Merge branch 'main' into zhuohan/remove-unnecessary-instance_id-setup

Signed-off-by: AuYang <459461160@qq.com>

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

@@ -13,9 +13,10 @@ 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
 
 
@@ -63,6 +64,9 @@ 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(
@@ -102,6 +106,8 @@ def test_lm_eval_correctness_param(config_filename, tp_size):
                 f"ground_truth={ground_truth:.3f} | "
                 f"measured={measured_value:.3f} | rtol={rtol}"
             )
-            success = success and np.isclose(ground_truth, measured_value, rtol=rtol)
+
+            min_acceptable = ground_truth * (1 - rtol)
+            success = success and measured_value >= min_acceptable
 
     assert success

.buildkite/performance-benchmarks/README.md

@@ -83,7 +83,6 @@ 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
@@ -33,6 +33,45 @@ 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
 # -----------------------------
@@ -52,19 +91,25 @@ 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)]
@@ -75,12 +120,25 @@ def compare_data_columns(
             "No common key columns found from info_cols across the input files."
         )
 
-    meta_added = False
+    union_index = None
+    metas: list[pd.DataFrame] = []
+    staged: list[tuple[str, pd.Series, pd.Series | None]] = []
 
     for file in files:
         df = pd.read_json(file, orient="records")
+        df = _normalize_concurrency_in_df(df, canonical="# of max concurrency.")
 
-        if drop_column in df.columns:
+        # 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:
             df = df.dropna(subset=[drop_column], ignore_index=True)
 
         for c in (
@@ -105,35 +163,61 @@ def compare_data_columns(
             meta = meta.groupby(level=key_cols, dropna=False).first()
 
         file_label = "/".join(file.split("/")[:-1]) or os.path.basename(file)
-        s = df_idx[data_column]
-        if not s.index.is_unique:
-            s = s.groupby(level=key_cols, dropna=False).mean()
+
+        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.name = file_label
 
-        if not meta_added:
-            frames.append(meta)
-            meta_added = True
-
+        name_s = None
         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)
 
-        frames.append(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)
         raw_data_cols.append(file_label)
-        compare_frames.append(s)
+        metric_series_aligned.append(s_aligned)
 
-        if len(compare_frames) >= 2:
-            base = compare_frames[0]
-            current = compare_frames[-1]
-            if "P99" in data_column or "Median" in data_column:
+        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):
                 ratio = base / current
             else:
                 ratio = current / base
             ratio = ratio.mask(base == 0)
-            ratio.name = f"Ratio 1 vs {len(compare_frames)}"
+            ratio.name = f"Ratio 1 vs {len(metric_series_aligned)}"
             frames.append(ratio)
 
     concat_df = pd.concat(frames, axis=1).reset_index(drop=True)
@@ -204,24 +288,10 @@ 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
+    df: pd.DataFrame,
+    threshold: float,
+    slack_pct: float = 0.0,
 ) -> pd.io.formats.style.Styler:
     conc_col = _find_concurrency_col(df)
     key_cols = [
@@ -234,12 +304,24 @@ def _highlight_threshold(
     ]
     conf_cols = [c for c in conf_cols if pd.api.types.is_numeric_dtype(df[c])]
 
-    return df.style.map(
-        lambda v: "background-color:#e6ffe6;font-weight:bold;"
-        if pd.notna(v) and v <= threshold
-        else "",
-        subset=conf_cols,
-    )
+    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)
 
 
 def highlight_ratio_columns(styler: pd.io.formats.style.Styler):
@@ -286,11 +368,30 @@ 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)
-    name = re.sub(r"[:\\/?*\[\]]", "_", name)
+
+    # Replace illegal characters with underscore.
+    trans = str.maketrans(
+        {
+            ":": "_",
+            "\\": "_",
+            "/": "_",
+            "?": "_",
+            "*": "_",
+            "[": "_",
+            "]": "_",
+        }
+    )
+    name = name.translate(trans)
+
+    # Strip quotes/spaces and collapse whitespace.
     name = name.strip().strip("'")
-    name = re.sub(r"\s+", " ", name)
+    name = " ".join(name.split())
+
     if not name:
         name = "sheet"
     return name[:31]
@@ -298,30 +399,57 @@ 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))
-    model = d.get("Model", "model")
-    model_short = str(model).split("/")[-1]
+
+    # Always keep input/output lengths (these are important).
     ilen = d.get("Input Len", "")
     olen = d.get("Output Len", "")
     lens = f"_{ilen}x{olen}" if ilen != "" and olen != "" else ""
+
+    # Shorten model name aggressively to make room for lens.
+    model = d.get("Model", "model")
+    leaf = str(model).split("/")[-1]
+
+    max_model_len = max(1, 31 - len(lens))
+    model_short = leaf[:max_model_len]
+
     return _sanitize_sheet_name(f"{model_short}{lens}")
 
 
 def _write_tables_to_excel_sheet(
     writer: pd.ExcelWriter, sheet: str, blocks: list[tuple[str, pd.DataFrame]]
 ):
-    startrow = 0
+    """Write all blocks to a sheet with a single to_excel() call.
+
+    Pandas+openpyxl can be extremely slow when called many times per sheet.
+    We flatten blocks into one table with a 'Section' column to keep structure
+    while making Excel generation fast and deterministic.
+    """
+    if not blocks:
+        pd.DataFrame().to_excel(writer, sheet_name=sheet, index=False)
+        return
+
+    combined_parts: list[pd.DataFrame] = []
     for title, df in blocks:
-        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
+        df2 = df.copy()
+        # Put the section label as the first column for readability.
+        df2.insert(0, "Section", title)
+        combined_parts.append(df2)
+
+    combined = pd.concat(combined_parts, axis=0, ignore_index=True, sort=False)
+    combined.to_excel(writer, sheet_name=sheet, index=False)
 
 
 def _safe_filename(s: str) -> str:
-    s = re.sub(r"[^\w\-.]+", "_", str(s).strip())
-    return s[:180] if len(s) > 180 else s
+    # 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
 
 
 # -----------------------------
@@ -428,7 +556,11 @@ 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
+    df: pd.DataFrame,
+    conc_col: str,
+    cfg_col: str,
+    threshold: float,
+    slack_pct: float = 0.0,
 ):
     if df is None or conc_col not in df.columns or cfg_col not in df.columns:
         return pd.NA
@@ -441,7 +573,14 @@ def _max_concurrency_ok(
     if d.empty:
         return pd.NA
 
-    ok = d[d[cfg_col] <= threshold]
+    # 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]
     if ok.empty:
         return pd.NA
 
@@ -507,15 +646,25 @@ 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)
+            _max_concurrency_ok(
+                ttft_group_df, conc_col, cfg, args.ttft_max_ms, args.ttft_slack_pct
+            )
             if ttft_group_df is not None
             else pd.NA
         )
         tpot_max = (
-            _max_concurrency_ok(tpot_group_df, conc_col, cfg, args.tpot_max_ms)
+            _max_concurrency_ok(
+                tpot_group_df, conc_col, cfg, args.tpot_max_ms, args.tpot_slack_pct
+            )
             if tpot_group_df is not None
             else pd.NA
         )
@@ -544,8 +693,8 @@ def build_valid_max_concurrency_summary_html(
         rows.append(
             {
                 "Configuration": cfg,
-                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} (TTFT ≤ {ttft_range})": ttft_max,
+                f"Max {conc_col} (TPOT ≤ {tpot_range})": tpot_max,
                 f"Max {conc_col} (Both)": both,
                 "Output Tput @ Both (tok/s)": tput_at_both,
                 "TTFT @ Both (ms)": ttft_at_both,
@@ -620,15 +769,24 @@ 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)
+            _max_concurrency_ok(
+                ttft_group_df, conc_col, cfg, args.ttft_max_ms, args.ttft_slack_pct
+            )
             if ttft_group_df is not None
             else pd.NA
         )
         tpot_max = (
-            _max_concurrency_ok(tpot_group_df, conc_col, cfg, args.tpot_max_ms)
+            _max_concurrency_ok(
+                tpot_group_df, conc_col, cfg, args.tpot_max_ms, args.tpot_slack_pct
+            )
             if tpot_group_df is not None
             else pd.NA
         )
@@ -657,8 +815,8 @@ def build_valid_max_concurrency_summary_df(
         rows.append(
             {
                 "Configuration": cfg,
-                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} (TTFT ≤ {ttft_range})": ttft_max,
+                f"Max {conc_col} (TPOT ≤ {tpot_range})": tpot_max,
                 f"Max {conc_col} (Both)": both,
                 "Output Tput @ Both (tok/s)": tput_at_both,
                 "TTFT @ Both (ms)": ttft_at_both,
@@ -751,7 +909,21 @@ def build_parser() -> argparse.ArgumentParser:
         help="Reference limit for TPOT plots (ms)",
     )
 
-    # ---- NEW: export options ----
+    # ---- SLA tolerance (slack) options ----
+    parser.add_argument(
+        "--ttft-slack-pct",
+        type=float,
+        default=5.0,
+        help="Allowed percentage above TTFT SLA (default: 5).",
+    )
+    parser.add_argument(
+        "--tpot-slack-pct",
+        type=float,
+        default=5.0,
+        help="Allowed percentage above TPOT SLA (default: 5).",
+    )
+
+    # ---- export options ----
     parser.add_argument(
         "--excel-out",
         type=str,
@@ -843,9 +1015,13 @@ def render_metric_table_html(
 
     metric_name = metric_label.lower()
     if "ttft" in metric_name:
-        styler = _highlight_threshold(display_group, args.ttft_max_ms)
+        styler = _highlight_threshold(
+            display_group, args.ttft_max_ms, args.ttft_slack_pct
+        )
     elif ("tpot" in metric_name) or ("median" in metric_name) or ("p99" in metric_name):
-        styler = _highlight_threshold(display_group, args.tpot_max_ms)
+        styler = _highlight_threshold(
+            display_group, args.tpot_max_ms, args.tpot_slack_pct
+        )
     else:
         styler = display_group.style
 
@@ -962,22 +1138,46 @@ def write_report_group_first(
         csv_dir.mkdir(parents=True, exist_ok=True)
 
     excel_path = args.excel_out or "perf_comparison.xlsx"
-    with pd.ExcelWriter(excel_path, engine="openpyxl") as xw:
+    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()
         # ---- Environment sheet (first) ----
         env_sheet = _sanitize_sheet_name("Environment")
         env_df = _load_env_df_for_inputs(args, files)
-        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)
+        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)
         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:
@@ -993,12 +1193,19 @@ 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
-                dedup_i = 1
-                while sheet in xw.sheets:
-                    dedup_i += 1
-                    sheet = _sanitize_sheet_name(f"{sheet_base}_{dedup_i}")
+                if do_excel:
+                    dedup_i = 1
+                    while sheet in used_sheets:
+                        dedup_i += 1
+                        suffix = f"_{dedup_i}"
+                        # Ensure uniqueness even when sheet names are truncated.
+                        base = str(sheet_base)
+                        keep = max(1, 31 - len(suffix))
+                        sheet = _sanitize_sheet_name(base[:keep] + suffix)
+                    used_sheets.add(sheet)
 
                 excel_blocks: list[tuple[str, pd.DataFrame]] = []
 
@@ -1059,7 +1266,7 @@ def write_report_group_first(
                         )
 
                         excel_blocks.append(
-                            (metric_label, display_group.reset_index(drop=True))
+                            (metric_label, group_df.reset_index(drop=True))
                         )
                         if csv_dir:
                             fn = _safe_filename(
@@ -1067,7 +1274,7 @@ def write_report_group_first(
                                     "/", "_"
                                 )
                             )
-                            display_group.to_csv(csv_dir / f"{fn}.csv", index=False)
+                            group_df.to_csv(csv_dir / f"{fn}.csv", index=False)
 
                     summary_html = build_valid_max_concurrency_summary_html(
                         tput_group_df=tput_group_df,
@@ -1097,9 +1304,13 @@ def write_report_group_first(
                             )
                             summary_df.to_csv(csv_dir / f"{fn}.csv", index=False)
 
-                _write_tables_to_excel_sheet(xw, sheet, excel_blocks)
+                if do_excel:
+                    _write_tables_to_excel_sheet(xw, sheet, excel_blocks)
 
-    print(f"Wrote Excel: {excel_path}")
+    if disable_excel:
+        print("Skipped Excel generation (VLLM_COMPARE_DISABLE_EXCEL=1).")
+    else:
+        print(f"Wrote Excel: {excel_path}")
     if csv_dir:
         print(f"Wrote CSVs under: {csv_dir}")
 

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

@@ -12,6 +12,13 @@ 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.
@@ -183,6 +190,304 @@ 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
   # $1: test type (latency or throughput)
@@ -347,10 +652,48 @@ run_serving_tests() {
     server_envs=$(echo "$params" | jq -r '.server_environment_variables')
     client_params=$(echo "$params" | jq -r '.client_parameters')
 
-    server_args=$(json2args "$server_params")
+    # 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 "$server_params_no_model")
+
     server_envs=$(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')
@@ -382,14 +725,13 @@ 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="$server_envs vllm serve \
+    server_command="$server_envs vllm serve $server_model \
       $server_args"
 
     # run the server
@@ -436,6 +778,14 @@ 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 \
@@ -444,8 +794,9 @@ 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 $client_remote_args "
+          $client_args_effective $client_remote_args "
 
         echo "Running test case $test_name with qps $qps"
         echo "Client command: $client_command"
@@ -467,6 +818,11 @@ run_serving_tests() {
         echo "$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
@@ -532,6 +888,7 @@ 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

@@ -0,0 +1,37 @@
+{
+  "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

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

@@ -72,17 +72,6 @@
         "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": {
@@ -105,17 +94,6 @@
         "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": {
@@ -139,14 +117,25 @@
       }
     },
     {
-      "test_name": "serving_llama8B_tp4_random_2048_128",
+      "test_name": "serving_llama8B_tp1_random_2048_2048",
       "server_parameters": {
-        "tensor_parallel_size": 4
+        "tensor_parallel_size": 1
       },
       "client_parameters": {
         "dataset_name": "random",
         "random-input-len": 2048,
-        "random-output-len": 128
+        "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
       }
     }
   ]

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

@@ -10,7 +10,6 @@
         "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,
@@ -37,7 +36,6 @@
         "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,
@@ -64,7 +62,6 @@
         "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,
@@ -91,7 +88,6 @@
         "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

@@ -5,7 +5,6 @@
         "server_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",
             "tensor_parallel_size": 1,
-            "swap_space": 16,
             "disable_log_stats": "",
             "load_format": "dummy"
         },
@@ -23,7 +22,6 @@
         "server_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
             "tensor_parallel_size": 4,
-            "swap_space": 16,
             "disable_log_stats": "",
             "load_format": "dummy"
         },
@@ -41,7 +39,6 @@
         "server_parameters": {
             "model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
             "tensor_parallel_size": 2,
-            "swap_space": 16,
             "disable_log_stats": "",
             "load_format": "dummy"
         },
@@ -59,7 +56,6 @@
         "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/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/offline_inference/basic/generate.py --model facebook/opt-125m"
+    python3 examples/basic/offline_inference/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 "$container_id" bash -c "
     export TORCH_COMPILE_DISABLE=1
     set -xve
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m" >> "$HOME"/test_basic.log
+    python3 examples/basic/offline_inference/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/offline_inference/basic/generate.py --model meta-llama/Llama-3.2-1B
+    python3 examples/basic/offline_inference/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/offline_inference/basic/generate.py --model facebook/opt-125m
+  cd vllm; timeout 120s python -u examples/basic/offline_inference/generate.py --model facebook/opt-125m
 '
 
 EXITCODE=$?

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

@@ -34,15 +34,15 @@ docker run \
     set -e
     echo $ZE_AFFINITY_MASK
     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
+    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
+    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
     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

.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")
+  PLATFORM_ARGS=("--no-async-scheduling" "--attention-backend=TRITON_ATTN")
   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

@@ -0,0 +1,248 @@
+#!/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/test-amd.yaml

@@ -42,6 +42,7 @@ steps:
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi325_1
   grade: Blocking
+  optional: true
   soft_fail: true
   source_file_dependencies:
   - requirements/nightly_torch_test.txt
@@ -67,6 +68,7 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   source_file_dependencies:
   - vllm/
@@ -97,6 +99,7 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   source_file_dependencies:
   - tests/standalone_tests/python_only_compile.sh
@@ -140,6 +143,7 @@ steps:
   timeout_in_minutes: 40
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   fast_check: true
@@ -503,6 +507,7 @@ steps:
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi325_1
   grade: Blocking
+  optional: true
   source_file_dependencies:
     - vllm/
     - tests/v1
@@ -520,6 +525,7 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/examples"
   source_file_dependencies:
@@ -529,12 +535,12 @@ steps:
   commands:
     - pip install tensorizer # for tensorizer test
     # for basic
-    - python3 offline_inference/basic/chat.py
-    - 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
+    - python3 basic/offline_inference/chat.py --attention-backend TRITON_ATTN
+    - 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
     # for multi-modal models
     - python3 offline_inference/audio_language.py --seed 0
     - python3 offline_inference/vision_language.py --seed 0
@@ -823,6 +829,7 @@ steps:
   timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   source_file_dependencies:
   - csrc/
@@ -936,6 +943,7 @@ steps:
   timeout_in_minutes: 25
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
@@ -1046,6 +1054,7 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
@@ -1059,6 +1068,7 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   source_file_dependencies:
   - vllm/
@@ -1072,6 +1082,7 @@ steps:
   timeout_in_minutes: 100
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
@@ -1090,6 +1101,7 @@ steps:
   timeout_in_minutes: 10
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
@@ -1169,7 +1181,7 @@ 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 -k 'not (Gemma3 or Qwen2VL or Qwen2_5_VL)'
-    - python3 examples/offline_inference/basic/chat.py
+    - python3 examples/basic/offline_inference/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
@@ -1355,6 +1367,7 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_2
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
@@ -1393,6 +1406,7 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
+  optional: true
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -1410,6 +1424,7 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
+  optional: true
   # grade: Blocking
   num_gpus: 4
   source_file_dependencies:
@@ -1461,6 +1476,7 @@ steps:
 - label: NixlConnector PD accuracy tests (Distributed) # 30min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
+  optional: true
   # grade: Blocking
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
@@ -1475,6 +1491,7 @@ steps:
 - label: DP EP NixlConnector PD accuracy tests (Distributed) # 15min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
+  optional: true
   # grade: Blocking
   timeout_in_minutes: 15
   working_dir: "/vllm-workspace/tests"
@@ -1486,6 +1503,20 @@ steps:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
     - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
+- label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs)
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_4
+  # grade: Blocking
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
+    - CROSS_LAYERS_BLOCKS=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
 ##### multi gpus test #####
 ##### A100 test #####
 
@@ -1625,8 +1656,8 @@ steps:
   - vllm/model_executor/layers/quantization/mxfp4.py
   - vllm/v1/attention/backends/flashinfer.py
   commands:
-    - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - VLLM_ROCM_USE_AITER_MHA=0 VLLM_ROCM_USE_AITER=1 VLLM_USE_AITER_UNIFIED_ATTENTION=1 pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
+  - uv pip install --system 'gpt-oss[eval]==0.0.5'
+  - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-gfx942.txt
 
 ##### EPLB Accuracy Tests #####
 - label: DeepSeek V2-Lite Accuracy
@@ -1765,6 +1796,7 @@ steps:
   # in /vllm/tools/pre_commit/generate_nightly_torch_test.py
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
+  optional: true
   soft_fail: true
   source_file_dependencies:
   - requirements/nightly_torch_test.txt
@@ -1775,6 +1807,7 @@ steps:
   timeout_in_minutes: 15
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - vllm/
   - tests/multimodal
@@ -1787,6 +1820,7 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - vllm/
   - tests/test_inputs.py
@@ -1816,6 +1850,7 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - tests/standalone_tests/python_only_compile.sh
   - setup.py
@@ -1826,6 +1861,7 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   fast_check: true
   torch_nightly: true
   source_file_dependencies:
@@ -1856,6 +1892,7 @@ steps:
   timeout_in_minutes: 40
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   fast_check: true
   torch_nightly: true
@@ -1873,6 +1910,7 @@ steps:
   timeout_in_minutes: 130
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   fast_check: true
   torch_nightly: true
@@ -1889,6 +1927,7 @@ steps:
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   fast_check: true
   torch_nightly: true
@@ -1907,6 +1946,7 @@ steps:
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   fast_check: true
   torch_nightly: true
@@ -1921,6 +1961,7 @@ steps:
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   fast_check: true
   torch_nightly: true
@@ -1999,6 +2040,7 @@ steps:
   timeout_in_minutes: 10
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_8
+  optional: true
   gpu: h100
   num_gpus: 8
   working_dir: "/vllm-workspace/tests"
@@ -2019,6 +2061,7 @@ steps:
 - label: EPLB Algorithm Test # 5min
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
+  optional: true
   timeout_in_minutes: 15
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -2030,6 +2073,7 @@ steps:
 - label: EPLB Execution Test # 10min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
+  optional: true
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -2044,6 +2088,7 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_2
+  optional: true
   num_gpus: 2
   source_file_dependencies:
   - vllm/
@@ -2085,12 +2130,13 @@ steps:
   commands:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
 
+
 - label: V1 Test e2e + engine # 65min
   timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental]
-  # The test uses 4 GPUs, but we schedule it on 8-GPU machines for stability.
-  # See discussion here: https://github.com/vllm-project/vllm/pull/31040
-  agent_pool: mi355_8
+  agent_pool: mi355_1
+  optional: true
+  # grade: Blocking
   source_file_dependencies:
     - vllm/
     - tests/v1
@@ -2100,10 +2146,39 @@ steps:
     - pytest -v -s v1/e2e
     - pytest -v -s v1/engine
 
+- label: V1 Test e2e (2 GPUs) # 65min
+  timeout_in_minutes: 90
+  mirror_hardwares: [amdexperimental]
+  agent_pool: mi355_2
+  optional: true
+  # grade: Blocking
+  source_file_dependencies:
+    - vllm/
+    - tests/v1
+  commands:
+    # Only run tests that need exactly 2 GPUs
+    - pytest -v -s v1/e2e/test_spec_decode.py -k "tensor_parallelism"
+
+- label: V1 Test e2e (4 GPUs) # 65min
+  timeout_in_minutes: 90
+  mirror_hardwares: [amdexperimental]
+  # The test uses 4 GPUs, but we schedule it on 8-GPU machines for stability.
+  # See discussion here: https://github.com/vllm-project/vllm/pull/31040
+  agent_pool: mi355_4
+  optional: true
+  # grade: Blocking
+  source_file_dependencies:
+    - vllm/
+    - tests/v1
+  commands:
+    # Only run tests that need 4 GPUs
+    - pytest -v -s v1/e2e/test_spec_decode.py -k "eagle_correctness_heavy"
+
 - label: V1 Test entrypoints # 35min
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
     - vllm/
     - tests/v1
@@ -2114,6 +2189,7 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
     - vllm/
     - tests/v1
@@ -2136,7 +2212,19 @@ steps:
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 
-# TODO: Add the "V1 Test attention (MI300)" test group
+- label: V1 Test attention (H100) # 10min
+  mirror_hardwares: [amdexperimental]
+  agent_pool: mi355_1
+  optional: true
+  timeout_in_minutes: 30
+  gpu: h100
+  source_file_dependencies:
+    - vllm/config/attention.py
+    - vllm/model_executor/layers/attention
+    - vllm/v1/attention
+    - tests/v1/attention
+  commands:
+    - pytest -v -s v1/attention
 
 - label: Batch Invariance Tests (H100) # 10min
   mirror_hardwares: [amdexperimental]
@@ -2186,6 +2274,7 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   working_dir: "/vllm-workspace/examples"
   source_file_dependencies:
   - vllm/entrypoints
@@ -2194,12 +2283,12 @@ steps:
   commands:
     - pip install tensorizer # for tensorizer test
     # for basic
-    - python3 offline_inference/basic/chat.py
-    - 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
+    - python3 basic/offline_inference/chat.py --attention-backend TRITON_ATTN
+    - 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
     # for multi-modal models
     - python3 offline_inference/audio_language.py --seed 0
     - python3 offline_inference/vision_language.py --seed 0
@@ -2220,6 +2309,7 @@ steps:
   timeout_in_minutes: 15
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - vllm/
   - tests/cuda
@@ -2231,6 +2321,7 @@ steps:
   timeout_in_minutes: 75
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - vllm/model_executor/layers
   - vllm/sampling_metadata.py
@@ -2263,6 +2354,7 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   torch_nightly: true
   source_file_dependencies:
     - vllm/
@@ -2279,6 +2371,7 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2294,6 +2387,7 @@ steps:
   timeout_in_minutes: 40
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
@@ -2311,6 +2405,7 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - tests/v1/cudagraph
   - vllm/v1/cudagraph_dispatcher.py
@@ -2324,6 +2419,7 @@ steps:
   timeout_in_minutes: 75
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - csrc/
   - tests/kernels/core
@@ -2335,6 +2431,7 @@ steps:
   timeout_in_minutes: 35
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - csrc/attention/
   - vllm/v1/attention
@@ -2349,6 +2446,7 @@ steps:
   timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - csrc/quantization/
   - vllm/model_executor/layers/quantization
@@ -2361,6 +2459,7 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - csrc/quantization/cutlass_w8a8/moe/
   - csrc/moe/
@@ -2377,6 +2476,7 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - csrc/mamba/
   - tests/kernels/mamba
@@ -2408,6 +2508,7 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - vllm/utils/import_utils.py
   - tests/kernels/helion/
@@ -2420,6 +2521,7 @@ steps:
   torch_nightly: true
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - vllm/engine/arg_utils.py
   - vllm/config/model.py
@@ -2436,6 +2538,7 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   working_dir: "/vllm-workspace/.buildkite"
   source_file_dependencies:
   - benchmarks/
@@ -2446,6 +2549,7 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - vllm/
   - tests/benchmarks/
@@ -2456,6 +2560,7 @@ steps:
   timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
@@ -2476,6 +2581,7 @@ steps:
   timeout_in_minutes: 75
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
@@ -2487,6 +2593,7 @@ steps:
   timeout_in_minutes: 15
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - csrc/
   - vllm/entrypoints/openai/
@@ -2503,6 +2610,7 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2515,6 +2623,7 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   torch_nightly: true
   source_file_dependencies:
   - vllm/model_executor/models/
@@ -2534,6 +2643,7 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2546,6 +2656,7 @@ steps:
 - label: Basic Models Test (Other CPU) # 5min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   timeout_in_minutes: 10
   torch_nightly: true
   source_file_dependencies:
@@ -2560,6 +2671,7 @@ steps:
   timeout_in_minutes: 25
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2573,6 +2685,7 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   torch_nightly: true
   source_file_dependencies:
   - vllm/model_executor/models/
@@ -2593,6 +2706,7 @@ steps:
   timeout_in_minutes: 75
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2662,6 +2776,7 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
@@ -2674,6 +2789,7 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
@@ -2685,6 +2801,7 @@ steps:
   timeout_in_minutes: 100
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_1
+  optional: true
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -2702,6 +2819,7 @@ steps:
   timeout_in_minutes: 10
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
   - vllm/multimodal/
@@ -2758,6 +2876,7 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_1
+  optional: true
   source_file_dependencies:
   - vllm/model_executor/layers/quantization
   - tests/models/quantization
@@ -2775,7 +2894,7 @@ 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 -k 'not (Gemma3 or Qwen2VL or Qwen2_5_VL)'
-    - python3 examples/offline_inference/basic/chat.py
+    - python3 examples/basic/offline_inference/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
@@ -2801,8 +2920,8 @@ steps:
   - vllm/v1/attention/selector.py
   - vllm/platforms/cuda.py
   commands:
-    rocm-smi
-    - python3 examples/offline_inference/basic/chat.py
+    - rocm-smi
+    - python3 examples/basic/offline_inference/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 
@@ -2909,6 +3028,7 @@ steps:
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_2
+  optional: true
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -2991,6 +3111,7 @@ steps:
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
   agent_pool: mi355_2
+  optional: true
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -3012,6 +3133,7 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_2
+  optional: true
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -3049,6 +3171,7 @@ steps:
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
+  optional: true
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -3065,6 +3188,7 @@ steps:
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
+  optional: true
   num_gpus: 4
   source_file_dependencies:
   - vllm/lora
@@ -3113,6 +3237,7 @@ steps:
 - label: NixlConnector PD accuracy tests (Distributed) # 30min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
+  optional: true
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -3126,6 +3251,7 @@ steps:
 - label: DP EP NixlConnector PD accuracy tests (Distributed) # 15min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi355_4
+  optional: true
   timeout_in_minutes: 15
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -3136,6 +3262,20 @@ steps:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
     - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
+- label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs)
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi355_4
+  # grade: Blocking
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
+    - CROSS_LAYERS_BLOCKS=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
 ##### multi gpus test #####
 ##### A100 test #####
 
@@ -3250,6 +3390,7 @@ steps:
 - label: ROCm LM Eval Large Models (8 Card)
   mirror_hardwares: [amdproduction]
   agent_pool: mi355_8
+  optional: true
   num_gpus: 8
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   commands:
@@ -3268,8 +3409,8 @@ steps:
   - vllm/model_executor/layers/quantization/mxfp4.py
   - vllm/v1/attention/backends/flashinfer.py
   commands:
-    - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - VLLM_ROCM_USE_AITER_MHA=0 VLLM_ROCM_USE_AITER=1 VLLM_USE_AITER_UNIFIED_ATTENTION=1 pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
+  - uv pip install --system 'gpt-oss[eval]==0.0.5'
+  - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-gfx950.txt
 
 ##### EPLB Accuracy Tests #####
 - label: DeepSeek V2-Lite Accuracy
@@ -3283,7 +3424,7 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
 
-- label: Qwen3-30B-A3B-FP8-block Accuracy (B200/MI355)
+- label: Qwen3-30B-A3B-FP8-block Accuracy (B200-MI355)
   mirror_hardwares: [amdexperimental, amdproduction, amdmi355]
   agent_pool: mi355_2
   timeout_in_minutes: 60
@@ -3305,7 +3446,7 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh 0.8 1319 8040
 
-- label: Attention Benchmarks Smoke Test (B200/MI355)
+- label: Attention Benchmarks Smoke Test (B200-MI355)
   device: b200
   mirror_hardwares: [amdexperimental, amdmi355]
   agent_pool: mi355_2

.buildkite/test_areas/basic_correctness.yaml

@@ -14,8 +14,3 @@ 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

@@ -101,8 +101,8 @@ 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 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"
+    # 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)"
 
 - label: Fusion E2E Config Sweep (H100)
   timeout_in_minutes: 30
@@ -132,9 +132,9 @@ steps:
   commands:
     - nvidia-smi
     # Run all models but only FLASHINFER, Inductor partition and native custom ops
-    # Qwen requires +quant_fp8 as -quant_fp8 rms+quant fusion is not supported
+    # Qwen/Deepseek 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 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 deepseek)) or llama-3)"
 
 - label: Fusion E2E TP2 Quick (H100)
   timeout_in_minutes: 20
@@ -150,8 +150,8 @@ 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"
-    - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "inductor_partition and not +rms_norm and not +quant_fp8"
+    - 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))"
 
 - label: Fusion E2E TP2 AR-RMS Config Sweep (H100)
   timeout_in_minutes: 40
@@ -205,7 +205,7 @@ steps:
   commands:
     - nvidia-smi
     # Run all models but only FLASHINFER, Inductor partition and native custom ops
-    # include qwen with +quant_fp8 as -quant_fp8 rms+quant fusion is not supported
+    # include qwen/deepseek with +quant_fp8 as -quant_fp8 rms+quant fusion is not supported
     # for ar-rms-quant-fp4, also sweep llama3
-    - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "(FLASHINFER and inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and qwen3)) or 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)"
+    - 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))"

.buildkite/test_areas/distributed.yaml

@@ -149,7 +149,7 @@ steps:
   num_devices: 2
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
-    # - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py --- failing, need to re-enable
+    - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py
     - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
     - pytest -v -s tests/v1/distributed/test_dbo.py
 

.buildkite/test_areas/entrypoints.yaml

@@ -24,11 +24,6 @@ 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
@@ -60,11 +55,6 @@ steps:
   - pytest -v -s entrypoints/instrumentator
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
   - pytest -v -s tool_use
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
 
 - label: Entrypoints Integration (Pooling)
   timeout_in_minutes: 50
@@ -75,11 +65,6 @@ steps:
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/pooling
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
 
 - label: Entrypoints Integration (Responses API)
   timeout_in_minutes: 50

.buildkite/test_areas/kernels.yaml

@@ -8,8 +8,9 @@ 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
+    - pytest -v -s kernels/core kernels/test_top_k_per_row.py kernels/test_concat_mla_q.py
 
 - label: Kernels Attention Test %N
   timeout_in_minutes: 35
@@ -96,7 +97,7 @@ steps:
   - vllm/platforms/cuda.py
   commands:
     - nvidia-smi
-    - python3 examples/offline_inference/basic/chat.py
+    - python3 examples/basic/offline_inference/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/misc.yaml

@@ -67,12 +67,13 @@ steps:
   - examples/
   commands:
     - pip install tensorizer # for tensorizer test
-    - 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 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
     # for multi-modal models
     - python3 offline_inference/audio_language.py --seed 0
     - python3 offline_inference/vision_language.py --seed 0
@@ -87,11 +88,6 @@ steps:
     - python3 offline_inference/spec_decode.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
     # https://github.com/vllm-project/vllm/pull/26682 uses slightly more memory in PyTorch 2.9+ causing this test to OOM in 1xL4 GPU
     - python3 offline_inference/spec_decode.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
 
 - label: Metrics, Tracing (2 GPUs)
   timeout_in_minutes: 20

.buildkite/test_areas/model_runner_v2.yaml

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

@@ -39,8 +39,3 @@ steps:
   - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
   - pytest -v -s models/test_oot_registration.py # it needs a clean process
   - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
-  mirror:
-    amd:
-      device: mi325_2
-      depends_on:
-      - image-build-amd

.github/mergify.yml

@@ -38,15 +38,13 @@ pull_request_rules:
 
         > [!TIP]
         > <details>
-        > <summary>Is <code>mypy</code> or <code>markdownlint</code> failing?</summary>
+        > <summary>Is <code>mypy</code> failing?</summary>
         > <br/>
-        > <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:
+        > <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:
         >
         > ```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>
 

.pre-commit-config.yaml

@@ -24,12 +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/igorshubovych/markdownlint-cli
-  rev: v0.45.0
+- repo: https://github.com/DavidAnson/markdownlint-cli2
+  rev: v0.21.0
   hooks:
-  - id: markdownlint
-    exclude: '.*\.inc\.md'
-    stages: [manual] # Only run in CI
+  - id: markdownlint-cli2
+    language_version: lts
+    args: [--fix]
 - repo: https://github.com/rhysd/actionlint
   rev: v1.7.7
   hooks:
@@ -55,7 +55,7 @@ repos:
       language: python
       types_or: [python, pyi]
       require_serial: true
-      additional_dependencies: ["mypy[faster-cache]==1.15.0", regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
+      additional_dependencies: ["mypy[faster-cache]==1.19.1", 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"

.readthedocs.yaml

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

CMakeLists.txt

@@ -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;gfx1200;gfx1201;gfx1150;gfx1151")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201")
 
 # ROCm installation prefix. Default to /opt/rocm but allow override via
 # -DROCM_PATH=/your/rocm/path when invoking cmake.

