[Misc] fix qwen3.5 config (#34604)

Signed-off-by: Seiji Eicher <seiji@anyscale.com>

.buildkite/hardware_tests/amd.yaml

@@ -1,6 +1,7 @@
-group: Hardware
+group: Hardware - AMD Build 
 steps:
   - label: "AMD: :docker: build image"
+    key: image-build-amd
     depends_on: []
     device: amd_cpu
     no_plugin: true

.buildkite/hardware_tests/arm.yaml

@@ -1,8 +0,0 @@
-group: Hardware
-steps:
-  - label: "Arm CPU Test"
-    soft_fail: true
-    device: arm_cpu
-    no_plugin: true
-    commands: 
-    - bash .buildkite/scripts/hardware_ci/run-cpu-test-arm.sh

.buildkite/hardware_tests/cpu.yaml

@@ -0,0 +1,100 @@
+group: CPU
+depends_on: []
+steps:
+- label: CPU-Kernel Tests
+  depends_on: []
+  soft_fail: true
+  device: intel_cpu
+  no_plugin: true
+  source_file_dependencies:
+  - csrc/cpu/
+  - cmake/cpu_extension.cmake
+  - CMakeLists.txt
+  - vllm/_custom_ops.py
+  - tests/kernels/attention/test_cpu_attn.py
+  - tests/kernels/moe/test_cpu_fused_moe.py
+  - tests/kernels/test_onednn.py
+  commands:
+    - |
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 20m "
+      pytest -x -v -s tests/kernels/attention/test_cpu_attn.py
+      pytest -x -v -s tests/kernels/moe/test_cpu_fused_moe.py
+      pytest -x -v -s tests/kernels/test_onednn.py"
+
+- label: CPU-Language Generation and Pooling Model Tests
+  depends_on: []
+  soft_fail: true
+  device: intel_cpu
+  no_plugin: true
+  source_file_dependencies:
+  - csrc/cpu/
+  - vllm/
+  - tests/models/language/generation/
+  - tests/models/language/pooling/
+  commands:
+    - |
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 30m "
+      pytest -x -v -s tests/models/language/generation -m cpu_model
+      pytest -x -v -s tests/models/language/pooling -m cpu_model"
+
+- label: CPU-Quantization Model Tests
+  depends_on: []
+  soft_fail: true
+  device: intel_cpu
+  no_plugin: true
+  source_file_dependencies:
+  - csrc/cpu/
+  - vllm/model_executor/layers/quantization/cpu_wna16.py
+  - vllm/model_executor/layers/quantization/gptq_marlin.py
+  - vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_int8.py
+  - vllm/model_executor/layers/quantization/kernels/scaled_mm/cpu.py
+  - vllm/model_executor/layers/quantization/kernels/mixed_precision/cpu.py
+  - tests/quantization/test_compressed_tensors.py
+  - tests/quantization/test_cpu_wna16.py
+  commands:
+    - |
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 20m "
+      pytest -x -v -s tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs
+      pytest -x -v -s tests/quantization/test_cpu_wna16.py"
+      
+- label: CPU-Distributed Tests
+  depends_on: []
+  soft_fail: true
+  device: intel_cpu
+  no_plugin: true
+  source_file_dependencies:
+  - csrc/cpu/shm.cpp
+  - vllm/v1/worker/cpu_worker.py
+  - vllm/v1/worker/gpu_worker.py
+  - vllm/v1/worker/cpu_model_runner.py
+  - vllm/v1/worker/gpu_model_runner.py
+  - vllm/platforms/cpu.py
+  - vllm/distributed/parallel_state.py
+  - vllm/distributed/device_communicators/cpu_communicator.py
+  commands:
+    - |
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 10m "
+      bash .buildkite/scripts/hardware_ci/run-cpu-distributed-smoke-test.sh"
+
+- label: CPU-Multi-Modal Model Tests %N
+  depends_on: []
+  soft_fail: true
+  device: intel_cpu
+  no_plugin: true
+  source_file_dependencies:
+  # - vllm/
+  - vllm/model_executor/layers/rotary_embedding
+  - tests/models/multimodal/generation/
+  commands:
+    - |
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
+      pytest -x -v -s tests/models/multimodal/generation --ignore=tests/models/multimodal/generation/test_pixtral.py -m cpu_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
+  parallelism: 2
+
+- label: "Arm CPU Test"
+  depends_on: []
+  soft_fail: true
+  device: arm_cpu
+  no_plugin: true
+  commands: 
+  - bash .buildkite/scripts/hardware_ci/run-cpu-test-arm.sh

.buildkite/hardware_tests/intel.yaml

@@ -1,13 +1,6 @@
 group: Hardware
 depends_on: ~
 steps:
-  - label: "Intel CPU Test"
-    soft_fail: true
-    device: intel_cpu
-    no_plugin: true
-    commands: 
-    - bash .buildkite/scripts/hardware_ci/run-cpu-test.sh
-  
   - label: "Intel HPU Test"
     soft_fail: true
     device: intel_hpu

.buildkite/image_build/image_build.yaml

@@ -3,6 +3,7 @@ steps:
   - label: ":docker: Build image"
     key: image-build
     depends_on: []
+    timeout_in_minutes: 600
     commands:
     - if [[ "$BUILDKITE_BRANCH" != "main" ]]; then .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $VLLM_USE_PRECOMPILED $VLLM_MERGE_BASE_COMMIT $IMAGE_TAG; fi
     - if [[ "$BUILDKITE_BRANCH" == "main" ]]; then .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $VLLM_USE_PRECOMPILED $VLLM_MERGE_BASE_COMMIT $IMAGE_TAG $IMAGE_TAG_LATEST; fi
@@ -41,7 +42,7 @@ steps:
           limit: 2
         - exit_status: -10  # Agent was lost
           limit: 2
-  
+
   - label: ":docker: Build CPU arm64 image"
     key: cpu-arm64-image-build
     depends_on: []

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

@@ -9,8 +9,10 @@ import json
 import os
 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
@@ -275,6 +277,131 @@ def _apply_two_decimals(
     return styler.format({c: "{:.2f}" for c in num_cols}, na_rep="")
 
 
+# -----------------------------
+# Export helpers (Excel + CSV)
+# -----------------------------
+def _sanitize_sheet_name(name: str) -> str:
+    """
+    Excel sheet constraints:
+      - max 31 chars
+      - cannot contain: : \ / ? * [ ]
+      - cannot be empty
+    """
+    name = "sheet" if name is None else str(name)
+    name = re.sub(r"[:\\/?*\[\]]", "_", name)
+    name = name.strip().strip("'")
+    name = re.sub(r"\s+", " ", name)
+    if not name:
+        name = "sheet"
+    return name[:31]
+
+
+def _group_to_sheet_base(group_cols: list[str], gkey_tuple) -> str:
+    d = dict(zip(group_cols, gkey_tuple))
+    model = d.get("Model", "model")
+    model_short = str(model).split("/")[-1]
+    ilen = d.get("Input Len", "")
+    olen = d.get("Output Len", "")
+    lens = f"_{ilen}x{olen}" if ilen != "" and olen != "" else ""
+    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
+    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
+
+
+def _safe_filename(s: str) -> str:
+    s = re.sub(r"[^\w\-.]+", "_", str(s).strip())
+    return s[:180] if len(s) > 180 else s
+
+
+# -----------------------------
+# vLLM environment export helper
+# -----------------------------
+def _parse_vllm_env_txt(env_path: Path) -> pd.DataFrame:
+    """Parse vllm_env.txt into a flat table (Section, Key, Value).
+
+    Supports:
+      - section headers as standalone lines (no ':' or '=')
+      - key-value lines like 'OS: Ubuntu ...'
+      - env var lines like 'HF_HOME=/data/hf'
+    """
+    lines = env_path.read_text(encoding="utf-8", errors="replace").splitlines()
+    section = "General"
+    rows: list[dict] = []
+
+    def set_section(s: str):
+        nonlocal section
+        s = (s or "").strip()
+        if s:
+            section = s
+
+    for raw in lines:
+        stripped = raw.strip()
+        if not stripped:
+            continue
+        # divider lines like =====
+        if set(stripped) <= {"="}:
+            continue
+
+        # section header heuristic: short standalone line
+        if ":" not in stripped and "=" not in stripped and len(stripped) <= 64:
+            if stripped.lower().startswith("collecting environment information"):
+                continue
+            set_section(stripped)
+            continue
+
+        # env var style: KEY=VALUE (and not a URL with :)
+        if "=" in stripped and ":" not in stripped:
+            k, v = stripped.split("=", 1)
+            k = k.strip()
+            v = v.strip()
+            if k:
+                rows.append({"Section": section, "Key": k, "Value": v})
+            continue
+
+        # key: value
+        if ":" in stripped:
+            k, v = stripped.split(":", 1)
+            k = k.strip()
+            v = v.strip()
+            if k:
+                rows.append({"Section": section, "Key": k, "Value": v})
+            continue
+
+    return pd.DataFrame(rows, columns=["Section", "Key", "Value"])
+
+
+def _load_env_df_for_inputs(args, files: list[str]) -> pd.DataFrame | None:
+    """Load vllm_env.txt next to the *original* input JSON file.
+
+    Note: when only one -f is provided, the script may split JSON into ./splits/...,
+    but vllm_env.txt typically lives next to the original benchmark_results.json.
+    """
+    base_dir: Path | None = None
+    if getattr(args, "file", None):
+        base_dir = Path(args.file[0]).resolve().parent
+    elif files:
+        base_dir = Path(files[0]).resolve().parent
+    if base_dir is None:
+        return None
+
+    env_path = base_dir / "vllm_env.txt"
+    if not env_path.exists():
+        return None
+    df = _parse_vllm_env_txt(env_path)
+    return df
+
+
 # -----------------------------
 # Valid max concurrency summary helpers
 # -----------------------------
@@ -428,7 +555,6 @@ def build_valid_max_concurrency_summary_html(
 
     summary_df = pd.DataFrame(rows)
 
-    # --- Coerce numeric columns so Styler doesn't miss them due to object dtype ---
     for c in summary_df.columns:
         if c == "Configuration":
             continue
@@ -436,12 +562,10 @@ def build_valid_max_concurrency_summary_html(
 
     both_col = f"Max {conc_col} (Both)"
 
-    # --- Strict 2-decimal formatting for ALL non-Configuration columns ---
     formatters = {}
     for c in summary_df.columns:
         if c == "Configuration":
             continue
-        # default argument binds per-column formatter correctly
         formatters[c] = lambda v: "" if pd.isna(v) else f"{float(v):.2f}"
 
     styler = summary_df.style.format(formatters)
@@ -460,6 +584,95 @@ def build_valid_max_concurrency_summary_html(
     return title + styler.to_html(table_attributes='border="1" class="dataframe"')
 
 
+def build_valid_max_concurrency_summary_df(
+    tput_group_df: pd.DataFrame | None,
+    ttft_group_df: pd.DataFrame | None,
+    tpot_group_df: pd.DataFrame | None,
+    conc_col: str,
+    args,
+) -> pd.DataFrame | None:
+    if ttft_group_df is None and tpot_group_df is None:
+        return None
+
+    ttft_cols = (
+        _config_value_columns(ttft_group_df, conc_col)
+        if ttft_group_df is not None
+        else []
+    )
+    tpot_cols = (
+        _config_value_columns(tpot_group_df, conc_col)
+        if tpot_group_df is not None
+        else []
+    )
+    tput_cols = (
+        _config_value_columns(tput_group_df, conc_col)
+        if tput_group_df is not None
+        else []
+    )
+
+    if ttft_group_df is not None and tpot_group_df is not None:
+        cfg_cols = [c for c in ttft_cols if c in tpot_cols]
+        if tput_group_df is not None:
+            cfg_cols = [c for c in cfg_cols if c in tput_cols] or cfg_cols
+    else:
+        cfg_cols = ttft_cols or tpot_cols
+
+    if not cfg_cols:
+        cfg_cols = sorted(set(ttft_cols) | set(tpot_cols) | set(tput_cols), key=str)
+
+    rows = []
+    for cfg in cfg_cols:
+        ttft_max = (
+            _max_concurrency_ok(ttft_group_df, conc_col, cfg, args.ttft_max_ms)
+            if ttft_group_df is not None
+            else pd.NA
+        )
+        tpot_max = (
+            _max_concurrency_ok(tpot_group_df, conc_col, cfg, args.tpot_max_ms)
+            if tpot_group_df is not None
+            else pd.NA
+        )
+        both = (
+            pd.NA
+            if (pd.isna(ttft_max) or pd.isna(tpot_max))
+            else min(ttft_max, tpot_max)
+        )
+
+        tput_at_both = (
+            _value_at_concurrency(tput_group_df, conc_col, cfg, both)
+            if tput_group_df is not None
+            else pd.NA
+        )
+        ttft_at_both = (
+            _value_at_concurrency(ttft_group_df, conc_col, cfg, both)
+            if ttft_group_df is not None
+            else pd.NA
+        )
+        tpot_at_both = (
+            _value_at_concurrency(tpot_group_df, conc_col, cfg, both)
+            if tpot_group_df is not None
+            else pd.NA
+        )
+
+        rows.append(
+            {
+                "Configuration": cfg,
+                f"Max {conc_col} (TTFT ≤ {args.ttft_max_ms:g} ms)": ttft_max,
+                f"Max {conc_col} (TPOT ≤ {args.tpot_max_ms:g} ms)": tpot_max,
+                f"Max {conc_col} (Both)": both,
+                "Output Tput @ Both (tok/s)": tput_at_both,
+                "TTFT @ Both (ms)": ttft_at_both,
+                "TPOT @ Both (ms)": tpot_at_both,
+            }
+        )
+
+    df = pd.DataFrame(rows)
+    for c in df.columns:
+        if c != "Configuration":
+            df[c] = pd.to_numeric(df[c], errors="coerce")
+    return df
+
+
 # -----------------------------
 # Plot helper
 # -----------------------------
@@ -537,6 +750,21 @@ def build_parser() -> argparse.ArgumentParser:
         default=100.0,
         help="Reference limit for TPOT plots (ms)",
     )
+
+    # ---- NEW: export options ----
+    parser.add_argument(
+        "--excel-out",
+        type=str,
+        default="perf_comparison.xlsx",
+        help="Write one sheet per (Model, Dataset, Input Len, Output Len).",
+    )
+    parser.add_argument(
+        "--csv-out-dir",
+        type=str,
+        default="",
+        help="If set, write per-group per-metric CSVs into this directory.",
+    )
+
     return parser
 
 
@@ -657,7 +885,6 @@ def maybe_write_plot(
         markers=True,
     )
 
-    # Ensure plot hover + y tick labels are also 2 decimals.
     fig.update_traces(hovertemplate="%{y:.2f}<extra></extra>")
     fig.update_yaxes(tickformat=".2f")
 
@@ -730,87 +957,151 @@ def write_report_group_first(
         for metric_label, (df, _) in metric_cache.items()
     }
 
-    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:
-            gkey_tuple = normalize_group_key(gkey)
-            suffix = build_group_suffix(group_cols_canonical, gkey_tuple)
-            sub_path = group_filename(gkey_tuple)
-            group_header = (
-                '<div style="font-size: 1.4em; font-weight: 700; '
-                'margin: 18px 0 10px 0;">'
-                f"{_html.escape(suffix)}"
-                "</div>\n"
-            )
+    csv_dir = Path(args.csv_out_dir) if args.csv_out_dir else None
+    if csv_dir:
+        csv_dir.mkdir(parents=True, exist_ok=True)
 
-            main_fh.write(group_header)
-            with open(sub_path, "w", encoding="utf-8") as sub_fh:
-                sub_fh.write('<meta charset="utf-8">\n')
-                sub_fh.write(group_header)
-                tput_group_df = None
-                ttft_group_df = None
-                tpot_group_df = None
-                conc_col = args.xaxis
+    excel_path = args.excel_out or "perf_comparison.xlsx"
+    with pd.ExcelWriter(excel_path, engine="openpyxl") as xw:
+        # ---- Environment sheet (first) ----
+        env_sheet = _sanitize_sheet_name("Environment")
+        env_df = _load_env_df_for_inputs(args, files)
+        if env_df is None or env_df.empty:
+            pd.DataFrame(
+                [
+                    {
+                        "Section": "Environment",
+                        "Key": "vllm_env.txt",
+                        "Value": "NOT FOUND (or empty)",
+                    }
+                ]
+            ).to_excel(xw, sheet_name=env_sheet, index=False)
+        else:
+            env_df.to_excel(xw, sheet_name=env_sheet, index=False)
+        with open("perf_comparison.html", "w", encoding="utf-8") as main_fh:
+            main_fh.write('<meta charset="utf-8">\n')
+            for gkey in group_keys:
+                gkey_tuple = normalize_group_key(gkey)
+                suffix = build_group_suffix(group_cols_canonical, gkey_tuple)
+                sub_path = group_filename(gkey_tuple)
+                group_header = (
+                    '<div style="font-size: 1.4em; font-weight: 700; '
+                    'margin: 18px 0 10px 0;">'
+                    f"{_html.escape(suffix)}"
+                    "</div>\n"
+                )
 
-                for metric_label in plan.data_cols:
-                    gb = metric_groupbys[metric_label]
-                    df_sorted, raw_data_cols = metric_cache[metric_label]
+                main_fh.write(group_header)
 
-                    try:
-                        group_df = gb.get_group(gkey)
-                    except KeyError:
-                        missing = (
-                            '<div style="font-size: 1.1em; font-weight: 600; '
-                            'margin: 10px 0;">'
-                            f"{_html.escape(metric_label)} — missing for this group"
-                            "</div>\n"
+                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}")
+
+                excel_blocks: list[tuple[str, pd.DataFrame]] = []
+
+                with open(sub_path, "w", encoding="utf-8") as sub_fh:
+                    sub_fh.write('<meta charset="utf-8">\n')
+                    sub_fh.write(group_header)
+                    tput_group_df = None
+                    ttft_group_df = None
+                    tpot_group_df = None
+                    conc_col = args.xaxis
+
+                    for metric_label in plan.data_cols:
+                        gb = metric_groupbys[metric_label]
+                        df_sorted, raw_data_cols = metric_cache[metric_label]
+
+                        try:
+                            group_df = gb.get_group(gkey)
+                        except KeyError:
+                            missing = (
+                                '<div style="font-size: 1.1em; font-weight: 600; '
+                                'margin: 10px 0;">'
+                                f"{_html.escape(metric_label)} — missing for this group"
+                                "</div>\n"
+                            )
+                            main_fh.write(missing)
+                            sub_fh.write(missing)
+                            continue
+
+                        if conc_col not in group_df.columns:
+                            conc_col = _find_concurrency_col(group_df)
+
+                        mn = metric_label.lower().strip()
+                        if "tok/s" in mn:
+                            tput_group_df = group_df
+                        elif "ttft" in mn:
+                            ttft_group_df = group_df
+                        elif mn in ("p99", "median") or "tpot" in mn:
+                            tpot_group_df = group_df
+
+                        display_group = group_df.drop(
+                            columns=group_cols_canonical, errors="ignore"
                         )
 
-                        main_fh.write(missing)
-                        sub_fh.write(missing)
-                        continue
+                        html = render_metric_table_html(
+                            display_group, metric_label, suffix, args
+                        )
+                        main_fh.write(html)
+                        sub_fh.write(html)
 
-                    if conc_col not in group_df.columns:
-                        conc_col = _find_concurrency_col(group_df)
+                        maybe_write_plot(
+                            main_fh,
+                            sub_fh,
+                            group_df=group_df,
+                            raw_data_cols=raw_data_cols,
+                            metric_label=metric_label,
+                            y_axis_col=y_axis_col,
+                            args=args,
+                        )
 
-                    mn = metric_label.lower().strip()
-                    if "tok/s" in mn:
-                        tput_group_df = group_df
-                    elif "ttft" in mn:
-                        ttft_group_df = group_df
-                    elif mn in ("p99", "median") or "tpot" in mn:
-                        tpot_group_df = group_df
+                        excel_blocks.append(
+                            (metric_label, display_group.reset_index(drop=True))
+                        )
+                        if csv_dir:
+                            fn = _safe_filename(
+                                f"{sheet}__{metric_label}".replace(" ", "_").replace(
+                                    "/", "_"
+                                )
+                            )
+                            display_group.to_csv(csv_dir / f"{fn}.csv", index=False)
 
-                    display_group = group_df.drop(
-                        columns=group_cols_canonical, errors="ignore"
-                    )
-
-                    html = render_metric_table_html(
-                        display_group, metric_label, suffix, args
-                    )
-                    main_fh.write(html)
-                    sub_fh.write(html)
-
-                    maybe_write_plot(
-                        main_fh,
-                        sub_fh,
-                        group_df=group_df,
-                        raw_data_cols=raw_data_cols,
-                        metric_label=metric_label,
-                        y_axis_col=y_axis_col,
+                    summary_html = build_valid_max_concurrency_summary_html(
+                        tput_group_df=tput_group_df,
+                        ttft_group_df=ttft_group_df,
+                        tpot_group_df=tpot_group_df,
+                        conc_col=conc_col,
                         args=args,
                     )
+                    if summary_html:
+                        main_fh.write(summary_html)
+                        sub_fh.write(summary_html)
 
-                summary_html = build_valid_max_concurrency_summary_html(
-                    tput_group_df=tput_group_df,
-                    ttft_group_df=ttft_group_df,
-                    tpot_group_df=tpot_group_df,
-                    conc_col=conc_col,
-                    args=args,
-                )
-                if summary_html:
-                    main_fh.write(summary_html)
-                    sub_fh.write(summary_html)
+                    summary_df = build_valid_max_concurrency_summary_df(
+                        tput_group_df=tput_group_df,
+                        ttft_group_df=ttft_group_df,
+                        tpot_group_df=tpot_group_df,
+                        conc_col=conc_col,
+                        args=args,
+                    )
+                    if summary_df is not None:
+                        excel_blocks.append(
+                            ("Valid Max Concurrency Summary", summary_df)
+                        )
+                        if csv_dir:
+                            fn = _safe_filename(
+                                f"{sheet}__Valid_Max_Concurrency_Summary"
+                            )
+                            summary_df.to_csv(csv_dir / f"{fn}.csv", index=False)
+
+                _write_tables_to_excel_sheet(xw, sheet, excel_blocks)
+
+    print(f"Wrote Excel: {excel_path}")
+    if csv_dir:
+        print(f"Wrote CSVs under: {csv_dir}")
 
 
 def main():

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

@@ -1,6 +1,4 @@
 #!/bin/bash
-
-# This script should be run inside the CI process
 # This script assumes that we are already inside the vllm/ directory
 # Benchmarking results will be available inside vllm/benchmarks/results/
 
@@ -9,6 +7,11 @@
 set -x
 set -o pipefail
 
+# Environment-driven debug controls (like ON_CPU=1)
+DRY_RUN="${DRY_RUN:-0}"
+MODEL_FILTER="${MODEL_FILTER:-}"
+DTYPE_FILTER="${DTYPE_FILTER:-}"
+
 check_gpus() {
   if command -v nvidia-smi; then
     # check the number of GPUs and GPU type.
@@ -112,13 +115,12 @@ json2envs() {
 }
 
 wait_for_server() {
-  # wait for vllm server to start
-  # return 1 if vllm server crashes
   local timeout_val="1200"
   timeout "$timeout_val" bash -c '
-    until curl -X POST localhost:8000/v1/completions; do
+    until curl -sf http://localhost:8000/v1/models >/dev/null; do
       sleep 1
-    done' && return 0 || return 1
+    done
+  '
 }
 
 kill_processes_launched_by_current_bash() {
@@ -252,37 +254,16 @@ run_benchmark_tests() {
   done
 }
 
-run_latency_tests() {
-  run_benchmark_tests "latency" "$1"
-}
+run_latency_tests() { run_benchmark_tests "latency" "$1"; }
+run_startup_tests() { run_benchmark_tests "startup" "$1"; }
+run_throughput_tests() { run_benchmark_tests "throughput" "$1"; }
 
-run_startup_tests() {
-  run_benchmark_tests "startup" "$1"
-}
-
-run_throughput_tests() {
-  run_benchmark_tests "throughput" "$1"
-}
-
-run_serving_tests() {
-  # run serving tests using `vllm bench serve` command
-  # $1: a json file specifying serving test cases
-  #
-  # Supported JSON formats:
-  # 1) Plain format: top-level array
-  #    [ { "test_name": "...", "server_parameters": {...}, ... }, ... ]
-  #
-  # 2) Default parameters field + plain format tests
-  #    {
-  #      "defaults": { ... },
-  #      "tests": [ { "test_name": "...", "server_parameters": {...}, ... }, ... ]
-  #    }
-
-  local serving_test_file
-  serving_test_file=$1
-
-  # Iterate over serving tests
-  jq -c '
+merge_serving_tests_stream() {
+  # Emit merged serving test objects, optionally filtered by MODEL_FILTER/DTYPE_FILTER in DRY_RUN mode.
+  # This helper does NOT modify JSON; it only filters the stream in dry-run mode.
+  local serving_test_file="$1"
+  # shellcheck disable=SC2016
+  local merged='
     if type == "array" then
       # Plain format: test cases array
       .[]
@@ -304,7 +285,50 @@ run_serving_tests() {
     else
       error("Unsupported serving test file format: must be array or object with .tests")
     end
-  ' "$serving_test_file" | while read -r params; do
+  '
+
+  jq -c "$merged" "$serving_test_file" | \
+  if [[ "${DRY_RUN:-0}" == "1" && ( "${MODEL_FILTER}${DTYPE_FILTER}" != "" ) ]]; then
+    jq -c --arg model "$MODEL_FILTER" --arg dtype "$DTYPE_FILTER" '
+      select((($model|length)==0)
+             or ((.server_parameters.model // "") == $model)
+             or ((.client_parameters.model // "") == $model))
+      | select((($dtype|length)==0) or ((.server_parameters.dtype // "") == $dtype))
+    '
+  else
+    cat
+  fi
+}
+
+run_serving_tests() {
+  # run serving tests using `vllm bench serve` command
+  # $1: a json file specifying serving test cases
+  #
+  # Supported JSON formats:
+  # 1) Plain format: top-level array
+  #    [ { "test_name": "...", "server_parameters": {...}, ... }, ... ]
+  #
+  # 2) Default parameters field + plain format tests
+  #    {
+  #      "defaults": { ... },
+  #      "tests": [ { "test_name": "...", "server_parameters": {...}, ... }, ... ]
+  #    }
+
+  local serving_test_file
+  serving_test_file=$1
+
+  # In dry-run mode, if filters are provided but no tests match, fail fast.
+  if [[ "${DRY_RUN:-0}" == "1" && ( "${MODEL_FILTER}${DTYPE_FILTER}" != "" ) ]]; then
+    local count
+    count=$(merge_serving_tests_stream "$serving_test_file" | wc -l | tr -d ' ')
+    if [[ "$count" -eq 0 ]]; then
+      echo "No matching serving tests found in $serving_test_file for model='$MODEL_FILTER' dtype='$DTYPE_FILTER'." >&2
+      return 0
+    fi
+  fi
+
+  # Iterate over serving tests (merged + optional filtered stream)
+  merge_serving_tests_stream "$serving_test_file" | while read -r params; do
     # get the test name, and append the GPU type back to it.
     test_name=$(echo "$params" | jq -r '.test_name')
     if [[ ! "$test_name" =~ ^serving_ ]]; then
@@ -373,7 +397,7 @@ run_serving_tests() {
     echo "Server command: $server_command"
     # support remote vllm server
     client_remote_args=""
-    if [[ -z "${REMOTE_HOST}" ]]; then
+    if [[ -z "${REMOTE_HOST}" && "${DRY_RUN:-0}" != "1" ]]; then
       bash -c "$server_command" &
       server_pid=$!
       # wait until the server is alive
@@ -384,6 +408,9 @@ run_serving_tests() {
         echo ""
         echo "vLLM failed to start within the timeout period."
       fi
+    elif [[ "${DRY_RUN:-0}" == "1" ]]; then
+        # dry-run: don't start server
+        echo "Dry Run."
     else
       server_command="Using Remote Server $REMOTE_HOST $REMOTE_PORT"
       if [[ ${REMOTE_PORT} ]]; then
@@ -402,9 +429,7 @@ run_serving_tests() {
     for qps in $qps_list; do
       # remove the surrounding single quote from qps
       if [[ "$qps" == *"inf"* ]]; then
-        echo "qps was $qps"
         qps="inf"
-        echo "now qps is $qps"
       fi
 
       # iterate over different max_concurrency
@@ -425,7 +450,9 @@ run_serving_tests() {
         echo "Running test case $test_name with qps $qps"
         echo "Client command: $client_command"
 
-        bash -c "$client_command"
+        if [[ "${DRY_RUN:-0}" != "1" ]]; then
+          bash -c "$client_command"
+        fi
 
         # record the benchmarking commands
         jq_output=$(jq -n \
@@ -443,12 +470,15 @@ run_serving_tests() {
     done
 
     # clean up
-    kill -9 $server_pid
-    kill_gpu_processes
+    if [[ "${DRY_RUN:-0}" != "1" ]]; then
+      kill -9 $server_pid
+      kill_gpu_processes
+    fi
   done
 }
 
 main() {
+
   local ARCH
   ARCH=''
   if [[ "$ON_CPU" == "1" ]]; then
@@ -458,7 +488,13 @@ main() {
      check_gpus
      ARCH="$arch_suffix"
   fi
-  check_hf_token
+
+  # DRY_RUN does not execute vLLM; do not require HF_TOKEN.
+  if [[ "${DRY_RUN:-0}" != "1" ]]; then
+    check_hf_token
+  else
+    echo "DRY_RUN=1 -> skip HF_TOKEN validation"
+  fi
 
   # dependencies
   (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
@@ -479,11 +515,16 @@ main() {
 
   # dump vllm info via vllm collect-env
   env_output=$(vllm collect-env)
-
   echo "$env_output" >"$RESULTS_FOLDER/vllm_env.txt"
 
   # benchmarking
-  run_serving_tests $QUICK_BENCHMARK_ROOT/tests/"${SERVING_JSON:-serving-tests$ARCH.json}"
+  run_serving_tests $QUICK_BENCHMARK_ROOT/tests/"${SERVING_JSON:-serving-tests$ARCH.json}" || exit $?
+
+  if [[ "${DRY_RUN:-0}" == "1" ]]; then
+    echo "DRY_RUN=1 -> skip latency/startup/throughput suites"
+    exit 0
+  fi
+
   run_latency_tests $QUICK_BENCHMARK_ROOT/tests/"${LATENCY_JSON:-latency-tests$ARCH.json}"
   run_startup_tests $QUICK_BENCHMARK_ROOT/tests/"${STARTUP_JSON:-startup-tests$ARCH.json}"
   run_throughput_tests $QUICK_BENCHMARK_ROOT/tests/"${THROUGHPUT_JSON:-throughput-tests$ARCH.json}"

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

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

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

@@ -0,0 +1,283 @@
+{
+  "defaults": {
+    "qps_list": [
+      "inf"
+    ],
+    "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
+    "server_environment_variables": {
+      "VLLM_RPC_TIMEOUT": 100000,
+      "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+      "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+      "VLLM_CPU_SGL_KERNEL": 1,
+      "VLLM_CPU_KVCACHE_SPACE": 40
+    },
+    "server_parameters": {
+      "model": "meta-llama/Llama-3.1-8B-Instruct",
+      "tensor_parallel_size": 1,
+      "dtype": "bfloat16",
+      "distributed_executor_backend": "mp",
+      "block_size": 128,
+      "trust_remote_code": "",
+      "disable_log_stats": "",
+      "max_num_batched_tokens": 2048,
+      "max_num_seqs": 256
+    },
+    "client_parameters": {
+      "model": "meta-llama/Llama-3.1-8B-Instruct",
+      "backend": "vllm",
+      "ignore-eos": "",
+      "num_prompts": 200
+    }
+  },
+  "tests": [
+    {
+      "test_name": "serving_llama8B_tp1_sharegpt",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "sharegpt",
+        "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp2_sharegpt",
+      "server_parameters": {
+        "tensor_parallel_size": 2
+      },
+      "client_parameters": {
+        "dataset_name": "sharegpt",
+        "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp1_random_128_128",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp2_random_128_128",
+      "server_parameters": {
+        "tensor_parallel_size": 2
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp4_random_128_128",
+      "server_parameters": {
+        "tensor_parallel_size": 4
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp1_random_128_2048",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 2048
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp2_random_128_2048",
+      "server_parameters": {
+        "tensor_parallel_size": 2
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 2048
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp4_random_128_2048",
+      "server_parameters": {
+        "tensor_parallel_size": 4
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 2048
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp1_random_2048_128",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 2048,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp2_random_2048_128",
+      "server_parameters": {
+        "tensor_parallel_size": 2
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 2048,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp4_random_2048_128",
+      "server_parameters": {
+        "tensor_parallel_size": 4
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 2048,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_int4_tp1_random_128_128",
+      "server_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_int4_tp2_random_128_128",
+      "server_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "tensor_parallel_size": 2
+      },
+      "client_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_int4_tp4_random_128_128",
+      "server_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "tensor_parallel_size": 4
+      },
+      "client_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama3B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "meta-llama/Llama-3.2-3B-Instruct",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "meta-llama/Llama-3.2-3B-Instruct",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_granite2B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "ibm-granite/granite-3.2-2b-instruct",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "ibm-granite/granite-3.2-2b-instruct",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_qwen1.7B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "Qwen/Qwen3-1.7B",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "Qwen/Qwen3-1.7B",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_qwen4B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "Qwen/Qwen3-4B",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "Qwen/Qwen3-4B",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_qwen8B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "Qwen/Qwen3-8B",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "Qwen/Qwen3-8B",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_glm9B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "zai-org/glm-4-9b-hf",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "zai-org/glm-4-9b-hf",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_gemma7B_tp1_random_128_128",
+      "server_parameters": {
+        "model": "google/gemma-7b",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "google/gemma-7b",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    }
+  ]
+}

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

@@ -148,136 +148,6 @@
         "random-input-len": 2048,
         "random-output-len": 128
       }
-    },
-    {
-      "test_name": "serving_llama8B_int4_tp1_random_128_128",
-      "server_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_int4_tp2_random_128_128",
-      "server_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "tensor_parallel_size": 2
-      },
-      "client_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama8B_int4_tp4_random_128_128",
-      "server_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "tensor_parallel_size": 4
-      },
-      "client_parameters": {
-        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_llama3B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "meta-llama/Llama-3.2-3B-Instruct",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "meta-llama/Llama-3.2-3B-Instruct",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_granite2B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "ibm-granite/granite-3.2-2b-instruct",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "ibm-granite/granite-3.2-2b-instruct",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_qwen1.7B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "Qwen/Qwen3-1.7B",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "Qwen/Qwen3-1.7B",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_qwen4B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "Qwen/Qwen3-4B",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "Qwen/Qwen3-4B",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_qwen8B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "Qwen/Qwen3-8B",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "Qwen/Qwen3-8B",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_glm9B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "zai-org/glm-4-9b-hf",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "zai-org/glm-4-9b-hf",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
-    },
-    {
-      "test_name": "serving_gemma7B_tp1_random_128_128",
-      "server_parameters": {
-        "model": "google/gemma-7b",
-        "tensor_parallel_size": 1
-      },
-      "client_parameters": {
-        "model": "google/gemma-7b",
-        "dataset_name": "random",
-        "random-input-len": 128,
-        "random-output-len": 128
-      }
     }
   ]
 }

.buildkite/scripts/annotate-release.sh

@@ -27,7 +27,7 @@ aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cpu-cp38-
 To download and upload the image:
 
 \`\`\`
-Download images:
+# Download images:
 
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64
@@ -35,8 +35,12 @@ docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu130
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm-base
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm
+docker pull public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION}
+docker pull public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:v${RELEASE_VERSION}
 
-Tag and push images:
+# Tag and push images:
+
+## CUDA
 
 docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64 vllm/vllm-openai:x86_64
 docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:latest-x86_64
@@ -62,34 +66,21 @@ docker tag vllm/vllm-openai:aarch64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-a
 docker push vllm/vllm-openai:latest-aarch64-cu130
 docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
 
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-rocm
-docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-rocm vllm/vllm-openai-rocm:latest
-docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-rocm vllm/vllm-openai-rocm:v${RELEASE_VERSION}-rocm
+## ROCm
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT} vllm/vllm-openai-rocm:latest
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT} vllm/vllm-openai-rocm:v${RELEASE_VERSION}
 docker push vllm/vllm-openai-rocm:latest
-docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}-rocm
+docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}
 
-Create multi-arch manifest:
 docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm-base vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base
 docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base vllm/vllm-openai-rocm:latest-base
 docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
 docker push vllm/vllm-openai-rocm:latest-base
 docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
 
-docker manifest rm vllm/vllm-openai:latest
-docker manifest create vllm/vllm-openai:latest vllm/vllm-openai:latest-x86_64 vllm/vllm-openai:latest-aarch64
-docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
-docker manifest push vllm/vllm-openai:latest
-docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}
-
-docker manifest rm vllm/vllm-openai:latest-cu130
-docker manifest create vllm/vllm-openai:latest-cu130 vllm/vllm-openai:latest-x86_64-cu130 vllm/vllm-openai:latest-aarch64-cu130
-docker manifest create vllm/vllm-openai:v${RELEASE_VERSION}-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
-docker manifest push vllm/vllm-openai:latest-cu130
-docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}-cu130
-
-# CPU images (vllm/vllm-openai-cpu)
-docker pull public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION}
-docker pull public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:v${RELEASE_VERSION}
+## CPU
 
 docker tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION} vllm/vllm-openai-cpu:x86_64
 docker tag vllm/vllm-openai-cpu:x86_64 vllm/vllm-openai-cpu:latest-x86_64
@@ -103,6 +94,20 @@ docker tag vllm/vllm-openai-cpu:arm64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-a
 docker push vllm/vllm-openai-cpu:latest-arm64
 docker push vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
 
+# Create multi-arch manifest:
+
+docker manifest rm vllm/vllm-openai:latest
+docker manifest create vllm/vllm-openai:latest vllm/vllm-openai:latest-x86_64 vllm/vllm-openai:latest-aarch64
+docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
+docker manifest push vllm/vllm-openai:latest
+docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}
+
+docker manifest rm vllm/vllm-openai:latest-cu130
+docker manifest create vllm/vllm-openai:latest-cu130 vllm/vllm-openai:latest-x86_64-cu130 vllm/vllm-openai:latest-aarch64-cu130
+docker manifest create vllm/vllm-openai:v${RELEASE_VERSION}-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
+docker manifest push vllm/vllm-openai:latest-cu130
+docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}-cu130
+
 docker manifest rm vllm/vllm-openai-cpu:latest || true
 docker manifest create vllm/vllm-openai-cpu:latest vllm/vllm-openai-cpu:latest-x86_64 vllm/vllm-openai-cpu:latest-arm64
 docker manifest create vllm/vllm-openai-cpu:v${RELEASE_VERSION} vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64

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

@@ -0,0 +1,26 @@
+#!/bin/bash
+set -euox pipefail
+
+echo "--- PP+TP"
+vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -pp=2 &
+server_pid=$!
+timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
+vllm bench serve \
+    --backend vllm \
+    --dataset-name random \
+    --model meta-llama/Llama-3.2-3B-Instruct \
+    --num-prompts 20 \
+    --endpoint /v1/completions
+kill -s SIGTERM $server_pid &
+
+echo "--- DP+TP"
+vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -dp=2 &
+server_pid=$!
+timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
+vllm bench serve \
+    --backend vllm \
+    --dataset-name random \
+    --model meta-llama/Llama-3.2-3B-Instruct \
+    --num-prompts 20 \
+    --endpoint /v1/completions
+kill -s SIGTERM $server_pid &

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

@@ -2,119 +2,19 @@
 
 # This script build the CPU docker image and run the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
-set -ex
+set -euox pipefail
 
 # allow to bind to different cores
 CORE_RANGE=${CORE_RANGE:-48-95}
-# used for TP/PP E2E test
-OMP_CORE_RANGE=${OMP_CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
+IMAGE_NAME="cpu-test-$NUMA_NODE"
+TIMEOUT_VAL=$1
+TEST_COMMAND=$2
 
-export CMAKE_BUILD_PARALLEL_LEVEL=32
-
-# Setup cleanup
-remove_docker_container() {
-    set -e;
-    docker rm -f cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"-avx2 || true;
-}
-trap remove_docker_container EXIT
-remove_docker_container
-
-# Try building the docker image
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --progress plain --tag cpu-test-"$NUMA_NODE" --target vllm-test -f docker/Dockerfile.cpu .
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --progress plain --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$NUMA_NODE"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
+# building the docker image
+echo "--- :docker: Building Docker image"
+docker build --progress plain --tag "$IMAGE_NAME" --target vllm-test -f docker/Dockerfile.cpu .
 
