatorBenchmark:
             if not should_use_fn(tensor):
                 return None
 
-            torch.accelerator.synchronize()
+            torch.cuda.synchronize()
             stream = torch.cuda.Stream()
             with torch.cuda.stream(stream):
                 graph_input = tensor.clone()
@@ -360,17 +360,17 @@ class CommunicatorBenchmark:
                         for _ in range(CUDA_GRAPH_CAPTURE_CYCLES):
                             allreduce_fn(graph_input)
 
-            torch.accelerator.synchronize()
+            torch.cuda.synchronize()
             for _ in range(num_warmup):
                 graph.replay()
-            torch.accelerator.synchronize()
+            torch.cuda.synchronize()
 
-            torch.accelerator.synchronize()
+            torch.cuda.synchronize()
             start_time = time.perf_counter()
 
             for _ in range(num_trials):
                 graph.replay()
-            torch.accelerator.synchronize()
+            torch.cuda.synchronize()
 
             end_time = time.perf_counter()
 
@@ -495,7 +495,7 @@ def main():
 
     # Set device
     device = torch.device(f"cuda:{rank}")
-    torch.accelerator.set_device_index(device)
+    torch.cuda.set_device(device)
 
     # Get CPU process group
     cpu_group = dist.new_group(backend="gloo")

benchmarks/kernels/benchmark_fused_collective.py

@@ -385,32 +385,32 @@ def benchmark_operation(
     # Warmup before graph capture
     for _ in range(warmup):
         operation_func(*args, **kwargs)
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     # Create CUDA graph
     graph = torch.cuda.CUDAGraph()
     num_op_per_cudagraph = 10
 
     # Use vLLM's graph_capture to make tensor_model_parallel_all_reduce graph-safe
-    device = torch.device(f"cuda:{torch.accelerator.current_device_index()}")
+    device = torch.device(f"cuda:{torch.cuda.current_device()}")
     with graph_capture(device=device), torch.cuda.graph(graph):
         for _ in range(num_op_per_cudagraph):
             operation_func(*args, **kwargs)
 
     # Graph warmup
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
     for _ in range(warmup):
         graph.replay()
 
     # Benchmark with CUDA graph
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
     start_time = time.perf_counter()
 
     for _ in range(trials // num_op_per_cudagraph):
         # operation_func(*args, **kwargs)
         graph.replay()
 
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
     end_time = time.perf_counter()
 
     avg_time_ms = ((end_time - start_time) / trials) * 1000
@@ -984,7 +984,7 @@ def main():
     world_size = int(os.environ["WORLD_SIZE"])
 
     device = torch.device(f"cuda:{rank}")
-    torch.accelerator.set_device_index(device)
+    torch.cuda.set_device(device)
     torch.set_default_device(device)
 
     init_distributed_environment()

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -50,7 +50,7 @@ def bench_run(
     per_out_ch: bool,
     mkn: tuple[int, int, int],
 ):
-    init_workspace_manager(torch.accelerator.current_device_index())
+    init_workspace_manager(torch.cuda.current_device())
     label = "Quant Matmul"
 
     sub_label = (
@@ -224,7 +224,7 @@ def bench_run(
     def replay_graph(graph, num_repeats):
         for _ in range(num_repeats):
             graph.replay()
-        torch.accelerator.synchronize()
+        torch.cuda.synchronize()
 
     cutlass_stream = torch.cuda.Stream()
     cutlass_graph = torch.cuda.CUDAGraph()
@@ -239,7 +239,7 @@ def bench_run(
             topk_weights,
             topk_ids,
         )
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     triton_stream = torch.cuda.Stream()
     triton_graph = torch.cuda.CUDAGraph()
@@ -254,7 +254,7 @@ def bench_run(
             w2_scale,
             a_scale,
         )
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     min_run_time = 5
     num_warmup = 5

benchmarks/kernels/benchmark_layernorm.py

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

benchmarks/kernels/benchmark_lora.py

@@ -1035,7 +1035,7 @@ def bench_optype(
     # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are set up
     for kwargs in kwargs_list:
         op_type.bench_fn()(**kwargs)
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     # Merge into a single kwargs and qualify arguments as ArgPool
     kwargs = {k: ArgPool([]) for k in kwargs_list[0]}

benchmarks/kernels/benchmark_mla_k_concat.py

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

benchmarks/kernels/benchmark_moe.py

@@ -54,7 +54,7 @@ def clear_triton_cache():
 