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/mla_runner.py

@@ -145,7 +145,6 @@ def create_minimal_vllm_config(
     cache_config = CacheConfig(
         block_size=block_size,
         gpu_memory_utilization=0.9,
-        swap_space=0,
         cache_dtype="auto",
         enable_prefix_caching=False,
     )

benchmarks/attention_benchmarks/runner.py

@@ -141,7 +141,6 @@ def _create_vllm_config(
     cache_config = CacheConfig(
         block_size=config.block_size,
         cache_dtype="auto",
-        swap_space=0,
     )
     cache_config.num_gpu_blocks = max_num_blocks
     cache_config.num_cpu_blocks = 0

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/kernels/bench_concat_mla_q.py

@@ -0,0 +1,98 @@
+# 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

@@ -0,0 +1,153 @@
+# 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_moe.py

@@ -626,7 +626,11 @@ class BenchmarkWorker:
             if visible_device != f"{self.device_id}":
                 need_device_guard = True
 
-        with torch.cuda.device(self.device_id) if need_device_guard else nullcontext():
+        with (
+            torch.accelerator.device_index(self.device_id)
+            if need_device_guard
+            else nullcontext()
+        ):
             for idx, config in enumerate(tqdm(search_space)):
                 try:
                     kernel_time = benchmark_config(

cmake/cpu_extension.cmake

@@ -79,7 +79,8 @@ else()
     find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
     find_isa(${CPUINFO} "bf16" ARM_BF16_FOUND) # Check for ARM BF16 support
     find_isa(${CPUINFO} "S390" S390_FOUND)
-    find_isa(${CPUINFO} "v" RVV_FOUND) # Check for RISC-V RVV support
+    find_isa(${CPUINFO} "zvfhmin" RVV_FP16_FOUND) # Check for RISC-V Vector FP16 support
+    find_isa(${CPUINFO} "zvfbfmin" RVV_BF16_FOUND) # Check for RISC-V Vector BF16 support
 
     # Support cross-compilation by allowing override via environment variables
     if (ENABLE_ARM_BF16)
@@ -142,11 +143,19 @@ elseif (S390_FOUND)
         "-march=native"
         "-mtune=native")
 elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
-    if(RVV_FOUND)
-	    message(FAIL_ERROR "Can't support rvv now.")
+    message(STATUS "RISC-V detected")
+    if(RVV_BF16_FOUND)
+        message(STATUS "BF16 extension detected")
+        set(MARCH_FLAGS -march=rv64gcv_zvfh_zfbfmin_zvfbfmin_zvl128b -mrvv-vector-bits=zvl -mabi=lp64d)
+        add_compile_definitions(RISCV_BF16_SUPPORT)
+    elseif (RVV_FP16_FOUND)
+        message(WARNING "BF16 functionality is not available")
+        set(MARCH_FLAGS -march=rv64gcv_zvfh_zvl128b -mrvv-vector-bits=zvl -mabi=lp64d)
     else()
+        message(STATUS "compile riscv with scalar")
         list(APPEND CXX_COMPILE_FLAGS "-march=rv64gc")
     endif()
+    list(APPEND CXX_COMPILE_FLAGS ${MARCH_FLAGS})
 else()
     message(FATAL_ERROR "vLLM CPU backend requires X86, Power9+ ISA, S390X ISA, ARMv8 or RISC-V support.")
 endif()

csrc/cache.h

@@ -74,6 +74,12 @@ void indexer_k_quant_and_cache(
     int64_t quant_block_size,     // quantization block size
     const std::string& scale_fmt);
 
+// Concatenate query nope and rope for MLA/DSA attention
+void concat_mla_q(
+    torch::Tensor& ql_nope,  // [num_tokens, num_heads, nope_dim]
+    torch::Tensor& q_pe,     // [num_tokens, num_heads, rope_dim]
+    torch::Tensor& q_out);   // [num_tokens, num_heads, nope_dim + rope_dim]
+
 // Extract function to gather quantized K cache
 void cp_gather_indexer_k_quant_cache(
     const torch::Tensor& kv_cache,  // [num_blocks, block_size, cache_stride]

csrc/cache_kernels.cu

@@ -8,6 +8,7 @@
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
 #include "quantization/vectorization_utils.cuh"
+#include "concat_mla_q.cuh"
 
 #ifdef USE_ROCM
   #include "quantization/w8a8/fp8/amd/quant_utils.cuh"
@@ -918,8 +919,8 @@ __global__ void gather_and_maybe_dequant_cache(
 // SCALAR_T is the data type of the destination tensor.
 // CACHE_T is the stored data type of kv-cache.
 // KV_DTYPE is the real data type of kv-cache.
-#define CALL_GATHER_CACHE(SCALAR_T, CACHE_T, KV_DTYPE)                        \
-  vllm::gather_and_maybe_dequant_cache<SCALAR_T, CACHE_T, KV_DTYPE, 576,      \
+#define CALL_GATHER_CACHE(SCALAR_T, CACHE_T, KV_DTYPE, ENTRY_SZ)              \
+  vllm::gather_and_maybe_dequant_cache<SCALAR_T, CACHE_T, KV_DTYPE, ENTRY_SZ, \
                                        thread_block_size>                     \
       <<<grid, block, 0, stream>>>(                                           \
           reinterpret_cast<CACHE_T*>(src_cache.data_ptr()),                   \
@@ -930,6 +931,12 @@ __global__ void gather_and_maybe_dequant_cache(
           dst_entry_stride, reinterpret_cast<const float*>(scale.data_ptr()), \
           seq_starts_ptr);
 
+#define CALL_GATHER_CACHE_576(SCALAR_T, CACHE_T, KV_DTYPE) \
+  CALL_GATHER_CACHE(SCALAR_T, CACHE_T, KV_DTYPE, 576)
+
+#define CALL_GATHER_CACHE_320(SCALAR_T, CACHE_T, KV_DTYPE) \
+  CALL_GATHER_CACHE(SCALAR_T, CACHE_T, KV_DTYPE, 320)
+
 // Gather sequences from the cache into the destination tensor.
 //  - cu_seq_lens contains the cumulative sequence lengths for each batch
 //  - block_table contains the cache block indices for each sequence
@@ -959,9 +966,10 @@ void gather_and_maybe_dequant_cache(
     TORCH_CHECK(seq_starts.value().dtype() == torch::kInt32,
                 "seq_starts must be int32");
   }
-  TORCH_CHECK(head_dim == 576,
-              "gather_and_maybe_dequant_cache only support the head_dim to 576 "
-              "for better performance")
+  TORCH_CHECK(
+      head_dim == 320 || head_dim == 576,
+      "gather_and_maybe_dequant_cache only support the head_dim to 320 or 576 "
+      "for better performance")
 
   TORCH_CHECK(src_cache.device() == dst.device(),
               "src_cache and dst must be on the same device");
@@ -986,7 +994,13 @@ void gather_and_maybe_dequant_cache(
   const int32_t* seq_starts_ptr =
       seq_starts.has_value() ? seq_starts.value().data_ptr<int32_t>() : nullptr;
 
-  DISPATCH_BY_KV_CACHE_DTYPE(dst.dtype(), kv_cache_dtype, CALL_GATHER_CACHE);
+  if (head_dim == 576) {
+    DISPATCH_BY_KV_CACHE_DTYPE(dst.dtype(), kv_cache_dtype,
+                               CALL_GATHER_CACHE_576);
+  } else {
+    DISPATCH_BY_KV_CACHE_DTYPE(dst.dtype(), kv_cache_dtype,
+                               CALL_GATHER_CACHE_320);
+  }
 }
 
 namespace vllm {
@@ -995,75 +1009,67 @@ namespace vllm {
 // Similar to cp_gather_cache but specifically for FP8->BF16 conversion
 __global__ void cp_gather_and_upconvert_fp8_kv_cache(
     const uint8_t* __restrict__ src_cache,    // [NUM_BLOCKS, BLOCK_SIZE, 656]
-    __nv_bfloat16* __restrict__ dst,          // [TOT_TOKENS, 576]
-    const int32_t* __restrict__ block_table,  // [BATCH, BLOCK_INDICES]
-    const int32_t* __restrict__ seq_lens,     // [BATCH]
-    const int32_t* __restrict__ workspace_starts,  // [BATCH]
-    const int32_t block_size, const int32_t head_dim,
-    const int64_t block_table_stride, const int64_t cache_block_stride,
-    const int64_t cache_entry_stride, const int64_t dst_entry_stride) {
-  const int64_t bid = blockIdx.x;  // Batch ID
-  const int32_t num_splits = gridDim.y;
-  const int32_t split = blockIdx.y;
-  const int32_t seq_start = workspace_starts[bid];
-  const int32_t seq_len = seq_lens[bid];
-  const int32_t tot_slots = seq_len;
-  const int32_t split_slots = cuda_utils::ceil_div(tot_slots, num_splits);
+    __nv_bfloat16* __restrict__ dst,          // [total_tokens, 576]
+    const int32_t* __restrict__ block_table,  // [num_reqs, BLOCK_INDICES]
+    const int32_t* __restrict__ workspace_starts,  // [num_reqs]
+    const int32_t num_reqs, const int32_t block_size,
+    const int32_t total_tokens, const int64_t block_table_stride,
+    const int64_t cache_block_stride, const int64_t cache_entry_stride,
+    const int64_t dst_entry_stride) {
+  const int flat_warp_id = (blockIdx.x * blockDim.x + threadIdx.x) >> 5;
+  if (flat_warp_id >= total_tokens) return;
+  const int lane_id = threadIdx.x & 31;
 
-  const int32_t split_start = split * split_slots;
-  const int32_t split_end = min((split + 1) * split_slots, tot_slots);
-
-  const bool is_active_split = (split_start < tot_slots);
-
-  if (!is_active_split) return;
-
-  // Adjust the pointer for the block_table for this batch
-  const int32_t batch_offset = bid * block_table_stride;
-  int32_t offset = split_start;
-  int32_t offset_div = offset / block_size;
-  offset = offset % block_size;
-  const int32_t* batch_block_table = block_table + batch_offset;
-
-  // Adjust dst pointer based on the cumulative sequence lengths
-  dst += seq_start * dst_entry_stride;
-
-  const int tid = threadIdx.x;
-
-  // Process each token in this split
-  for (int pid = split_start; pid < split_end; ++pid) {
-    auto block_id = batch_block_table[offset_div];
-    const uint8_t* token_ptr =
-        src_cache + block_id * cache_block_stride + offset * cache_entry_stride;
-    __nv_bfloat16* dst_ptr = dst + pid * dst_entry_stride;
-
-    // FP8 format: 512 bytes fp8 + 16 bytes scales + 128 bytes rope (64 bf16)
-    const uint8_t* no_pe_ptr = token_ptr;
-    const float* scales_ptr = reinterpret_cast<const float*>(token_ptr + 512);
-    const __nv_bfloat16* rope_ptr =
-        reinterpret_cast<const __nv_bfloat16*>(token_ptr + 512 + 16);
-
-    // Parallelize fp8 dequant (512 elements) and rope copy (64 elements)
-    if (tid < 512) {
-      // FP8 dequantization
-      const int tile = tid >> 7;  // each tile is 128 elements
-      const float scale = scales_ptr[tile];
-      const uint8_t val = no_pe_ptr[tid];
-      dst_ptr[tid] =
-          fp8::scaled_convert<__nv_bfloat16, uint8_t,
-                              vllm::Fp8KVCacheDataType::kFp8E4M3>(val, scale);
-    } else if (tid < 576) {
-      // Rope copy (64 bf16 elements)
-      const int rope_idx = tid - 512;
-      dst_ptr[512 + rope_idx] = rope_ptr[rope_idx];
-    }
-
-    // Move to next token
-    offset += 1;
-    if (offset == block_size) {
-      offset_div += 1;
-      offset = 0;
-    }
+  // Binary search to find which request owns this output token
+  int lo = 0, hi = num_reqs - 1;
+  while (lo < hi) {
+    int mid = (lo + hi + 1) >> 1;
+    if (workspace_starts[mid] <= flat_warp_id)
+      lo = mid;
+    else
+      hi = mid - 1;
   }
+  const int req_id = lo;
+
+  // Compute physical token address via block table
+  const int out_token_id = flat_warp_id;
+  const int token_offset = out_token_id - workspace_starts[req_id];
+  const int cache_block_idx = token_offset / block_size;
+  const int offset_in_block = token_offset % block_size;
+  const int physical_block =
+      block_table[req_id * block_table_stride + cache_block_idx];
+
+  const uint8_t* token_ptr = src_cache + physical_block * cache_block_stride +
+                             offset_in_block * cache_entry_stride;
+
+  const int4* nope_src = reinterpret_cast<const int4*>(token_ptr);
+  const int4 fp8_data = nope_src[lane_id];
+
+  const float* scales_ptr = reinterpret_cast<const float*>(token_ptr + 512);
+  const float scale = scales_ptr[lane_id >> 3];
+
+  const uint2 fp8_lo = make_uint2(fp8_data.x, fp8_data.y);
+  const uint2 fp8_hi = make_uint2(fp8_data.z, fp8_data.w);
+#ifdef USE_ROCM
+  const bf16_8_t bf16_lo =
+      fp8::scaled_vec_conversion<bf16_8_t, uint2>(fp8_lo, scale);
+  const bf16_8_t bf16_hi =
+      fp8::scaled_vec_conversion<bf16_8_t, uint2>(fp8_hi, scale);
+#else
+  const bf16_8_t bf16_lo =
+      fp8::scaled_vec_conversion<bf16_8_t, uint2>(fp8_lo, scale, __NV_E4M3);
+  const bf16_8_t bf16_hi =
+      fp8::scaled_vec_conversion<bf16_8_t, uint2>(fp8_hi, scale, __NV_E4M3);
+#endif
+
+  __nv_bfloat16* dst_ptr = dst + out_token_id * dst_entry_stride;
+  int4* nope_dst = reinterpret_cast<int4*>(dst_ptr) + lane_id * 2;
+  nope_dst[0] = *reinterpret_cast<const int4*>(&bf16_lo);
+  nope_dst[1] = *reinterpret_cast<const int4*>(&bf16_hi);
+
+  const int* rope_src = reinterpret_cast<const int*>(token_ptr + 528);
+  int* rope_dst = reinterpret_cast<int*>(dst_ptr + 512);
+  rope_dst[lane_id] = rope_src[lane_id];
 }
 
 template <typename scalar_t>
@@ -1257,15 +1263,16 @@ void cp_gather_and_upconvert_fp8_kv_cache(
     src_ptr = reinterpret_cast<const uint8_t*>(src_cache.data_ptr());
   }
 
-  // Decide on the number of splits based on the batch size
-  int num_splits = batch_size > 128 ? 2 : batch_size > 64 ? 4 : 16;
-  dim3 grid(batch_size, num_splits);
-  dim3 block(576);
+  const int total_tokens = dst.size(0);
+  constexpr int warps_per_block = 8;
+  const int grid_size = (total_tokens + warps_per_block - 1) / warps_per_block;
+  const int block_size_threads = warps_per_block * 32;  // 256 threads
 
-  vllm::cp_gather_and_upconvert_fp8_kv_cache<<<grid, block, 0, stream>>>(
+  vllm::cp_gather_and_upconvert_fp8_kv_cache<<<grid_size, block_size_threads, 0,
+                                               stream>>>(
       src_ptr, reinterpret_cast<__nv_bfloat16*>(dst.data_ptr()),
-      block_table.data_ptr<int32_t>(), seq_lens.data_ptr<int32_t>(),
-      workspace_starts.data_ptr<int32_t>(), block_size, head_dim,
+      block_table.data_ptr<int32_t>(), workspace_starts.data_ptr<int32_t>(),
+      static_cast<int32_t>(batch_size), block_size, total_tokens,
       block_table_stride, cache_block_stride, cache_entry_stride,
       dst_entry_stride);
 }
@@ -1365,3 +1372,43 @@ void cp_gather_indexer_k_quant_cache(
     CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(32);
   }
 }
+
+// Concatenate ql_nope and q_pe into a contiguous q_out tensor for MLA/DSA.
+// Replaces torch.cat((ql_nope, q_pe), dim=-1).
+void concat_mla_q(torch::Tensor& ql_nope,  // [num_tokens, num_heads, nope_dim]
+                  torch::Tensor& q_pe,     // [num_tokens, num_heads, rope_dim]
+                  torch::Tensor& q_out     // [num_tokens, num_heads, nope_dim +
+                                           // rope_dim]
+) {
+  const int num_tokens = ql_nope.size(0);
+  const int num_heads = ql_nope.size(1);
+  const int nope_dim = ql_nope.size(2);
+  const int rope_dim = q_pe.size(2);
+
+  TORCH_CHECK(nope_dim % 512 == 0, "nope_dim must be a multiple of 512, got ",
+              nope_dim);
+  TORCH_CHECK(rope_dim == 64, "rope_dim must be 64, got ", rope_dim);
+  TORCH_CHECK(q_out.size(2) == nope_dim + rope_dim);
+
+  TORCH_CHECK(ql_nope.stride(2) == 1, "ql_nope must have stride 1 in dim 2");
+  TORCH_CHECK(q_pe.stride(2) == 1, "q_pe must have stride 1 in dim 2");
+  TORCH_CHECK(q_out.stride(2) == 1, "q_out must have stride 1 in dim 2");
+
+  if (num_tokens == 0) return;
+
+  constexpr int warps_per_block = 8;
+  const int total_warps = num_tokens * num_heads;
+  const int grid_size = (total_warps + warps_per_block - 1) / warps_per_block;
+  const int block_size = warps_per_block * 32;
+
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(ql_nope));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  VLLM_DISPATCH_FLOATING_TYPES(ql_nope.scalar_type(), "concat_mla_q", [&] {
+    vllm::ConcatMLAQKernel<scalar_t, 512><<<grid_size, block_size, 0, stream>>>(
+        q_out.data_ptr<scalar_t>(), ql_nope.data_ptr<scalar_t>(),
+        q_pe.data_ptr<scalar_t>(), num_tokens, num_heads, q_out.stride(0),
+        q_out.stride(1), ql_nope.stride(0), ql_nope.stride(1), q_pe.stride(0),
+        q_pe.stride(1));
+  });
+}

csrc/concat_mla_q.cuh

@@ -0,0 +1,60 @@
+#ifndef CONCAT_MLA_Q_CUH_
+#define CONCAT_MLA_Q_CUH_
+
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+
+#include "cuda_vec_utils.cuh"
+
+namespace vllm {
+
+// Concatenates ql_nope [num_tokens, num_heads, NOPE_DIM] and
+// q_pe [num_tokens, num_heads, 64]
+// into q_out [num_tokens, num_heads, NOPE_DIM+64].
+// Currently instantiated only for NOPE_DIM=512.
+// Rope dim is hardcoded to 64 (DeepSeek V3.2 MLA)
+template <typename DType, int NOPE_DIM>
+__global__ void ConcatMLAQKernel(
+    DType* __restrict__ q_out, const DType* __restrict__ ql_nope,
+    const DType* __restrict__ q_pe, const int num_tokens, const int num_heads,
+    const int64_t out_stride_0, const int64_t out_stride_1,
+    const int64_t nope_stride_0, const int64_t nope_stride_1,
+    const int64_t pe_stride_0, const int64_t pe_stride_1) {
+  const int flat_warp_id = (blockIdx.x * blockDim.x + threadIdx.x) >> 5;
+  if (flat_warp_id >= num_tokens * num_heads) return;
+
+  const int token_id = flat_warp_id / num_heads;
+  const int head_id = flat_warp_id % num_heads;
+  const int lane_id = threadIdx.x & 31;
+
+  constexpr bool use_256b = VLLM_256B_PTX_ENABLED;
+  constexpr int nope_vec_loads =
+      NOPE_DIM * sizeof(DType) / (VecTraits<use_256b>::ARCH_MAX_VEC_SIZE * 32);
+
+  const DType* nope_src =
+      ql_nope + token_id * nope_stride_0 + head_id * nope_stride_1;
+  DType* nope_dst = q_out + token_id * out_stride_0 + head_id * out_stride_1;
+
+#pragma unroll
+  for (int i = 0; i < nope_vec_loads; i++) {
+    const int offset = i * 32 + lane_id;
+    if constexpr (use_256b) {
+      st256_cs(reinterpret_cast<u32x8_t*>(nope_dst) + offset,
+               ld256_cs(reinterpret_cast<const u32x8_t*>(nope_src) + offset));
+    } else {
+      st128_cs(reinterpret_cast<int4*>(nope_dst) + offset,
+               ld128_cs(reinterpret_cast<const int4*>(nope_src) + offset));
+    }
+  }
+
+  const int* rope_src = reinterpret_cast<const int*>(
+      q_pe + token_id * pe_stride_0 + head_id * pe_stride_1);
+  int* rope_dst = reinterpret_cast<int*>(q_out + token_id * out_stride_0 +
+                                         head_id * out_stride_1 + NOPE_DIM);
+
+  st32_cs(rope_dst + lane_id, ld32_cs(rope_src + lane_id));
+}
+
+}  // namespace vllm
+
+#endif  // CONCAT_MLA_Q_CUH_

csrc/cpu/cpu_types.hpp

@@ -13,6 +13,9 @@
 #elif defined(__aarch64__)
   // arm implementation
   #include "cpu_types_arm.hpp"
+#elif defined(__riscv_v)
+  // riscv implementation
+  #include "cpu_types_riscv.hpp"
 #else
   #warning "unsupported vLLM cpu implementation, vLLM will compile with scalar"
   #include "cpu_types_scalar.hpp"