 # Run the image, setting --shm-size=4g for tensor parallel.
-docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=16 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
-docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=16 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-avx2
-
-function cpu_tests() {
-  set -e
-  export NUMA_NODE=$2
-
-  # list packages
-  docker exec cpu-test-"$NUMA_NODE"-avx2 bash -c "
-    set -e
-    pip list"
-
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
-    set -e
-    pip list"
-
-  # offline inference
-  docker exec cpu-test-"$NUMA_NODE"-avx2 bash -c "
-    set -e
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
-
-  # Run kernel tests
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
-    set -e
-    pytest -x -v -s tests/kernels/attention/test_cpu_attn.py
-    pytest -x -v -s tests/kernels/moe/test_cpu_fused_moe.py
-    pytest -x -v -s tests/kernels/test_onednn.py"
-
-  # Run basic model test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
-    set -e
-    # Note: disable until supports V1
-    # pytest -x -v -s tests/kernels/attention/test_cache.py -m cpu_model
-    # pytest -x -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
-
-    pytest -x -v -s tests/models/language/generation -m cpu_model
-    VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model
-
-    pytest -x -v -s tests/models/language/pooling -m cpu_model
-    pytest -x -v -s tests/models/multimodal/generation \
-                --ignore=tests/models/multimodal/generation/test_pixtral.py \
-                -m cpu_model"
-
-  # Run compressed-tensor test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
-    set -e
-    pytest -x -s -v \
-    tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs"
-
-  # Run AWQ/GPTQ test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
-    set -e
-    pytest -x -s -v \
-    tests/quantization/test_cpu_wna16.py"
-
-  # Run multi-lora tests
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
-    set -e
-    pytest -x -s -v \
-    tests/lora/test_qwenvl.py"
-
-  # online serving: tp+pp
-  docker exec cpu-test-"$NUMA_NODE" bash -c '
-    set -e
-    VLLM_CPU_OMP_THREADS_BIND=$E2E_OMP_THREADS VLLM_CPU_SGL_KERNEL=1 vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -pp=2 &
-    server_pid=$!
-    timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
-    vllm bench serve \
-      --backend vllm \
-      --dataset-name random \
-      --model meta-llama/Llama-3.2-3B-Instruct \
-      --num-prompts 20 \
-      --endpoint /v1/completions
-    kill -s SIGTERM $server_pid &'
-
-  # online serving: tp+dp
-  docker exec cpu-test-"$NUMA_NODE" bash -c '
-    set -e
-    VLLM_CPU_OMP_THREADS_BIND=$E2E_OMP_THREADS VLLM_CPU_SGL_KERNEL=1 vllm serve meta-llama/Llama-3.2-3B-Instruct -tp=2 -dp=2 &
-    server_pid=$!
-    timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
-    vllm bench serve \
-      --backend vllm \
-      --dataset-name random \
-      --model meta-llama/Llama-3.2-3B-Instruct \
-      --num-prompts 20 \
-      --endpoint /v1/completions
-    kill -s SIGTERM $server_pid &'
-}
-
-# All of CPU tests are expected to be finished less than 40 mins.
-export -f cpu_tests
-timeout 2.5h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+docker run --rm --cpuset-cpus=$CORE_RANGE --cpuset-mems=$NUMA_NODE -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN -e VLLM_CPU_KVCACHE_SPACE=16 -e VLLM_CPU_CI_ENV=1 -e VLLM_CPU_SIM_MULTI_NUMA=1 --shm-size=4g $IMAGE_NAME \
+        timeout $TIMEOUT_VAL bash -c "set -euox pipefail; echo \"--- Print packages\"; pip list; echo \"--- Running tests\"; ${TEST_COMMAND}"

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

@@ -39,6 +39,8 @@ docker run \
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --attention-backend=TRITON_ATTN
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --quantization fp8
+    python3 examples/offline_inference/basic/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4  --block-size 64 --enforce-eager
     python3 examples/offline_inference/basic/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2
     python3 examples/offline_inference/basic/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
     cd tests

.buildkite/test-amd.yaml

@@ -70,6 +70,7 @@ steps:
   - vllm/
   - tests/test_inputs.py
   - tests/test_outputs.py
+  - tests/test_pooling_params.py
   - tests/multimodal
   - tests/renderers
   - tests/standalone_tests/lazy_imports.py
@@ -82,6 +83,7 @@ steps:
   - python3 standalone_tests/lazy_imports.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
+  - pytest -v -s test_pooling_params.py
   - pytest -v -s -m 'cpu_test' multimodal
   - pytest -v -s renderers
   - pytest -v -s tokenizers_
@@ -130,7 +132,7 @@ steps:
   - tests/entrypoints/
   commands:
   - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/openai --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/instrumentator --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
+  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/openai --ignore=entrypoints/rpc --ignore=entrypoints/instrumentator --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
 
 - label: Entrypoints Integration Test (LLM) # 30min
   timeout_in_minutes: 40
@@ -177,14 +179,14 @@ steps:
   torch_nightly: true
   source_file_dependencies:
   - vllm/
-  - tests/entrypoints/sleep
   - tests/entrypoints/rpc
+  - tests/entrypoints/instrumentator
   - tests/tool_use
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/sleep
+  - pytest -v -s entrypoints/instrumentator
+  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
   - pytest -v -s tool_use
-  - PYTHONPATH=/vllm-workspace  pytest -v -s entrypoints/rpc
 
 - label: Entrypoints Integration Test (Pooling)
   timeout_in_minutes: 50
@@ -231,6 +233,7 @@ steps:
   - tests/compile/fullgraph/test_basic_correctness.py
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
+  - examples/offline_inference/new_weight_syncing/
   - tests/examples/offline_inference/data_parallel.py
   - tests/v1/distributed
   - tests/v1/engine/test_engine_core_client.py
@@ -266,10 +269,16 @@ steps:
   - pytest -v -s distributed/test_symm_mem_allreduce.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
+  # OLD rlhf examples
   - pushd ../examples/offline_inference
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
   - popd
+  # NEW rlhf examples
+  - pushd ../examples/offline_inference/new_weight_syncing
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_async_new_apis.py
+  - popd
 
 - label: Distributed Tests (8 GPUs) # 4min
   timeout_in_minutes: 10
@@ -505,7 +514,7 @@ steps:
     - 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/pooling/vision_language_pooling.py --seed 0
+    - 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
@@ -525,6 +534,7 @@ steps:
   - tests/cuda
   commands:
     - pytest -v -s cuda/test_cuda_context.py
+    - pytest -v -s cuda/test_platform_no_cuda_init.py
 
 - label: Samplers Test # 56min
   timeout_in_minutes: 75
@@ -854,10 +864,11 @@ steps:
   torch_nightly: true
   source_file_dependencies:
   - vllm/
+  - tests/models/test_terratorch.py
   - tests/models/test_transformers.py
   - tests/models/test_registry.py
   commands:
-    - pytest -v -s models/test_transformers.py models/test_registry.py
+    - pytest -v -s models/test_terratorch.py models/test_transformers.py models/test_registry.py
 
 - label: Basic Models Test (Other CPU) # 5min
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -1180,16 +1191,16 @@ steps:
   - vllm/model_executor/layers/layernorm.py
   - vllm/model_executor/layers/activation.py
   - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/test_fusion_attn.py
-  - tests/compile/test_silu_mul_quant_fusion.py
-  - tests/compile/distributed/test_fusion_all_reduce.py
+  - tests/compile/passes/test_fusion_attn.py
+  - tests/compile/passes/test_silu_mul_quant_fusion.py
+  - tests/compile/passes/distributed/test_fusion_all_reduce.py
   - tests/compile/fullgraph/test_full_graph.py
   commands:
     - nvidia-smi
-    - pytest -v -s tests/compile/test_fusion_attn.py
-    - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
+    - pytest -v -s tests/compile/passes/test_fusion_attn.py
+    - pytest -v -s tests/compile/passes/test_silu_mul_quant_fusion.py
     # this runner has 2 GPUs available even though num_gpus=2 is not set
-    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
+    - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
 
     # # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
     # # Wrap with quotes to escape yaml
@@ -1323,7 +1334,7 @@ steps:
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
   - VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
-  - pytest -v -s distributed/test_sequence_parallel.py
+  - pytest -v -s compile/correctness_e2e/test_sequence_parallel.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
   - pytest -v -s v1/worker/test_worker_memory_snapshot.py
 
@@ -1546,15 +1557,15 @@ steps:
   working_dir: "/vllm-workspace/"
   num_gpus: 2
   commands:
-    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_async_tp.py
-    - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
-    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
+    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/passes/distributed/test_async_tp.py
+    - pytest -v -s tests/compile/passes/distributed/test_sequence_parallelism.py
+    - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
     #- pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
     # - "VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'"
     # Old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293
     # in favor of new tests in fusions_e2e. We avoid replicating the new jobs in this file as it's deprecated.
 
-    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/distributed/test_sequence_parallel.py
+    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/correctness_e2e/test_sequence_parallel.py
     - pytest -v -s tests/distributed/test_context_parallel.py
     - HIP_VISIBLE_DEVICES=0,1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=allgather_reducescatter --disable-nccl-for-dp-synchronization
     - pytest -v -s tests/v1/distributed/test_dbo.py

.buildkite/test-pipeline.yaml

@@ -63,6 +63,7 @@ steps:
   - vllm/
   - tests/test_inputs.py
   - tests/test_outputs.py
+  - tests/test_pooling_params.py
   - tests/multimodal
   - tests/renderers
   - tests/standalone_tests/lazy_imports.py
@@ -75,6 +76,7 @@ steps:
   - python3 standalone_tests/lazy_imports.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
+  - pytest -v -s test_pooling_params.py
   - pytest -v -s -m 'cpu_test' multimodal
   - pytest -v -s renderers
   - pytest -v -s tokenizers_
@@ -116,7 +118,7 @@ steps:
   - tests/entrypoints/
   commands:
   - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
+  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
 
 - label: Entrypoints Integration Test (LLM) # 30min
   timeout_in_minutes: 40
@@ -146,7 +148,7 @@ steps:
   - tests/entrypoints/test_chat_utils
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/  --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/instrumentator --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/  --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses
   - pytest -v -s entrypoints/test_chat_utils.py
 
 - label: Entrypoints Integration Test (API Server 2)
@@ -157,13 +159,13 @@ steps:
   torch_nightly: true
   source_file_dependencies:
   - vllm/
-  - tests/entrypoints/sleep
   - tests/entrypoints/rpc
+  - tests/entrypoints/instrumentator
   - tests/tool_use
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/sleep
-  - PYTHONPATH=/vllm-workspace  pytest -v -s entrypoints/rpc
+  - pytest -v -s entrypoints/instrumentator
+  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
   - pytest -v -s tool_use
 
 - label: Entrypoints Integration Test (Pooling)
@@ -204,6 +206,7 @@ steps:
   - tests/compile/fullgraph/test_basic_correctness.py
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
+  - examples/offline_inference/new_weight_syncing/
   - tests/examples/offline_inference/data_parallel.py
   - tests/v1/distributed
   - tests/v1/engine/test_engine_core_client.py
@@ -238,10 +241,16 @@ steps:
   - pytest -v -s distributed/test_symm_mem_allreduce.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
+  # OLD rlhf examples
   - pushd ../examples/offline_inference
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
   - popd
+  # NEW rlhf examples
+  - pushd ../examples/offline_inference/new_weight_syncing
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_async_new_apis.py
+  - popd
 
 - label: Distributed Tests (8 GPUs) # 4min
   timeout_in_minutes: 10
@@ -444,7 +453,7 @@ steps:
     - 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/pooling/vision_language_pooling.py --seed 0
+    - 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
@@ -510,6 +519,7 @@ steps:
   # However, find does not normally propagate error codes, so we combine it with xargs
   # (using -0 for proper path handling)
   - "find compile/ -maxdepth 1 -name 'test_*.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
+  - pytest -s -v compile/passes --ignore compile/passes/distributed
 
 - label: PyTorch Fullgraph Smoke Test # 15min
   timeout_in_minutes: 30
@@ -795,10 +805,11 @@ steps:
   torch_nightly: true
   source_file_dependencies:
   - vllm/
+  - tests/models/test_terratorch.py
   - tests/models/test_transformers.py
   - tests/models/test_registry.py
   commands:
-    - pytest -v -s models/test_transformers.py models/test_registry.py
+    - pytest -v -s models/test_terratorch.py models/test_transformers.py models/test_registry.py
 
 - label: Basic Models Test (Other CPU) # 5min
   timeout_in_minutes: 10
@@ -851,7 +862,7 @@ steps:
   commands:
     # Install fast path packages for testing against transformers
     # Note: also needed to run plamo2 model in vLLM
-    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.2.5'
+    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
     # Shard hybrid language model tests
     - pytest -v -s models/language/generation \
@@ -870,7 +881,7 @@ steps:
   commands:
     # Install fast path packages for testing against transformers
     # Note: also needed to run plamo2 model in vLLM
-    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.2.5'
+    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
     - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
@@ -1070,14 +1081,14 @@ steps:
   - vllm/model_executor/layers/quantization/input_quant_fp8.py
   - tests/compile/test_fusion_attn.py
   - tests/compile/test_silu_mul_quant_fusion.py
-  - tests/compile/distributed/test_fusion_all_reduce.py
+  - tests/compile/passes/distributed/test_fusion_all_reduce.py
   - tests/compile/fullgraph/test_full_graph.py
   commands:
     - nvidia-smi
     - pytest -v -s tests/compile/test_fusion_attn.py
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
     # this runner has 2 GPUs available even though num_gpus=2 is not set
-    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
+    - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
     #  # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
     #  # Wrap with quotes to escape yaml
     #  - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
@@ -1144,6 +1155,8 @@ steps:
   - pytest -v -s distributed/test_shm_broadcast.py
   - pytest -v -s distributed/test_shm_buffer.py
   - pytest -v -s distributed/test_shm_storage.py
+  - pytest -v -s distributed/test_packed_tensor.py
+  - pytest -v -s distributed/test_weight_transfer.py
 
 - label: 2 Node Tests (4 GPUs in total) # 16min
   timeout_in_minutes: 30
@@ -1409,8 +1422,8 @@ steps:
   commands:
     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
     # Run sequence parallel tests
-    - pytest -v -s tests/distributed/test_sequence_parallel.py
-    - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
+    - pytest -v -s tests/compile/correctness_e2e/test_sequence_parallel.py
+    - pytest -v -s tests/compile/passes/distributed/test_sequence_parallelism.py
 
 - label: Distributed Tests (H100) # optional
   gpu: h100
@@ -1418,7 +1431,7 @@ steps:
   working_dir: "/vllm-workspace/"
   num_gpus: 2
   commands:
-    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_async_tp.py
+    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/passes/distributed/test_async_tp.py
     - pytest -v -s tests/distributed/test_context_parallel.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/basic_correctness.yaml

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

.buildkite/test_areas/benchmarks.yaml

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

.buildkite/test_areas/compile.yaml

@@ -2,7 +2,7 @@ group: Compile
 depends_on: 
   - image-build
 steps:
-- label: Sequence Parallel Tests (2 GPUs)
+- label: Sequence Parallel Correctness Tests (2 GPUs)
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/"
   num_devices: 2
@@ -11,12 +11,12 @@ steps:
   - vllm/compilation/
   - vllm/v1/worker/
   - vllm/v1/cudagraph_dispatcher.py
-  - tests/distributed/test_sequence_parallel.py
+  - tests/compile/correctness_e2e/test_sequence_parallel.py
   commands:
   - export VLLM_TEST_CLEAN_GPU_MEMORY=1
-  - pytest -v -s tests/distributed/test_sequence_parallel.py
+  - pytest -v -s tests/compile/correctness_e2e/test_sequence_parallel.py
 
-- label: Sequence Parallel Tests (2xH100)
+- label: Sequence Parallel Correctness Tests (2xH100)
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/"
   device: h100
@@ -24,24 +24,30 @@ steps:
   num_devices: 2
   commands:
   - export VLLM_TEST_CLEAN_GPU_MEMORY=1
-  - pytest -v -s tests/distributed/test_sequence_parallel.py
+  - pytest -v -s tests/compile/correctness_e2e/test_sequence_parallel.py
+
+- label: AsyncTP Correctness Tests (2xH100)
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/"
+  device: h100
+  optional: true
+  num_devices: 2
+  commands:
+  - export VLLM_TEST_CLEAN_GPU_MEMORY=1
+  - pytest -v -s tests/compile/correctness_e2e/test_async_tp.py
 
 - label: Distributed Compile Unit Tests (2xH100)
-  timeout_in_minutes: 40
+  timeout_in_minutes: 20
   working_dir: "/vllm-workspace/"
   device: h100
   num_devices: 2
   source_file_dependencies:
   - vllm/compilation/
   - vllm/model_executor/layers
-  - tests/compile/distributed/test_fusion_all_reduce.py
-  - tests/compile/distributed/test_sequence_parallelism.py
-  - tests/compile/distributed/test_async_tp.py
+  - tests/compile/passes/distributed/
   commands:
   - export VLLM_TEST_CLEAN_GPU_MEMORY=1
-  - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
-  - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
-  - pytest -v -s tests/compile/distributed/test_async_tp.py
+  - pytest -s -v tests/compile/passes/distributed
 
 - label: Fusion and Compile Unit Tests (B200)
   timeout_in_minutes: 20
@@ -55,17 +61,17 @@ steps:
   - vllm/model_executor/layers/attention/attention.py
   - vllm/v1/attention/backends/flashinfer.py
   - vllm/compilation/ # TODO(luka) limit to vllm/compilation/passes
-  - tests/compile/test_fusion_attn.py
-  - tests/compile/test_silu_mul_quant_fusion.py
-  - tests/compile/distributed/test_fusion_all_reduce.py
+  - tests/compile/passes/test_fusion_attn.py
+  - tests/compile/passes/test_silu_mul_quant_fusion.py
+  - tests/compile/passes/distributed/test_fusion_all_reduce.py
   - tests/compile/fullgraph/test_full_graph.py
   commands:
     # b200 runners are limited, so we limit the tests to the minimum set only supported on Blackwell
     - nvidia-smi
-    - pytest -v -s tests/compile/test_fusion_attn.py -k FLASHINFER
-    - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
+    - pytest -v -s tests/compile/passes/test_fusion_attn.py -k FLASHINFER
+    - pytest -v -s tests/compile/passes/test_silu_mul_quant_fusion.py
     # this runner has 2 GPUs available even though num_devices=2 is not set
-    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
+    - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
     # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
     # TODO(luka) move to H100 once pass tests run on H100
     - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
@@ -115,13 +121,10 @@ steps:
   optional: true
   commands:
     - nvidia-smi
-    # Run all models and attn backends but only Inductor partition and native custom ops
-    # -k "inductor_partition and not +rms_norm and not +quant_fp8"
+    # 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
-    # -k "inductor_partition and not +rms_norm and +quant_fp8 and qwen3"
-    # Run just llama3 (fp8 & fp4) for all config combinations
-    # -k "llama-3"
-    - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and not +rms_norm and not +quant_fp8" -k "inductor_partition and not +rms_norm and +quant_fp8 and qwen3" -k "llama-3"
+    # 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)"
 
 - label: Fusion E2E TP2 Quick (H100)
   timeout_in_minutes: 20
@@ -156,7 +159,7 @@ steps:
     - tests/compile/fusions_e2e/
   commands:
     - nvidia-smi
-    # Run just llama3 (fp4 & fp8 & bf16) for all config combinations
+    # Run just llama3 (fp8 & bf16) for all config combinations
     - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "llama-3"
 
 - label: Fusion E2E TP2 AsyncTP Config Sweep (H100)
@@ -191,7 +194,8 @@ steps:
     - tests/compile/fusions_e2e/
   commands:
     - nvidia-smi
-    # Run all models and attn backends but only Inductor partition and native custom ops
+    # 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
     # for ar-rms-quant-fp4, also sweep llama3
-    - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "inductor_partition and not +rms_norm and not +quant_fp8" -k "Llama-3.1-8B-Instruct-FP4"
-    - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "inductor_partition and not +rms_norm and not +quant_fp8"
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "(FLASHINFER and inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and qwen3)) or Llama-3.1-8B-Instruct-FP4"
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "FLASHINFER and inductor_partition and not +rms_norm and (not +quant_fp8 or +quant_fp8 and qwen3)"

.buildkite/test_areas/cuda.yaml

@@ -9,6 +9,7 @@ steps:
   - tests/cuda
   commands:
     - pytest -v -s cuda/test_cuda_context.py
+    - pytest -v -s cuda/test_platform_no_cuda_init.py
 
 - label: Cudagraph
   timeout_in_minutes: 20

.buildkite/test_areas/distributed.yaml

@@ -62,6 +62,7 @@ steps:
   - tests/compile/fullgraph/test_basic_correctness.py
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
+  - examples/offline_inference/new_weight_syncing/
   - tests/examples/offline_inference/data_parallel.py
   - tests/v1/distributed
   - tests/v1/engine/test_engine_core_client.py
@@ -96,9 +97,14 @@ steps:
   - pytest -v -s distributed/test_symm_mem_allreduce.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
+  # OLD rlhf examples
   - cd ../examples/offline_inference
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
+  # NEW rlhf examples
+  - cd new_weight_syncing
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_async_new_apis.py
 
 - label: Distributed Tests (8 GPUs)(H100)
   timeout_in_minutes: 10

.buildkite/test_areas/entrypoints.yaml

@@ -24,6 +24,11 @@ steps:
   - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
 
 - label: Entrypoints Integration (API Server 1)
   timeout_in_minutes: 130
@@ -42,15 +47,13 @@ steps:
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
-  - tests/tool_use
-  - tests/entrypoints/sleep
-  - tests/entrypoints/instrumentator
   - tests/entrypoints/rpc
+  - tests/entrypoints/instrumentator
+  - tests/tool_use
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
   - pytest -v -s entrypoints/instrumentator
-  - pytest -v -s entrypoints/sleep
+  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
   - pytest -v -s tool_use
 
 - label: Entrypoints Integration (Pooling)

.buildkite/test_areas/misc.yaml

@@ -72,7 +72,7 @@ steps:
     - 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/pooling/vision_language_pooling.py --seed 0
+    - 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
@@ -122,6 +122,7 @@ steps:
   - vllm/
   - tests/test_inputs.py
   - tests/test_outputs.py
+  - tests/test_pooling_params.py
   - tests/multimodal
   - tests/renderers
   - tests/standalone_tests/lazy_imports.py
@@ -134,6 +135,7 @@ steps:
   - python3 standalone_tests/lazy_imports.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
+  - pytest -v -s test_pooling_params.py
   - pytest -v -s -m 'cpu_test' multimodal
   - pytest -v -s renderers
   - pytest -v -s tokenizers_

.buildkite/test_areas/models_basic.yaml

@@ -4,7 +4,6 @@ depends_on:
 steps:
 - label: Basic Models Tests (Initialization)
   timeout_in_minutes: 45
-  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -16,7 +15,6 @@ steps:
 
 - label: Basic Models Tests (Extra Initialization) %N
   timeout_in_minutes: 45
-  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/model_executor/models/
@@ -33,10 +31,17 @@ steps:
   timeout_in_minutes: 45
   source_file_dependencies:
   - vllm/
+  - tests/models/test_terratorch.py
   - tests/models/test_transformers.py
   - tests/models/test_registry.py
   commands:
-    - pytest -v -s models/test_transformers.py models/test_registry.py
+    - pytest -v -s models/test_terratorch.py models/test_transformers.py models/test_registry.py
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+    
 
 - label: Basic Models Test (Other CPU) # 5min
   depends_on: 

.buildkite/test_areas/models_language.yaml

@@ -4,7 +4,6 @@ depends_on:
 steps:
 - label: Language Models Tests (Standard)
   timeout_in_minutes: 25
-  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -16,7 +15,6 @@ steps:
 
 - label: Language Models Tests (Extra Standard) %N
   timeout_in_minutes: 45
-  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/model_executor/models/
@@ -32,7 +30,6 @@ steps:
 
 - label: Language Models Tests (Hybrid) %N
   timeout_in_minutes: 75
-  mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -40,7 +37,7 @@ steps:
   commands:
     # Install fast path packages for testing against transformers
     # Note: also needed to run plamo2 model in vLLM
-    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.2.5'
+    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
     # Shard hybrid language model tests
     - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
@@ -48,7 +45,6 @@ steps:
 
 - label: Language Models Test (Extended Generation) # 80min
   timeout_in_minutes: 110
-  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/
@@ -56,13 +52,12 @@ steps:
   commands:
     # Install fast path packages for testing against transformers
     # Note: also needed to run plamo2 model in vLLM
-    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.2.5'
+    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
     - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
 - label: Language Models Test (PPL)
   timeout_in_minutes: 110
-  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/
@@ -72,7 +67,6 @@ steps:
 
 - label: Language Models Test (Extended Pooling)  # 36min
   timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/
@@ -82,7 +76,6 @@ steps:
 
 - label: Language Models Test (MTEB)
   timeout_in_minutes: 110
-  mirror_hardwares: [amdexperimental]
   optional: true
   source_file_dependencies:
   - vllm/

.buildkite/test_areas/pytorch.yaml

@@ -3,7 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: PyTorch Compilation Unit Tests
-  timeout_in_minutes: 30
+  timeout_in_minutes: 10
   source_file_dependencies:
     - vllm/
     - tests/compile
@@ -17,6 +17,14 @@ steps:
   # (using -0 for proper path handling)
   - "find compile/ -maxdepth 1 -name 'test_*.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
 
+- label: PyTorch Compilation Passes Unit Tests
+  timeout_in_minutes: 20
+  source_file_dependencies:
+    - vllm/
+    - tests/compile/passes
+  commands:
+  - pytest -s -v compile/passes --ignore compile/passes/distributed
+
 - label: PyTorch Fullgraph Smoke Test
   timeout_in_minutes: 35
   source_file_dependencies:

.buildkite/test_areas/samplers.yaml

@@ -12,3 +12,10 @@ steps:
   commands:
     - pytest -v -s samplers
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+      commands:
+      - pytest -v -s -m 'not skip_v1' samplers

.github/CODEOWNERS

@@ -2,7 +2,9 @@
 # for more info about CODEOWNERS file
 
 # This lists cover the "core" components of vLLM that require careful review
-/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @njhill @22quinn
+/vllm/compilation @zou3519 @youkaichao @ProExpertProg
+/vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery
+/vllm/lora @jeejeelee
 /vllm/model_executor/layers/attention @LucasWilkinson
 /vllm/model_executor/layers/fused_moe @mgoin @pavanimajety
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety
@@ -11,18 +13,34 @@
 /vllm/model_executor/layers/batch_invariant.py @yewentao256 
 /vllm/multimodal @DarkLight1337 @ywang96 @NickLucche @tjtanaa
 /vllm/vllm_flash_attn @LucasWilkinson
-/vllm/lora @jeejeelee
-/vllm/reasoning @aarnphm @chaunceyjiang
-/vllm/entrypoints @aarnphm @chaunceyjiang
-/vllm/tool_parsers @aarnphm @chaunceyjiang
-/vllm/compilation @zou3519 @youkaichao @ProExpertProg
-/vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery
 CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
 # so spam a lot of people
 /vllm/config @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
-/vllm/config/cache.py @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg @heheda12345
+/vllm/config/cache.py @heheda12345
+
+# Entrypoints
+/vllm/entrypoints/anthropic @mgoin @DarkLight1337
+/vllm/entrypoints/cli @hmellor @mgoin @DarkLight1337 @russellb
+/vllm/entrypoints/mcp @heheda12345
+/vllm/entrypoints/openai @aarnphm @chaunceyjiang @DarkLight1337 @russellb
+/vllm/entrypoints/openai/realtime @njhill
+/vllm/entrypoints/openai/speech_to_text @NickLucche
+/vllm/entrypoints/pooling @noooop
+/vllm/entrypoints/sagemaker @DarkLight1337
+/vllm/entrypoints/serve @njhill
+/vllm/entrypoints/*.py @njhill
+/vllm/entrypoints/chat_utils.py @DarkLight1337
+/vllm/entrypoints/llm.py @DarkLight1337
+
+# Input/Output Processing
+/vllm/sampling_params.py @njhill @NickLucche
+/vllm/pooling_params.py @noooop @DarkLight1337
+/vllm/tokenizers @DarkLight1337 @njhill
+/vllm/renderers @DarkLight1337 @njhill
+/vllm/reasoning @aarnphm @chaunceyjiang
+/vllm/tool_parsers @aarnphm @chaunceyjiang
 
 # vLLM V1
 /vllm/v1/attention @LucasWilkinson
@@ -115,8 +133,8 @@ mkdocs.yaml @hmellor
 /vllm/model_executor/models/mixtral*.py @patrickvonplaten
 /vllm/model_executor/models/voxtral*.py @patrickvonplaten
 /vllm/model_executor/models/pixtral*.py @patrickvonplaten
+/vllm/tokenizers/mistral.py @patrickvonplaten
 /vllm/transformers_utils/configs/mistral.py @patrickvonplaten
-/vllm/transformers_utils/tokenizers/mistral.py @patrickvonplaten
 
 # Kernels
 /vllm/v1/attention/ops/chunked_prefill_paged_decode.py @tdoublep
@@ -152,9 +170,7 @@ mkdocs.yaml @hmellor
 /examples/pooling @noooop
 /tests/models/*/pooling* @noooop
 /tests/entrypoints/pooling @noooop
-/vllm/entrypoints/pooling @noooop
 /vllm/config/pooler.py @noooop
-/vllm/pooling_params.py @noooop
 /vllm/model_executor/layers/pooler @noooop
 
 # Security guide and policies

.github/workflows/cleanup_pr_body.yml

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

.gitignore

@@ -238,3 +238,6 @@ ep_kernels_workspace/
 vllm/grpc/vllm_engine_pb2.py
 vllm/grpc/vllm_engine_pb2_grpc.py
 vllm/grpc/vllm_engine_pb2.pyi
+
+# Ignore generated cpu headers 
+csrc/cpu/cpu_attn_dispatch_generated.h

.pre-commit-config.yaml

@@ -121,24 +121,9 @@ repos:
     name: Update Dockerfile dependency graph
     entry: tools/pre_commit/update-dockerfile-graph.sh
     language: script
-  - id: enforce-import-regex-instead-of-re
-    name: Enforce import regex as re
-    entry: python tools/pre_commit/enforce_regex_import.py
-    language: python
-    types: [python]
-    pass_filenames: false
-    additional_dependencies: [regex]
-  # forbid directly import triton
-  - id: forbid-direct-triton-import
-    name: "Forbid direct 'import triton'"
-    entry: python tools/pre_commit/check_triton_import.py
-    language: python
-    types: [python]
-    pass_filenames: false
-    additional_dependencies: [regex]
-  - id: check-pickle-imports
-    name: Prevent new pickle/cloudpickle imports
-    entry: python tools/pre_commit/check_pickle_imports.py
+  - id: check-forbidden-imports
+    name: Check for forbidden imports
+    entry: python tools/pre_commit/check_forbidden_imports.py
     language: python
     types: [python]
     additional_dependencies: [regex]
@@ -158,6 +143,11 @@ repos:
     name: Check attention backend documentation is up to date
     entry: python tools/pre_commit/generate_attention_backend_docs.py --check
     language: python
+  - id: check-boolean-context-manager
+    name: Check for boolean ops in with-statements
+    entry: python tools/pre_commit/check_boolean_context_manager.py
+    language: python
+    types: [python]
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

.readthedocs.yaml

@@ -9,13 +9,14 @@ build:
     python: "3.12"
   jobs:
     post_checkout:
-      - git fetch --unshallow || true
+      - git fetch origin main --unshallow --no-tags --filter=blob:none || true
+    pre_create_environment:
+      - pip install uv
+    create_environment:
+      - uv venv $READTHEDOCS_VIRTUALENV_PATH
+    install:
+      - uv pip install --python $READTHEDOCS_VIRTUALENV_PATH/bin/python --no-cache-dir -r requirements/docs.txt 
 
 mkdocs:
   configuration: mkdocs.yaml
   fail_on_warning: true
-
-# Optionally declare the Python requirements required to build your docs
-python:
-  install:
-    - requirements: requirements/docs.txt

CMakeLists.txt

@@ -56,8 +56,8 @@ endif()
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.9.1")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.9.1")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.10.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.10.0")
 
 #
 # Try to find python package with an executable that exactly matches
@@ -293,6 +293,7 @@ set(VLLM_EXT_SRC
   "csrc/fused_qknorm_rope_kernel.cu"
   "csrc/layernorm_quant_kernels.cu"
   "csrc/sampler.cu"
+  "csrc/topk.cu"
   "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/w8a8/int8/scaled_quant.cu"
@@ -433,7 +434,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_BF16_KERNEL_SRC})
     endif()
 
-    if (MARLIN_SM75_ARCHS) 
+    if (MARLIN_SM75_ARCHS)
       file(GLOB MARLIN_TEMPLATE_SM75_KERNEL_SRC "csrc/quantization/marlin/sm75_kernel_*.cu")
       set_gencode_flags_for_srcs(
         SRCS "${MARLIN_TEMPLATE_SM75_KERNEL_SRC}"
@@ -445,7 +446,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_SM75_KERNEL_SRC})
     endif()
 
-    if (MARLIN_FP8_ARCHS) 
+    if (MARLIN_FP8_ARCHS)
       file(GLOB MARLIN_TEMPLATE_FP8_KERNEL_SRC "csrc/quantization/marlin/sm89_kernel_*.cu")
       set_gencode_flags_for_srcs(
         SRCS "${MARLIN_TEMPLATE_FP8_KERNEL_SRC}"
@@ -1042,7 +1043,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_SRC})
     endif()
 
-    if (MARLIN_MOE_SM75_ARCHS) 
+    if (MARLIN_MOE_SM75_ARCHS)
       file(GLOB MARLIN_MOE_SM75_SRC "csrc/moe/marlin_moe_wna16/sm75_kernel_*.cu")
       set_gencode_flags_for_srcs(
         SRCS "${MARLIN_MOE_SM75_SRC}"

benchmarks/README.md

@@ -11,7 +11,7 @@ This directory used to contain vLLM's benchmark scripts and utilities for perfor
 
 ## Usage
 
-For detailed usage instructions, examples, and dataset information, see the [Benchmark CLI documentation](https://docs.vllm.ai/en/latest/contributing/benchmarks.html#benchmark-cli).
+For detailed usage instructions, examples, and dataset information, see the [Benchmark CLI documentation](https://docs.vllm.ai/en/latest/benchmarking/cli/#benchmark-cli).
 
 For full CLI reference see:
 

benchmarks/attention_benchmarks/batch_spec.py

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

benchmarks/attention_benchmarks/benchmark.py

@@ -43,6 +43,7 @@ from common import (
     ModelParameterSweep,
     ParameterSweep,
     ResultsFormatter,
+    batch_spec_sort_key,
     is_mla_backend,
 )
 
@@ -218,10 +219,13 @@ def run_model_parameter_sweep(
                         by_param_and_spec[key].append(r)
                         break
 
-    # Sort by param value then spec
+    # Sort by param value then spec (batch_size, q_len, kv_len)
     sorted_keys = sorted(
         by_param_and_spec.keys(),
-        key=lambda x: (int(x[0]) if x[0].isdigit() else x[0], x[1]),
+        key=lambda x: (
+            int(x[0]) if x[0].isdigit() else x[0],
+            batch_spec_sort_key(x[1]),
+        ),
     )
 
     current_param_value = None
@@ -330,7 +334,7 @@ def run_parameter_sweep(
                 by_spec[spec] = []
             by_spec[spec].append(r)
 
-    for spec in sorted(by_spec.keys()):
+    for spec in sorted(by_spec.keys(), key=batch_spec_sort_key):
         results = by_spec[spec]
         best = min(results, key=lambda r: r.mean_time)
         console.print(
@@ -496,15 +500,18 @@ def main():
         if "description" in yaml_config:
             console.print(f"[dim]{yaml_config['description']}[/]")
 
-        # Override args with YAML values
-        # (YAML takes precedence unless CLI arg was explicitly set)
-        # Backend(s)
-        if "backend" in yaml_config:
-            args.backend = yaml_config["backend"]
-            args.backends = None
-        elif "backends" in yaml_config:
-            args.backends = yaml_config["backends"]
-            args.backend = None
+        # Override args with YAML values, but CLI args take precedence
+        # Check if CLI provided backends (they would be non-None and not default)
+        cli_backends_provided = args.backends is not None or args.backend is not None
+
+        # Backend(s) - only use YAML if CLI didn't specify
+        if not cli_backends_provided:
+            if "backend" in yaml_config:
+                args.backend = yaml_config["backend"]
+                args.backends = None
+            elif "backends" in yaml_config:
+                args.backends = yaml_config["backends"]
+                args.backend = None
 
         # Check for special modes
         if "mode" in yaml_config:
@@ -544,13 +551,15 @@ def main():
             args.num_kv_heads = model.get("num_kv_heads", args.num_kv_heads)
             args.block_size = model.get("block_size", args.block_size)
 
-        # Benchmark settings
-        if "benchmark" in yaml_config:
-            bench = yaml_config["benchmark"]
-            args.device = bench.get("device", args.device)
-            args.repeats = bench.get("repeats", args.repeats)
-            args.warmup_iters = bench.get("warmup_iters", args.warmup_iters)
-            args.profile_memory = bench.get("profile_memory", args.profile_memory)
+        # Benchmark settings (top-level keys)
+        if "device" in yaml_config:
+            args.device = yaml_config["device"]
+        if "repeats" in yaml_config:
+            args.repeats = yaml_config["repeats"]
+        if "warmup_iters" in yaml_config:
+            args.warmup_iters = yaml_config["warmup_iters"]
+        if "profile_memory" in yaml_config:
+            args.profile_memory = yaml_config["profile_memory"]
 
         # Parameter sweep configuration
         if "parameter_sweep" in yaml_config:

benchmarks/attention_benchmarks/common.py

@@ -12,16 +12,36 @@ from typing import Any
 
 import numpy as np
 import torch
+from batch_spec import get_batch_type, parse_batch_spec
 from rich.console import Console
 from rich.table import Table
 
+
+def batch_spec_sort_key(spec: str) -> tuple[int, int, int]:
+    """
+    Extract sorting key from batch spec: (batch_size, max_q_len, max_kv_len).
+
+    This ensures results are sorted by batch size first, then query length,
+    then sequence length, rather than alphabetically.
+    """
+    try:
+        requests = parse_batch_spec(spec)
+        batch_size = len(requests)
+        max_q_len = max(r.q_len for r in requests) if requests else 0
+        max_kv_len = max(r.kv_len for r in requests) if requests else 0
+        return (batch_size, max_q_len, max_kv_len)
+    except Exception:
+        # Fallback for unparseable specs
+        return (0, 0, 0)
+
+
 # Mock classes for vLLM attention infrastructure
 
 
 class MockHfConfig:
     """Mock HuggingFace config that satisfies vLLM's requirements."""
 
-    def __init__(self, mla_dims: dict):
+    def __init__(self, mla_dims: dict, index_topk: int | None = None):
         self.num_attention_heads = mla_dims["num_q_heads"]
         self.num_key_value_heads = mla_dims["num_kv_heads"]
         self.hidden_size = mla_dims["head_dim"] * mla_dims["num_q_heads"]
@@ -32,6 +52,8 @@ class MockHfConfig:
         self.qk_rope_head_dim = mla_dims["qk_rope_head_dim"]
         self.v_head_dim = mla_dims["v_head_dim"]
         self.qk_head_dim = mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"]
+        if index_topk is not None:
+            self.index_topk = index_topk
 
     def get_text_config(self):
         return self
@@ -82,6 +104,38 @@ class MockKVBProj:
         return (result,)  # Return as tuple to match ColumnParallelLinear API
 
 
+class MockIndexer:
+    """Mock Indexer for sparse MLA backends.
+
+    Provides topk_indices_buffer that sparse MLA backends use to determine
+    which KV cache slots to attend to for each token.
+    """
+
+    def __init__(
+        self,
+        max_num_tokens: int,
+        topk_tokens: int,
+        device: torch.device,
+    ):
+        self.topk_tokens = topk_tokens
+        self.topk_indices_buffer = torch.zeros(
+            (max_num_tokens, topk_tokens),
+            dtype=torch.int32,
+            device=device,
+        )
+
+    def fill_random_indices(self, num_tokens: int, max_kv_len: int):
+        """Fill topk_indices_buffer with random valid indices for benchmarking."""
+        indices = torch.randint(
+            0,
+            max_kv_len,
+            (num_tokens, self.topk_tokens),
+            dtype=torch.int32,
+            device=self.topk_indices_buffer.device,
+        )
+        self.topk_indices_buffer[:num_tokens] = indices
+
+
 class MockLayer(AttentionLayerBase):
     """Mock attention layer with scale parameters and impl.
 
@@ -316,14 +370,19 @@ class ResultsFormatter:
             backends: List of backend names being compared
             compare_to_fastest: Show percentage comparison to fastest
         """
-        # Group by batch spec
+        # Group by batch spec, preserving first-occurrence order
         by_spec = {}
+        specs_order = []
         for r in results:
             spec = r.config.batch_spec
             if spec not in by_spec:
                 by_spec[spec] = {}
+                specs_order.append(spec)
             by_spec[spec][r.config.backend] = r
 
+        # Sort specs by (batch_size, q_len, kv_len) instead of alphabetically
+        specs_order = sorted(by_spec.keys(), key=batch_spec_sort_key)
+
         # Create shortened backend names for display
         def shorten_backend_name(name: str) -> str:
             """Shorten long backend names for table display."""
@@ -337,6 +396,8 @@ class ResultsFormatter:
 
         table = Table(title="Attention Benchmark Results")
         table.add_column("Batch\nSpec", no_wrap=True)
+        table.add_column("Type", no_wrap=True)
+        table.add_column("Batch\nSize", justify="right", no_wrap=True)
 
         multi = len(backends) > 1
         for backend in backends:
@@ -350,12 +411,14 @@ class ResultsFormatter:
                 table.add_column(col_rel, justify="right", no_wrap=False)
 
         # Add rows
-        for spec in sorted(by_spec.keys()):
+        for spec in specs_order:
             spec_results = by_spec[spec]
             times = {b: r.mean_time for b, r in spec_results.items() if r.success}
             best_time = min(times.values()) if times else 0.0
 
-            row = [spec]
+            batch_type = get_batch_type(spec)
+            batch_size = len(parse_batch_spec(spec))
+            row = [spec, batch_type, str(batch_size)]
             for backend in backends:
                 if backend in spec_results:
                     r = spec_results[backend]
@@ -486,10 +549,11 @@ def get_attention_scale(head_dim: int) -> float:
 
 def is_mla_backend(backend: str) -> bool:
     """
-    Check if backend is an MLA backend using the backend's is_mla() property.
+    Check if backend is an MLA backend using the AttentionBackendEnum.
 
     Args:
-        backend: Backend name (e.g., "CUTLASS_MLA", "FLASHINFER_MLA")
+        backend: Backend name matching AttentionBackendEnum exactly
+        (e.g., "FLASHMLA_SPARSE")
 
     Returns:
         True if the backend is an MLA backend, False otherwise
@@ -497,7 +561,8 @@ def is_mla_backend(backend: str) -> bool:
     from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
     try:
-        backend_class = AttentionBackendEnum[backend.upper()].get_class()
+        backend_enum = AttentionBackendEnum[backend]
+        backend_class = backend_enum.get_class()
         return backend_class.is_mla()
-    except (KeyError, ValueError, ImportError):
+    except (KeyError, ValueError, ImportError, AttributeError):
         return False

benchmarks/attention_benchmarks/configs/mla_decode.yaml

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

benchmarks/attention_benchmarks/configs/mla_mixed_batch.yaml

@@ -45,10 +45,10 @@ batch_specs:
   - "4q4k_60q1s4k"          # 4 prefill + 60 decode
 
 backends:
-  - cutlass_mla
-  - flashinfer_mla
-  - flashattn_mla   # Hopper only
-  - flashmla        # Hopper only
+  - CUTLASS_MLA
+  - FLASHINFER_MLA
+  - FLASH_ATTN_MLA   # Hopper only
+  - FLASHMLA         # Hopper only
 
 device: "cuda:0"
 repeats: 5

benchmarks/attention_benchmarks/configs/mla_prefill.yaml

@@ -0,0 +1,62 @@
+# MLA prefill-only benchmark configuration for sparse backends
+
+model:
+  name: "deepseek-v3"
+  num_layers: 60
+  num_q_heads: 128
+  num_kv_heads: 1
+  head_dim: 576
+  kv_lora_rank: 512
+  qk_nope_head_dim: 128
+  qk_rope_head_dim: 64
+  v_head_dim: 128
+  block_size: 128
+
+# Model parameter sweep: simulate tensor parallelism by varying num_q_heads
+# TP=1: 128 heads, TP=2: 64 heads, TP=4: 32 heads, TP=8: 16 heads
+model_parameter_sweep:
+  param_name: "num_q_heads"
+  values: [128, 64, 32, 16]
+  label_format: "{backend}_{value}h"
+
+batch_specs:
+  # Pure prefill
+  - "1q512"
+  - "1q1k"
+  - "1q2k"
+  - "1q4k"
+  - "1q8k"
+
+  # Batched pure prefill
+  - "2q512"
+  - "2q1k"
+  - "2q2k"
+  - "2q4k"
+  - "2q8k"
+  - "4q512"
+  - "4q1k"
+  - "4q2k"
+  - "4q4k"
+  - "4q8k"
+  - "8q512"
+  - "8q1k"
+  - "8q2k"
+  - "8q4k"
+  - "8q8k"
+
+  # Extend
+  - "1q512s4k"
+  - "1q512s8k"
+  - "1q1ks8k"
+  - "1q2ks8k"
+  - "1q2ks16k"
+  - "1q4ks16k"
+
+backends:
+  - FLASHMLA_SPARSE
+  - FLASHINFER_MLA_SPARSE
+
+device: "cuda:0"
+repeats: 10
+warmup_iters: 3
+profile_memory: true

benchmarks/attention_benchmarks/configs/reorder_threshold.yaml

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

benchmarks/attention_benchmarks/configs/speculative_decode.yaml

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

benchmarks/attention_benchmarks/configs/standard_attention.yaml

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

benchmarks/attention_benchmarks/mla_runner.py

@@ -8,14 +8,13 @@ This module provides helpers for running MLA backends without
 needing full VllmConfig integration.
 """
 
-import importlib
-
 import numpy as np
 import torch
 from batch_spec import parse_batch_spec
 from common import (
     BenchmarkResult,
     MockHfConfig,
+    MockIndexer,
     MockKVBProj,
     MockLayer,
     setup_mla_dims,
@@ -62,6 +61,7 @@ def create_minimal_vllm_config(
     block_size: int = 128,
     max_num_seqs: int = 256,
     mla_dims: dict | None = None,
+    index_topk: int | None = None,
 ) -> VllmConfig:
     """
     Create minimal VllmConfig for MLA benchmarks.
@@ -73,6 +73,8 @@ def create_minimal_vllm_config(
         max_num_seqs: Maximum number of sequences
         mla_dims: Optional custom MLA dimensions dict. If not provided, uses
                   setup_mla_dims(model_name)
+        index_topk: Optional topk value for sparse MLA backends. If provided,
+                    the config will include index_topk for sparse attention.
 