     # Clear CUDA memory cache
     if torch.cuda.is_available():
-        torch.accelerator.empty_cache()
+        torch.cuda.empty_cache()
 
     # Try to clear Triton's runtime cache
     try:
@@ -304,19 +304,19 @@ def benchmark_config(
 
     # JIT compilation & warmup
     run()
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     # Capture 10 invocations with CUDA graph
     graph = torch.cuda.CUDAGraph()
     with torch.cuda.graph(graph):
         for _ in range(10):
             run()
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     # Warmup
     for _ in range(5):
         graph.replay()
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     start_event = torch.Event(enable_timing=True)
     end_event = torch.Event(enable_timing=True)
@@ -324,7 +324,7 @@ def benchmark_config(
     latencies: list[float] = []
     for i in range(num_iters):
         prepare(i)
-        torch.accelerator.synchronize()
+        torch.cuda.synchronize()
 
         start_event.record()
         graph.replay()
@@ -626,11 +626,7 @@ class BenchmarkWorker:
             if visible_device != f"{self.device_id}":
                 need_device_guard = True
 
-        with (
-            torch.accelerator.device_index(self.device_id)
-            if need_device_guard
-            else nullcontext()
-        ):
+        with torch.cuda.device(self.device_id) if need_device_guard else nullcontext():
             for idx, config in enumerate(tqdm(search_space)):
                 try:
                     kernel_time = benchmark_config(
@@ -750,20 +746,17 @@ def get_weight_block_size_safety(config, default_value=None):
 
 
 def get_model_params(config):
-    architectures = getattr(config, "architectures", None) or [type(config).__name__]
-    architecture = architectures[0]
-
-    if architecture == "DbrxForCausalLM":
+    if config.architectures[0] == "DbrxForCausalLM":
         E = config.ffn_config.moe_num_experts
         topk = config.ffn_config.moe_top_k
         intermediate_size = config.ffn_config.ffn_hidden_size
         hidden_size = config.hidden_size
-    elif architecture == "JambaForCausalLM":
+    elif config.architectures[0] == "JambaForCausalLM":
         E = config.num_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
         hidden_size = config.hidden_size
-    elif architecture in (
+    elif config.architectures[0] in (
         "DeepseekV2ForCausalLM",
         "DeepseekV3ForCausalLM",
         "DeepseekV32ForCausalLM",
@@ -777,7 +770,7 @@ def get_model_params(config):
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
         hidden_size = config.hidden_size
-    elif architecture in (
+    elif config.architectures[0] in (
         "Qwen2MoeForCausalLM",
         "Qwen3MoeForCausalLM",
         "Qwen3NextForCausalLM",
@@ -786,27 +779,23 @@ def get_model_params(config):
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
         hidden_size = config.hidden_size
-    elif architecture in (
-        "Qwen3VLMoeForConditionalGeneration",
-        "Qwen3_5MoeForConditionalGeneration",
-        "Qwen3_5MoeTextConfig",
-    ):
+    elif config.architectures[0] == "Qwen3VLMoeForConditionalGeneration":
         text_config = config.get_text_config()
         E = text_config.num_experts
         topk = text_config.num_experts_per_tok
         intermediate_size = text_config.moe_intermediate_size
         hidden_size = text_config.hidden_size
-    elif architecture == "HunYuanMoEV1ForCausalLM":
+    elif config.architectures[0] == "HunYuanMoEV1ForCausalLM":
         E = config.num_experts
         topk = config.moe_topk[0]
         intermediate_size = config.moe_intermediate_size[0]
         hidden_size = config.hidden_size
-    elif architecture == "Qwen3OmniMoeForConditionalGeneration":
+    elif config.architectures[0] == "Qwen3OmniMoeForConditionalGeneration":
         E = config.thinker_config.text_config.num_experts
         topk = config.thinker_config.text_config.num_experts_per_tok
         intermediate_size = config.thinker_config.text_config.moe_intermediate_size
         hidden_size = config.thinker_config.text_config.hidden_size
-    elif architecture == "PixtralForConditionalGeneration":
+    elif config.architectures[0] == "PixtralForConditionalGeneration":
         # Pixtral can contain different LLM architectures,
         # recurse to get their parameters
         return get_model_params(config.get_text_config())
@@ -821,23 +810,6 @@ def get_model_params(config):
     return E, topk, intermediate_size, hidden_size
 