csrc/cpu/cpu_types_riscv.hpp

@@ -0,0 +1,832 @@
+#ifndef CPU_TYPES_RISCV_HPP
+#define CPU_TYPES_RISCV_HPP
+
+#include <algorithm>
+#include <cmath>
+#include <cstring>
+#include <iostream>
+#include <limits>
+#include <riscv_vector.h>
+#include <torch/all.h>
+
+// ============================================================================
+// Vector Register Type Definitions (VLEN=128 bits)
+// ============================================================================
+
+typedef vfloat16m1_t fixed_vfloat16m1_t
+    __attribute__((riscv_rvv_vector_bits(128)));
+typedef vfloat16m2_t fixed_vfloat16m2_t
+    __attribute__((riscv_rvv_vector_bits(256)));
+
+typedef vfloat32m1_t fixed_vfloat32m1_t
+    __attribute__((riscv_rvv_vector_bits(128)));
+typedef vfloat32m2_t fixed_vfloat32m2_t
+    __attribute__((riscv_rvv_vector_bits(256)));
+typedef vfloat32m4_t fixed_vfloat32m4_t
+    __attribute__((riscv_rvv_vector_bits(512)));
+typedef vfloat32m8_t fixed_vfloat32m8_t
+    __attribute__((riscv_rvv_vector_bits(1024)));
+
+typedef vint32m2_t fixed_vint32m2_t __attribute__((riscv_rvv_vector_bits(256)));
+typedef vint32m4_t fixed_vint32m4_t __attribute__((riscv_rvv_vector_bits(512)));
+
+typedef vuint16m1_t fixed_vuint16m1_t
+    __attribute__((riscv_rvv_vector_bits(128)));
+typedef vuint16m2_t fixed_vuint16m2_t
+    __attribute__((riscv_rvv_vector_bits(256)));
+typedef vuint16m4_t fixed_vuint16m4_t
+    __attribute__((riscv_rvv_vector_bits(512)));
+
+#ifdef RISCV_BF16_SUPPORT
+typedef vbfloat16m1_t fixed_vbfloat16m1_t
+    __attribute__((riscv_rvv_vector_bits(128)));
+typedef vbfloat16m2_t fixed_vbfloat16m2_t
+    __attribute__((riscv_rvv_vector_bits(256)));
+typedef vbfloat16m4_t fixed_vbfloat16m4_t
+    __attribute__((riscv_rvv_vector_bits(512)));
+#endif
+
+namespace vec_op {
+
+#ifdef RISCV_BF16_SUPPORT
+  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
+    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
+    AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)  \
+    AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+#else
+  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
+    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
+    AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
+#endif
+
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
+
+#define FORCE_INLINE __attribute__((always_inline)) inline
+
+namespace {
+template <typename T, T... indexes, typename F>
+constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F&& f) {
+  (f(std::integral_constant<T, indexes>{}), ...);
+};
+}  // namespace
+
+template <typename T, T count, typename F,
+          typename = std::enable_if_t<std::is_invocable_v<F, T>>>
+constexpr void unroll_loop(F&& f) {
+  unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
+}
+
+template <typename T>
+struct Vec {
+  constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; };
+};
+
+struct FP32Vec8;
+struct FP32Vec16;
+
+// ============================================================================
+// FP16 Implementation
+// ============================================================================
+
+struct FP16Vec8 : public Vec<FP16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  fixed_vfloat16m1_t reg;
+
+  explicit FP16Vec8(const void* ptr)
+      : reg(__riscv_vle16_v_f16m1(static_cast<const _Float16*>(ptr),
+                                  VEC_ELEM_NUM)) {};
+
+  explicit FP16Vec8(const FP32Vec8&);
+
+  void save(void* ptr) const {
+    __riscv_vse16_v_f16m1(static_cast<_Float16*>(ptr), reg, VEC_ELEM_NUM);
+  }
+  void save(void* ptr, int elem_num) const {
+    __riscv_vse16_v_f16m1(static_cast<_Float16*>(ptr), reg, elem_num);
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(_Float16);
+    __riscv_vsse16_v_f16m1(static_cast<_Float16*>(ptr), byte_stride, reg,
+                           VEC_ELEM_NUM);
+  }
+};
+
+struct FP16Vec16 : public Vec<FP16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  fixed_vfloat16m2_t reg;
+
+  explicit FP16Vec16(const void* ptr)
+      : reg(__riscv_vle16_v_f16m2(static_cast<const _Float16*>(ptr),
+                                  VEC_ELEM_NUM)) {};
+
+  explicit FP16Vec16(const FP32Vec16& vec);
+
+  void save(void* ptr) const {
+    __riscv_vse16_v_f16m2(static_cast<_Float16*>(ptr), reg, VEC_ELEM_NUM);
+  }
+  void save(void* ptr, int elem_num) const {
+    __riscv_vse16_v_f16m2(static_cast<_Float16*>(ptr), reg, elem_num);
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(_Float16);
+    __riscv_vsse16_v_f16m2(static_cast<_Float16*>(ptr), byte_stride, reg,
+                           VEC_ELEM_NUM);
+  }
+};
+
+// ============================================================================
+// BF16 Implementation
+// ============================================================================
+
+#ifdef RISCV_BF16_SUPPORT
+
+FORCE_INLINE fixed_vuint16m1_t bf16_to_u16(fixed_vbfloat16m1_t v) {
+  return __riscv_vreinterpret_v_bf16m1_u16m1(v);
+}
+FORCE_INLINE fixed_vuint16m2_t bf16_to_u16(fixed_vbfloat16m2_t v) {
+  return __riscv_vreinterpret_v_bf16m2_u16m2(v);
+}
+FORCE_INLINE fixed_vuint16m4_t bf16_to_u16(fixed_vbfloat16m4_t v) {
+  return __riscv_vreinterpret_v_bf16m4_u16m4(v);
+}
+
+struct BF16Vec8 : public Vec<BF16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  fixed_vbfloat16m1_t reg;
+
+  explicit BF16Vec8(const void* ptr)
+      : reg(__riscv_vreinterpret_v_u16m1_bf16m1(__riscv_vle16_v_u16m1(
+            reinterpret_cast<const uint16_t*>(ptr), VEC_ELEM_NUM))) {};
+
+  explicit BF16Vec8(fixed_vbfloat16m1_t data) : reg(data) {};
+  explicit BF16Vec8(const FP32Vec8&);
+
+  void save(void* ptr) const {
+    __riscv_vse16_v_u16m1(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
+                          VEC_ELEM_NUM);
+  }
+  void save(void* ptr, int elem_num) const {
+    __riscv_vse16_v_u16m1(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
+                          elem_num);
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    __riscv_vsse16_v_u16m1(reinterpret_cast<uint16_t*>(ptr), byte_stride,
+                           bf16_to_u16(reg), VEC_ELEM_NUM);
+  }
+};
+
+struct BF16Vec16 : public Vec<BF16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  fixed_vbfloat16m2_t reg;
+
+  explicit BF16Vec16(const void* ptr)
+      : reg(__riscv_vreinterpret_v_u16m2_bf16m2(__riscv_vle16_v_u16m2(
+            reinterpret_cast<const uint16_t*>(ptr), VEC_ELEM_NUM))) {};
+
+  explicit BF16Vec16(fixed_vbfloat16m2_t data) : reg(data) {};
+  explicit BF16Vec16(const FP32Vec16&);
+
+  void save(void* ptr) const {
+    __riscv_vse16_v_u16m2(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
+                          VEC_ELEM_NUM);
+  }
+  void save(void* ptr, int elem_num) const {
+    __riscv_vse16_v_u16m2(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
+                          elem_num);
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    __riscv_vsse16_v_u16m2(reinterpret_cast<uint16_t*>(ptr), byte_stride,
+                           bf16_to_u16(reg), VEC_ELEM_NUM);
+  }
+};
+
+struct BF16Vec32 : public Vec<BF16Vec32> {
+  constexpr static int VEC_ELEM_NUM = 32;
+  fixed_vbfloat16m4_t reg;
+
+  explicit BF16Vec32(const void* ptr)
+      : reg(__riscv_vreinterpret_v_u16m4_bf16m4(__riscv_vle16_v_u16m4(
+            reinterpret_cast<const uint16_t*>(ptr), VEC_ELEM_NUM))) {};
+
+  explicit BF16Vec32(fixed_vbfloat16m4_t data) : reg(data) {};
+
+  explicit BF16Vec32(const BF16Vec8& v) {
+    fixed_vuint16m1_t u16_val = bf16_to_u16(v.reg);
+    fixed_vuint16m4_t u16_combined =
+        __riscv_vcreate_v_u16m1_u16m4(u16_val, u16_val, u16_val, u16_val);
+    reg = __riscv_vreinterpret_v_u16m4_bf16m4(u16_combined);
+  };
+
+  void save(void* ptr) const {
+    __riscv_vse16_v_u16m4(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
+                          VEC_ELEM_NUM);
+  }
+  void save(void* ptr, int elem_num) const {
+    __riscv_vse16_v_u16m4(reinterpret_cast<uint16_t*>(ptr), bf16_to_u16(reg),
+                          elem_num);
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    __riscv_vsse16_v_u16m4(reinterpret_cast<uint16_t*>(ptr), byte_stride,
+                           bf16_to_u16(reg), VEC_ELEM_NUM);
+  }
+};
+
+#else
+// ============================================================================
+// BF16 Fallback Implementation (FP32 Simulation)
+// ============================================================================
+
+struct BF16Vec8 : public Vec<BF16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  fixed_vfloat32m2_t reg_fp32;
+  explicit BF16Vec8(const void* ptr) {
+    const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
+    float tmp[8];
+    for (int i = 0; i < 8; ++i) {
+      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
+      std::memcpy(&tmp[i], &v, 4);
+    }
+    reg_fp32 = __riscv_vle32_v_f32m2(tmp, 8);
+  }
+  explicit BF16Vec8(const FP32Vec8&);
+  void save(void* ptr) const {
+    float tmp[8];
+    __riscv_vse32_v_f32m2(tmp, reg_fp32, 8);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < 8; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+  void save(void* ptr, int elem_num) const {
+    float tmp[8];
+    __riscv_vse32_v_f32m2(tmp, reg_fp32, 8);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < elem_num; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    float tmp[8];
+    __riscv_vse32_v_f32m2(tmp, reg_fp32, 8);
+    uint8_t* u8 = static_cast<uint8_t*>(ptr);
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    for (int i = 0; i < 8; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      uint16_t val = static_cast<uint16_t>(v >> 16);
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+    }
+  }
+};
+
+struct BF16Vec16 : public Vec<BF16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  fixed_vfloat32m4_t reg_fp32;
+  explicit BF16Vec16(const void* ptr) {
+    const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
+    float tmp[16];
+    for (int i = 0; i < 16; ++i) {
+      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
+      std::memcpy(&tmp[i], &v, 4);
+    }
+    reg_fp32 = __riscv_vle32_v_f32m4(tmp, 16);
+  }
+  explicit BF16Vec16(const FP32Vec16&);
+  void save(void* ptr) const {
+    float tmp[16];
+    __riscv_vse32_v_f32m4(tmp, reg_fp32, 16);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < 16; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+  void save(void* ptr, int elem_num) const {
+    float tmp[16];
+    __riscv_vse32_v_f32m4(tmp, reg_fp32, 16);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < elem_num; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    float tmp[16];
+    __riscv_vse32_v_f32m4(tmp, reg_fp32, 16);
+    uint8_t* u8 = static_cast<uint8_t*>(ptr);
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    for (int i = 0; i < 16; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      uint16_t val = static_cast<uint16_t>(v >> 16);
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+    }
+  }
+};
+
+struct BF16Vec32 : public Vec<BF16Vec32> {
+  constexpr static int VEC_ELEM_NUM = 32;
+  fixed_vfloat32m8_t reg_fp32;
+
+  explicit BF16Vec32(const void* ptr) {
+    const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
+    float tmp[32];
+    for (int i = 0; i < 32; ++i) {
+      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
+      std::memcpy(&tmp[i], &v, 4);
+    }
+    reg_fp32 = __riscv_vle32_v_f32m8(tmp, 32);
+  }
+
+  explicit BF16Vec32(const BF16Vec8& v) {
+    float tmp_small[8];
+    __riscv_vse32_v_f32m2(tmp_small, v.reg_fp32, 8);
+    float tmp_large[32];
+    for (int i = 0; i < 4; ++i) {
+      std::memcpy(tmp_large + (i * 8), tmp_small, 8 * sizeof(float));
+    }
+    reg_fp32 = __riscv_vle32_v_f32m8(tmp_large, 32);
+  }
+
+  void save(void* ptr) const {
+    float tmp[32];
+    __riscv_vse32_v_f32m8(tmp, reg_fp32, 32);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < 32; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+
+  void save(void* ptr, int elem_num) const {
+    float tmp[32];
+    __riscv_vse32_v_f32m8(tmp, reg_fp32, 32);
+    uint16_t* u16 = static_cast<uint16_t*>(ptr);
+    for (int i = 0; i < elem_num; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      u16[i] = static_cast<uint16_t>(v >> 16);
+    }
+  }
+
+  void save_strided(void* ptr, ptrdiff_t stride) const {
+    float tmp[32];
+    __riscv_vse32_v_f32m8(tmp, reg_fp32, 32);
+    uint8_t* u8 = static_cast<uint8_t*>(ptr);
+    ptrdiff_t byte_stride = stride * sizeof(uint16_t);
+    for (int i = 0; i < 32; ++i) {
+      uint32_t v;
+      std::memcpy(&v, &tmp[i], 4);
+      uint16_t val = static_cast<uint16_t>(v >> 16);
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+    }
+  }
+};
+#endif
+
+// ============================================================================
+// FP32 Implementation
+// ============================================================================
+
+struct FP32Vec4 : public Vec<FP32Vec4> {
+  constexpr static int VEC_ELEM_NUM = 4;
+  fixed_vfloat32m1_t reg;
+  explicit FP32Vec4(float v) : reg(__riscv_vfmv_v_f_f32m1(v, VEC_ELEM_NUM)) {};
+  explicit FP32Vec4() : reg(__riscv_vfmv_v_f_f32m1(0.0f, VEC_ELEM_NUM)) {};
+  explicit FP32Vec4(const float* ptr)
+      : reg(__riscv_vle32_v_f32m1(ptr, VEC_ELEM_NUM)) {};
+  explicit FP32Vec4(fixed_vfloat32m1_t data) : reg(data) {};
+  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {};
+  void save(float* ptr) const { __riscv_vse32_v_f32m1(ptr, reg, VEC_ELEM_NUM); }
+  void save(float* ptr, int elem_num) const {
+    __riscv_vse32_v_f32m1(ptr, reg, elem_num);
+  }
+};
+
+struct FP32Vec8 : public Vec<FP32Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  fixed_vfloat32m2_t reg;
+
+  explicit FP32Vec8(float v) : reg(__riscv_vfmv_v_f_f32m2(v, VEC_ELEM_NUM)) {};
+  explicit FP32Vec8() : reg(__riscv_vfmv_v_f_f32m2(0.0f, VEC_ELEM_NUM)) {};
+  explicit FP32Vec8(const float* ptr)
+      : reg(__riscv_vle32_v_f32m2(ptr, VEC_ELEM_NUM)) {};
+  explicit FP32Vec8(fixed_vfloat32m2_t data) : reg(data) {};
+  explicit FP32Vec8(const FP32Vec8& data) : reg(data.reg) {};
+  explicit FP32Vec8(const FP16Vec8& v)
+      : reg(__riscv_vfwcvt_f_f_v_f32m2(v.reg, VEC_ELEM_NUM)) {};
+  explicit FP32Vec8(fixed_vfloat16m1_t v)
+      : reg(__riscv_vfwcvt_f_f_v_f32m2(v, VEC_ELEM_NUM)) {};
+
+#ifdef RISCV_BF16_SUPPORT
+  explicit FP32Vec8(fixed_vbfloat16m1_t v)
+      : reg(__riscv_vfwcvtbf16_f_f_v_f32m2(v, VEC_ELEM_NUM)) {};
+  explicit FP32Vec8(const BF16Vec8& v)
+      : reg(__riscv_vfwcvtbf16_f_f_v_f32m2(v.reg, VEC_ELEM_NUM)) {};
+#else
+  explicit FP32Vec8(const BF16Vec8& v) : reg(v.reg_fp32) {};
+#endif
+
+  float reduce_sum() const {
+    fixed_vfloat32m1_t scalar = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    scalar = __riscv_vfredusum_vs_f32m2_f32m1(reg, scalar, VEC_ELEM_NUM);
+    return __riscv_vfmv_f_s_f32m1_f32(scalar);
+  }
+
+  FP32Vec8 operator*(const FP32Vec8& b) const {
+    return FP32Vec8(__riscv_vfmul_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec8 operator+(const FP32Vec8& b) const {
+    return FP32Vec8(__riscv_vfadd_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec8 operator-(const FP32Vec8& b) const {
+    return FP32Vec8(__riscv_vfsub_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec8 operator/(const FP32Vec8& b) const {
+    return FP32Vec8(__riscv_vfdiv_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
+  }
+
+  FP32Vec8 min(const FP32Vec8& b) const {
+    return FP32Vec8(__riscv_vfmin_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec8 max(const FP32Vec8& b) const {
+    return FP32Vec8(__riscv_vfmax_vv_f32m2(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec8 abs() const {
+    return FP32Vec8(__riscv_vfabs_v_f32m2(reg, VEC_ELEM_NUM));
+  }
+
+  FP32Vec8 min(const FP32Vec8& b, int elem_num) const {
+    return FP32Vec8(__riscv_vfmin_vv_f32m2(reg, b.reg, elem_num));
+  }
+  FP32Vec8 max(const FP32Vec8& b, int elem_num) const {
+    return FP32Vec8(__riscv_vfmax_vv_f32m2(reg, b.reg, elem_num));
+  }
+
+  FP32Vec8 clamp(const FP32Vec8& min_v, const FP32Vec8& max_v) const {
+    fixed_vfloat32m2_t temp =
+        __riscv_vfmax_vv_f32m2(min_v.reg, reg, VEC_ELEM_NUM);
+    return FP32Vec8(__riscv_vfmin_vv_f32m2(max_v.reg, temp, VEC_ELEM_NUM));
+  }
+
+  void save(float* ptr) const { __riscv_vse32_v_f32m2(ptr, reg, VEC_ELEM_NUM); }
+  void save(float* ptr, int elem_num) const {
+    __riscv_vse32_v_f32m2(ptr, reg, elem_num);
+  }
+  void save_strided(float* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(float);
+    __riscv_vsse32_v_f32m2(ptr, byte_stride, reg, VEC_ELEM_NUM);
+  }
+
+  FP32Vec8 exp() const {
+    const float inv_ln2 = 1.44269504088896341f;
+    fixed_vfloat32m2_t x_scaled =
+        __riscv_vfmul_vf_f32m2(reg, inv_ln2, VEC_ELEM_NUM);
+    fixed_vint32m2_t n_int = __riscv_vfcvt_x_f_v_i32m2(x_scaled, VEC_ELEM_NUM);
+    fixed_vfloat32m2_t n_float = __riscv_vfcvt_f_x_v_f32m2(n_int, VEC_ELEM_NUM);
+
+    fixed_vfloat32m2_t r =
+        __riscv_vfsub_vv_f32m2(x_scaled, n_float, VEC_ELEM_NUM);
+
+    fixed_vfloat32m2_t poly =
+        __riscv_vfmv_v_f_f32m2(0.001333355810164f, VEC_ELEM_NUM);
+    poly = __riscv_vfmul_vv_f32m2(poly, r, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m2(poly, 0.009618129107628f, VEC_ELEM_NUM);
+    poly = __riscv_vfmul_vv_f32m2(poly, r, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m2(poly, 0.055504108664821f, VEC_ELEM_NUM);
+    poly = __riscv_vfmul_vv_f32m2(poly, r, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m2(poly, 0.240226506959101f, VEC_ELEM_NUM);
+    poly = __riscv_vfmul_vv_f32m2(poly, r, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m2(poly, 0.693147180559945f, VEC_ELEM_NUM);
+    poly = __riscv_vfmul_vv_f32m2(poly, r, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m2(poly, 1.0f, VEC_ELEM_NUM);
+
+    fixed_vint32m2_t biased_exp =
+        __riscv_vadd_vx_i32m2(n_int, 127, VEC_ELEM_NUM);
+    biased_exp = __riscv_vmax_vx_i32m2(biased_exp, 0, VEC_ELEM_NUM);
+    fixed_vint32m2_t exponent_bits =
+        __riscv_vsll_vx_i32m2(biased_exp, 23, VEC_ELEM_NUM);
+    fixed_vfloat32m2_t scale =
+        __riscv_vreinterpret_v_i32m2_f32m2(exponent_bits);
+
+    return FP32Vec8(__riscv_vfmul_vv_f32m2(poly, scale, VEC_ELEM_NUM));
+  }
+
+  FP32Vec8 tanh() const {
+    fixed_vfloat32m2_t x_clamped = __riscv_vfmin_vf_f32m2(
+        __riscv_vfmax_vf_f32m2(reg, -9.0f, VEC_ELEM_NUM), 9.0f, VEC_ELEM_NUM);
+    fixed_vfloat32m2_t x2 =
+        __riscv_vfmul_vf_f32m2(x_clamped, 2.0f, VEC_ELEM_NUM);
+    FP32Vec8 exp_val = FP32Vec8(x2).exp();
+    fixed_vfloat32m2_t num =
+        __riscv_vfsub_vf_f32m2(exp_val.reg, 1.0f, VEC_ELEM_NUM);
+    fixed_vfloat32m2_t den =
+        __riscv_vfadd_vf_f32m2(exp_val.reg, 1.0f, VEC_ELEM_NUM);
+    return FP32Vec8(__riscv_vfdiv_vv_f32m2(num, den, VEC_ELEM_NUM));
+  }
+
+  FP32Vec8 er() const {
+    const float p = 0.3275911f, a1 = 0.254829592f, a2 = -0.284496736f,
+                a3 = 1.421413741f, a4 = -1.453152027f, a5 = 1.061405429f;
+    fixed_vfloat32m2_t abs_x = __riscv_vfabs_v_f32m2(reg, VEC_ELEM_NUM);
+
+    fixed_vfloat32m2_t t = __riscv_vfadd_vf_f32m2(
+        __riscv_vfmul_vf_f32m2(abs_x, p, VEC_ELEM_NUM), 1.0f, VEC_ELEM_NUM);
+    t = __riscv_vfrdiv_vf_f32m2(t, 1.0f, VEC_ELEM_NUM);
+
+    fixed_vfloat32m2_t poly = __riscv_vfmv_v_f_f32m2(a5, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m2(__riscv_vfmul_vv_f32m2(poly, t, VEC_ELEM_NUM),
+                                  a4, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m2(__riscv_vfmul_vv_f32m2(poly, t, VEC_ELEM_NUM),
+                                  a3, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m2(__riscv_vfmul_vv_f32m2(poly, t, VEC_ELEM_NUM),
+                                  a2, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m2(__riscv_vfmul_vv_f32m2(poly, t, VEC_ELEM_NUM),
+                                  a1, VEC_ELEM_NUM);
+    poly = __riscv_vfmul_vv_f32m2(poly, t, VEC_ELEM_NUM);
+
+    fixed_vfloat32m2_t exp_val =
+        FP32Vec8(__riscv_vfneg_v_f32m2(
+                     __riscv_vfmul_vv_f32m2(abs_x, abs_x, VEC_ELEM_NUM),
+                     VEC_ELEM_NUM))
+            .exp()
+            .reg;
+    fixed_vfloat32m2_t res = __riscv_vfrsub_vf_f32m2(
+        __riscv_vfmul_vv_f32m2(poly, exp_val, VEC_ELEM_NUM), 1.0f,
+        VEC_ELEM_NUM);
+
+    vbool16_t mask = __riscv_vmflt_vf_f32m2_b16(reg, 0.0f, VEC_ELEM_NUM);
+    return FP32Vec8(__riscv_vfneg_v_f32m2_m(mask, res, VEC_ELEM_NUM));
+  }
+};
+
+struct FP32Vec16 : public Vec<FP32Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  fixed_vfloat32m4_t reg;
+
+  explicit FP32Vec16(float v) : reg(__riscv_vfmv_v_f_f32m4(v, VEC_ELEM_NUM)) {};
+  explicit FP32Vec16() : reg(__riscv_vfmv_v_f_f32m4(0.0f, VEC_ELEM_NUM)) {};
+  explicit FP32Vec16(const float* ptr)
+      : reg(__riscv_vle32_v_f32m4(ptr, VEC_ELEM_NUM)) {};
+  explicit FP32Vec16(fixed_vfloat32m4_t data) : reg(data) {};
+  explicit FP32Vec16(const FP32Vec8& data)
+      : reg(__riscv_vcreate_v_f32m2_f32m4(data.reg, data.reg)) {};
+  explicit FP32Vec16(const FP32Vec16& data) : reg(data.reg) {};
+  explicit FP32Vec16(const FP16Vec16& v);
+
+#ifdef RISCV_BF16_SUPPORT
+  explicit FP32Vec16(fixed_vbfloat16m2_t v)
+      : reg(__riscv_vfwcvtbf16_f_f_v_f32m4(v, VEC_ELEM_NUM)) {};
+  explicit FP32Vec16(const BF16Vec16& v)
+      : reg(__riscv_vfwcvtbf16_f_f_v_f32m4(v.reg, VEC_ELEM_NUM)) {};
+#else
+  explicit FP32Vec16(const BF16Vec16& v) : reg(v.reg_fp32) {};
+#endif
+
+  FP32Vec16 operator+(const FP32Vec16& b) const {
+    return FP32Vec16(__riscv_vfadd_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec16 operator-(const FP32Vec16& b) const {
+    return FP32Vec16(__riscv_vfsub_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec16 operator*(const FP32Vec16& b) const {
+    return FP32Vec16(__riscv_vfmul_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec16 operator/(const FP32Vec16& b) const {
+    return FP32Vec16(__riscv_vfdiv_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
+  }
+
+  FP32Vec16 fma(const FP32Vec16& a, const FP32Vec16& b) const {
+    return FP32Vec16(__riscv_vfmacc_vv_f32m4(reg, a.reg, b.reg, VEC_ELEM_NUM));
+  }
+
+  float reduce_sum() const {
+    fixed_vfloat32m1_t scalar = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    scalar = __riscv_vfredusum_vs_f32m4_f32m1(reg, scalar, VEC_ELEM_NUM);
+    return __riscv_vfmv_f_s_f32m1_f32(scalar);
+  }
+
+  float reduce_max() const {
+    fixed_vfloat32m1_t scalar =
+        __riscv_vfmv_s_f_f32m1(std::numeric_limits<float>::lowest(), 1);
+    scalar = __riscv_vfredmax_vs_f32m4_f32m1(reg, scalar, VEC_ELEM_NUM);
+    return __riscv_vfmv_f_s_f32m1_f32(scalar);
+  }
+
+  float reduce_min() const {
+    fixed_vfloat32m1_t scalar =
+        __riscv_vfmv_s_f_f32m1(std::numeric_limits<float>::max(), 1);
+    scalar = __riscv_vfredmin_vs_f32m4_f32m1(reg, scalar, VEC_ELEM_NUM);
+    return __riscv_vfmv_f_s_f32m1_f32(scalar);
+  }
+
+  template <int group_size>
+  float reduce_sub_sum(int idx) {
+    static_assert(VEC_ELEM_NUM % group_size == 0);
+    const int start = idx * group_size;
+    vuint32m4_t indices = __riscv_vid_v_u32m4(VEC_ELEM_NUM);
+    vbool8_t mask = __riscv_vmand_mm_b8(
+        __riscv_vmsgeu_vx_u32m4_b8(indices, start, VEC_ELEM_NUM),
+        __riscv_vmsltu_vx_u32m4_b8(indices, start + group_size, VEC_ELEM_NUM),
+        VEC_ELEM_NUM);
+    fixed_vfloat32m1_t scalar = __riscv_vfmv_s_f_f32m1(0.0f, 1);
+    scalar =
+        __riscv_vfredusum_vs_f32m4_f32m1_m(mask, reg, scalar, VEC_ELEM_NUM);
+    return __riscv_vfmv_f_s_f32m1_f32(scalar);
+  };
+
+  FP32Vec16 max(const FP32Vec16& b) const {
+    return FP32Vec16(__riscv_vfmax_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec16 min(const FP32Vec16& b) const {
+    return FP32Vec16(__riscv_vfmin_vv_f32m4(reg, b.reg, VEC_ELEM_NUM));
+  }
+  FP32Vec16 abs() const {
+    return FP32Vec16(__riscv_vfabs_v_f32m4(reg, VEC_ELEM_NUM));
+  }
+
+  FP32Vec16 clamp(const FP32Vec16& min_v, const FP32Vec16& max_v) const {
+    return FP32Vec16(__riscv_vfmin_vv_f32m4(
+        max_v.reg, __riscv_vfmax_vv_f32m4(min_v.reg, reg, VEC_ELEM_NUM),
+        VEC_ELEM_NUM));
+  }
+
+  void save(float* ptr) const { __riscv_vse32_v_f32m4(ptr, reg, VEC_ELEM_NUM); }
+  void save(float* ptr, int elem_num) const {
+    __riscv_vse32_v_f32m4(ptr, reg, elem_num);
+  }
+  void save_strided(float* ptr, ptrdiff_t stride) const {
+    ptrdiff_t byte_stride = stride * sizeof(float);
+    __riscv_vsse32_v_f32m4(ptr, byte_stride, reg, VEC_ELEM_NUM);
+  }
+
+  FP32Vec16 exp() const {
+    const float inv_ln2 = 1.44269504088896341f;
+    fixed_vfloat32m4_t x_scaled =
+        __riscv_vfmul_vf_f32m4(reg, inv_ln2, VEC_ELEM_NUM);
+    fixed_vint32m4_t n_int = __riscv_vfcvt_x_f_v_i32m4(x_scaled, VEC_ELEM_NUM);
+    fixed_vfloat32m4_t n_float = __riscv_vfcvt_f_x_v_f32m4(n_int, VEC_ELEM_NUM);
+    fixed_vfloat32m4_t r =
+        __riscv_vfsub_vv_f32m4(x_scaled, n_float, VEC_ELEM_NUM);
+
+    fixed_vfloat32m4_t poly =
+        __riscv_vfmv_v_f_f32m4(0.001333355810164f, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, r, VEC_ELEM_NUM),
+                                  0.009618129107628f, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, r, VEC_ELEM_NUM),
+                                  0.055504108664821f, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, r, VEC_ELEM_NUM),
+                                  0.240226506959101f, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, r, VEC_ELEM_NUM),
+                                  0.693147180559945f, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, r, VEC_ELEM_NUM),
+                                  1.0f, VEC_ELEM_NUM);
+
+    fixed_vint32m4_t biased_exp = __riscv_vmax_vx_i32m4(
+        __riscv_vadd_vx_i32m4(n_int, 127, VEC_ELEM_NUM), 0, VEC_ELEM_NUM);
+    fixed_vfloat32m4_t scale = __riscv_vreinterpret_v_i32m4_f32m4(
+        __riscv_vsll_vx_i32m4(biased_exp, 23, VEC_ELEM_NUM));
+
+    return FP32Vec16(__riscv_vfmul_vv_f32m4(poly, scale, VEC_ELEM_NUM));
+  }
+
+  FP32Vec16 tanh() const {
+    fixed_vfloat32m4_t x_clamped = __riscv_vfmin_vf_f32m4(
+        __riscv_vfmax_vf_f32m4(reg, -9.0f, VEC_ELEM_NUM), 9.0f, VEC_ELEM_NUM);
+    FP32Vec16 exp_val =
+        FP32Vec16(__riscv_vfmul_vf_f32m4(x_clamped, 2.0f, VEC_ELEM_NUM)).exp();
+    return FP32Vec16(__riscv_vfdiv_vv_f32m4(
+        __riscv_vfsub_vf_f32m4(exp_val.reg, 1.0f, VEC_ELEM_NUM),
+        __riscv_vfadd_vf_f32m4(exp_val.reg, 1.0f, VEC_ELEM_NUM), VEC_ELEM_NUM));
+  }
+
+  FP32Vec16 er() const {
+    const float p = 0.3275911f, a1 = 0.254829592f, a2 = -0.284496736f,
+                a3 = 1.421413741f, a4 = -1.453152027f, a5 = 1.061405429f;
+    fixed_vfloat32m4_t abs_x = __riscv_vfabs_v_f32m4(reg, VEC_ELEM_NUM);
+    fixed_vfloat32m4_t t = __riscv_vfrdiv_vf_f32m4(
+        __riscv_vfadd_vf_f32m4(__riscv_vfmul_vf_f32m4(abs_x, p, VEC_ELEM_NUM),
+                               1.0f, VEC_ELEM_NUM),
+        1.0f, VEC_ELEM_NUM);
+
+    fixed_vfloat32m4_t poly = __riscv_vfmv_v_f_f32m4(a5, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, t, VEC_ELEM_NUM),
+                                  a4, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, t, VEC_ELEM_NUM),
+                                  a3, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, t, VEC_ELEM_NUM),
+                                  a2, VEC_ELEM_NUM);
+    poly = __riscv_vfadd_vf_f32m4(__riscv_vfmul_vv_f32m4(poly, t, VEC_ELEM_NUM),
+                                  a1, VEC_ELEM_NUM);
+    poly = __riscv_vfmul_vv_f32m4(poly, t, VEC_ELEM_NUM);
+
+    fixed_vfloat32m4_t exp_val =
+        FP32Vec16(__riscv_vfneg_v_f32m4(
+                      __riscv_vfmul_vv_f32m4(abs_x, abs_x, VEC_ELEM_NUM),
+                      VEC_ELEM_NUM))
+            .exp()
+            .reg;
+    fixed_vfloat32m4_t res = __riscv_vfrsub_vf_f32m4(
+        __riscv_vfmul_vv_f32m4(poly, exp_val, VEC_ELEM_NUM), 1.0f,
+        VEC_ELEM_NUM);
+
+    vbool8_t mask = __riscv_vmflt_vf_f32m4_b8(reg, 0.0f, VEC_ELEM_NUM);
+    return FP32Vec16(__riscv_vfneg_v_f32m4_m(mask, res, VEC_ELEM_NUM));
+  }
+};
+
+// ============================================================================
+// Type Traits & Global Helpers
+// ============================================================================
+
+template <typename T>
+struct VecType {
+  using vec_type = void;
+  using vec_t = void;
+};
+
+template <typename T>
+using vec_t = typename VecType<T>::vec_type;
+
+template <>
+struct VecType<float> {
+  using vec_type = FP32Vec8;
+  using vec_t = FP32Vec8;
+};
+template <>
+struct VecType<c10::Half> {
+  using vec_type = FP16Vec8;
+  using vec_t = FP16Vec8;
+};
+template <>
+struct VecType<c10::BFloat16> {
+  using vec_type = BF16Vec8;
+  using vec_t = BF16Vec8;
+};
+
+template <typename T>
+void storeFP32(float v, T* ptr) {
+  *ptr = v;
+}
+template <>
+inline void storeFP32<c10::Half>(float v, c10::Half* ptr) {
+  *reinterpret_cast<_Float16*>(ptr) = static_cast<_Float16>(v);
+}
+
+inline FP16Vec16::FP16Vec16(const FP32Vec16& v) {
+  reg = __riscv_vfncvt_f_f_w_f16m2(v.reg, VEC_ELEM_NUM);
+}
+inline FP16Vec8::FP16Vec8(const FP32Vec8& v) {
+  reg = __riscv_vfncvt_f_f_w_f16m1(v.reg, VEC_ELEM_NUM);
+}
+inline FP32Vec16::FP32Vec16(const FP16Vec16& v) {
+  reg = __riscv_vfwcvt_f_f_v_f32m4(v.reg, VEC_ELEM_NUM);
+}
+inline void fma(FP32Vec16& acc, const FP32Vec16& a, const FP32Vec16& b) {
+  acc = acc.fma(a, b);
+}
+
+#ifdef RISCV_BF16_SUPPORT
+template <>
+inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
+  *ptr = static_cast<__bf16>(v);
+};
+inline BF16Vec8::BF16Vec8(const FP32Vec8& v)
+    : reg(__riscv_vfncvtbf16_f_f_w_bf16m1(v.reg, VEC_ELEM_NUM)) {};
+inline BF16Vec16::BF16Vec16(const FP32Vec16& v)
+    : reg(__riscv_vfncvtbf16_f_f_w_bf16m2(v.reg, VEC_ELEM_NUM)) {};
+#else
+template <>
+inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
+  uint32_t val;
+  std::memcpy(&val, &v, 4);
+  *reinterpret_cast<uint16_t*>(ptr) = static_cast<uint16_t>(val >> 16);
+}
+inline BF16Vec8::BF16Vec8(const FP32Vec8& v) : reg_fp32(v.reg) {}
+inline BF16Vec16::BF16Vec16(const FP32Vec16& v) : reg_fp32(v.reg) {}
+#endif
+
+inline void prefetch(const void* addr) { __builtin_prefetch(addr, 0, 1); }
+
+}  // namespace vec_op
+
+#ifndef CPU_KERNEL_GUARD_IN
+  #define CPU_KERNEL_GUARD_IN(NAME)
+#endif
+
+#ifndef CPU_KERNEL_GUARD_OUT
+  #define CPU_KERNEL_GUARD_OUT(NAME)
+#endif
+
+#endif  // CPU_TYPES_RISCV_HPP

csrc/cuda_vec_utils.cuh

@@ -196,7 +196,6 @@ __forceinline__ __device__ u32x8_t ld256_cs(const u32x8_t* addr) {
   return val;
 #else
   assert(false && "ld256_cs requires SM100+ with CUDA 12.9+");
-  return {};
 #endif
 }
 
@@ -211,23 +210,51 @@ __forceinline__ __device__ void st256_cs(u32x8_t* addr, u32x8_t val) {
 #endif
 }
 
-// 32-bit cache-streaming (.cs) load / store  — SM100+ only.
+// 32-bit load / store.
+__device__ __forceinline__ int ld32(const int* addr) { return __ldg(addr); }
+
+__device__ __forceinline__ void st32(int* addr, int val) { *addr = val; }
+
+// 32-bit cache-streaming (.cs) load / store.
+// Falls back to ld32/st32 on ROCm (no .cs hint).
 __forceinline__ __device__ int ld32_cs(const int* addr) {
-#if VLLM_256B_PTX_ENABLED
   int val;
+#ifndef USE_ROCM
   asm volatile("ld.global.cs.b32 %0, [%1];" : "=r"(val) : "l"(addr));
-  return val;
 #else
-  assert(false && "ld32_cs requires SM100+ with CUDA 12.9+");
-  return 0;
+  val = ld32(addr);
 #endif
+  return val;
 }
 
 __forceinline__ __device__ void st32_cs(int* addr, int val) {
-#if VLLM_256B_PTX_ENABLED
+#ifndef USE_ROCM
   asm volatile("st.global.cs.b32 [%0], %1;" ::"l"(addr), "r"(val));
 #else
-  assert(false && "st32_cs requires SM100+ with CUDA 12.9+");
+  st32(addr, val);
+#endif
+}
+
+// 128-bit cache-streaming (.cs) load / store.
+// Falls back to ld128/st128 on ROCm (no .cs hint).
+__forceinline__ __device__ int4 ld128_cs(const int4* addr) {
+  int4 val;
+#ifndef USE_ROCM
+  asm volatile("ld.global.cs.v4.u32 {%0,%1,%2,%3}, [%4];"
+               : "=r"(val.x), "=r"(val.y), "=r"(val.z), "=r"(val.w)
+               : "l"(addr));
+#else
+  ld128(val, addr);
+#endif
+  return val;
+}
+
+__forceinline__ __device__ void st128_cs(int4* addr, int4 val) {
+#ifndef USE_ROCM
+  asm volatile("st.global.cs.v4.u32 [%0], {%1,%2,%3,%4};" ::"l"(addr),
+               "r"(val.x), "r"(val.y), "r"(val.z), "r"(val.w));
+#else
+  st128(val, addr);
 #endif
 }
 
@@ -260,7 +287,7 @@ __device__ __forceinline__ void ld256_cg_or_zero(u32x8_t& val, const void* ptr,
 
 __device__ __forceinline__ void ld128_cg_or_zero(uint4& val, const void* ptr,
                                                  bool pred) {
-#if VLLM_256B_PTX_ENABLED
+#ifndef USE_ROCM
   uint32_t r0, r1, r2, r3;
 
   asm volatile(
@@ -278,7 +305,7 @@ __device__ __forceinline__ void ld128_cg_or_zero(uint4& val, const void* ptr,
 
   val = uint4{r0, r1, r2, r3};
 #else
-  assert(false && "ld128_cg_or_zero requires SM100+ with CUDA 12.9+");
+  assert(false && "ld128_cg_or_zero is not supported on ROCm");
 #endif
 }
 

csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu

@@ -15,31 +15,33 @@ __device__ void rms_norm_dynamic_per_token_quant_vec(
     scalar_t const* __restrict__ input,   // [..., hidden_size]
     scalar_t const* __restrict__ weight,  // [hidden_size]
     float const* scale_ub, float const var_epsilon, int32_t const hidden_size,
-    scalar_t* __restrict__ residual = nullptr) {
+    int32_t const input_stride, scalar_t* __restrict__ residual = nullptr) {
   float rms = 0.0f;
   float token_scale = 0.0f;
 
   // Compute rms
   vllm::vectorized::compute_rms<scalar_t, has_residual>(
-      &rms, input, hidden_size, var_epsilon, residual);
+      &rms, input, hidden_size, input_stride, var_epsilon, residual);
 
   // Compute scale
   vllm::vectorized::compute_dynamic_per_token_scales<scalar_t, scalar_out_t,
                                                      has_residual>(
       &token_scale, scales, input, weight, rms, scale_ub, hidden_size,
-      residual);
+      input_stride, residual);
 
   // RMS Norm + Quant
   if constexpr (std::is_same_v<scalar_out_t, int8_t>) {
     token_scale = 1.0f / token_scale;
     vllm::vectorized::norm_and_quant<scalar_t, scalar_out_t, true,
-                                     has_residual>(
-        out, input, weight, rms, &token_scale, hidden_size, residual);
+                                     has_residual>(out, input, weight, rms,
+                                                   &token_scale, hidden_size,
+                                                   input_stride, residual);
   } else {
     // FP8 - Do not invert token_scale for exact match with FBGemm
     vllm::vectorized::norm_and_quant<scalar_t, scalar_out_t, false,
-                                     has_residual>(
-        out, input, weight, rms, &token_scale, hidden_size, residual);
+                                     has_residual>(out, input, weight, rms,
+                                                   &token_scale, hidden_size,
+                                                   input_stride, residual);
   }
 }
 
@@ -51,38 +53,40 @@ __global__ void rms_norm_dynamic_per_token_quant_kernel(
     scalar_t const* __restrict__ input,   // [..., hidden_size]
     scalar_t const* __restrict__ weight,  // [hidden_size]
     float const* scale_ub, float const var_epsilon, int32_t const hidden_size,
-    scalar_t* __restrict__ residual = nullptr) {
+    int32_t const input_stride, scalar_t* __restrict__ residual = nullptr) {
   // For vectorization, token_input and token_output pointers need to be
   // aligned at 8-byte and 4-byte addresses respectively.
-  bool const can_vectorize = hidden_size % 4 == 0;
+  bool const can_vectorize = hidden_size % 4 == 0 and input_stride % 4 == 0;
 
   if (can_vectorize) {
     return rms_norm_dynamic_per_token_quant_vec<scalar_t, scalar_out_t,
                                                 has_residual>(
         out, scales, input, weight, scale_ub, var_epsilon, hidden_size,
-        residual);
+        input_stride, residual);
   }
 
   float rms = 0.0f;
   float token_scale = 0.0f;
 
   // Compute RMS
-  vllm::compute_rms<scalar_t, has_residual>(&rms, input, hidden_size,
-                                            var_epsilon, residual);
+  vllm::compute_rms<scalar_t, has_residual>(
+      &rms, input, hidden_size, input_stride, var_epsilon, residual);
   // Compute Scale
   vllm::compute_dynamic_per_token_scales<scalar_t, scalar_out_t, has_residual>(
       &token_scale, scales, input, weight, rms, scale_ub, hidden_size,
-      residual);
+      input_stride, residual);
 
   // RMS Norm + Quant
   if constexpr (std::is_same_v<scalar_out_t, int8_t>) {
     token_scale = 1.0f / token_scale;
     vllm::norm_and_quant<scalar_t, scalar_out_t, true, has_residual>(
-        out, input, weight, rms, &token_scale, hidden_size, residual);
+        out, input, weight, rms, &token_scale, hidden_size, input_stride,
+        residual);
   } else {
     // FP8 - Do not invert s_token_scale for exact match with FBGemm
     vllm::norm_and_quant<scalar_t, scalar_out_t, false, has_residual>(
-        out, input, weight, rms, &token_scale, hidden_size, residual);
+        out, input, weight, rms, &token_scale, hidden_size, input_stride,
+        residual);
   }
 }
 
@@ -97,19 +101,20 @@ __global__ void rms_norm_per_block_quant_kernel(
     scalar_t const* __restrict__ input,   // [..., hidden_size]
     scalar_t const* __restrict__ weight,  // [hidden_size]
     float const* scale_ub, float const var_epsilon, int32_t const hidden_size,
-    scalar_t* __restrict__ residual = nullptr, int64_t outer_scale_stride = 1) {
+    int32_t const input_stride, scalar_t* __restrict__ residual = nullptr,
+    int64_t outer_scale_stride = 1) {
   float rms;
   // Compute RMS
   // Always able to vectorize due to constraints on hidden_size
   vllm::vectorized::compute_rms<scalar_t, has_residual>(
-      &rms, input, hidden_size, var_epsilon, residual);
+      &rms, input, hidden_size, input_stride, var_epsilon, residual);
 
   // Compute Scale
   // Always able to vectorize due to constraints on hidden_size and group_size
   vllm::vectorized::compute_dynamic_per_token_scales<
       scalar_t, scalar_out_t, has_residual, is_scale_transposed, group_size>(
-      nullptr, scales, input, weight, rms, scale_ub, hidden_size, residual,
-      outer_scale_stride);
+      nullptr, scales, input, weight, rms, scale_ub, hidden_size, input_stride,
+      residual, outer_scale_stride);
 
   // RMS Norm + Quant
   // Always able to vectorize due to constraints on hidden_size
@@ -120,7 +125,7 @@ __global__ void rms_norm_per_block_quant_kernel(
   vllm::vectorized::norm_and_quant<
       scalar_t, scalar_out_t, std::is_same_v<scalar_out_t, int8_t>,
       has_residual, is_scale_transposed, group_size>(
-      out, input, weight, rms, scales, hidden_size, residual,
+      out, input, weight, rms, scales, hidden_size, input_stride, residual,
       outer_scale_stride);
 }
 
@@ -137,6 +142,7 @@ void rms_norm_dynamic_per_token_quant_dispatch(
     std::optional<at::Tensor> const& scale_ub,
     std::optional<at::Tensor>& residual) {
   int32_t hidden_size = input.size(-1);
+  int32_t input_stride = input.view({-1, hidden_size}).stride(0);
   auto num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
@@ -153,7 +159,7 @@ void rms_norm_dynamic_per_token_quant_dispatch(
                   out.data_ptr<scalar_t>(), scales.data_ptr<float>(),
                   input.data_ptr<scalar_in_t>(), weight.data_ptr<scalar_in_t>(),
                   scale_ub.has_value() ? scale_ub->data_ptr<float>() : nullptr,
-                  var_epsilon, hidden_size,
+                  var_epsilon, hidden_size, input_stride,
                   has_residual ? residual->data_ptr<scalar_in_t>() : nullptr);
         });
   });
@@ -170,7 +176,9 @@ void rms_norm_dynamic_per_token_quant(
                                         ? c10::ScalarType::Float8_e4m3fn
                                         : c10::ScalarType::Float8_e4m3fnuz;
   TORCH_CHECK(out.dtype() == kFp8Type || out.dtype() == torch::kInt8);
-  TORCH_CHECK(out.is_contiguous() && input.is_contiguous());
+  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK(input.stride(-1) == 1,
+              "Input must be contiguous in the last dimension");
 
   if (scale_ub.has_value()) {
     TORCH_CHECK(out.dtype() == kFp8Type);
@@ -179,6 +187,7 @@ void rms_norm_dynamic_per_token_quant(
   TORCH_CHECK(scales.dtype() == torch::kFloat32);
   if (residual) {
     TORCH_CHECK(residual->scalar_type() == input.scalar_type());
+    TORCH_CHECK(residual->is_contiguous());
   }
 
   VLLM_DISPATCH_FLOATING_TYPES(
@@ -200,6 +209,15 @@ void rms_norm_per_block_quant_dispatch(
     std::optional<at::Tensor> const& scale_ub,
     std::optional<at::Tensor>& residual, bool is_scale_transposed) {
   int32_t hidden_size = input.size(-1);
+  int32_t input_stride = input.view({-1, hidden_size}).stride(0);
+
+  TORCH_CHECK(hidden_size % 4 == 0,
+              "Hidden size must be divisible by 4 for vectorized access");
+  TORCH_CHECK(input_stride % 4 == 0,
+              "Input stride must be divisible by 4 for vectorized access");
+  TORCH_CHECK(group_size % 4 == 0,
+              "Group size must be divisible by 4 for vectorized access");
+
   auto num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
@@ -225,7 +243,7 @@ void rms_norm_per_block_quant_dispatch(
                             weight.data_ptr<scalar_in_t>(),
                             scale_ub.has_value() ? scale_ub->data_ptr<float>()
                                                  : nullptr,
-                            var_epsilon, hidden_size,
+                            var_epsilon, hidden_size, input_stride,
                             has_residual ? residual->data_ptr<scalar_in_t>()
                                          : nullptr,
                             scales.stride(1));
@@ -246,7 +264,9 @@ void rms_norm_per_block_quant(torch::Tensor& out, torch::Tensor const& input,
                                         ? c10::ScalarType::Float8_e4m3fn
                                         : c10::ScalarType::Float8_e4m3fnuz;
   TORCH_CHECK(out.dtype() == kFp8Type || out.dtype() == torch::kInt8);
-  TORCH_CHECK(out.is_contiguous() && input.is_contiguous());
+  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK(input.stride(-1) == 1,
+              "Input must be contiguous in the last dimension");
 
   if (scale_ub.has_value()) {
     TORCH_CHECK(out.dtype() == kFp8Type);
@@ -255,6 +275,7 @@ void rms_norm_per_block_quant(torch::Tensor& out, torch::Tensor const& input,
   TORCH_CHECK(scales.dtype() == torch::kFloat32);
   if (residual) {
     TORCH_CHECK(residual->scalar_type() == input.scalar_type());
+    TORCH_CHECK(residual->is_contiguous());
   }
 