     Returns:
         VllmConfig for benchmarking
@@ -82,7 +84,7 @@ def create_minimal_vllm_config(
         mla_dims = setup_mla_dims(model_name)
 
     # Create mock HF config first (avoids downloading from HuggingFace)
-    mock_hf_config = MockHfConfig(mla_dims)
+    mock_hf_config = MockHfConfig(mla_dims, index_topk=index_topk)
 
     # Create a temporary minimal config.json to avoid HF downloads
     # This ensures consistent ModelConfig construction without network access
@@ -120,16 +122,12 @@ def create_minimal_vllm_config(
             seed=0,
             max_model_len=32768,
             quantization=None,
-            quantization_param_path=None,
             enforce_eager=False,
-            max_context_len_to_capture=None,
-            max_seq_len_to_capture=8192,
             max_logprobs=20,
             disable_sliding_window=False,
             skip_tokenizer_init=True,
             served_model_name=None,
             limit_mm_per_prompt=None,
-            use_async_output_proc=True,
             config_format="auto",
         )
     finally:
@@ -180,56 +178,65 @@ def create_minimal_vllm_config(
 # ============================================================================
 
 
-# Backend name to class name prefix mapping
-_BACKEND_NAME_MAP = {
-    "flashattn_mla": "FlashAttnMLA",
-    "flashmla": "FlashMLA",
-    "flashinfer_mla": "FlashInferMLA",
-    "cutlass_mla": "CutlassMLA",
-}
-
-# Special properties that differ from defaults
+# Backend-specific properties that can't be inferred from the backend class
+# Keys are AttentionBackendEnum names (uppercase)
 _BACKEND_PROPERTIES = {
-    "flashmla": {
+    "FLASHMLA": {
         "query_format": "concat",  # Single concatenated tensor (vs tuple)
-        "block_size": 64,  # FlashMLA uses fixed block size
     },
-    "flashinfer_mla": {
-        "block_size": 64,  # FlashInfer MLA only supports 32 or 64
+    "FLASHMLA_SPARSE": {
+        "query_format": "concat",  # Single concatenated tensor (vs tuple)
     },
 }
 
 
 def _get_backend_config(backend: str) -> dict:
     """
-    Get backend configuration using naming conventions.
+    Get backend configuration from AttentionBackendEnum.
 
-    All MLA backends follow the pattern:
-    - Module: vllm.v1.attention.backends.mla.{backend}
-    - Impl: {Name}Impl
-    - Metadata: {Name}Metadata (or MLACommonMetadata)
-    - DecodeMetadata: {Name}DecodeMetadata (or MLACommonDecodeMetadata)
-    - MetadataBuilder: {Name}MetadataBuilder
+    Uses the registry to get the backend class and extract configuration
+    from its methods (get_impl_cls, get_builder_cls, is_sparse, etc.).
+
+    Args:
+        backend: Backend name matching AttentionBackendEnum exactly
+        (e.g., "FLASHMLA_SPARSE")
+
+    Returns:
+        Dict with backend configuration
     """
-    if backend not in _BACKEND_NAME_MAP:
-        raise ValueError(f"Unknown backend: {backend}")
+    from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
-    name = _BACKEND_NAME_MAP[backend]
+    try:
+        backend_enum = AttentionBackendEnum[backend]
+        backend_class = backend_enum.get_class()
+    except (KeyError, ValueError) as e:
+        valid_backends = [e.name for e in AttentionBackendEnum if e.name != "CUSTOM"]
+        raise ValueError(
+            f"Unknown backend: {backend}. "
+            f"Valid MLA backends: {[b for b in valid_backends if 'MLA' in b]}"
+        ) from e
+
+    # Get block size from backend class
+    block_sizes = backend_class.get_supported_kernel_block_sizes()
+    # Use first supported block size (backends typically support one for MLA)
+    block_size = block_sizes[0] if block_sizes else None
+    if hasattr(block_size, "value"):
+        # Handle MultipleOf enum
+        block_size = None
+
+    # Check if sparse via class method if available
+    is_sparse = getattr(backend_class, "is_sparse", lambda: False)()
+
+    # Get properties that can't be inferred
     props = _BACKEND_PROPERTIES.get(backend, {})
 
-    # Check if backend uses common metadata (FlashInfer, CUTLASS)
-    uses_common = backend in ("flashinfer_mla", "cutlass_mla")
-
     return {
-        "module": f"vllm.v1.attention.backends.mla.{backend}",
-        "impl_class": f"{name}Impl",
-        "metadata_class": "MLACommonMetadata" if uses_common else f"{name}Metadata",
-        "decode_metadata_class": "MLACommonDecodeMetadata"
-        if uses_common
-        else f"{name}DecodeMetadata",
-        "builder_class": f"{name}MetadataBuilder",
+        "backend_class": backend_class,
+        "impl_class": backend_class.get_impl_cls(),
+        "builder_class": backend_class.get_builder_cls(),
         "query_format": props.get("query_format", "tuple"),
-        "block_size": props.get("block_size", None),
+        "block_size": block_size,
+        "is_sparse": is_sparse,
     }
 
 
@@ -447,22 +454,26 @@ def _create_backend_impl(
     mla_dims: dict,
     vllm_config: VllmConfig,
     device: torch.device,
+    max_num_tokens: int = 8192,
+    index_topk: int | None = None,
 ):
     """
     Create backend implementation instance.
 
     Args:
-        backend_cfg: Backend configuration dict
+        backend_cfg: Backend configuration dict from _get_backend_config()
         mla_dims: MLA dimension configuration
         vllm_config: VllmConfig instance
         device: Target device
+        max_num_tokens: Maximum number of tokens for sparse indexer buffer
+        index_topk: Topk value for sparse MLA backends
 
     Returns:
-        Tuple of (impl, layer, builder_instance)
+        Tuple of (impl, layer, builder_instance, indexer)
     """
-    # Import backend classes
-    backend_module = importlib.import_module(backend_cfg["module"])
-    impl_class = getattr(backend_module, backend_cfg["impl_class"])
+    # Get classes from backend config (already resolved by _get_backend_config)
+    impl_class = backend_cfg["impl_class"]
+    builder_class = backend_cfg["builder_class"]
 
     # Calculate scale
     scale = 1.0 / np.sqrt(mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"])
@@ -474,26 +485,44 @@ def _create_backend_impl(
         v_head_dim=mla_dims["v_head_dim"],
     )
 
+    # Create indexer for sparse backends
+    indexer = None
+    if backend_cfg.get("is_sparse", False):
+        if index_topk is None:
+            index_topk = 2048  # Default topk for sparse MLA
+        indexer = MockIndexer(
+            max_num_tokens=max_num_tokens,
+            topk_tokens=index_topk,
+            device=device,
+        )
+
+    # Build impl kwargs
+    impl_kwargs = {
+        "num_heads": mla_dims["num_q_heads"],
+        "head_size": mla_dims["head_dim"],
+        "scale": scale,
+        "num_kv_heads": mla_dims["num_kv_heads"],
+        "alibi_slopes": None,
+        "sliding_window": None,
+        "kv_cache_dtype": "auto",
+        "logits_soft_cap": None,
+        "attn_type": "decoder",
+        "kv_sharing_target_layer_name": None,
+        "q_lora_rank": None,
+        "kv_lora_rank": mla_dims["kv_lora_rank"],
+        "qk_nope_head_dim": mla_dims["qk_nope_head_dim"],
+        "qk_rope_head_dim": mla_dims["qk_rope_head_dim"],
+        "qk_head_dim": mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"],
+        "v_head_dim": mla_dims["v_head_dim"],
+        "kv_b_proj": mock_kv_b_proj,
+    }
+
+    # Add indexer for sparse backends
+    if indexer is not None:
+        impl_kwargs["indexer"] = indexer
+
     # Create impl
-    impl = impl_class(
-        num_heads=mla_dims["num_q_heads"],
-        head_size=mla_dims["head_dim"],
-        scale=scale,
-        num_kv_heads=mla_dims["num_kv_heads"],
-        alibi_slopes=None,
-        sliding_window=None,
-        kv_cache_dtype="auto",
-        logits_soft_cap=None,
-        attn_type="decoder",
-        kv_sharing_target_layer_name=None,
-        q_lora_rank=None,
-        kv_lora_rank=mla_dims["kv_lora_rank"],
-        qk_nope_head_dim=mla_dims["qk_nope_head_dim"],
-        qk_rope_head_dim=mla_dims["qk_rope_head_dim"],
-        qk_head_dim=mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"],
-        v_head_dim=mla_dims["v_head_dim"],
-        kv_b_proj=mock_kv_b_proj,
-    )
+    impl = impl_class(**impl_kwargs)
 
     # Initialize DCP attributes
     if not hasattr(impl, "dcp_world_size") or impl.dcp_world_size in (None, -1):
@@ -515,9 +544,7 @@ def _create_backend_impl(
 
     # Create builder instance if needed
     builder_instance = None
-    if backend_cfg["builder_class"]:
-        builder_class = getattr(backend_module, backend_cfg["builder_class"])
-
+    if builder_class:
         # Populate static_forward_context so builder can find the layer
         # MockLayer inherits from AttentionLayerBase, so isinstance checks pass
         vllm_config.compilation_config.static_forward_context = {"placeholder": layer}
@@ -529,7 +556,7 @@ def _create_backend_impl(
             device=device,
         )
 
-    return impl, layer, builder_instance
+    return impl, layer, builder_instance, indexer
 
 
 # ============================================================================
@@ -594,6 +621,7 @@ def _run_single_benchmark(
     backend_cfg: dict,
     mla_dims: dict,
     device: torch.device,
+    indexer=None,
 ) -> BenchmarkResult:
     """
     Run a single benchmark iteration.
@@ -606,6 +634,7 @@ def _run_single_benchmark(
         backend_cfg: Backend configuration dict
         mla_dims: MLA dimension configuration
         device: Target device
+        indexer: Optional MockIndexer for sparse backends
 
     Returns:
         BenchmarkResult with timing statistics
@@ -613,7 +642,9 @@ def _run_single_benchmark(
     # Parse batch spec
     requests = parse_batch_spec(config.batch_spec)
     q_lens = [r.q_len for r in requests]
+    kv_lens = [r.kv_len for r in requests]
     total_q = sum(q_lens)
+    max_kv_len = max(kv_lens)
 
     # Determine block size
     block_size = backend_cfg["block_size"] or config.block_size
@@ -641,8 +672,16 @@ def _run_single_benchmark(
         torch.bfloat16,
     )
 
-    # Determine which forward method to use based on metadata
-    if metadata.decode is not None:
+    # Fill indexer with random indices for sparse backends
+    is_sparse = backend_cfg.get("is_sparse", False)
+    if is_sparse and indexer is not None:
+        indexer.fill_random_indices(total_q, max_kv_len)
+
+    # Determine which forward method to use
+    if is_sparse:
+        # Sparse backends use forward_mqa
+        forward_fn = lambda: impl.forward_mqa(decode_inputs, kv_cache, metadata, layer)
+    elif metadata.decode is not None:
         forward_fn = lambda: impl._forward_decode(
             decode_inputs, kv_cache, metadata, layer
         )
@@ -693,11 +732,13 @@ def _run_single_benchmark(
 def _run_mla_benchmark_batched(
     backend: str,
     configs_with_params: list[tuple],  # [(config, threshold, num_splits), ...]
+    index_topk: int = 2048,
 ) -> list[BenchmarkResult]:
     """
     Unified batched MLA benchmark runner for all backends.
 
-    Works for: flashattn_mla, flashmla, flashinfer_mla, cutlass_mla
+    Works for: flashattn_mla, flashmla, flashinfer_mla, cutlass_mla,
+               flashinfer_mla_sparse, flashmla_sparse
 
     This function reuses backend initialization across multiple benchmarks
     to avoid setup/teardown overhead.
@@ -707,6 +748,7 @@ def _run_mla_benchmark_batched(
         configs_with_params: List of (config, threshold, num_splits) tuples
             - threshold: reorder_batch_threshold (FlashAttn/FlashMLA only)
             - num_splits: num_kv_splits (CUTLASS only)
+        index_topk: Topk value for sparse MLA backends (default 2048)
 
     Returns:
         List of BenchmarkResult objects
@@ -730,19 +772,27 @@ def _run_mla_benchmark_batched(
     if mla_dims is None:
         mla_dims = setup_mla_dims("deepseek-v3")
 
+    # Determine if this is a sparse backend
+    is_sparse = backend_cfg.get("is_sparse", False)
+
     # Create and set vLLM config for MLA (reused across all benchmarks)
     vllm_config = create_minimal_vllm_config(
         model_name="deepseek-v3",  # Used only for model path
         block_size=block_size,
         mla_dims=mla_dims,  # Use custom dims from config or default
+        index_topk=index_topk if is_sparse else None,
     )
 
     results = []
 
     with set_current_vllm_config(vllm_config):
-        # Create backend impl, layer, and builder (reused across benchmarks)
-        impl, layer, builder_instance = _create_backend_impl(
-            backend_cfg, mla_dims, vllm_config, device
+        # Create backend impl, layer, builder, and indexer (reused across benchmarks)
+        impl, layer, builder_instance, indexer = _create_backend_impl(
+            backend_cfg,
+            mla_dims,
+            vllm_config,
+            device,
+            index_topk=index_topk if is_sparse else None,
         )
 
         # Run each benchmark with the shared impl
@@ -768,6 +818,7 @@ def _run_mla_benchmark_batched(
                     backend_cfg,
                     mla_dims,
                     device,
+                    indexer=indexer,
                 )
                 results.append(result)
 
@@ -793,20 +844,24 @@ def run_mla_benchmark(
     config,
     reorder_batch_threshold: int | None = None,
     num_kv_splits: int | None = None,
+    index_topk: int = 2048,
 ) -> BenchmarkResult | list[BenchmarkResult]:
     """
     Unified MLA benchmark runner for all backends.
 
-    Works for: flashattn_mla, flashmla, flashinfer_mla, cutlass_mla
+    Works for: flashattn_mla, flashmla, flashinfer_mla, cutlass_mla,
+               flashinfer_mla_sparse, flashmla_sparse
 
     Always uses batched execution internally for optimal performance.
 
     Args:
-        backend: Backend name (flashattn_mla, flashmla, flashinfer_mla, cutlass_mla)
+        backend: Backend name (flashattn_mla, flashmla, flashinfer_mla, cutlass_mla,
+                 flashinfer_mla_sparse, flashmla_sparse)
         config: BenchmarkConfig or list of (BenchmarkConfig, param) tuples
         reorder_batch_threshold: Threshold override for FlashAttn/FlashMLA
                                  (single config mode only)
         num_kv_splits: Number of KV splits for CUTLASS (single config mode only)
+        index_topk: Topk value for sparse MLA backends (default 2048)
 
     Returns:
         BenchmarkResult (single mode) or list of BenchmarkResult (batched mode)
@@ -816,9 +871,9 @@ def run_mla_benchmark(
         # Already in batched format
         if len(config) > 0 and isinstance(config[0], tuple):
             # Format: [(cfg, param), ...] where param is threshold or num_splits
-            if backend in ("flashattn_mla", "flashmla"):
+            if backend in ("flashattn_mla", "flashmla", "flashmla_sparse"):
                 configs_with_params = [(cfg, param, None) for cfg, param in config]
-            else:  # cutlass_mla or flashinfer_mla
+            else:  # cutlass_mla, flashinfer_mla, or sparse backends
                 configs_with_params = [(cfg, None, param) for cfg, param in config]
         else:
             # Format: [cfg, ...] - just configs
@@ -830,7 +885,7 @@ def run_mla_benchmark(
         return_single = True
 
     # Use unified batched execution
-    results = _run_mla_benchmark_batched(backend, configs_with_params)
+    results = _run_mla_benchmark_batched(backend, configs_with_params, index_topk)
 
     # Return single result or list based on input
     return results[0] if return_single else results

benchmarks/attention_benchmarks/runner.py

@@ -8,7 +8,9 @@ This module provides helpers for running standard attention backends
 (FlashAttention, Triton, FlashInfer) with real vLLM integration.
 """
 
+import logging
 import types
+from contextlib import contextmanager
 
 import numpy as np
 import torch
@@ -24,8 +26,13 @@ from vllm.config import (
     ParallelConfig,
     SchedulerConfig,
     VllmConfig,
+    set_current_vllm_config,
+)
+from vllm.v1.attention.backends.utils import (
+    CommonAttentionMetadata,
+    get_kv_cache_layout,
+    set_kv_cache_layout,
 )
-from vllm.v1.attention.backends.utils import CommonAttentionMetadata
 from vllm.v1.kv_cache_interface import FullAttentionSpec
 
 # ============================================================================
@@ -33,37 +40,41 @@ from vllm.v1.kv_cache_interface import FullAttentionSpec
 # ============================================================================
 
 
-_BACKEND_CONFIG = {
-    "flash": {
-        "module": "vllm.v1.attention.backends.flash_attn",
-        "backend_class": "FlashAttentionBackend",
-        "dtype": torch.float16,
-        "cache_layout": "standard",
-        # ^ [2, num_blocks, block_size, num_kv_heads, head_dim]
-    },
-    "triton": {
-        "module": "vllm.v1.attention.backends.triton_attn",
-        "backend_class": "TritonAttentionBackend",
-        "dtype": torch.float32,
-        "cache_layout": "standard",
-    },
-    "flashinfer": {
-        "module": "vllm.v1.attention.backends.flashinfer",
-        "backend_class": "FlashInferBackend",
-        "dtype": torch.float16,
-        "cache_layout": "flashinfer",
-        # ^ [num_blocks, 2, block_size, num_kv_heads, head_dim]
-    },
-}
-
-
 def _get_backend_config(backend: str) -> dict:
-    if backend not in _BACKEND_CONFIG:
+    """
+    Get backend configuration from AttentionBackendEnum.
+
+    Args:
+        backend: Backend name matching AttentionBackendEnum exactly
+                 (e.g., "FLASH_ATTN", "TRITON_ATTN", "FLASHINFER")
+
+    Returns:
+        Dict with backend_class
+    """
+    from vllm.v1.attention.backends.registry import AttentionBackendEnum
+
+    try:
+        backend_enum = AttentionBackendEnum[backend]
+        backend_class = backend_enum.get_class()
+    except (KeyError, ValueError) as e:
+        valid_backends = [b.name for b in AttentionBackendEnum if b.name != "CUSTOM"]
         raise ValueError(
-            f"Unknown backend: {backend}. "
-            f"Available: {', '.join(_BACKEND_CONFIG.keys())}"
-        )
-    return _BACKEND_CONFIG[backend]
+            f"Unknown backend: {backend}. Valid backends: {valid_backends}"
+        ) from e
+
+    return {"backend_class": backend_class}
+
+
+@contextmanager
+def log_warnings_and_errors_only():
+    """Temporarily set vLLM logger to WARNING level."""
+    logger = logging.getLogger("vllm")
+    old_level = logger.level
+    logger.setLevel(logging.WARNING)
+    try:
+        yield
+    finally:
+        logger.setLevel(old_level)
 
 
 # ============================================================================
@@ -88,11 +99,7 @@ def _build_common_attn_metadata(
     query_start_loc_cpu = query_start_loc.cpu()
 
     seq_lens = torch.tensor(kv_lens, dtype=torch.int32, device=device)
-    seq_lens_cpu = seq_lens.cpu()
-    max_seq_len = int(seq_lens_cpu.max())
-
-    context_lens = [kv - q for kv, q in zip(kv_lens, q_lens)]
-    num_computed_tokens_cpu = torch.tensor(context_lens, dtype=torch.int32)
+    max_seq_len = int(seq_lens.max().item())
 
     max_blocks = (max(kv_lens) + block_size - 1) // block_size
     num_blocks = batch_size * max_blocks
@@ -107,8 +114,6 @@ def _build_common_attn_metadata(
         query_start_loc=query_start_loc,
         query_start_loc_cpu=query_start_loc_cpu,
         seq_lens=seq_lens,
-        seq_lens_cpu=seq_lens_cpu,
-        num_computed_tokens_cpu=num_computed_tokens_cpu,
         num_reqs=batch_size,
         num_actual_tokens=total_tokens,
         max_query_len=max_query_len,
@@ -121,7 +126,6 @@ def _build_common_attn_metadata(
 
 def _create_vllm_config(
     config: BenchmarkConfig,
-    dtype: torch.dtype,
     max_num_blocks: int,
 ) -> VllmConfig:
     """Create a VllmConfig for benchmarking with mock model methods."""
@@ -129,7 +133,7 @@ def _create_vllm_config(
         model="meta-llama/Meta-Llama-3-8B",
         tokenizer="meta-llama/Meta-Llama-3-8B",
         trust_remote_code=False,
-        dtype=dtype,
+        dtype="auto",  # Use model's native dtype
         seed=0,
         max_model_len=1024,
     )
@@ -198,15 +202,12 @@ def _create_backend_impl(
     backend_cfg: dict,
     config: BenchmarkConfig,
     device: torch.device,
+    dtype: torch.dtype,
 ):
     """Create backend implementation instance."""
-    import importlib
-
-    backend_module = importlib.import_module(backend_cfg["module"])
-    backend_class = getattr(backend_module, backend_cfg["backend_class"])
+    backend_class = backend_cfg["backend_class"]
 
     scale = get_attention_scale(config.head_dim)
-    dtype = backend_cfg["dtype"]
 
     impl = backend_class.get_impl_cls()(
         num_heads=config.num_q_heads,
@@ -227,7 +228,7 @@ def _create_backend_impl(
 
     layer = MockLayer(device, kv_cache_spec=kv_cache_spec)
 
-    return backend_class, impl, layer, dtype
+    return backend_class, impl, layer
 
 
 def _create_metadata_builder(
@@ -235,11 +236,44 @@ def _create_metadata_builder(
     kv_cache_spec: FullAttentionSpec,
     vllm_config: VllmConfig,
     device: torch.device,
+    backend_name: str = "",
 ):
     """Create metadata builder instance."""
-    return backend_class.get_builder_cls()(
+    layer_names = ["layer_0"]
+    builder_cls = backend_class.get_builder_cls()
+
+    # Flashinfer needs get_per_layer_parameters mocked since we don't have
+    # real model layers registered
+    if backend_name == "FLASHINFER":
+        import unittest.mock
+
+        from vllm.v1.attention.backends.utils import PerLayerParameters
+
+        def mock_get_per_layer_parameters(vllm_config, layer_names, impl_cls):
+            head_size = vllm_config.model_config.get_head_size()
+            return {
+                layer_name: PerLayerParameters(
+                    window_left=-1,  # No sliding window
+                    logits_soft_cap=0.0,  # No soft cap
+                    sm_scale=1.0 / (head_size**0.5),  # Standard scale
+                )
+                for layer_name in layer_names
+            }
+
+        with unittest.mock.patch(
+            "vllm.v1.attention.backends.flashinfer.get_per_layer_parameters",
+            mock_get_per_layer_parameters,
+        ):
+            return builder_cls(
+                kv_cache_spec=kv_cache_spec,
+                layer_names=layer_names,
+                vllm_config=vllm_config,
+                device=device,
+            )
+
+    return builder_cls(
         kv_cache_spec=kv_cache_spec,
-        layer_names=["layer_0"],
+        layer_names=layer_names,
         vllm_config=vllm_config,
         device=device,
     )
@@ -281,39 +315,44 @@ def _create_input_tensors(
 def _create_kv_cache(
     config: BenchmarkConfig,
     max_num_blocks: int,
-    cache_layout: str,
+    backend_class,
     device: torch.device,
     dtype: torch.dtype,
 ) -> list:
-    """Create KV cache tensors for all layers."""
-    if cache_layout == "flashinfer":
-        # FlashInfer layout: [num_blocks, 2, block_size, num_kv_heads, head_dim]
-        cache_list = [
-            torch.zeros(
-                max_num_blocks,
-                2,
-                config.block_size,
-                config.num_kv_heads,
-                config.head_dim,
-                device=device,
-                dtype=dtype,
-            )
-            for _ in range(config.num_layers)
-        ]
-    else:
-        # Standard layout: [2, num_blocks, block_size, num_kv_heads, head_dim]
-        cache_list = [
-            torch.zeros(
-                2,
-                max_num_blocks,
-                config.block_size,
-                config.num_kv_heads,
-                config.head_dim,
-                device=device,
-                dtype=dtype,
-            )
-            for _ in range(config.num_layers)
-        ]
+    """Create KV cache tensors for all layers using the backend's methods.
+
+    Uses the backend's get_kv_cache_shape() and get_kv_cache_stride_order()
+    to create the cache with the correct shape and memory layout.
+    """
+    # Get the logical shape from the backend
+    cache_shape = backend_class.get_kv_cache_shape(
+        num_blocks=max_num_blocks,
+        block_size=config.block_size,
+        num_kv_heads=config.num_kv_heads,
+        head_size=config.head_dim,
+    )
+
+    # Get the stride order for custom memory layout
+    try:
+        stride_order = backend_class.get_kv_cache_stride_order()
+        assert len(stride_order) == len(cache_shape)
+    except (AttributeError, NotImplementedError):
+        stride_order = tuple(range(len(cache_shape)))
+
+    # Permute shape to physical layout order
+    physical_shape = tuple(cache_shape[i] for i in stride_order)
+
+    # Compute inverse permutation to get back to logical view
+    inv_order = [stride_order.index(i) for i in range(len(stride_order))]
+
+    cache_list = []
+    for _ in range(config.num_layers):
+        # Allocate in physical layout order (contiguous in memory)
+        cache = torch.zeros(*physical_shape, device=device, dtype=dtype)
+        # Permute to logical view
+        cache = cache.permute(*inv_order)
+        cache_list.append(cache)
+
     return cache_list
 
 
@@ -396,7 +435,7 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
     """
     Run standard attention benchmark with real kernels.
 
-    Supports: flash, triton, flashinfer
+    Supports: FLASH_ATTN, TRITON_ATTN, FLASHINFER
 
     Args:
         config: Benchmark configuration
@@ -411,60 +450,79 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
 
     requests = parse_batch_spec(config.batch_spec)
 
-    if config.backend == "flashinfer":
+    if config.backend == "FLASHINFER":
         requests = reorder_for_flashinfer(requests)
 
     q_lens = [r.q_len for r in requests]
     kv_lens = [r.kv_len for r in requests]
     total_q = sum(q_lens)
     max_kv = max(kv_lens)
+    batch_size = len(q_lens)
 
-    max_num_blocks = (max_kv + config.block_size - 1) // config.block_size
+    # Calculate total blocks needed: batch_size * max_blocks_per_request
+    max_blocks_per_request = (max_kv + config.block_size - 1) // config.block_size
+    max_num_blocks = batch_size * max_blocks_per_request
 
-    backend_class, impl, layer, dtype = _create_backend_impl(
-        backend_cfg, config, device
-    )
+    # Suppress vLLM logs during setup to reduce spam
+    with log_warnings_and_errors_only():
+        # Create vllm_config first - uses model's native dtype via "auto"
+        vllm_config = _create_vllm_config(config, max_num_blocks)
+        dtype = vllm_config.model_config.dtype
 
-    common_metadata = _build_common_attn_metadata(
-        q_lens, kv_lens, config.block_size, device
-    )
+        # Wrap everything in set_current_vllm_config context
+        # This is required for backends like flashinfer that need global config
+        with set_current_vllm_config(vllm_config):
+            backend_class, impl, layer = _create_backend_impl(
+                backend_cfg, config, device, dtype
+            )
 
-    kv_cache_spec = FullAttentionSpec(
-        block_size=config.block_size,
-        num_kv_heads=config.num_kv_heads,
-        head_size=config.head_dim,
-        dtype=dtype,
-    )
+            # Set KV cache layout if the backend requires a specific one
+            # (e.g., FlashInfer requires HND on SM100/Blackwell for TRTLLM attention)
+            required_layout = backend_class.get_required_kv_cache_layout()
+            if required_layout is not None:
+                set_kv_cache_layout(required_layout)
+                get_kv_cache_layout.cache_clear()
 
-    vllm_config = _create_vllm_config(config, dtype, max_num_blocks)
+            common_metadata = _build_common_attn_metadata(
+                q_lens, kv_lens, config.block_size, device
+            )
 
-    builder = _create_metadata_builder(
-        backend_class, kv_cache_spec, vllm_config, device
-    )
+            kv_cache_spec = FullAttentionSpec(
+                block_size=config.block_size,
+                num_kv_heads=config.num_kv_heads,
+                head_size=config.head_dim,
+                dtype=dtype,
+            )
 
-    attn_metadata = builder.build(
-        common_prefix_len=0,
-        common_attn_metadata=common_metadata,
-    )
+            builder = _create_metadata_builder(
+                backend_class, kv_cache_spec, vllm_config, device, config.backend
+            )
 
-    q_list, k_list, v_list = _create_input_tensors(config, total_q, device, dtype)
+            attn_metadata = builder.build(
+                common_prefix_len=0,
+                common_attn_metadata=common_metadata,
+            )
 
-    cache_list = _create_kv_cache(
-        config, max_num_blocks, backend_cfg["cache_layout"], device, dtype
-    )
+            q_list, k_list, v_list = _create_input_tensors(
+                config, total_q, device, dtype
+            )
 
-    times, mem_stats = _run_single_benchmark(
-        config,
-        impl,
-        layer,
-        q_list,
-        k_list,
-        v_list,
-        cache_list,
-        attn_metadata,
-        device,
-        dtype,
-    )
+            cache_list = _create_kv_cache(
+                config, max_num_blocks, backend_class, device, dtype
+            )
+
+            times, mem_stats = _run_single_benchmark(
+                config,
+                impl,
+                layer,
+                q_list,
+                k_list,
+                v_list,
+                cache_list,
+                attn_metadata,
+                device,
+                dtype,
+            )
 
     mean_time = np.mean(times)
     throughput = total_q / mean_time if mean_time > 0 else 0

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -11,6 +11,7 @@ import torch
 import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from tests.kernels.moe.utils import make_dummy_moe_config
 from vllm import _custom_ops as ops
+from vllm.model_executor.layers.fused_moe.activation import MoEActivation
 from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
 from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassExpertsFp8
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
@@ -161,7 +162,7 @@ def bench_run(
                 w2_fp8q_cutlass,
                 topk_weights,
                 topk_ids,
-                activation="silu",
+                activation=MoEActivation.SILU,
                 global_num_experts=num_experts,
             )
     torch.cuda.synchronize()

benchmarks/kernels/benchmark_fused_collective.py

@@ -5,7 +5,7 @@
 Benchmark for FlashInfer fused collective operations vs standard operations.
 
 This benchmark compares:
-1. FlashInfer's trtllm_allreduce_fusion (fused allreduce + rmsnorm + optional quant)
+1. FlashInfer's allreduce_fusion (fused allreduce + rmsnorm + optional quant)
 2. Standard tensor_model_parallel_all_reduce + separate rmsnorm/quant operations
 
 Usage with torchrun:
@@ -24,7 +24,6 @@ import torch.distributed as dist  # type: ignore
 
 from vllm.config.vllm import CompilationConfig, VllmConfig, set_current_vllm_config
 from vllm.distributed import (
-    get_tp_group,
     tensor_model_parallel_all_reduce,
 )
 from vllm.distributed.parallel_state import (
@@ -52,11 +51,12 @@ logger = init_logger(__name__)
 try:
     import flashinfer.comm as flashinfer_comm  # type: ignore
 
-    if not hasattr(flashinfer_comm, "trtllm_allreduce_fusion"):
+    if not (
+        hasattr(flashinfer_comm, "allreduce_fusion")
+        and hasattr(flashinfer_comm, "create_allreduce_fusion_workspace")
+    ):
         flashinfer_comm = None
-        logger.warning(
-            "FlashInfer comm module found but missing trtllm_allreduce_fusion"
-        )
+        logger.warning("FlashInfer comm module found but missing allreduce_fusion API")
 except ImportError:
     flashinfer_comm = None
     logger.warning("FlashInfer not found, only benchmarking standard operations")
@@ -75,7 +75,7 @@ _FI_MAX_SIZES = {
 }
 
 # Global workspace tensor for FlashInfer
-_FI_WORKSPACE_TENSOR = None
+_FI_WORKSPACE = None
 
 
 def setup_flashinfer_workspace(
@@ -83,10 +83,10 @@ def setup_flashinfer_workspace(
     rank: int,
     hidden_dim: int,
     max_token_num: int,
-    use_fp32_lamport: bool = False,
+    dtype: torch.dtype,
 ):
     """Setup FlashInfer workspace for fused allreduce operations."""
-    global _FI_WORKSPACE_TENSOR
+    global _FI_WORKSPACE
 
     if flashinfer_comm is None:
         return None, None
@@ -96,33 +96,29 @@ def setup_flashinfer_workspace(
         return None, None
 
     try:
-        # Create IPC workspace
-        ipc_handles, workspace_tensor = (
-            flashinfer_comm.trtllm_create_ipc_workspace_for_all_reduce_fusion(
-                tp_rank=rank,
-                tp_size=world_size,
-                max_token_num=max_token_num,
-                hidden_dim=hidden_dim,
-                group=get_tp_group().device_group,
-                use_fp32_lamport=use_fp32_lamport,
-            )
+        workspace = flashinfer_comm.create_allreduce_fusion_workspace(
+            backend="trtllm",
+            world_size=world_size,
+            rank=rank,
+            max_token_num=max_token_num,
+            hidden_dim=hidden_dim,
+            dtype=dtype,
         )
 
-        _FI_WORKSPACE_TENSOR = workspace_tensor
-        return ipc_handles, workspace_tensor
+        _FI_WORKSPACE = workspace
+        return workspace
     except Exception as e:
         logger.error("Failed to setup FlashInfer workspace: %s", e)
-        return None, None
+        return None
 
 
-def cleanup_flashinfer_workspace(ipc_handles):
+def cleanup_flashinfer_workspace(workspace):
     """Cleanup FlashInfer workspace."""
-    if flashinfer_comm is None or ipc_handles is None:
+    if flashinfer_comm is None or workspace is None:
         return
 
     try:
-        group = get_tp_group().device_group
-        flashinfer_comm.trtllm_destroy_ipc_workspace_for_all_reduce(ipc_handles, group)
+        workspace.destroy()
     except Exception as e:
         logger.error("Failed to cleanup FlashInfer workspace: %s", e)
 
@@ -132,25 +128,15 @@ class FlashInferFusedAllReduceParams:
 
     def __init__(
         self,
-        rank: int,
-        world_size: int,
-        use_fp32_lamport: bool = False,
         max_token_num: int = 1024,
     ):
-        self.rank = rank
-        self.world_size = world_size
-        self.use_fp32_lamport = use_fp32_lamport
-        self.trigger_completion_at_end = True
         self.launch_with_pdl = True
         self.fp32_acc = True
         self.max_token_num = max_token_num
 
     def get_trtllm_fused_allreduce_kwargs(self):
         return {
-            "world_rank": self.rank,
-            "world_size": self.world_size,
             "launch_with_pdl": self.launch_with_pdl,
-            "trigger_completion_at_end": self.trigger_completion_at_end,
             "fp32_acc": self.fp32_acc,
         }
 
@@ -165,7 +151,7 @@ def flashinfer_fused_allreduce_rmsnorm(
     norm_out: torch.Tensor | None = None,
 ):
     """FlashInfer fused allreduce + rmsnorm operation."""
-    if flashinfer_comm is None or _FI_WORKSPACE_TENSOR is None:
+    if flashinfer_comm is None or _FI_WORKSPACE is None:
         raise RuntimeError("FlashInfer not available or workspace not initialized")
 
     if norm_out is None:
@@ -174,18 +160,15 @@ def flashinfer_fused_allreduce_rmsnorm(
     else:
         residual_out = input_tensor
 
-    flashinfer_comm.trtllm_allreduce_fusion(
-        allreduce_in=input_tensor,
-        token_num=input_tensor.shape[0],
+    flashinfer_comm.allreduce_fusion(
+        input=input_tensor,
+        workspace=_FI_WORKSPACE,
+        pattern=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNorm,
         residual_in=residual,
         residual_out=residual_out,
         norm_out=norm_out,
         rms_gamma=rms_gamma,
         rms_eps=rms_eps,
-        hidden_dim=input_tensor.shape[-1],
-        workspace_ptrs=_FI_WORKSPACE_TENSOR,
-        pattern_code=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNorm,
-        allreduce_out=None,
         quant_out=None,
         scale_out=None,
         layout_code=flashinfer_comm.QuantizationSFLayout.SWIZZLED_128x4,
@@ -207,7 +190,7 @@ def flashinfer_fused_allreduce_rmsnorm_fp8_quant(
     quant_out: torch.Tensor | None = None,
 ):
     """FlashInfer fused allreduce + rmsnorm + FP8 quantization."""
-    if flashinfer_comm is None or _FI_WORKSPACE_TENSOR is None:
+    if flashinfer_comm is None or _FI_WORKSPACE is None:
         raise RuntimeError("FlashInfer not available or workspace not initialized")
 
     if norm_out is None:
@@ -216,18 +199,15 @@ def flashinfer_fused_allreduce_rmsnorm_fp8_quant(
     else:
         residual_out = input_tensor
 
-    flashinfer_comm.trtllm_allreduce_fusion(
-        allreduce_in=input_tensor,
-        token_num=input_tensor.shape[0],
+    flashinfer_comm.allreduce_fusion(
+        input=input_tensor,
+        workspace=_FI_WORKSPACE,
+        pattern=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNormFP8Quant,
         residual_in=residual,
         residual_out=residual_out,
         norm_out=norm_out,
         rms_gamma=rms_gamma,
         rms_eps=rms_eps,
-        hidden_dim=input_tensor.shape[-1],
-        workspace_ptrs=_FI_WORKSPACE_TENSOR,
-        pattern_code=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNormFP8Quant,
-        allreduce_out=None,
         quant_out=quant_out,
         scale_out=None,
         layout_code=flashinfer_comm.QuantizationSFLayout.SWIZZLED_128x4,
@@ -250,7 +230,7 @@ def flashinfer_fused_allreduce_rmsnorm_fp4_quant(
     norm_out: torch.Tensor | None = None,
 ):
     """FlashInfer fused allreduce + rmsnorm + FP4 quantization."""
-    if flashinfer_comm is None or _FI_WORKSPACE_TENSOR is None:
+    if flashinfer_comm is None or _FI_WORKSPACE is None:
         raise RuntimeError("FlashInfer not available or workspace not initialized")
 
     if norm_out is None:
@@ -259,18 +239,15 @@ def flashinfer_fused_allreduce_rmsnorm_fp4_quant(
     else:
         residual_out = input_tensor
 
-    flashinfer_comm.trtllm_allreduce_fusion(
-        allreduce_in=input_tensor,
-        token_num=input_tensor.shape[0],
+    flashinfer_comm.allreduce_fusion(
+        input=input_tensor,
+        workspace=_FI_WORKSPACE,
+        pattern=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNormFP4Quant,
         residual_in=residual,
         residual_out=residual_out,
         norm_out=norm_out,
         rms_gamma=rms_gamma,
         rms_eps=rms_eps,
-        hidden_dim=input_tensor.shape[-1],
-        workspace_ptrs=_FI_WORKSPACE_TENSOR,
-        pattern_code=flashinfer_comm.AllReduceFusionPattern.kARResidualRMSNormFP4Quant,
-        allreduce_out=None,
         quant_out=quant_out,
         scale_out=output_scale,
         layout_code=flashinfer_comm.QuantizationSFLayout.SWIZZLED_128x4,
@@ -1040,23 +1017,31 @@ def main():
     configs = list(itertools.product(args.num_tokens, dtypes, residual_options))
 
     # Setup FlashInfer workspace if available
-    ipc_handles = None
+    workspace = None
     allreduce_params = None
 
     if flashinfer_comm is not None:
         # Use the largest hidden dimension for workspace setup
+        max_element_size = max(torch.finfo(dt).bits // 8 for dt in dtypes)
+        workspace_dtype = (
+            torch.float32
+            if max_element_size == 4
+            else (torch.bfloat16 if torch.bfloat16 in dtypes else torch.float16)
+        )
         max_num_token = _FI_MAX_SIZES.get(world_size) // (
-            args.hidden_dim * world_size * 2
+            args.hidden_dim * max_element_size
         )
 
-        ipc_handles, workspace_tensor = setup_flashinfer_workspace(
-            world_size, rank, args.hidden_dim, max_num_token
+        workspace = setup_flashinfer_workspace(
+            world_size,
+            rank,
+            args.hidden_dim,
+            max_num_token,
+            dtype=workspace_dtype,
         )
 
-        if workspace_tensor is not None:
+        if workspace is not None:
             allreduce_params = FlashInferFusedAllReduceParams(
-                rank=rank,
-                world_size=world_size,
                 max_token_num=max_num_token,
             )
 
@@ -1119,8 +1104,8 @@ def main():
 
     finally:
         # Cleanup
-        if ipc_handles is not None:
-            cleanup_flashinfer_workspace(ipc_handles)
+        if workspace is not None:
+            cleanup_flashinfer_workspace(workspace)
 
         dist.barrier()
 

benchmarks/kernels/benchmark_moe.py

@@ -16,6 +16,7 @@ import torch
 from ray.experimental.tqdm_ray import tqdm
 
 from vllm.model_executor.layers.fused_moe import fused_topk
+from vllm.model_executor.layers.fused_moe.activation import MoEActivation
 from vllm.model_executor.layers.fused_moe.config import (
     FusedMoEConfig,
     FusedMoEParallelConfig,
@@ -99,13 +100,38 @@ def benchmark_config(
     dtype: torch.dtype,
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
+    use_int4_w4a16: bool = False,
     num_iters: int = 100,
     block_quant_shape: list[int] = None,
     use_deep_gemm: bool = False,
 ) -> float:
     init_dtype = torch.float16 if use_fp8_w8a8 else dtype
     x = torch.randn(num_tokens, hidden_size, dtype=dtype)
-    if use_int8_w8a16:
+    if use_int4_w4a16:
+        # Int4 packed weights: 2 int4 values per uint8 byte
+        # K dimension is packed (halved)
+        intermediate_size = shard_intermediate_size // 2  # after silu_and_mul
+        w1 = torch.randint(
+            0,
+            255,
+            (
+                num_experts,
+                shard_intermediate_size,
+                hidden_size // 2,  # int4 packing
+            ),
+            dtype=torch.uint8,
+        )
+        w2 = torch.randint(
+            0,
+            255,
+            (
+                num_experts,
+                hidden_size,
+                intermediate_size // 2,  # int4 packing
+            ),
+            dtype=torch.uint8,
+        )
+    elif use_int8_w8a16:
         w1 = torch.randint(
             -127,
             127,
@@ -139,7 +165,20 @@ def benchmark_config(
     w2_scale = None
     a1_scale = None
     a2_scale = None
-    if use_int8_w8a16:
+    if use_int4_w4a16:
+        if block_quant_shape is None:
+            raise ValueError("block_quant_shape is required for int4_w4a16")
+        group_size = block_quant_shape[1]
+        # Scales shape: (E, N, K // group_size) in fp16
+        w1_scale = torch.rand(
+            (num_experts, shard_intermediate_size, hidden_size // group_size),
+            dtype=dtype,
+        )
+        w2_scale = torch.rand(
+            (num_experts, hidden_size, intermediate_size // group_size),
+            dtype=dtype,
+        )
+    elif use_int8_w8a16:
         w1_scale = torch.randn(
             (num_experts, 2 * shard_intermediate_size), dtype=torch.float32
         )
@@ -198,6 +237,7 @@ def benchmark_config(
             a1_scale=a1_scale,
             a2_scale=a2_scale,
             block_shape=block_quant_shape,
+            weight_dtype="int4" if use_int4_w4a16 else None,
         )
 
         deep_gemm_experts = None
@@ -211,7 +251,8 @@ def benchmark_config(
                         hidden_dim=hidden_size,
                         intermediate_size_per_partition=shard_intermediate_size,
                         num_local_experts=num_experts,
-                        activation="silu",
+                        num_logical_experts=num_experts,
+                        activation=MoEActivation.SILU,
                         moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(),
                         in_dtype=init_dtype,
                         routing_method=RoutingMethodType.TopK,
@@ -226,9 +267,10 @@ def benchmark_config(
                 x, input_gating, topk, renormalize=not use_deep_gemm
             )
 
+            inplace = not disable_inplace()
             if use_deep_gemm:
                 return deep_gemm_experts(
-                    x, w1, w2, topk_weights, topk_ids, inplace=True
+                    x, w1, w2, topk_weights, topk_ids, inplace=inplace
                 )
             return fused_experts(
                 x,
@@ -236,7 +278,7 @@ def benchmark_config(
                 w2,
                 topk_weights,
                 topk_ids,
-                inplace=True,
+                inplace=inplace,
                 quant_config=quant_config,
             )
 
