Unified Llama 4 Scout/Maverick BI-V150 deployment bundle

2026-10-26 16:24:16 +08:00
commit caddeadd15
16 changed files with 1201 additions and 0 deletions
--- a/apply_patch.py
+++ b/apply_patch.py
@@ -0,0 +1,117 @@
+#!/usr/bin/env python3
+"""Apply the CoreX vLLM patches needed by Llama 4 Scout / Maverick.
+
+The patches in ./patches were generated against the exact base image
+`registry.iluvatar.com.cn:10443/customer/sz/vllm0.17.0-4.4.0-x86:v7`
+(vLLM 0.17.0+corex.20260612070235).
+
+Scout needs only the chunked-attention patch.  Maverick additionally needs the
+FP8 quantization / FP8 MoE patches and the `corexx_fp8_moe` fallback module.
+One image contains the union of both patch sets; the FP8 paths are only used
+when the mounted model is FP8.
+
+The script is idempotent: it checks for a marker in each target file and skips
+patches that are already applied.
+"""
+
+from __future__ import annotations
+
+import ast
+import os
+import pathlib
+import shutil
+import subprocess
+import sys
+
+PATCH_DIR = pathlib.Path(
+    os.environ.get("LLAMA4_PATCH_DIR", "/opt/llama4/patches")
+)
+
+# LLAMA4_PKG_ROOT is only needed for local testing against a copy of the
+# vendor files.  In the image it is derived from the installed vllm package.
+_pkg_env = os.environ.get("LLAMA4_PKG_ROOT")
+if _pkg_env:
+    PKG = pathlib.Path(_pkg_env)
+else:
+    import vllm
+
+    PKG = pathlib.Path(vllm.__file__).resolve().parent.parent
+
+# patch file -> (target file relative to the package root, applied marker)
+TARGETS = {
+    "01_fp8_quant.patch": (
+        "vllm/_custom_ops.py",
+        "COREXX_FP8_QUANT_PATCH",
+    ),
+    "02_fp8_moe.patch": (
+        "vllm/model_executor/layers/fused_moe/fused_moe.py",
+        "COREXX_FP8_MOE_PATCH",
+    ),
+    "03_llama4_chunked_attn.patch": (
+        "vllm/v1/attention/backends/utils.py",
+        "seq_lens_np=seqlens_k_local",
+    ),
+}
+
+
+def apply_one(patch_name: str, target_rel: str, marker: str) -> None:
+    target = PKG / target_rel
+    if not target.exists():
+        sys.exit(f"target file not found: {target}")
+    if marker in target.read_text():
+        print(f"[skip] {target} (already patched)")
+        return
+
+    patch_file = PATCH_DIR / patch_name
+    if not patch_file.exists():
+        sys.exit(f"missing patch file: {patch_file}")
+
+    result = subprocess.run(
+        [
+            "patch",
+            "-p1",
+            "--forward",
+            "--batch",
+            "--no-backup-if-mismatch",
+            "-d",
+            str(PKG),
+            "-i",
+            str(patch_file),
+        ],
+        text=True,
+        capture_output=True,
+    )
+    if result.stdout:
+        print(result.stdout, end="")
+    if result.returncode != 0:
+        if result.stderr:
+            print(result.stderr, file=sys.stderr)
+        sys.exit(f"patch failed: {patch_name}")
+    if marker not in target.read_text():
+        sys.exit(f"patch {patch_name} applied but marker not found in {target}")
+    print(f"[ok]   {target}")
+
+
+def main() -> None:
+    for patch_name, (target_rel, marker) in TARGETS.items():
+        apply_one(patch_name, target_rel, marker)
+
+    src = PATCH_DIR / "corexx_fp8_moe.py"
+    dst = PKG / "vllm/model_executor/layers/fused_moe/corexx_fp8_moe.py"
+    if not src.exists():
+        sys.exit(f"missing module: {src}")
+    shutil.copy2(src, dst)
+    print(f"[copy] {src} -> {dst}")
+
+    for path in (
+        PKG / "vllm/_custom_ops.py",
+        PKG / "vllm/model_executor/layers/fused_moe/fused_moe.py",
+        dst,
+        PKG / "vllm/v1/attention/backends/utils.py",
+    ):
+        ast.parse(path.read_text())
+    print("all patches applied; syntax OK")
+
+
+if __name__ == "__main__":
+    main()