   TORCH_CHECK(group_size == 128 || group_size == 64,

csrc/quantization/fused_kernels/layernorm_utils.cuh

@@ -16,14 +16,17 @@ namespace vllm {
 // has_residual must be true, if residual is not a nullptr
 template <typename scalar_t, bool has_residual = false>
 __device__ void compute_rms(float* rms, scalar_t const* __restrict__ input,
-                            int32_t const hidden_size, float const epsilon,
+                            int32_t const hidden_size,
+                            int32_t const input_stride, float const epsilon,
                             scalar_t const* __restrict__ residual = nullptr) {
+  int64_t const input_token_offset =
+      blockIdx.x * static_cast<int64_t>(input_stride);
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
   // sum of squares
   float ss = 0.0f;
 
   for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
-    float x = static_cast<float>(input[token_offset + i]);
+    float x = static_cast<float>(input[input_token_offset + i]);
     if constexpr (has_residual) {
       x += static_cast<float>(residual[token_offset + i]);
     }
@@ -73,15 +76,20 @@ __device__ void compute_dynamic_per_token_scales(
     float* __restrict__ token_scale, float* __restrict__ all_token_scales,
     scalar_t const* __restrict__ input, scalar_t const* __restrict__ weight,
     float const rms, float const* __restrict__ scale_ub,
-    int32_t const hidden_size, scalar_t const* __restrict__ residual = nullptr,
+    int32_t const hidden_size, int32_t const input_stride,
+    scalar_t const* __restrict__ residual = nullptr,
     int32_t const group_size = 0, int64_t outer_scale_stride = 1) {
   float block_absmax_val_maybe = 0.0f;
   constexpr scalar_out_t qmax{quant_type_max_v<scalar_out_t>};
   __syncthreads();
+
+  int64_t const input_token_offset =
+      blockIdx.x * static_cast<int64_t>(input_stride);
+  int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
+
   if (group_size > 0) {
-    __shared__ float s_max_vals[1024];
-    int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
     int64_t num_groups = hidden_size / group_size;
+    __shared__ float s_max_vals[1024];
     int64_t const threads_per_group = blockDim.x / num_groups;
     int64_t const thread_in_group = threadIdx.x % threads_per_group;
     int64_t const group_offset = threadIdx.x / threads_per_group * group_size;
@@ -89,7 +97,7 @@ __device__ void compute_dynamic_per_token_scales(
     int64_t const thread_end =
         min(group_offset + group_size, static_cast<int64_t>(hidden_size));
     for (auto i = thread_offset; i < thread_end; i += threads_per_group) {
-      float x = static_cast<float>(input[token_offset + i]);
+      float x = static_cast<float>(input[input_token_offset + i]);
       if constexpr (has_residual) {
         x += static_cast<float>(residual[token_offset + i]);
       }
@@ -144,10 +152,8 @@ __device__ void compute_dynamic_per_token_scales(
     }
     __syncthreads();
   } else {
-    int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
-
     for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
-      float x = static_cast<float>(input[token_offset + i]);
+      float x = static_cast<float>(input[input_token_offset + i]);
       if constexpr (has_residual) {
         x += static_cast<float>(residual[token_offset + i]);
       }
@@ -185,12 +191,15 @@ template <typename scalar_t, typename scalar_out_t, bool is_scale_inverted,
 __device__ void norm_and_quant(
     scalar_out_t* __restrict__ output, scalar_t const* __restrict__ input,
     scalar_t const* __restrict__ weight, float const rms, float* const scale,
-    int32_t const hidden_size, scalar_t* __restrict__ residual = nullptr,
-    int32_t const group_size = 0, int64_t outer_scale_stride = 1) {
+    int32_t const hidden_size, int32_t const input_stride,
+    scalar_t* __restrict__ residual = nullptr, int32_t const group_size = 0,
+    int64_t outer_scale_stride = 1) {
+  int64_t const input_token_offset =
+      blockIdx.x * static_cast<int64_t>(input_stride);
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
 
   for (auto i = threadIdx.x; i < hidden_size; i += blockDim.x) {
-    float x = static_cast<float>(input[token_offset + i]);
+    float x = static_cast<float>(input[input_token_offset + i]);
     if constexpr (has_residual) {
       x += static_cast<float>(residual[token_offset + i]);
       residual[token_offset + i] = static_cast<scalar_t>(x);
@@ -224,13 +233,16 @@ namespace vectorized {
 // hidden_size must be a multiple of 4
 template <typename scalar_t, bool has_residual = false>
 __device__ void compute_rms(float* rms, scalar_t const* __restrict__ input,
-                            int32_t const hidden_size, float const epsilon,
+                            int32_t const hidden_size,
+                            int32_t const input_stride, float const epsilon,
                             scalar_t const* __restrict__ residual = nullptr) {
+  int64_t const input_token_offset =
+      blockIdx.x * static_cast<int64_t>(input_stride);
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
 
   // Vectorized input/output to better utilize memory bandwidth.
   vec4_t<scalar_t> const* vec_input =
-      reinterpret_cast<vec4_t<scalar_t> const*>(&input[token_offset]);
+      reinterpret_cast<vec4_t<scalar_t> const*>(&input[input_token_offset]);
   vec4_t<scalar_t> const* vec_residual = nullptr;
   if constexpr (has_residual) {
     vec_residual =
@@ -288,7 +300,8 @@ __device__ void compute_dynamic_per_token_scales(
     float* __restrict__ token_scale, float* __restrict__ all_token_scales,
     scalar_t const* __restrict__ input, scalar_t const* __restrict__ weight,
     float const rms, float const* __restrict__ scale_ub,
-    int32_t const hidden_size, scalar_t const* __restrict__ residual = nullptr,
+    int32_t const hidden_size, int32_t const input_stride,
+    scalar_t const* __restrict__ residual = nullptr,
     int64_t outer_scale_stride = 1) {
   constexpr scalar_out_t qmax{quant_type_max_v<scalar_out_t>};
 
@@ -300,10 +313,13 @@ __device__ void compute_dynamic_per_token_scales(
   vec4_t<scalar_t> const* vec_weight = nullptr;
   vec4_t<scalar_t> const* vec_residual = nullptr;
 
+  int64_t const input_token_offset =
+      blockIdx.x * static_cast<int64_t>(input_stride);
+  int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
+
   if constexpr (group_size > 0) {
     __shared__ float s_max_vals[1024];
 
-    int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
     int64_t const num_groups = hidden_size / group_size;
     int64_t const threads_per_group = blockDim.x / num_groups;
     int64_t const thread_in_group = threadIdx.x % threads_per_group;
@@ -312,7 +328,8 @@ __device__ void compute_dynamic_per_token_scales(
     int64_t const thread_offset = group_offset + thread_in_group;
     int64_t const thread_end = min(group_offset + (group_size >> 2),
                                    static_cast<int64_t>(hidden_size >> 2));
-    vec_input = reinterpret_cast<vec4_t<scalar_t> const*>(&input[token_offset]);
+    vec_input =
+        reinterpret_cast<vec4_t<scalar_t> const*>(&input[input_token_offset]);
     vec_weight = reinterpret_cast<vec4_t<scalar_t> const*>(weight);
     if constexpr (has_residual) {
       vec_residual =
@@ -396,8 +413,8 @@ __device__ void compute_dynamic_per_token_scales(
     __syncthreads();
 
   } else {
-    int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
-    vec_input = reinterpret_cast<vec4_t<scalar_t> const*>(&input[token_offset]);
+    vec_input =
+        reinterpret_cast<vec4_t<scalar_t> const*>(&input[input_token_offset]);
     vec_weight = reinterpret_cast<vec4_t<scalar_t> const*>(weight);
     if constexpr (has_residual) {
       vec_residual =
@@ -462,18 +479,18 @@ __device__ void compute_dynamic_per_token_scales(
 template <typename scalar_t, typename scalar_out_t, bool is_scale_inverted,
           bool has_residual = false, bool is_scale_transposed = false,
           int32_t group_size = 0>
-__device__ void norm_and_quant(scalar_out_t* __restrict__ output,
-                               scalar_t const* __restrict__ input,
-                               scalar_t const* __restrict__ weight,
-                               float const rms, float* const scale,
-                               int32_t const hidden_size,
-                               scalar_t* __restrict__ residual = nullptr,
-                               int64_t outer_scale_stride = 1) {
+__device__ void norm_and_quant(
+    scalar_out_t* __restrict__ output, scalar_t const* __restrict__ input,
+    scalar_t const* __restrict__ weight, float const rms, float* const scale,
+    int32_t const hidden_size, int32_t const input_stride,
+    scalar_t* __restrict__ residual = nullptr, int64_t outer_scale_stride = 1) {
+  int64_t const input_token_offset =
+      blockIdx.x * static_cast<int64_t>(input_stride);
   int64_t const token_offset = blockIdx.x * static_cast<int64_t>(hidden_size);
 
   // Vectorized input/output/weight/residual to better utilize memory bandwidth.
   vec4_t<scalar_t> const* vec_input =
-      reinterpret_cast<vec4_t<scalar_t> const*>(&input[token_offset]);
+      reinterpret_cast<vec4_t<scalar_t> const*>(&input[input_token_offset]);
   vec4_t<scalar_t> const* vec_weight =
       reinterpret_cast<vec4_t<scalar_t> const*>(weight);
   q8x4_t<scalar_out_t>* vec_output =

csrc/rocm/skinny_gemms.cu

@@ -12,6 +12,7 @@
 #include "../cuda_compat.h"
 #include "dispatch_utils.h"
 #include "quantization/w8a8/fp8/common.cuh"
+#include "core/batch_invariant.hpp"
 
 // TODO(rasmith): The kernels in this file are susceptible to integer overflow
 // issues, do not take strides, and are unable to handle PyTorch tensors that
@@ -1224,17 +1225,14 @@ torch::Tensor wvSplitK(const at::Tensor& in_a, const at::Tensor& in_b,
 #if defined(__gfx950__)
   #define WVSPLITKRC_1KPASS
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
-          int UNRL, int N, int GrpsShrB, int CHUNKK>
+          int UNRL, int N, int GrpsShrB, int CHUNKK, int DTRMNSTC>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
     __attribute__((amdgpu_waves_per_eu(1, 1)))
-    wvSplitKrc_(const int actlN, const int K, const int M, const int Bx,
-                const int By, const scalar_t* __restrict__ B,
-                const scalar_t* __restrict__ A,
-                const scalar_t* __restrict__ BIAS, float* glbl, scalar_t* C,
-                const int CuCount) {
-  // Use upper half of glbl buffer for atomic reduce counting
-  int* cntr = (int*)(&glbl[M * N]);
-
+    wvSplitKrc_(const int actlN, const int K, const int Kap, const int M,
+                const int Bx, const int By, const scalar_t* __restrict__ A,
+                const scalar_t* __restrict__ B,
+                const scalar_t* __restrict__ BIAS, float* glbl, int* cntr,
+                scalar_t* C, const int CuCount) {
   constexpr int NTILE = 16;
   constexpr int APAD = 1;
   constexpr int ASTRD = 64;
@@ -1425,11 +1423,11 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
           unsigned int kOffcp = min__(K - A_CHUNK, k_str + kOff);
           for (unsigned int n = 0; n < N; n += CHUNKK * sprdN) {
             __builtin_amdgcn_global_load_lds(
-                (int*)(&A[min__(
-                    K * actlN - A_CHUNK,
-                    kOffcp + K * (n / CHUNKK +
-                                  (N / CHUNKK) * (threadIdx.x / (64 / CHUNKK)) +
-                                  (threadIdx.y % sprdN)))]),
+                (int*)(&A[min__(Kap * actlN - A_CHUNK,
+                                kOffcp + Kap * (n / CHUNKK +
+                                                (N / CHUNKK) * (threadIdx.x /
+                                                                (64 / CHUNKK)) +
+                                                (threadIdx.y % sprdN)))]),
                 (int*)(&s[(k +
                            kFitPdd * ((n / CHUNKK) + (threadIdx.y % sprdN)))]),
                 16, 0, 0);
@@ -1533,45 +1531,98 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
     }
   }
 
+  union flt4 {
+    scalar8 s8;
+    float2 f2[2];
+    float4 f4;
+  };
   if (m + (threadIdx.x % 16) < M) {
     int my_cntr;
     int mindx = m + (threadIdx.x % 16);
     int g_mindx = m * 4 + (threadIdx.x % 64);  // coalesced atomic reduction
     scalar_t biases[N / NTILE / GrpsShrB][4] = {};
     // Atomic add the output, read biases
-    for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++)
-      for (uint32_t j = 0; j < 4; j++) {
-        // int nindx = (j + (threadIdx.x / 16) * 4) + nt * NTILE +
-        //             (N / GrpsShrB) * (threadIdx.y % GrpsShrB);
-        // int adr = mindx + M * nindx;
-        int g_nindx =
-            j + (nt * NTILE + (N / GrpsShrB) * (threadIdx.y % GrpsShrB)) / 4;
-        int g_adr = g_mindx + M * g_nindx * 4;
-        atomicAdd(&glbl[g_adr], sum4[nt][0][j]);
+    for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++) {
+      int g_nindx =
+          (nt * NTILE + (N / GrpsShrB) * (threadIdx.y % GrpsShrB)) / 4;
+      int g_adr = g_mindx * 4 + 0 + M * g_nindx * 4;
+      if (DTRMNSTC) {
+        flt4 flt4_ = {.s8 = sum4[nt][0]};
+        __hip_atomic_store((float2*)&glbl[g_adr + M * N * (m0 / Mmod)],
+                           flt4_.f2[0], __ATOMIC_RELAXED,
+                           __HIP_MEMORY_SCOPE_AGENT);
+        __hip_atomic_store((float2*)&glbl[g_adr + 2 + M * N * (m0 / Mmod)],
+                           flt4_.f2[1], __ATOMIC_RELAXED,
+                           __HIP_MEMORY_SCOPE_AGENT);
+      } else {
+        for (uint32_t j = 0; j < 4; j++)
+          atomicAdd((&glbl[g_adr + j]), sum4[nt][0][j]);
       }
+    }
+
+    __atomic_signal_fence(__ATOMIC_SEQ_CST);
+    asm volatile("s_waitcnt vmcnt(0)" ::: "memory");
+    __atomic_signal_fence(__ATOMIC_SEQ_CST);
+
     int nindx_ = (0 + (threadIdx.x / 16) * 4) + 0 * NTILE +
                  (N / GrpsShrB) * (threadIdx.y % GrpsShrB);
     int adr_ = mindx + M * nindx_ / 4;
-    // Update the complete counter
     my_cntr = atomicAdd(&cntr[adr_], 1);
-    float vals[N / NTILE / GrpsShrB][4] = {};
+
+    // make sure LDS is free for write out staging
+    if (DTRMNSTC) __syncthreads();
+
+    // Update the complete counter
+    flt4 vals[N / NTILE / GrpsShrB] = {};
     // If we're the last k-shard, read back the value and convert...
     if (my_cntr + 1 == k_rnd) {
-      if (BIAS)
-        for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++) {
-          for (uint32_t j = 0; j < 4; j++) {
-            int nindx = (j + (threadIdx.x / 16) * 4) + nt * NTILE +
-                        (N / GrpsShrB) * (threadIdx.y % GrpsShrB);
-            biases[nt][j] = BIAS[(mindx % Bx) + (nindx % By) * Bx];
+      cntr[adr_] = 0;  // clear for next round
+      if constexpr (DTRMNSTC) {
+  #pragma unroll
+        for (int ks = 0; ks < k_rnd; ks++) {
+          for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++) {
+            int g_nindx =
+                (nt * NTILE + (N / GrpsShrB) * (threadIdx.y % GrpsShrB)) / 4;
+            int g_adr = g_mindx * 4 + 0 + M * g_nindx * 4;
+            __builtin_amdgcn_global_load_lds(
+                (float4*)(&glbl[g_adr + M * N * ks]),
+                &(((float4*)s)[(threadIdx.y * THRDS) + ks * THRDS * 4 +
+                               nt * THRDS * 4 * k_rnd]),
+                16, 0, 0);
           }
         }
-      for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++) {
-        for (uint32_t j = 0; j < 4; j++) {
-          int g_nindx =
-              j + (nt * NTILE + (N / GrpsShrB) * (threadIdx.y % GrpsShrB)) / 4;
-          int g_adr = g_mindx + M * g_nindx * 4;
-          vals[nt][j] = glbl[g_adr];
+        if (BIAS)
+          for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++) {
+            for (uint32_t j = 0; j < 4; j++) {
+              int nindx = (j + (threadIdx.x / 16) * 4) + nt * NTILE +
+                          (N / GrpsShrB) * (threadIdx.y % GrpsShrB);
+              biases[nt][j] = BIAS[(mindx % Bx) + (nindx % By) * Bx];
+            }
+          }
+        asm volatile("s_waitcnt 0");
+        for (int ks = 0; ks < k_rnd; ks++) {
+          for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++) {
+            float4 eval = ((float4*)s)[(threadIdx.x + threadIdx.y * THRDS) +
+                                       ks * THRDS * 4 + nt * THRDS * 4 * k_rnd];
+            vals[nt].f4 += eval;
+          }
         }
+      } else {
+        for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++) {
+          int g_nindx =
+              (nt * NTILE + (N / GrpsShrB) * (threadIdx.y % GrpsShrB)) / 4;
+          int g_adr = g_mindx * 4 + 0 + M * g_nindx * 4;
+          vals[nt].f4 = *(float4*)(&glbl[g_adr]);
+          *(float4*)(&glbl[g_adr]) = {};  // clear out for next round
+        }
+        if (BIAS)
+          for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++) {
+            for (uint32_t j = 0; j < 4; j++) {
+              int nindx = (j + (threadIdx.x / 16) * 4) + nt * NTILE +
+                          (N / GrpsShrB) * (threadIdx.y % GrpsShrB);
+              biases[nt][j] = BIAS[(mindx % Bx) + (nindx % By) * Bx];
+            }
+          }
       }
       __builtin_amdgcn_sched_barrier(0);
       for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++) {
@@ -1581,11 +1632,11 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
           if (nindx < actlN) {
             int adr = mindx + M * nindx;
             if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
-              vals[nt][j] += __bfloat162float(biases[nt][j]);
-              C[adr] = __float2bfloat16(vals[nt][j]);
+              vals[nt].s8[j] += __bfloat162float(biases[nt][j]);
+              C[adr] = __float2bfloat16(vals[nt].s8[j]);
             } else {
-              vals[nt][j] += __half2float(biases[nt][j]);
-              C[adr] = __float2half(vals[nt][j]);
+              vals[nt].s8[j] += __half2float(biases[nt][j]);
+              C[adr] = __float2half(vals[nt].s8[j]);
             }
           }
         }
@@ -1604,21 +1655,25 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
 }
 #else   // !defined(__HIP__GFX9__) TODO: Add NAVI support
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
-          int UNRL, int N, int GrpsShrB, int CHUNKK>
-__global__ void wvSplitKrc_(const int actlN, const int K, const int M,
-                            const int Bx, const int By, const scalar_t* B,
-                            const scalar_t* __restrict__ A,
+          int UNRL, int N, int GrpsShrB, int CHUNKK, int DTRMNSTC>
+__global__ void wvSplitKrc_(const int actlN, const int K, const int Kap,
+                            const int M, const int Bx, const int By,
+                            const scalar_t* B, const scalar_t* __restrict__ A,
                             const scalar_t* __restrict__ BIAS, float* glbl,
-                            // int* cntr,
-                            scalar_t* C, const int CuCount){UNREACHABLE_CODE}
+                            int* cntr, scalar_t* C,
+                            const int CuCount){UNREACHABLE_CODE}
 #endif  // defined(__HIP__GFX9__) TODO: Add NAVI support
 
 torch::Tensor wvSplitKrc(const at::Tensor& in_a, const at::Tensor& in_b,
                          const std::optional<at::Tensor>& in_bias,
                          const int64_t CuCount) {
-  auto M_in = in_a.size(0);
-  auto N_in = in_b.size(0);
-  auto K_in = in_a.size(1);
+  int _DTRMNSTC = 1;  // vllm::vllm_is_batch_invariant();
+
+  auto M_in = in_b.size(0);
+  auto N_in = in_a.size(0);
+  auto K_in = in_b.size(1);
+  auto Kap_in = in_a.stride(0);
+
   auto Bx_in =
       (in_bias.has_value() && in_bias->numel() > 0)
           ? (in_bias->sizes().size() == 2) ? in_bias->size(1) : in_bias->size(0)
@@ -1635,13 +1690,9 @@ torch::Tensor wvSplitKrc(const at::Tensor& in_a, const at::Tensor& in_b,
 
   auto out_c = torch::empty(
       {N_in, M_in},
-      torch::TensorOptions().dtype(in_b.dtype()).device(in_b.device()));
+      torch::TensorOptions().dtype(in_a.dtype()).device(in_a.device()));
 
   auto N_p2 = 1U << (32 - __builtin_clz(N_in - 1));
-  auto axl_glbl = torch::empty(
-      {N_p2 + N_p2 / 4, M_in + M_in / 4},
-      torch::TensorOptions().dtype(torch::kFloat32).device(in_b.device()));
-  axl_glbl.zero_();  // disable for FAST_UNSAFE_RDC_INIT
 
   dim3 grid(CuCount);
 
@@ -1649,55 +1700,70 @@ torch::Tensor wvSplitKrc(const at::Tensor& in_a, const at::Tensor& in_b,
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   // const int max_lds_len = get_lds_size() / 2;
 
+  // With 64 Ms per CU (each of 4 SIMDs working on a 16x16 tile),
+  // and each working on a 512-shard of K, how many CUs would we need?
+  int rndup_cus = ((M_in + 64 - 1) / 64) * ((K_in + 512 - 1) / 512);
+
+  // How many of 4 waves in a group can work on same 16 Ms at same time? First
+  // try to maximize this. This reduces the Ms each group works on, i.e.
+  // increasing the number of CUs needed.
+  int GrpsShrB = min(N_p2 / 16, 4);
+
+  // Given the above, how many CUs would we need?
+  int CuNeeded = rndup_cus * GrpsShrB;
+
+  if (CuNeeded > CuCount) throw std::runtime_error("Invalid wvSplitKrc size");
+
+  // Can we increase SplitK by shrinking the K-shared to 256?
+  int chunkk = (CuNeeded * 2 <= CuCount) ? 2 : 1;
+
+  static torch::Tensor axl_glbl =
+      torch::zeros(
+          128 * 1024 * (_DTRMNSTC ? 12 : 1),
+          torch::TensorOptions().dtype(torch::kFloat32).device(in_a.device()))
+          .detach();
+  static torch::Tensor axl_cntr =
+      torch::zeros(
+          128 * 1024 * (_DTRMNSTC ? 12 : 1) / 4,
+          torch::TensorOptions().dtype(torch::kInt).device(in_a.device()))
+          .detach();
+  auto glbl = axl_glbl.data_ptr<float>();
+  auto cntr = axl_cntr.data_ptr<int>();
+
 #define WVSPLITKrc(_N, _GrpsShrB, _CHUNKK)                                     \
   {                                                                            \
     dim3 block(64, 4);                                                         \
-    wvSplitKrc_<fptype, 64, 16, 4, 8, 1, _N, _GrpsShrB, _CHUNKK>               \
-        <<<grid, block, 0, stream>>>(N_in, K_in, M_in, Bx_in, By_in, af4, bf4, \
-                                     biasf4, glbl, c, CuCount);                \
+    if (_DTRMNSTC)                                                             \
+      wvSplitKrc_<fptype, 64, 16, 4, 8, 1, _N, _GrpsShrB, _CHUNKK, 1>          \
+          <<<grid, block, 0, stream>>>(N_in, K_in, Kap_in, M_in, Bx_in, By_in, \
+                                       af4, bf4, biasf4, glbl, cntr, c,        \
+                                       CuCount);                               \
+    else                                                                       \
+      wvSplitKrc_<fptype, 64, 16, 4, 8, 1, _N, _GrpsShrB, _CHUNKK, 0>          \
+          <<<grid, block, 0, stream>>>(N_in, K_in, Kap_in, M_in, Bx_in, By_in, \
+                                       af4, bf4, biasf4, glbl, cntr, c,        \
+                                       CuCount);                               \
   }
 
-  AT_DISPATCH_REDUCED_FLOATING_TYPES(in_b.scalar_type(), "wvSplitKrc", [&] {
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(in_a.scalar_type(), "wvSplitKrc", [&] {
     using fptype = typename scalar<scalar_t>::type;
-    fptype* af4 = reinterpret_cast<fptype*>(in_a.data_ptr());
+    const fptype* af4 = reinterpret_cast<const fptype*>(in_a.data_ptr());
     const fptype* bf4 = reinterpret_cast<const fptype*>(in_b.data_ptr());
     const fptype* biasf4 =
         (in_bias.has_value() && in_bias->numel() > 0)
             ? reinterpret_cast<const fptype*>(in_bias->data_ptr())
             : nullptr;
     fptype* c = reinterpret_cast<fptype*>(out_c.data_ptr());
-    auto glbl = axl_glbl.data_ptr<float>();
-
-    // With 64 Ms per CU (each of 4 SIMDs working on a 16x16 tile),
-    // and each working on a 512-shard of K, how many CUs would we need?
-    int rndup_cus = ((M_in + 64 - 1) / 64) * ((K_in + 512 - 1) / 512);
-
-    // How many of 4 waves in a group can work on same 16 Ms at same time? First
-    // try to maximize this. This reduces the Ms each group works on, i.e.
-    // increasing the number of CUs needed.
-    int GrpsShrB = min(N_p2 / 16, 4);
-
-    // Given the above, how many CUs would we need?
-    int CuNeeded = rndup_cus * GrpsShrB;
-
-    if (CuNeeded > CuCount) std::runtime_error("Invalid wvSplitKrc size");
-
-    // Can we increase SplitK by shrinking the K-shared to 256?
-    int chunkk = (CuNeeded * 2 <= CuCount) ? 2 : 1;
 
     switch (N_p2) {
       case 16:
         WVSPLITKrc(16, 1, 1) break;
       case 32:
-        if (chunkk == 2)
-          WVSPLITKrc(32, 2, 2) else if (chunkk == 1) WVSPLITKrc(32, 2, 1) break;
+        if (chunkk == 2) WVSPLITKrc(32, 2, 2) else WVSPLITKrc(32, 2, 1) break;
       case 64:
-        if (chunkk == 2)
-          WVSPLITKrc(64, 4, 2) else if (chunkk == 1) WVSPLITKrc(64, 4, 1) break;
+        if (chunkk == 2) WVSPLITKrc(64, 4, 2) else WVSPLITKrc(64, 4, 1) break;
       case 128:
-        if (chunkk == 2)
-          WVSPLITKrc(128, 4, 2) else if (chunkk == 1)
-              WVSPLITKrc(128, 4, 1) break;
+        if (chunkk == 2) WVSPLITKrc(128, 4, 2) else WVSPLITKrc(128, 4, 1) break;
       default:
         throw std::runtime_error(
             "Unsupported N value: " + std::to_string(M_in) + "," +

csrc/torch_bindings.cpp

@@ -802,6 +802,10 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
   cache_ops.impl("indexer_k_quant_and_cache", torch::kCUDA,
                  &indexer_k_quant_and_cache);
 
+  cache_ops.def(
+      "concat_mla_q(Tensor ql_nope, Tensor q_pe, Tensor! q_out) -> ()");
+  cache_ops.impl("concat_mla_q", torch::kCUDA, &concat_mla_q);
+
   cache_ops.def(
       "cp_gather_indexer_k_quant_cache(Tensor kv_cache, Tensor! dst_k, Tensor! "
       "dst_scale, Tensor block_table, Tensor cu_seq_lens) -> ()");

docs/benchmarking/cli.md

@@ -18,7 +18,7 @@ th {
 </style>
 
 | Dataset | Online | Offline | Data Path |
-|---------|--------|---------|-----------|
+| ------- | ------ | ------- | --------- |
 | ShareGPT | ✅ | ✅ | `wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json` |
 | ShareGPT4V (Image) | ✅ | ✅ | `wget https://huggingface.co/datasets/Lin-Chen/ShareGPT4V/resolve/main/sharegpt4v_instruct_gpt4-vision_cap100k.json`<br>Note that the images need to be downloaded separately. For example, to download COCO's 2017 Train images:<br>`wget http://images.cocodataset.org/zips/train2017.zip` |
 | ShareGPT4Video (Video) | ✅ | ✅ | `git clone https://huggingface.co/datasets/ShareGPT4Video/ShareGPT4Video` |
@@ -383,14 +383,14 @@ The `--burstiness` parameter mathematically controls request arrival patterns us
 
 Load Pattern Recommendations by Use Case:
 
-| Use Case           | Burstiness   | Request Rate    | Max Concurrency | Description                                               |
-| ---                | ---          | ---             | ---             | ---                                                       |
+| Use Case           | Burstiness   | Request Rate    | Max Concurrency | Description                                                                        |
+| ---                | ---          | ---             | ---             | ---                                                                                |
 | Maximum Throughput | N/A          | Infinite        | Limited         | **Most common**: Simulates load balancer/gateway limits with unlimited user demand |
-| Realistic Testing  | 1.0          | Moderate (5-20) | Infinite        | Natural Poisson traffic patterns for baseline performance |
-| Stress Testing     | 0.1-0.5      | High (20-100)   | Infinite        | Challenging burst patterns to test resilience             |
-| Latency Profiling  | 2.0-5.0      | Low (1-10)      | Infinite        | Uniform load for consistent timing analysis               |
-| Capacity Planning  | 1.0          | Variable        | Limited         | Test resource limits with realistic constraints           |
-| SLA Validation     | 1.0          | Target rate     | SLA limit       | Production-like constraints for compliance testing        |
+| Realistic Testing  | 1.0          | Moderate (5-20) | Infinite        | Natural Poisson traffic patterns for baseline performance                          |
+| Stress Testing     | 0.1-0.5      | High (20-100)   | Infinite        | Challenging burst patterns to test resilience                                      |
+| Latency Profiling  | 2.0-5.0      | Low (1-10)      | Infinite        | Uniform load for consistent timing analysis                                        |
+| Capacity Planning  | 1.0          | Variable        | Limited         | Test resource limits with realistic constraints                                    |
+| SLA Validation     | 1.0          | Target rate     | SLA limit       | Production-like constraints for compliance testing                                 |
 
 These load patterns help evaluate different aspects of your vLLM deployment, from basic performance characteristics to resilience under challenging traffic conditions.
 
@@ -941,7 +941,7 @@ Benchmark per-stage latency of the multimodal (MM) input processor pipeline, inc
 The benchmark measures the following stages for each request:
 
 | Stage | Description |
-|-------|-------------|
+| ----- | ----------- |
 | `get_mm_hashes_secs` | Time spent hashing multimodal inputs |
 | `get_cache_missing_items_secs` | Time spent looking up the processor cache |
 | `apply_hf_processor_secs` | Time spent in the HuggingFace processor |

docs/benchmarking/dashboard.md

@@ -39,6 +39,12 @@ When run, benchmark script generates results under **benchmark/results** folder,
 - `THROUGHPUT_JSON`: JSON file to use for the throughout tests. Default value is empty string (use default file).
 - `REMOTE_HOST`: IP for the remote vLLM service to benchmark. Default value is empty string.
 - `REMOTE_PORT`: Port for the remote vLLM service to benchmark. Default value is empty string.
+- `PROMPTS_PER_CONCURRENCY`: Multiplier to compute `num_prompts` for serving tests (`num_prompts = max_concurrency × value`). Overrides JSON `num_prompts`. Default is NULL.
+- `ENABLE_ADAPTIVE_CONCURRENCY`: set the value to '1' to enable adaptive SLA-based concurrency search after the static serving max_concurrency sweep. Default value is 0.
+- `SLA_TTFT_MS`: default TTFT SLA threshold in milliseconds for adaptive concurrency search. Default value is 3000.
+- `SLA_TPOT_MS`: default TPOT SLA threshold in milliseconds for adaptive concurrency search. Default value is 100.
+- `ADAPTIVE_MAX_PROBES`: maximum number of extra adaptive search probes. Default value is 8.
+- `ADAPTIVE_MAX_CONCURRENCY`: maximum allowed concurrency during adaptive search. Default value is 1024.
 