@@ -478,6 +520,7 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
+        use_int4_w4a16: bool = False,
         block_quant_shape: list[int] = None,
         use_deep_gemm: bool = False,
     ) -> tuple[dict[str, int], float]:
@@ -485,7 +528,10 @@ class BenchmarkWorker:
 
         set_random_seed(self.seed)
         dtype_str = _get_config_dtype_str(
-            dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
+            dtype,
+            use_int8_w8a16=use_int8_w8a16,
+            use_fp8_w8a8=use_fp8_w8a8,
+            use_int4_w4a16=use_int4_w4a16,
         )
         # NOTE(woosuk): The current naming convention uses w2.shape[2], which
         # is the intermediate size after silu_and_mul.
@@ -516,6 +562,7 @@ class BenchmarkWorker:
             dtype,
             use_fp8_w8a8,
             use_int8_w8a16,
+            use_int4_w4a16=use_int4_w4a16,
             num_iters=100,
             block_quant_shape=block_quant_shape,
             use_deep_gemm=use_deep_gemm,
@@ -532,6 +579,7 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
+        use_int4_w4a16: bool,
         search_space: list[dict[str, int]],
         block_quant_shape: list[int],
         use_deep_gemm: bool,
@@ -542,7 +590,7 @@ class BenchmarkWorker:
         best_config = None
         best_time = float("inf")
         if current_platform.is_rocm():
-            is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
+            is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16 or use_int4_w4a16)
             search_space = prune_rocm_search_space(
                 num_tokens,
                 shard_intermediate_size,
@@ -571,6 +619,7 @@ class BenchmarkWorker:
                         dtype,
                         use_fp8_w8a8,
                         use_int8_w8a16,
+                        use_int4_w4a16,
                         num_iters=20,
                         block_quant_shape=block_quant_shape,
                         use_deep_gemm=use_deep_gemm,
@@ -618,6 +667,7 @@ def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
             else {}
         ),
         **({"kpack": config["kpack"]} if "kpack" in config else {}),
+        **({"SPLIT_K": config["SPLIT_K"]} if "SPLIT_K" in config else {}),
     }
 
 
@@ -630,11 +680,15 @@ def save_configs(
     dtype: torch.dtype,
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
+    use_int4_w4a16: bool,
     block_quant_shape: list[int],
     save_dir: str,
 ) -> None:
     dtype_str = _get_config_dtype_str(
-        dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
+        dtype,
+        use_int8_w8a16=use_int8_w8a16,
+        use_fp8_w8a8=use_fp8_w8a8,
+        use_int4_w4a16=use_int4_w4a16,
     )
 
     # NOTE(woosuk): The current naming convention uses w2.shape[2], which
@@ -686,6 +740,7 @@ def get_model_params(config):
         "DeepseekV2ForCausalLM",
         "DeepseekV3ForCausalLM",
         "DeepseekV32ForCausalLM",
+        "GlmMoeDsaForCausalLM",
         "Glm4MoeForCausalLM",
         "Glm4MoeLiteForCausalLM",
         "NemotronHForCausalLM",
@@ -735,6 +790,38 @@ def get_model_params(config):
     return E, topk, intermediate_size, hidden_size
 
 
+def get_quantization_group_size(config) -> int | None:
+    """Extract the quantization group size from the HF model config.
+
+    This reads directly from the HuggingFace config object (as returned by
+    ``get_config()``), not from vLLM's quantization config classes.
+
+    Supports AWQ/GPTQ-style configs (direct 'group_size' key) and
+    compressed-tensors configs (nested inside 'config_groups').
+    """
+    quantization_config = getattr(config, "quantization_config", {})
+    if not isinstance(quantization_config, dict):
+        return None
+    # AWQ / GPTQ style: group_size is a top-level key
+    gs = quantization_config.get("group_size")
+    if gs is not None:
+        return gs
+    # compressed-tensors style: group_size is nested in config_groups
+    config_groups = quantization_config.get("config_groups", {})
+    if not isinstance(config_groups, dict):
+        return None
+    for group_cfg in config_groups.values():
+        if not isinstance(group_cfg, dict):
+            continue
+        weights = group_cfg.get("weights", {})
+        if not isinstance(weights, dict):
+            continue
+        gs = weights.get("group_size")
+        if gs is not None:
+            return gs
+    return None
+
+
 def main(args: argparse.Namespace):
     print(args)
 
@@ -753,7 +840,20 @@ def main(args: argparse.Namespace):
     dtype = torch.float16 if current_platform.is_rocm() else config.dtype
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"
     use_int8_w8a16 = args.dtype == "int8_w8a16"
+    use_int4_w4a16 = args.dtype == "int4_w4a16"
     block_quant_shape = get_weight_block_size_safety(config)
+    if use_int4_w4a16:
+        group_size = get_quantization_group_size(config)
+        if group_size is None:
+            raise ValueError(
+                "Could not determine group_size from model config. "
+                "The model's quantization_config must contain a 'group_size' "
+                "field (AWQ/GPTQ) or 'config_groups.*.weights.group_size' "
+                "(compressed-tensors)."
+            )
+        # For int4_w4a16, block_shape = [0, group_size]
+        # block_shape[0]=0 means no block quantization on N dimension
+        block_quant_shape = [0, group_size]
 
     if args.batch_size is None:
         batch_sizes = [
@@ -807,8 +907,20 @@ def main(args: argparse.Namespace):
         return ray.get(outputs)
 
     if args.tune:
-        is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
-        search_space = get_configs_compute_bound(is_fp16, block_quant_shape)
+        # int4_w4a16 weights are uint8-packed, not fp16; treat like fp8 for
+        # search space generation (no matrix_instr_nonkdim/kpack exploration).
+        is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16 or use_int4_w4a16)
+        # For int4_w4a16, the group_size constraint on BLOCK_SIZE_K does not
+        # apply: the gptq_awq kernel handles arbitrary BLOCK_SIZE_K regardless
+        # of group_size. Skip block_quant_shape filtering to keep the full
+        # search space (e.g. BLOCK_SIZE_K=64 with group_size=128).
+        tune_block_quant_shape = None if use_int4_w4a16 else block_quant_shape
+        search_space = get_configs_compute_bound(is_fp16, tune_block_quant_shape)
+        if use_int4_w4a16:
+            # SPLIT_K is a required kernel constexpr for gptq_awq kernel;
+            # only SPLIT_K=1 is used at runtime, so fix it during tuning.
+            for cfg in search_space:
+                cfg["SPLIT_K"] = 1
         print(f"Start tuning over {len(search_space)} configurations...")
         if use_deep_gemm:
             raise ValueError(
@@ -828,6 +940,7 @@ def main(args: argparse.Namespace):
                     dtype,
                     use_fp8_w8a8,
                     use_int8_w8a16,
+                    use_int4_w4a16,
                     search_space,
                     block_quant_shape,
                     use_deep_gemm,
@@ -847,6 +960,7 @@ def main(args: argparse.Namespace):
             dtype,
             use_fp8_w8a8,
             use_int8_w8a16,
+            use_int4_w4a16,
             block_quant_shape,
             args.save_dir,
         )
@@ -865,6 +979,7 @@ def main(args: argparse.Namespace):
                     dtype,
                     use_fp8_w8a8,
                     use_int8_w8a16,
+                    use_int4_w4a16,
                     block_quant_shape,
                     use_deep_gemm,
                 )
@@ -887,7 +1002,10 @@ if __name__ == "__main__":
     )
     parser.add_argument("--enable-expert-parallel", "-enable-ep", action="store_true")
     parser.add_argument(
-        "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto"
+        "--dtype",
+        type=str,
+        choices=["auto", "fp8_w8a8", "int8_w8a16", "int4_w4a16"],
+        default="auto",
     )
     parser.add_argument("--use-deep-gemm", action="store_true")
     parser.add_argument(

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -44,10 +44,8 @@ def benchmark_permute(
     hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype)
     # output_hidden_states = torch.empty_like(hidden_states)
     if use_fp8_w8a8:
-        align_block_size = 128  # deepgemm needs 128 m aligned block
         qhidden_states, scale = _fp8_quantize(hidden_states, None, None)
     else:
-        align_block_size = None
         qhidden_states = hidden_states
 
     gating_output = torch.randn(num_iters, num_tokens, num_experts, dtype=torch.float32)
@@ -67,7 +65,6 @@ def benchmark_permute(
             topk_ids=topk_ids,
             n_expert=num_experts,
             expert_map=None,
-            align_block_size=align_block_size,
         )
 
     # JIT compilation & warmup
@@ -117,10 +114,8 @@ def benchmark_unpermute(
     # init_dtype = torch.float16 if use_fp8_w8a8 else dtype
     hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype)
     if use_fp8_w8a8:
-        align_block_size = 128  # deepgemm needs 128 m aligned block
         qhidden_states, scale = _fp8_quantize(hidden_states, None, None)
     else:
-        align_block_size = None
         qhidden_states = hidden_states
 
     input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32)
@@ -142,7 +137,6 @@ def benchmark_unpermute(
             topk_ids=topk_ids,
             n_expert=num_experts,
             expert_map=None,
-            align_block_size=align_block_size,
         )
         # convert to fp16/bf16 as gemm output
         return (

cmake/external_projects/triton_kernels.cmake

@@ -1,9 +1,9 @@
 # Install OpenAI triton_kernels from https://github.com/triton-lang/triton/tree/main/python/triton_kernels
 
-set(DEFAULT_TRITON_KERNELS_TAG "v3.5.0")
+set(DEFAULT_TRITON_KERNELS_TAG "v3.6.0")
 
 # Set TRITON_KERNELS_SRC_DIR for use with local development with vLLM. We expect TRITON_KERNELS_SRC_DIR to
-# be directly set to the triton_kernels python directory. 
+# be directly set to the triton_kernels python directory.
 if (DEFINED ENV{TRITON_KERNELS_SRC_DIR})
   message(STATUS "[triton_kernels] Fetch from $ENV{TRITON_KERNELS_SRC_DIR}")
   FetchContent_Declare(
@@ -24,7 +24,7 @@ else()
   )
 endif()
 
-# Fetch content 
+# Fetch content
 FetchContent_MakeAvailable(triton_kernels)
 
 if (NOT triton_kernels_SOURCE_DIR)
@@ -47,7 +47,7 @@ install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/third_party/tr
 ## Copy .py files to install directory.
 install(DIRECTORY
         ${TRITON_KERNELS_PYTHON_DIR}
-        DESTINATION 
+        DESTINATION
         vllm/third_party/triton_kernels/
         COMPONENT triton_kernels
         FILES_MATCHING PATTERN "*.py")

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 2adfc8c2177c5b0e8ddeedfd5a8990d80eb496ff
+          GIT_TAG 5824e6e2008271063c3229ab3e7032bd74abbbc6
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

csrc/activation_kernels.cu

@@ -9,6 +9,111 @@
 
 namespace vllm {
 
+struct alignas(32) u32x8_t {
+  uint32_t u0, u1, u2, u3, u4, u5, u6, u7;
+};
+
+__device__ __forceinline__ void ld256(u32x8_t& val, const u32x8_t* ptr) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000
+  asm volatile("ld.global.nc.v8.u32 {%0,%1,%2,%3,%4,%5,%6,%7}, [%8];\n"
+               : "=r"(val.u0), "=r"(val.u1), "=r"(val.u2), "=r"(val.u3),
+                 "=r"(val.u4), "=r"(val.u5), "=r"(val.u6), "=r"(val.u7)
+               : "l"(ptr));
+#else
+  const uint4* uint_ptr = reinterpret_cast<const uint4*>(ptr);
+  uint4 top_half = __ldg(&uint_ptr[0]);
+  uint4 bottom_half = __ldg(&uint_ptr[1]);
+  val.u0 = top_half.x;
+  val.u1 = top_half.y;
+  val.u2 = top_half.z;
+  val.u3 = top_half.w;
+  val.u4 = bottom_half.x;
+  val.u5 = bottom_half.y;
+  val.u6 = bottom_half.z;
+  val.u7 = bottom_half.w;
+#endif
+}
+
+__device__ __forceinline__ void st256(u32x8_t& val, u32x8_t* ptr) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000
+  asm volatile("st.global.v8.u32 [%0], {%1,%2,%3,%4,%5,%6,%7,%8};\n"
+               :
+               : "l"(ptr), "r"(val.u0), "r"(val.u1), "r"(val.u2), "r"(val.u3),
+                 "r"(val.u4), "r"(val.u5), "r"(val.u6), "r"(val.u7)
+               : "memory");
+#else
+  uint4* uint_ptr = reinterpret_cast<uint4*>(ptr);
+  uint_ptr[0] = make_uint4(val.u0, val.u1, val.u2, val.u3);
+  uint_ptr[1] = make_uint4(val.u4, val.u5, val.u6, val.u7);
+#endif
+}
+
+template <bool support_256>
+struct VecTraits;
+
+template <>
+struct VecTraits<true> {
+  static constexpr int ARCH_MAX_VEC_SIZE = 32;
+  using vec_t = u32x8_t;
+};
+
+template <>
+struct VecTraits<false> {
+  static constexpr int ARCH_MAX_VEC_SIZE = 16;
+  using vec_t = int4;
+};
+
+template <typename T>
+struct PackedTraits;
+
+template <>
+struct PackedTraits<c10::BFloat16> {
+  using packed_t = __nv_bfloat162;
+};
+
+template <>
+struct PackedTraits<c10::Half> {
+  using packed_t = __half2;
+};
+
+template <>
+struct PackedTraits<float> {
+  using packed_t = float2;
+};
+
+template <typename packed_t>
+__device__ __forceinline__ float2 cast_to_float2(const packed_t& val) {
+  if constexpr (std::is_same_v<packed_t, __nv_bfloat162>) {
+    return __bfloat1622float2(val);
+  } else if constexpr (std::is_same_v<packed_t, __half2>) {
+    return __half22float2(val);
+  } else if constexpr (std::is_same_v<packed_t, float2>) {
+    return float2(val);
+  }
+}
+
+template <typename packed_t>
+__device__ __forceinline__ packed_t cast_to_packed(const float2& val) {
+  if constexpr (std::is_same_v<packed_t, __nv_bfloat162>) {
+    return __float22bfloat162_rn(val);
+  } else if constexpr (std::is_same_v<packed_t, __half2>) {
+    return __float22half2_rn(val);
+  } else if constexpr (std::is_same_v<packed_t, float2>) {
+    return float2(val);
+  }
+}
+
+template <typename packed_t>
+__device__ __forceinline__ packed_t packed_mul(const packed_t& x,
+                                               const packed_t& y) {
+  if constexpr (std::is_same_v<packed_t, __nv_bfloat162> ||
+                std::is_same_v<packed_t, __half2>) {
+    return __hmul2(x, y);
+  } else if constexpr (std::is_same_v<packed_t, float2>) {
+    return make_float2(x.x * y.x, x.y * y.y);
+  }
+}
+
 template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
           bool act_first>
 __device__ __forceinline__ scalar_t compute(const scalar_t& x,
@@ -16,52 +121,69 @@ __device__ __forceinline__ scalar_t compute(const scalar_t& x,
   return act_first ? ACT_FN(x) * y : x * ACT_FN(y);
 }
 
+template <typename packed_t, packed_t (*PACKED_ACT_FN)(const packed_t&),
+          bool act_first>
+__device__ __forceinline__ packed_t packed_compute(const packed_t& x,
+                                                   const packed_t& y) {
+  return act_first ? packed_mul(PACKED_ACT_FN(x), y)
+                   : packed_mul(x, PACKED_ACT_FN(y));
+}
+
 // Check if all pointers are 16-byte aligned for int4 vectorized access
-__device__ __forceinline__ bool is_16byte_aligned(const void* ptr) {
+__host__ __device__ __forceinline__ bool is_16byte_aligned(const void* ptr) {
   return (reinterpret_cast<uintptr_t>(ptr) & 15) == 0;
 }
 
+// Check if all pointers are 16-byte aligned for longlong4_32a vectorized access
+__host__ __device__ __forceinline__ bool is_32byte_aligned(const void* ptr) {
+  return (reinterpret_cast<uintptr_t>(ptr) & 31) == 0;
+}
+
 // Activation and gating kernel template.
-template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
-          bool act_first>
+template <typename scalar_t, typename packed_t,
+          scalar_t (*ACT_FN)(const scalar_t&),
+          packed_t (*PACKED_ACT_FN)(const packed_t&), bool act_first,
+          bool use_vec, bool use_256b = false>
 __global__ void act_and_mul_kernel(
     scalar_t* __restrict__ out,          // [..., d]
     const scalar_t* __restrict__ input,  // [..., 2, d]
     const int d) {
-  constexpr int VEC_SIZE = 16 / sizeof(scalar_t);
-  const int64_t token_idx = blockIdx.x;
-  const scalar_t* x_ptr = input + token_idx * 2 * d;
+  const scalar_t* x_ptr = input + blockIdx.x * 2 * d;
   const scalar_t* y_ptr = x_ptr + d;
-  scalar_t* out_ptr = out + token_idx * d;
+  scalar_t* out_ptr = out + blockIdx.x * d;
 
-  // Check alignment for 128-bit vectorized access.
-  // All three pointers must be 16-byte aligned for safe int4 operations.
-  const bool aligned = is_16byte_aligned(x_ptr) && is_16byte_aligned(y_ptr) &&
-                       is_16byte_aligned(out_ptr);
+  if constexpr (use_vec) {
+    // Fast path: 128-bit/256-bit vectorized loop
+    using vec_t = typename VecTraits<use_256b>::vec_t;
+    constexpr int ARCH_MAX_VEC_SIZE = VecTraits<use_256b>::ARCH_MAX_VEC_SIZE;
+    constexpr int VEC_SIZE = ARCH_MAX_VEC_SIZE / sizeof(packed_t);
 
-  if (aligned && d >= VEC_SIZE) {
-    // Fast path: 128-bit vectorized loop
-    const int4* x_vec = reinterpret_cast<const int4*>(x_ptr);
-    const int4* y_vec = reinterpret_cast<const int4*>(y_ptr);
-    int4* out_vec = reinterpret_cast<int4*>(out_ptr);
-    const int num_vecs = d / VEC_SIZE;
-    const int vec_end = num_vecs * VEC_SIZE;
+    const vec_t* x_vec = reinterpret_cast<const vec_t*>(x_ptr);
+    const vec_t* y_vec = reinterpret_cast<const vec_t*>(y_ptr);
+    vec_t* out_vec = reinterpret_cast<vec_t*>(out_ptr);
+    const int num_vecs = d / 2 / VEC_SIZE;
 
     for (int i = threadIdx.x; i < num_vecs; i += blockDim.x) {
-      int4 x = VLLM_LDG(&x_vec[i]), y = VLLM_LDG(&y_vec[i]), r;
-      auto* xp = reinterpret_cast<scalar_t*>(&x);
-      auto* yp = reinterpret_cast<scalar_t*>(&y);
-      auto* rp = reinterpret_cast<scalar_t*>(&r);
+      vec_t x, y;
+      if constexpr (use_256b) {
+        ld256(x, &x_vec[i]);
+        ld256(y, &y_vec[i]);
+      } else {
+        x = VLLM_LDG(&x_vec[i]);
+        y = VLLM_LDG(&y_vec[i]);
+      }
+      auto* xp = reinterpret_cast<packed_t*>(&x);
+      auto* yp = reinterpret_cast<packed_t*>(&y);
 #pragma unroll
       for (int j = 0; j < VEC_SIZE; j++) {
-        rp[j] = compute<scalar_t, ACT_FN, act_first>(xp[j], yp[j]);
+        xp[j] =
+            packed_compute<packed_t, PACKED_ACT_FN, act_first>(xp[j], yp[j]);
+      }
+      if constexpr (use_256b) {
+        st256(x, &out_vec[i]);
+      } else {
+        out_vec[i] = x;
       }
-      out_vec[i] = r;
-    }
-    // Scalar cleanup for remaining elements
-    for (int i = vec_end + threadIdx.x; i < d; i += blockDim.x) {
-      out_ptr[i] = compute<scalar_t, ACT_FN, act_first>(VLLM_LDG(&x_ptr[i]),
-                                                        VLLM_LDG(&y_ptr[i]));
     }
   } else {
     // Scalar fallback for unaligned data or small d
@@ -79,6 +201,15 @@ __device__ __forceinline__ T silu_kernel(const T& x) {
   return (T)(((float)x) / (1.0f + expf((float)-x)));
 }
 
+template <typename packed_t>
+__device__ __forceinline__ packed_t packed_silu_kernel(const packed_t& val) {
+  // x * sigmoid(x)
+  float2 fval = cast_to_float2(val);
+  fval.x = fval.x / (1.0f + expf(-fval.x));
+  fval.y = fval.y / (1.0f + expf(-fval.y));
+  return cast_to_packed<packed_t>(fval);
+}
+
 template <typename T>
 __device__ __forceinline__ T gelu_kernel(const T& x) {
   // Equivalent to PyTorch GELU with 'none' approximation.
@@ -89,6 +220,18 @@ __device__ __forceinline__ T gelu_kernel(const T& x) {
   return (T)(f * 0.5f * (1.0f + ::erf(f * ALPHA)));
 }
 
+template <typename packed_t>
+__device__ __forceinline__ packed_t packed_gelu_kernel(const packed_t& val) {
+  // Equivalent to PyTorch GELU with 'none' approximation.
+  // Refer to:
+  // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L36-L38
+  constexpr float ALPHA = M_SQRT1_2;
+  float2 fval = cast_to_float2(val);
+  fval.x = fval.x * 0.5f * (1.0f + ::erf(fval.x * ALPHA));
+  fval.y = fval.y * 0.5f * (1.0f + ::erf(fval.y * ALPHA));
+  return cast_to_packed<packed_t>(fval);
+}
+
 template <typename T>
 __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
   // Equivalent to PyTorch GELU with 'tanh' approximation.
@@ -102,32 +245,83 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
   return (T)(0.5f * f * (1.0f + ::tanhf(inner)));
 }
 
+template <typename packed_t>
+__device__ __forceinline__ packed_t
+packed_gelu_tanh_kernel(const packed_t& val) {
+  // Equivalent to PyTorch GELU with 'tanh' approximation.
+  // Refer to:
+  // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L25-L30
+  float2 fval = cast_to_float2(val);
+  constexpr float BETA = M_SQRT2 * M_2_SQRTPI * 0.5f;
+  constexpr float KAPPA = 0.044715;
+
+  float x_cube = fval.x * fval.x * fval.x;
+  float inner = BETA * (fval.x + KAPPA * x_cube);
+  fval.x = 0.5f * fval.x * (1.0f + ::tanhf(inner));
+
+  x_cube = fval.y * fval.y * fval.y;
+  inner = BETA * (fval.y + KAPPA * x_cube);
+  fval.y = 0.5f * fval.y * (1.0f + ::tanhf(inner));
+  return cast_to_packed<packed_t>(fval);
+}
+
 }  // namespace vllm
 
 // Launch activation and gating kernel.
 // Use ACT_FIRST (bool) indicating whether to apply the activation function
 // first.
-#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, ACT_FIRST)                 \
-  int d = input.size(-1) / 2;                                            \
-  int64_t num_tokens = input.numel() / input.size(-1);                   \
-  dim3 grid(num_tokens);                                                 \
-  dim3 block(std::min(d, 1024));                                         \
-  if (num_tokens == 0) {                                                 \
-    return;                                                              \
-  }                                                                      \
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));      \
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();          \
-  VLLM_DISPATCH_FLOATING_TYPES(                                          \
-      input.scalar_type(), "act_and_mul_kernel", [&] {                   \
-        vllm::act_and_mul_kernel<scalar_t, KERNEL<scalar_t>, ACT_FIRST>  \
-            <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),       \
-                                         input.data_ptr<scalar_t>(), d); \
-      });
+#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, PACKED_KERNEL, ACT_FIRST)     \
+  auto dtype = input.scalar_type();                                         \
+  int d = input.size(-1) / 2;                                               \
+  int64_t num_tokens = input.numel() / input.size(-1);                      \
+  if (num_tokens == 0) {                                                    \
+    return;                                                                 \
+  }                                                                         \
+  dim3 grid(num_tokens);                                                    \
+  int cc_major = at::cuda::getCurrentDeviceProperties()->major;             \
+  int support_vec = (cc_major >= 10 && num_tokens > 128) ? 32 : 16;         \
+  int vec_size = support_vec / at::elementSize(dtype);                      \
+  const bool use_vec = (d % vec_size == 0);                                 \
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));         \
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();             \
+  if (use_vec) {                                                            \
+    dim3 block(std::min(d / vec_size, 1024));                               \
+    if (cc_major >= 10 && num_tokens > 128) {                               \
+      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {       \
+        vllm::act_and_mul_kernel<                                           \
+            scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,      \
+            KERNEL<scalar_t>,                                               \
+            PACKED_KERNEL<typename vllm::PackedTraits<scalar_t>::packed_t>, \
+            ACT_FIRST, true, true><<<grid, block, 0, stream>>>(             \
+            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);       \
+      });                                                                   \
+    } else {                                                                \
+      VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {       \
+        vllm::act_and_mul_kernel<                                           \
+            scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,      \
+            KERNEL<scalar_t>,                                               \
+            PACKED_KERNEL<typename vllm::PackedTraits<scalar_t>::packed_t>, \
+            ACT_FIRST, true, false><<<grid, block, 0, stream>>>(            \
+            out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);       \
+      });                                                                   \
+    }                                                                       \
+  } else {                                                                  \
+    dim3 block(std::min(d, 1024));                                          \
+    VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {         \
+      vllm::act_and_mul_kernel<                                             \
+          scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,        \
+          KERNEL<scalar_t>,                                                 \
+          PACKED_KERNEL<typename vllm::PackedTraits<scalar_t>::packed_t>,   \
+          ACT_FIRST, false><<<grid, block, 0, stream>>>(                    \
+          out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d);         \
+    });                                                                     \
+  }
 
 void silu_and_mul(torch::Tensor& out,    // [..., d]
                   torch::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, true);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
+                                true);
 }
 
 void mul_and_silu(torch::Tensor& out,    // [..., d]
@@ -135,19 +329,22 @@ void mul_and_silu(torch::Tensor& out,    // [..., d]
 {
   // The difference between mul_and_silu and silu_and_mul is that mul_and_silu
   // applies the silu to the latter half of the input.
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, false);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
+                                false);
 }
 
 void gelu_and_mul(torch::Tensor& out,    // [..., d]
                   torch::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel, true);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel, vllm::packed_gelu_kernel,
+                                true);
 }
 
 void gelu_tanh_and_mul(torch::Tensor& out,    // [..., d]
                        torch::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel, true);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel,
+                                vllm::packed_gelu_tanh_kernel, true);
 }
 
 namespace vllm {
@@ -158,42 +355,57 @@ __device__ __forceinline__ T fatrelu_kernel(const T& x, const float threshold) {
   return (T)(f > threshold ? f : 0.0f);
 }
 
-template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&, const float)>
+template <typename packed_t>
+__device__ __forceinline__ packed_t
+packed_fatrelu_kernel(const packed_t& val, const float threshold) {
+  float2 fval = cast_to_float2(val);
+  fval.x = fval.x > threshold ? fval.x : 0.0f;
+  fval.y = fval.y > threshold ? fval.y : 0.0f;
+  return cast_to_packed<packed_t>(fval);
+}
+
+template <typename scalar_t, typename packed_t,
+          scalar_t (*ACT_FN)(const scalar_t&, const float),
+          packed_t (*PACKED_ACT_FN)(const packed_t&, const float), bool use_vec,
+          bool use_256b = false>
 __global__ void act_and_mul_kernel_with_param(
     scalar_t* __restrict__ out, const scalar_t* __restrict__ input, const int d,
     const float param) {
-  constexpr int VEC_SIZE = 16 / sizeof(scalar_t);
-  const int64_t token_idx = blockIdx.x;
-  const scalar_t* x_ptr = input + token_idx * 2 * d;
+  const scalar_t* x_ptr = input + blockIdx.x * 2 * d;
   const scalar_t* y_ptr = x_ptr + d;
-  scalar_t* out_ptr = out + token_idx * d;
+  scalar_t* out_ptr = out + blockIdx.x * d;
 
-  // Check alignment for 128-bit vectorized access
-  const bool aligned = is_16byte_aligned(x_ptr) && is_16byte_aligned(y_ptr) &&
-                       is_16byte_aligned(out_ptr);
+  if constexpr (use_vec) {
+    // Fast path: 128-bit/256-bit vectorized loop
+    using vec_t = typename VecTraits<use_256b>::vec_t;
+    constexpr int ARCH_MAX_VEC_SIZE = VecTraits<use_256b>::ARCH_MAX_VEC_SIZE;
+    constexpr int VEC_SIZE = ARCH_MAX_VEC_SIZE / sizeof(packed_t);
 
-  if (aligned && d >= VEC_SIZE) {
-    // Fast path: 128-bit vectorized loop
-    const int4* x_vec = reinterpret_cast<const int4*>(x_ptr);
-    const int4* y_vec = reinterpret_cast<const int4*>(y_ptr);
-    int4* out_vec = reinterpret_cast<int4*>(out_ptr);
-    const int num_vecs = d / VEC_SIZE;
-    const int vec_end = num_vecs * VEC_SIZE;
+    const vec_t* x_vec = reinterpret_cast<const vec_t*>(x_ptr);
+    const vec_t* y_vec = reinterpret_cast<const vec_t*>(y_ptr);
+    vec_t* out_vec = reinterpret_cast<vec_t*>(out_ptr);
+    const int num_vecs = d / 2 / VEC_SIZE;
 
     for (int i = threadIdx.x; i < num_vecs; i += blockDim.x) {
-      int4 x = VLLM_LDG(&x_vec[i]), y = VLLM_LDG(&y_vec[i]), r;
-      auto* xp = reinterpret_cast<scalar_t*>(&x);
-      auto* yp = reinterpret_cast<scalar_t*>(&y);
-      auto* rp = reinterpret_cast<scalar_t*>(&r);
+      vec_t x, y;
+      if constexpr (use_256b) {
+        ld256(x, &x_vec[i]);
+        ld256(y, &y_vec[i]);
+      } else {
+        x = VLLM_LDG(&x_vec[i]);
+        y = VLLM_LDG(&y_vec[i]);
+      }
+      auto* xp = reinterpret_cast<packed_t*>(&x);
+      auto* yp = reinterpret_cast<packed_t*>(&y);
 #pragma unroll
       for (int j = 0; j < VEC_SIZE; j++) {
-        rp[j] = ACT_FN(xp[j], param) * yp[j];
+        xp[j] = packed_mul(PACKED_ACT_FN(xp[j], param), yp[j]);
+      }
+      if constexpr (use_256b) {
+        st256(x, &out_vec[i]);
+      } else {
+        out_vec[i] = x;
       }
-      out_vec[i] = r;
-    }
-    // Scalar cleanup for remaining elements
-    for (int i = vec_end + threadIdx.x; i < d; i += blockDim.x) {
-      out_ptr[i] = ACT_FN(VLLM_LDG(&x_ptr[i]), param) * VLLM_LDG(&y_ptr[i]);
     }
   } else {
     // Scalar fallback for unaligned data or small d
@@ -276,20 +488,58 @@ __global__ void swigluoai_and_mul_kernel(
 
 }  // namespace vllm
 
-#define LAUNCH_ACTIVATION_GATE_KERNEL_WITH_PARAM(KERNEL, PARAM)         \
-  int d = input.size(-1) / 2;                                           \
-  int64_t num_tokens = input.numel() / input.size(-1);                  \
-  dim3 grid(num_tokens);                                                \
-  dim3 block(std::min(d, 1024));                                        \
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));     \
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();         \
-  VLLM_DISPATCH_FLOATING_TYPES(                                         \
-      input.scalar_type(), "act_and_mul_kernel_with_param", [&] {       \
-        vllm::act_and_mul_kernel_with_param<scalar_t, KERNEL<scalar_t>> \
-            <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),      \
-                                         input.data_ptr<scalar_t>(), d, \
-                                         PARAM);                        \
-      });
+#define LAUNCH_ACTIVATION_GATE_KERNEL_WITH_PARAM(KERNEL, PACKED_KERNEL, PARAM) \
+  auto dtype = input.scalar_type();                                            \
+  int d = input.size(-1) / 2;                                                  \
+  int64_t num_tokens = input.numel() / input.size(-1);                         \
+  if (num_tokens == 0) {                                                       \
+    return;                                                                    \
+  }                                                                            \
+  dim3 grid(num_tokens);                                                       \
+  int cc_major = at::cuda::getCurrentDeviceProperties()->major;                \
+  int support_vec = (cc_major >= 10 && num_tokens > 128) ? 32 : 16;            \
+  int vec_size = support_vec / at::elementSize(dtype);                         \
+  const bool use_vec = (d % vec_size == 0);                                    \
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));            \
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                \
+  if (use_vec) {                                                               \
+    dim3 block(std::min(d / vec_size, 1024));                                  \
+    if (cc_major >= 10 && num_tokens > 128) {                                  \
+      VLLM_DISPATCH_FLOATING_TYPES(                                            \
+          dtype, "act_and_mul_kernel_with_param", [&] {                        \
+            vllm::act_and_mul_kernel_with_param<                               \
+                scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,     \
+                KERNEL<scalar_t>,                                              \
+                PACKED_KERNEL<                                                 \
+                    typename vllm::PackedTraits<scalar_t>::packed_t>,          \
+                true, true><<<grid, block, 0, stream>>>(                       \
+                out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d,       \
+                PARAM);                                                        \
+          });                                                                  \
+    } else {                                                                   \
+      VLLM_DISPATCH_FLOATING_TYPES(                                            \
+          dtype, "act_and_mul_kernel_with_param", [&] {                        \
+            vllm::act_and_mul_kernel_with_param<                               \
+                scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,     \
+                KERNEL<scalar_t>,                                              \
+                PACKED_KERNEL<                                                 \
+                    typename vllm::PackedTraits<scalar_t>::packed_t>,          \
+                true, false><<<grid, block, 0, stream>>>(                      \
+                out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d,       \
+                PARAM);                                                        \
+          });                                                                  \
+    }                                                                          \
+  } else {                                                                     \
+    dim3 block(std::min(d, 1024));                                             \
+    VLLM_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel_with_param", [&] { \
+      vllm::act_and_mul_kernel_with_param<                                     \
+          scalar_t, typename vllm::PackedTraits<scalar_t>::packed_t,           \
+          KERNEL<scalar_t>,                                                    \
+          PACKED_KERNEL<typename vllm::PackedTraits<scalar_t>::packed_t>,      \
+          false><<<grid, block, 0, stream>>>(                                  \
+          out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), d, PARAM);     \
+    });                                                                        \
+  }
 
 #define LAUNCH_SIGLUOAI_AND_MUL(KERNEL, ALPHA, LIMIT)                          \
   int d = input.size(-1) / 2;                                                  \
@@ -309,7 +559,8 @@ __global__ void swigluoai_and_mul_kernel(
 void fatrelu_and_mul(torch::Tensor& out,    // [..., d],
                      torch::Tensor& input,  // [..., 2 * d]
                      double threshold) {
-  LAUNCH_ACTIVATION_GATE_KERNEL_WITH_PARAM(vllm::fatrelu_kernel, threshold);
+  LAUNCH_ACTIVATION_GATE_KERNEL_WITH_PARAM(
+      vllm::fatrelu_kernel, vllm::packed_fatrelu_kernel, threshold);
 }
 void swigluoai_and_mul(torch::Tensor& out,    // [..., d]
                        torch::Tensor& input,  // [..., 2 * d]
@@ -319,39 +570,41 @@ void swigluoai_and_mul(torch::Tensor& out,    // [..., d]
 namespace vllm {
 
 // Element-wise activation kernel template.
-template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&)>
+template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&), bool use_vec,
+          bool use_256b = false>
 __global__ void activation_kernel(
     scalar_t* __restrict__ out,          // [..., d]
     const scalar_t* __restrict__ input,  // [..., d]
     const int d) {
-  constexpr int VEC_SIZE = 16 / sizeof(scalar_t);
-  const int64_t token_idx = blockIdx.x;
-  const scalar_t* in_ptr = input + token_idx * d;
-  scalar_t* out_ptr = out + token_idx * d;
+  const scalar_t* in_ptr = input + blockIdx.x * d;
+  scalar_t* out_ptr = out + blockIdx.x * d;
 
-  // Check alignment for 128-bit vectorized access
-  const bool aligned = is_16byte_aligned(in_ptr) && is_16byte_aligned(out_ptr);
-
-  if (aligned && d >= VEC_SIZE) {
-    // Fast path: 128-bit vectorized loop
-    const int4* in_vec = reinterpret_cast<const int4*>(in_ptr);
-    int4* out_vec = reinterpret_cast<int4*>(out_ptr);
+  if constexpr (use_vec) {
+    // Fast path: 128-bit/256-bit vectorized loop
+    using vec_t = typename VecTraits<use_256b>::vec_t;
+    constexpr int ARCH_MAX_VEC_SIZE = VecTraits<use_256b>::ARCH_MAX_VEC_SIZE;
+    constexpr int VEC_SIZE = ARCH_MAX_VEC_SIZE / sizeof(scalar_t);
+    const vec_t* in_vec = reinterpret_cast<const vec_t*>(in_ptr);
+    vec_t* out_vec = reinterpret_cast<vec_t*>(out_ptr);
     const int num_vecs = d / VEC_SIZE;
-    const int vec_end = num_vecs * VEC_SIZE;
 
     for (int i = threadIdx.x; i < num_vecs; i += blockDim.x) {
-      int4 v = VLLM_LDG(&in_vec[i]), r;
+      vec_t v;
+      if constexpr (use_256b) {
+        ld256(v, &in_vec[i]);
+      } else {
+        v = VLLM_LDG(&in_vec[i]);
+      }
       auto* vp = reinterpret_cast<scalar_t*>(&v);
-      auto* rp = reinterpret_cast<scalar_t*>(&r);
 #pragma unroll
       for (int j = 0; j < VEC_SIZE; j++) {
-        rp[j] = ACT_FN(vp[j]);
+        vp[j] = ACT_FN(vp[j]);
+      }
+      if constexpr (use_256b) {
+        st256(v, &out_vec[i]);
+      } else {
+        out_vec[i] = v;
       }
-      out_vec[i] = r;
-    }
-    // Scalar cleanup for remaining elements
-    for (int i = vec_end + threadIdx.x; i < d; i += blockDim.x) {
-      out_ptr[i] = ACT_FN(VLLM_LDG(&in_ptr[i]));
     }
   } else {
     // Scalar fallback for unaligned data or small d
@@ -365,18 +618,43 @@ __global__ void activation_kernel(
 }  // namespace vllm
 
 // Launch element-wise activation kernel.
-#define LAUNCH_ACTIVATION_KERNEL(KERNEL)                                       \
-  int d = input.size(-1);                                                      \
-  int64_t num_tokens = input.numel() / d;                                      \
-  dim3 grid(num_tokens);                                                       \
-  dim3 block(std::min(d, 1024));                                               \
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));            \
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                \
-  VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "activation_kernel", [&] { \
-    vllm::activation_kernel<scalar_t, KERNEL<scalar_t>>                        \
-        <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),                 \
-                                     input.data_ptr<scalar_t>(), d);           \
-  });
+#define LAUNCH_ACTIVATION_KERNEL(KERNEL)                                 \
+  auto dtype = input.scalar_type();                                      \
+  int d = input.size(-1);                                                \
+  int64_t num_tokens = input.numel() / input.size(-1);                   \
+  if (num_tokens == 0) {                                                 \
+    return;                                                              \
+  }                                                                      \
+  dim3 grid(num_tokens);                                                 \
+  int cc_major = at::cuda::getCurrentDeviceProperties()->major;          \
+  int support_vec = (cc_major >= 10 && num_tokens > 128) ? 32 : 16;      \
+  int vec_size = support_vec / at::elementSize(dtype);                   \
+  const bool use_vec = (d % vec_size == 0);                              \
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));      \
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();          \
+  if (use_vec) {                                                         \
+    dim3 block(std::min(d / vec_size, 1024));                            \
+    if (cc_major >= 10 && num_tokens > 128) {                            \
+      VLLM_DISPATCH_FLOATING_TYPES(dtype, "activation_kernel", [&] {     \
+        vllm::activation_kernel<scalar_t, KERNEL<scalar_t>, true, true>  \
+            <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),       \
+                                         input.data_ptr<scalar_t>(), d); \
+      });                                                                \
+    } else {                                                             \
+      VLLM_DISPATCH_FLOATING_TYPES(dtype, "activation_kernel", [&] {     \
+        vllm::activation_kernel<scalar_t, KERNEL<scalar_t>, true, false> \
+            <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),       \
+                                         input.data_ptr<scalar_t>(), d); \
+      });                                                                \
+    }                                                                    \
+  } else {                                                               \
+    dim3 block(std::min(d, 1024));                                       \
+    VLLM_DISPATCH_FLOATING_TYPES(dtype, "activation_kernel", [&] {       \
+      vllm::activation_kernel<scalar_t, KERNEL<scalar_t>, false>         \
+          <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),         \
+                                       input.data_ptr<scalar_t>(), d);   \
+    });                                                                  \
+  }
 
 namespace vllm {
 

csrc/cache_kernels.cu

@@ -1234,8 +1234,13 @@ void cp_gather_and_upconvert_fp8_kv_cache(
               "src_cache and seq_lens must be on the same device");
   TORCH_CHECK(src_cache.device() == workspace_starts.device(),
               "src_cache and workspace_starts must be on the same device");
-
-  TORCH_CHECK(src_cache.dtype() == torch::kUInt8, "src_cache must be uint8");
+  auto dtype = src_cache.scalar_type();
+  TORCH_CHECK(
+      dtype == at::ScalarType::Byte ||               // uint8
+          dtype == at::ScalarType::Float8_e4m3fn ||  // fp8 e4m3
+          dtype == at::ScalarType::Float8_e5m2,      // fp8 e5m2
+      "src_cache must be uint8, float8_e4m3fn, or float8_e5m2, but got ",
+      src_cache.dtype());
   TORCH_CHECK(dst.dtype() == torch::kBFloat16, "dst must be bfloat16");
   TORCH_CHECK(head_dim == 576, "head_dim must be 576 for MLA");
 
@@ -1244,14 +1249,21 @@ void cp_gather_and_upconvert_fp8_kv_cache(
   int64_t cache_entry_stride = src_cache.stride(1);
   int64_t dst_entry_stride = dst.stride(0);
 
+  const uint8_t* src_ptr = nullptr;
+  if (dtype == at::ScalarType::Byte) {
+    src_ptr = src_cache.data_ptr<uint8_t>();
+  } else {
+    // float8_e4m3fn or float8_e5m2
+    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);
 
   vllm::cp_gather_and_upconvert_fp8_kv_cache<<<grid, block, 0, stream>>>(
-      src_cache.data_ptr<uint8_t>(),
-      reinterpret_cast<__nv_bfloat16*>(dst.data_ptr()),
+      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_stride, cache_block_stride, cache_entry_stride,

csrc/cpu/cpu_attn_impl.hpp

@@ -821,7 +821,7 @@ struct VecTypeTrait<c10::BFloat16> {
   using vec_t = vec_op::BF16Vec16;
 };
 
-#if !defined(__powerpc__) && !defined(__s390x__)
+#if !defined(__powerpc__)
 template <>
 struct VecTypeTrait<c10::Half> {
   using vec_t = vec_op::FP16Vec16;
@@ -1107,7 +1107,8 @@ class AttentionMainLoop {
           if (sliding_window_left != -1) {
             pos = std::max(pos, curr_token_pos - sliding_window_left);
           }
-          return pos;
+          // Clamp to tile end to avoid OOB when window starts past the tile
+          return std::min(pos, kv_tile_end_pos);
         }();
 
         int32_t right_kv_pos = [&]() {

csrc/cpu/cpu_attn_neon.hpp

@@ -4,6 +4,9 @@
 #include "cpu_attn_impl.hpp"
 #include <arm_neon.h>
 #include <type_traits>
+#ifdef ARM_BF16_SUPPORT
+  #include "cpu_attn_neon_bfmmla.hpp"
+#endif
 namespace cpu_attention {
 
 namespace {
@@ -57,7 +60,7 @@ FORCE_INLINE void load_row8_B_as_f32<c10::BFloat16>(const c10::BFloat16* p,
 #endif
 }
 
-// Mx8, with 1 <= M <= 8 , K streamed, unroll-by-4 with NEON FMLAs
+// Mx8, with 1 <= M <= 8 , K streamed, unroll-by-4 with ASIMD FMLAs
 // #Loads = (K // 4) * (M + 4 * sizeof(kv_cache_t) / 2)
 // #FMLAs = (K // 4) * (4 * 2 * M)
 // We have (4 * 2 * M) FMLAs for (M + 4 * sizeof(kv_cache_t) / 2) loads
@@ -381,6 +384,18 @@ class AttentionImpl<ISA::NEON, scalar_t, head_dim> {
     }
   }
 };
+
+#ifdef ARM_BF16_SUPPORT
+// For BF16 on Arm, reuse the BFMMLA kernels with 32-token alignment.
+template <int64_t head_dim>
+class AttentionImpl<ISA::NEON, c10::BFloat16, head_dim>
+    : public AttentionImplNEONBFMMLA<BLOCK_SIZE_ALIGNMENT, ISA::NEON,
+                                     head_dim> {};
+#endif
 }  // namespace cpu_attention
 