 
-def resolve_dtype(config) -> torch.dtype:
-    if current_platform.is_rocm():
-        return torch.float16
-
-    dtype = getattr(config, "dtype", None)
-    if dtype is not None:
-        return dtype
-
-    if hasattr(config, "get_text_config"):
-        text_config = config.get_text_config()
-        dtype = getattr(text_config, "dtype", None)
-        if dtype is not None:
-            return dtype
-
-    return torch.bfloat16
-
-
 def get_quantization_group_size(config) -> int | None:
     """Extract the quantization group size from the HF model config.
 
@@ -885,7 +857,7 @@ def main(args: argparse.Namespace):
     else:
         ensure_divisibility(intermediate_size, args.tp_size, "intermediate_size")
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    dtype = resolve_dtype(config)
+    dtype = torch.float16 if current_platform.is_rocm() else config.dtype
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"
     use_int8_w8a16 = args.dtype == "int8_w8a16"
     use_int4_w4a16 = args.dtype == "int4_w4a16"

benchmarks/kernels/benchmark_moe_defaults.py

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

benchmarks/kernels/benchmark_moe_permute_unpermute.py

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

benchmarks/kernels/benchmark_mrope.py

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

benchmarks/kernels/benchmark_paged_attention.py

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

benchmarks/kernels/benchmark_per_token_group_quant.py

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

benchmarks/kernels/benchmark_quant.py

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

benchmarks/kernels/benchmark_reshape_and_cache.py

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

benchmarks/kernels/benchmark_reshape_and_cache_flash.py

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

benchmarks/kernels/benchmark_router_gemm.py

@@ -1,134 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import torch
-import torch.nn.functional as F
-
-from vllm import _custom_ops as ops
-from vllm.platforms import current_platform
-from vllm.transformers_utils.config import get_config
-from vllm.triton_utils import triton
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-
-# Dimensions supported by the DSV3 specialized kernel
-DSV3_SUPPORTED_NUM_EXPERTS = [256, 384]
-DSV3_SUPPORTED_HIDDEN_SIZES = [7168]
-
-# Dimensions supported by the gpt-oss specialized kernel
-GPT_OSS_SUPPORTED_NUM_EXPERTS = [32, 128]
-GPT_OSS_SUPPORTED_HIDDEN_SIZES = [2880]
-
-
-def get_batch_size_range(max_batch_size):
-    return [2**x for x in range(14) if 2**x <= max_batch_size]
-
-
-def get_model_params(config):
-    if config.architectures[0] in (
-        "DeepseekV2ForCausalLM",
-        "DeepseekV3ForCausalLM",
-        "DeepseekV32ForCausalLM",
-    ):
-        num_experts = config.n_routed_experts
-        hidden_size = config.hidden_size
-    elif config.architectures[0] in ("GptOssForCausalLM",):
-        num_experts = config.num_local_experts
-        hidden_size = config.hidden_size
-    else:
-        raise ValueError(f"Unsupported architecture: {config.architectures}")
-    return num_experts, hidden_size
-
-
-def get_benchmark(model, max_batch_size, trust_remote_code):
-    @triton.testing.perf_report(
-        triton.testing.Benchmark(
-            x_names=["batch_size"],
-            x_vals=get_batch_size_range(max_batch_size),
-            x_log=False,
-            line_arg="provider",
-            line_vals=[
-                "torch",
-                "vllm",
-            ],
-            line_names=["PyTorch", "vLLM"],
-            styles=([("blue", "-"), ("red", "-")]),
-            ylabel="TFLOPs",
-            plot_name=f"{model} router gemm throughput",
-            args={},
-        )
-    )
-    def benchmark(batch_size, provider):
-        config = get_config(model=model, trust_remote_code=trust_remote_code)
-        num_experts, hidden_size = get_model_params(config)
-
-        mat_a = torch.randn(
-            (batch_size, hidden_size), dtype=torch.bfloat16, device="cuda"
-        ).contiguous()
-        mat_b = torch.randn(
-            (num_experts, hidden_size), dtype=torch.bfloat16, device="cuda"
-        ).contiguous()
-        bias = torch.randn(
-            num_experts, dtype=torch.bfloat16, device="cuda"
-        ).contiguous()
-
-        is_hopper_or_blackwell = current_platform.is_device_capability(
-            90
-        ) or current_platform.is_device_capability_family(100)
-        allow_dsv3_router_gemm = (
-            is_hopper_or_blackwell
-            and num_experts in DSV3_SUPPORTED_NUM_EXPERTS
-            and hidden_size in DSV3_SUPPORTED_HIDDEN_SIZES
-        )
-        allow_gpt_oss_router_gemm = (
-            is_hopper_or_blackwell
-            and num_experts in GPT_OSS_SUPPORTED_NUM_EXPERTS
-            and hidden_size in GPT_OSS_SUPPORTED_HIDDEN_SIZES
-        )
-
-        has_bias = False
-        if allow_gpt_oss_router_gemm:
-            has_bias = True
-
-        quantiles = [0.5, 0.2, 0.8]
-
-        if provider == "torch":
-
-            def runner():
-                if has_bias:
-                    F.linear(mat_a, mat_b, bias)
-                else:
-                    F.linear(mat_a, mat_b)
-        elif provider == "vllm":
-
-            def runner():
-                if allow_dsv3_router_gemm:
-                    ops.dsv3_router_gemm(mat_a, mat_b, torch.bfloat16)
-                elif allow_gpt_oss_router_gemm:
-                    ops.gpt_oss_router_gemm(mat_a, mat_b, bias)
-                else:
-                    raise ValueError("Unsupported router gemm")
-
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            runner, quantiles=quantiles
-        )
-
-        def tflops(t_ms):
-            flops = 2 * batch_size * hidden_size * num_experts
-            return flops / (t_ms * 1e-3) / 1e12
-
-        return tflops(ms), tflops(max_ms), tflops(min_ms)
-
-    return benchmark
-
-
-if __name__ == "__main__":
-    parser = FlexibleArgumentParser()
-    parser.add_argument("--model", type=str, default="openai/gpt-oss-20b")
-    parser.add_argument("--max-batch-size", default=16, type=int)
-    parser.add_argument("--trust-remote-code", action="store_true")
-    args = parser.parse_args()
-
-    # Get the benchmark function
-    benchmark = get_benchmark(args.model, args.max_batch_size, args.trust_remote_code)
-    # Run performance benchmark
-    benchmark.run(print_data=True)