 ### Visualization
 
@@ -60,12 +66,12 @@ Here is an example using the script to compare result_a and result_b with max co
 
 ***Output Tput (tok/s) — Model : [ meta-llama/Llama-3.1-8B-Instruct ] , Dataset Name : [ random ] , Input Len : [ 2048.0 ] , Output Len : [ 2048.0 ]***
 
-|    | # of max concurrency | qps  | results_a/benchmark_results.json | results_b/benchmark_results.json | perf_ratio        |
-|----|------|-----|-----------|----------|----------|
-| 0  | 12 | inf | 24.98   | 186.03 |  7.45 |
-| 1  | 16 | inf|  25.49  | 246.92 | 9.69 |
-| 2  | 24 | inf| 27.74  | 293.34 |  10.57 |
-| 3  | 32 | inf| 28.61  |306.69 | 10.72 |
+| | # of max concurrency | qps | results_a/benchmark_results.json | results_b/benchmark_results.json | perf_ratio |
+| | -------------------- | --- | -------------------------------- | -------------------------------- | ---------- |
+| 0 | 12 | inf | 24.98 | 186.03 |  7.45 |
+| 1 | 16 | inf |  25.49 | 246.92 | 9.69 |
+| 2 | 24 | inf | 27.74 | 293.34 |  10.57 |
+| 3 | 32 | inf | 28.61 |306.69 | 10.72 |
 
 ***compare-json-results.py – Command-Line Parameters***  
 

docs/cli/bench/mm_processor.md

@@ -29,7 +29,7 @@ vllm bench mm-processor \
 ## Measured Stages
 
 | Stage | Description |
-|-------|-------------|
+| ----- | ----------- |
 | `get_mm_hashes_secs` | Time spent hashing multimodal inputs |
 | `get_cache_missing_items_secs` | Time spent looking up the processor cache |
 | `apply_hf_processor_secs` | Time spent in the HuggingFace processor |

docs/cli/json_tip.inc.md

@@ -1,3 +1,4 @@
+<!-- markdownlint-disable MD041 -->
 When passing JSON CLI arguments, the following sets of arguments are equivalent:
 
 - `--json-arg '{"key1": "value1", "key2": {"key3": "value2"}}'`
@@ -6,4 +7,4 @@ When passing JSON CLI arguments, the following sets of arguments are equivalent:
 Additionally, list elements can be passed individually using `+`:
 
 - `--json-arg '{"key4": ["value3", "value4", "value5"]}'`
-- `--json-arg.key4+ value3 --json-arg.key4+='value4,value5'`
+- `--json-arg.key4+ value3 --json-arg.key4+='value4,value5'`

docs/configuration/optimization.md

@@ -293,7 +293,7 @@ llm = LLM(
 Based on the configuration, the content of the multi-modal caches on `P0` and `P1` are as follows:
 
 | mm_processor_cache_type | Cache Type | `P0` Cache | `P1` Engine Cache | `P1` Worker Cache | Max. Memory |
-|-------------------|-------------|------------|------------|-------------|-------------|
+| ----------------- | ----------- | ---------- | ---------- | ----------- | ----------- |
 | lru | Processor Caching | K + V | N/A | N/A | `mm_processor_cache_gb * data_parallel_size` |
 | lru | Key-Replicated Caching | K | K + V | N/A | `mm_processor_cache_gb * api_server_count` |
 | shm | Shared Memory Caching | K | N/A | V | `mm_processor_cache_gb * api_server_count` |

docs/contributing/README.md

@@ -94,7 +94,6 @@ vLLM's `pre-commit` hooks will now run automatically every time you commit.
     Some `pre-commit` hooks only run in CI. If you need to, you can run them locally with:
 
     ```bash
-    pre-commit run --hook-stage manual markdownlint
     pre-commit run --hook-stage manual mypy-3.10
     ```
 

docs/contributing/ci/update_pytorch_version.md

@@ -66,12 +66,12 @@ This complicates the process as we cannot use the out-of-the-box
 - Important indexes at the moment include:
 
 | Platform | `--extra-index-url` |
-|----------|-----------------|
-| CUDA 12.8| [https://download.pytorch.org/whl/cu128](https://download.pytorch.org/whl/cu128)|
-| CPU      | [https://download.pytorch.org/whl/cpu](https://download.pytorch.org/whl/cpu)|
+| -------- | ------------------- |
+| CUDA 12.8 | [https://download.pytorch.org/whl/cu128](https://download.pytorch.org/whl/cu128) |
+| CPU | [https://download.pytorch.org/whl/cpu](https://download.pytorch.org/whl/cpu) |
 | ROCm 6.2 | [https://download.pytorch.org/whl/rocm6.2.4](https://download.pytorch.org/whl/rocm6.2.4) |
 | ROCm 6.3 | [https://download.pytorch.org/whl/rocm6.3](https://download.pytorch.org/whl/rocm6.3) |
-| XPU      | [https://download.pytorch.org/whl/xpu](https://download.pytorch.org/whl/xpu) |
+| XPU | [https://download.pytorch.org/whl/xpu](https://download.pytorch.org/whl/xpu) |
 
 - Update the below files to match the CUDA version from step 1. This makes sure that the release vLLM wheel is tested on CI.
     - `.buildkite/release-pipeline.yaml`

docs/contributing/deprecation_policy.md

@@ -66,7 +66,7 @@ stages will be removed.
 Assume a feature is deprecated in `v0.9.0`.
 
 | Release       | Status                                                                                          |
-|---------------|-------------------------------------------------------------------------------------------------|
+| ------------- | ----------------------------------------------------------------------------------------------- |
 | `v0.9.0`      | Feature is deprecated with clear removal version listed.                                        |
 | `v0.10.0`     | Feature is now off by default, throws an error when used, and can be re-enabled for legacy use. |
 | `v0.11.0`     | Feature is removed.                                                                             |

docs/deployment/frameworks/helm.md

@@ -49,7 +49,7 @@ chart **including persistent volumes** and deletes the release.
 The following table describes configurable parameters of the chart in `values.yaml`:
 
 | Key | Type | Default | Description |
-|-----|------|---------|-------------|
+| --- | ---- | ------- | ----------- |
 | autoscaling | object | {"enabled":false,"maxReplicas":100,"minReplicas":1,"targetCPUUtilizationPercentage":80} | Autoscaling configuration |
 | autoscaling.enabled | bool | false | Enable autoscaling |
 | autoscaling.maxReplicas | int | 100 | Maximum replicas |

docs/deployment/integrations/kuberay.md

@@ -6,7 +6,7 @@ A Ray cluster can be declared in YAML, and the operator then handles pod schedul
 ## Why KubeRay instead of manual scripts?
 
 | Feature | Manual scripts | KubeRay |
-|---------|-----------------------------------------------------------|---------|
+| ------- | --------------------------------------------------------- | ------- |
 | Cluster bootstrap | Manually SSH into every node and run a script | One command to create or update the whole cluster: `kubectl apply -f cluster.yaml` |
 | Autoscaling | Manual | Automatically patches CRDs for adjusting cluster size |
 | Upgrades | Tear down & re-create manually | Blue/green deployment updates supported |

docs/design/arch_overview.md

@@ -119,7 +119,7 @@ The code can be found in [vllm/v1/engine/coordinator.py](../../vllm/v1/engine/co
 For a deployment with `N` GPUs, `TP` tensor parallel size, `DP` data parallel size, and `A` API server count:
 
 | Process Type | Count | Notes |
-|---|---|---|
+| - | - | - |
 | API Server | `A` (default `DP`) | Handles HTTP requests and input processing |
 | Engine Core | `DP` (default 1) | Scheduler and KV cache management |
 | GPU Worker | `N` (= `DP x PP x TP`) | One per GPU, executes model forward passes |

docs/design/attention_backends.md

@@ -101,7 +101,7 @@ Priority is **1 = highest** (tried first).
 **Blackwell (SM 10.x):**
 
 | Priority | Backend |
-|----------|---------|
+| -------- | ------- |
 | 1 | `FLASHINFER` |
 | 2 | `FLASH_ATTN` |
 | 3 | `TRITON_ATTN` |
@@ -110,7 +110,7 @@ Priority is **1 = highest** (tried first).
 **Ampere/Hopper (SM 8.x-9.x):**
 
 | Priority | Backend |
-|----------|---------|
+| -------- | ------- |
 | 1 | `FLASH_ATTN` |
 | 2 | `FLASHINFER` |
 | 3 | `TRITON_ATTN` |
@@ -121,7 +121,7 @@ Priority is **1 = highest** (tried first).
 **Blackwell (SM 10.x):**
 
 | Priority | Backend |
-|----------|---------|
+| -------- | ------- |
 | 1 | `FLASHINFER_MLA` |
 | 2 | `CUTLASS_MLA` |
 | 3 | `FLASH_ATTN_MLA` |
@@ -133,7 +133,7 @@ Priority is **1 = highest** (tried first).
 **Ampere/Hopper (SM 8.x-9.x):**
 
 | Priority | Backend |
-|----------|---------|
+| -------- | ------- |
 | 1 | `FLASH_ATTN_MLA` |
 | 2 | `FLASHMLA` |
 | 3 | `FLASHINFER_MLA` |
@@ -145,7 +145,7 @@ Priority is **1 = highest** (tried first).
 ## Legend
 
 | Column | Description |
-|--------|-------------|
+| ------ | ----------- |
 | **Dtypes** | Supported model data types (fp16, bf16, fp32) |
 | **KV Dtypes** | Supported KV cache data types (`auto`, `fp8`, `fp8_e4m3`, etc.) |
 | **Block Sizes** | Supported KV cache block sizes (%N means multiples of N) |
@@ -162,20 +162,20 @@ Priority is **1 = highest** (tried first).
 ## Standard Attention (MHA, MQA, GQA) Backends
 
 | Backend | Version | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | MM Prefix | DCP | Attention Types | Compute Cap. |
-|---------|---------|--------|-----------|-------------|------------|------|-----------|-----|-----------------|--------------|
-| `CPU_ATTN` |  | fp16, bf16, fp32 | `auto` | Any | 32, 64, 80, 96, 112, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | All | N/A |
+| ------- | ------- | ------ | --------- | ----------- | ---------- | ---- | --------- | --- | --------------- | ------------ |
+| `CPU_ATTN` | | fp16, bf16, fp32 | `auto` | Any | 32, 64, 80, 96, 112, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | All | N/A |
 | `FLASHINFER` | Native† | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ❌ | ❌ | ✅ | Decoder | 7.x-9.x |
 | `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ✅ | ❌ | ✅ | Decoder | 10.x |
 | `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥8.0 |
 | `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ❌ | ✅ | All | 9.x |
 | `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥10.0 |
-| `FLASH_ATTN_DIFFKV` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ✅ | Decoder | Any |
-| `FLEX_ATTENTION` |  | fp16, bf16, fp32 | `auto`, `bfloat16` | Any | Any | ❌ | ✅ | ❌ | Decoder, Encoder Only | Any |
-| `ROCM_AITER_FA` |  | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32 | 64, 128, 256 | ❌ | ❌ | ❌ | Decoder, Enc-Dec | N/A |
-| `ROCM_AITER_UNIFIED_ATTN` |  | fp16, bf16 | `auto` | %16 | Any | ✅ | ✅ | ❌ | All | N/A |
-| `ROCM_ATTN` |  | fp16, bf16, fp32 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 544 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ✅ | ✅ | ❌ | All | N/A |
-| `TREE_ATTN` |  | fp16, bf16 | `auto` | %16 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | Any |
-| `TRITON_ATTN` |  | fp16, bf16, fp32 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
+| `FLASH_ATTN_DIFFKV` | | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ✅ | Decoder | Any |
+| `FLEX_ATTENTION` | | fp16, bf16, fp32 | `auto`, `bfloat16` | Any | Any | ❌ | ✅ | ❌ | Decoder, Encoder Only | Any |
+| `ROCM_AITER_FA` | | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32 | 64, 128, 256 | ❌ | ❌ | ❌ | Decoder, Enc-Dec | N/A |
+| `ROCM_AITER_UNIFIED_ATTN` | | fp16, bf16 | `auto` | %16 | Any | ✅ | ✅ | ❌ | All | N/A |
+| `ROCM_ATTN` | | fp16, bf16, fp32 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ✅ | ✅ | ❌ | All | N/A |
+| `TREE_ATTN` | | fp16, bf16 | `auto` | %16 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | Any |
+| `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
 
 > **†** FlashInfer uses TRTLLM attention on Blackwell (SM100), which supports sinks. Disable via `--attention-config.use_trtllm_attention=0`.
 >
@@ -191,10 +191,10 @@ The prefill backend is selected at runtime based on hardware and
 configuration.
 
 | Backend | Description | Compute Cap. | Enable | Disable | Notes |
-|---------|-------------|--------------|--------|---------|-------|
+| ------- | ----------- | ------------ | ------ | ------- | ----- |
 | TRT-LLM Ragged‡ | TensorRT-LLM ragged attention | 10.x | Default on SM100 | `-ac.use_trtllm_ragged_deepseek_prefill=0` | DeepSeek R1 dims only |
 | FlashInfer | FlashInfer CUTLASS backend | 10.x | `-ac.disable_flashinfer_prefill=0` | `-ac.disable_flashinfer_prefill=1` | DeepSeek R1 dims only |
-| cuDNN | cuDNN-based attention | 10.x | `-ac.use_cudnn_prefill=1` | `-ac.use_cudnn_prefill=0` |  |
+| cuDNN | cuDNN-based attention | 10.x | `-ac.use_cudnn_prefill=1` | `-ac.use_cudnn_prefill=0` | |
 | FlashAttention | FlashAttention varlen (FA2/FA3) | Any | Default fallback | Use other backends | FA3 on SM90, FA2 otherwise |
 
 > **‡** TRT-LLM Ragged is the default on Blackwell (SM100).
@@ -203,14 +203,15 @@ configuration.
 ### Decode Backends
 
 | Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Sparse | MM Prefix | DCP | Attention Types | Compute Cap. |
-|---------|--------|-----------|-------------|------------|------|--------|-----------|-----|-----------------|--------------|
+| ------- | ------ | --------- | ----------- | ---------- | ---- | ------ | --------- | --- | --------------- | ------------ |
 | `CUTLASS_MLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 128 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
 | `FLASHINFER_MLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | 10.x |
-| `FLASHINFER_MLA_SPARSE` | fp16, bf16 | `auto`, `bfloat16` | 32, 64 | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 10.x |
+| `FLASHINFER_MLA_SPARSE` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 10.x |
 | `FLASHMLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 64 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x-10.x |
 | `FLASHMLA_SPARSE` | bf16 | `auto`, `bfloat16`, `fp8_ds_mla` | 64 | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
 | `FLASH_ATTN_MLA` | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x |
 | `ROCM_AITER_MLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 1 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
 | `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto`, `bfloat16` | 1 | Any | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
 | `ROCM_AITER_TRITON_MLA` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
-| `TRITON_MLA` | fp16, bf16 | `auto`, `bfloat16` | Any | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
+| `TRITON_MLA` | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
+| `XPU_MLA_SPARSE` | fp16, bf16 | `auto`, `bfloat16` | Any | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | Any |

docs/design/cuda_graphs.md

@@ -174,18 +174,18 @@ Suppose we have hybrid attention backends (e.g., in mamba mixer models). In that
 The following table lists backends that support full CUDA Graphs at the time of writing.
 
 | Attention Backend | cudagraph_support | Comments |
-|:---|:---|:---|
+| :---------------- | :---------------- | :------- |
 | FlashAttention v2 | `UNIFORM_BATCH` | Actually `ALWAYS` but workaround to fallback to `FULL_AND_PIECEWISE` for performance reason |
 | FlashAttention v3 | `ALWAYS` | has unified routine for both batches, so `FULL` mode is good |
 | Triton Attention | `ALWAYS` | prefer `FULL_AND_PIECEWISE` since it has different kernels for prefill/mixed and pure decode batches |
-| AITER FlashAttention | `UNIFORM_BATCH`| |
+| AITER FlashAttention | `UNIFORM_BATCH` | |
 | FlashInfer | `UNIFORM_SINGLE_TOKEN_DECODE` | Will be set to `UNIFORM_BATCH` when using TRTLLM attention on Blackwell |
 | FlashMLA | `UNIFORM_BATCH` | |
 | FlashInferMLA | `UNIFORM_BATCH` | |
 | FlashInferMLASparse | `UNIFORM_BATCH` | |
 | AITER MLA | `UNIFORM_SINGLE_TOKEN_DECODE` | |
 | CUTLASS MLA | `UNIFORM_SINGLE_TOKEN_DECODE` | |
-| Mamba attention| `UNIFORM_SINGLE_TOKEN_DECODE` | |
+| Mamba attention | `UNIFORM_SINGLE_TOKEN_DECODE` | |
 
 Unlisted backends are all declared as `NEVER`.
 

docs/design/debug_vllm_compile.md

@@ -5,12 +5,12 @@ TL;DR:
 - use tlparse to acquire torch.compile logs. Include these logs in bug reports and/or support asks.
 - The vLLM-torch.compile integration is multiple pieces. vLLM exposes flags to turn off each piece:
 
-| Online Flag | Offline Flag   |      Result |
-|----------|----------|-------------|
-| --enforce-eager | enforce_eager=True |  Turn off torch.compile and CUDAGraphs |
-| -cc.mode=0 | mode=CompilationMode.NONE |  Turn off torch.compile only |
-| -cc.cudagraph_mode=NONE | compilation_config=CompilationConfig(cudagraph_mode=CUDAGraphMode.NONE) |  Turn off CUDAGraphs only |
-| -cc.backend=eager | compilation_config=CompilationConfig(backend='eager') |  Turn off TorchInductor |
+| Online Flag | Offline Flag | Result |
+| ----------- | ------------ | ------ |
+| --enforce-eager | enforce_eager=True | Turn off torch.compile and CUDAGraphs |
+| -cc.mode=0 | mode=CompilationMode.NONE | Turn off torch.compile only |
+| -cc.cudagraph_mode=NONE | compilation_config=CompilationConfig(cudagraph_mode=CUDAGraphMode.NONE) | Turn off CUDAGraphs only |
+| -cc.backend=eager | compilation_config=CompilationConfig(backend='eager') | Turn off TorchInductor |
 
 ## vLLM-torch.compile overview
 

docs/design/fusions.md

@@ -19,7 +19,7 @@ or just on the low or high end.
     If tuning performance by hand, always benchmark your exact use-case with and without the fusion to verify the impact.
 
 | Fusion                                                                         | `PassConfig` flag            | Fused operations                               | Default at                     | E2E Speedup        | Fullgraph | `num_tokens` |
-|--------------------------------------------------------------------------------|------------------------------|------------------------------------------------|--------------------------------|--------------------|-----------|--------------|
+| ------------------------------------------------------------------------------ | ---------------------------- | ---------------------------------------------- | ------------------------------ | ------------------ | --------- | ------------ |
 | [AllReduce + RMSNorm](#allreduce--rmsnorm-fuse_allreduce_rms)                  | `fuse_allreduce_rms`         | All-reduce → RMSNorm (+residual_add) (→ quant) | O2 (Hopper/Blackwell + TP > 1) | 5-20%              | No        | Low          |
 | [Attention + Quant](#attention--quantization-fuse_attn_quant)                  | `fuse_attn_quant`            | Attention output → FP8/NVFP4 quant             | Off by default                 | 3-7%               | Yes       | Always       |
 | [RoPE + KV-Cache Update](#rope--kv-cache-update-fuse_rope_kvcache)             | `fuse_rope_kvcache`          | Rotary embedding → KV cache write              | O1 (ROCm/AITER only)           | TBD                | No        | Low          |
@@ -37,7 +37,7 @@ The table below lists the quantization schemes supported by each fusion on each
 [#36066](https://github.com/vllm-project/vllm/issues/36066)
 
 | Fusion                       | SM100 (Blackwell)                        | SM90 (Hopper)                            | SM89 (Ada)                               | SM80 (Ampere) | ROCm                                     |
-|------------------------------|------------------------------------------|------------------------------------------|------------------------------------------|---------------|------------------------------------------|
+| ---------------------------- | ---------------------------------------- | ---------------------------------------- | ---------------------------------------- | ------------- | ---------------------------------------- |
 | `fuse_allreduce_rms`         | FP16/BF16, FP8 static, NVFP4             | FP16/BF16, FP8 static                    | —                                        | —             | —                                        |
 | `fuse_attn_quant`\*          | FP8 static\*, NVFP4\*                    | FP8 static\*                             | FP8 static\*                             | —             | FP8 static\*                             |
 | `fuse_rope_kvcache`          | —                                        | —                                        | —                                        | —             | FP16/BF16                                |

docs/design/metrics.md

@@ -507,10 +507,10 @@ longer relevant in v1:
 - `vllm:num_requests_swapped`
 - `vllm:cpu_cache_usage_perc`
 
-In this mode, when a request is preempted (e.g. to make room in KV
-cache to complete other requests), we swap kv cache blocks out to CPU
-memory. This is also known as "KV cache offloading" and is configured
-with `--swap-space` and `--preemption-mode`.
+In this mode, when a request was preempted (e.g. to make room in KV
+cache to complete other requests), kv cache blocks were swapped out to
+CPU memory. The `--swap-space` flag has been removed as this feature
+is no longer used in V1.
 
 Historically, [vLLM has long supported beam search](https://github.com/vllm-project/vllm/issues/6226). The
 SequenceGroup encapsulated the idea of N Sequences which

docs/design/moe_kernel_features.md

@@ -31,7 +31,7 @@ th {
 </style>
 
 | Backend | Output act. format | Quant. types | Quant. format | Async | Apply Weight On Input | Subclass |
-|---------|--------------------|--------------|---------------|-------|-----------------------|-----------|
+| ------- | ------------------ | ------------ | ------------- | ----- | --------------------- | --------- |
 | naive | standard | all<sup>1</sup> | G,A,T | N | <sup>6</sup> | [layer.py][vllm.model_executor.layers.fused_moe.layer.FusedMoE] |
 | deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize.DeepEPHTPrepareAndFinalize] |
 | deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize.DeepEPLLPrepareAndFinalize] |
@@ -78,7 +78,7 @@ Most experts flavors include an equivalent modular interface which will be a sub
 To be used with a particular `FusedMoEPrepareAndFinalizeModular` subclass, MoE kernels must have compatible activation formats, quantization types and quantization formats.
 
 | Kernel | Input act. format | Quant. types | Quant. format | Activation function | Apply Weight On Input | Modular | Source |
-|--------|-------------------|--------------|---------------|---------------------|-----------------------|---------|--------|
+| ------ | ----------------- | ------------ | ------------- | ------------------- | --------------------- | ------- | ------ |
 | triton | standard | all<sup>1</sup> | G,A,T | silu, gelu,</br>swigluoai,</br>silu_no_mul,</br>gelu_no_mul | Y | Y | [`fused_experts`][vllm.model_executor.layers.fused_moe.fused_moe.fused_experts],</br>[`TritonExperts`][vllm.model_executor.layers.fused_moe.fused_moe.TritonExperts] |
 | triton (batched) | batched | all<sup>1</sup> | G,A,T | silu, gelu | <sup>6</sup> | Y | [`BatchedTritonExperts`][vllm.model_executor.layers.fused_moe.fused_batched_moe.BatchedTritonExperts] |
 | deep gemm | standard,</br>batched | fp8 | G(128),A,T | silu, gelu | <sup>6</sup> | Y | </br>[`DeepGemmExperts`][vllm.model_executor.layers.fused_moe.deep_gemm_moe.DeepGemmExperts],</br>[`BatchedDeepGemmExperts`][vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe.BatchedDeepGemmExperts] |
@@ -105,7 +105,7 @@ To be used with a particular `FusedMoEPrepareAndFinalizeModular` subclass, MoE k
 The following table shows "families" of modular kernels that are intended to work together. There are some combinations which may work but have not yet been tested, e.g. flashinfer with other fp8 experts. Note that the "naive" backend will work with any non-modular experts.
 
 | backend | `FusedMoEPrepareAndFinalizeModular` subclasses | `FusedMoEExpertsModular` subclasses |
-|---------|-----------------------------------------|----------------------------------------------|
-| deepep_high_throughput | `DeepEPHTPrepareAndFinalize` |  `DeepGemmExperts`,</br>`TritonExperts`,</br>`TritonOrDeepGemmExperts`,</br>`CutlassExpertsFp8`, </br>`MarlinExperts` |
-| deepep_low_latency | `DeepEPLLPrepareAndFinalize` |  `BatchedDeepGemmExperts`,</br>`BatchedTritonExperts`,</br>`CutlassBatchedExpertsFp8`,</br>`BatchedMarlinExperts` |
+| ------- | ---------------------------------------------- | ----------------------------------- |
+| deepep_high_throughput | `DeepEPHTPrepareAndFinalize` | `DeepGemmExperts`,</br>`TritonExperts`,</br>`TritonOrDeepGemmExperts`,</br>`CutlassExpertsFp8`, </br>`MarlinExperts` |
+| deepep_low_latency | `DeepEPLLPrepareAndFinalize` | `BatchedDeepGemmExperts`,</br>`BatchedTritonExperts`,</br>`CutlassBatchedExpertsFp8`,</br>`BatchedMarlinExperts` |
 | flashinfer | `FlashInferCutlassMoEPrepareAndFinalize` | `FlashInferExperts` |

docs/design/plugin_system.md

@@ -155,4 +155,4 @@ The interface for the model/module may change during vLLM's development. If you
     - `use_v1` parameter in `Platform.get_attn_backend_cls` is deprecated. It has been removed in v0.13.0.
     - `_Backend` in `vllm.attention` is deprecated. It has been removed in v0.13.0. Please use `vllm.v1.attention.backends.registry.register_backend` to add new attention backend to `AttentionBackendEnum` instead.
     - `seed_everything` platform interface is deprecated. It has been removed in v0.16.0. Please use `vllm.utils.torch_utils.set_random_seed` instead.
-    - `prompt` in `Platform.validate_request` is deprecated and will be removed in v0.18.0.
+    - `prompt` in `Platform.validate_request` is deprecated. It has been removed in v0.18.0.

docs/design/torch_compile_multimodal.md

@@ -26,7 +26,7 @@ This feature is off by default, but can be enabled by setting `compile_mm_encode
 
 To compile a multimodal component such as an encoder, we follow the same mechanism as the LLM text backbone, with a few additional scaffoldings:
 
-1. The `@support_torch_compile` decorator should include `enable_if=should_torch_compile_mm_vit`. This will gate the compilation behind our
+1. The `@support_torch_compile` decorator should include `enable_if=should_torch_compile_mm_encoder`. This will gate the compilation behind our
 `compile_mm_encoder` configuration
 
 2. `with set_model_tag("<component_name>", is_encoder=True)` context manager should be used around the nn.Module's instantiation. Since torch.compile

docs/features/README.md

@@ -37,7 +37,7 @@ th:not(:first-child) {
 </style>
 
 | Feature | [CP](../configuration/optimization.md#chunked-prefill) | [APC](automatic_prefix_caching.md) | [LoRA](lora.md) | [SD](speculative_decoding/README.md) | CUDA graph | [pooling](../models/pooling_models.md) | <abbr title="Encoder-Decoder Models">enc-dec</abbr> | <abbr title="Logprobs">logP</abbr> | <abbr title="Prompt Logprobs">prmpt logP</abbr> | <abbr title="Async Output Processing">async output</abbr> | multi-step | <abbr title="Multimodal Inputs">mm</abbr> | best-of | beam-search | [prompt-embeds](prompt_embeds.md) |
-|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
+| - | - | - | - | - | - | - | - | - | - | - | - | - | - | - | - |
 | [CP](../configuration/optimization.md#chunked-prefill) | ✅ | | | | | | | | | | | | | | |
 | [APC](automatic_prefix_caching.md) | ✅ | ✅ | | | | | | | | | | | | | |
 | [LoRA](lora.md) | ✅ | ✅ | ✅ | | | | | | | | | | | | |
@@ -59,23 +59,23 @@ th:not(:first-child) {
 
 ### Feature x Hardware
 
-| Feature                                                   | Volta               | Turing    | Ampere    | Ada    | Hopper     | CPU                | AMD    | Intel GPU |
-|-----------------------------------------------------------|---------------------|-----------|-----------|--------|------------|--------------------|--------| ------------|
-| [CP](../configuration/optimization.md#chunked-prefill)                                     | [❌](https://github.com/vllm-project/vllm/issues/2729) | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
-| [APC](automatic_prefix_caching.md)                        | [❌](https://github.com/vllm-project/vllm/issues/3687) | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
-| [LoRA](lora.md)                                           | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
-| [SD](speculative_decoding/README.md)                        | ✅                  | ✅        | ✅        | ✅     | ✅        | ❌                  | ✅     | ✅        |
-| CUDA graph                                                | ✅                  | ✅        | ✅        | ✅     | ✅        | ❌                  | ✅     | [❌](https://github.com/vllm-project/vllm/issues/26970)        |
-| [pooling](../models/pooling_models.md)                    | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
-| <abbr title="Encoder-Decoder Models">enc-dec</abbr>       | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ❌     | ✅        |
-| [mm](multimodal_inputs.md)                                | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
-| [prompt-embeds](prompt_embeds.md)                         | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ❔     | ✅        |
-| <abbr title="Logprobs">logP</abbr>                        | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
-| <abbr title="Prompt Logprobs">prmpt logP</abbr>           | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
-| <abbr title="Async Output Processing">async output</abbr> | ✅                  | ✅        | ✅        | ✅     | ✅        | ❌                  | ❌     | ✅        |
-| multi-step                                                | ✅                  | ✅        | ✅        | ✅     | ✅        | [❌](https://github.com/vllm-project/vllm/issues/8477) | ✅     | ✅        |
-| best-of                                                   | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
-| beam-search                                               | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅        |
+| Feature | Volta | Turing | Ampere | Ada | Hopper | CPU | AMD | Intel GPU |
+| ------- | ----- | ------ | ------ | --- | ------ | --- | --- | --------- |
+| [CP](../configuration/optimization.md#chunked-prefill) | [❌](https://github.com/vllm-project/vllm/issues/2729) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| [APC](automatic_prefix_caching.md) | [❌](https://github.com/vllm-project/vllm/issues/3687) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| [LoRA](lora.md) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| [SD](speculative_decoding/README.md) | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
+| CUDA graph | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | [❌](https://github.com/vllm-project/vllm/issues/26970) |
+| [pooling](../models/pooling_models.md) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| <abbr title="Encoder-Decoder Models">enc-dec</abbr> | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ |
+| [mm](multimodal_inputs.md) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| [prompt-embeds](prompt_embeds.md) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❔ | ✅ |
+| <abbr title="Logprobs">logP</abbr> | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| <abbr title="Prompt Logprobs">prmpt logP</abbr> | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| <abbr title="Async Output Processing">async output</abbr> | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ |
+| multi-step | ✅ | ✅ | ✅ | ✅ | ✅ | [❌](https://github.com/vllm-project/vllm/issues/8477) | ✅ | ✅ |
+| best-of | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
+| beam-search | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
 
 !!! note
     For information on feature support on Google TPU, please refer to the [TPU-Inference Recommended Models and Features](https://docs.vllm.ai/projects/tpu/en/latest/recommended_models_features/) documentation.

docs/features/disagg_prefill.md

@@ -44,6 +44,12 @@ For NixlConnector, you may also specify one or multiple NIXL_Backend. Such as:
   --kv-transfer-config '{"kv_connector":"OffloadingConnector","kv_role":"kv_both","kv_connector_extra_config":{"block_size": 64, "cpu_bytes_to_use": 1000000000}}'
   ```
 
+- **FlexKVConnectorV1**: refer to [examples/offline_inference/prefix_caching_flexkv.py](../../examples/offline_inference/prefix_caching_flexkv.py) for the example usage of FlexKVConnectorV1. FlexKV is a distributed KV Store and multi-level cache management system for ultra-large-scale LLM inference.
+
+  ```bash
+  --kv-transfer-config '{"kv_connector":"FlexKVConnectorV1","kv_role":"kv_both"}'
+  ```
+
 ## Benchmarks
 
 Please refer to [benchmarks/disagg_benchmarks](../../benchmarks/disagg_benchmarks) for disaggregated prefilling benchmarks.

docs/features/interleaved_thinking.md

@@ -20,9 +20,9 @@ With interleaved thinking, the model can:
 vLLM currently supports the following interleaved thinking models:
 
 | Model Series | Reasoning Parser Name |
-|--------------|-----------------------|
-| moonshotai/Kimi-K2-Thinking    |  kimi_k2  |
-| MiniMaxAI/MiniMax-M2           |  minimax_m2  |
+| ------------ | --------------------- |
+| moonshotai/Kimi-K2-Thinking | kimi_k2 |
+| MiniMaxAI/MiniMax-M2 | minimax_m2 |
 
 ## Example Usage
 

docs/features/quantization/README.md

@@ -44,16 +44,16 @@ th:not(:first-child) {
 }
 </style>
 
-| Implementation        | Volta   | Turing   | Ampere   | Ada   | Hopper   | AMD GPU   | Intel GPU   | x86 CPU   |
-|-----------------------|---------|----------|----------|-------|----------|-----------|-------------|-----------|
-| AWQ                   | ❌      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ✅︎          | ✅︎        |
-| GPTQ                  | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ✅︎          | ✅︎        |
-| Marlin (GPTQ/AWQ/FP8/FP4) | ❌      | ✅︎*       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ❌        |
-| INT8 (W8A8)           | ❌      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ✅︎        |
-| FP8 (W8A8)            | ❌      | ❌       | ❌       | ✅︎    | ✅︎       | ✅︎         | ❌          | ❌        |
-| bitsandbytes          | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ❌        |
-| DeepSpeedFP           | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ❌        |
-| GGUF                  | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ✅︎         | ❌          | ❌        |
+| Implementation            | Volta | Turing | Ampere | Ada | Hopper | AMD GPU | Intel GPU | x86 CPU |
+| ------------------------- | ----- | ------ | ------ | --- | ------ | ------- | --------- | ------- |
+| AWQ                       | ❌    | ✅︎     | ✅︎     | ✅︎  | ✅︎     | ❌      | ✅︎        | ✅︎      |
+| GPTQ                      | ✅︎    | ✅︎     | ✅︎     | ✅︎  | ✅︎     | ❌      | ✅︎        | ✅︎      |
+| Marlin (GPTQ/AWQ/FP8/FP4) | ❌    | ✅︎*    | ✅︎     | ✅︎  | ✅︎     | ❌      | ❌        | ❌      |
+| INT8 (W8A8)               | ❌    | ✅︎     | ✅︎     | ✅︎  | ✅︎     | ❌      | ❌        | ✅︎      |
+| FP8 (W8A8)                | ❌    | ❌     | ❌     | ✅︎  | ✅︎     | ✅︎      | ❌        | ❌      |
+| bitsandbytes              | ✅︎    | ✅︎     | ✅︎     | ✅︎  | ✅︎     | ❌      | ❌        | ❌      |
+| DeepSpeedFP               | ✅︎    | ✅︎     | ✅︎     | ✅︎  | ✅︎     | ❌      | ❌        | ❌      |
+| GGUF                      | ✅︎    | ✅︎     | ✅︎     | ✅︎  | ✅︎     | ✅︎      | ❌        | ❌      |
 
 - Volta refers to SM 7.0, Turing to SM 7.5, Ampere to SM 8.0/8.6, Ada to SM 8.9, and Hopper to SM 9.0.
 - ✅︎ indicates that the quantization method is supported on the specified hardware.
@@ -131,7 +131,7 @@ class MyQuantConfig(QuantizationConfig):
 Your custom `QuantizationConfig` subclass must implement these abstract methods:
 
 | Method | Description |
-|--------|-------------|
+| ------ | ----------- |
 | `get_name()` | Returns the name of the quantization method |
 | `get_supported_act_dtypes()` | Returns list of supported activation dtypes (e.g., `torch.float16`) |
 | `get_min_capability()` | Returns minimum GPU compute capability (e.g., 80 for Ampere, -1 for no restriction) |

docs/features/quantization/fp8.md

@@ -114,7 +114,7 @@ Here's an example of the resulting scores:
 