-#endif  // #ifndef CPU_ATTN_NEON_HPP
+#undef BLOCK_SIZE_ALIGNMENT
+#undef HEAD_SIZE_ALIGNMENT
+#undef MAX_Q_HEAD_NUM_PER_ITER
+
+#endif  // #ifndef CPU_ATTN_ASIMD_HPP

csrc/cpu/cpu_attn_neon_bfmmla.hpp

@@ -0,0 +1,682 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#ifndef CPU_ATTN_NEON_BFMMLA_HPP
+#define CPU_ATTN_NEON_BFMMLA_HPP
+
+#include "cpu_attn_impl.hpp"
+
+#include <arm_neon.h>
+
+#include <cstdint>
+#include <vector>
+
+namespace cpu_attention {
+
+namespace {
+
+// BFMMLA tile dimensions
+constexpr int32_t TILE_ROWS = 2;  // M dimension
+constexpr int32_t TILE_K = 4;     // K reduction
+constexpr int32_t TILE_COLS = 2;  // N dimension (column-pair)
+
+// Derived constants
+constexpr int32_t OUTPUT_COLS_PER_BLOCK = 8;   // 4 column-pairs
+constexpr int32_t K_TOKENS_PER_GROUP = 8;      // Tokens grouped in K cache
+constexpr int32_t V_TOKENS_PER_ROW_BLOCK = 4;  // Tokens per V cache row block
+constexpr int32_t K_INNER_STRIDE = K_TOKENS_PER_GROUP * TILE_K;
+constexpr int32_t V_INNER_STRIDE = V_TOKENS_PER_ROW_BLOCK * TILE_COLS;
+constexpr int32_t PACK_ELEMENTS_PER_K_CHUNK = TILE_ROWS * TILE_K;  // A packing
+
+// Matrix Packing and Accumulator
+// Reshape two rows of Q into BFMMLA-friendly interleaved
+// Input:  row0 = [a0,a1,a2,a3], row1 = [b0,b1,b2,b3]
+// Output: [a0,a1,a2,a3,b0,b1,b2,b3, a4,a5,a6,a7,b4,b5,b6,b7]
+// For K tail (K % TILE_K != 0): pads with zeros to complete the final chunk
+FORCE_INLINE void reshape_Q_2xK_for_bfmmla(const c10::BFloat16* __restrict r0,
+                                           const c10::BFloat16* __restrict r1,
+                                           c10::BFloat16* __restrict dst,
+                                           int32_t K) {
+  const uint16_t* s0 = reinterpret_cast<const uint16_t*>(r0);
+  const uint16_t* s1 = reinterpret_cast<const uint16_t*>(r1);
+  uint16_t* d = reinterpret_cast<uint16_t*>(dst);
+
+  // Process TILE_K elements at a time (PACK_ELEMENTS_PER_K_CHUNK output)
+  int32_t k = 0;
+  for (; k + TILE_K <= K; k += TILE_K, d += PACK_ELEMENTS_PER_K_CHUNK) {
+    vst1q_u16(d, vcombine_u16(vld1_u16(s0 + k), vld1_u16(s1 + k)));
+  }
+
+  // Handle K tail: pack remaining elements with zero-padding
+  const int32_t tail = K - k;
+  if (tail > 0) {
+    // Pack remaining tail elements: [r0[k..k+tail-1], pad, r1[k..k+tail-1],
+    // pad]
+    for (int32_t t = 0; t < tail; ++t) {
+      d[t] = s0[k + t];
+      d[t + TILE_K] = s1[k + t];
+    }
+    // Zero-pad the rest
+    for (int32_t t = tail; t < TILE_K; ++t) {
+      d[t] = 0;
+      d[t + TILE_K] = 0;
+    }
+  }
+}
+
+// 2x2 accumulator load/store with compile-time row count
+template <int32_t m_rows>
+FORCE_INLINE float32x4_t load_acc_2x2(float* base, int64_t ldc, int col_off) {
+  static_assert(m_rows == 1 || m_rows == 2);
+  float32x2_t row0 = vld1_f32(base + col_off);
+  float32x2_t row1 =
+      (m_rows == 2) ? vld1_f32(base + ldc + col_off) : vdup_n_f32(0.f);
+  return vcombine_f32(row0, row1);
+}
+
+template <int32_t m_rows>
+FORCE_INLINE void store_acc_2x2(float32x4_t acc, float* base, int64_t ldc,
+                                int col_off) {
+  static_assert(m_rows == 1 || m_rows == 2);
+  vst1_f32(base + col_off, vget_low_f32(acc));
+  if constexpr (m_rows == 2) {
+    vst1_f32(base + ldc + col_off, vget_high_f32(acc));
+  }
+}
+
+// Initialize 4 column-pair accumulators for 2 rows (8 columns total)
+#define INIT_ACC_ROWPAIR_4(a0, a1, a2, a3, Crow, ldc, m_rows, accum) \
+  do {                                                               \
+    if (accum) {                                                     \
+      if (m_rows == 2) {                                             \
+        a0 = load_acc_2x2<2>(Crow, ldc, 0);                          \
+        a1 = load_acc_2x2<2>(Crow, ldc, 2);                          \
+        a2 = load_acc_2x2<2>(Crow, ldc, 4);                          \
+        a3 = load_acc_2x2<2>(Crow, ldc, 6);                          \
+      } else {                                                       \
+        a0 = load_acc_2x2<1>(Crow, ldc, 0);                          \
+        a1 = load_acc_2x2<1>(Crow, ldc, 2);                          \
+        a2 = load_acc_2x2<1>(Crow, ldc, 4);                          \
+        a3 = load_acc_2x2<1>(Crow, ldc, 6);                          \
+      }                                                              \
+    } else {                                                         \
+      a0 = a1 = a2 = a3 = vdupq_n_f32(0.f);                          \
+    }                                                                \
+  } while (0)
+
+// Store 4 column-pair accumulators back to C matrix
+#define STORE_ACC_ROWPAIR_4(a0, a1, a2, a3, Crow, ldc, m_rows) \
+  do {                                                         \
+    if (m_rows == 2) {                                         \
+      store_acc_2x2<2>(a0, Crow, ldc, 0);                      \
+      store_acc_2x2<2>(a1, Crow, ldc, 2);                      \
+      store_acc_2x2<2>(a2, Crow, ldc, 4);                      \
+      store_acc_2x2<2>(a3, Crow, ldc, 6);                      \
+    } else {                                                   \
+      store_acc_2x2<1>(a0, Crow, ldc, 0);                      \
+      store_acc_2x2<1>(a1, Crow, ldc, 2);                      \
+      store_acc_2x2<1>(a2, Crow, ldc, 4);                      \
+      store_acc_2x2<1>(a3, Crow, ldc, 6);                      \
+    }                                                          \
+  } while (0)
+
+// Perform 4 BFMMLA operations: acc += A @ B for 4 column-pairs
+#define BFMMLA_COMPUTE_4(r0, r1, r2, r3, a, b0, b1, b2, b3) \
+  do {                                                      \
+    r0 = vbfmmlaq_f32(r0, a, b0);                           \
+    r1 = vbfmmlaq_f32(r1, a, b1);                           \
+    r2 = vbfmmlaq_f32(r2, a, b2);                           \
+    r3 = vbfmmlaq_f32(r3, a, b3);                           \
+  } while (0)
+
+// Micro-kernel: updates a small fixed tile using BFMMLA.
+// RP = number of row-pairs (1,2,4)
+// Computes C[TILE_ROWS*RP, OUTPUT_COLS_PER_BLOCK] += A_packed @ B.
+// A_packed interleaves RP row-pairs; B layout is driven by the attention phase:
+// - AttentionGemmPhase::QK -> token-column layout (Q @ K^T)
+// - AttentionGemmPhase::PV -> token-row layout (P @ V)
+// K_static < 0 enables runtime K (PV only)
+template <int32_t RP, int32_t K_static, AttentionGemmPhase phase>
+FORCE_INLINE void gemm_rowpairs_x8_bfmmla_neon(
+    const bfloat16_t* const* __restrict A_packed_rp,
+    const int32_t* __restrict m_rows_rp, const bfloat16_t* __restrict B_blk,
+    float* __restrict C, int64_t ldc, bool accumulate, int64_t b_stride,
+    int32_t K_runtime = 0) {
+  static_assert(RP == 1 || RP == 2 || RP == 4, "RP must be 1,2,4");
+  static_assert(K_static < 0 || K_static % TILE_K == 0,
+                "K must be divisible by TILE_K");
+  static_assert(K_static >= 0 || phase == AttentionGemmPhase::PV,
+                "Runtime K only supported for PV");
+
+  constexpr bool runtime_k = (K_static < 0);
+  const int32_t K_iters =
+      runtime_k ? (K_runtime / TILE_K) : (K_static / TILE_K);
+  const int32_t K_tail = runtime_k ? (K_runtime % TILE_K) : 0;
+
+  if (!runtime_k) {
+    // Help the compiler fold away unused K_runtime when K is compile-time
+    (void)K_runtime;
+  }
+
+  auto* C_al = C;
+  const auto* B_al = B_blk;
+
+  // Setup A pointers
+  const bfloat16_t* a_ptr[4] = {
+      A_packed_rp[0],
+      (RP >= 2) ? A_packed_rp[1] : nullptr,
+      (RP >= 4) ? A_packed_rp[2] : nullptr,
+      (RP >= 4) ? A_packed_rp[3] : nullptr,
+  };
+
+  // Setup B pointers based on layout
+  const bfloat16_t* b_ptr[4];
+  if constexpr (phase == AttentionGemmPhase::PV) {
+    b_ptr[0] = B_blk + 0 * b_stride;
+    b_ptr[1] = B_blk + 1 * b_stride;
+    b_ptr[2] = B_blk + 2 * b_stride;
+    b_ptr[3] = B_blk + 3 * b_stride;
+  }
+
+  float32x4_t acc[4][4];
+
+// Initialize accumulators
+#define INIT_RP(rp)                                                            \
+  if constexpr (RP > rp) {                                                     \
+    INIT_ACC_ROWPAIR_4(acc[rp][0], acc[rp][1], acc[rp][2], acc[rp][3],         \
+                       C_al + (rp * 2) * ldc, ldc, m_rows_rp[rp], accumulate); \
+  }
+  INIT_RP(0);
+  INIT_RP(1);
+  INIT_RP(2);
+  INIT_RP(3);
+#undef INIT_RP
+
+  // Main compute loop
+  for (int32_t ki = 0; ki < K_iters; ++ki) {
+    bfloat16x8_t b0, b1, b2, b3;
+    if constexpr (phase == AttentionGemmPhase::PV) {
+      b0 = vld1q_bf16(b_ptr[0] + ki * V_INNER_STRIDE);
+      b1 = vld1q_bf16(b_ptr[1] + ki * V_INNER_STRIDE);
+      b2 = vld1q_bf16(b_ptr[2] + ki * V_INNER_STRIDE);
+      b3 = vld1q_bf16(b_ptr[3] + ki * V_INNER_STRIDE);
+    } else {
+      const bfloat16_t* b_base = B_al + ki * b_stride;
+      b0 = vld1q_bf16(b_base + 0 * V_INNER_STRIDE);
+      b1 = vld1q_bf16(b_base + 1 * V_INNER_STRIDE);
+      b2 = vld1q_bf16(b_base + 2 * V_INNER_STRIDE);
+      b3 = vld1q_bf16(b_base + 3 * V_INNER_STRIDE);
+    }
+
+#define COMPUTE_RP(rp)                                                       \
+  if constexpr (RP > rp) {                                                   \
+    bfloat16x8_t a = vld1q_bf16(a_ptr[rp] + ki * PACK_ELEMENTS_PER_K_CHUNK); \
+    BFMMLA_COMPUTE_4(acc[rp][0], acc[rp][1], acc[rp][2], acc[rp][3], a, b0,  \
+                     b1, b2, b3);                                            \
+  }
+    COMPUTE_RP(0);
+    COMPUTE_RP(1);
+    COMPUTE_RP(2);
+    COMPUTE_RP(3);
+#undef COMPUTE_RP
+  }
+
+  // K tail for runtime PV: fallback path
+  if constexpr (runtime_k) {
+    if (K_tail > 0) {
+      const int32_t tail_offset = K_iters * V_INNER_STRIDE;
+      const int32_t a_tail_offset = K_iters * PACK_ELEMENTS_PER_K_CHUNK;
+      for (int32_t kt = 0; kt < K_tail; ++kt) {
+        float32x4_t b_vecs[4];
+        for (int32_t p = 0; p < 4; ++p) {
+          const bfloat16_t* bp = b_ptr[p] + tail_offset + kt * TILE_COLS;
+          const float b0 = vcvtah_f32_bf16(bp[0]);
+          const float b1 = vcvtah_f32_bf16(bp[1]);
+          const float32x2_t b_pair = vset_lane_f32(b1, vdup_n_f32(b0), 1);
+          b_vecs[p] = vcombine_f32(b_pair, b_pair);
+        }
+
+#define TAIL_RP(rp)                                                     \
+  if constexpr (RP > rp) {                                              \
+    const bfloat16_t* ap = A_packed_rp[rp] + a_tail_offset;             \
+    float a_row0 = vcvtah_f32_bf16(ap[kt]);                             \
+    float a_row1 =                                                      \
+        (m_rows_rp[rp] == 2) ? vcvtah_f32_bf16(ap[kt + TILE_K]) : 0.0f; \
+    const float32x4_t a_vec =                                           \
+        vcombine_f32(vdup_n_f32(a_row0), vdup_n_f32(a_row1));           \
+    for (int32_t p = 0; p < 4; ++p) {                                   \
+      acc[rp][p] = vmlaq_f32(acc[rp][p], a_vec, b_vecs[p]);             \
+    }                                                                   \
+  }
+        TAIL_RP(0);
+        TAIL_RP(1);
+        TAIL_RP(2);
+        TAIL_RP(3);
+#undef TAIL_RP
+      }
+    }
+  }
+
+  // Store results
+#define STORE_RP(rp)                                                    \
+  if constexpr (RP > rp) {                                              \
+    STORE_ACC_ROWPAIR_4(acc[rp][0], acc[rp][1], acc[rp][2], acc[rp][3], \
+                        C_al + (rp * 2) * ldc, ldc, m_rows_rp[rp]);     \
+  }
+  STORE_RP(0);
+  STORE_RP(1);
+  STORE_RP(2);
+  STORE_RP(3);
+#undef STORE_RP
+}
+
+// Meso-kernel: packs a small MBxK slice of A, then tiles over N and calls the
+// micro-kernel for each OUTPUT_COLS_PER_BLOCK chunk. K_static < 0 enables
+// runtime K (PV only).
+template <int32_t MB, int32_t N, int32_t K_static, AttentionGemmPhase phase>
+FORCE_INLINE void gemm_packA_compute_MB_xN(
+    const c10::BFloat16* __restrict A, const c10::BFloat16* __restrict B,
+    float* __restrict C, int32_t K_runtime, int64_t lda, int64_t ldc,
+    int64_t b_layout_stride, int64_t b_reduction_stride, bool accumulate) {
+  static_assert(MB >= 1 && MB <= 8, "MB must be in [1,8]");
+  static_assert(N % OUTPUT_COLS_PER_BLOCK == 0,
+                "N must be a multiple of OUTPUT_COLS_PER_BLOCK");
+  static_assert(K_static < 0 || K_static % TILE_K == 0,
+                "K must be divisible by TILE_K");
+  static_assert(K_static >= 0 || phase == AttentionGemmPhase::PV,
+                "Runtime K only supported for PV");
+
+  constexpr bool runtime_k = (K_static < 0);
+  const int32_t K_val = runtime_k ? K_runtime : K_static;
+
+  // Keep small packs on-stack to avoid heap churn
+  constexpr int32_t STACK_PACK_STRIDE =
+      (1024 / TILE_K) * PACK_ELEMENTS_PER_K_CHUNK;
+
+  constexpr int32_t ROW_PAIRS = (MB + 1) / TILE_ROWS;
+  const int32_t pack_stride =
+      runtime_k ? ((K_val + TILE_K - 1) / TILE_K) * PACK_ELEMENTS_PER_K_CHUNK
+                : (K_static / TILE_K) * PACK_ELEMENTS_PER_K_CHUNK;
+
+  alignas(64) c10::BFloat16 A_packed_stack[ROW_PAIRS * STACK_PACK_STRIDE];
+  std::vector<c10::BFloat16> A_packed_heap;
+  c10::BFloat16* A_packed =
+      (pack_stride <= STACK_PACK_STRIDE)
+          ? A_packed_stack
+          : (A_packed_heap.resize(ROW_PAIRS * pack_stride),
+             A_packed_heap.data());
+
+  for (int32_t rp = 0; rp < ROW_PAIRS; ++rp) {
+    const int32_t m = rp * TILE_ROWS;
+    const int32_t m_rows = (m + 1 < MB) ? TILE_ROWS : 1;
+    const c10::BFloat16* A0 = A + m * lda;
+    const c10::BFloat16* A1 = (m_rows == TILE_ROWS) ? (A + (m + 1) * lda) : A0;
+    reshape_Q_2xK_for_bfmmla(A0, A1, A_packed + rp * pack_stride, K_val);
+  }
+
+  for (int32_t n = 0; n < N; n += OUTPUT_COLS_PER_BLOCK) {
+    const c10::BFloat16* B_blk_c10 =
+        (phase == AttentionGemmPhase::PV)
+            ? (B + (n / TILE_COLS) * b_layout_stride)
+            : (B + (n / OUTPUT_COLS_PER_BLOCK) * b_layout_stride);
+    const bfloat16_t* B_blk = reinterpret_cast<const bfloat16_t*>(B_blk_c10);
+
+    // Process row-pairs in groups of 4, 2, then 1
+    int32_t row_pair_idx = 0;
+
+#define PROCESS_RP_GROUP(group_size)                                       \
+  for (; row_pair_idx + (group_size - 1) < ROW_PAIRS;                      \
+       row_pair_idx += group_size) {                                       \
+    const bfloat16_t* Ap[group_size];                                      \
+    int32_t mr[group_size];                                                \
+    for (int32_t i = 0; i < group_size; ++i) {                             \
+      Ap[i] = reinterpret_cast<const bfloat16_t*>(                         \
+          A_packed + (row_pair_idx + i) * pack_stride);                    \
+      mr[i] = (((row_pair_idx + i) * TILE_ROWS + 1) < MB) ? TILE_ROWS : 1; \
+    }                                                                      \
+    float* C_blk = C + (row_pair_idx * TILE_ROWS) * ldc + n;               \
+    if constexpr (runtime_k) {                                             \
+      gemm_rowpairs_x8_bfmmla_neon<group_size, -1, phase>(                 \
+          Ap, mr, B_blk, C_blk, ldc, accumulate, b_layout_stride, K_val);  \
+    } else {                                                               \
+      gemm_rowpairs_x8_bfmmla_neon<group_size, K_static, phase>(           \
+          Ap, mr, B_blk, C_blk, ldc, accumulate,                           \
+          (phase == AttentionGemmPhase::PV) ? b_layout_stride              \
+                                            : b_reduction_stride);         \
+    }                                                                      \
+  }
+
+    PROCESS_RP_GROUP(4);
+    PROCESS_RP_GROUP(2);
+    PROCESS_RP_GROUP(1);
+#undef PROCESS_RP_GROUP
+  }
+}
+
+// Macro-kernel: iterates over M in MB={8,4,2,1} chunks.
+// Supports compile-time K specialization when K >= 0; otherwise uses runtime K
+// (runtime K path is only supported for PV).
+template <AttentionGemmPhase phase, int32_t N, int32_t K = -1>
+FORCE_INLINE void gemm_macro_neon_bfmmla(
+    const c10::BFloat16* __restrict A, const c10::BFloat16* __restrict B,
+    float* __restrict C, int32_t M, int32_t K_runtime, int64_t lda, int64_t ldc,
+    int64_t b_layout_stride, int64_t b_reduction_stride, bool accumulate) {
+  static_assert(N % OUTPUT_COLS_PER_BLOCK == 0,
+                "N must be a multiple of OUTPUT_COLS_PER_BLOCK");
+
+  if constexpr (K >= 0) {
+    static_assert(K % TILE_K == 0, "K must be divisible by TILE_K");
+    for (int32_t m = 0; m < M;) {
+      const int32_t rem = M - m;
+      const c10::BFloat16* A_blk = A + m * lda;
+      float* C_blk = C + m * ldc;
+
+#define DISPATCH_MB(mb)                                                   \
+  gemm_packA_compute_MB_xN<mb, N, K, phase>(A_blk, B, C_blk, 0, lda, ldc, \
+                                            b_layout_stride,              \
+                                            b_reduction_stride, accumulate)
+
+      if (rem >= 8) {
+        DISPATCH_MB(8);
+        m += 8;
+      } else if (rem >= 4) {
+        DISPATCH_MB(4);
+        m += 4;
+      } else if (rem >= 2) {
+        DISPATCH_MB(2);
+        m += 2;
+      } else {
+        DISPATCH_MB(1);
+        m += 1;
+      }
+#undef DISPATCH_MB
+    }
+  } else {
+    static_assert(phase == AttentionGemmPhase::PV,
+                  "Runtime K specialization only supported for PV.");
+    const int32_t K_val = K_runtime;
+
+    for (int32_t m = 0; m < M;) {
+      const int32_t rem = M - m;
+      const c10::BFloat16* A_blk = A + m * lda;
+      float* C_blk = C + m * ldc;
+
+#define DISPATCH_MB_RUNTIME(mb)                                                \
+  gemm_packA_compute_MB_xN<mb, N, -1, phase>(A_blk, B, C_blk, K_val, lda, ldc, \
+                                             b_layout_stride,                  \
+                                             b_reduction_stride, accumulate)
+
+      if (rem >= 8) {
+        DISPATCH_MB_RUNTIME(8);
+        m += 8;
+      } else if (rem >= 4) {
+        DISPATCH_MB_RUNTIME(4);
+        m += 4;
+      } else if (rem >= 2) {
+        DISPATCH_MB_RUNTIME(2);
+        m += 2;
+      } else {
+        DISPATCH_MB_RUNTIME(1);
+        m += 1;
+      }
+#undef DISPATCH_MB_RUNTIME
+    }
+  }
+}
+
+#undef INIT_ACC_ROWPAIR_4
+#undef STORE_ACC_ROWPAIR_4
+#undef BFMMLA_COMPUTE_4
+
+}  // namespace
+
+// TileGemm Adapter for Attention
+
+template <typename kv_cache_t, int32_t BlockTokens, int32_t HeadDim>
+class TileGemmNEONBFMMLA {
+ public:
+  template <AttentionGemmPhase phase, int32_t head_dim_ct>
+  FORCE_INLINE static void gemm(const int32_t m_size, void* __restrict__ a_tile,
+                                kv_cache_t* __restrict__ b_tile,
+                                float* __restrict__ c_tile, const int64_t lda,
+                                [[maybe_unused]] const int64_t ldb,
+                                const int64_t ldc,
+                                [[maybe_unused]] const int32_t block_size,
+                                [[maybe_unused]] const int32_t dynamic_k_size,
+                                const bool accum_c) {
+    static_assert(BlockTokens % OUTPUT_COLS_PER_BLOCK == 0);
+    // BFMMLA kernels require compile-time head_dim; keep head_dim_ct only for
+    // API parity with other tile_gemm implementations.
+    if constexpr (head_dim_ct >= 0) {
+      static_assert(head_dim_ct == HeadDim,
+                    "BFMMLA expects head_dim_ct to match HeadDim; PV passes "
+                    "-1 for API parity.");
+    }
+
+    if constexpr (phase == AttentionGemmPhase::QK) {
+      const int64_t b_reduction_stride = K_INNER_STRIDE;
+      const int64_t b_token_block_stride = (HeadDim / TILE_K) * K_INNER_STRIDE;
+
+      gemm_macro_neon_bfmmla<AttentionGemmPhase::QK, BlockTokens, HeadDim>(
+          reinterpret_cast<const c10::BFloat16*>(a_tile), b_tile, c_tile,
+          m_size, 0, lda, ldc, b_token_block_stride, b_reduction_stride,
+          accum_c);
+    } else {
+      const int64_t b_pair_stride =
+          (block_size / V_TOKENS_PER_ROW_BLOCK) * V_INNER_STRIDE;
+
+      // PV gemm with runtime K specialization
+      switch (dynamic_k_size) {
+        case 32:
+          gemm_macro_neon_bfmmla<AttentionGemmPhase::PV, HeadDim, 32>(
+              reinterpret_cast<const c10::BFloat16*>(a_tile), b_tile, c_tile,
+              m_size, 32, lda, ldc, b_pair_stride, 0, accum_c);
+          break;
+        case 128:
+          gemm_macro_neon_bfmmla<AttentionGemmPhase::PV, HeadDim, 128>(
+              reinterpret_cast<const c10::BFloat16*>(a_tile), b_tile, c_tile,
+              m_size, 128, lda, ldc, b_pair_stride, 0, accum_c);
+          break;
+        case 256:
+          gemm_macro_neon_bfmmla<AttentionGemmPhase::PV, HeadDim, 256>(
+              reinterpret_cast<const c10::BFloat16*>(a_tile), b_tile, c_tile,
+              m_size, 256, lda, ldc, b_pair_stride, 0, accum_c);
+          break;
+        default:
+          gemm_macro_neon_bfmmla<AttentionGemmPhase::PV, HeadDim>(
+              reinterpret_cast<const c10::BFloat16*>(a_tile), b_tile, c_tile,
+              m_size, dynamic_k_size, lda, ldc, b_pair_stride, 0, accum_c);
+          break;
+      }
+    }
+  }
+};
+
+// Shared ASIMD BFMMLA implementation (BF16 only). The block size alignment and
+// ISA tag are template parameters so we can reuse the same kernels for
+// different NEON configurations.
+template <int64_t block_size_alignment, ISA isa_type, int64_t head_dim>
+class AttentionImplNEONBFMMLA {
+ public:
+  using query_t = c10::BFloat16;
+  using q_buffer_t = c10::BFloat16;
+  using kv_cache_t = c10::BFloat16;
+  using logits_buffer_t = float;
+  using partial_output_buffer_t = float;
+  using prob_buffer_t = c10::BFloat16;
+
+  static constexpr int64_t BlockSizeAlignment = block_size_alignment;
+  // HeadDimAlignment equals head_dim so that the PV phase processes
+  // the full head dimension in a single gemm call.
+  static constexpr int64_t HeadDimAlignment = head_dim;
+  static constexpr int64_t MaxQHeadNumPerIteration = 16;
+  static constexpr int64_t HeadDim = head_dim;
+  static constexpr ISA ISAType = isa_type;
+  static constexpr bool scale_on_logits = false;
+
+  static_assert(HeadDim % OUTPUT_COLS_PER_BLOCK == 0);
+  static_assert(BlockSizeAlignment % OUTPUT_COLS_PER_BLOCK == 0);
+  static_assert(HeadDim % TILE_K == 0, "HeadDim must be a multiple of TILE_K");
+
+ public:
+  template <template <typename tile_gemm_t> typename attention>
+  FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
+    attention<
+        TileGemmNEONBFMMLA<kv_cache_t, static_cast<int32_t>(BlockSizeAlignment),
+                           static_cast<int32_t>(HeadDim)>>
+        attention_iteration;
+    attention_iteration(CPU_ATTENTION_PARAMS);
+  }
+
+  // Key cache stride per token group (TokenColumn layout; QK)
+  static constexpr int64_t k_cache_token_group_stride(
+      [[maybe_unused]] const int32_t block_size) {
+    static_assert(BlockSizeAlignment % K_TOKENS_PER_GROUP == 0);
+    return (BlockSizeAlignment / K_TOKENS_PER_GROUP) *
+           ((head_dim / TILE_K) * K_INNER_STRIDE);
+  }
+
+  // Value cache stride per token group (TokenRow layout; PV)
+  static constexpr int64_t v_cache_token_group_stride(
+      [[maybe_unused]] const int32_t block_size) {
+    static_assert(BlockSizeAlignment % V_TOKENS_PER_ROW_BLOCK == 0);
+    return (BlockSizeAlignment / V_TOKENS_PER_ROW_BLOCK) * V_INNER_STRIDE;
+  }
+
+  // The stride to move to the "next" head_dim group
+  // is the full V cache size per head, since HeadDimAlignment == head_dim.
+  // Hence, the stride is not used in this case
+  static constexpr int64_t v_cache_head_group_stride(
+      [[maybe_unused]] const int32_t block_size) {
+    return head_dim * block_size;
+  }
+
+  // Convert Q heads to BF16 and apply scale factor using native BF16 intrinsics
+  static void copy_q_heads_tile(c10::BFloat16* __restrict__ src,
+                                c10::BFloat16* __restrict__ q_buffer,
+                                const int32_t q_num,
+                                const int32_t q_heads_per_kv,
+                                const int64_t q_num_stride,
+                                const int64_t q_head_stride, float scale) {
+    constexpr int32_t dim = static_cast<int32_t>(head_dim);
+    const float32x4_t scale_vec = vdupq_n_f32(scale);
+
+    for (int32_t qi = 0; qi < q_num; ++qi) {
+      for (int32_t hi = 0; hi < q_heads_per_kv; ++hi) {
+        c10::BFloat16* __restrict__ curr_q =
+            src + qi * q_num_stride + hi * q_head_stride;
+        c10::BFloat16* __restrict__ dst =
+            q_buffer + qi * q_heads_per_kv * head_dim + hi * head_dim;
+
+        for (int32_t i = 0; i < dim; i += OUTPUT_COLS_PER_BLOCK) {
+          bfloat16x8_t in8 =
+              vld1q_bf16(reinterpret_cast<const bfloat16_t*>(curr_q + i));
+          float32x4_t lo = vmulq_f32(vcvtq_low_f32_bf16(in8), scale_vec);
+          float32x4_t hi = vmulq_f32(vcvtq_high_f32_bf16(in8), scale_vec);
+
+          bfloat16x4_t lo_b = vcvt_bf16_f32(lo);
+          bfloat16x4_t hi_b = vcvt_bf16_f32(hi);
+          bfloat16x8_t out = vcombine_bf16(lo_b, hi_b);
+          vst1q_bf16(reinterpret_cast<bfloat16_t*>(dst + i), out);
+        }
+      }
+    }
+  }
+
+ public:
+  // Reshape and cache K/V into BFMMLA-optimized layouts
+  // K cache:
+  // [block_size/K_TOKENS_PER_GROUP][head_dim/TILE_K][K_INNER_STRIDE]
+  // - TokenColumn
+  // V cache:
+  // [head_dim/TILE_COLS][block_size/V_TOKENS_PER_ROW_BLOCK][V_INNER_STRIDE]
+  // - TokenRows
+  static void reshape_and_cache(
+      const c10::BFloat16* __restrict__ key,
+      const c10::BFloat16* __restrict__ value,
+      c10::BFloat16* __restrict__ key_cache,
+      c10::BFloat16* __restrict__ value_cache,
+      const int64_t* __restrict__ slot_mapping, const int64_t token_num,
+      const int64_t key_token_num_stride, const int64_t value_token_num_stride,
+      const int64_t head_num, const int64_t key_head_num_stride,
+      const int64_t value_head_num_stride,
+      [[maybe_unused]] const int64_t num_blocks,
+      const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
+      const int64_t block_size,
+      [[maybe_unused]] const int64_t block_size_stride) {
+    const int64_t k_block_stride = (head_dim / TILE_K) * K_INNER_STRIDE;
+    const int64_t v_pair_stride =
+        (block_size / V_TOKENS_PER_ROW_BLOCK) * V_INNER_STRIDE;
+
+#pragma omp parallel for
+    for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
+      for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
+        const int64_t pos = slot_mapping[token_idx];
+        if (pos < 0) continue;
+
+        const int64_t block_idx = pos / block_size;
+        const int64_t block_offset = pos % block_size;
+
+        // Key cache: TokenColumn QK
+        {
+          const c10::BFloat16* __restrict key_src =
+              key + token_idx * key_token_num_stride +
+              head_idx * key_head_num_stride;
+
+          c10::BFloat16* __restrict key_base = key_cache +
+                                               block_idx * num_blocks_stride +
+                                               head_idx * cache_head_num_stride;
+
+          const int64_t block_in_block = block_offset / K_TOKENS_PER_GROUP;
+          const int64_t pair_in_block =
+              (block_offset % K_TOKENS_PER_GROUP) / TILE_COLS;
+          const int64_t lane_base = (block_offset & 1) ? TILE_K : 0;
+
+          c10::BFloat16* __restrict block_base =
+              key_base + block_in_block * k_block_stride;
+
+          for (int64_t hd4 = 0; hd4 < head_dim / TILE_K; ++hd4) {
+            uint16_t* dst_u16 = reinterpret_cast<uint16_t*>(
+                block_base + hd4 * K_INNER_STRIDE +
+                pair_in_block * V_INNER_STRIDE + lane_base);
+            const uint16_t* src_u16 =
+                reinterpret_cast<const uint16_t*>(key_src + hd4 * TILE_K);
+            vst1_u16(dst_u16, vld1_u16(src_u16));
+          }
+        }
+
+        // Value cache: TokenRow PV
+        {
+          const c10::BFloat16* __restrict value_src =
+              value + token_idx * value_token_num_stride +
+              head_idx * value_head_num_stride;
+
+          c10::BFloat16* __restrict value_base =
+              value_cache + block_idx * num_blocks_stride +
+              head_idx * cache_head_num_stride;
+
+          const int64_t row_block = block_offset / V_TOKENS_PER_ROW_BLOCK;
+          const int64_t lane = block_offset & (V_TOKENS_PER_ROW_BLOCK - 1);
+
+          c10::BFloat16* __restrict row_block_base =
+              value_base + row_block * V_INNER_STRIDE;
+
+          for (int64_t hd2 = 0; hd2 < head_dim / TILE_COLS; ++hd2) {
+            c10::BFloat16* __restrict dst_val =
+                row_block_base + hd2 * v_pair_stride;
+
+            const uint16_t* src_u16 =
+                reinterpret_cast<const uint16_t*>(value_src);
+            uint16_t* dst_u16 = reinterpret_cast<uint16_t*>(dst_val);
+            dst_u16[lane] = src_u16[hd2 * TILE_COLS + 0];
+            dst_u16[lane + V_TOKENS_PER_ROW_BLOCK] =
+                src_u16[hd2 * TILE_COLS + 1];
+          }
+        }
+      }
+    }
+  }
+};
+
+}  // namespace cpu_attention
+
+#endif  // CPU_ATTN_ASIMD_BFMMLA_HPP

csrc/cpu/cpu_fused_moe.cpp

@@ -147,7 +147,7 @@ void fused_moe_impl(scalar_t* __restrict__ output, scalar_t* __restrict__ input,
                     const int32_t token_num, const int32_t expert_num,
                     const int32_t topk_num, const int32_t input_size_13,
                     const int32_t output_size_13, const int32_t input_size_2,
-                    const int32_t output_size_2) {
+                    const int32_t output_size_2, const bool skip_weighted) {
   using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
   constexpr int32_t gemm_n_tile_size = gemm_t::NSize;
   constexpr int32_t gemm_m_tile_size = gemm_t::MaxMSize;
@@ -582,6 +582,11 @@ void fused_moe_impl(scalar_t* __restrict__ output, scalar_t* __restrict__ input,
         scalar_t* __restrict__ curr_output_buffer =
             output + token_id * output_size_2;
 
+        if (skip_weighted) {
+          // Only for topk_num == 1
+          *curr_weight = 1.0f;
+        }
+
         if (topk_num > 1) {
           {
             int32_t w2_output_idx = curr_expand_token_id_index_buffer[0];
@@ -699,7 +704,7 @@ void cpu_fused_moe(
     const std::optional<torch::Tensor>& w2_bias,  // [expert_num, output_size_2]
     const torch::Tensor& topk_weights,            // [token_num, k], float32
     const torch::Tensor& topk_id,                 // [token_num, k], int32
-    const std::string& act, const std::string& isa) {
+    const bool skip_weighted, const std::string& act, const std::string& isa) {
   const int32_t token_num = input.size(0);
   const int32_t input_size_13 = input.size(1);
   const int64_t input_stride = input.stride(0);
@@ -711,6 +716,8 @@ void cpu_fused_moe(
   const int32_t topk_num = topk_id.size(1);
   const FusedMOEAct act_type = get_act_type(act);
   cpu_utils::ISA isa_type = cpu_utils::get_isa(isa);
+  TORCH_CHECK(!skip_weighted || topk_num == 1,
+              "skip_weighted is only supported for topk=1 on CPU");
 
   VLLM_DISPATCH_FLOATING_TYPES(w13.scalar_type(), "cpu_fused_moe", [&]() {
     CPU_ISA_DISPATCH_IMPL(isa_type, [&]() {
@@ -721,7 +728,7 @@ void cpu_fused_moe(
           w2_bias.has_value() ? w2_bias->data_ptr<scalar_t>() : nullptr,
           topk_weights.data_ptr<float>(), topk_id.data_ptr<int32_t>(), act_type,
           token_num, expert_num, topk_num, input_size_13, output_size_13,
-          input_size_2, output_size_2);
+          input_size_2, output_size_2, skip_weighted);
     });
   });
 }

csrc/cpu/cpu_types_vxe.hpp

@@ -16,10 +16,12 @@ namespace vec_op {
 #define vec_sr(a, b) ((a) >> (b))  // Vector Shift Right Algebraic
 #define vec_sl(a, b) ((a) << (b))  // Vector Shift Left
 
-// FIXME: FP16 is not fully supported in Torch-CPU
-#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
-  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
-  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+// NOTE: FP16 (Half) is supported on s390x via custom bit-manipulation
+// conversion. PyTorch itself lacks native s390x FP16 support.
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)            \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)    \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
 
 #define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
   AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
@@ -86,6 +88,39 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
   }
 };
 
+struct FP16Vec8 : public Vec<FP16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+
+  __vector signed short reg;
+
+  explicit FP16Vec8(const void* ptr) : reg(*(__vector signed short*)ptr) {}
+  explicit FP16Vec8(const FP32Vec8&);
+
+  void save(void* ptr) const {
+    *reinterpret_cast<__vector signed short*>(ptr) = reg;
+  }
+};
+
+struct FP16Vec16 : public Vec<FP16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+
+  ss16x8x2_t reg;
+
+  explicit FP16Vec16(const void* ptr) {
+    // Load 256 bits (16 FP16 values) in two parts
+    reg.val[0] = (__vector signed short)vec_xl(0, (signed short*)ptr);
+    reg.val[1] = (__vector signed short)vec_xl(16, (signed short*)ptr);
+  }
+
+  explicit FP16Vec16(const FP32Vec16&);
+
+  void save(void* ptr) const {
+    // Save 256 bits in two parts
+    vec_xst(reg.val[0], 0, (signed short*)ptr);
+    vec_xst(reg.val[1], 16, (signed short*)ptr);
+  }
+};
+
 struct BF16Vec16 : public Vec<BF16Vec16> {
   constexpr static int VEC_ELEM_NUM = 16;
 
@@ -108,6 +143,92 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
 
 const static __vector signed short zero = vec_splats((signed short)0);
 
+FORCE_INLINE __vector float fp16_to_fp32_bits(__vector unsigned int x) {
+  const __vector unsigned int mask_sign = {0x8000, 0x8000, 0x8000, 0x8000};
+  const __vector unsigned int mask_exp = {0x7C00, 0x7C00, 0x7C00, 0x7C00};
+  const __vector unsigned int mask_mant = {0x03FF, 0x03FF, 0x03FF, 0x03FF};
+  const __vector unsigned int bias_adj = {112, 112, 112, 112};
+  const __vector unsigned int exp_max_fp16 = {0x1F, 0x1F, 0x1F,
+                                              0x1F};  // FP16 NaN/Inf exponent
+  const __vector unsigned int exp_max_fp32 = {0xFF, 0xFF, 0xFF,
+                                              0xFF};  // FP32 NaN/Inf exponent
+
+  __vector unsigned int s = (x & mask_sign) << 16;
+  __vector unsigned int e = (x & mask_exp) >> 10;
+  __vector unsigned int m = (x & mask_mant) << 13;
+
+  // Check for NaN/Inf: exponent = 0x1F in FP16
+  __vector __bool int is_nan_inf = vec_cmpeq(e, exp_max_fp16);
+
+  // Normal: adjust bias; NaN/Inf: set to 0xFF
+  __vector unsigned int e_normal = e + bias_adj;
+  e = vec_sel(e_normal, exp_max_fp32, is_nan_inf);
+
+  return (__vector float)(s | (e << 23) | m);
+}
+
+FORCE_INLINE __vector unsigned int fp32_to_fp16_bits(__vector float f_in) {
+  __vector unsigned int in = (__vector unsigned int)f_in;
+
+  const __vector unsigned int mask_sign_32 = {0x80000000, 0x80000000,
+                                              0x80000000, 0x80000000};
+  const __vector unsigned int mask_exp_32 = {0x7F800000, 0x7F800000, 0x7F800000,
+                                             0x7F800000};
+  const __vector unsigned int mask_mant_32 = {0x007FFFFF, 0x007FFFFF,
+                                              0x007FFFFF, 0x007FFFFF};
+
+  // Use SIGNED integers for exponent math to handle underflow check
+  const __vector signed int bias_adj = {112, 112, 112, 112};
+  const __vector signed int zero = {0, 0, 0, 0};
+  const __vector signed int max_exp = {31, 31, 31, 31};  // Max FP16 exp
+  const __vector unsigned int exp_max_fp32 = {0xFF, 0xFF, 0xFF, 0xFF};
+  const __vector unsigned int exp_max_fp16 = {0x1F, 0x1F, 0x1F, 0x1F};
+
+  __vector unsigned int s = (in & mask_sign_32) >> 16;
+  __vector unsigned int e_u = (in & mask_exp_32) >> 23;
+
+  // Check for NaN/Inf: exponent = 0xFF in FP32
+  __vector __bool int is_nan_inf = vec_cmpeq(e_u, exp_max_fp32);
+
+  __vector signed int e_s = (__vector signed int)e_u;
+  e_s = vec_sub(e_s, bias_adj);
+  e_s = vec_max(e_s, zero);
+  e_s = vec_min(e_s, max_exp);
+  __vector unsigned int e_normal = (__vector unsigned int)e_s;
+
+  __vector unsigned int e_final = vec_sel(e_normal, exp_max_fp16, is_nan_inf);
+
+  const __vector unsigned int one_v = {1, 1, 1, 1};
+  const __vector unsigned int mask_sticky = {0xFFF, 0xFFF, 0xFFF, 0xFFF};
+
+  __vector unsigned int round_bit = (in >> 12) & one_v;
+  __vector unsigned int sticky = in & mask_sticky;
+  __vector unsigned int m = (in & mask_mant_32) >> 13;
+  __vector unsigned int lsb = m & one_v;  // LSB of mantissa for tie-breaking
+
+  // Round up if: round_bit && (sticky || lsb)
+  __vector __bool int sticky_nonzero =
+      vec_cmpgt(sticky, (__vector unsigned int){0, 0, 0, 0});
+  __vector __bool int lsb_set = vec_cmpeq(lsb, one_v);
+  __vector __bool int round_up =
+      vec_and(vec_cmpeq(round_bit, one_v), vec_or(sticky_nonzero, lsb_set));
+
+  m = vec_sel(m, m + one_v, round_up);
+
+  const __vector unsigned int mant_mask = {0x3FF, 0x3FF, 0x3FF, 0x3FF};
+  const __vector unsigned int max_normal_exp = {0x1E, 0x1E, 0x1E, 0x1E};
+  __vector __bool int mant_overflows = vec_cmpgt(m, mant_mask);
+  __vector __bool int would_overflow_to_inf =
+      vec_and(mant_overflows, vec_cmpeq(e_final, max_normal_exp));
+  __vector unsigned int e_inc = vec_min(e_final + one_v, exp_max_fp16);
+  e_final = vec_sel(e_final, e_inc, mant_overflows);
+  m = vec_and(m, mant_mask);
+  e_final = vec_sel(e_final, max_normal_exp, would_overflow_to_inf);
+  m = vec_sel(m, mant_mask, would_overflow_to_inf);
+
+  return s | (e_final << 10) | m;
+}
+
 struct BF16Vec32 : public Vec<BF16Vec32> {
   constexpr static int VEC_ELEM_NUM = 32;
 