benchmarks/kernels/benchmark_silu_mul_fp8_quant.py

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

benchmarks/kernels/benchmark_trtllm_decode_attention.py

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

benchmarks/kernels/benchmark_trtllm_prefill_attention.py

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

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -177,18 +177,18 @@ def benchmark_config(
     def run():
         w8a8_block_matmul(A, B, As, Bs, block_size, config, out_dtype)
 
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
     # JIT complication & warmup
     for _ in range(5):
         run()
-    torch.accelerator.synchronize()
+    torch.cuda.synchronize()
 
     start_event = torch.Event(enable_timing=True)
     end_event = torch.Event(enable_timing=True)
 
     latencies: list[float] = []
     for i in range(num_iters):
-        torch.accelerator.synchronize()
+        torch.cuda.synchronize()
         start_event.record()
         run()
         end_event.record()
@@ -285,7 +285,7 @@ def tune_on_gpu(args_dict):
     weight_shapes = args_dict["weight_shapes"]
     args = args_dict["args"]
 
-    torch.accelerator.set_device_index(gpu_id)
+    torch.cuda.set_device(gpu_id)
     print(f"Starting tuning on GPU {gpu_id} with batch sizes {batch_sizes}")
 
     block_n = args.block_n
@@ -334,7 +334,7 @@ def distribute_batch_sizes(batch_sizes, num_gpus):
 
 def main(args):
     print(args)
-    num_gpus = torch.accelerator.device_count()
+    num_gpus = torch.cuda.device_count()
     if num_gpus == 0:
         raise RuntimeError("No GPU available for tuning")
     print(f"Found {num_gpus} GPUs for parallel tuning")

benchmarks/kernels/cpu/benchmark_cpu_attn.py

@@ -27,7 +27,7 @@ def get_attn_isa(
     else:
         if current_platform.get_cpu_architecture() == CpuArchEnum.ARM:
             return "neon"
-        elif torch.cpu._is_amx_tile_supported():
+        elif torch._C._cpu._is_amx_tile_supported():
             return "amx"
         else:
             return "vec"