 ```text
 |Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
-|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
+| --- |------:| -------------- |-----:| --------- | - |----:| - |-----:|
 |gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.768|±  |0.0268|
 |     |       |strict-match    |     5|exact_match|↑  |0.768|±  |0.0268|
 ```

docs/features/reasoning_outputs.md

@@ -12,7 +12,7 @@ Reasoning models return an additional `reasoning` field in their outputs, which
 vLLM currently supports the following reasoning models:
 
 | Model Series | Parser Name | Structured Output Support | Tool Calling |
-|--------------|-------------|------------------|-------------|
+| ------------ | ----------- | ---------------- | ----------- |
 | [DeepSeek R1 series](https://huggingface.co/collections/deepseek-ai/deepseek-r1-678e1e131c0169c0bc89728d) | `deepseek_r1` | `json`, `regex` | ❌ |
 | [DeepSeek-V3.1](https://huggingface.co/collections/deepseek-ai/deepseek-v31-68a491bed32bd77e7fca048f) | `deepseek_v3` | `json`, `regex` | ❌ |
 | [ERNIE-4.5-VL series](https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT) | `ernie45` | `json`, `regex` | ❌ |

docs/getting_started/installation/README.md

@@ -16,4 +16,4 @@ vLLM supports the following hardware platforms:
 
 vLLM supports third-party hardware plugins that live **outside** the main `vllm` repository. These follow the [Hardware-Pluggable RFC](../../design/plugin_system.md).
 
-A list of all supported hardware can be found on the [vllm.ai website](https://vllm.ai/#hardware). If you want to add new hardware, please contact us on [Slack](https://slack.vllm.ai/) or [Email](mailto:collaboration@vllm.ai).
+A list of all supported hardware can be found on the [vllm.ai website](https://vllm.ai/#compatibility). If you want to add new hardware, please contact us on [Slack](https://slack.vllm.ai/) or [Email](mailto:collaboration@vllm.ai).

docs/getting_started/installation/cpu.apple.inc.md

@@ -1,4 +1,5 @@
-# --8<-- [start:installation]
+<!-- markdownlint-disable MD041 -->
+--8<-- [start:installation]
 
 vLLM has experimental support for macOS with Apple Silicon. For now, users must build from source to natively run on macOS.
 
@@ -7,23 +8,23 @@ Currently the CPU implementation for macOS supports FP32 and FP16 datatypes.
 !!! tip "GPU-Accelerated Inference with vLLM-Metal"
     For GPU-accelerated inference on Apple Silicon using Metal, check out [vllm-metal](https://github.com/vllm-project/vllm-metal), a community-maintained hardware plugin that uses MLX as the compute backend.
 
-# --8<-- [end:installation]
-# --8<-- [start:requirements]
+--8<-- [end:installation]
+--8<-- [start:requirements]
 
 - OS: `macOS Sonoma` or later
 - SDK: `XCode 15.4` or later with Command Line Tools
 - Compiler: `Apple Clang >= 15.0.0`
 
-# --8<-- [end:requirements]
-# --8<-- [start:set-up-using-python]
+--8<-- [end:requirements]
+--8<-- [start:set-up-using-python]
 
-# --8<-- [end:set-up-using-python]
-# --8<-- [start:pre-built-wheels]
+--8<-- [end:set-up-using-python]
+--8<-- [start:pre-built-wheels]
 
 Currently, there are no pre-built Apple silicon CPU wheels.
 
-# --8<-- [end:pre-built-wheels]
-# --8<-- [start:build-wheel-from-source]
+--8<-- [end:pre-built-wheels]
+--8<-- [start:build-wheel-from-source]
 
 After installation of XCode and the Command Line Tools, which include Apple Clang, execute the following commands to build and install vLLM from source.
 
@@ -36,7 +37,7 @@ uv pip install -e .
 
 !!! tip
     The `--index-strategy unsafe-best-match` flag is needed to resolve dependencies across multiple package indexes (PyTorch CPU index and PyPI). Without this flag, you may encounter `typing-extensions` version conflicts.
-    
+
     The term "unsafe" refers to the package resolution strategy, not security. By default, `uv` only searches the first index where a package is found to prevent dependency confusion attacks. This flag allows `uv` to search all configured indexes to find the best compatible versions. Since both PyTorch and PyPI are trusted package sources, using this strategy is safe and appropriate for vLLM installation.
 
 !!! note
@@ -77,14 +78,14 @@ uv pip install -e .
     ```
     On Apple Clang 16 you should see: `#define __cplusplus 201703L`
 
-# --8<-- [end:build-wheel-from-source]
-# --8<-- [start:pre-built-images]
+--8<-- [end:build-wheel-from-source]
+--8<-- [start:pre-built-images]
 
 Currently, there are no pre-built Arm silicon CPU images.
 
-# --8<-- [end:pre-built-images]
-# --8<-- [start:build-image-from-source]
+--8<-- [end:pre-built-images]
+--8<-- [start:build-image-from-source]
 
-# --8<-- [end:build-image-from-source]
-# --8<-- [start:extra-information]
-# --8<-- [end:extra-information]
+--8<-- [end:build-image-from-source]
+--8<-- [start:extra-information]
+--8<-- [end:extra-information]

docs/getting_started/installation/cpu.arm.inc.md

@@ -1,19 +1,20 @@
-# --8<-- [start:installation]
+<!-- markdownlint-disable MD041 -->
+--8<-- [start:installation]
 
 vLLM offers basic model inferencing and serving on Arm CPU platform, with support for NEON, data types FP32, FP16 and BF16.
 
-# --8<-- [end:installation]
-# --8<-- [start:requirements]
+--8<-- [end:installation]
+--8<-- [start:requirements]
 
 - OS: Linux
 - Compiler: `gcc/g++ >= 12.3.0` (optional, recommended)
 - Instruction Set Architecture (ISA): NEON support is required
 
-# --8<-- [end:requirements]
-# --8<-- [start:set-up-using-python]
+--8<-- [end:requirements]
+--8<-- [start:set-up-using-python]
 
-# --8<-- [end:set-up-using-python]
-# --8<-- [start:pre-built-wheels]
+--8<-- [end:set-up-using-python]
+--8<-- [start:pre-built-wheels]
 
 Pre-built vLLM wheels for Arm are available since version 0.11.2. These wheels contain pre-compiled C++ binaries.
 
@@ -43,13 +44,14 @@ uv pip install https://github.com/vllm-project/vllm/releases/download/v${VLLM_VE
 
 The `uv` approach works for vLLM `v0.6.6` and later. A unique feature of `uv` is that packages in `--extra-index-url` have [higher priority than the default index](https://docs.astral.sh/uv/pip/compatibility/#packages-that-exist-on-multiple-indexes). If the latest public release is `v0.6.6.post1`, `uv`'s behavior allows installing a commit before `v0.6.6.post1` by specifying the `--extra-index-url`. In contrast, `pip` combines packages from `--extra-index-url` and the default index, choosing only the latest version, which makes it difficult to install a development version prior to the released version.
 
-**Install the latest code**
+#### Install the latest code
 
 LLM inference is a fast-evolving field, and the latest code may contain bug fixes, performance improvements, and new features that are not released yet. To allow users to try the latest code without waiting for the next release, vLLM provides working pre-built Arm CPU wheels for every commit since `v0.11.2` on <https://wheels.vllm.ai/nightly>. For native CPU wheels, this index should be used:
 
-* `https://wheels.vllm.ai/nightly/cpu/vllm`
+- `https://wheels.vllm.ai/nightly/cpu/vllm`
 
 To install from nightly index, run:
+
 ```bash
 uv pip install vllm --extra-index-url https://wheels.vllm.ai/nightly/cpu --index-strategy first-index
 ```
@@ -64,7 +66,7 @@ uv pip install vllm --extra-index-url https://wheels.vllm.ai/nightly/cpu --index
     pip install https://wheels.vllm.ai/4fa7ce46f31cbd97b4651694caf9991cc395a259/vllm-0.13.0rc2.dev104%2Bg4fa7ce46f.cpu-cp38-abi3-manylinux_2_35_aarch64.whl # current nightly build (the filename will change!)
     ```
 
-**Install specific revisions**
+#### Install specific revisions
 
 If you want to access the wheels for previous commits (e.g. to bisect the behavior change, performance regression), you can specify the commit hash in the URL:
 
@@ -73,8 +75,8 @@ export VLLM_COMMIT=730bd35378bf2a5b56b6d3a45be28b3092d26519 # use full commit ha
 uv pip install vllm --extra-index-url https://wheels.vllm.ai/${VLLM_COMMIT}/cpu --index-strategy first-index
 ```
 
-# --8<-- [end:pre-built-wheels]
-# --8<-- [start:build-wheel-from-source]
+--8<-- [end:pre-built-wheels]
+--8<-- [start:build-wheel-from-source]
 
 First, install the recommended compiler. We recommend using `gcc/g++ >= 12.3.0` as the default compiler to avoid potential problems. For example, on Ubuntu 22.4, you can run:
 
@@ -133,8 +135,8 @@ Testing has been conducted on AWS Graviton3 instances for compatibility.
     export LD_PRELOAD="$TC_PATH:$LD_PRELOAD"
     ```
 
-# --8<-- [end:build-wheel-from-source]
-# --8<-- [start:pre-built-images]
+--8<-- [end:build-wheel-from-source]
+--8<-- [start:pre-built-images]
 
 To pull the latest image from Docker Hub:
 
@@ -170,10 +172,10 @@ export VLLM_COMMIT=6299628d326f429eba78736acb44e76749b281f5 # use full commit ha
 docker pull public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:${VLLM_COMMIT}-arm64-cpu
 ```
 
-# --8<-- [end:pre-built-images]
-# --8<-- [start:build-image-from-source]
+--8<-- [end:pre-built-images]
+--8<-- [start:build-image-from-source]
 
-## Building for your target ARM CPU
+#### Building for your target ARM CPU
 
 ```bash
 docker build -f docker/Dockerfile.cpu \
@@ -189,9 +191,9 @@ docker build -f docker/Dockerfile.cpu \
     - `VLLM_CPU_ARM_BF16=true` - Force-enable ARM BF16 support (build with BF16 regardless of build system capabilities)
     - `VLLM_CPU_ARM_BF16=false` - Rely on auto-detection (default)
 
-### Examples
+##### Examples
 
-**Auto-detection build (native ARM)**
+###### Auto-detection build (native ARM)
 
 ```bash
 # Building on ARM64 system - platform auto-detected
@@ -200,7 +202,7 @@ docker build -f docker/Dockerfile.cpu \
         --target vllm-openai .
 ```
 
-**Cross-compile for ARM with BF16 support**
+###### Cross-compile for ARM with BF16 support
 
 ```bash
 # Building on ARM64 for newer ARM CPUs with BF16
@@ -210,7 +212,7 @@ docker build -f docker/Dockerfile.cpu \
         --target vllm-openai .
 ```
 
-**Cross-compile from x86_64 to ARM64 with BF16**
+###### Cross-compile from x86_64 to ARM64 with BF16
 
 ```bash
 # Requires Docker buildx with ARM emulation (QEMU)
@@ -226,7 +228,7 @@ docker buildx build -f docker/Dockerfile.cpu \
 !!! note "ARM BF16 requirements"
     ARM BF16 support requires ARMv8.6-A or later (FEAT_BF16). Supported on AWS Graviton3/4, AmpereOne, and other recent ARM processors.
 
-## Launching the OpenAI server
+#### Launching the OpenAI server
 
 ```bash
 docker run --rm \
@@ -245,6 +247,6 @@ docker run --rm \
 !!! tip "Alternative to --privileged"
     Instead of `--privileged=true`, use `--cap-add SYS_NICE --security-opt seccomp=unconfined` for better security.
 
-# --8<-- [end:build-image-from-source]
-# --8<-- [start:extra-information]
-# --8<-- [end:extra-information]
+--8<-- [end:build-image-from-source]
+--8<-- [start:extra-information]
+--8<-- [end:extra-information]

docs/getting_started/installation/cpu.md

@@ -1,3 +1,7 @@
+---
+toc_depth: 3
+---
+
 # CPU
 
 vLLM is a Python library that supports the following CPU variants. Select your CPU type to see vendor specific instructions:
@@ -255,7 +259,7 @@ ON_CPU=1 SERVING_JSON=serving-tests-cpu-text.json DRY_RUN=1 MODEL_FILTER=meta-ll
 
     # On this platform, it is recommended to only bind openMP threads on logical CPU cores 0-7 or 8-15
     $ export VLLM_CPU_OMP_THREADS_BIND=0-7
-    $ python examples/offline_inference/basic/basic.py
+    $ python examples/basic/offline_inference/basic.py
     ```
 
 - When deploying vLLM CPU backend on a multi-socket machine with NUMA and enable tensor parallel or pipeline parallel, each NUMA node is treated as a TP/PP rank. So be aware to set CPU cores of a single rank on the same NUMA node to avoid cross NUMA node memory access.

docs/getting_started/installation/cpu.s390x.inc.md

@@ -1,27 +1,28 @@
-# --8<-- [start:installation]
+<!-- markdownlint-disable MD041 -->
+--8<-- [start:installation]
 
 vLLM has experimental support for s390x architecture on IBM Z platform. For now, users must build from source to natively run on IBM Z platform.
 
 Currently, the CPU implementation for s390x architecture supports FP32 datatype only.
 
-# --8<-- [end:installation]
-# --8<-- [start:requirements]
+--8<-- [end:installation]
+--8<-- [start:requirements]
 
 - OS: `Linux`
 - SDK: `gcc/g++ >= 12.3.0` or later with Command Line Tools
 - Instruction Set Architecture (ISA): VXE support is required. Works with Z14 and above.
 - Build install python packages: `pyarrow`, `torch` and `torchvision`
 
-# --8<-- [end:requirements]
-# --8<-- [start:set-up-using-python]
+--8<-- [end:requirements]
+--8<-- [start:set-up-using-python]
 
-# --8<-- [end:set-up-using-python]
-# --8<-- [start:pre-built-wheels]
+--8<-- [end:set-up-using-python]
+--8<-- [start:pre-built-wheels]
 
 Currently, there are no pre-built IBM Z CPU wheels.
 
-# --8<-- [end:pre-built-wheels]
-# --8<-- [start:build-wheel-from-source]
+--8<-- [end:pre-built-wheels]
+--8<-- [start:build-wheel-from-source]
 
 Install the following packages from the package manager before building the vLLM. For example on RHEL 9.4:
 
@@ -65,13 +66,13 @@ Execute the following commands to build and install vLLM from source.
             pip install dist/*.whl
     ```
 
-# --8<-- [end:build-wheel-from-source]
-# --8<-- [start:pre-built-images]
+--8<-- [end:build-wheel-from-source]
+--8<-- [start:pre-built-images]
 
 Currently, there are no pre-built IBM Z CPU images.
 
-# --8<-- [end:pre-built-images]
-# --8<-- [start:build-image-from-source]
+--8<-- [end:pre-built-images]
+--8<-- [start:build-image-from-source]
 
 ```bash
 docker build -f docker/Dockerfile.s390x \
@@ -93,6 +94,6 @@ docker run --rm \
 !!! tip
     An alternative of `--privileged true` is `--cap-add SYS_NICE --security-opt seccomp=unconfined`.
 
-# --8<-- [end:build-image-from-source]
-# --8<-- [start:extra-information]
-# --8<-- [end:extra-information]
+--8<-- [end:build-image-from-source]
+--8<-- [start:extra-information]
+--8<-- [end:extra-information]

docs/getting_started/installation/cpu.x86.inc.md

@@ -1,9 +1,10 @@
-# --8<-- [start:installation]
+<!-- markdownlint-disable MD041 -->
+--8<-- [start:installation]
 
 vLLM supports basic model inferencing and serving on x86 CPU platform, with data types FP32, FP16 and BF16.
 
-# --8<-- [end:installation]
-# --8<-- [start:requirements]
+--8<-- [end:installation]
+--8<-- [start:requirements]
 
 - OS: Linux
 - CPU flags: `avx512f` (Recommended), `avx512_bf16` (Optional), `avx512_vnni` (Optional)
@@ -11,11 +12,11 @@ vLLM supports basic model inferencing and serving on x86 CPU platform, with data
 !!! tip
     Use `lscpu` to check the CPU flags.
 
-# --8<-- [end:requirements]
-# --8<-- [start:set-up-using-python]
+--8<-- [end:requirements]
+--8<-- [start:set-up-using-python]
 
-# --8<-- [end:set-up-using-python]
-# --8<-- [start:pre-built-wheels]
+--8<-- [end:set-up-using-python]
+--8<-- [start:pre-built-wheels]
 
 Pre-built vLLM wheels for x86 with AVX512 are available since version 0.13.0. To install release wheels:
 
@@ -25,6 +26,7 @@ export VLLM_VERSION=$(curl -s https://api.github.com/repos/vllm-project/vllm/rel
 # use uv
 uv pip install https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}+cpu-cp38-abi3-manylinux_2_35_x86_64.whl --torch-backend cpu
 ```
+
 ??? console "pip"
     ```bash
     # use pip
@@ -46,7 +48,7 @@ uv pip install https://github.com/vllm-project/vllm/releases/download/v${VLLM_VE
     export LD_PRELOAD="$TC_PATH:$IOMP_PATH:$LD_PRELOAD"
     ```
 
-**Install the latest code**
+#### Install the latest code
 
 To install the wheel built from the latest main branch:
 
@@ -54,7 +56,7 @@ To install the wheel built from the latest main branch:
 uv pip install vllm --extra-index-url https://wheels.vllm.ai/nightly/cpu --index-strategy first-index --torch-backend cpu
 ```
 
-**Install specific revisions**
+#### Install specific revisions
 
 If you want to access the wheels for previous commits (e.g. to bisect the behavior change, performance regression), you can specify the commit hash in the URL:
 
@@ -63,8 +65,8 @@ export VLLM_COMMIT=730bd35378bf2a5b56b6d3a45be28b3092d26519 # use full commit ha
 uv pip install vllm --extra-index-url https://wheels.vllm.ai/${VLLM_COMMIT}/cpu --index-strategy first-index --torch-backend cpu
 ```
 
-# --8<-- [end:pre-built-wheels]
-# --8<-- [start:build-wheel-from-source]
+--8<-- [end:pre-built-wheels]
+--8<-- [start:build-wheel-from-source]
 
 Install recommended compiler. We recommend to use `gcc/g++ >= 12.3.0` as the default compiler to avoid potential problems. For example, on Ubuntu 22.4, you can run:
 
@@ -158,8 +160,8 @@ uv pip install dist/*.whl
     ]
     ```
 
-# --8<-- [end:build-wheel-from-source]
-# --8<-- [start:pre-built-images]
+--8<-- [end:build-wheel-from-source]
+--8<-- [start:pre-built-images]
 
 You can pull the latest available CPU image from Docker Hub:
 
@@ -189,10 +191,10 @@ vllm/vllm-openai-cpu:latest-x86_64 <args...>
 !!! warning
     If deploying the pre-built images on machines without `avx512f`, `avx512_bf16`, or `avx512_vnni` support, an `Illegal instruction` error may be raised. See the build-image-from-source section below for build arguments to match your target CPU capabilities.
 
-# --8<-- [end:pre-built-images]
-# --8<-- [start:build-image-from-source]
+--8<-- [end:pre-built-images]
+--8<-- [start:build-image-from-source]
 
-## Building for your target CPU
+#### Building for your target CPU
 
 ```bash
 docker build -f docker/Dockerfile.cpu \
@@ -212,15 +214,15 @@ docker build -f docker/Dockerfile.cpu \
     - `VLLM_CPU_{ISA}=true` - Force-enable the instruction set (build with ISA regardless of build system capabilities)
     - `VLLM_CPU_{ISA}=false` - Rely on auto-detection (default)
 
-### Examples
+##### Examples
 
-**Auto-detection build (default)**
+###### Auto-detection build (default)
 
 ```bash
 docker build -f docker/Dockerfile.cpu --tag vllm-cpu-env --target vllm-openai .
 ```
 
-**Cross-compile for AVX512**
+###### Cross-compile for AVX512
 
 ```bash
 docker build -f docker/Dockerfile.cpu \
@@ -231,7 +233,7 @@ docker build -f docker/Dockerfile.cpu \
         --target vllm-openai .
 ```
 
-**Cross-compile for AVX2**
+###### Cross-compile for AVX2
 
 ```bash
 docker build -f docker/Dockerfile.cpu \
@@ -240,7 +242,7 @@ docker build -f docker/Dockerfile.cpu \
         --target vllm-openai .
 ```
 
-## Launching the OpenAI server
+#### Launching the OpenAI server
 
 ```bash
 docker run --rm \
@@ -255,6 +257,6 @@ docker run --rm \
             other vLLM OpenAI server arguments
 ```
 
-# --8<-- [end:build-image-from-source]
-# --8<-- [start:extra-information]
-# --8<-- [end:extra-information]
+--8<-- [end:build-image-from-source]
+--8<-- [start:extra-information]
+--8<-- [end:extra-information]

docs/getting_started/installation/gpu.cuda.inc.md

@@ -1,14 +1,15 @@
-# --8<-- [start:installation]
+<!-- markdownlint-disable MD041 MD051 -->
+--8<-- [start:installation]
 
 vLLM contains pre-compiled C++ and CUDA (12.8) binaries.
 
-# --8<-- [end:installation]
-# --8<-- [start:requirements]
+--8<-- [end:installation]
+--8<-- [start:requirements]
 
 - GPU: compute capability 7.0 or higher (e.g., V100, T4, RTX20xx, A100, L4, H100, etc.)
 
-# --8<-- [end:requirements]
-# --8<-- [start:set-up-using-python]
+--8<-- [end:requirements]
+--8<-- [start:set-up-using-python]
 
 !!! note
     PyTorch installed via `conda` will statically link `NCCL` library, which can cause issues when vLLM tries to use `NCCL`. See <https://github.com/vllm-project/vllm/issues/8420> for more details.
@@ -17,8 +18,8 @@ In order to be performant, vLLM has to compile many cuda kernels. The compilatio
 
 Therefore, it is recommended to install vLLM with a **fresh new** environment. If either you have a different CUDA version or you want to use an existing PyTorch installation, you need to build vLLM from source. See [below](#build-wheel-from-source) for more details.
 
-# --8<-- [end:set-up-using-python]
-# --8<-- [start:pre-built-wheels]
+--8<-- [end:set-up-using-python]
+--8<-- [start:pre-built-wheels]
 
 ```bash
 uv pip install vllm --torch-backend=auto
@@ -49,8 +50,8 @@ uv pip install https://github.com/vllm-project/vllm/releases/download/v${VLLM_VE
 
 LLM inference is a fast-evolving field, and the latest code may contain bug fixes, performance improvements, and new features that are not released yet. To allow users to try the latest code without waiting for the next release, vLLM provides wheels for every commit since `v0.5.3` on <https://wheels.vllm.ai/nightly>. There are multiple indices that could be used:
 
-* `https://wheels.vllm.ai/nightly`: the default variant (CUDA with version specified in `VLLM_MAIN_CUDA_VERSION`) built with the last commit on the `main` branch. Currently it is CUDA 12.9.
-* `https://wheels.vllm.ai/nightly/<variant>`: all other variants. Now this includes `cu130`, and `cpu`. The default variant (`cu129`) also has a subdirectory to keep consistency.
+- `https://wheels.vllm.ai/nightly`: the default variant (CUDA with version specified in `VLLM_MAIN_CUDA_VERSION`) built with the last commit on the `main` branch. Currently it is CUDA 12.9.
+- `https://wheels.vllm.ai/nightly/<variant>`: all other variants. Now this includes `cu130`, and `cpu`. The default variant (`cu129`) also has a subdirectory to keep consistency.
 
 To install from nightly index, run:
 
@@ -82,8 +83,8 @@ uv pip install vllm \
     --extra-index-url https://wheels.vllm.ai/${VLLM_COMMIT} # add variant subdirectory here if needed
 ```
 
-# --8<-- [end:pre-built-wheels]
-# --8<-- [start:build-wheel-from-source]
+--8<-- [end:pre-built-wheels]
+--8<-- [start:build-wheel-from-source]
 
 #### Set up using Python-only build (without compilation) {#python-only-build}
 
@@ -116,9 +117,9 @@ uv pip install --editable .
 
 There are more environment variables to control the behavior of Python-only build:
 
-* `VLLM_PRECOMPILED_WHEEL_LOCATION`: specify the exact wheel URL or local file path of a pre-compiled wheel to use. All other logic to find the wheel will be skipped.
-* `VLLM_PRECOMPILED_WHEEL_COMMIT`: override the commit hash to download the pre-compiled wheel. It can be `nightly` to use the last **already built** commit on the main branch.
-* `VLLM_PRECOMPILED_WHEEL_VARIANT`: specify the variant subdirectory to use on the nightly index, e.g., `cu129`, `cu130`, `cpu`. If not specified, the variant is auto-detected based on your system's CUDA version (from PyTorch or nvidia-smi). You can also set `VLLM_MAIN_CUDA_VERSION` to override auto-detection.
+- `VLLM_PRECOMPILED_WHEEL_LOCATION`: specify the exact wheel URL or local file path of a pre-compiled wheel to use. All other logic to find the wheel will be skipped.
+- `VLLM_PRECOMPILED_WHEEL_COMMIT`: override the commit hash to download the pre-compiled wheel. It can be `nightly` to use the last **already built** commit on the main branch.
+- `VLLM_PRECOMPILED_WHEEL_VARIANT`: specify the variant subdirectory to use on the nightly index, e.g., `cu129`, `cu130`, `cpu`. If not specified, the variant is auto-detected based on your system's CUDA version (from PyTorch or nvidia-smi). You can also set `VLLM_MAIN_CUDA_VERSION` to override auto-detection.
 
 You can find more information about vLLM's wheels in [Install the latest code](#install-the-latest-code).
 
@@ -236,8 +237,8 @@ export VLLM_TARGET_DEVICE=empty
 uv pip install -e .
 ```
 
-# --8<-- [end:build-wheel-from-source]
-# --8<-- [start:pre-built-images]
+--8<-- [end:build-wheel-from-source]
+--8<-- [start:pre-built-images]
 
 vLLM offers an official Docker image for deployment.
 The image can be used to run OpenAI compatible server and is available on Docker Hub as [vllm/vllm-openai](https://hub.docker.com/r/vllm/vllm-openai/tags).
@@ -314,8 +315,8 @@ docker run --runtime nvidia --gpus all \
 
 This will automatically configure `LD_LIBRARY_PATH` to point to the compatibility libraries before loading PyTorch and other dependencies.
 
-# --8<-- [end:pre-built-images]
-# --8<-- [start:build-image-from-source]
+--8<-- [end:pre-built-images]
+--8<-- [start:build-image-from-source]
 
 You can build and run vLLM from source via the provided [docker/Dockerfile](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile). To build vLLM:
 
@@ -415,9 +416,9 @@ The argument `vllm/vllm-openai` specifies the image to run, and should be replac
 !!! note
     **For version 0.4.1 and 0.4.2 only** - the vLLM docker images under these versions are supposed to be run under the root user since a library under the root user's home directory, i.e. `/root/.config/vllm/nccl/cu12/libnccl.so.2.18.1` is required to be loaded during runtime. If you are running the container under a different user, you may need to first change the permissions of the library (and all the parent directories) to allow the user to access it, then run vLLM with environment variable `VLLM_NCCL_SO_PATH=/root/.config/vllm/nccl/cu12/libnccl.so.2.18.1` .
 
-# --8<-- [end:build-image-from-source]
-# --8<-- [start:supported-features]
+--8<-- [end:build-image-from-source]
+--8<-- [start:supported-features]
 
 See [Feature x Hardware](../../features/README.md#feature-x-hardware) compatibility matrix for feature support information.
 
-# --8<-- [end:supported-features]
+--8<-- [end:supported-features]

docs/getting_started/installation/gpu.md

@@ -88,8 +88,7 @@ vLLM is a Python library that supports the following GPU variants. Select your G
 
 ### Pre-built images
 
-<!-- markdownlint-disable MD025 -->
-# --8<-- [start:pre-built-images]
+--8<-- [start:pre-built-images]
 
 === "NVIDIA CUDA"
 
@@ -103,15 +102,11 @@ vLLM is a Python library that supports the following GPU variants. Select your G
 
     --8<-- "docs/getting_started/installation/gpu.xpu.inc.md:pre-built-images"
 
-# --8<-- [end:pre-built-images]
-<!-- markdownlint-enable MD025 -->
+--8<-- [end:pre-built-images]
 
-<!-- markdownlint-disable MD001 -->
 ### Build image from source
-<!-- markdownlint-enable MD001 -->
 
-<!-- markdownlint-disable MD025 -->
-# --8<-- [start:build-image-from-source]
+--8<-- [start:build-image-from-source]
 
 === "NVIDIA CUDA"
 
@@ -125,8 +120,7 @@ vLLM is a Python library that supports the following GPU variants. Select your G
 
     --8<-- "docs/getting_started/installation/gpu.xpu.inc.md:build-image-from-source"
 
-# --8<-- [end:build-image-from-source]
-<!-- markdownlint-enable MD025 -->
+--8<-- [end:build-image-from-source]
 
 ## Supported features
 

docs/getting_started/installation/gpu.rocm.inc.md

@@ -1,23 +1,24 @@
-# --8<-- [start:installation]
+<!-- markdownlint-disable MD041 MD051 -->
+--8<-- [start:installation]
 
 vLLM supports AMD GPUs with ROCm 6.3 or above. Pre-built wheels are available for ROCm 7.0.
 
-# --8<-- [end:installation]
-# --8<-- [start:requirements]
+--8<-- [end:installation]
+--8<-- [start:requirements]
 
 - GPU: MI200s (gfx90a), MI300 (gfx942), MI350 (gfx950), Radeon RX 7900 series (gfx1100/1101), Radeon RX 9000 series (gfx1200/1201), Ryzen AI MAX / AI 300 Series (gfx1151/1150)
 - ROCm 6.3 or above
     - MI350 requires ROCm 7.0 or above
     - Ryzen AI MAX / AI 300 Series requires ROCm 7.0.2 or above
 
-# --8<-- [end:requirements]
-# --8<-- [start:set-up-using-python]
+--8<-- [end:requirements]
+--8<-- [start:set-up-using-python]
 
 The vLLM wheel bundles PyTorch and all required dependencies, and you should use the included PyTorch for compatibility. Because vLLM compiles many ROCm kernels to ensure a validated, high‑performance stack, the resulting binaries may not be compatible with other ROCm or PyTorch builds.
 If you need a different ROCm version or want to use an existing PyTorch installation, you’ll need to build vLLM from source.  See [below](#build-wheel-from-source) for more details.
 
-# --8<-- [end:set-up-using-python]
-# --8<-- [start:pre-built-wheels]
+--8<-- [end:set-up-using-python]
+--8<-- [start:pre-built-wheels]
 
 To install the latest version of vLLM for Python 3.12, ROCm 7.0 and `glibc >= 2.35`.
 
@@ -34,7 +35,7 @@ To install a specific version and ROCm variant of vLLM wheel.
 uv pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/0.15.0/rocm700
 ```
 
-!!! warning "Caveats for using `pip`" 
+!!! warning "Caveats for using `pip`"
 
     We recommend leveraging `uv` to install vLLM wheel. Using `pip` to install from custom indices is cumbersome, because `pip` combines packages from `--extra-index-url` and the default index, choosing only the latest version, which makes it difficult to install wheel from custom index if exact versions of all packages are specified exactly. In contrast, `uv` gives the extra index [higher priority than the default index](https://docs.astral.sh/uv/pip/compatibility/#packages-that-exist-on-multiple-indexes).
 
@@ -44,8 +45,8 @@ uv pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/0.15.0/rocm700
     pip install vllm==0.15.0+rocm700 --extra-index-url https://wheels.vllm.ai/rocm/0.15.0/rocm700
     ```
 
-# --8<-- [end:pre-built-wheels]
-# --8<-- [start:build-wheel-from-source]
+--8<-- [end:pre-built-wheels]
+--8<-- [start:build-wheel-from-source]
 
 !!! tip
     - If you found that the following installation step does not work for you, please refer to [docker/Dockerfile.rocm_base](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm_base). Dockerfile is a form of installation steps.
@@ -104,7 +105,6 @@ uv pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/0.15.0/rocm700
     !!! note
         - The validated `$FA_BRANCH` can be found in the [docker/Dockerfile.rocm_base](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm_base).
 
-
 3. Optionally, if you choose to build AITER yourself to use a certain branch or commit, you can build AITER using the following steps:
 
     ```bash
@@ -120,7 +120,6 @@ uv pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/0.15.0/rocm700
         - You will need to config the `$AITER_BRANCH_OR_COMMIT` for your purpose.
         - The validated `$AITER_BRANCH_OR_COMMIT` can be found in the [docker/Dockerfile.rocm_base](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm_base).
 
-
 4. Optionally, if you want to use MORI for EP or PD disaggregation, you can install [MORI](https://github.com/ROCm/mori) using the following steps:
 
     ```bash
@@ -135,7 +134,6 @@ uv pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/0.15.0/rocm700
         - You will need to config the `$MORI_BRANCH_OR_COMMIT` for your purpose.
         - The validated `$MORI_BRANCH_OR_COMMIT` can be found in the [docker/Dockerfile.rocm_base](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm_base).
 
-
 5. Build vLLM. For example, vLLM on ROCM 7.0 can be built with the following steps:
 
     ???+ console "Commands"
@@ -171,8 +169,8 @@ uv pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/0.15.0/rocm700
     - For MI300x (gfx942) users, to achieve optimal performance, please refer to [MI300x tuning guide](https://rocm.docs.amd.com/en/latest/how-to/tuning-guides/mi300x/index.html) for performance optimization and tuning tips on system and workflow level.
       For vLLM, please refer to [vLLM performance optimization](https://rocm.docs.amd.com/en/latest/how-to/rocm-for-ai/inference-optimization/vllm-optimization.html).
 
-# --8<-- [end:build-wheel-from-source]
-# --8<-- [start:pre-built-images]
+--8<-- [end:build-wheel-from-source]
+--8<-- [start:pre-built-images]
 
 vLLM offers an official Docker image for deployment.
 The image can be used to run OpenAI compatible server and is available on Docker Hub as [vllm/vllm-openai-rocm](https://hub.docker.com/r/vllm/vllm-openai-rocm/tags).
@@ -217,8 +215,8 @@ rocm/vllm-dev:nightly
     Please check [LLM inference performance validation on AMD Instinct MI300X](https://rocm.docs.amd.com/en/latest/how-to/performance-validation/mi300x/vllm-benchmark.html)
     for instructions on how to use this prebuilt docker image.
 
-# --8<-- [end:pre-built-images]
-# --8<-- [start:build-image-from-source]
+--8<-- [end:pre-built-images]
+--8<-- [start:build-image-from-source]
 
 You can build and run vLLM from source via the provided [docker/Dockerfile.rocm](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm).
 
@@ -271,7 +269,6 @@ To build vllm on ROCm 7.0 for MI200 and MI300 series, you can use the default (w
 DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile.rocm -t vllm/vllm-openai-rocm .
 ```
 
-
 To run vLLM with the custom-built Docker image:
 
 ```bash
@@ -308,9 +305,9 @@ To use the docker image as base for development, you can launch it in interactiv
         vllm/vllm-openai-rocm
     ```
 
-# --8<-- [end:build-image-from-source]
-# --8<-- [start:supported-features]
+--8<-- [end:build-image-from-source]
+--8<-- [start:supported-features]
 
 See [Feature x Hardware](../../features/README.md#feature-x-hardware) compatibility matrix for feature support information.
 
-# --8<-- [end:supported-features]
+--8<-- [end:supported-features]

docs/getting_started/installation/gpu.xpu.inc.md

@@ -1,32 +1,32 @@
-# --8<-- [start:installation]
+<!-- markdownlint-disable MD041 -->
+--8<-- [start:installation]
 
 vLLM initially supports basic model inference and serving on Intel GPU platform.
 
-# --8<-- [end:installation]
-# --8<-- [start:requirements]
+--8<-- [end:installation]
+--8<-- [start:requirements]
 
 - Supported Hardware: Intel Data Center GPU, Intel ARC GPU
-- OneAPI requirements: oneAPI 2025.3
-- Dependency: [vllm-xpu-kernels](https://github.com/vllm-project/vllm-xpu-kernels): a package provide all necessary vllm custom kernel when running vLLM on Intel GPU platform, 
+- Dependency: [vllm-xpu-kernels](https://github.com/vllm-project/vllm-xpu-kernels): a package provide all necessary vllm custom kernel when running vLLM on Intel GPU platform,
 - Python: 3.12
 !!! warning
     The provided vllm-xpu-kernels whl is Python3.12 specific so this version is a MUST.
 
-# --8<-- [end:requirements]
-# --8<-- [start:set-up-using-python]
+--8<-- [end:requirements]
+--8<-- [start:set-up-using-python]
 
 There is no extra information on creating a new Python environment for this device.
 
-# --8<-- [end:set-up-using-python]
-# --8<-- [start:pre-built-wheels]
+--8<-- [end:set-up-using-python]
+--8<-- [start:pre-built-wheels]
 
 Currently, there are no pre-built XPU wheels.
 
-# --8<-- [end:pre-built-wheels]
-# --8<-- [start:build-wheel-from-source]
+--8<-- [end:pre-built-wheels]
+--8<-- [start:build-wheel-from-source]
 
-- First, install required [driver](https://dgpu-docs.intel.com/driver/installation.html#installing-gpu-drivers) and [Intel OneAPI](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) 2025.3 or later.
-- Second, install Python packages for vLLM XPU backend building:
+- First, install required [driver](https://dgpu-docs.intel.com/driver/installation.html#installing-gpu-drivers).
+- Second, install Python packages for vLLM XPU backend building (Intel OneAPI dependencies are installed automatically as part of `torch-xpu`, see [PyTorch XPU get started](https://docs.pytorch.org/docs/stable/notes/get_start_xpu.html)):
 