@@ -180,6 +301,18 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
     reg.val[1] = (__vector float)vec_mergel(v.reg, zero);
   }
 
+  explicit FP32Vec8(const FP16Vec8& v) {
+    // Cast to UNSIGNED short vector to prevent sign-extension during unpack
+    __vector unsigned short raw_u = (__vector unsigned short)v.reg;
+
+    // Unpack 8x16-bit to two 4x32-bit vectors (Zero extended)
+    __vector unsigned int raw_hi = (__vector unsigned int)vec_unpackh(raw_u);
+    __vector unsigned int raw_lo = (__vector unsigned int)vec_unpackl(raw_u);
+
+    reg.val[0] = fp16_to_fp32_bits(raw_hi);
+    reg.val[1] = fp16_to_fp32_bits(raw_lo);
+  }
+
   float reduce_sum() const {
     AliasReg ar;
     ar.reg = reg;
@@ -531,6 +664,22 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     reg.val[3] = (__vector float)vec_mergel(v.reg.val[1], zero);
   }
 
+  explicit FP32Vec16(const FP16Vec16& v) {
+    __vector unsigned int raw_hi_0 =
+        (__vector unsigned int)vec_unpackh(v.reg.val[0]);
+    __vector unsigned int raw_lo_0 =
+        (__vector unsigned int)vec_unpackl(v.reg.val[0]);
+    reg.val[0] = fp16_to_fp32_bits(raw_hi_0);
+    reg.val[1] = fp16_to_fp32_bits(raw_lo_0);
+
+    __vector unsigned int raw_hi_1 =
+        (__vector unsigned int)vec_unpackh(v.reg.val[1]);
+    __vector unsigned int raw_lo_1 =
+        (__vector unsigned int)vec_unpackl(v.reg.val[1]);
+    reg.val[2] = fp16_to_fp32_bits(raw_hi_1);
+    reg.val[3] = fp16_to_fp32_bits(raw_lo_1);
+  }
+
   explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
 
   FP32Vec16 operator*(const FP32Vec16& b) const {
@@ -628,8 +777,10 @@ struct VecType<c10::BFloat16> {
   using vec_type = BF16Vec8;
 };
 
-// On s390x, FP16 (Half) is not natively supported, use FP32 vectors instead
-using FP16Vec16 = FP32Vec16;
+template <>
+struct VecType<c10::Half> {
+  using vec_type = FP16Vec8;
+};
 
 template <typename T>
 void storeFP32(float v, T* ptr) {
@@ -650,6 +801,52 @@ inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
   *ptr = *(v_ptr + 1);
 }
 
+template <>
+inline void storeFP32<::c10::Half>(float v, ::c10::Half* ptr) {
+  // Use bit-manipulation for IEEE FP32 to FP16 conversion since vector
+  // intrinsics for FP32 to FP16 conversion does not use IEEE rounding and can
+  // produce incorrect results for some inputs. Process each of the 4 vectors
+  // separately.
+  uint32_t in;
+  std::memcpy(&in, &v, sizeof(in));
+
+  uint32_t s = (in & 0x80000000) >> 16;  // Sign
+  uint32_t e = (in & 0x7F800000) >> 23;  // Exponent
+  uint32_t round_bit = (in >> 12) & 1;
+  uint32_t sticky = (in & 0xFFF) != 0;  // Any bits in [11..0]
+  uint32_t m = (in & 0x007FFFFF) >> 13;
+  uint32_t lsb = m & 1;  // LSB of mantissa for tie-breaking
+
+  // Check for NaN/Inf before rounding
+  bool is_nan_inf = (e == 0xFF);
+
+  if (round_bit && (sticky || lsb)) {
+    m++;
+    // Handle mantissa overflow: if m overflows 10 bits, increment exponent
+    if (m > 0x3FF) {
+      m = 0;
+      e++;
+    }
+  }
+
+  if (is_nan_inf) {
+    // NaN/Inf: preserve it
+    e = 0x1F;
+  } else {
+    // Normal: adjust bias (127 - 15), flush subnormals to zero
+    e = (e >= 112) ? (e - 112) : 0;
+    // If exponent overflows to Inf range, saturate to max normal FP16 value
+    if (e > 0x1E) {
+      e = 0x1E;   // Max normal exponent
+      m = 0x3FF;  // Max mantissa
+    }
+  }
+
+  uint16_t fp16 = (uint16_t)(s | (e << 10) | m);
+
+  *reinterpret_cast<uint16_t*>(ptr) = fp16;
+}
+
 #ifndef __VEC_CLASS_FP_NAN
   #define __VEC_CLASS_FP_NAN (1 << 6)
 #endif
@@ -803,6 +1000,44 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
   reg.val[1] = (__vector signed short)vec_perm(inp2, inp3, omask);
 }
 
+inline FP16Vec8::FP16Vec8(const FP32Vec8& v) {
+  // Use bit-manipulation for IEEE FP32 to FP16 conversion since vector
+  // intrinsics for FP32 to FP16 conversion does not use IEEE rounding and can
+  // produce incorrect results for some inputs. Process each of the 4 vectors
+  // separately.
+  __vector unsigned int res_hi = fp32_to_fp16_bits(v.reg.val[0]);
+  __vector unsigned int res_lo = fp32_to_fp16_bits(v.reg.val[1]);
+
+  const __vector unsigned char perm_pack = {
+      2,  3,  6,  7,  10, 11, 14, 15,  // Select lower 2 bytes from res_hi
+      18, 19, 22, 23, 26, 27, 30, 31   // Select lower 2 bytes from res_lo
+  };
+
+  reg = vec_perm((__vector signed short)res_hi, (__vector signed short)res_lo,
+                 perm_pack);
+}
+
+inline FP16Vec16::FP16Vec16(const FP32Vec16& v) {
+  // Use bit-manipulation for IEEE FP32 to FP16 conversion since vector
+  // intrinsics for FP32 to FP16 conversion does not use IEEE rounding and can
+  // produce incorrect results for some inputs. Process each of the 4 vectors
+  // separately.
+  __vector unsigned int res_0 = fp32_to_fp16_bits(v.reg.val[0]);
+  __vector unsigned int res_1 = fp32_to_fp16_bits(v.reg.val[1]);
+  __vector unsigned int res_2 = fp32_to_fp16_bits(v.reg.val[2]);
+  __vector unsigned int res_3 = fp32_to_fp16_bits(v.reg.val[3]);
+
+  const __vector unsigned char perm_pack = {
+      2,  3,  6,  7,  10, 11, 14, 15,  // Lower 2 bytes from first vector
+      18, 19, 22, 23, 26, 27, 30, 31   // Lower 2 bytes from second vector
+  };
+
+  reg.val[0] = vec_perm((__vector signed short)res_0,
+                        (__vector signed short)res_1, perm_pack);
+  reg.val[1] = vec_perm((__vector signed short)res_2,
+                        (__vector signed short)res_3, perm_pack);
+}
+
 // 1D softmax over `n` elements in `input`, writes result to `output`.
 // Uses FP32Vec8 for main body, scalar tail handling.
 // Requirement: n > 0

csrc/cpu/dnnl_helper.cpp

@@ -237,12 +237,20 @@ W8A8MatMulPrimitiveHandler::W8A8MatMulPrimitiveHandler(const Args& args)
   };
   dnnl::memory::desc original_b_md({b_k_size_, b_n_size_}, b_type_,
                                    {b_k_stride_, b_n_stride_});
+#ifdef __aarch64__
+  // dummy M size for prepacking weights
+  // Prepacking weights improves performance and avoid runtime reorders
+  constexpr dnnl_dim_t kProbeM = 128;
+#else
+  constexpr dnnl_dim_t kProbeM = DNNL_RUNTIME_DIM_VAL;
+#endif
+
   prepack_weight(args.b_ptr, original_b_md,
                  create_primitive_desc(
-                     MSizeCacheKey{.a_m_size = DNNL_RUNTIME_DIM_VAL,
+                     MSizeCacheKey{.a_m_size = kProbeM,
                                    .use_bias = false,
                                    .bias_type = dnnl::memory::data_type::undef},
-                     true)
+                     /*first_time=*/true)
                      .weights_desc());
   init_runtime_memory_cache(args);
 }

csrc/cpu/mla_decode.cpp

@@ -18,8 +18,8 @@ struct KernelVecType<float> {
 
 template <>
 struct KernelVecType<c10::Half> {
-#if defined(__powerpc64__) || defined(__s390x__)
-  // Power and s390x architecture-specific vector types
+#if defined(__powerpc64__)
+  // Power specific vector types
   using qk_load_vec_type = vec_op::FP32Vec16;
   using qk_vec_type = vec_op::FP32Vec16;
   using v_load_vec_type = vec_op::FP32Vec16;
@@ -38,7 +38,7 @@ struct KernelVecType<c10::BFloat16> {
   using qk_vec_type = vec_op::BF16Vec32;
   using v_load_vec_type = vec_op::BF16Vec16;
 };
-#elif defined(__aarch64__)
+#else
 template <>
 struct KernelVecType<c10::BFloat16> {
   using qk_load_vec_type = vec_op::BF16Vec16;

csrc/cpu/torch_bindings.cpp

@@ -119,8 +119,8 @@ void cpu_fused_moe(torch::Tensor& output, const torch::Tensor& input,
                    const std::optional<torch::Tensor>& w13_bias,
                    const std::optional<torch::Tensor>& w2_bias,
                    const torch::Tensor& topk_weights,
-                   const torch::Tensor& topk_id, const std::string& act,
-                   const std::string& isa);
+                   const torch::Tensor& topk_id, const bool skip_weighted,
+                   const std::string& act, const std::string& isa);
 
 TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // vLLM custom ops
@@ -320,6 +320,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def(
       "cpu_fused_moe(Tensor(a0!) output, Tensor input, Tensor w13, Tensor w2, "
       "Tensor? w13_bias, Tensor? w2_bias, Tensor topk_weights, Tensor topk_id, "
+      "bool skip_weighted, "
       "str act, str isa) -> ()");
   ops.impl("cpu_fused_moe", torch::kCPU, &cpu_fused_moe);
 #endif

csrc/cuda_view.cu

@@ -2,33 +2,58 @@
 #include <torch/cuda.h>
 #include <cuda_runtime.h>
 
-// This function assumes that `cpu_tensor` is a CPU tensor allocated with pinned
-// memory, and that UVA (Unified Virtual Addressing) is enabled.
+// This function assumes that `cpu_tensor` is a CPU tensor,
+// and that UVA (Unified Virtual Addressing) is enabled.
 torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor) {
   TORCH_CHECK(cpu_tensor.device().is_cpu(), "Input tensor must be on CPU");
 
-  // Get raw host pointer from CPU tensor
-  void* host_ptr = cpu_tensor.data_ptr();
+  // handle empty tensor
+  if (cpu_tensor.numel() == 0) {
+    return torch::empty(cpu_tensor.sizes(),
+                        cpu_tensor.options().device(torch::kCUDA));
+  }
+
+  if (cpu_tensor.is_pinned()) {
+    // If CPU tensor is pinned, directly get the device pointer.
+    void* host_ptr = const_cast<void*>(cpu_tensor.data_ptr());
+    void* device_ptr = nullptr;
+    cudaError_t err = cudaHostGetDevicePointer(&device_ptr, host_ptr, 0);
+    TORCH_CHECK(err == cudaSuccess,
+                "cudaHostGetDevicePointer failed: ", cudaGetErrorString(err));
+
+    return torch::from_blob(
+        device_ptr, cpu_tensor.sizes(), cpu_tensor.strides(),
+        [base = cpu_tensor](void*) {},  // keep cpu tensor alive
+        cpu_tensor.options().device(torch::kCUDA));
+  }
+
+  // If CPU tensor is not pinned, allocate a new pinned memory buffer.
+  torch::Tensor contiguous_cpu = cpu_tensor.contiguous();
+  size_t nbytes = contiguous_cpu.nbytes();
+
+  void* host_ptr = nullptr;
+  cudaError_t err = cudaHostAlloc(&host_ptr, nbytes, cudaHostAllocMapped);
+  if (err != cudaSuccess) {
+    AT_ERROR("cudaHostAlloc failed: ", cudaGetErrorString(err));
+  }
+
+  err = cudaMemcpy(host_ptr, contiguous_cpu.data_ptr(), nbytes,
+                   cudaMemcpyDefault);
+  if (err != cudaSuccess) {
+    cudaFreeHost(host_ptr);
+    AT_ERROR("cudaMemcpy failed: ", cudaGetErrorString(err));
+  }
 
-  // Get a device pointer corresponding to the pinned host memory
   void* device_ptr = nullptr;
-  cudaError_t err = cudaHostGetDevicePointer(&device_ptr, host_ptr, 0);
-  TORCH_CHECK(err == cudaSuccess,
-              "cudaHostGetDevicePointer failed: ", cudaGetErrorString(err));
+  err = cudaHostGetDevicePointer(&device_ptr, host_ptr, 0);
+  if (err != cudaSuccess) {
+    cudaFreeHost(host_ptr);
+    AT_ERROR("cudaHostGetDevicePointer failed: ", cudaGetErrorString(err));
+  }
 
-  // We'll use the same sizes, strides, and dtype as the CPU tensor.
-  // TODO: check if layout is respected.
-  auto sizes = cpu_tensor.sizes();
-  auto strides = cpu_tensor.strides();
-  auto options = cpu_tensor.options().device(torch::kCUDA);
+  auto deleter = [host_ptr](void*) { cudaFreeHost(host_ptr); };
 
-  // use default no-op deleter, since the memory is owned by the original CPU
-  // tensor
-  torch::Tensor cuda_tensor =
-      torch::from_blob(device_ptr, sizes, strides, options);
-
-  TORCH_CHECK(cuda_tensor.device().is_cuda(),
-              "Resulting tensor is not on CUDA device");
-
-  return cuda_tensor;
-}
+  return torch::from_blob(device_ptr, contiguous_cpu.sizes(),
+                          contiguous_cpu.strides(), deleter,
+                          contiguous_cpu.options().device(torch::kCUDA));
+}

csrc/cutlass_extensions/common.hpp

@@ -152,3 +152,14 @@ struct enable_sm120_only : Kernel {
 #endif
   }
 };
+
+// SM12x family includes SM120 (RTX 5090) and SM121 (DGX Spark GB10)
+template <typename Kernel>
+struct enable_sm120_family : Kernel {
+  template <typename... Args>
+  CUTLASS_DEVICE void operator()(Args&&... args) {
+#if defined __CUDA_ARCH__ && (__CUDA_ARCH__ >= 1200 && __CUDA_ARCH__ < 1300)
+    Kernel::operator()(std::forward<Args>(args)...);
+#endif
+  }
+};

csrc/moe/moe_permute_unpermute_op.cu

@@ -14,12 +14,10 @@ void moe_permute(
     const torch::Tensor& token_expert_indices,       // [n_token, topk]
     const std::optional<torch::Tensor>& expert_map,  // [n_expert]
     int64_t n_expert, int64_t n_local_expert, int64_t topk,
-    const std::optional<int64_t>& align_block_size,
     torch::Tensor& permuted_input,             // [permuted_size, hidden]
     torch::Tensor& expert_first_token_offset,  // [n_local_expert + 1]
     torch::Tensor& inv_permuted_idx,           // [n_token, topk]
-    torch::Tensor& permuted_idx,               // [permute_size]
-    torch::Tensor& m_indices) {                // [align_expand_m]
+    torch::Tensor& permuted_idx) {             // [permute_size]
   TORCH_CHECK(expert_first_token_offset.scalar_type() == at::ScalarType::Long,
               "expert_first_token_offset must be int64");
   TORCH_CHECK(topk_ids.scalar_type() == at::ScalarType::Int,
@@ -34,8 +32,6 @@ void moe_permute(
               "token_expert_indices shape must be same as inv_permuted_idx");
   auto n_token = input.sizes()[0];
   auto n_hidden = input.sizes()[1];
-  auto align_block_size_value =
-      align_block_size.has_value() ? align_block_size.value() : -1;
   auto stream = at::cuda::getCurrentCUDAStream().stream();
   const long sorter_size =
       CubKeyValueSorter::getWorkspaceSize(n_token * topk, n_expert);
@@ -73,42 +69,15 @@ void moe_permute(
       get_ptr<int64_t>(expert_first_token_offset), n_token, n_expert,
       n_local_expert, topk, sorter, get_ptr<int>(sort_workspace), stream);
 
-  // DeepGEMM: use getMIndices kernel to compute
-  // 1) align_expert_first_token_offset (aligned prefix offsets)
-  // 2) m_indices (expert id for each aligned row)
-  // eg. expert0: 3, expert1: 5, expert2: 2 tokens respectively
-  // expert_first_token_offset = [0, 3, 8, 10], align_block_size = 4
-  // expert0: 3->4, expert1: 5->8, expert2: 2->4
-  // align_expert_first_token_offset = [0, 4, 12, 16]
-  // so m_indices = [0,0,0,0, 1,1,1,1,1,1,1,1, 2,2,2,2]
-  torch::Tensor align_expert_first_token_offset;
-  const int64_t* aligned_expert_first_token_offset_ptr = nullptr;
-  if (align_block_size.has_value()) {
-    align_expert_first_token_offset =
-        torch::zeros_like(expert_first_token_offset);
-    getMIndices(get_ptr<int64_t>(expert_first_token_offset),
-                get_ptr<int64_t>(align_expert_first_token_offset),
-                get_ptr<int>(m_indices), n_local_expert, align_block_size_value,
-                stream);
-    aligned_expert_first_token_offset_ptr =
-        get_ptr<int64_t>(align_expert_first_token_offset);
-  }
-
   // dispatch expandInputRowsKernelLauncher
   MOE_DISPATCH(input.scalar_type(), [&] {
     expandInputRowsKernelLauncher<scalar_t>(
         get_ptr<scalar_t>(input), get_ptr<scalar_t>(permuted_input),
         get_ptr<int>(permuted_experts_id), get_ptr<int>(sorted_row_idx),
         get_ptr<int>(inv_permuted_idx), get_ptr<int>(permuted_idx),
-        get_ptr<int64_t>(expert_first_token_offset),
-        aligned_expert_first_token_offset_ptr, n_token, valid_num_ptr, n_hidden,
-        topk, n_local_expert, align_block_size_value, stream);
+        get_ptr<int64_t>(expert_first_token_offset), n_token, valid_num_ptr,
+        n_hidden, topk, n_local_expert, stream);
   });
-
-  // this is only required for DeepGemm and not required for CUTLASS group gemm
-  if (align_block_size.has_value()) {
-    expert_first_token_offset.copy_(align_expert_first_token_offset);
-  }
 }
 
 void moe_unpermute(
@@ -201,16 +170,13 @@ void shuffle_rows(const torch::Tensor& input_tensor,
 
 #else
 
-void moe_permute(const torch::Tensor& input, const torch::Tensor& topk_weights,
-                 torch::Tensor& topk_ids,
+void moe_permute(const torch::Tensor& input, const torch::Tensor& topk_ids,
                  const torch::Tensor& token_expert_indices,
                  const std::optional<torch::Tensor>& expert_map,
                  int64_t n_expert, int64_t n_local_expert, int64_t topk,
-                 const std::optional<int64_t>& align_block_size,
                  torch::Tensor& permuted_input,
                  torch::Tensor& expert_first_token_offset,
-                 torch::Tensor& src_row_id2dst_row_id_map,
-                 torch::Tensor& m_indices) {
+                 torch::Tensor& inv_permuted_idx, torch::Tensor& permuted_idx) {
   TORCH_CHECK(false, "moe_permute is not supported on CUDA < 12.0");
 }
 

csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.cu

@@ -168,64 +168,4 @@ void preprocessTopkIdLauncher(int* topk_id_ptr, int size,
       topk_id_ptr, size, expert_map_ptr, num_experts);
 }
 
-template <bool ALIGN_BLOCK_SIZE>
-__global__ void getMIndicesKernel(int64_t* expert_first_token_offset,
-                                  int64_t* align_expert_first_token_offset,
-                                  int* m_indices, const int num_local_expert,
-                                  const int align_block_size) {
-  int eidx = blockIdx.x;
-  int tidx = threadIdx.x;
-  extern __shared__ int64_t smem_expert_first_token_offset[];
-  for (int i = tidx; i <= num_local_expert; i += blockDim.x) {
-    smem_expert_first_token_offset[i] = __ldg(expert_first_token_offset + i);
-  }
-  __syncthreads();
-  auto last_token_offset = smem_expert_first_token_offset[eidx + 1];
-  auto first_token_offset = smem_expert_first_token_offset[eidx];
-  int n_token_in_expert = last_token_offset - first_token_offset;
-
-  if constexpr (ALIGN_BLOCK_SIZE) {
-    n_token_in_expert = (n_token_in_expert + align_block_size - 1) /
-                        align_block_size * align_block_size;
-    // round up to ALIGN_BLOCK_SIZE
-    int64_t accumulate_align_offset = 0;
-    for (int i = 1; i <= eidx + 1; i++) {
-      int n_token = smem_expert_first_token_offset[i] -
-                    smem_expert_first_token_offset[i - 1];
-      accumulate_align_offset =
-          accumulate_align_offset + (n_token + align_block_size - 1) /
-                                        align_block_size * align_block_size;
-      if (i == eidx) {
-        first_token_offset = accumulate_align_offset;
-      }
-      // last block store align_expert_first_token_offset
-      if (eidx == num_local_expert - 1 && threadIdx.x == 0) {
-        align_expert_first_token_offset[i] = accumulate_align_offset;
-      }
-    }
-  }
-  for (int idx = tidx; idx < n_token_in_expert; idx += blockDim.x) {
-    // update m_indice with expert id
-    m_indices[first_token_offset + idx] = eidx;
-  }
-}
-
-void getMIndices(int64_t* expert_first_token_offset,
-                 int64_t* align_expert_first_token_offset, int* m_indices,
-                 int num_local_expert, const int align_block_size,
-                 cudaStream_t stream) {
-  int block = 256;
-  int grid = num_local_expert;
-  int smem_size = sizeof(int64_t) * (num_local_expert + 1);
-  if (align_block_size == -1) {
-    getMIndicesKernel<false><<<grid, block, smem_size, stream>>>(
-        expert_first_token_offset, align_expert_first_token_offset, m_indices,
-        num_local_expert, align_block_size);
-  } else {
-    getMIndicesKernel<true><<<grid, block, smem_size, stream>>>(
-        expert_first_token_offset, align_expert_first_token_offset, m_indices,
-        num_local_expert, align_block_size);
-  }
-}
-
 #endif

csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.h

@@ -60,10 +60,9 @@ void expandInputRowsKernelLauncher(
     T const* unpermuted_input, T* permuted_output, int* sorted_experts,
     int const* expanded_dest_row_to_expanded_source_row,
     int* expanded_source_row_to_expanded_dest_row, int* permuted_idx,
-    int64_t const* expert_first_token_offset,
-    int64_t const* aligned_expert_first_token_offset, int64_t const num_rows,
+    int64_t const* expert_first_token_offset, int64_t const num_rows,
     int64_t const* num_valid_tokens_ptr, int64_t const cols, int const k,
-    int num_local_experts, const int& align_block_size, cudaStream_t stream);
+    int num_local_experts, cudaStream_t stream);
 
 template <class T, class OutputType>
 void finalizeMoeRoutingKernelLauncher(
@@ -76,9 +75,4 @@ void preprocessTopkIdLauncher(int* topk_id_ptr, int size,
                               const int* expert_map_ptr, int num_experts,
                               cudaStream_t stream);
 
-void getMIndices(int64_t* expert_first_token_offset,
-                 int64_t* align_expert_first_token_offset, int* m_indices,
-                 int num_local_expert, const int align_block_size,
-                 cudaStream_t stream);
-
 #include "moe_permute_unpermute_kernel.inl"

csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.inl

@@ -1,14 +1,13 @@
 #pragma once
 
-template <typename T, bool CHECK_SKIPPED, bool ALIGN_BLOCK_SIZE>
+template <typename T, bool CHECK_SKIPPED>
 __global__ void expandInputRowsKernel(
     T const* unpermuted_input, T* permuted_output, int* sorted_experts,
     int const* expanded_dest_row_to_expanded_source_row,
     int* expanded_source_row_to_expanded_dest_row, int* permuted_idx,
-    int64_t const* expert_first_token_offset,
-    int64_t const* aligned_expert_first_token_offset, int64_t const num_rows,
+    int64_t const* expert_first_token_offset, int64_t const num_rows,
     int64_t const* num_dest_rows, int64_t const cols, int64_t k,
-    int num_local_experts, int align_block_size) {
+    int num_local_experts) {
   // Reverse permutation map.
   // I do this so that later, we can use the source -> dest map to do the k-way
   // reduction and unpermuting. I need the reverse map for that reduction to
@@ -19,24 +18,6 @@ __global__ void expandInputRowsKernel(
       expanded_dest_row_to_expanded_source_row[expanded_dest_row];
   int expert_id = sorted_experts[expanded_dest_row];
 
-  if constexpr (ALIGN_BLOCK_SIZE) {
-    // convert (unaligned) expanded_dest_row -> aligned expanded_dest_row.
-    // aligned_expert_first_token_offset[e] provides the aligned prefix start
-    // for expert e. For non-local experts we map to the end (total aligned M).
-    int64_t aligned_base = 0;
-    int64_t token_offset_in_expert = 0;
-    if (expert_id >= num_local_experts) {
-      aligned_base =
-          __ldg(aligned_expert_first_token_offset + num_local_experts);
-      token_offset_in_expert = 0;
-    } else {
-      aligned_base = __ldg(aligned_expert_first_token_offset + expert_id);
-      token_offset_in_expert =
-          expanded_dest_row - __ldg(expert_first_token_offset + expert_id);
-    }
-    expanded_dest_row = aligned_base + token_offset_in_expert;
-  }
-
   if (threadIdx.x == 0) {
     assert(expanded_dest_row <= INT32_MAX);
     expanded_source_row_to_expanded_dest_row[expanded_source_row] =
@@ -76,29 +57,25 @@ void expandInputRowsKernelLauncher(
     T const* unpermuted_input, T* permuted_output, int* sorted_experts,
     int const* expanded_dest_row_to_expanded_source_row,
     int* expanded_source_row_to_expanded_dest_row, int* permuted_idx,
-    int64_t const* expert_first_token_offset,
-    int64_t const* aligned_expert_first_token_offset, int64_t const num_rows,
+    int64_t const* expert_first_token_offset, int64_t const num_rows,
     int64_t const* num_valid_tokens_ptr, int64_t const cols, int const k,
-    int num_local_experts, const int& align_block_size, cudaStream_t stream) {
+    int num_local_experts, cudaStream_t stream) {
   int64_t const blocks = num_rows * k;
   int64_t const threads = 256;
-  using FuncPtr = decltype(&expandInputRowsKernel<T, true, true>);
-  FuncPtr func_map[2][2] = {
-      {&expandInputRowsKernel<T, false, false>,
-       &expandInputRowsKernel<T, false, true>},
-      {&expandInputRowsKernel<T, true, false>,
-       &expandInputRowsKernel<T, true, true>},
+  using FuncPtr = decltype(&expandInputRowsKernel<T, true>);
+  FuncPtr func_map[2] = {
+      &expandInputRowsKernel<T, false>,
+      &expandInputRowsKernel<T, true>,
   };
   bool is_check_skip = num_valid_tokens_ptr != nullptr;
-  bool is_align_block_size = align_block_size != -1;
-  auto func = func_map[is_check_skip][is_align_block_size];
+  auto func = func_map[is_check_skip];
 
   func<<<blocks, threads, 0, stream>>>(
       unpermuted_input, permuted_output, sorted_experts,
       expanded_dest_row_to_expanded_source_row,
       expanded_source_row_to_expanded_dest_row, permuted_idx,
-      expert_first_token_offset, aligned_expert_first_token_offset, num_rows,
-      num_valid_tokens_ptr, cols, k, num_local_experts, align_block_size);
+      expert_first_token_offset, num_rows, num_valid_tokens_ptr, cols, k,
+      num_local_experts);
 }
 
 template <class T, class U>

csrc/moe/torch_bindings.cpp

@@ -99,9 +99,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
       "moe_permute(Tensor input, Tensor topk_ids,"
       "Tensor token_expert_indices, Tensor? expert_map, int n_expert,"
       "int n_local_expert,"
-      "int topk, int? align_block_size,Tensor! permuted_input, Tensor! "
+      "int topk, Tensor! permuted_input, Tensor! "
       "expert_first_token_offset, Tensor! inv_permuted_idx, Tensor! "
-      "permuted_idx, Tensor! m_indices)->()");
+      "permuted_idx)->()");
 
   m.def(
       "moe_unpermute(Tensor permuted_hidden_states, Tensor topk_weights,"

csrc/ops.h

@@ -114,6 +114,10 @@ void top_k_per_row_decode(const torch::Tensor& logits, int64_t next_n,
                           int64_t numRows, int64_t stride0, int64_t stride1,
                           int64_t topK);
 
+void large_context_topk(const torch::Tensor& score, torch::Tensor& indices,
+                        const torch::Tensor& lengths,
+                        std::optional<torch::Tensor> row_starts_opt);
+
 void rms_norm_static_fp8_quant(torch::Tensor& out, torch::Tensor& input,
                                torch::Tensor& weight, torch::Tensor& scale,
                                double epsilon);

csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8_dispatch.cuh

@@ -103,7 +103,8 @@ struct cutlass_3x_gemm_fp8_blockwise {
           MainloopScheduler
       >::CollectiveOp;
 
-  using KernelType = enable_sm120_only<cutlass::gemm::kernel::GemmUniversal<
+  // SM12x family to support both SM120 (RTX 5090) and SM121 (DGX Spark)
+  using KernelType = enable_sm120_family<cutlass::gemm::kernel::GemmUniversal<
       Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue>>;
 
   struct GemmKernel : public KernelType {};

csrc/rocm/skinny_gemms.cu

@@ -1365,13 +1365,12 @@ torch::Tensor wvSplitK(const at::Tensor& in_a, const at::Tensor& in_b,
   return out_c;
 }
 
-#if defined(__gfx950__)  // TODO: Add NAVI support
-  // This version targets big A[] cases, where it is much larger than LDS
-  // capacity
+// This version targets cases skinny where CUs are not filled
+// Wave-SplitK is used with reduction done via atomics.
+#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 UNRL, int N, int GrpsShrB, int CHUNKK>
 __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,
@@ -1383,12 +1382,11 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   int* cntr = (int*)(&glbl[M * N]);
 
   constexpr int NTILE = 16;
-  constexpr int WVLDS_ = (NTILE * THRDS * A_CHUNK);
   constexpr int APAD = 1;
   constexpr int ASTRD = 64;
   constexpr int BPAD = 1;
-  constexpr int BSTRD = 64;
-  constexpr int WVLDS = ((WVLDS_ + (WVLDS_ / BSTRD) * 4 * BPAD));
+  constexpr int WVLDS_ = THRDS * A_CHUNK / CHUNKK;
+  constexpr int WVLDS = ((WVLDS_ + A_CHUNK * BPAD)) * YTILE;
 
   constexpr int max_lds_len = LDS_SIZE / 2;
 
@@ -1442,17 +1440,17 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         break;
     }
   #else
-  int constexpr kFit = 512;
+  int constexpr kFit = 512 / CHUNKK;
   int constexpr kfitsPerRdc = 1;
   #endif
 
-  bool doRdc = (kfitsPerRdc * kFit < K);
+  bool doRdc = true;  // Assuming (kfitsPerRdc * kFit < K) is always true
   uint32_t numCuWithFullK =
       ((M + (WvPrGrp * YTILE / GrpsShrB) - 1) / (WvPrGrp * YTILE / GrpsShrB));
   uint32_t Mmod = numCuWithFullK * (WvPrGrp * YTILE / GrpsShrB);
 
   // given above k-split, find this wave's position
-  uint32_t kFitPdd = kFit + (kFit / ASTRD) * APAD;
+  uint32_t kFitPdd = kFit * CHUNKK + ((kFit * CHUNKK) / ASTRD) * APAD;
   uint32_t m0 = (blockIdx.x * WvPrGrp / GrpsShrB) * YTILE;
   uint32_t m1 = ((threadIdx.y % WvPrGrp) / GrpsShrB) * YTILE;
   uint32_t m = (m0 + m1) % Mmod;
@@ -1460,8 +1458,8 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   uint32_t k_end = (m0 / Mmod + 1) * kFit * kfitsPerRdc;
   const uint32_t k_rnd = (K + kFit * kfitsPerRdc - 1) / (kFit * kfitsPerRdc);
 
-  scalar8 sum4[N / NTILE / GrpsShrB][1];
-  bigType bigB_[YTILE / GrpsShrB][UNRL];
+  scalar8 sum4[N / NTILE / GrpsShrB][1] = {0};
+  bigType bigB_[YTILE / GrpsShrB / CHUNKK][UNRL];
   const uint32_t bLoader = (threadIdx.y % GrpsShrB);
   uint32_t kBase = 0;
   if (k_str >= K) return;
@@ -1498,12 +1496,15 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
     #pragma unroll
   for (uint32_t k2 = 0; k2 < UNRL; k2++) {
     uint32_t k = k_str + k2 * THRDS * A_CHUNK;
-    uint32_t k_ = k + threadIdx.x * A_CHUNK;
+    uint32_t k_ = k + (threadIdx.x % (THRDS / CHUNKK)) * A_CHUNK;
     const scalar_t* B_ = &B[min__(k_, K - A_CHUNK)];
     #pragma unroll
-    for (uint32_t y = 0; y < YTILE / GrpsShrB; y++)
-      bigB_[y][k2].h8 = (loadnt(
-          (scalar8*)(&B_[min__(y * GrpsShrB + bLoader + m, M - 1) * K])));
+    for (uint32_t y = 0; y < YTILE / GrpsShrB; y += CHUNKK)
+      bigB_[y / CHUNKK][k2].h8 = (loadnt(
+          (scalar8*)(&B_[min__((y + threadIdx.x / (THRDS / CHUNKK)) * GrpsShrB +
+                                   bLoader + m,
+                               M - 1) *
+                         K])));
   }
   {
   #else
@@ -1556,48 +1557,50 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       if (reloada) {
   #endif
         constexpr int sprdN = 4;
-        const uint32_t thrd = ((threadIdx.y / sprdN) * THRDS + threadIdx.x);
+        const uint32_t thrd = threadIdx.x % (THRDS / CHUNKK);
 
   #ifndef WVSPLITKRC_1KPASS
     #pragma unroll
-        for (int k = 0; k < kFit; k += THRDS * (WvPrGrp / sprdN) * A_CHUNK) {
+        for (int k = 0; k < kFit;
+             k += (THRDS * (WvPrGrp / sprdN) * A_CHUNK) / CHUNKK) {
   #else
         const unsigned int k = 0;
         {
   #endif
           unsigned int kOff = k + (thrd * A_CHUNK);
           unsigned int kOffcp = min__(K - A_CHUNK, k_str + kOff);
-          const unsigned int k_in = kOffcp + ((threadIdx.y % sprdN)) * K;
-          const unsigned int k_ot = kOff + ((threadIdx.y % sprdN)) * kFitPdd;
-          for (unsigned int n = 0; n < N / 2; n += sprdN) {
-            __builtin_amdgcn_global_load_lds((int*)(&A[k_in + n * K]),
-                                             (int*)(&s[(k_ot + n * kFitPdd)]),
-                                             16, 0, 0);
-            if (((threadIdx.y % sprdN)) + n + N / 2 >= actlN) continue;
+          for (unsigned int n = 0; n < N; n += CHUNKK * sprdN) {
             __builtin_amdgcn_global_load_lds(
-                (int*)(&A[k_in + (n + N / 2) * K]),
-                (int*)(&s[(k_ot + (n + N / 2) * kFitPdd)]), 16, 0, 0);
+                (int*)(&A[min__(
+                    K * actlN - A_CHUNK,
+                    kOffcp + K * (n / CHUNKK +
+                                  (N / CHUNKK) * (threadIdx.x / (64 / CHUNKK)) +
+                                  (threadIdx.y % sprdN)))]),
+                (int*)(&s[(k +
+                           kFitPdd * ((n / CHUNKK) + (threadIdx.y % sprdN)))]),
+                16, 0, 0);
           }
 
           // Stage loaded B[] to LDS for MFMA swizzling...
           for (uint32_t k2 = 0; k2 < UNRL; k2++) {
             uint32_t k = k1 + k2 * THRDS * A_CHUNK;
-            uint32_t k_ = k + threadIdx.x * A_CHUNK;
+            uint32_t k_ = k + (threadIdx.x % (THRDS / CHUNKK)) * A_CHUNK;
             const bool oob_k = (k_ >= K);
-            for (uint32_t y = 0; y < YTILE / GrpsShrB; y++) {
-              uint32_t idx = threadIdx.x * 4 +
-                             (y * GrpsShrB + bLoader) * ((THRDS + BPAD) * 4);
+            for (uint32_t y = 0; y < YTILE / GrpsShrB; y += CHUNKK) {
+              uint32_t idx =
+                  (threadIdx.x % (THRDS / CHUNKK)) * 4 +
+                  ((y + threadIdx.x / (THRDS / CHUNKK)) * GrpsShrB + bLoader) *
+                      ((THRDS / CHUNKK + BPAD) * 4);
               // zero out if oob
               *((scalar8*)&myStg[idx]) =
-                  (oob_k || (y * GrpsShrB + bLoader + m >= M))
+                  (oob_k)  // TODO: ever necessary (y*GrpsShrB+bLoader+m>=M) ?
                       ? 0
-                      : bigB_[y][k2].h8;
+                      : bigB_[y / CHUNKK][k2].h8;
             }
           }
         }
       }
     }
-
   #ifndef WVSPLITKRC_1KPASS
     // Fire load of next B[] chunk...
     if ((k1 + THRDS * A_CHUNK * UNRL < k_end) &&
@@ -1608,40 +1611,50 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         uint32_t k_ = k + threadIdx.x * A_CHUNK;
         const scalar_t* B_ = &B[min__(k_, K - A_CHUNK)];
     #pragma unroll
-        for (uint32_t y = 0; y < YTILE / GrpsShrB; y++)
-          bigB_[y][k2].h8 = (loadnt(
-              (scalar8*)(&B_[min__(y * GrpsShrB + bLoader + m, M - 1) * K])));
+        for (uint32_t y = 0; y < YTILE / GrpsShrB; y += CHUNKK)
+          bigB_[y / CHUNKK][k2].h8 = (loadnt(
+              (scalar8*)(&B_[min__((y + threadIdx.x / (THRDS / CHUNKK)) *
+                                           GrpsShrB +
+                                       bLoader + m,
+                                   M - 1) *
+                             K])));
       }
   #endif
 
     // B[] staging is cooperative across GrpsShrB, so sync here before reading
-    // back
+    // back. This wait is currently inserted by compiler, but not gauranteed.
+    asm volatile("s_waitcnt 0");
     __syncthreads();
 
     // read back B[] swizzled for MFMA...
-    bigType bigB[YTILE][UNRL];
+    bigType bigB[YTILE / CHUNKK][UNRL];
     for (uint32_t k2 = 0; k2 < UNRL; k2++) {
-      for (uint32_t y = 0; y < YTILE; y++) {
-        unsigned int idx = (threadIdx.x % YTILE) * ((THRDS + BPAD) * 4) +
-                           (threadIdx.x / YTILE) * 4 + y * 16;
+      for (uint32_t y = 0; y < YTILE / CHUNKK; y++) {
+        unsigned int idx =
+            (threadIdx.x % YTILE) * ((THRDS / CHUNKK + BPAD) * 4) +
+            (threadIdx.x / YTILE) * 4 + y * 16;
         bigB[y][k2].h8 = *((scalar8*)&myStg[idx]);
       }
     }
 
     // rReadback A[] swizzled for MFMA...
-    bigType bigA[N / GrpsShrB][UNRL];
+    bigType bigA[N / GrpsShrB / CHUNKK][UNRL];
   #pragma unroll
     for (uint32_t k2 = 0; k2 < UNRL; k2++) {
       uint32_t k = k1 + k2 * THRDS * A_CHUNK - kBase - k_str;
   #pragma unroll
       for (uint32_t nt = 0; nt < N / GrpsShrB; nt += NTILE)
   #pragma unroll
-        for (uint32_t n = 0; n < NTILE; n++) {
-          uint32_t idxa = (nt + (threadIdx.x % NTILE) +
-                           (N / GrpsShrB) * (threadIdx.y % GrpsShrB)) *
-                              kFitPdd +
-                          A_CHUNK * ((threadIdx.x / NTILE) + n * 4) + k;
-          bigA[nt + n][k2] = *((const bigType*)(&(s[idxa])));
+        for (uint32_t n = 0; n < NTILE / CHUNKK; n++) {
+          uint32_t idxa =
+              ((nt + (N / GrpsShrB) * (threadIdx.y % GrpsShrB)) % (N / CHUNKK) +
+               (threadIdx.x % NTILE)) *
+                  kFitPdd +
+              ((nt + (N / GrpsShrB) * (threadIdx.y % GrpsShrB)) /
+               (N / CHUNKK)) *
+                  A_CHUNK * (64 / CHUNKK) +
+              A_CHUNK * ((threadIdx.x / NTILE) + n * 4) + k;
+          bigA[nt / CHUNKK + n][k2] = *((const bigType*)(&(s[idxa])));
         }
     }
 
@@ -1650,152 +1663,75 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
     for (uint32_t k2 = 0; k2 < UNRL; k2++) {
   #pragma unroll
       for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++) {
-        if constexpr (std::is_same_v<scalar_t, half>) {
-          sum4[nt][0] = __builtin_amdgcn_mfma_f32_16x16x16f16(
-              bigA[nt * NTILE + 0][k2].h4[0], bigB[0][k2].h4[0],
-              (k1 == k_str) ? ((scalar8){0}) : sum4[nt][0], 0, 0, 0);
-          sum4[nt][0] = __builtin_amdgcn_mfma_f32_16x16x16f16(
-              bigA[nt * NTILE + 0][k2].h4[1], bigB[0][k2].h4[1], sum4[nt][0], 0,
-              0, 0);
-        } else {  // bf16
-          sum4[nt][0] = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(
-              bigA[nt * NTILE + 0][k2].h4[0], bigB[0][k2].h4[0],
-              (k1 == k_str) ? ((scalar8){0}) : sum4[nt][0], 0, 0, 0);
-          sum4[nt][0] = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(
-              bigA[nt * NTILE + 0][k2].h4[1], bigB[0][k2].h4[1], sum4[nt][0], 0,
-              0, 0);
-        }
   #pragma unroll
-        for (uint32_t j = 1; j < YTILE; j++) {
+        for (uint32_t j = 0; j < YTILE / CHUNKK; j++) {
           if constexpr (std::is_same_v<scalar_t, half>) {
-            sum4[nt][0] = __builtin_amdgcn_mfma_f32_16x16x16f16(
-                bigA[nt * NTILE + j][k2].h4[0], bigB[j][k2].h4[0], sum4[nt][0],
-                0, 0, 0);
-            sum4[nt][0] = __builtin_amdgcn_mfma_f32_16x16x16f16(
-                bigA[nt * NTILE + j][k2].h4[1], bigB[j][k2].h4[1], sum4[nt][0],
-                0, 0, 0);
+            sum4[nt][0] = __builtin_amdgcn_mfma_f32_16x16x32_f16(
+                bigA[nt * (YTILE / CHUNKK) + j][k2].h8, bigB[j][k2].h8,
+                sum4[nt][0], 0, 0, 0);
           } else {  // bf16
-            sum4[nt][0] = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(
-                bigA[nt * NTILE + j][k2].h4[0], bigB[j][k2].h4[0], sum4[nt][0],
-                0, 0, 0);
-            sum4[nt][0] = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(
-                bigA[nt * NTILE + j][k2].h4[1], bigB[j][k2].h4[1], sum4[nt][0],
-                0, 0, 0);
+            sum4[nt][0] = __builtin_amdgcn_mfma_f32_16x16x32_bf16(
+                bigA[nt * (YTILE / CHUNKK) + j][k2].h8, bigB[j][k2].h8,
+                sum4[nt][0], 0, 0, 0);
           }
         }
       }
     }
   }
 
