上传文件至 /

Dockerfile

@@ -0,0 +1,21 @@
+FROM harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0.11.0
+
+# 修复1 transformers 库级 patch
+COPY patch.py /tmp/patch.py
+RUN python3 /tmp/patch.py
+
+# 修复2 Triton backend head_size 验证绕过
+COPY patch_triton.py /tmp/patch_triton.py
+RUN python3 /tmp/patch_triton.py
+
+# 修复3 运行时 tokenizer 配置修复脚本
+COPY detect_tokenizer.py /opt/detect_tokenizer.py
+COPY fix_tokenizer.py /opt/fix_tokenizer.py
+
+# 修复4 head_size 检测，自动切换 attention backend
+COPY detect_head_size.py /opt/detect_head_size.py
+
+COPY entrypoint.sh /opt/entrypoint.sh
+RUN chmod +x /opt/entrypoint.sh
+
+ENTRYPOINT ["/opt/entrypoint.sh"]

detect_head_size.py

@@ -0,0 +1,27 @@
+import json, os, sys
+
+MODEL_DIR = os.environ.get("MODEL_DIR", "/model")
+cfg_path = os.path.join(MODEL_DIR, "config.json")
+
+if not os.path.exists(cfg_path):
+    sys.exit(0)
+
+with open(cfg_path) as f:
+    cfg = json.load(f)
+
+head_size = cfg.get("head_dim")
+if head_size is None:
+    hs = cfg.get("hidden_size")
+    nh = cfg.get("num_attention_heads")
+    if hs and nh:
+        head_size = hs // nh
+
+if head_size is None:
+    sys.exit(0)
+
+SUPPORTED = {32, 64, 96, 128, 160, 192, 224, 256}
+if head_size not in SUPPORTED:
+    print(head_size)
+    sys.exit(2)
+
+sys.exit(0)

detect_tokenizer.py

@@ -0,0 +1,25 @@
+import os 
+import json
+
+def detect(model_dir):
+    cfg_path = os.path.join(model_dir, "tokenizer_config.json")
+
+    if os.path.exists(cfg_path):
+        with open(cfg_path) as f:
+            cfg = json.load(f)
+        cls = cfg.get("tokenizer_class", "")
+    else:
+        cls = ""
+
+    files = os.listdir(model_dir)
+
+    if "tokenizer.json" in files:
+        return "fast", cls
+
+    if "tokenizer.model" in files:
+        return "sentencepiece", cls
+
+    if "vocab.json" in files and "merges.txt" in files:
+        return "bpe", cls
+
+    return "unknown", cls

entrypoint.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+set -e
+
+MODEL_DIR=${MODEL_DIR:-/model}
+FIX_TOKENIZER_DIR=/tmp/fixed_tokenizer
+
+echo "[entrypoint] fixing tokenizer..."
+python3 /opt/fix_tokenizer.py
+
+echo "[entrypoint] checking head_size..."
+set +e
+HEAD_OUT=$(python3 /opt/detect_head_size.py)
+RC=$?
+set -e
+
+if [ "$RC" = "2" ] && [ -n "$HEAD_OUT" ]; then
+    export VLLM_USE_FLASH_ATTN_PA=0
+    echo "[entrypoint] head_size=$HEAD_OUT not in FlashAttention whitelist, switching to Triton backend (VLLM_USE_FLASH_ATTN_PA=0)"
+fi
+
+echo "[entrypoint] starting vllm..."
+exec vllm serve "$MODEL_DIR" \
+  --tokenizer "$FIX_TOKENIZER_DIR" \
+  "$@"

fix_tokenizer.py

@@ -0,0 +1,64 @@
+import os 
+import shutil
+import json
+import transformers
+import inspect
+from detect_tokenizer import detect
+
+MODEL_DIR = os.environ.get("MODEL_DIR", "/model")
+OUT_DIR = os.environ.get("FIX_TOKENIZER_DIR", "/tmp/fixed_tokenizer")
+
+os.makedirs(OUT_DIR, exist_ok=True)
+
+def copy_if_exists(name):
+    src = os.path.join(MODEL_DIR, name)
+    if os.path.exists(src):
+        shutil.copy(src, OUT_DIR)
+
+for f in [
+    "tokenizer.json",
+    "tokenizer_config.json",
+    "special_tokens_map.json",
+    "vocab.json",
+    "merges.txt",
+    "tokenizer.model",
+]:
+    copy_if_exists(f)
+
+typ, orig_cls = detect(MODEL_DIR)
+
+cfg_path = os.path.join(OUT_DIR, "tokenizer_config.json")
+
+if os.path.exists(cfg_path):
+    with open(cfg_path) as f:
+        cfg = json.load(f)
+else:
+    cfg = {}
+
+VALID_CLASSES = {
+    name for name, obj in inspect.getmembers(transformers)
+    if inspect.isclass(obj) and "Tokenizer" in name
+}
+
+BAD_CLASSES = {"TokenizersBackend", "TiktokenTokenizer"}
+
+FALLBACK = {
+    "fast": "PreTrainedTokenizerFast",
+    "sentencepiece": "LlamaTokenizer",
+    "bpe": "GPT2TokenizerFast",
+}
+
+if orig_cls and orig_cls in VALID_CLASSES and orig_cls not in BAD_CLASSES:
+    print(f"[fix_tokenizer] tokenizer_class '{orig_cls}' is valid, skip override")
+else:
+    fallback = FALLBACK.get(typ, "PreTrainedTokenizerFast")
+    if orig_cls:
+        print(f"[fix] override bad tokenizer_class: {orig_cls} → {fallback}")
+    else:
+        print(f"[fix] tokenizer_class missing, set to: {fallback}")
+    cfg["tokenizer_class"] = fallback
+
+with open(cfg_path, "w") as f:
+    json.dump(cfg, f)
+
+print(f"[fix_tokenizer] done → {OUT_DIR}")