 ```bash
 git clone https://github.com/vllm-project/vllm.git
@@ -35,19 +35,32 @@ pip install --upgrade pip
 pip install -v -r requirements/xpu.txt
 ```
 
-- Then, build and install vLLM XPU backend:
+- Then, install the correct Triton package for Intel XPU.
+
+    The default `triton` package (for NVIDIA GPUs) may be installed as a transitive dependency (e.g., via `xgrammar`). For Intel XPU, you must replace it with `triton-xpu`:
+
+    ```bash
+    pip uninstall -y triton triton-xpu
+    pip install triton-xpu==3.6.0 --extra-index-url https://download.pytorch.org/whl/xpu
+    ```
+
+    !!! note
+        - `triton` (without suffix) is for NVIDIA GPUs only. On XPU, using it instead of `triton-xpu` can cause correctness or runtime issues.
+        - For torch 2.10 (the version used in `requirements/xpu.txt`), the matching package is `triton-xpu==3.6.0`. If you use a different version of torch, check the corresponding `triton-xpu` version in [docker/Dockerfile.xpu](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.xpu).
+
+- Finally, build and install vLLM XPU backend:
 
 ```bash
 VLLM_TARGET_DEVICE=xpu pip install --no-build-isolation -e . -v
 ```
 
-# --8<-- [end:build-wheel-from-source]
-# --8<-- [start:pre-built-images]
+--8<-- [end:build-wheel-from-source]
+--8<-- [start:pre-built-images]
 
 Currently, we release prebuilt XPU images at docker [hub](https://hub.docker.com/r/intel/vllm/tags) based on vLLM released version. For more information, please refer release [note](https://github.com/intel/ai-containers/blob/main/vllm).
 
-# --8<-- [end:pre-built-images]
-# --8<-- [start:build-image-from-source]
+--8<-- [end:pre-built-images]
+--8<-- [start:build-image-from-source]
 
 ```bash
 docker build -f docker/Dockerfile.xpu -t vllm-xpu-env --shm-size=4g .
@@ -61,8 +74,8 @@ docker run -it \
              vllm-xpu-env
 ```
 
-# --8<-- [end:build-image-from-source]
-# --8<-- [start:supported-features]
+--8<-- [end:build-image-from-source]
+--8<-- [start:supported-features]
 
 XPU platform supports **tensor parallel** inference/serving and also supports **pipeline parallel** as a beta feature for online serving. For **pipeline parallel**, we support it on single node with mp as the backend. For example, a reference execution like following:
 
@@ -77,9 +90,9 @@ vllm serve facebook/opt-13b \
 
 By default, a ray instance will be launched automatically if no existing one is detected in the system, with `num-gpus` equals to `parallel_config.world_size`. We recommend properly starting a ray cluster before execution, referring to the [examples/online_serving/run_cluster.sh](https://github.com/vllm-project/vllm/blob/main/examples/online_serving/run_cluster.sh) helper script.
 
-# --8<-- [end:supported-features]
-# --8<-- [start:distributed-backend]
+--8<-- [end:supported-features]
+--8<-- [start:distributed-backend]
 
 XPU platform uses **torch-ccl** for torch<2.8 and **xccl** for torch>=2.8 as distributed backend, since torch 2.8 supports **xccl** as built-in backend for XPU.
 
-# --8<-- [end:distributed-backend]
+--8<-- [end:distributed-backend]

docs/getting_started/installation/python_env_setup.inc.md

@@ -1,3 +1,4 @@
+<!-- markdownlint-disable MD041 -->
 It's recommended to use [uv](https://docs.astral.sh/uv/), a very fast Python environment manager, to create and manage Python environments. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following commands:
 
 ```bash

docs/getting_started/quickstart.md

@@ -75,7 +75,7 @@ This guide will help you quickly get started with vLLM to perform:
 
 ## Offline Batched Inference
 
-With vLLM installed, you can start generating texts for list of input prompts (i.e. offline batch inferencing). See the example script: [examples/offline_inference/basic/basic.py](../../examples/offline_inference/basic/basic.py)
+With vLLM installed, you can start generating texts for list of input prompts (i.e. offline batch inferencing). See the example script: [examples/basic/offline_inference/basic.py](../../examples/basic/offline_inference/basic.py)
 
 The first line of this example imports the classes [LLM][vllm.LLM] and [SamplingParams][vllm.SamplingParams]:
 
@@ -228,7 +228,7 @@ Since this server is compatible with OpenAI API, you can use it as a drop-in rep
     print("Completion result:", completion)
     ```
 
-A more detailed client example can be found here: [examples/offline_inference/basic/basic.py](../../examples/offline_inference/basic/basic.py)
+A more detailed client example can be found here: [examples/basic/offline_inference/basic.py](../../examples/basic/offline_inference/basic.py)
 
 ### OpenAI Chat Completions API with vLLM
 

docs/maybe_skip_pr_build.sh

@@ -19,6 +19,6 @@ if [[ "$HTTP_CODE" -ne 200 ]]; then
 elif grep -qE '"name": *"(documentation|ready)"' /tmp/pr_response.json; then
   echo "Found required label, proceeding with build."
 else
-  echo "PR #${READTHEDOCS_VERSION} lacks 'documentation' or 'ready' label, skipping build."
-  exit 183
+  echo "PR #${READTHEDOCS_VERSION} lacks 'documentation' or 'ready' label, cancelling build."
+  exit 1
 fi

docs/mkdocs/javascript/reo.js

@@ -1,3 +0,0 @@
-// Reo.Dev documentation tracking
-// https://docs.reo.dev/integrations/tracking-beacon/install-javascript-for-documentation
-!function(){var e,t,n;e="d5c4337961ef0ac",t=function(){Reo.init({clientID:"d5c4337961ef0ac"})},(n=document.createElement("script")).src="https://static.reo.dev/"+e+"/reo.js",n.defer=!0,n.onload=t,document.head.appendChild(n)}();

docs/models/generative_models.md

@@ -59,7 +59,7 @@ for output in outputs:
     By default, vLLM will use sampling parameters recommended by model creator by applying the `generation_config.json` from the huggingface model repository if it exists. In most cases, this will provide you with the best results by default if [SamplingParams][vllm.SamplingParams] is not specified.
 
     However, if vLLM's default sampling parameters are preferred, please pass `generation_config="vllm"` when creating the [LLM][vllm.LLM] instance.
-A code example can be found here: [examples/offline_inference/basic/basic.py](../../examples/offline_inference/basic/basic.py)
+A code example can be found here: [examples/basic/offline_inference/basic.py](../../examples/basic/offline_inference/basic.py)
 
 ### `LLM.beam_search`
 
@@ -121,7 +121,7 @@ and automatically applies the model's [chat template](https://huggingface.co/doc
         print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
     ```
 
-A code example can be found here: [examples/offline_inference/basic/chat.py](../../examples/offline_inference/basic/chat.py)
+A code example can be found here: [examples/basic/offline_inference/chat.py](../../examples/basic/offline_inference/chat.py)
 
 If the model doesn't have a chat template or you want to specify another one,
 you can explicitly pass a chat template:

docs/models/hardware_supported_models/cpu.md

@@ -2,32 +2,32 @@
 
 ## Validated Hardware
 
-| Hardware                                 |
-| ----------------------------------------- |
-| [Intel® Xeon® 6 Processors](https://www.intel.com/content/www/us/en/products/details/processors/xeon.html)                   |
-| [Intel® Xeon® 5 Processors](https://www.intel.com/content/www/us/en/products/docs/processors/xeon/5th-gen-xeon-scalable-processors.html)              |
+| Hardware |
+| -------- |
+| [Intel® Xeon® 6 Processors](https://www.intel.com/content/www/us/en/products/details/processors/xeon.html) |
+| [Intel® Xeon® 5 Processors](https://www.intel.com/content/www/us/en/products/docs/processors/xeon/5th-gen-xeon-scalable-processors.html) |
 
 ## Recommended Models
 
 ### Text-only Language Models
 
 | Model                                | Architecture                             | Supported |
-|--------------------------------------|-------------------------------------------|-----------|
-| meta-llama/Llama-3.1-8B-Instruct     | LlamaForCausalLM                          | ✅        |
-| meta-llama/Llama-3.2-3B-Instruct     | LlamaForCausalLM                          | ✅        |
-| ibm-granite/granite-3.2-2b-instruct  | GraniteForCausalLM                        | ✅        |
-| Qwen/Qwen3-1.7B                      | Qwen3ForCausalLM                          | ✅        |
-| Qwen/Qwen3-4B                        | Qwen3ForCausalLM                          | ✅        |
-| Qwen/Qwen3-8B                        | Qwen3ForCausalLM                          | ✅        |
-| zai-org/glm-4-9b-hf                  | GLMForCausalLM                            | ✅        |
-| google/gemma-7b                      | GemmaForCausalLM                          | ✅        |
+| ------------------------------------ | ---------------------------------------- | --------- |
+| meta-llama/Llama-3.1-8B-Instruct     | LlamaForCausalLM                         | ✅        |
+| meta-llama/Llama-3.2-3B-Instruct     | LlamaForCausalLM                         | ✅        |
+| ibm-granite/granite-3.2-2b-instruct  | GraniteForCausalLM                       | ✅        |
+| Qwen/Qwen3-1.7B                      | Qwen3ForCausalLM                         | ✅        |
+| Qwen/Qwen3-4B                        | Qwen3ForCausalLM                         | ✅        |
+| Qwen/Qwen3-8B                        | Qwen3ForCausalLM                         | ✅        |
+| zai-org/glm-4-9b-hf                  | GLMForCausalLM                           | ✅        |
+| google/gemma-7b                      | GemmaForCausalLM                         | ✅        |
 
 ### Multimodal Language Models
 
 | Model                                | Architecture                             | Supported |
-|--------------------------------------|-------------------------------------------|-----------|
-| Qwen/Qwen2.5-VL-7B-Instruct          | Qwen2VLForConditionalGeneration           | ✅        |
-| openai/whisper-large-v3              | WhisperForConditionalGeneration           | ✅        |
+| ------------------------------------ | ---------------------------------------- | --------- |
+| Qwen/Qwen2.5-VL-7B-Instruct          | Qwen2VLForConditionalGeneration          | ✅        |
+| openai/whisper-large-v3              | WhisperForConditionalGeneration          | ✅        |
 
 ✅ Runs and optimized.  
 🟨 Runs and correct but not optimized to green yet.  

docs/models/hardware_supported_models/xpu.md

@@ -2,9 +2,9 @@
 
 ## Validated Hardware
 
-| Hardware                                 |
-| ----------------------------------------- |
-| [Intel® Arc™ Pro B-Series Graphics](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/workstations/b-series/overview.html)                   |
+| Hardware |
+| -------- |
+| [Intel® Arc™ Pro B-Series Graphics](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/workstations/b-series/overview.html) |
 
 ## Recommended Models
 
@@ -12,53 +12,53 @@
 
 | Model                                     | Architecture                                         | FP16 | Dynamic FP8 | MXFP4 |
 | ----------------------------------------- | ---------------------------------------------------- | ---- | ----------- | ----- |
-| openai/gpt-oss-20b                        | GPTForCausalLM                                       |      |             | ✅     |
-| openai/gpt-oss-120b                       | GPTForCausalLM                                       |      |             | ✅     |
-| deepseek-ai/DeepSeek-R1-Distill-Llama-8B  | LlamaForCausalLM                                     | ✅    | ✅           |       |
-| deepseek-ai/DeepSeek-R1-Distill-Qwen-14B  | QwenForCausalLM                                      | ✅    | ✅           |       |
-| deepseek-ai/DeepSeek-R1-Distill-Qwen-32B  | QwenForCausalLM                                      | ✅    | ✅           |       |
-| deepseek-ai/DeepSeek-R1-Distill-Llama-70B | LlamaForCausalLM                                     | ✅    | ✅           |       |
-| Qwen/Qwen2.5-72B-Instruct                 | Qwen2ForCausalLM                                     | ✅    | ✅           |       |
-| Qwen/Qwen3-14B                            | Qwen3ForCausalLM                                     | ✅    | ✅           |       |
-| Qwen/Qwen3-32B                            | Qwen3ForCausalLM                                     | ✅    | ✅           |       |
-| Qwen/Qwen3-30B-A3B                        | Qwen3ForCausalLM                                     | ✅    | ✅           |       |
-| Qwen/Qwen3-30B-A3B-GPTQ-Int4              | Qwen3ForCausalLM                                     | ✅    | ✅           |       |
-| Qwen/Qwen3-coder-30B-A3B-Instruct         | Qwen3ForCausalLM                                     | ✅    | ✅           |       |
-| Qwen/QwQ-32B                              | QwenForCausalLM                                      | ✅    | ✅           |       |
-| deepseek-ai/DeepSeek-V2-Lite              | DeepSeekForCausalLM                                  | ✅    | ✅           |       |
-| meta-llama/Llama-3.1-8B-Instruct          | LlamaForCausalLM                                     | ✅    | ✅           |       |
-| baichuan-inc/Baichuan2-13B-Chat           | BaichuanForCausalLM                                  | ✅    | ✅           |       |
-| THUDM/GLM-4-9B-chat                       | GLMForCausalLM                                       | ✅    | ✅           |       |
-| THUDM/CodeGeex4-All-9B                    | CodeGeexForCausalLM                                  | ✅    | ✅           |       |
-| chuhac/TeleChat2-35B                      | LlamaForCausalLM (TeleChat2 based on Llama arch)     | ✅    | ✅           |       |
-| 01-ai/Yi1.5-34B-Chat                      | YiForCausalLM                                        | ✅    | ✅           |       |
-| THUDM/CodeGeex4-All-9B                    | CodeGeexForCausalLM                                  | ✅    | ✅           |       |
-| deepseek-ai/DeepSeek-Coder-33B-base       | DeepSeekCoderForCausalLM                             | ✅    | ✅           |       |
-| baichuan-inc/Baichuan2-13B-Chat           | BaichuanForCausalLM                                  | ✅    | ✅           |       |
-| meta-llama/Llama-2-13b-chat-hf            | LlamaForCausalLM                                     | ✅    | ✅           |       |
-| THUDM/CodeGeex4-All-9B                    | CodeGeexForCausalLM                                  | ✅    | ✅           |       |
-| Qwen/Qwen1.5-14B-Chat                     | QwenForCausalLM                                      | ✅    | ✅           |       |
-| Qwen/Qwen1.5-32B-Chat                     | QwenForCausalLM                                      | ✅    | ✅           |       |
+| openai/gpt-oss-20b                        | GPTForCausalLM                                       |      |             | ✅    |
+| openai/gpt-oss-120b                       | GPTForCausalLM                                       |      |             | ✅    |
+| deepseek-ai/DeepSeek-R1-Distill-Llama-8B  | LlamaForCausalLM                                     | ✅   | ✅          |       |
+| deepseek-ai/DeepSeek-R1-Distill-Qwen-14B  | QwenForCausalLM                                      | ✅   | ✅          |       |
+| deepseek-ai/DeepSeek-R1-Distill-Qwen-32B  | QwenForCausalLM                                      | ✅   | ✅          |       |
+| deepseek-ai/DeepSeek-R1-Distill-Llama-70B | LlamaForCausalLM                                     | ✅   | ✅          |       |
+| Qwen/Qwen2.5-72B-Instruct                 | Qwen2ForCausalLM                                     | ✅   | ✅          |       |
+| Qwen/Qwen3-14B                            | Qwen3ForCausalLM                                     | ✅   | ✅          |       |
+| Qwen/Qwen3-32B                            | Qwen3ForCausalLM                                     | ✅   | ✅          |       |
+| Qwen/Qwen3-30B-A3B                        | Qwen3ForCausalLM                                     | ✅   | ✅          |       |
+| Qwen/Qwen3-30B-A3B-GPTQ-Int4              | Qwen3ForCausalLM                                     | ✅   | ✅          |       |
+| Qwen/Qwen3-coder-30B-A3B-Instruct         | Qwen3ForCausalLM                                     | ✅   | ✅          |       |
+| Qwen/QwQ-32B                              | QwenForCausalLM                                      | ✅   | ✅          |       |
+| deepseek-ai/DeepSeek-V2-Lite              | DeepSeekForCausalLM                                  | ✅   | ✅          |       |
+| meta-llama/Llama-3.1-8B-Instruct          | LlamaForCausalLM                                     | ✅   | ✅          |       |
+| baichuan-inc/Baichuan2-13B-Chat           | BaichuanForCausalLM                                  | ✅   | ✅          |       |
+| THUDM/GLM-4-9B-chat                       | GLMForCausalLM                                       | ✅   | ✅          |       |
+| THUDM/CodeGeex4-All-9B                    | CodeGeexForCausalLM                                  | ✅   | ✅          |       |
+| chuhac/TeleChat2-35B                      | LlamaForCausalLM (TeleChat2 based on Llama arch)     | ✅   | ✅          |       |
+| 01-ai/Yi1.5-34B-Chat                      | YiForCausalLM                                        | ✅   | ✅          |       |
+| THUDM/CodeGeex4-All-9B                    | CodeGeexForCausalLM                                  | ✅   | ✅          |       |
+| deepseek-ai/DeepSeek-Coder-33B-base       | DeepSeekCoderForCausalLM                             | ✅   | ✅          |       |
+| baichuan-inc/Baichuan2-13B-Chat           | BaichuanForCausalLM                                  | ✅   | ✅          |       |
+| meta-llama/Llama-2-13b-chat-hf            | LlamaForCausalLM                                     | ✅   | ✅          |       |
+| THUDM/CodeGeex4-All-9B                    | CodeGeexForCausalLM                                  | ✅   | ✅          |       |
+| Qwen/Qwen1.5-14B-Chat                     | QwenForCausalLM                                      | ✅   | ✅          |       |
+| Qwen/Qwen1.5-32B-Chat                     | QwenForCausalLM                                      | ✅   | ✅          |       |
 
 ### Multimodal Language Models
 
 | Model                        | Architecture                     | FP16 | Dynamic FP8 | MXFP4 |
 | ---------------------------- | -------------------------------- | ---- | ----------- | ----- |
-| OpenGVLab/InternVL3_5-8B     | InternVLForConditionalGeneration | ✅    | ✅           |       |
-| OpenGVLab/InternVL3_5-14B    | InternVLForConditionalGeneration | ✅    | ✅           |       |
-| OpenGVLab/InternVL3_5-38B    | InternVLForConditionalGeneration | ✅    | ✅           |       |
-| Qwen/Qwen2-VL-7B-Instruct    | Qwen2VLForConditionalGeneration  | ✅    | ✅           |       |
-| Qwen/Qwen2.5-VL-72B-Instruct | Qwen2VLForConditionalGeneration  | ✅    | ✅           |       |
-| Qwen/Qwen2.5-VL-32B-Instruct | Qwen2VLForConditionalGeneration  | ✅    | ✅           |       |
-| THUDM/GLM-4v-9B              | GLM4vForConditionalGeneration    | ✅    | ✅           |       |
-| openbmb/MiniCPM-V-4          | MiniCPMVForConditionalGeneration | ✅    | ✅           |       |
+| OpenGVLab/InternVL3_5-8B     | InternVLForConditionalGeneration | ✅   | ✅          |       |
+| OpenGVLab/InternVL3_5-14B    | InternVLForConditionalGeneration | ✅   | ✅          |       |
+| OpenGVLab/InternVL3_5-38B    | InternVLForConditionalGeneration | ✅   | ✅          |       |
+| Qwen/Qwen2-VL-7B-Instruct    | Qwen2VLForConditionalGeneration  | ✅   | ✅          |       |
+| Qwen/Qwen2.5-VL-72B-Instruct | Qwen2VLForConditionalGeneration  | ✅   | ✅          |       |
+| Qwen/Qwen2.5-VL-32B-Instruct | Qwen2VLForConditionalGeneration  | ✅   | ✅          |       |
+| THUDM/GLM-4v-9B              | GLM4vForConditionalGeneration    | ✅   | ✅          |       |
+| openbmb/MiniCPM-V-4          | MiniCPMVForConditionalGeneration | ✅   | ✅          |       |
 
 ### Embedding and Reranker Language Models
 
 | Model                   | Architecture                   | FP16 | Dynamic FP8 | MXFP4 |
 | ----------------------- | ------------------------------ | ---- | ----------- | ----- |
-| Qwen/Qwen3-Embedding-8B | Qwen3ForTextEmbedding          | ✅    | ✅           |       |
-| Qwen/Qwen3-Reranker-8B  | Qwen3ForSequenceClassification | ✅    | ✅           |       |
+| Qwen/Qwen3-Embedding-8B | Qwen3ForTextEmbedding          | ✅   | ✅          |       |
+| Qwen/Qwen3-Reranker-8B  | Qwen3ForSequenceClassification | ✅   | ✅          |       |
 
 ✅ Runs and optimized.  
 🟨 Runs and correct but not optimized to green yet.  

docs/models/pooling_models.md

@@ -31,7 +31,7 @@ vLLM will attempt to automatically convert the model according to the architectu
 shown in the table below.
 
 | Architecture                                    | `--convert` | Supported pooling tasks               |
-|-------------------------------------------------|-------------|---------------------------------------|
+| ----------------------------------------------- | ----------- | ------------------------------------- |
 | `*ForTextEncoding`, `*EmbeddingModel`, `*Model` | `embed`     | `token_embed`, `embed`                |
 | `*ForRewardModeling`, `*RewardModel`            | `embed`     | `token_embed`, `embed`                |
 | `*For*Classification`, `*ClassificationModel`   | `classify`  | `token_classify`, `classify`, `score` |
@@ -46,7 +46,7 @@ Each pooling model in vLLM supports one or more of these tasks according to
 enabling the corresponding APIs:
 
 | Task             | APIs                                                                          |
-|------------------|-------------------------------------------------------------------------------|
+| ---------------- | ----------------------------------------------------------------------------- |
 | `embed`          | `LLM.embed(...)`, `LLM.score(...)`\*, `LLM.encode(..., pooling_task="embed")` |
 | `classify`       | `LLM.classify(...)`, `LLM.encode(..., pooling_task="classify")`               |
 | `score`          | `LLM.score(...)`                                                              |
@@ -69,7 +69,7 @@ If the model has been converted via `--convert` (see above),
 the pooler assigned to each task has the following attributes by default:
 
 | Task       | Pooling Type | Normalization | Softmax |
-|------------|--------------|---------------|---------|
+| ---------- | ------------ | ------------- | ------- |
 | `embed`    | `LAST`       | ✅︎            | ❌      |
 | `classify` | `LAST`       | ❌            | ✅︎      |
 
@@ -99,7 +99,7 @@ embeds = output.outputs.embedding
 print(f"Embeddings: {embeds!r} (size={len(embeds)})")
 ```
 
-A code example can be found here: [examples/offline_inference/basic/embed.py](../../examples/offline_inference/basic/embed.py)
+A code example can be found here: [examples/basic/offline_inference/embed.py](../../examples/basic/offline_inference/embed.py)
 
 ### `LLM.classify`
 
@@ -116,7 +116,7 @@ probs = output.outputs.probs
 print(f"Class Probabilities: {probs!r} (size={len(probs)})")
 ```
 
-A code example can be found here: [examples/offline_inference/basic/classify.py](../../examples/offline_inference/basic/classify.py)
+A code example can be found here: [examples/basic/offline_inference/classify.py](../../examples/basic/offline_inference/classify.py)
 
 ### `LLM.score`
 
@@ -140,7 +140,7 @@ score = output.outputs.score
 print(f"Score: {score}")
 ```
 
-A code example can be found here: [examples/offline_inference/basic/score.py](../../examples/offline_inference/basic/score.py)
+A code example can be found here: [examples/basic/offline_inference/score.py](../../examples/basic/offline_inference/score.py)
 
 ### `LLM.reward`
 
@@ -156,7 +156,7 @@ data = output.outputs.data
 print(f"Data: {data!r}")
 ```
 
-A code example can be found here: [examples/offline_inference/basic/reward.py](../../examples/offline_inference/basic/reward.py)
+A code example can be found here: [examples/basic/offline_inference/reward.py](../../examples/basic/offline_inference/reward.py)
 
 ### `LLM.encode`
 
@@ -314,7 +314,7 @@ An OpenAI client example can be found here: [examples/pooling/embed/openai_embed
 vLLM supports ColBERT models with multiple encoder backbones:
 
 | Architecture | Backbone | Example HF Models |
-|---|---|---|
+| - | - | - |
 | `HF_ColBERT` | BERT | `answerdotai/answerai-colbert-small-v1`, `colbert-ir/colbertv2.0` |
 | `ColBERTModernBertModel` | ModernBERT | `lightonai/GTE-ModernColBERT-v1` |
 | `ColBERTJinaRobertaModel` | Jina XLM-RoBERTa | `jinaai/jina-colbert-v2` |
@@ -379,7 +379,7 @@ An example can be found here: [examples/pooling/score/colbert_rerank_online.py](
 ColQwen3 is based on [ColPali](https://arxiv.org/abs/2407.01449), which extends ColBERT's late interaction approach to **multi-modal** inputs. While ColBERT operates on text-only token embeddings, ColPali/ColQwen3 can embed both **text and images** (e.g. PDF pages, screenshots, diagrams) into per-token L2-normalized vectors and compute relevance via MaxSim scoring. ColQwen3 specifically uses Qwen3-VL as its vision-language backbone.
 
 | Architecture | Backbone | Example HF Models |
-|---|---|---|
+| - | - | - |
 | `ColQwen3` | Qwen3-VL | `TomoroAI/tomoro-colqwen3-embed-4b`, `TomoroAI/tomoro-colqwen3-embed-8b` |
 | `OpsColQwen3Model` | Qwen3-VL | `OpenSearch-AI/Ops-Colqwen3-4B`, `OpenSearch-AI/Ops-Colqwen3-8B` |
 | `Qwen3VLNemotronEmbedModel` | Qwen3-VL | `nvidia/nemotron-colembed-vl-4b-v2`, `nvidia/nemotron-colembed-vl-8b-v2` |
@@ -507,7 +507,7 @@ Llama Nemotron VL Embedding models combine the bidirectional Llama embedding bac
 single-vector embeddings from text and/or images.
 
 | Architecture | Backbone | Example HF Models |
-|---|---|---|
+| - | - | - |
 | `LlamaNemotronVLModel` | Bidirectional Llama + SigLIP | `nvidia/llama-nemotron-embed-vl-1b-v2` |
 
 Start the server:
@@ -567,7 +567,7 @@ Llama Nemotron VL reranker models combine the same bidirectional Llama + SigLIP
 backbone with a sequence-classification head for cross-encoder scoring and reranking.
 
 | Architecture | Backbone | Example HF Models |
-|---|---|---|
+| - | - | - |
 | `LlamaNemotronVLForSequenceClassification` | Bidirectional Llama + SigLIP | `nvidia/llama-nemotron-rerank-vl-1b-v2` |
 
 Start the server:

docs/models/supported_models.md

@@ -179,7 +179,7 @@ class MyConfig(PretrainedConfig):
 Some model architectures are supported via vLLM plugins. These plugins extend vLLM's capabilities through the [plugin system](../design/plugin_system.md).
 
 | Architecture | Models | Plugin Repository |
-|--------------|--------|-------------------|
+| ------------ | ------ | ----------------- |
 | `BartForConditionalGeneration` | BART | [bart-plugin](https://github.com/vllm-project/bart-plugin) |
 | `Florence2ForConditionalGeneration` | Florence-2 | [bart-plugin](https://github.com/vllm-project/bart-plugin) |
 
@@ -363,7 +363,7 @@ th {
 </style>
 
 | Architecture | Models | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|-------------------|----------------------|---------------------------|
+| ------------ | ------ | ----------------- | -------------------- | ------------------------- |
 | `AfmoeForCausalLM` | Afmoe | TBA | ✅︎ | ✅︎ |
 | `ApertusForCausalLM` | Apertus | `swiss-ai/Apertus-8B-2509`, `swiss-ai/Apertus-70B-Instruct-2509`, etc. | ✅︎ | ✅︎ |
 | `AquilaForCausalLM` | Aquila, Aquila2 | `BAAI/Aquila-7B`, `BAAI/AquilaChat-7B`, etc. | ✅︎ | ✅︎ |
@@ -387,7 +387,7 @@ th {
 | `Dots1ForCausalLM` | dots.llm1 | `rednote-hilab/dots.llm1.base`, `rednote-hilab/dots.llm1.inst`, etc. | | ✅︎ |
 | `DotsOCRForCausalLM` | dots_ocr | `rednote-hilab/dots.ocr` | ✅︎ | ✅︎ |
 | `Ernie4_5ForCausalLM` | Ernie4.5 | `baidu/ERNIE-4.5-0.3B-PT`, etc. | ✅︎ | ✅︎ |
-| `Ernie4_5_MoeForCausalLM` | Ernie4.5MoE | `baidu/ERNIE-4.5-21B-A3B-PT`, `baidu/ERNIE-4.5-300B-A47B-PT`, etc. |✅︎| ✅︎ |
+| `Ernie4_5_MoeForCausalLM` | Ernie4.5MoE | `baidu/ERNIE-4.5-21B-A3B-PT`, `baidu/ERNIE-4.5-300B-A47B-PT`, etc. | ✅︎ | ✅︎ |
 | `ExaoneForCausalLM` | EXAONE-3 | `LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct`, etc. | ✅︎ | ✅︎ |
 | `ExaoneMoEForCausalLM` | K-EXAONE | `LGAI-EXAONE/K-EXAONE-236B-A23B`, etc. | | |
 | `Exaone4ForCausalLM` | EXAONE-4 | `LGAI-EXAONE/EXAONE-4.0-32B`, etc. | ✅︎ | ✅︎ |
@@ -427,18 +427,18 @@ th {
 | `Jais2ForCausalLM` | Jais2 | `inceptionai/Jais-2-8B-Chat`, `inceptionai/Jais-2-70B-Chat`, etc. | | ✅︎ |
 | `JambaForCausalLM` | Jamba | `ai21labs/AI21-Jamba-1.5-Large`, `ai21labs/AI21-Jamba-1.5-Mini`, `ai21labs/Jamba-v0.1`, etc. | ✅︎ | ✅︎ |
 | `KimiLinearForCausalLM` | Kimi-Linear-48B-A3B-Base, Kimi-Linear-48B-A3B-Instruct | `moonshotai/Kimi-Linear-48B-A3B-Base`, `moonshotai/Kimi-Linear-48B-A3B-Instruct` | | ✅︎ |
-| `Lfm2ForCausalLM`  | LFM2  | `LiquidAI/LFM2-1.2B`, `LiquidAI/LFM2-700M`, `LiquidAI/LFM2-350M`, etc. | ✅︎ | ✅︎ |
-| `Lfm2MoeForCausalLM`  | LFM2MoE  | `LiquidAI/LFM2-8B-A1B-preview`, etc. | ✅︎ | ✅︎ |
+| `Lfm2ForCausalLM` | LFM2 | `LiquidAI/LFM2-1.2B`, `LiquidAI/LFM2-700M`, `LiquidAI/LFM2-350M`, etc. | ✅︎ | ✅︎ |
+| `Lfm2MoeForCausalLM` | LFM2MoE | `LiquidAI/LFM2-8B-A1B-preview`, etc. | ✅︎ | ✅︎ |
 | `LlamaForCausalLM` | Llama 3.1, Llama 3, Llama 2, LLaMA, Yi | `meta-llama/Meta-Llama-3.1-405B-Instruct`, `meta-llama/Meta-Llama-3.1-70B`, `meta-llama/Meta-Llama-3-70B-Instruct`, `meta-llama/Llama-2-70b-hf`, `01-ai/Yi-34B`, etc. | ✅︎ | ✅︎ |
 | `LongcatFlashForCausalLM` | LongCat-Flash | `meituan-longcat/LongCat-Flash-Chat`, `meituan-longcat/LongCat-Flash-Chat-FP8` | ✅︎ | ✅︎ |
 | `MambaForCausalLM` | Mamba | `state-spaces/mamba-130m-hf`, `state-spaces/mamba-790m-hf`, `state-spaces/mamba-2.8b-hf`, etc. | | ✅︎ |
 | `Mamba2ForCausalLM` | Mamba2 | `mistralai/Mamba-Codestral-7B-v0.1`, etc. | | ✅︎ |
 | `MiMoForCausalLM` | MiMo | `XiaomiMiMo/MiMo-7B-RL`, etc. | ✅︎ | ✅︎ |
-| `MiMoV2FlashForCausalLM` | MiMoV2Flash | `XiaomiMiMo/MiMo-V2-Flash`, etc. | ︎| ✅︎ |
+| `MiMoV2FlashForCausalLM` | MiMoV2Flash | `XiaomiMiMo/MiMo-V2-Flash`, etc. | | ✅︎ |
 | `MiniCPMForCausalLM` | MiniCPM | `openbmb/MiniCPM-2B-sft-bf16`, `openbmb/MiniCPM-2B-dpo-bf16`, `openbmb/MiniCPM-S-1B-sft`, etc. | ✅︎ | ✅︎ |
 | `MiniCPM3ForCausalLM` | MiniCPM3 | `openbmb/MiniCPM3-4B`, etc. | ✅︎ | ✅︎ |
 | `MiniMaxForCausalLM` | MiniMax-Text | `MiniMaxAI/MiniMax-Text-01-hf`, etc. | | |
-| `MiniMaxM2ForCausalLM` | MiniMax-M2, MiniMax-M2.1 |`MiniMaxAI/MiniMax-M2`, etc. | ✅︎ | ✅︎ |
+| `MiniMaxM2ForCausalLM` | MiniMax-M2, MiniMax-M2.1 | `MiniMaxAI/MiniMax-M2`, etc. | ✅︎ | ✅︎ |
 | `MistralForCausalLM` | Ministral-3, Mistral, Mistral-Instruct | `mistralai/Ministral-3-3B-Instruct-2512`, `mistralai/Mistral-7B-v0.1`, `mistralai/Mistral-7B-Instruct-v0.1`, etc. | ✅︎ | ✅︎ |
 | `MistralLarge3ForCausalLM` | Mistral-Large-3-675B-Base-2512, Mistral-Large-3-675B-Instruct-2512 | `mistralai/Mistral-Large-3-675B-Base-2512`, `mistralai/Mistral-Large-3-675B-Instruct-2512`, etc. | ✅︎ | ✅︎ |
 | `MixtralForCausalLM` | Mixtral-8x7B, Mixtral-8x7B-Instruct | `mistralai/Mixtral-8x7B-v0.1`, `mistralai/Mixtral-8x7B-Instruct-v0.1`, `mistral-community/Mixtral-8x22B-v0.1`, etc. | ✅︎ | ✅︎ |
@@ -453,9 +453,9 @@ th {
 | `OPTForCausalLM` | OPT, OPT-IML | `facebook/opt-66b`, `facebook/opt-iml-max-30b`, etc. | ✅︎ | ✅︎ |
 | `OrionForCausalLM` | Orion | `OrionStarAI/Orion-14B-Base`, `OrionStarAI/Orion-14B-Chat`, etc. | | ✅︎ |
 | `OuroForCausalLM` | ouro | `ByteDance/Ouro-1.4B`, `ByteDance/Ouro-2.6B`, etc. | ✅︎ | |
-| `PanguEmbeddedForCausalLM` |openPangu-Embedded-7B | `FreedomIntelligence/openPangu-Embedded-7B-V1.1` | ✅︎ | ✅︎ |
-| `PanguProMoEV2ForCausalLM` |openpangu-pro-moe-v2 | | ✅︎ | ✅︎ |
-| `PanguUltraMoEForCausalLM` |openpangu-ultra-moe-718b-model | `FreedomIntelligence/openPangu-Ultra-MoE-718B-V1.1` | ✅︎ | ✅︎ |
+| `PanguEmbeddedForCausalLM` | openPangu-Embedded-7B | `FreedomIntelligence/openPangu-Embedded-7B-V1.1` | ✅︎ | ✅︎ |
+| `PanguProMoEV2ForCausalLM` | openpangu-pro-moe-v2 | | ✅︎ | ✅︎ |
+| `PanguUltraMoEForCausalLM` | openpangu-ultra-moe-718b-model | `FreedomIntelligence/openPangu-Ultra-MoE-718B-V1.1` | ✅︎ | ✅︎ |
 | `PhiForCausalLM` | Phi | `microsoft/phi-1_5`, `microsoft/phi-2`, etc. | ✅︎ | ✅︎ |
 | `Phi3ForCausalLM` | Phi-4, Phi-3 | `microsoft/Phi-4-mini-instruct`, `microsoft/Phi-4`, `microsoft/Phi-3-mini-4k-instruct`, `microsoft/Phi-3-mini-128k-instruct`, `microsoft/Phi-3-medium-128k-instruct`, etc. | ✅︎ | ✅︎ |
 | `PhiMoEForCausalLM` | Phi-3.5-MoE | `microsoft/Phi-3.5-MoE-instruct`, etc. | ✅︎ | ✅︎ |
@@ -469,13 +469,15 @@ th {
 | `Qwen3MoeForCausalLM` | Qwen3MoE | `Qwen/Qwen3-30B-A3B`, etc. | ✅︎ | ✅︎ |
 | `Qwen3NextForCausalLM` | Qwen3NextMoE | `Qwen/Qwen3-Next-80B-A3B-Instruct`, etc. | ✅︎ | ✅︎ |
 | `RWForCausalLM` | Falcon RW | `tiiuae/falcon-40b`, etc. | | ✅︎ |
+| `SarvamMoEForCausalLM` | Sarvam 2 | `sarvamai/sarvam2-30b-a3b`, etc. | ✅︎ | ✅︎ |
+| `SarvamMLAForCausalLM` | Sarvam 2 | `sarvamai/sarvam2-105b-a9b`, etc. | | ✅︎ |
 | `SeedOssForCausalLM` | SeedOss | `ByteDance-Seed/Seed-OSS-36B-Instruct`, etc. | ✅︎ | ✅︎ |
 | `SolarForCausalLM` | Solar Pro | `upstage/solar-pro-preview-instruct`, etc. | ✅︎ | ✅︎ |
 | `StableLmForCausalLM` | StableLM | `stabilityai/stablelm-3b-4e1t`, `stabilityai/stablelm-base-alpha-7b-v2`, etc. | | |
 | `StableLMEpochForCausalLM` | StableLM Epoch | `stabilityai/stablelm-zephyr-3b`, etc. | | ✅︎ |
 | `Starcoder2ForCausalLM` | Starcoder2 | `bigcode/starcoder2-3b`, `bigcode/starcoder2-7b`, `bigcode/starcoder2-15b`, etc. | | ✅︎ |
 | `Step1ForCausalLM` | Step-Audio | `stepfun-ai/Step-Audio-EditX`, etc. | ✅︎ | ✅︎ |
-| `Step3p5ForCausalLM` | Step-3.5-flash | `stepfun-ai/Step-3.5-Flash`, etc. |  | ✅︎ |
+| `Step3p5ForCausalLM` | Step-3.5-flash | `stepfun-ai/Step-3.5-Flash`, etc. | | ✅︎ |
 | `TeleChatForCausalLM` | TeleChat | `chuhac/TeleChat2-35B`, etc. | ✅︎ | ✅︎ |
 | `TeleChat2ForCausalLM` | TeleChat2 | `Tele-AI/TeleChat2-3B`, `Tele-AI/TeleChat2-7B`, `Tele-AI/TeleChat2-35B`, etc. | ✅︎ | ✅︎ |
 | `TeleFLMForCausalLM` | TeleFLM | `CofeAI/FLM-2-52B-Instruct-2407`, `CofeAI/Tele-FLM`, etc. | ✅︎ | ✅︎ |
@@ -490,7 +492,7 @@ th {
 Some models are supported only via the [Transformers modeling backend](#transformers). The purpose of the table below is to acknowledge models which we officially support in this way. The logs will say that the Transformers modeling backend is being used, and you will see no warning that this is fallback behaviour. This means that, if you have issues with any of the models listed below, please [make an issue](https://github.com/vllm-project/vllm/issues/new/choose) and we'll do our best to fix it!
 