-  if (!doRdc) {
-    if (m + (threadIdx.x % 16) < M) {
-      scalar_t biases[N / NTILE / GrpsShrB][4] = {0};
+  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]);
+      }
+    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] = {};
+    // 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 mindx = m + (threadIdx.x % 16);
             int nindx = (j + (threadIdx.x / 16) * 4) + nt * NTILE +
                         (N / GrpsShrB) * (threadIdx.y % GrpsShrB);
-            biases[nt][j] = BIAS[(mindx % Bx) + (nindx % By) * M];
+            biases[nt][j] = BIAS[(mindx % Bx) + (nindx % By) * Bx];
           }
         }
       for (uint32_t nt = 0; nt < N / NTILE / GrpsShrB; nt++) {
         for (uint32_t j = 0; j < 4; j++) {
-          int mindx = m + (threadIdx.x % 16);
-          int nindx = (j + (threadIdx.x / 16) * 4) + nt * NTILE +
-                      (N / GrpsShrB) * (threadIdx.y % GrpsShrB);
-          int adr = mindx + M * nindx;
-          if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
-            if (BIAS) sum4[nt][0][j] += __bfloat162float(biases[nt][j]);
-            C[adr] = __float2bfloat16(sum4[nt][0][j]);
-          } else {
-            if (BIAS) sum4[nt][0][j] += __half2float(biases[nt][j]);
-            C[adr] = __float2half(sum4[nt][0][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];
         }
       }
-    }
-  } else {
-    if (m + (threadIdx.x % 16) < M) {
-      int my_cntr;
-      if (!BIAS) {
-        int mindx = m + (threadIdx.x % 16);
-        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);
+      __builtin_amdgcn_sched_barrier(0);
+      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);
+          if (nindx < actlN) {
             int adr = mindx + M * nindx;
-            atomicAdd(&glbl[adr], sum4[nt][0][j]);
-          }
-        int nindx_ = (0 + (threadIdx.x / 16) * 4) + 0 * NTILE +
-                     (N / GrpsShrB) * (threadIdx.y % GrpsShrB);
-        int adr_ = mindx + M * nindx_ / 4;
-        my_cntr = atomicAdd(&cntr[adr_], 1);
-        float vals[N / NTILE / GrpsShrB][4] = {};
-        if (my_cntr + 1 == k_rnd) {
-          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;
-              vals[nt][j] = glbl[adr];
-            }
-          }
-          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);
-              if (nindx >= actlN) break;
-              int adr = mindx + M * nindx;
-              if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
-                C[adr] = __float2bfloat16(vals[nt][j]);
-              } else {
-                C[adr] = __float2half(vals[nt][j]);
-              }
-            }
-          }
-        }
-      } else {
-        int mindx = m + (threadIdx.x % 16);
-        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;
-            atomicAdd(&glbl[adr], sum4[nt][0][j]);
-            biases[nt][j] = BIAS[(mindx % Bx) + (nindx % By) * M];
-          }
-        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] = {};
-        // If we're the last k-shard, read back the value and convert...
-        if (my_cntr + 1 == k_rnd) {
-          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;
-              vals[nt][j] = glbl[adr];
-            }
-          }
-          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);
-              if (nindx >= actlN) break;
-              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]);
-              } else {
-                vals[nt][j] += __half2float(biases[nt][j]);
-                C[adr] = __float2half(vals[nt][j]);
-              }
+            if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
+              vals[nt][j] += __bfloat162float(biases[nt][j]);
+              C[adr] = __float2bfloat16(vals[nt][j]);
+            } else {
+              vals[nt][j] += __half2float(biases[nt][j]);
+              C[adr] = __float2half(vals[nt][j]);
             }
           }
         }
@@ -1814,7 +1750,7 @@ __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 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,
@@ -1859,10 +1795,10 @@ 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;
 
-#define WVSPLITKrc(_WvPrGrp, _YTILE, _UNRL, _N, _GrpsShrB)                     \
+#define WVSPLITKrc(_N, _GrpsShrB, _CHUNKK)                                     \
   {                                                                            \
-    dim3 block(64, _WvPrGrp);                                                  \
-    wvSplitKrc_<fptype, 64, _YTILE, _WvPrGrp, 8, _UNRL, _N, _GrpsShrB>         \
+    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);                \
   }
@@ -1877,15 +1813,37 @@ torch::Tensor wvSplitKrc(const at::Tensor& in_a, const at::Tensor& in_b,
             : 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(4, 16, 1, 16, 1) break;
+        WVSPLITKrc(16, 1, 1) break;
       case 32:
-        WVSPLITKrc(4, 16, 1, 32, 2) break;
+        if (chunkk == 2)
+          WVSPLITKrc(32, 2, 2) else if (chunkk == 1) WVSPLITKrc(32, 2, 1) break;
       case 64:
-        WVSPLITKrc(4, 16, 1, 64, 2) break;
+        if (chunkk == 2)
+          WVSPLITKrc(64, 4, 2) else if (chunkk == 1) WVSPLITKrc(64, 4, 1) break;
       case 128:
-        WVSPLITKrc(4, 16, 1, 128, 4) break;
+        if (chunkk == 2)
+          WVSPLITKrc(128, 4, 2) else if (chunkk == 1)
+              WVSPLITKrc(128, 4, 1) break;
       default:
         throw std::runtime_error(
             "Unsupported N value: " + std::to_string(M_in) + "," +
@@ -1899,8 +1857,9 @@ torch::Tensor wvSplitKrc(const at::Tensor& in_a, const at::Tensor& in_b,
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
-    wvSplitKQ_hf_sml_(const int K, const int Kp, const int M, const int Bx,
-                      const int By, const fp8_t* B, const fp8_t* __restrict__ A,
+    wvSplitKQ_hf_sml_(const int K, const int Kap, const int Kbp, const int M,
+                      const int Bx, const int By, const fp8_t* B,
+                      const fp8_t* __restrict__ A,
                       const scalar_t* __restrict__ BIAS, scalar_t* C,
                       const float* __restrict__ s_A,
                       const float* __restrict__ s_B, const int _WvPrGrp,
@@ -1924,9 +1883,14 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   __shared__ fp8_t s[max_lds_len];
 
   for (uint32_t k = (threadIdx.y * THRDS + threadIdx.x) * A_CHUNK;
-       k < min__(K * N, max_lds_len); k += THRDS * WvPrGrp * A_CHUNK) {
+       k < min__(Kap * N, max_lds_len); k += THRDS * WvPrGrp * A_CHUNK) {
+  #if defined(__gfx950__)
+    __builtin_amdgcn_global_load_lds((int*)(&A[k]), (int*)(&s[k]), 16, 0, 0);
+  #else
     *((bigType*)(&s[k])) = *((bigType*)(&A[k]));
+  #endif
   }
+  asm volatile("s_waitcnt vmcnt(0)");
   __syncthreads();
 
   if (threadIdx.y >= _WvPrGrp) return;
@@ -1934,37 +1898,24 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   uint32_t m = (blockIdx.x * _WvPrGrp + (threadIdx.y % _WvPrGrp)) * YTILE;
 
   using floatx16 = __attribute__((__vector_size__(16 * sizeof(float)))) float;
-  floatx16 sum[N][YTILE];
   float sA = *s_A;
   float sB = *s_B;
 
   while (m < M) {
-    for (int i = 0; i < YTILE; i++)
-      for (int n = 0; n < N; n++) sum[n][i] = {0.f};
-
-    bigType bigA[N][UNRL];
-    bigType bigB[YTILE][UNRL];
-
+    floatx16 sum[N][YTILE] = {};
     for (uint32_t k1 = 0; k1 < K; k1 += THRDS * A_CHUNK * UNRL) {
-  #pragma unroll
-      for (uint32_t k2 = 0; k2 < UNRL; k2++) {
-  #pragma unroll
-        for (uint32_t n = 0; n < N; ++n) bigA[n][k2].h8 = {0.f};
-  #pragma unroll
-        for (uint32_t y = 0; y < YTILE; ++y) bigB[y][k2].h8 = {0.f};
-      }
+      bigType bigA[N][UNRL] = {};
+      bigType bigB[YTILE][UNRL];
 
       // Fetch the weight matrix from memory!
   #pragma unroll
       for (uint32_t k2 = 0; k2 < UNRL; k2++) {
         uint32_t k = k1 + k2 * THRDS * A_CHUNK;
         uint32_t k_ = k + threadIdx.x * A_CHUNK;
-        if (k_ >= K) break;
-
-        const fp8_t* B_ = &B[(m + 0) * Kp + k_];
+        const fp8_t* B_ = &B[min__(k_, K - A_CHUNK)];
   #pragma unroll
         for (uint32_t y = 0; y < YTILE; ++y) {
-          bigB[y][k2].h8 = (loadnt((scalar8*)(&B_[y * Kp])));
+          bigB[y][k2].h8 = (loadnt((scalar8*)(&B_[min__(y + m, M - 1) * Kbp])));
         }
       }
 
@@ -1975,16 +1926,13 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         uint32_t k_ = k + threadIdx.x * A_CHUNK;
         if (k_ >= K) break;
         for (int n = 0; n < N; n++) {
-          bigA[n][k2] = *((const bigType*)(&(s[k_ + K * n])));
+          bigA[n][k2] = *((const bigType*)(&(s[k_ + Kap * n])));
         }
       }
 
   // Do the matrix multiplication in interleaved manner
   #pragma unroll
       for (uint32_t k2 = 0; k2 < UNRL; k2++) {
-        uint32_t k = k1 + k2 * THRDS * A_CHUNK;
-        if (k >= K) break;
-
         for (uint32_t n = 0; n < N; n++) {
           for (int i = 0; i < A_CHUNK; i += 8) {
             for (int y = 0; y < YTILE; ++y) {
@@ -2002,48 +1950,27 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       for (int y = 0; y < YTILE; y++) {
         float accm0 = sum[n][y][0];
         float accm16 = sum[n][y][8];
-        asm("v_add_f32 %0, %2, %3 row_shl:1 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][1]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:1 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][9]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:2 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][2]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:2 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][10]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:3 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][3]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:3 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][11]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:8 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][4]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:8 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][12]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:9 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][5]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:9 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][13]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:10 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][6]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:10 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][14]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:11 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][7]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:11 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][15]), "v"(accm16));
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][1], 0x101, 0xf, 0xf,
+                                          1);  // row_shl1
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][9], 0x101, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][2], 0x102, 0xf, 0xf,
+                                          1);  // row_shl2
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][10], 0x102, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][3], 0x103, 0xf, 0xf,
+                                          1);  // row_shl3
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][11], 0x103, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][4], 0x108, 0xf, 0xf,
+                                          1);  // row_shl8
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][12], 0x108, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][5], 0x109, 0xf, 0xf,
+                                          1);  // row_shl9
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][13], 0x109, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][6], 0x10a, 0xf, 0xf,
+                                          1);  // row_shl10
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][14], 0x10a, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][7], 0x10b, 0xf, 0xf,
+                                          1);  // row_shl11
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][15], 0x10b, 0xf, 0xf, 1);
         accm0 += __shfl(accm0, 36);
         accm16 += __shfl(accm16, 52);
         sum[n][y][0] = accm0 + __shfl(accm16, 16);
@@ -2051,19 +1978,23 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
     }
 
     if (threadIdx.x == 0) {
+      scalar_t biases[N][YTILE] = {};
+      if (BIAS)
+        for (int n = 0; n < N; n++) {
+          for (int y = 0; y < YTILE; y++) {
+            biases[n][y] = BIAS[(m + y) % Bx + (n % By) * Bx];
+          }
+        }
       for (int n = 0; n < N; n++) {
         for (int y = 0; y < YTILE; y++) {
           if (y + m >= M) break;  // To avoid mem access fault.
           sum[n][y][0] *= sA * sB;
           if constexpr (std::is_same_v<scalar_t, half>) {
-            if (BIAS)
-              sum[n][y][0] += __half2float(BIAS[(m + y) % Bx + (n % By) * M]);
+            sum[n][y][0] += __half2float(biases[n][y]);
           } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
-            if (BIAS)
-              sum[n][y][0] +=
-                  __bfloat162float(BIAS[(m + y) % Bx + (n % By) * M]);
+            sum[n][y][0] += __bfloat162float(biases[n][y]);
           }
-          C[m + y + n * M] = __float2s<scalar_t>(sum[n][y][0]);  // * sA * sB);
+          C[m + y + n * M] = __float2s<scalar_t>(sum[n][y][0]);
         }
       }
     }
@@ -2074,9 +2005,9 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
 #else   // !defined(__HIP__MI3XX__) TODO: Add NAVI support
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
-__global__ void wvSplitKQ_hf_sml_(const int K, const int Kp, const int M,
-                                  const int Bx, const int By, const fp8_t* B,
-                                  const fp8_t* __restrict__ A,
+__global__ void wvSplitKQ_hf_sml_(const int K, const int Kap, const int Kbp,
+                                  const int M, const int Bx, const int By,
+                                  const fp8_t* B, const fp8_t* __restrict__ A,
                                   const scalar_t* __restrict__ BIAS,
                                   scalar_t* C, const float* __restrict__ s_A,
                                   const float* __restrict__ s_B,
@@ -2089,8 +2020,9 @@ __global__ void wvSplitKQ_hf_sml_(const int K, const int Kp, const int M,
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
-    wvSplitKQ_hf_(const int K, const int Kp, const int M, const int Bx,
-                  const int By, const fp8_t* B, const fp8_t* __restrict__ A,
+    wvSplitKQ_hf_(const int K, const int Kap, const int Kbp, const int M,
+                  const int Bx, const int By, const fp8_t* B,
+                  const fp8_t* __restrict__ A,
                   const scalar_t* __restrict__ BIAS, scalar_t* C,
                   const float* __restrict__ s_A, const float* __restrict__ s_B,
                   const int _WvPrGrp, const int CuCount) {
@@ -2113,9 +2045,14 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   __shared__ fp8_t s[max_lds_len];
 
   for (uint32_t k = (threadIdx.y * THRDS + threadIdx.x) * A_CHUNK;
-       k < min__(K * N, max_lds_len); k += THRDS * WvPrGrp * A_CHUNK) {
+       k < min__(Kap * N, max_lds_len); k += THRDS * WvPrGrp * A_CHUNK) {
+  #if defined(__gfx950__)
+    __builtin_amdgcn_global_load_lds((int*)(&A[k]), (int*)(&s[k]), 16, 0, 0);
+  #else
     *((bigType*)(&s[k])) = *((bigType*)(&A[k]));
+  #endif
   }
+  asm volatile("s_waitcnt vmcnt(0)");
   __syncthreads();
 
   if (threadIdx.y >= _WvPrGrp) return;
@@ -2123,29 +2060,23 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   uint32_t m = (blockIdx.x * _WvPrGrp + (threadIdx.y % _WvPrGrp)) * YTILE;
 
   using floatx16 = __attribute__((__vector_size__(16 * sizeof(float)))) float;
-  floatx16 sum[N][YTILE];
   float sA = *s_A;
   float sB = *s_B;
 
   while (m < M) {
-    for (int i = 0; i < YTILE; i++)
-      for (int n = 0; n < N; n++) sum[n][i] = {0};
-
-    bigType bigA[N][UNRL];
-    bigType bigB[YTILE][UNRL];
-
+    floatx16 sum[N][YTILE] = {};
     for (uint32_t k1 = 0; k1 < K; k1 += THRDS * A_CHUNK * UNRL) {
+      bigType bigA[N][UNRL] = {};
+      bigType bigB[YTILE][UNRL];
+
       // Fetch the weight matrix from memory!
   #pragma unroll
       for (uint32_t k2 = 0; k2 < UNRL; k2++) {
         uint32_t k = k1 + k2 * THRDS * A_CHUNK;
         uint32_t k_ = k + threadIdx.x * A_CHUNK;
-        if (k_ >= K) break;
-
-        const fp8_t* B_ = &B[(m + 0) * Kp + k_];
+        const fp8_t* B_ = &B[min__(k_, K - A_CHUNK)];
         for (int y = 0; y < YTILE; ++y) {
-          if (y + m >= M) break;  // To avoid mem access fault.
-          bigB[y][k2].h8 = (loadnt((scalar8*)(&B_[y * Kp])));
+          bigB[y][k2].h8 = (loadnt((scalar8*)(&B_[min__(y + m, M - 1) * Kbp])));
         }
       }
 
@@ -2156,20 +2087,16 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         uint32_t k_ = k + threadIdx.x * A_CHUNK;
         if (k_ >= K) break;
         for (int n = 0; n < N; n++) {
-          if (k_ + K * n < max_lds_len)
-            bigA[n][k2] = *((const bigType*)(&(s[k_ + K * n])));
+          if (k_ + Kap * n < max_lds_len)
+            bigA[n][k2] = *((const bigType*)(&(s[k_ + Kap * n])));
           else
-            bigA[n][k2] = *((const bigType*)(&(A[k_ + K * n])));
+            bigA[n][k2] = *((const bigType*)(&(A[k_ + Kap * n])));
         }
       }
 
   // Do the matrix multiplication in interleaved manner
   #pragma unroll
       for (uint32_t k2 = 0; k2 < UNRL; k2++) {
-        uint32_t k = k1 + k2 * THRDS * A_CHUNK;
-        uint32_t k_ = k + threadIdx.x * A_CHUNK;
-        if (k_ >= K) break;
-
         for (uint32_t n = 0; n < N; n++) {
           for (int i = 0; i < A_CHUNK; i += 8) {
             for (int y = 0; y < YTILE; ++y) {
@@ -2187,48 +2114,27 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       for (int y = 0; y < YTILE; y++) {
         float accm0 = sum[n][y][0];
         float accm16 = sum[n][y][8];
-        asm("v_add_f32 %0, %2, %3 row_shl:1 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][1]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:1 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][9]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:2 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][2]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:2 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][10]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:3 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][3]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:3 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][11]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:8 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][4]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:8 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][12]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:9 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][5]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:9 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][13]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:10 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][6]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:10 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][14]), "v"(accm16));
-        asm("v_add_f32 %0, %2, %3 row_shl:11 bound_ctrl:0 "
-            : "=v"(accm0)
-            : "0"(accm0), "v"(sum[n][y][7]), "v"(accm0));
-        asm("v_add_f32 %0, %2, %3 row_shl:11 bound_ctrl:0 "
-            : "=v"(accm16)
-            : "0"(accm16), "v"(sum[n][y][15]), "v"(accm16));
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][1], 0x101, 0xf, 0xf,
+                                          1);  // row_shl1
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][9], 0x101, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][2], 0x102, 0xf, 0xf,
+                                          1);  // row_shl2
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][10], 0x102, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][3], 0x103, 0xf, 0xf,
+                                          1);  // row_shl3
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][11], 0x103, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][4], 0x108, 0xf, 0xf,
+                                          1);  // row_shl8
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][12], 0x108, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][5], 0x109, 0xf, 0xf,
+                                          1);  // row_shl9
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][13], 0x109, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][6], 0x10a, 0xf, 0xf,
+                                          1);  // row_shl10
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][14], 0x10a, 0xf, 0xf, 1);
+        accm0 += __builtin_amdgcn_mov_dpp(sum[n][y][7], 0x10b, 0xf, 0xf,
+                                          1);  // row_shl11
+        accm16 += __builtin_amdgcn_mov_dpp(sum[n][y][15], 0x10b, 0xf, 0xf, 1);
         accm0 += __shfl(accm0, 36);
         accm16 += __shfl(accm16, 52);
         sum[n][y][0] = accm0 + __shfl(accm16, 16);
@@ -2236,17 +2142,21 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
     }
 
     if (threadIdx.x == 0) {
+      scalar_t biases[N][YTILE] = {};
+      if (BIAS)
+        for (int n = 0; n < N; n++) {
+          for (int y = 0; y < YTILE; y++) {
+            biases[n][y] = BIAS[(m + y) % Bx + (n % By) * Bx];
+          }
+        }
       for (int n = 0; n < N; n++) {
         for (int y = 0; y < YTILE; y++) {
           if (y + m >= M) break;  // To avoid mem access fault.
           sum[n][y][0] *= sA * sB;
           if constexpr (std::is_same_v<scalar_t, half>) {
-            if (BIAS)
-              sum[n][y][0] += __half2float(BIAS[(m + y) % Bx + (n % By) * M]);
+            sum[n][y][0] += __half2float(biases[n][y]);
           } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
-            if (BIAS)
-              sum[n][y][0] +=
-                  __bfloat162float(BIAS[(m + y) % Bx + (n % By) * M]);
+            sum[n][y][0] += __bfloat162float(biases[n][y]);
           }
           C[m + y + n * M] = __float2s<scalar_t>(sum[n][y][0]);
         }
@@ -2259,9 +2169,9 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
 #else   // !defined(__HIP__MI3XX__) TODO: Add NAVI support
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
-__global__ void wvSplitKQ_hf_(const int K, const int Kp, const int M,
-                              const int Bx, const int By, const fp8_t* B,
-                              const fp8_t* __restrict__ A,
+__global__ void wvSplitKQ_hf_(const int K, const int Kap, const int Kbp,
+                              const int M, const int Bx, const int By,
+                              const fp8_t* B, const fp8_t* __restrict__ A,
                               const scalar_t* __restrict__ BIAS, scalar_t* C,
                               const float* __restrict__ s_A,
                               const float* __restrict__ s_B, const int _WvPrGrp,
@@ -2270,17 +2180,18 @@ __global__ void wvSplitKQ_hf_(const int K, const int Kp, const int M,
 }
 #endif  // defined(__HIP__MI3XX__) TODO: Add NAVI support
 
-void wvSplitKQ(const at::Tensor& in_a, const at::Tensor& in_b,
+void wvSplitKQ(const at::Tensor& in_b, const at::Tensor& in_a,
                const std::optional<at::Tensor>& in_bias, at::Tensor& out_c,
                const at::Tensor& scale_a, const at::Tensor& scale_b,
                const int64_t CuCount) {
   static c10::ScalarType kFp8Type = is_fp8_ocp()
                                         ? c10::ScalarType::Float8_e4m3fn
                                         : c10::ScalarType::Float8_e4m3fnuz;
-  auto M_in = in_a.size(0);
-  auto K_in = in_a.size(1);
-  auto N_in = in_b.size(0);
-  auto Kp_in = in_a.stride(0);
+  auto M_in = in_b.size(0);
+  auto K_in = in_b.size(1);
+  auto N_in = in_a.size(0);
+  auto Kap_in = in_a.stride(0);
+  auto Kbp_in = in_b.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)
@@ -2300,23 +2211,22 @@ void wvSplitKQ(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();
 
-#define WVSPLITKQ(_WvPrGrp, _YTILEs, _YTILEm, _YTILEb, _UNRLs, _UNRLm, _UNRLb, \
-                  _N)                                                          \
-  {                                                                            \
-    dim3 block(64, _WvPrGrp);                                                  \
-    if ((K_in * N_in <= max_lds_len) && (M_in % _YTILEs == 0)) {               \
-      int __wvPrGrp = mindiv(M_in, CuCount * _YTILEs, _WvPrGrp);               \
-      wvSplitKQ_hf_sml_<fptype, fp8_t, 64, _YTILEs, _WvPrGrp, 16, _UNRLs, _N>  \
-          <<<grid, block, 0, stream>>>(K_in, Kp_in, M_in, Bx_in, By_in, a_ptr, \
-                                       b_ptr, bias_ptr, c_ptr, s_a, s_b,       \
-                                       __wvPrGrp, CuCount);                    \
-    } else {                                                                   \
-      int __wvPrGrp = mindiv(M_in, CuCount * _YTILEm, _WvPrGrp);               \
-      wvSplitKQ_hf_<fptype, fp8_t, 64, _YTILEm, _WvPrGrp, 16, _UNRLm, _N>      \
-          <<<grid, block, 0, stream>>>(K_in, Kp_in, M_in, Bx_in, By_in, a_ptr, \
-                                       b_ptr, bias_ptr, c_ptr, s_a, s_b,       \
-                                       __wvPrGrp, CuCount);                    \
-    }                                                                          \
+#define WVSPLITKQ(_WvPrGrp, _YTILEs, _YTILEm, _UNRLs, _UNRLm, _N)             \
+  {                                                                           \
+    dim3 block(64, _WvPrGrp);                                                 \
+    if ((Kap_in * N_in <= max_lds_len) && (M_in % _YTILEs == 0)) {            \
+      int __wvPrGrp = min(_WvPrGrp, mindiv(M_in, CuCount * _YTILEs, 16));     \
+      wvSplitKQ_hf_sml_<fptype, fp8_t, 64, _YTILEs, _WvPrGrp, 16, _UNRLs, _N> \
+          <<<grid, block, 0, stream>>>(K_in, Kap_in, Kbp_in, M_in, Bx_in,     \
+                                       By_in, b_ptr, a_ptr, bias_ptr, c_ptr,  \
+                                       s_a, s_b, __wvPrGrp, CuCount);         \
+    } else {                                                                  \
+      int __wvPrGrp = min(_WvPrGrp, mindiv(M_in, CuCount * _YTILEm, 16));     \
+      wvSplitKQ_hf_<fptype, fp8_t, 64, _YTILEm, _WvPrGrp, 16, _UNRLm, _N>     \
+          <<<grid, block, 0, stream>>>(K_in, Kap_in, Kbp_in, M_in, Bx_in,     \
+                                       By_in, b_ptr, a_ptr, bias_ptr, c_ptr,  \
+                                       s_a, s_b, __wvPrGrp, CuCount);         \
+    }                                                                         \
   }
 
   AT_DISPATCH_REDUCED_FLOATING_TYPES(out_c.scalar_type(), "wvSplitKQ", [&] {
@@ -2332,16 +2242,16 @@ void wvSplitKQ(const at::Tensor& in_a, const at::Tensor& in_b,
                           : nullptr;
       switch (N_in) {
         case 1:
-          WVSPLITKQ(16, 2, 2, 2, 2, 2, 2, 1)
+          WVSPLITKQ(12, 2, 2, 2, 2, 1)
           break;
         case 2:
-          WVSPLITKQ(16, 2, 2, 2, 2, 2, 2, 2)
+          WVSPLITKQ(12, 2, 2, 2, 2, 2)
           break;
         case 3:
-          WVSPLITKQ(16, 4, 7, 7, 1, 1, 1, 3)
+          WVSPLITKQ(8, 2, 2, 1, 1, 3)
           break;
         case 4:
-          WVSPLITKQ(16, 4, 7, 7, 1, 1, 1, 4)
+          WVSPLITKQ(4, 2, 2, 1, 1, 4)
           break;
         default:
           throw std::runtime_error(

csrc/sampler.cu

@@ -725,4 +725,4 @@ void top_k_per_row_prefill(const torch::Tensor& logits,
                      static_cast<int>(stride0), static_cast<int>(stride1),
                      static_cast<int>(topK), kSortingAlgorithmThreshold);
   }
-}
+}

csrc/sparse/cutlass/sparse_scaled_mm_entry.cu

@@ -6,11 +6,11 @@
 #include "cutlass_extensions/common.hpp"
 
 bool cutlass_sparse_scaled_mm_supported(int64_t cuda_device_capability) {
-  // sparse CUTLASS kernels need at least
+  // sparse CUTLASS kernels need exactly hopper and are not forward compatible
   //   CUDA 12.2 and SM90 (Hopper)
 
 #if defined CUDA_VERSION
-  return CUDA_VERSION >= 12020 && cuda_device_capability >= 90;
+  return CUDA_VERSION >= 12020 && cuda_device_capability == 90;
 #endif
 
   return false;
@@ -98,7 +98,7 @@ std::vector<torch::Tensor> cutlass_sparse_compress(torch::Tensor const& a) {
 
   TORCH_CHECK_NOT_IMPLEMENTED(
       false,
-      "No compiled cutlass_sparse_compress for a compute capability less than "
+      "No compiled cutlass_sparse_compress for a compute capability equal to "
       "CUDA device capability: ",
       version_num);
 }

csrc/topk.cu

@@ -0,0 +1,373 @@
+// Portions of this file are adapted from SGLang PR:
+// https://github.com/sgl-project/sglang/pull/11194
+// and
+// https://github.com/sgl-project/sglang/pull/17747
+
+#include "cuda_compat.h"
+#include "dispatch_utils.h"
+
+#include <torch/cuda.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#ifndef USE_ROCM
+  #include <cub/cub.cuh>
+#else
+  #include <hipcub/hipcub.hpp>
+#endif
+
+namespace vllm {
+
+constexpr int TopK = 2048;              // DeepSeek V3 sparse attention top-k
+constexpr int kThreadsPerBlock = 1024;  // Threads per block
+
+// Shared memory budget
+#if defined(USE_ROCM)
+constexpr size_t kSmem = 48 * 1024;  // ROCm default: 48KB
+#else
+// Reduced from 128KB to 32KB to improve occupancy.
+// Each radix pass needs at most ~TopK candidates in the threshold bin,
+// so 4K entries per round (2 rounds = 8K entries = 32KB) is sufficient.
+constexpr size_t kSmem = 8 * 1024 * sizeof(uint32_t);  // 32KB (bytes)
+#endif
+
+struct FastTopKParams {
+  const float* __restrict__ input;         // [batch, seq_len] Logits
+  const int32_t* __restrict__ row_starts;  // [batch] Offset into each row
+                                           // (optional)
+  int32_t* __restrict__ indices;           // [batch, TopK] Output top-k indices
+  int32_t* __restrict__ lengths;           // [batch] Sequence lengths per row
+  int64_t input_stride;                    // Stride between rows
+};
+
+__device__ __forceinline__ auto convert_to_uint32_v2(float x) -> uint32_t {
+  uint32_t bits = __float_as_uint(x);
+  return (bits & 0x80000000u) ? ~bits : (bits | 0x80000000u);
+}
+
+__device__ __forceinline__ auto convert_to_uint8(float x) -> uint8_t {
+  __half h = __float2half_rn(x);
+  uint16_t bits = __half_as_ushort(h);
+  uint16_t key = (bits & 0x8000) ? static_cast<uint16_t>(~bits)
+                                 : static_cast<uint16_t>(bits | 0x8000);
+  return static_cast<uint8_t>(key >> 8);
+}
+
+__device__ void naive_topk_cuda(const float* __restrict__ logits,
+                                int32_t* __restrict__ output_indices,
+                                int32_t seq_len) {
+  const int thread_id = threadIdx.x;
+  for (int i = thread_id; i < TopK; i += kThreadsPerBlock) {
+    output_indices[i] = (i < seq_len) ? i : -1;
+  }
+}
+
+// Adapted from:
+// https://github.com/sgl-project/sglang/blob/v0.5.8/sgl-kernel/csrc/elementwise/topk.cu#L87
+// by: DarkSharpness
+// which at the same time is an optimized topk kernel copied from tilelang
+// kernel
+__device__ void fast_topk_cuda_tl(
+    const float* __restrict__ logits,  // Input logits [seq_len]
+    int* __restrict__ output_indices,  // Output top-k indices [TopK]
+    int logits_offset,                 // Starting offset in logits array
+    int seq_len)                       // Number of valid logits to process
+{
+  constexpr int RADIX = 256;
+  constexpr int MAX_BUFFERED_ITEMS = kSmem / (2 * sizeof(int));
+
+  alignas(128) __shared__ int shared_histogram[2][RADIX + 128];
+  alignas(128) __shared__ int shared_output_count;
+  alignas(128) __shared__ int shared_threshold_bin;
+  alignas(128) __shared__ int shared_buffered_count[2];
+
+  extern __shared__ int buffered_indices[][MAX_BUFFERED_ITEMS];
+
+  const int thread_id = threadIdx.x;
+  int remaining_k = TopK;
+
+  // Pass 0: Build coarse 8-bit histogram using FP16 high bits
+  if (thread_id < RADIX + 1) {
+    shared_histogram[0][thread_id] = 0;
+  }
+  __syncthreads();
+
+  for (int idx = thread_id; idx < seq_len; idx += kThreadsPerBlock) {
+    const auto bin = convert_to_uint8(logits[idx + logits_offset]);
+    ::atomicAdd(&shared_histogram[0][bin], 1);
+  }
+  __syncthreads();
+
+  // Helper: Compute cumulative sum (suffix sum) over histogram using ping-pong
+  // buffers
+  auto compute_cumulative_sum = [&]() {
+    static_assert(1 << 8 == RADIX,
+                  "Radix must be 256 for 8 unrolled iterations");
+#pragma unroll 8
+    for (int i = 0; i < 8; ++i) {
+      if (C10_LIKELY(thread_id < RADIX)) {
+        const int stride = 1 << i;
+        const int src_buffer = i & 1;
+        const int dst_buffer = src_buffer ^ 1;
+
+        int value = shared_histogram[src_buffer][thread_id];
+        if (thread_id < RADIX - stride) {
+          value += shared_histogram[src_buffer][thread_id + stride];
+        }
+        shared_histogram[dst_buffer][thread_id] = value;
+      }
+      __syncthreads();
+    }
+  };
+
+  compute_cumulative_sum();
+
+  // Find threshold bin where cumsum crosses remaining_k
+  if (thread_id < RADIX && shared_histogram[0][thread_id] > remaining_k &&
+      shared_histogram[0][thread_id + 1] <= remaining_k) {
+    shared_threshold_bin = thread_id;
+    shared_buffered_count[0] = 0;
+    shared_output_count = 0;
+  }
+  __syncthreads();
+
+  const int threshold_bin = shared_threshold_bin;
+  remaining_k -= shared_histogram[0][threshold_bin + 1];
+
+  // Early exit if threshold bin perfectly matches remaining_k
+  if (remaining_k == 0) {
+    for (int idx = thread_id; idx < seq_len; idx += kThreadsPerBlock) {
+      const int bin = convert_to_uint8(logits[idx + logits_offset]);
+      if (bin > threshold_bin) {
+        const int output_pos = ::atomicAdd(&shared_output_count, 1);
+        output_indices[output_pos] = idx;
+      }
+    }
+    __syncthreads();
+    return;
+  }
+
+  // Prepare for refinement passes: Process threshold bin
+  __syncthreads();
+  if (thread_id < RADIX + 1) {
+    shared_histogram[0][thread_id] = 0;
+  }
+  __syncthreads();
+
+  // Scan all elements and:
+  // 1. Write indices > threshold_bin to output
+  // 2. Buffer indices == threshold_bin for refinement
+  // 3. Build histogram for next refinement pass (fused optimization)
+  for (int idx = thread_id; idx < seq_len; idx += kThreadsPerBlock) {
+    const float logit_value = logits[idx + logits_offset];
+    const int bin = convert_to_uint8(logit_value);
+
+    if (bin > threshold_bin) {
+      // in top-k, write to output
+      const int output_pos = ::atomicAdd(&shared_output_count, 1);
+      output_indices[output_pos] = idx;
+    } else if (bin == threshold_bin) {
+      // Candidate for top-k, needs refinement
+      const int buffer_pos = ::atomicAdd(&shared_buffered_count[0], 1);
+      if (C10_LIKELY(buffer_pos < MAX_BUFFERED_ITEMS)) {
+        buffered_indices[0][buffer_pos] = idx;
+        // Fused: Build histogram for next pass
+        const uint32_t fp32_bits = convert_to_uint32_v2(logit_value);
+        const int next_bin = (fp32_bits >> 24) & 0xFF;
+        ::atomicAdd(&shared_histogram[0][next_bin], 1);
+      }
+    }
+  }
+  __syncthreads();
+
+  // ============================================================================
+  // Passes 1-4: Refine using 8-bit passes over FP32 bits
+  // ============================================================================
+  // FP32 bits [31:0] split into 4 bytes processed MSB-first:
+  // Pass 1: bits [31:24], Pass 2: bits [23:16], Pass 3: bits [15:8], Pass 4:
+  // bits [7:0]
+#pragma unroll 4
+  for (int pass = 0; pass < 4; ++pass) {
+    __shared__ int shared_final_k;  // For final pass: remaining slots to fill
+    const int src_buffer = pass % 2;
+    const int dst_buffer = src_buffer ^ 1;
+
+    // Clamp buffered count to prevent overflow
+    const int raw_buffered = shared_buffered_count[src_buffer];
+    const int num_buffered =
+        (raw_buffered < MAX_BUFFERED_ITEMS) ? raw_buffered : MAX_BUFFERED_ITEMS;
+
+    compute_cumulative_sum();
+
+    // Find threshold bin for this pass
+    if (thread_id < RADIX && shared_histogram[0][thread_id] > remaining_k &&
+        shared_histogram[0][thread_id + 1] <= remaining_k) {
+      shared_threshold_bin = thread_id;
+      shared_buffered_count[dst_buffer] = 0;
+      shared_final_k = remaining_k - shared_histogram[0][thread_id + 1];
+    }
+    __syncthreads();
+
+    const int threshold_bin = shared_threshold_bin;
+    remaining_k -= shared_histogram[0][threshold_bin + 1];
+
+    // Bit offset for this pass: 24, 16, 8, 0
+    const int bit_offset = 24 - pass * 8;
+
+    // Early exit if threshold bin perfectly matches
+    if (remaining_k == 0) {
+      for (int i = thread_id; i < num_buffered; i += kThreadsPerBlock) {
+        const int idx = buffered_indices[src_buffer][i];
+        const uint32_t fp32_bits =
+            convert_to_uint32_v2(logits[idx + logits_offset]);
+        const int bin = (fp32_bits >> bit_offset) & 0xFF;
+        if (bin > threshold_bin) {
+          const int output_pos = ::atomicAdd(&shared_output_count, 1);
+          output_indices[output_pos] = idx;
+        }
+      }
+      __syncthreads();
+      break;
+    }
+
+    // Continue refinement
+    __syncthreads();
+    if (thread_id < RADIX + 1) {
+      shared_histogram[0][thread_id] = 0;
+    }
+    __syncthreads();
+
+    for (int i = thread_id; i < num_buffered; i += kThreadsPerBlock) {
+      const int idx = buffered_indices[src_buffer][i];
+      const float logit_value = logits[idx + logits_offset];
+      const uint32_t fp32_bits = convert_to_uint32_v2(logit_value);
+      const int bin = (fp32_bits >> bit_offset) & 0xFF;
+
+      if (bin > threshold_bin) {
+        // Definitely in top-k
+        const int output_pos = ::atomicAdd(&shared_output_count, 1);
+        output_indices[output_pos] = idx;
+      } else if (bin == threshold_bin) {
+        if (pass == 3) {
+          // Final pass (bits [7:0]): No more refinement possible
+          // Fill remaining slots in reverse order to maintain descending order
+          const int slot = ::atomicAdd(&shared_final_k, -1);
+          if (slot > 0) {
+            output_indices[TopK - slot] = idx;
+          }
+        } else {
+          // Buffer for next pass and build next histogram
+          const int buffer_pos =
+              ::atomicAdd(&shared_buffered_count[dst_buffer], 1);
+          if (C10_LIKELY(buffer_pos < MAX_BUFFERED_ITEMS)) {
+            buffered_indices[dst_buffer][buffer_pos] = idx;
+            // Fused: Build histogram for next pass
+            const int next_bit_offset = bit_offset - 8;
+            const int next_bin = (fp32_bits >> next_bit_offset) & 0xFF;
+            ::atomicAdd(&shared_histogram[0][next_bin], 1);
+          }
+        }
+      }
+    }
+    __syncthreads();
+  }
+}
+
+__global__ __launch_bounds__(kThreadsPerBlock) void topk_kernel(
+    const FastTopKParams params) {
+  const auto& [input, row_starts, indices, lengths, input_stride] = params;
+  const uint64_t batch_idx = blockIdx.x;
+  const int logits_offset = row_starts == nullptr ? 0 : row_starts[batch_idx];
+  const int seq_len = lengths[batch_idx];
+  int* output_indices = indices + batch_idx * TopK;
+  const float* logits = input + batch_idx * input_stride;
+
+  if (seq_len <= TopK) {
+    // Shortcut: All elements are in top-k
+    return naive_topk_cuda(logits, output_indices, seq_len);
+  } else {
+    return fast_topk_cuda_tl(logits, output_indices, logits_offset, seq_len);
+  }
+}
+
+FastTopKParams get_params(
+    const at::Tensor& score, const at::Tensor& lengths,
+    std::optional<at::Tensor> row_starts_opt = std::nullopt,
+    std::optional<at::Tensor> indices_opt = std::nullopt) {
+  const int64_t batch_size = score.size(0);
+
+  TORCH_CHECK(score.dim() == 2 && score.stride(1) == 1,
+              "score must be 2D with contiguous rows");
+  TORCH_CHECK(lengths.dim() == 1 && lengths.is_contiguous() &&
+                  lengths.size(0) == batch_size,
+              "lengths must be 1D contiguous with size matching batch");
+
+  const int32_t* row_starts_ptr = nullptr;
+  if (row_starts_opt.has_value()) {
+    const auto& row_starts = *row_starts_opt;
+    TORCH_CHECK(row_starts.dim() == 1 && row_starts.size(0) == batch_size,
+                "row_starts must be 1D with size matching batch");
+    row_starts_ptr = row_starts.data_ptr<int32_t>();
+  }
+
+  int32_t* indices_ptr = nullptr;
+  if (indices_opt.has_value()) {
+    const auto& indices = *indices_opt;
+    TORCH_CHECK(indices.dim() == 2 && indices.is_contiguous() &&
+                    indices.size(0) == batch_size && indices.size(1) == TopK,
+                "indices must be 2D contiguous [batch, TopK]");
+    indices_ptr = indices.data_ptr<int32_t>();
+  }
+
+  return FastTopKParams{
+      .input = score.data_ptr<float>(),
+      .row_starts = row_starts_ptr,
+      .indices = indices_ptr,
+      .lengths = lengths.data_ptr<int32_t>(),
+      .input_stride = score.stride(0),
+  };
+}
+
+template <auto* kernel_func, size_t smem_bytes>
+void setup_kernel_smem_once() {
+  static const cudaError_t result = []() -> cudaError_t {
+#ifdef USE_ROCM
+    auto func_ptr = reinterpret_cast<const void*>(kernel_func);
+#else
+    auto func_ptr = kernel_func;
+#endif
+    return cudaFuncSetAttribute(
+        func_ptr, cudaFuncAttributeMaxDynamicSharedMemorySize, smem_bytes);
+  }();
+
+  TORCH_CHECK(
+      result == cudaSuccess,
+      "Failed to set kernel shared memory limit: ", cudaGetErrorString(result));
+}
+
+}  // namespace vllm
+
+void large_context_topk(
+    const torch::Tensor& logits, torch::Tensor& indices,
+    const torch::Tensor& seq_lens,
+    std::optional<torch::Tensor> row_starts = std::nullopt) {
+  TORCH_CHECK(logits.is_cuda(), "logits must be a CUDA tensor");
+  TORCH_CHECK(indices.is_cuda(), "indices must be a CUDA tensor");
+  TORCH_CHECK(seq_lens.is_cuda(), "seq_lens must be a CUDA tensor");
+  if (row_starts.has_value()) {
+    TORCH_CHECK(row_starts->is_cuda(), "row_starts must be a CUDA tensor");
+  }
+
+  const auto params = vllm::get_params(logits, seq_lens, row_starts, indices);
+  const int64_t batch_size = logits.size(0);
+
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  const dim3 grid(static_cast<uint32_t>(batch_size));
+  const dim3 block(vllm::kThreadsPerBlock);
+
+  vllm::setup_kernel_smem_once<vllm::topk_kernel, vllm::kSmem>();
+  vllm::topk_kernel<<<grid, block, vllm::kSmem, stream>>>(params);
+
+  const cudaError_t result = cudaGetLastError();
+  TORCH_CHECK(result == cudaSuccess,
+              "large_context_topk kernel failed: ", cudaGetErrorString(result));
+}

csrc/torch_bindings.cpp

@@ -190,6 +190,12 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "int numRows, int stride0, int stride1, int topK) -> ()");
   ops.impl("top_k_per_row_decode", torch::kCUDA, &top_k_per_row_decode);
 
+  ops.def(
+      "large_context_topk(Tensor score, Tensor indices, Tensor lengths, "
+      "Tensor? "
+      "row_starts_opt) -> ()");
+  ops.impl("large_context_topk", torch::kCUDA, &large_context_topk);
+
   // Layernorm-quant
   // Apply Root Mean Square (RMS) Normalization to the input tensor.
   ops.def(

docker/Dockerfile

@@ -320,7 +320,7 @@ WORKDIR /workspace
 
 # Build DeepGEMM wheel
 # Default moved here from tools/install_deepgemm.sh for centralized version management
-ARG DEEPGEMM_GIT_REF=594953acce41793ae00a1233eb516044d604bcb6
+ARG DEEPGEMM_GIT_REF=477618cd51baffca09c4b0b87e97c03fe827ef03
 COPY tools/install_deepgemm.sh /tmp/install_deepgemm.sh
 RUN --mount=type=cache,target=/root/.cache/uv \
     mkdir -p /tmp/deepgemm/dist && \
@@ -582,7 +582,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # This is ~1.1GB and only changes when FlashInfer version bumps
 # https://docs.flashinfer.ai/installation.html
 # From versions.json: .flashinfer.version
-ARG FLASHINFER_VERSION=0.6.2
+ARG FLASHINFER_VERSION=0.6.3
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system flashinfer-cubin==${FLASHINFER_VERSION} \
     && uv pip install --system flashinfer-jit-cache==${FLASHINFER_VERSION} \

docker/Dockerfile.cpu

@@ -134,7 +134,6 @@ WORKDIR /vllm-workspace
 # Copy test requirements
 COPY requirements/test.in requirements/cpu-test.in
 
-# TODO: Update to 2.9.0 when there is a new build for intel_extension_for_pytorch for that version
 RUN \
     sed -i '/mamba_ssm/d' requirements/cpu-test.in && \
     remove_packages_not_supported_on_aarch64() { \

docker/Dockerfile.nightly_torch

@@ -217,13 +217,13 @@ RUN pip install setuptools==75.6.0 packaging==23.2 ninja==1.11.1.3 build==1.2.2.
 
