Initial commit: Llama 4 Scout BI-V150 deployment bundle

2026-10-23 13:41:41 +08:00
commit 7ce7159d36
6 changed files with 296 additions and 0 deletions
--- a/apply_patch.py
+++ b/apply_patch.py
@@ -0,0 +1,77 @@
+#!/usr/bin/env python3
+"""
+修复 CoreX 版 vLLM 0.17.0 的 Llama 4 chunked local attention bug。
+
+症状（第一个真实请求时）：
+  TypeError: CommonAttentionMetadata.__init__() missing 2 required positional
+  arguments: 'key_start_loc' and 'seq_lens_np'
+
+原因：
+  vllm/v1/attention/backends/utils.py 的
+  make_local_attention_virtual_batches() 在构造 CommonAttentionMetadata 时
+  没有传新版本 dataclass 要求的 key_start_loc / seq_lens_np。
+
+本脚本幂等：已经打过补丁会直接返回 0。
+"""
+
+import os
+import sys
+
+try:
+    import vllm
+except ImportError as e:
+    print(f"ERROR: 无法 import vllm: {e}", file=sys.stderr)
+    sys.exit(1)
+
+TARGET = os.path.join(
+    os.path.dirname(vllm.__file__),
+    "v1",
+    "attention",
+    "backends",
+    "utils.py",
+)
+
+MARKER = "seq_lens_np=seqlens_k_local"
+
+OLD = """        query_start_loc=query_start_loc_cpu.to(device=device, non_blocking=True),
+        seq_lens=seq_lens_cpu.to(device=device, non_blocking=True),"""
+
+NEW = """        query_start_loc=query_start_loc_cpu.to(device=device, non_blocking=True),
+        key_start_loc=torch.from_numpy(
+            np.concatenate(
+                (np.zeros(1, dtype=np.int32), np.cumsum(seqlens_k_local, dtype=np.int32))
+            )
+        ).to(device=device, non_blocking=True),
+        seq_lens=seq_lens_cpu.to(device=device, non_blocking=True),
+        seq_lens_np=seqlens_k_local,"""
+
+
+def main() -> int:
+    if not os.path.exists(TARGET):
+        print(f"ERROR: 找不到目标文件: {TARGET}", file=sys.stderr)
+        return 1
+
+    with open(TARGET, "r", encoding="utf-8") as f:
+        src = f.read()
+
+    if MARKER in src:
+        print(f"patch already applied: {TARGET}")
+        return 0
+
+    if OLD not in src:
+        print(
+            "ERROR: 目标代码块未找到，vLLM 版本可能不同。\n"
+            f"请手动检查: {TARGET}",
+            file=sys.stderr,
+        )
+        return 1
+
+    with open(TARGET, "w", encoding="utf-8") as f:
+        f.write(src.replace(OLD, NEW, 1))
+
+    print(f"patched: {TARGET}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())