 | Architecture | Models | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|-------------------|----------------------|---------------------------|
+| ------------ | ------ | ----------------- | -------------------- | ------------------------- |
 | `SmolLM3ForCausalLM` | SmolLM3 | `HuggingFaceTB/SmolLM3-3B` | ✅︎ | ✅︎ |
 
 !!! note
@@ -509,16 +511,16 @@ See [this page](./pooling_models.md) for more information on how to use pooling
 These models primarily support the [`LLM.embed`](./pooling_models.md#llmembed) API.
 
 | Architecture | Models | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|-------------------|----------------------|---------------------------|
+| ------------ | ------ | ----------------- | -------------------- | ------------------------- |
 | `BertModel`<sup>C</sup> | BERT-based | `BAAI/bge-base-en-v1.5`, `Snowflake/snowflake-arctic-embed-xs`, etc. | | |
 | `BertSpladeSparseEmbeddingModel` | SPLADE | `naver/splade-v3` | | |
 | `Gemma2Model`<sup>C</sup> | Gemma 2-based | `BAAI/bge-multilingual-gemma2`, etc. | ✅︎ | ✅︎ |
 | `Gemma3TextModel`<sup>C</sup> | Gemma 3-based | `google/embeddinggemma-300m`, etc. | ✅︎ | ✅︎ |
 | `GritLM` | GritLM | `parasail-ai/GritLM-7B-vllm`. | ✅︎ | ✅︎ |
-| `GteModel`<sup>C</sup> | Arctic-Embed-2.0-M | `Snowflake/snowflake-arctic-embed-m-v2.0`. |  |  |
-| `GteNewModel`<sup>C</sup> | mGTE-TRM (see note) | `Alibaba-NLP/gte-multilingual-base`, etc. |  |  |
-| `ModernBertModel`<sup>C</sup> | ModernBERT-based | `Alibaba-NLP/gte-modernbert-base`, etc. |  |  |
-| `NomicBertModel`<sup>C</sup> | Nomic BERT | `nomic-ai/nomic-embed-text-v1`, `nomic-ai/nomic-embed-text-v2-moe`, `Snowflake/snowflake-arctic-embed-m-long`, etc. |  |  |
+| `GteModel`<sup>C</sup> | Arctic-Embed-2.0-M | `Snowflake/snowflake-arctic-embed-m-v2.0`. | | |
+| `GteNewModel`<sup>C</sup> | mGTE-TRM (see note) | `Alibaba-NLP/gte-multilingual-base`, etc. | | |
+| `ModernBertModel`<sup>C</sup> | ModernBERT-based | `Alibaba-NLP/gte-modernbert-base`, etc. | | |
+| `NomicBertModel`<sup>C</sup> | Nomic BERT | `nomic-ai/nomic-embed-text-v1`, `nomic-ai/nomic-embed-text-v2-moe`, `Snowflake/snowflake-arctic-embed-m-long`, etc. | | |
 | `LlamaBidirectionalModel`<sup>C</sup> | Llama-based with bidirectional attention | `nvidia/llama-nemotron-embed-1b-v2`, etc. | ✅︎ | ✅︎ |
 | `LlamaModel`<sup>C</sup>, `LlamaForCausalLM`<sup>C</sup>, `MistralModel`<sup>C</sup>, etc. | Llama-based | `intfloat/e5-mistral-7b-instruct`, etc. | ✅︎ | ✅︎ |
 | `Qwen2Model`<sup>C</sup>, `Qwen2ForCausalLM`<sup>C</sup> | Qwen2-based | `ssmits/Qwen2-7B-Instruct-embed-base` (see note), `Alibaba-NLP/gte-Qwen2-7B-instruct` (see note), etc. | ✅︎ | ✅︎ |
@@ -553,7 +555,7 @@ of the whole prompt are extracted from the normalized hidden state corresponding
 These models primarily support the [`LLM.classify`](./pooling_models.md#llmclassify) API.
 
 | Architecture | Models | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|-------------------|----------------------|---------------------------|
+| ------------ | ------ | ----------------- | -------------------- | ------------------------- |
 | `JambaForSequenceClassification` | Jamba | `ai21labs/Jamba-tiny-reward-dev`, etc. | ✅︎ | ✅︎ |
 | `GPT2ForSequenceClassification` | GPT2 | `nie3e/sentiment-polish-gpt2-small` | | |
 | `*Model`<sup>C</sup>, `*ForCausalLM`<sup>C</sup>, etc. | Generative models | N/A | \* | \* |
@@ -570,7 +572,7 @@ Cross-encoder and reranker models are a subset of classification models that acc
 These models primarily support the [`LLM.score`](./pooling_models.md#llmscore) API.
 
 | Architecture | Models | Example HF Models | Score template (see note) | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|-------------------|---------------------------|-----------------------------|-----------------------------------------|
+| ------------ | ------ | ----------------- | ------------------------- | --------------------------- | --------------------------------------- |
 | `BertForSequenceClassification` | BERT-based | `cross-encoder/ms-marco-MiniLM-L-6-v2`, etc. | N/A | | |
 | `GemmaForSequenceClassification` | Gemma-based | `BAAI/bge-reranker-v2-gemma`(see note), etc. | [bge-reranker-v2-gemma.jinja](../../examples/pooling/score/template/bge-reranker-v2-gemma.jinja) | ✅︎ | ✅︎ |
 | `GteNewForSequenceClassification` | mGTE-TRM (see note) | `Alibaba-NLP/gte-multilingual-reranker-base`, etc. | N/A | | |
@@ -620,7 +622,7 @@ These models primarily support the [`LLM.score`](./pooling_models.md#llmscore) A
 These models primarily support the [`LLM.reward`](./pooling_models.md#llmreward) API.
 
 | Architecture | Models | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|-------------------|----------------------|---------------------------|
+| ------------ | ------ | ----------------- | -------------------- | ------------------------- |
 | `InternLM2ForRewardModel` | InternLM2-based | `internlm/internlm2-1_8b-reward`, `internlm/internlm2-7b-reward`, etc. | ✅︎ | ✅︎ |
 | `LlamaForCausalLM` | Llama-based | `peiyi9979/math-shepherd-mistral-7b-prm`, etc. | ✅︎ | ✅︎ |
 | `Qwen2ForRewardModel` | Qwen2-based | `Qwen/Qwen2.5-Math-RM-72B`, etc. | ✅︎ | ✅︎ |
@@ -635,9 +637,9 @@ These models primarily support the [`LLM.reward`](./pooling_models.md#llmreward)
 These models primarily support the [`LLM.encode`](./pooling_models.md#llmencode) API.
 
 | Architecture | Models | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|-------------------|-----------------------------|-----------------------------------------|
-| `BertForTokenClassification` | bert-based | `boltuix/NeuroBERT-NER` (see note), etc. |  |  |
-| `ModernBertForTokenClassification` | ModernBERT-based | `disham993/electrical-ner-ModernBERT-base` |  |  |
+| ------------ | ------ | ----------------- | --------------------------- | --------------------------------------- |
+| `BertForTokenClassification` | bert-based | `boltuix/NeuroBERT-NER` (see note), etc. | | |
+| `ModernBertForTokenClassification` | ModernBERT-based | `disham993/electrical-ner-ModernBERT-base` | | |
 
 !!! note
     Named Entity Recognition (NER) usage, please refer to [examples/pooling/token_classify/ner_offline.py](../../examples/pooling/token_classify/ner_offline.py), [examples/pooling/token_classify/ner_online.py](../../examples/pooling/token_classify/ner_online.py).
@@ -676,7 +678,7 @@ See [this page](generative_models.md) for more information on how to use generat
 These models primarily accept the [`LLM.generate`](./generative_models.md#llmgenerate) API. Chat/Instruct models additionally support the [`LLM.chat`](./generative_models.md#llmchat) API.
 
 | Architecture | Models | Inputs | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|--------|-------------------|----------------------|---------------------------|
+| ------------ | ------ | ------ | ----------------- | -------------------- | ------------------------- |
 | `AriaForConditionalGeneration` | Aria | T + I<sup>+</sup> | `rhymes-ai/Aria` | | |
 | `AudioFlamingo3ForConditionalGeneration` | AudioFlamingo3 | T + A | `nvidia/audio-flamingo-3-hf`, `nvidia/music-flamingo-2601-hf` | ✅︎ | ✅︎ |
 | `AyaVisionForConditionalGeneration` | Aya Vision | T + I<sup>+</sup> | `CohereLabs/aya-vision-8b`, `CohereLabs/aya-vision-32b`, etc. | | ✅︎ |
@@ -696,9 +698,10 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `GLM4VForCausalLM`<sup>^</sup> | GLM-4V | T + I | `zai-org/glm-4v-9b`, `zai-org/cogagent-9b-20241220`, etc. | ✅︎ | ✅︎ |
 | `Glm4vForConditionalGeneration` | GLM-4.1V-Thinking | T + I<sup>E+</sup> + V<sup>E+</sup> | `zai-org/GLM-4.1V-9B-Thinking`, etc. | ✅︎ | ✅︎ |
 | `Glm4vMoeForConditionalGeneration` | GLM-4.5V | T + I<sup>E+</sup> + V<sup>E+</sup> | `zai-org/GLM-4.5V`, etc. | ✅︎ | ✅︎ |
-| `GlmOcrForConditionalGeneration` | GLM-OCR | T + I<sup>E+</sup>  | `zai-org/GLM-OCR`, etc. | ✅︎ | ✅︎ |
+| `GlmOcrForConditionalGeneration` | GLM-OCR | T + I<sup>E+</sup> | `zai-org/GLM-OCR`, etc. | ✅︎ | ✅︎ |
 | `GraniteSpeechForConditionalGeneration` | Granite Speech | T + A | `ibm-granite/granite-speech-3.3-8b` | ✅︎ | ✅︎ |
 | `HCXVisionForCausalLM` | HyperCLOVAX-SEED-Vision-Instruct-3B | T + I<sup>+</sup> + V<sup>+</sup> | `naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B` | | |
+| `HCXVisionV2ForCausalLM` | HyperCLOVAX-SEED-Think-32B | T + I<sup>+</sup> + V<sup>+</sup> | `naver-hyperclovax/HyperCLOVAX-SEED-Think-32B` | | |
 | `H2OVLChatModel` | H2OVL | T + I<sup>E+</sup> | `h2oai/h2ovl-mississippi-800m`, `h2oai/h2ovl-mississippi-2b`, etc. | | ✅︎ |
 | `HunYuanVLForConditionalGeneration` | HunyuanOCR | T + I<sup>E+</sup> | `tencent/HunyuanOCR`, etc. | ✅︎ | ✅︎ |
 | `Idefics3ForConditionalGeneration` | Idefics3 | T + I | `HuggingFaceM4/Idefics3-8B-Llama3`, etc. | ✅︎ | |
@@ -710,9 +713,10 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `KananaVForConditionalGeneration` | Kanana-V | T + I<sup>+</sup> | `kakaocorp/kanana-1.5-v-3b-instruct`, etc. | | ✅︎ |
 | `KeyeForConditionalGeneration` | Keye-VL-8B-Preview | T + I<sup>E+</sup> + V<sup>E+</sup> | `Kwai-Keye/Keye-VL-8B-Preview` | ✅︎ | ✅︎ |
 | `KeyeVL1_5ForConditionalGeneration` | Keye-VL-1_5-8B | T + I<sup>E+</sup> + V<sup>E+</sup> | `Kwai-Keye/Keye-VL-1_5-8B` | ✅︎ | ✅︎ |
-| `KimiVLForConditionalGeneration` | Kimi-VL-A3B-Instruct, Kimi-VL-A3B-Thinking | T + I<sup>+</sup> | `moonshotai/Kimi-VL-A3B-Instruct`, `moonshotai/Kimi-VL-A3B-Thinking` | | ✅︎ |
+| `KimiAudioForConditionalGeneration` | Kimi-Audio | T + A<sup>+</sup> | `moonshotai/Kimi-Audio-7B-Instruct` | | ✅︎ |
 | `KimiK25ForConditionalGeneration` | Kimi-K2.5 | T + I<sup>+</sup> | `moonshotai/Kimi-K2.5` | | ✅︎ |
-| `LightOnOCRForConditionalGeneration`  | LightOnOCR-1B  | T + I<sup>+</sup> | `lightonai/LightOnOCR-1B`, etc | ✅︎ | ✅︎ |
+| `KimiVLForConditionalGeneration` | Kimi-VL-A3B-Instruct, Kimi-VL-A3B-Thinking | T + I<sup>+</sup> | `moonshotai/Kimi-VL-A3B-Instruct`, `moonshotai/Kimi-VL-A3B-Thinking` | | ✅︎ |
+| `LightOnOCRForConditionalGeneration` | LightOnOCR-1B | T + I<sup>+</sup> | `lightonai/LightOnOCR-1B`, etc | ✅︎ | ✅︎ |
 | `Lfm2VlForConditionalGeneration` | LFM2-VL | T + I<sup>+</sup> | `LiquidAI/LFM2-VL-450M`, `LiquidAI/LFM2-VL-3B`, `LiquidAI/LFM2-VL-8B-A1B`, etc. | ✅︎ | ✅︎ |
 | `Llama4ForConditionalGeneration` | Llama 4 | T + I<sup>+</sup> | `meta-llama/Llama-4-Scout-17B-16E-Instruct`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct`, etc. | ✅︎ | ✅︎ |
 | `Llama_Nemotron_Nano_VL` | Llama Nemotron Nano VL | T + I<sup>E+</sup> | `nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1` | ✅︎ | ✅︎ |
@@ -729,7 +733,7 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `Molmo2ForConditionalGeneration` | Molmo2 | T + I<sup>+</sup> / V | `allenai/Molmo2-4B`, `allenai/Molmo2-8B`, `allenai/Molmo2-O-7B` | ✅︎ | ✅︎ |
 | `NVLM_D_Model` | NVLM-D 1.0 | T + I<sup>+</sup> | `nvidia/NVLM-D-72B`, etc. | | ✅︎ |
 | `OpenCUAForConditionalGeneration` | OpenCUA-7B | T + I<sup>E+</sup> | `xlangai/OpenCUA-7B` | ✅︎ | ✅︎ |
-| `OpenPanguVLForConditionalGeneration` | openpangu-VL | T + I<sup>E+</sup> + V<sup>E+</sup> |`FreedomIntelligence/openPangu-VL-7B` | ✅︎ | ✅︎ |
+| `OpenPanguVLForConditionalGeneration` | openpangu-VL | T + I<sup>E+</sup> + V<sup>E+</sup> | `FreedomIntelligence/openPangu-VL-7B` | ✅︎ | ✅︎ |
 | `Ovis` | Ovis2, Ovis1.6 | T + I<sup>+</sup> | `AIDC-AI/Ovis2-1B`, `AIDC-AI/Ovis1.6-Llama3.2-3B`, etc. | | ✅︎ |
 | `Ovis2_5` | Ovis2.5 | T + I<sup>+</sup> + V | `AIDC-AI/Ovis2.5-9B`, etc. | | |
 | `Ovis2_6ForCausalLM` | Ovis2.6 | T + I<sup>+</sup> + V | `AIDC-AI/Ovis2.6-2B`, etc. | | |
@@ -762,7 +766,7 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 Some models are supported only via the [Transformers modeling backend](#transformers). The purpose of the table below is to acknowledge models which we officially support in this way. The logs will say that the Transformers modeling backend is being used, and you will see no warning that this is fallback behaviour. This means that, if you have issues with any of the models listed below, please [make an issue](https://github.com/vllm-project/vllm/issues/new/choose) and we'll do our best to fix it!
 
 | Architecture | Models | Inputs | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|--------|-------------------|-----------------------------|-----------------------------------------|
+| ------------ | ------ | ------ | ----------------- | --------------------------- | --------------------------------------- |
 | `Emu3ForConditionalGeneration` | Emu3 | T + I | `BAAI/Emu3-Chat-hf` | ✅︎ | ✅︎ |
 
 <sup>^</sup> You need to set the architecture name via `--hf-overrides` to match the one in vLLM.</br>
@@ -793,7 +797,7 @@ Some models are supported only via the [Transformers modeling backend](#transfor
 Speech2Text models trained specifically for Automatic Speech Recognition.
 
 | Architecture | Models | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|-------------------|----------------------|---------------------------|
+| ------------ | ------ | ----------------- | -------------------- | ------------------------- |
 | `FireRedASR2ForConditionalGeneration` | FireRedASR2 | `allendou/FireRedASR2-LLM-vllm`, etc. | | |
 | `FunASRForConditionalGeneration` | FunASR | `allendou/Fun-ASR-Nano-2512-vllm`, etc. | | |
 | `Gemma3nForConditionalGeneration` | Gemma3n | `google/gemma-3n-E2B-it`, `google/gemma-3n-E4B-it`, etc. | | |
@@ -821,7 +825,7 @@ These models primarily support the [`LLM.embed`](./pooling_models.md#llmembed) A
 The following table lists those that are tested in vLLM.
 
 | Architecture | Models | Inputs | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|--------|-------------------|----------------------|---------------------------|
+| ------------ | ------ | ------ | ----------------- | -------------------- | ------------------------- |
 | `CLIPModel` | CLIP | T / I | `openai/clip-vit-base-patch32`, `openai/clip-vit-large-patch14`, etc. | | |
 | `ColModernVBertForRetrieval` | ColModernVBERT | T / I | `ModernVBERT/colmodernvbert-merged` | | |
 | `LlamaNemotronVLModel` | Llama Nemotron Embedding + SigLIP | T + I | `nvidia/llama-nemotron-embed-vl-1b-v2` | | |
@@ -842,7 +846,7 @@ Cross-encoder and reranker models are a subset of classification models that acc
 These models primarily support the [`LLM.score`](./pooling_models.md#llmscore) API.
 
 | Architecture | Models | Inputs | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
-|--------------|--------|--------|-------------------|----------------------|---------------------------|
+| ------------ | ------ | ------ | ----------------- | -------------------- | ------------------------- |
 | `JinaVLForSequenceClassification` | JinaVL-based | T + I<sup>E+</sup> | `jinaai/jina-reranker-m0`, etc. | ✅︎ | ✅︎ |
 | `LlamaNemotronVLForSequenceClassification` | Llama Nemotron Reranker + SigLIP | T + I<sup>E+</sup> | `nvidia/llama-nemotron-rerank-vl-1b-v2` | | |
 | `Qwen3VLForSequenceClassification` | Qwen3-VL-Reranker | T + I<sup>E+</sup> + V<sup>E+</sup> | `Qwen/Qwen3-VL-Reranker-2B`(see note), etc. | ✅︎ | ✅︎ |

docs/serving/expert_parallel_deployment.md

@@ -17,7 +17,7 @@ Before using EP, you need to install the necessary dependencies. We are actively
 vLLM provides multiple communication backends for EP. Use `--all2all-backend` to select one:
 
 | Backend | Use Case | Features | Best For |
-|---------|----------|----------|----------|
+| ------- | -------- | -------- | -------- |
 | `allgather_reducescatter` | Default backend | Standard all2all using allgather/reducescatter primitives | General purpose, works with any EP+DP configuration |
 | `deepep_high_throughput` | Multi-node prefill | Grouped GEMM with continuous layout, optimized for prefill | Prefill-dominated workloads, high-throughput scenarios |
 | `deepep_low_latency` | Multi-node decode | CUDA graph support, masked layout, optimized for decode | Decode-dominated workloads, low-latency scenarios |
@@ -48,7 +48,7 @@ Where:
 When EP is enabled, different layers in MoE models behave differently:
 
 | Layer Type | Behavior | Parallelism Used |
-|------------|----------|------------------|
+| ---------- | -------- | ---------------- |
 | **Expert (MoE) Layers** | Sharded across all EP ranks | Expert Parallel (EP) of size `TP × DP` |
 | **Attention Layers** | Behavior depends on TP size | See below |
 
@@ -146,9 +146,9 @@ When enabled, vLLM collects load statistics with every forward pass and periodic
 Configure EPLB with the `--eplb-config` argument, which accepts a JSON string. The available keys and their descriptions are:
 
 | Parameter | Description | Default |
-|-----------|-------------|---------|
-| `window_size`| Number of engine steps to track for rebalancing decisions | 1000 |
-| `step_interval`| Frequency of rebalancing (every N engine steps) | 3000 |
+| --------- | ----------- | ------- |
+| `window_size` | Number of engine steps to track for rebalancing decisions | 1000 |
+| `step_interval` | Frequency of rebalancing (every N engine steps) | 3000 |
 | `log_balancedness` | Log balancedness metrics (avg tokens per expert ÷ max tokens per expert) | `false` |
 | `num_redundant_experts` | Additional global experts per EP rank beyond equal distribution | `0` |
 | `use_async` | Use non-blocking EPLB for reduced latency overhead | `false` |

docs/serving/integrations/claude_code.md

@@ -60,6 +60,9 @@ The environment variables:
 !!! tip
     You can add these environment variables to your shell profile (e.g., `.bashrc`, `.zshrc`), Claude Code configuration file (`~/.claude/settings.json`), or create a wrapper script for convenience.
 
+!!! warning
+    Claude Code recently started injecting a per-request hash in the system prompt, which can defeat [prefix caching](../../design/prefix_caching.md) because the prompt changes on every request, causing greatly reduced performance. This is addressed automatically in vLLM versions > 0.17.1 but for older versions `"CLAUDE_CODE_ATTRIBUTION_HEADER": "0"` should be added to the `"env"` section of `~/.claude/settings.json` (see this [blog post](https://unsloth.ai/docs/basics/claude-code#fixing-90-slower-inference-in-claude-code) from Unsloth).
+
 ## Testing the Setup
 
 Once Claude Code launches, try a simple prompt to verify the connection:

docs/serving/integrations/llamaindex.md

@@ -17,7 +17,7 @@ llm = Vllm(
     model="microsoft/Orca-2-7b",
     tensor_parallel_size=4,
     max_new_tokens=100,
-    vllm_kwargs={"swap_space": 1, "gpu_memory_utilization": 0.5},
+    vllm_kwargs={"gpu_memory_utilization": 0.5},
 )
 ```
 

docs/serving/openai_compatible_server.md

@@ -190,7 +190,7 @@ vllm serve NousResearch/Meta-Llama-3-8B-Instruct --enable-offline-docs
 Our Completions API is compatible with [OpenAI's Completions API](https://platform.openai.com/docs/api-reference/completions);
 you can use the [official OpenAI Python client](https://github.com/openai/openai-python) to interact with it.
 
-Code example: [examples/online_serving/openai_completion_client.py](../../examples/online_serving/openai_completion_client.py)
+Code example: [examples/basic/online_serving/openai_completion_client.py](../../examples/basic/online_serving/openai_completion_client.py)
 
 #### Extra parameters
 
@@ -221,7 +221,7 @@ see our [Multimodal Inputs](../features/multimodal_inputs.md) guide for more inf
 
 - *Note: `image_url.detail` parameter is not supported.*
 
-Code example: [examples/online_serving/openai_chat_completion_client.py](../../examples/online_serving/openai_chat_completion_client.py)
+Code example: [examples/basic/online_serving/openai_chat_completion_client.py](../../examples/basic/online_serving/openai_chat_completion_client.py)
 
 #### Extra parameters
 
@@ -439,6 +439,8 @@ you can use the [official OpenAI Python client](https://github.com/openai/openai
 
 Code example: [examples/online_serving/openai_transcription_client.py](../../examples/online_serving/openai_transcription_client.py)
 
+NOTE: beam search is currently supported in the transcriptions endpoint for encoder-decoder multimodal models, e.g., whisper, but highly inefficient as work for handling the encoder/decoder cache is actively ongoing. This is an active point of ongoing optimization and will be handled properly in the very near future.
+
 #### API Enforced Limits
 
 Set the maximum audio file size (in MB) that VLLM will accept, via the
@@ -596,7 +598,7 @@ Audio must be sent as base64-encoded PCM16 audio at 16kHz sample rate, mono chan
 #### Client → Server Events
 
 | Event | Description |
-|-------|-------------|
+| ----- | ----------- |
 | `input_audio_buffer.append` | Send base64-encoded audio chunk: `{"type": "input_audio_buffer.append", "audio": "<base64>"}` |
 | `input_audio_buffer.commit` | Trigger transcription processing or end: `{"type": "input_audio_buffer.commit", "final": bool}` |
 | `session.update` | Configure session: `{"type": "session.update", "model": "model-name"}` |
@@ -604,7 +606,7 @@ Audio must be sent as base64-encoded PCM16 audio at 16kHz sample rate, mono chan
 #### Server → Client Events
 
 | Event | Description |
-|-------|-------------|
+| ----- | ----------- |
 | `session.created` | Connection established with session ID and timestamp |
 | `transcription.delta` | Incremental transcription text: `{"type": "transcription.delta", "delta": "text"}` |
 | `transcription.done` | Final transcription with usage stats |

docs/serving/parallelism_scaling.md

@@ -68,6 +68,12 @@ vLLM uses Ray to manage the distributed execution of tasks across multiple nodes
 
 Ray also offers high-level APIs for large-scale [offline batch inference](https://docs.ray.io/en/latest/data/working-with-llms.html) and [online serving](https://docs.ray.io/en/latest/serve/llm) that can leverage vLLM as the engine. These APIs add production-grade fault tolerance, scaling, and distributed observability to vLLM workloads.
 
+Ray is an optional dependency. Install it explicitly before using Ray-based execution, for example:
+
+```bash
+pip install "ray[cgraph]"
+```
+
 For details, see the [Ray documentation](https://docs.ray.io/en/latest/index.html).
 
 ### Ray cluster setup with containers

docs/usage/security.md

@@ -41,20 +41,20 @@ Key points from the PyTorch security guide:
 - Messages are sent unencrypted
 - Connections are accepted from anywhere without checks
 
-### Security Recommendations
+## Security Recommendations
 
-#### 1. **Network Isolation:**
+### 1. **Network Isolation:**
 
 - Deploy vLLM nodes on a dedicated, isolated network
 - Use network segmentation to prevent unauthorized access
 - Implement appropriate firewall rules
 
-#### 2. **Configuration Best Practices:**
+### 2. **Configuration Best Practices:**
 
 - Always set `VLLM_HOST_IP` to a specific IP address rather than using defaults
 - Configure firewalls to only allow necessary ports between nodes
 
-#### 3. **Access Control:**
+### 3. **Access Control:**
 
 - Restrict physical and network access to the deployment environment
 - Implement proper authentication and authorization for management interfaces
@@ -66,6 +66,18 @@ Restrict domains that vLLM can access for media URLs by setting
 `--allowed-media-domains` to prevent Server-Side Request Forgery (SSRF) attacks.
 (e.g. `--allowed-media-domains upload.wikimedia.org github.com www.bogotobogo.com`)
 
+Without domain restrictions, a malicious user could supply URLs that:
+
+- **Target internal services**: Access internal network endpoints, cloud metadata
+  services (e.g. `169.254.169.254`), or other services not intended to be
+  publicly reachable (SSRF).
+- **Consume excessive resources**: Point to extremely large files or slow
+  endpoints, causing the server to download unbounded amounts of data and
+  exhausting memory, disk, or network bandwidth.
+
+By explicitly allowlisting only the domains you expect media to come from, you
+significantly reduce the attack surface for these types of abuse.
+
 Also, consider setting `VLLM_MEDIA_URL_ALLOW_REDIRECTS=0` to prevent HTTP
 redirects from being followed to bypass domain restrictions.
 

docs/usage/v1_guide.md

@@ -83,13 +83,13 @@ based on assigned priority, with FCFS as a tie-breaker), configurable via the
 
 ### Hardware
 
-| Hardware         | Status                                        |
-|------------------|-----------------------------------------------|
-| **NVIDIA**       | <nobr>🟢</nobr>                               |
-| **AMD**          | <nobr>🟢</nobr>                               |
-| **INTEL GPU**    | <nobr>🟢</nobr>                               |
-| **TPU**          | <nobr>🟢</nobr>                               |
-| **CPU**          | <nobr>🟢</nobr>                               |
+| Hardware      | Status          |
+| --------------| --------------- |
+| **NVIDIA**    | <nobr>🟢</nobr> |
+| **AMD**       | <nobr>🟢</nobr> |
+| **INTEL GPU** | <nobr>🟢</nobr> |
+| **TPU**       | <nobr>🟢</nobr> |
+| **CPU**       | <nobr>🟢</nobr> |
 
 !!! note
 
@@ -104,13 +104,13 @@ based on assigned priority, with FCFS as a tie-breaker), configurable via the
 
 ### Models
 
-| Model Type                  | Status                                                                  |
-|-----------------------------|-------------------------------------------------------------------------|
-| **Decoder-only Models**     | <nobr>🟢</nobr>                                                         |
-| **Encoder-Decoder Models**  | <nobr>🟢 (Whisper), 🔴 (Others) </nobr>                                |
-| **Pooling Models**          | <nobr>🟢</nobr>                                                         |
-| **Mamba Models**            | <nobr>🟢</nobr>                                                         |
-| **Multimodal Models**       | <nobr>🟢</nobr>                                                         |
+| Model Type                 | Status                                  |
+| -------------------------- | --------------------------------------- |
+| **Decoder-only Models**    | <nobr>🟢</nobr>                         |
+| **Encoder-Decoder Models** | <nobr>🟢 (Whisper), 🔴 (Others) </nobr> |
+| **Pooling Models**         | <nobr>🟢</nobr>                         |
+| **Mamba Models**           | <nobr>🟢</nobr>                         |
+| **Multimodal Models**      | <nobr>🟢</nobr>                         |
 
 See below for the status of models that are not yet supported or have more features planned in V1.
 
@@ -145,7 +145,7 @@ following a similar pattern by implementing support through the [plugin system](
 ### Features
 
 | Feature                                     | Status                                                                            |
-|---------------------------------------------|-----------------------------------------------------------------------------------|
+| ------------------------------------------- | --------------------------------------------------------------------------------- |
 | **Prefix Caching**                          | <nobr>🟢 Functional</nobr>                                                        |
 | **Chunked Prefill**                         | <nobr>🟢 Functional</nobr>                                                        |
 | **LoRA**                                    | <nobr>🟢 Functional</nobr>                                                        |

examples/basic/offline_inference/README.md

@@ -1,4 +1,4 @@
-# Basic
+# Offline Inference
 
 The `LLM` class provides the primary Python interface for doing offline inference, which is interacting with a model without using a separate model inference server.
 
@@ -7,31 +7,31 @@ The `LLM` class provides the primary Python interface for doing offline inferenc
 The first script in this example shows the most basic usage of vLLM. If you are new to Python and vLLM, you should start here.
 
 ```bash
-python examples/offline_inference/basic/basic.py
+python examples/basic/offline_inference/basic.py
 ```
 
 The rest of the scripts include an [argument parser](https://docs.python.org/3/library/argparse.html), which you can use to pass any arguments that are compatible with [`LLM`](https://docs.vllm.ai/en/latest/api/offline_inference/llm.html). Try running the script with `--help` for a list of all available arguments.
 
 ```bash
-python examples/offline_inference/basic/classify.py
+python examples/basic/offline_inference/classify.py
 ```
 
 ```bash
-python examples/offline_inference/basic/embed.py
+python examples/basic/offline_inference/embed.py
 ```
 
 ```bash
-python examples/offline_inference/basic/score.py
+python examples/basic/offline_inference/score.py
 ```
 
 The chat and generate scripts also accept the [sampling parameters](https://docs.vllm.ai/en/latest/api/inference_params.html#sampling-parameters): `max_tokens`, `temperature`, `top_p` and `top_k`.
 
 ```bash
-python examples/offline_inference/basic/chat.py
+python examples/basic/offline_inference/chat.py
 ```
 
 ```bash
-python examples/offline_inference/basic/generate.py
+python examples/basic/offline_inference/generate.py
 ```
 
 ## Features