 
 # build flashinfer for torch nightly from source around 10 mins
-# release version: v0.6.2
+# release version: v0.6.3
 # todo(elainewy): cache flashinfer build result for faster build
 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
     --mount=type=cache,target=/root/.cache/uv \
     echo "git clone flashinfer..." \
-    && git clone --depth 1 --branch v0.6.2 --recursive https://github.com/flashinfer-ai/flashinfer.git \
+    && git clone --depth 1 --branch v0.6.3 --recursive https://github.com/flashinfer-ai/flashinfer.git \
     && cd flashinfer \
     && git submodule update --init --recursive \
     && echo "finish git clone flashinfer..." \

docker/Dockerfile.rocm_base

@@ -9,7 +9,7 @@ ARG PYTORCH_AUDIO_BRANCH="v2.9.0"
 ARG PYTORCH_AUDIO_REPO="https://github.com/pytorch/audio.git"
 ARG FA_BRANCH="0e60e394"
 ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git"
-ARG AITER_BRANCH="6af8b687"
+ARG AITER_BRANCH="v0.1.10.post2"
 ARG AITER_REPO="https://github.com/ROCm/aiter.git"
 ARG MORI_BRANCH="2d02c6a9"
 ARG MORI_REPO="https://github.com/ROCm/mori.git"
@@ -239,7 +239,7 @@ RUN pip install pyyaml && cd aiter \
            export HIP_CLANG_PATH=/opt/sccache-wrappers \
            && sccache --show-stats; \
        fi \
-    && PREBUILD_KERNELS=1 GPU_ARCHS=${AITER_ROCM_ARCH} python3 setup.py bdist_wheel --dist-dir=dist \
+    && GPU_ARCHS=${AITER_ROCM_ARCH} python3 setup.py bdist_wheel --dist-dir=dist \
     && if [ "$USE_SCCACHE" = "1" ]; then sccache --show-stats; fi \
     && ls /app/aiter/dist/*.whl
 RUN mkdir -p /app/install && cp /app/aiter/dist/*.whl /app/install

docker/Dockerfile.xpu

@@ -1,5 +1,10 @@
 FROM intel/deep-learning-essentials:2025.3.2-0-devel-ubuntu24.04 AS vllm-base
 
+WORKDIR /workspace/
+
+ARG PYTHON_VERSION=3.12
+ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/xpu"
+
 RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
     echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \
     add-apt-repository -y ppa:kobuk-team/intel-graphics
@@ -22,13 +27,16 @@ RUN apt clean && apt-get update -y && \
     python3.12-dev \
     python3-pip
 
-RUN update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.12 1
-RUN update-alternatives --install /usr/bin/python python /usr/bin/python3.12 1
-
 RUN apt update && apt upgrade -y && \
     apt install -y libze1 libze-dev libze-intel-gpu1 intel-opencl-icd libze-intel-gpu-raytracing intel-ocloc && \
     apt install -y intel-oneapi-compiler-dpcpp-cpp-2025.3
 
+ENV PATH="/root/.local/bin:$PATH"
+ENV VIRTUAL_ENV="/opt/venv"
+ENV UV_PYTHON_INSTALL_DIR=/opt/uv/python
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh
+RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
+ENV PATH="$VIRTUAL_ENV/bin:$PATH"
 
 # This oneccl contains the BMG support which is not the case for default version of oneapi 2025.2.
 ARG ONECCL_INSTALLER="intel-oneccl-2021.15.7.8_offline.sh"
@@ -44,20 +52,31 @@ SHELL ["bash", "-c"]
 CMD ["bash", "-c", "source /root/.bashrc && exec bash"]
 
 WORKDIR /workspace/vllm
-COPY requirements/xpu.txt /workspace/vllm/requirements/xpu.txt
-COPY requirements/common.txt /workspace/vllm/requirements/common.txt
 
-# suppress the python externally managed environment error
-RUN python3 -m pip config set global.break-system-packages true
+ENV UV_HTTP_TIMEOUT=500
 
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install --no-cache-dir \
-    -r requirements/xpu.txt
+# Configure package index for XPU
+ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
+ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
+ENV UV_INDEX_STRATEGY="unsafe-best-match"
+ENV UV_LINK_MODE="copy"
 
-# arctic-inference is built from source which needs torch-xpu properly installed
-# used for suffix method speculative decoding
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install --no-cache-dir arctic-inference==0.1.1
+RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,src=requirements/common.txt,target=/workspace/vllm/requirements/common.txt \
+    --mount=type=bind,src=requirements/xpu.txt,target=/workspace/vllm/requirements/xpu.txt \
+    uv pip install --upgrade pip && \
+    uv pip install -r requirements/xpu.txt
+
+ # used for suffix method speculative decoding
+ # build deps for proto + nanobind-based extensions to set up the build environment
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install grpcio-tools protobuf nanobind
+ # arctic-inference is built from source which needs torch-xpu properly installed first
+RUN --mount=type=cache,target=/root/.cache/uv \
+    source /opt/intel/oneapi/setvars.sh --force && \
+    source /opt/intel/oneapi/ccl/2021.15/env/vars.sh --force && \
+    export CMAKE_PREFIX_PATH="$(python -c 'import site; print(site.getsitepackages()[0])'):${CMAKE_PREFIX_PATH}" && \
+    uv pip install --no-build-isolation arctic-inference==0.1.1
 
 ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/lib/"
 
@@ -69,33 +88,32 @@ RUN --mount=type=bind,source=.git,target=.git \
 ENV VLLM_TARGET_DEVICE=xpu
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
 
-RUN --mount=type=cache,target=/root/.cache/pip \
+RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
-    pip install --no-build-isolation .
+    uv pip install --no-build-isolation .
 
 CMD ["/bin/bash"]
 
 FROM vllm-base AS vllm-openai
 
 # install additional dependencies for openai api server
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install accelerate hf_transfer pytest pytest_asyncio lm_eval[api] modelscope
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install accelerate hf_transfer pytest pytest_asyncio lm_eval[api] modelscope
 
 # install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
+RUN uv pip install -e tests/vllm_test_utils
 
 # install nixl from source code
 ENV NIXL_VERSION=0.7.0
-RUN python3 /workspace/vllm/tools/install_nixl_from_source_ubuntu.py
+RUN python /workspace/vllm/tools/install_nixl_from_source_ubuntu.py
 
 # FIX triton
-RUN --mount=type=cache,target=/root/.cache/pip pip uninstall triton triton-xpu -y && pip install triton-xpu==3.6.0 --extra-index-url=https://download.pytorch.org/whl/xpu
-
-# PyJWT-2.7.0 will influence some wheel behaviors, remove its dist-info to avoid conflicts
-RUN rm /usr/lib/python3/dist-packages/PyJWT-2.7.0.dist-info/ -rf
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip uninstall triton triton-xpu && \
+    uv pip install triton-xpu==3.6.0
 
 # remove torch bundled oneccl to avoid conflicts
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip uninstall oneccl oneccl-devel -y
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip uninstall oneccl oneccl-devel
 
 ENTRYPOINT ["vllm", "serve"]

docker/versions.json

@@ -50,7 +50,7 @@
       "default": "cuda"
     },
     "DEEPGEMM_GIT_REF": {
-      "default": "594953acce41793ae00a1233eb516044d604bcb6"
+      "default": "477618cd51baffca09c4b0b87e97c03fe827ef03"
     },
     "PPLX_COMMIT_HASH": {
       "default": "12cecfd"
@@ -68,7 +68,7 @@
       "default": "true"
     },
     "FLASHINFER_VERSION": {
-      "default": "0.6.2"
+      "default": "0.6.3"
     },
     "GDRCOPY_CUDA_VERSION": {
       "default": "12.8"

docs/benchmarking/cli.md

@@ -30,6 +30,7 @@ th {
 | HuggingFace-Other | ✅ | ✅ | `lmms-lab/LLaVA-OneVision-Data`, `Aeala/ShareGPT_Vicuna_unfiltered` |
 | HuggingFace-MTBench | ✅ | ✅ | `philschmid/mt-bench` |
 | HuggingFace-Blazedit | ✅ | ✅ | `vdaita/edit_5k_char`, `vdaita/edit_10k_char` |
+| HuggingFace-ASR | ✅ | ✅ | `openslr/librispeech_asr`, `facebook/voxpopuli`,  `LIUM/tedlium`, `edinburghcstr/ami`,        `speechcolab/gigaspeech`,        `kensho/spgispeech` |
 | Spec Bench | ✅ | ✅ | `wget https://raw.githubusercontent.com/hemingkx/Spec-Bench/refs/heads/main/data/spec_bench/question.jsonl` |
 | Custom | ✅ | ✅ | Local file: `data.jsonl` |
 | Custom MM | ✅ | ✅ | Local file: `mm_data.jsonl` |
@@ -299,6 +300,22 @@ vllm bench serve \
     --blazedit-max-distance 0.99
 ```
 
+`openslr/librispeech_asr`, `facebook/voxpopuli`, `LIUM/tedlium`, `edinburghcstr/ami`, `speechcolab/gigaspeech`, `kensho/spgispeech`
+
+```bash
+vllm bench serve \
+    --model openai/whisper-large-v3-turbo \
+    --backend openai-audio \
+    --dataset-name hf \
+    --dataset-path facebook/voxpopuli --hf-subset en --hf-split test --no-stream --trust-remote-code \
+    --num-prompts 99999999 \
+    --no-oversample \
+    --endpoint /v1/audio/transcriptions \
+    --ready-check-timeout-sec 600 \
+    --save-result \
+    --max-concurrency 512
+```
+
 #### Running With Sampling Parameters
 
 When using OpenAI-compatible backends such as `vllm`, optional sampling

docs/configuration/optimization.md

@@ -291,6 +291,52 @@ Based on the configuration, the content of the multi-modal caches on `P0` and `P
 K: Stores the hashes of multi-modal items
 V: Stores the processed tensor data of multi-modal items
 
+## CPU Resources for GPU Deployments
+
+vLLM V1 uses a multi-process architecture (see [V1 Process Architecture](../design/arch_overview.md#v1-process-architecture)) where each process requires CPU resources. Underprovisioning CPU cores is a common source of performance degradation, especially in virtualized environments.
+
+### Minimum CPU Requirements
+
+For a deployment with `N` GPUs, there are at minimum:
+
+- **1 API server process** -- handles HTTP requests, tokenization, and input processing
+- **1 engine core process** -- runs the scheduler and coordinates GPU workers
+- **N GPU worker processes** -- one per GPU, executes model forward passes
+
+This means there are always at least **`2 + N` processes** competing for CPU time.
+
+!!! warning
+    Using fewer physical CPU cores than processes will cause contention and significantly degrade throughput and latency. The engine core process runs a busy loop and is particularly sensitive to CPU starvation.
+
+The minimum is `2 + N` physical cores (1 for the API server, 1 for the engine core, and 1 per GPU worker). In practice, allocating more cores improves performance because the OS, PyTorch background threads, and other system processes also need CPU time.
+
+!!! important
+    Please note we are referring to **physical CPU cores** here. If your system has hyperthreading enabled, then 1 vCPU = 1 hyperthread = 1/2 physical CPU core, so you need `2 x (2 + N)` minimum vCPUs.
+
+### Data Parallel and Multi-API Server Deployments
+
+When using data parallelism or multiple API servers, the CPU requirements increase:
+
+```console
+Minimum physical cores = A + DP + N + (1 if DP > 1 else 0)
+```
+
+where `A` is the API server count (defaults to `DP`), `DP` is the data parallel size, and `N` is the total number of GPUs. For example, with `DP=4, TP=2` on 8 GPUs:
+
+```console
+4 API servers + 4 engine cores + 8 GPU workers + 1 DP coordinator = 17 processes
+```
+
+### Performance Impact
+
+CPU underprovisioning particularly impacts:
+
+- **Input processing throughput** -- tokenization, chat template rendering, and multi-modal data loading all run on CPU
+- **Scheduling latency** -- the engine core scheduler runs on CPU and directly affects how quickly new tokens are dispatched to the GPU workers
+- **Output processing** -- detokenization, networking, and especially streaming token responses use CPU cycles
+
+If you observe that GPU utilization is lower than expected, CPU contention may be the bottleneck. Increasing the number of available CPU cores and even the clock speed can significantly improve end-to-end performance.
+
 ## Attention Backend Selection
 
 vLLM supports multiple attention backends optimized for different hardware and use cases. The backend is automatically selected based on your GPU architecture, model type, and configuration, but you can also manually specify one for optimal performance.

docs/contributing/model/basic.md

@@ -138,7 +138,7 @@ These models should follow the same instructions as case (1), but they should in
 
 For case (3), we recommend looking at the implementation of [`MiniMaxText01ForCausalLM`](../../../vllm/model_executor/models/minimax_text_01.py) or [`Lfm2ForCausalLM`](../../../vllm/model_executor/models/lfm2.py) as a reference, which use custom "mamba-like" layers `MiniMaxText01LinearAttention` and `ShortConv` respectively.
 Please follow the same guidelines as case (2) for implementing these models.
-We use "mamba-like" to refer to layers that posses a state that is updated in-place, rather than being appended-to (like KV cache for attention).
+We use "mamba-like" to refer to layers that possess a state that is updated in-place, rather than being appended-to (like KV cache for attention).
 For implementing new custom mamba-like layers, one should inherit from `MambaBase` and implement the methods `get_state_dtype`, `get_state_shape` to calculate the data types and state shapes at runtime, as well as `mamba_type` and `get_attn_backend`.
 It is also necessary to implement the "attention meta-data" class which handles the meta-data that is common across all layers.
 Please see [`LinearAttentionMetadata`](../../../vllm/v1/attention/backends/linear_attn.py) or [`ShortConvAttentionMetadata`](../../../vllm/v1/attention/backends/short_conv_attn.py) for examples of this.

docs/contributing/model/multimodal.md

@@ -739,7 +739,7 @@ Each [PromptUpdate][vllm.multimodal.processing.PromptUpdate] instance specifies
         ```
 
     However, this is not entirely correct. After `FuyuImageProcessor.preprocess_with_tokenizer_info` is called,
-    a BOS token (`<s>`) is also added to the promopt:
+    a BOS token (`<s>`) is also added to the prompt:
 
     ??? code
 

docs/deployment/docker.md

@@ -1,161 +1,13 @@
+---
+toc_depth: 2
+---
+
 # Using Docker
 
-## Use vLLM's Official Docker Image
+## 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).
+--8<-- "docs/getting_started/installation/gpu.md:pre-built-images"
 
-```bash
-docker run --runtime nvidia --gpus all \
-    -v ~/.cache/huggingface:/root/.cache/huggingface \
-    --env "HF_TOKEN=$HF_TOKEN" \
-    -p 8000:8000 \
-    --ipc=host \
-    vllm/vllm-openai:latest \
-    --model Qwen/Qwen3-0.6B
-```
+## Build image from source
 
-This image can also be used with other container engines such as [Podman](https://podman.io/).
-
-```bash
-podman run --device nvidia.com/gpu=all \
-  -v ~/.cache/huggingface:/root/.cache/huggingface \
-  --env "HF_TOKEN=$HF_TOKEN" \
-  -p 8000:8000 \
-  --ipc=host \
-  docker.io/vllm/vllm-openai:latest \
-  --model Qwen/Qwen3-0.6B
-```
-
-You can add any other [engine-args](../configuration/engine_args.md) you need after the image tag (`vllm/vllm-openai:latest`).
-
-!!! note
-    You can either use the `ipc=host` flag or `--shm-size` flag to allow the
-    container to access the host's shared memory. vLLM uses PyTorch, which uses shared
-    memory to share data between processes under the hood, particularly for tensor parallel inference.
-
-!!! note
-    Optional dependencies are not included in order to avoid licensing issues (e.g. <https://github.com/vllm-project/vllm/issues/8030>).
-
-    If you need to use those dependencies (having accepted the license terms),
-    create a custom Dockerfile on top of the base image with an extra layer that installs them:
-
-    ```Dockerfile
-    FROM vllm/vllm-openai:v0.11.0
-
-    # e.g. install the `audio` optional dependencies
-    # NOTE: Make sure the version of vLLM matches the base image!
-    RUN uv pip install --system vllm[audio]==0.11.0
-    ```
-
-!!! tip
-    Some new models may only be available on the main branch of [HF Transformers](https://github.com/huggingface/transformers).
-
-    To use the development version of `transformers`, create a custom Dockerfile on top of the base image
-    with an extra layer that installs their code from source:
-
-    ```Dockerfile
-    FROM vllm/vllm-openai:latest
-
-    RUN uv pip install --system git+https://github.com/huggingface/transformers.git
-    ```
-
-## Building vLLM's Docker Image from Source
-
-You can build and run vLLM from source via the provided [docker/Dockerfile](../../docker/Dockerfile). To build vLLM:
-
-```bash
-# optionally specifies: --build-arg max_jobs=8 --build-arg nvcc_threads=2
-DOCKER_BUILDKIT=1 docker build . \
-    --target vllm-openai \
-    --tag vllm/vllm-openai \
-    --file docker/Dockerfile
-```
-
-!!! note
-    By default vLLM will build for all GPU types for widest distribution. If you are just building for the
-    current GPU type the machine is running on, you can add the argument `--build-arg torch_cuda_arch_list=""`
-    for vLLM to find the current GPU type and build for that.
-
-    If you are using Podman instead of Docker, you might need to disable SELinux labeling by
-    adding `--security-opt label=disable` when running `podman build` command to avoid certain [existing issues](https://github.com/containers/buildah/discussions/4184).
-
-!!! note
-    If you have not changed any C++ or CUDA kernel code, you can use precompiled wheels to significantly reduce Docker build time.
-
-    *   **Enable the feature** by adding the build argument: `--build-arg VLLM_USE_PRECOMPILED="1"`.
-    *   **How it works**: By default, vLLM automatically finds the correct wheels from our [Nightly Builds](../contributing/ci/nightly_builds.md) by using the merge-base commit with the upstream `main` branch.
-    *   **Override commit**: To use wheels from a specific commit, provide the `--build-arg VLLM_PRECOMPILED_WHEEL_COMMIT=<commit_hash>` argument.
-
-    For a detailed explanation, refer to the documentation on 'Set up using Python-only build (without compilation)' part in [Build wheel from source](../contributing/ci/nightly_builds.md#precompiled-wheels-usage), these args are similar.
-
-## Building for Arm64/aarch64
-
-A docker container can be built for aarch64 systems such as the Nvidia Grace-Hopper and Grace-Blackwell. Using the flag `--platform "linux/arm64"` will build for arm64.
-
-!!! note
-    Multiple modules must be compiled, so this process can take a while. Recommend using `--build-arg max_jobs=` & `--build-arg nvcc_threads=`
-    flags to speed up build process. However, ensure your `max_jobs` is substantially larger than `nvcc_threads` to get the most benefits.
-    Keep an eye on memory usage with parallel jobs as it can be substantial (see example below).
-
-??? console "Command"
-
-    ```bash
-    # Example of building on Nvidia GH200 server. (Memory usage: ~15GB, Build time: ~1475s / ~25 min, Image size: 6.93GB)
-    DOCKER_BUILDKIT=1 docker build . \
-    --file docker/Dockerfile \
-    --target vllm-openai \
-    --platform "linux/arm64" \
-    -t vllm/vllm-gh200-openai:latest \
-    --build-arg max_jobs=66 \
-    --build-arg nvcc_threads=2 \
-    --build-arg torch_cuda_arch_list="9.0 10.0+PTX" \
-    --build-arg RUN_WHEEL_CHECK=false
-    ```
-
-For (G)B300, we recommend using CUDA 13, as shown in the following command.
-
-??? console "Command"
-
-    ```bash
-    DOCKER_BUILDKIT=1 docker build \
-    --build-arg CUDA_VERSION=13.0.1 \
-    --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 \
-    --build-arg max_jobs=256 \
-    --build-arg nvcc_threads=2 \
-    --build-arg RUN_WHEEL_CHECK=false \
-    --build-arg torch_cuda_arch_list='9.0 10.0+PTX' \
-    --platform "linux/arm64" \
-    --tag vllm/vllm-gb300-openai:latest \
-    --target vllm-openai \
-    -f docker/Dockerfile \
-    .
-    ```
-
-!!! note
-    If you are building the `linux/arm64` image on a non-ARM host (e.g., an x86_64 machine), you need to ensure your system is set up for cross-compilation using QEMU. This allows your host machine to emulate ARM64 execution.
-
-    Run the following command on your host machine to register QEMU user static handlers:
-
-    ```bash
-    docker run --rm --privileged multiarch/qemu-user-static --reset -p yes
-    ```
-
-    After setting up QEMU, you can use the `--platform "linux/arm64"` flag in your `docker build` command.
-
-## Use the custom-built vLLM Docker image
-
-To run vLLM with the custom-built Docker image:
-
-```bash
-docker run --runtime nvidia --gpus all \
-    -v ~/.cache/huggingface:/root/.cache/huggingface \
-    -p 8000:8000 \
-    --env "HF_TOKEN=<secret>" \
-    vllm/vllm-openai <args...>
-```
-
-The argument `vllm/vllm-openai` specifies the image to run, and should be replaced with the name of the custom-built image (the `-t` tag from the build command).
-
-!!! 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<-- "docs/getting_started/installation/gpu.md:build-image-from-source"

docs/design/arch_overview.md

@@ -78,6 +78,73 @@ That code can be found in [vllm/entrypoints/openai/api_server.py](../../vllm/ent
 
 More details on the API server can be found in the [OpenAI-Compatible Server](../serving/openai_compatible_server.md) document.
 
+## V1 Process Architecture
+
+vLLM V1 uses a multi-process architecture to separate concerns and maximize throughput. Understanding this architecture is important for properly sizing CPU resources in your deployment. The key processes are:
+
+### API Server Process
+
+The API server process handles HTTP requests (e.g., the OpenAI-compatible API), performs input processing (tokenization, multi-modal data loading), and streams results back to clients. It communicates with the engine core process(es) via ZMQ sockets.
+
+By default, there is **1 API server process**, but when data parallelism is used, the API server count automatically scales to match the data parallel size. This can also be manually configured with the `--api-server-count` flag. Each API server connects to **all** engine cores via ZMQ in a many-to-many topology, enabling any API server to route requests to any engine core. Each API server process uses multiple CPU threads for media loading (controlled by `VLLM_MEDIA_LOADING_THREAD_COUNT`, default 8).
+
+The code can be found in [vllm/entrypoints/openai/api_server.py](../../vllm/entrypoints/openai/api_server.py) and [vllm/v1/utils.py](../../vllm/v1/utils.py).
+
+### Engine Core Process
+
+The engine core process runs the scheduler, manages KV cache, and coordinates model execution across GPU workers. It runs a busy loop that continuously schedules requests and dispatches work to the GPU workers.
+
+There is **1 engine core process per data parallel rank**. For example, with `--data-parallel-size 4`, there are 4 engine core processes.
+
+The code can be found in [vllm/v1/engine/core.py](../../vllm/v1/engine/core.py) and [vllm/v1/engine/utils.py](../../vllm/v1/engine/utils.py).
+
+### GPU Worker Processes
+
+Each GPU is managed by a dedicated worker process. The worker process loads model weights, executes forward passes, and manages GPU memory. Workers communicate with the engine core process that owns them.
+
+There is **1 worker process per GPU**. The total number of GPU worker processes equals `tensor_parallel_size x pipeline_parallel_size` per engine core.
+
+The code can be found in [vllm/v1/executor/multiproc_executor.py](../../vllm/v1/executor/multiproc_executor.py) and [vllm/v1/worker/gpu_worker.py](../../vllm/v1/worker/gpu_worker.py).
+
+### DP Coordinator Process (conditional)
+
+When using data parallelism (`--data-parallel-size > 1`), an additional coordinator process manages load balancing across DP ranks and coordinates synchronized forward passes for MoE models.
+
+There is **1 DP coordinator process** (only when data parallelism is enabled).
+
+The code can be found in [vllm/v1/engine/coordinator.py](../../vllm/v1/engine/coordinator.py).
+
+### Process Count Summary
+
+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 TP`) | One per GPU, executes model forward passes |
+| DP Coordinator | 1 if `DP > 1`, else 0 | Load balancing across DP ranks |
+| **Total** | **`A + DP + N` (+ 1 if DP > 1)** | |
+
+For example, a typical single-node deployment with 4 GPUs (`vllm serve -tp=4`) has:
+
+- 1 API server + 1 engine core + 4 GPU workers = **6 processes**
+
+<figure markdown="1">
+![V1 Process Architecture - TP=4](../assets/design/arch_overview/v1_process_architecture_tp4.png)
+</figure>
+
+A data parallel deployment with 8 GPUs (`vllm serve -tp=2 -dp=4`) has:
+
+- 4 API servers + 4 engine cores + 8 GPU workers + 1 DP coordinator = **17 processes**
+
+<figure markdown="1">
+![V1 Process Architecture - TP=2, DP=4](../assets/design/arch_overview/v1_process_architecture_tp2_dp4.png)
+</figure>
+
+For CPU resource sizing recommendations, see
+[CPU Resources for GPU Deployments](../configuration/optimization.md#cpu-resources-for-gpu-deployments).
+
 ## LLM Engine
 
 The `LLMEngine` and `AsyncLLMEngine` classes are central to the functioning of

docs/design/attention_backends.md

@@ -128,6 +128,7 @@ Priority is **1 = highest** (tried first).
 | 4 | `FLASHMLA` |
 | 5 | `TRITON_MLA` |
 | 6 | `FLASHMLA_SPARSE` |
+| 7 | `FLASHINFER_MLA_SPARSE` |
 
 **Ampere/Hopper (SM 8.x-9.x):**
 
@@ -152,6 +153,7 @@ Priority is **1 = highest** (tried first).
 | **Sink** | Attention sink support (for StreamingLLM) |
 | **Sparse** | Sparse attention support (MLA only) |
 | **MM Prefix** | Multimodal prefix full attention support |
+| **DCP** | Decode Context Parallelism support (`--decode-context-parallel-size`) |
 | **Attention Types** | Supported attention patterns (Decoder, Encoder, Enc-Dec) |
 | **Compute Cap.** | Required CUDA compute capability (N/A for non-CUDA backends) |
 
@@ -159,20 +161,20 @@ Priority is **1 = highest** (tried first).
 
 ## Standard Attention (MHA, MQA, GQA) Backends
 
-| Backend | Version | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | MM Prefix | Attention Types | Compute Cap. |
-|---------|---------|--------|-----------|-------------|------------|------|-----------|-----------------|--------------|
-| `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_DIFFKV` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | Decoder | Any |
-| `FLEX_ATTENTION` |  | fp16, bf16, fp32 | `auto`, `bfloat16` | Any | Any | ❌ | ✅ | Decoder, Encoder Only | Any |
-| `ROCM_AITER_FA` |  | fp16, bf16 | `auto` | 16, 32 | 64, 128, 256 | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_UNIFIED_ATTN` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | Decoder | N/A |
-| `ROCM_ATTN` |  | fp16, bf16, fp32 | `auto` | 16, 32, 544 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | Decoder | N/A |
-| `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 |
+| 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 |
+| `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_DIFFKV` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ✅ | Decoder | Any |
+| `FLEX_ATTENTION` |  | fp16, bf16, fp32 | `auto`, `bfloat16` | Any | Any | ❌ | ✅ | ❌ | Decoder, Encoder Only | Any |
+| `ROCM_AITER_FA` |  | fp16, bf16 | `auto` | 16, 32 | 64, 128, 256 | ❌ | ❌ | ❌ | Decoder | N/A |
+| `ROCM_AITER_UNIFIED_ATTN` |  | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | Decoder | N/A |
+| `ROCM_ATTN` |  | fp16, bf16, fp32 | `auto` | 16, 32, 544 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | N/A |
+| `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`.
 >
@@ -199,14 +201,15 @@ configuration.
 
 ### Decode Backends
 
-| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Sparse | MM Prefix | 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 |
-| `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` | 1 | Any | ❌ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto` | Any | 576 | ❌ | ❌ | ❌ | 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 |
+| 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 |
+| `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` | 1 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
+| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto` | Any | 576 | ❌ | ❌ | ❌ | ❌ | 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 |

docs/design/io_processor_plugins.md

@@ -14,8 +14,26 @@ IOProcessorOutput = TypeVar("IOProcessorOutput")
 
 class IOProcessor(ABC, Generic[IOProcessorInput, IOProcessorOutput]):
     def __init__(self, vllm_config: VllmConfig):
+        super().__init__()
+
         self.vllm_config = vllm_config
 
+    @abstractmethod
+    def parse_data(self, data: object) -> IOProcessorInput:
+        raise NotImplementedError
+
+    def merge_sampling_params(
+        self,
+        params: SamplingParams | None = None,
+    ) -> SamplingParams:
+        return params or SamplingParams()
+
+    def merge_pooling_params(
+        self,
+        params: PoolingParams | None = None,
+    ) -> PoolingParams:
+        return params or PoolingParams()
+
     @abstractmethod
     def pre_process(
         self,
@@ -55,29 +73,13 @@ class IOProcessor(ABC, Generic[IOProcessorInput, IOProcessorOutput]):
             [(i, item) async for i, item in model_output], key=lambda output: output[0]
         )
         collected_output = [output[1] for output in sorted_output]
-        return self.post_process(collected_output, request_id, **kwargs)
-
-    @abstractmethod
-    def parse_request(self, request: Any) -> IOProcessorInput:
-        raise NotImplementedError
-
-    def validate_or_generate_params(
-        self, params: SamplingParams | PoolingParams | None = None
-    ) -> SamplingParams | PoolingParams:
-        return params or PoolingParams()
-
-    @abstractmethod
-    def output_to_response(
-        self, plugin_output: IOProcessorOutput
-    ) -> IOProcessorResponse:
-        raise NotImplementedError
+        return self.post_process(collected_output, request_id=request_id, **kwargs)
 ```
 
-The `parse_request` method is used for validating the user prompt and converting it into the input expected by the `pre_process`/`pre_process_async` methods.
+The `parse_data` method is used for validating the user data and converting it into the input expected by the `pre_process*` methods.
+The `merge_sampling_params` and `merge_pooling_params` methods merge input `SamplingParams` or `PoolingParams` (if any) with the default one.
 The `pre_process*` methods take the validated plugin input to generate vLLM's model prompts for regular inference.
 The `post_process*` methods take `PoolingRequestOutput` objects as input and generate a custom plugin output.
-The `validate_or_generate_params` method is used for validating with the plugin any `SamplingParameters`/`PoolingParameters` received with the user request, or to generate new ones if none are specified. The function always returns the validated/generated parameters.
-The `output_to_response` method is used only for online serving and converts the plugin output to the `IOProcessorResponse` type that is then returned by the API Server. The implementation of the `/pooling` serving endpoint is available here [vllm/entrypoints/openai/serving_pooling.py](../../vllm/entrypoints/pooling/pooling/serving.py).
 
 An example implementation of a plugin that enables generating geotiff images with the PrithviGeospatialMAE model is available [here](https://github.com/IBM/terratorch/tree/main/terratorch/vllm/plugins/segmentation). Please, also refer to our online ([examples/pooling/plugin/prithvi_geospatial_mae_online.py](../../examples/pooling/plugin/prithvi_geospatial_mae_online.py)) and offline ([examples/pooling/plugin/prithvi_geospatial_mae_io_processor.py](../../examples/pooling/plugin/prithvi_geospatial_mae_io_processor.py)) inference examples.
 

docs/design/moe_kernel_features.md

@@ -32,7 +32,7 @@ th {
 
 | Backend | Output act. format | Quant. types | Quant. format | Async | Apply Weight On Input | Subclass |
 |---------|--------------------|--------------|---------------|-------|-----------------------|-----------|
-| naive | standard | all<sup>1</sup> | G,A,T | N | <sup>6</sup> | [layer.py][vllm.model_executor.layers.fused_moe.layer.FusedMoE.forward_impl] |
+| naive | standard | all<sup>1</sup> | G,A,T | N | <sup>6</sup> | [layer.py][vllm.model_executor.layers.fused_moe.layer.FusedMoE |
 | pplx | batched | fp8,int8 | G,A,T | Y | Y | [`PplxPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.pplx_prepare_finalize.PplxPrepareAndFinalize] |
 | deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize.DeepEPLLPrepareAndFinalize] |
 | deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize.DeepEPHTPrepareAndFinalize] |

docs/design/plugin_system.md

@@ -24,7 +24,7 @@ vLLM's plugin system uses the standard Python `entry_points` mechanism. This mec
             ["register_dummy_model = vllm_add_dummy_model:register"]
         })
 
-    # inside `vllm_add_dummy_model.py` file
+    # inside `vllm_add_dummy_model/__init__.py` file
     def register():
         from vllm import ModelRegistry
 
@@ -45,7 +45,7 @@ Every plugin has three parts:
 
 ## Types of supported plugins
 
-- **General plugins** (with group name `vllm.general_plugins`): The primary use case for these plugins is to register custom, out-of-the-tree models into vLLM. This is done by calling `ModelRegistry.register_model` to register the model inside the plugin function.
+- **General plugins** (with group name `vllm.general_plugins`): The primary use case for these plugins is to register custom, out-of-the-tree models into vLLM. This is done by calling `ModelRegistry.register_model` to register the model inside the plugin function. For an example of an official model plugin, see the [bart-plugin](https://github.com/vllm-project/bart-plugin) which adds support for `BartForConditionalGeneration`.
 
 - **Platform plugins** (with group name `vllm.platform_plugins`): The primary use case for these plugins is to register custom, out-of-the-tree platforms into vLLM. The plugin function should return `None` when the platform is not supported in the current environment, or the platform class's fully qualified name when the platform is supported.
 

docs/features/multimodal_inputs.md

@@ -510,7 +510,7 @@ Our OpenAI-compatible server accepts multi-modal data via the [Chat Completions
     If no fallback is available, an error is raised and you have to provide the chat template manually via the `--chat-template` argument.
 
     For certain models, we provide alternative chat templates inside [examples](../../examples).
-    For example, VLM2Vec uses [examples/template_vlm2vec_phi3v.jinja](../../examples/template_vlm2vec_phi3v.jinja) which is different from the default one for Phi-3-Vision.
+    For example, VLM2Vec uses [examples/pooling/embed/template/vlm2vec_phi3v.jinja](../../examples/pooling/embed/template/vlm2vec_phi3v.jinja) which is different from the default one for Phi-3-Vision.
 
 ### Image Inputs
 
@@ -521,7 +521,7 @@ First, launch the OpenAI-compatible server:
 
 ```bash
 vllm serve microsoft/Phi-3.5-vision-instruct --runner generate \
-  --trust-remote-code --max-model-len 4096 --limit-mm-per-prompt '{"image":2}'
+  --trust-remote-code --max-model-len 4096 --limit-mm-per-prompt.image 2
 ```
 
 Then, you can use the OpenAI client as follows:

docs/features/nixl_connector_usage.md

@@ -213,6 +213,15 @@ Support use case: Prefill with 'HND' and decode with 'NHD' with experimental con
 --kv-transfer-config '{..., "enable_permute_local_kv":"True"}'
 ```
 
+### Cross layers blocks
+
+By default, this feature is disabled. On attention backends that support this feature, each logical block is contiguous in physical memory. This reduces the number of buffers that need to be transferred.
+To enable this feature:
+
+```bash
+--kv-transfer-config '{..., "kv_connector_extra_config": {"enable_cross_layers_blocks": "True"}}'
+```
+
 ## Example Scripts/Code
 
 Refer to these example scripts in the vLLM repository:

docs/features/quantization/README.md

@@ -48,7 +48,7 @@ th:not(:first-child) {
 |-----------------------|---------|----------|----------|-------|----------|-----------|-------------|-----------|
 | AWQ                   | ❌      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ✅︎          | ✅︎        |
 | GPTQ                  | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ✅︎          | ✅︎        |
-| Marlin (GPTQ/AWQ/FP8) | ❌      | ❌       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ❌        |
+| Marlin (GPTQ/AWQ/FP8/FP4) | ❌      | ✅︎*       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ❌        |
 | INT8 (W8A8)           | ❌      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ✅︎        |
 | FP8 (W8A8)            | ❌      | ❌       | ❌       | ✅︎    | ✅︎       | ✅︎         | ❌          | ❌        |
 | bitsandbytes          | ✅︎      | ✅︎       | ✅︎       | ✅︎    | ✅︎       | ❌         | ❌          | ❌        |
@@ -59,6 +59,7 @@ th:not(:first-child) {
 - ✅︎ indicates that the quantization method is supported on the specified hardware.
 - ❌ indicates that the quantization method is not supported on the specified hardware.
 - All Intel Gaudi quantization support has been migrated to [vLLM-Gaudi](https://github.com/vllm-project/vllm-gaudi).
+- *Turing does not support Marlin MXFP4.
 
 !!! note
     For information on quantization 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/quantization/fp8.md

@@ -2,7 +2,7 @@
 
 vLLM supports FP8 (8-bit floating point) weight and activation quantization using hardware acceleration on GPUs such as Nvidia H100 and AMD MI300x.
 Currently, only Hopper and Ada Lovelace GPUs are officially supported for W8A8.
-Ampere GPUs are supported for W8A16 (weight-only FP8) utilizing Marlin kernels.
+Turing/Ampere GPUs are supported for W8A16 (weight-only FP8) utilizing Marlin kernels.
 Quantization of models with FP8 allows for a 2x reduction in model memory requirements and up to a 1.6x improvement in throughput with minimal impact on accuracy.
 
 Please visit the HF collection of [quantized FP8 checkpoints of popular LLMs ready to use with vLLM](https://huggingface.co/collections/neuralmagic/fp8-llms-for-vllm-666742ed2b78b7ac8df13127).
@@ -13,8 +13,8 @@ The FP8 types typically supported in hardware have two distinct representations,
 - **E5M2**: Consists of 1 sign bit, 5 exponent bits, and 2 bits of mantissa. It can store values up to +/-57344, +/- `inf`, and `nan`. The tradeoff for the increased dynamic range is lower precision of the stored values.
 
 !!! note
-    FP8 computation is supported on NVIDIA GPUs with compute capability > 8.9 (Ada Lovelace, Hopper).
-    FP8 models will run on compute capability > 8.0 (Ampere) as weight-only W8A16, utilizing FP8 Marlin.
+    FP8 computation is supported on NVIDIA GPUs with compute capability >= 8.9 (Ada Lovelace, Hopper).
+    FP8 models will run on compute capability >= 7.5 (Turing) as weight-only W8A16, utilizing FP8 Marlin.
 
 ## Installation
 

docs/features/quantization/modelopt.md

@@ -17,6 +17,7 @@ following `quantization.quant_algo` values:
 - `FP8_PER_CHANNEL_PER_TOKEN`: per-channel weight scale and dynamic per-token activation quantization.
 - `FP8_PB_WO` (ModelOpt may emit `fp8_pb_wo`): block-scaled FP8 weight-only (typically 128×128 blocks).
 - `NVFP4`: ModelOpt NVFP4 checkpoints (use `quantization="modelopt_fp4"`).
+- `MXFP8`: ModelOpt MXFP8 checkpoints (use `quantization="modelopt_mxfp8"`).
 
 ## Quantizing HuggingFace Models with PTQ
 

docs/features/spec_decode/README.md

@@ -1,10 +1,5 @@
 # Speculative Decoding
 
-!!! warning
-    Please note that speculative decoding in vLLM is not yet optimized and does
-    not usually yield inter-token latency reductions for all prompt datasets or sampling parameters.
-    The work to optimize it is ongoing and can be followed here: <https://github.com/vllm-project/vllm/issues/4630>
-
 !!! warning
     Currently, speculative decoding in vLLM is not compatible with pipeline parallelism.
 

docs/getting_started/installation/cpu.md

@@ -176,7 +176,7 @@ For the full and up-to-date list of models validated on CPU platforms, please se
 
 ### How to find benchmark configuration examples for supported CPU models?
 
-For any model listed under [Supported Models on CPU](../../models/hardware_supported_models/cpu.md), optimized runtime configurations are provided in the vLLM Benchmark Suite’s CPU test cases, defined in [cpu test cases](../../../.buildkite/performance-benchmarks/tests/serving-tests-cpu.json)
+For any model listed under [Supported Models on CPU](../../models/hardware_supported_models/cpu.md), optimized runtime configurations are provided in the vLLM Benchmark Suite’s CPU test cases, defined in cpu test cases as serving-tests-cpu.json. Full test cases for Text-only models, Multi-Modal models and Embedded models are in cpu Text-Only test cases as serving-tests-cpu-text.json, cpu Multi-Modal test cases as serving-tests-cpu-multimodal.json and cpu Embedded test cases as serving-tests-cpu-embed.json.  
 For details on how these optimized configurations are determined, see: [performance-benchmark-details](../../../.buildkite/performance-benchmarks/README.md#performance-benchmark-details).
 To benchmark the supported models using these optimized settings, follow the steps in [running vLLM Benchmark Suite manually](../../benchmarking/dashboard.md#manually-trigger-the-benchmark) and run the Benchmark Suite on a CPU environment.  
 
@@ -199,6 +199,28 @@ lscpu | grep "NUMA node(s):" | awk '{print $3}'
 For performance reference, users may also consult the [vLLM Performance Dashboard](https://hud.pytorch.org/benchmark/llms?repoName=vllm-project%2Fvllm&deviceName=cpu)
 , which publishes default-model CPU results produced using the same Benchmark Suite.
 
+#### Dry-Run
+
+For users only need to get the optimized runtime configurations without running benchmark, a Dry-Run mode is provided.
+By passing an environment variable DRY_RUN=1 with run-performance-benchmarks.sh,
+all commands will be generated under `./benchmark/results/`.
+
+```bash
+ON_CPU=1 DRY_RUN=1 bash .buildkite/performance-benchmarks/scripts/run-performance-benchmarks.sh
+```
+
+By providing different JSON file, users can get runtime configurations for different models such as Embedded Models.
+
+```bash
+ON_CPU=1 SERVING_JSON=serving-tests-cpu-embed.json DRY_RUN=1 bash .buildkite/performance-benchmarks/scripts/run-performance-benchmarks.sh
+```
+
+By providing MODEL_FILTER and DTYPE_FILTER, only commands for related model ID and Data Type will be generated.
+
+```bash
+ON_CPU=1 SERVING_JSON=serving-tests-cpu-text.json DRY_RUN=1 MODEL_FILTER=meta-llama/Llama-3.1-8B-Instruct DTYPE_FILTER=bfloat16  bash .buildkite/performance-benchmarks/scripts/run-performance-benchmarks.sh
+```
+
 ### How to decide `VLLM_CPU_OMP_THREADS_BIND`?
 
 - Default `auto` thread-binding is recommended for most cases. Ideally, each OpenMP thread will be bound to a dedicated physical core respectively, threads of each rank will be bound to the same NUMA node respectively, and 1 CPU per rank will be reserved for other vLLM components when `world_size > 1`. If you have any performance problems or unexpected binding behaviours, please try to bind threads as following.