基于曦望官方 vLLM 镜像构建模型服务镜像

Dockerfile

@@ -1,4 +1,25 @@
-FROM harbor.4pd.io/modelhubxc/enginex-sunrise/vllm-fix-tokenizer:v1.1.1
+FROM registry.maas.sunrise-ai.com/public/vllm:S2-v1.1.1
+ENV LD_LIBRARY_PATH=/usr/local/pccl/lib:\
+/usr/local/tangrt/targets/linux-x86_64/lib:\
+/usr/local/tangrt/targets/linux-x86_64/lib/stub:\
+/root/pt200/gcc-11.3.0/install/lib64:\
+/root:/root/gcc-11.5.0/lib64:\
+/usr/local/pccl/lib:\
+/usr/local/tangrt/targets/linux-x86_64/lib:\
+/usr/local/tangrt/targets/linux-x86_64/lib/stub:\
+/usr/local/tangrt/lib/linux-x86_64:\
+/root/pt200/gcc-11.3.0/install/lib64:\
+/root:\
+/usr/lib64:\
+/usr/local/lib/python3.10/site-packages/torch/lib
+ENV TORCH_DEVICE_BACKEND_AUTOLOAD=0
+ENV PATH=/root/gcc-11.5.0/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
+ENV PYTHONPATH=/sunrise_code/vllm:/sunrise_code/sunrise_vllm:/usr/local/lib/python3.10/site-packages:
+RUN ln -sf /usr/local/bin/python3.10 /usr/bin/python3
+COPY fix_tokenizer.py /opt/
+COPY detect_tokenizer.py /opt/
+COPY entrypoint.sh /opt/
 WORKDIR /model
 COPY . /model
+RUN chmod +x /opt/entrypoint.sh
 ENTRYPOINT ["/opt/entrypoint.sh"]

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 +1,39 @@
+#!/bin/bash
+set -e
+
+MODEL_DIR=${1:-/model}
+shift || true
+
+FIX_TOKENIZER_DIR=/tmp/fixed_tokenizer
+AUTO_FIX=${AUTO_FIX_TOKENIZER:-auto}
+
+echo "[entrypoint] model dir: $MODEL_DIR"
+
+NEED_FIX=0
+
+if [ "$AUTO_FIX" = "1" ] || [ "$AUTO_FIX" = "true" ]; then
+    NEED_FIX=1
+elif [ "$AUTO_FIX" = "auto" ]; then
+    if [ -f "$MODEL_DIR/tokenizer_config.json" ]; then
+        if grep -q "TokenizersBackend\|TiktokenTokenizer" "$MODEL_DIR/tokenizer_config.json"; then
+            NEED_FIX=1
+        fi
+        # 检测 extra_special_tokens 是否为 list 格式
+        if grep -q '"extra_special_tokens":\s*\[' "$MODEL_DIR/tokenizer_config.json"; then
+            NEED_FIX=1
+        fi
+    fi
+fi
+
+if [ $NEED_FIX -eq 1 ]; then
+    echo "[entrypoint] fixing tokenizer..."
+    python3 /opt/fix_tokenizer.py
+    TOKENIZER_ARG="--tokenizer $FIX_TOKENIZER_DIR"
+else
+    echo "[entrypoint] tokenizer OK, skip fix"
+    TOKENIZER_ARG=""
+fi
+
+echo "[entrypoint] starting vllm..."
+
+exec vllm serve "$MODEL_DIR" $TOKENIZER_ARG "$@"

fix_tokenizer.py

@@ -0,0 +1,72 @@
+import os
+import shutil
+import json
+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)
+
+# 复制所有可能相关文件
+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 = {}
+
+# ===== 自动修复策略 =====
+if typ == "fast":
+    cfg["tokenizer_class"] = "PreTrainedTokenizerFast"
+    cfg["from_slow"] = False
+    cfg.pop("backend", None)
+
+elif typ == "sentencepiece":
+    cfg["tokenizer_class"] = "LlamaTokenizer"
+
+elif typ == "bpe":
+    cfg["tokenizer_class"] = "GPT2TokenizerFast"
+
+else:
+    cfg["tokenizer_class"] = "PreTrainedTokenizerFast"
+
+# 特殊 case 修复
+bad_classes = [
+    "TokenizersBackend",
+    "TiktokenTokenizer",
+]
+
+if orig_cls in bad_classes:
+    print(f"[fix] override bad tokenizer_class: {orig_cls} → {cfg['tokenizer_class']}")
+    print(f"[fix] override from_slow: {cfg['from_slow']}")
+
+# 修复 extra_special_tokens: list → dict 格式
+if "extra_special_tokens" in cfg and isinstance(cfg["extra_special_tokens"], list):
+    orig_list = cfg["extra_special_tokens"]
+    cfg["extra_special_tokens"] = {token: token for token in orig_list}
+    print(f"[fix] converted extra_special_tokens from list ({len(orig_list)} items) to dict format")
+
+# 写回
+with open(cfg_path, "w") as f:
+    json.dump(cfg, f)
+
+print(f"[fix_tokenizer] done → {OUT_DIR}")