commit 8022cd4b2e486fec564446d1e260f38f9fcd20ae
Author: 4paradigm <4paradigm@4paradigmdeMacBook-Pro.local>
Date:   Tue Jul 21 10:18:53 2026 +0800

    fix tokenizer

diff --git a/.gitea/workflows/docker-build-push.yml b/.gitea/workflows/docker-build-push.yml
new file mode 100644
index 0000000..9c5c285
--- /dev/null
+++ b/.gitea/workflows/docker-build-push.yml
@@ -0,0 +1,132 @@
+name: Docker Build and Push
+
+on:
+  push:
+    tags:
+      - "v*"
+
+jobs:
+  docker:
+    runs-on: amd64-ubuntu-24.04
+
+    steps:
+      - name: Clone repository
+        run: |
+          git clone "${{ gitea.server_url }}/${{ gitea.repository }}.git" .
+          git checkout "${{ gitea.ref_name }}"
+
+      - name: Set image metadata
+        run: |
+          IMAGE_NAME="$(echo "${{ gitea.repository }}" | tr '[:upper:]' '[:lower:]' | tr '_' '-')"
+          IMAGE="${DOCKER_REGISTRY}/${DOCKER_USERNAME}/${IMAGE_NAME}:${{ gitea.ref_name }}"
+
+          echo "IMAGE_NAME=${IMAGE_NAME}" >> "$GITEA_ENV"
+          echo "IMAGE=${IMAGE}" >> "$GITEA_ENV"
+
+      - name: Load and Validate Task Info
+        run: |
+          set -a
+          . .gitea/workflows/task_info.env
+          set +a
+
+          for name in FRAMEWORK GPU_TYPE TASK_TYPE; do
+            eval "value=\${${name}:-}"
+            if [ "$name" = "FRAMEWORK" ] && [ -z "$value" ]; then
+              echo "${name} is empty in .gitea/workflows/task_info.env"
+              exit 1
+            fi
+
+            echo "${name}=${value}" >> "$GITEA_ENV"
+          done
+
+      - name: Validate Image Verify Metadata
+        run: |
+          if [ -z "${FIXED_TOKEN:-}" ]; then
+            echo "FIXED_TOKEN is not configured on runner"
+            exit 1
+          fi
+
+          if ! response="$(curl --silent --show-error --location --get 'https://modelhub.org.cn/adminApi/image-verify/validate' \
+            --header "Xc-Token: ${FIXED_TOKEN}" \
+            --data-urlencode "gpuType=${GPU_TYPE:-}" \
+            --data-urlencode "taskType=${TASK_TYPE:-}")"; then
+            echo "failed to call image verify validate API"
+            exit 1
+          fi
+
+          VALIDATE_RESPONSE="$response" python3 - <<'PY'
+          import json
+          import os
+          import sys
+
+          raw = os.environ.get("VALIDATE_RESPONSE", "")
+          try:
+              body = json.loads(raw)
+          except json.JSONDecodeError:
+              print("image verify validate API returned invalid JSON")
+              print(raw)
+              sys.exit(1)
+
+          if body.get("code") == 0 and body.get("data") is True:
+              print("image verify metadata validation passed")
+              sys.exit(0)
+
+          message = body.get("message") or "unknown error"
+          print(f"image verify metadata validation failed: {message}")
+          print(raw)
+          sys.exit(1)
+          PY
+
+      - name: Login to Docker Registry
+        run: |
+          echo "$DOCKER_PASSWORD" | docker login "$DOCKER_REGISTRY" \
+            -u "$DOCKER_USERNAME" \
+            --password-stdin
+
+      - name: Build Docker Image
+        run: |
+          docker build -t "$IMAGE" .
+
+      - name: Push Docker Image
+        run: |
+          for attempt in 1 2 3; do
+            echo "Starting docker push attempt ${attempt}/3 for ${IMAGE}"
+            docker push "$IMAGE" &
+            PUSH_PID=$!
+
+            while kill -0 "$PUSH_PID" 2>/dev/null; do
+              echo "docker push is still running at $(date -u '+%Y-%m-%dT%H:%M:%SZ')"
+              sleep 60
+            done
+
+            if wait "$PUSH_PID"; then
+              echo "docker push completed successfully"
+              exit 0
+            fi
+
+            echo "docker push failed on attempt ${attempt}/3"
+            sleep 30
+          done
+
+          echo "docker push failed after 3 attempts"
+          exit 1
+
+      - name: Notify Image Verify
+        run: |
+          if [ -z "${FIXED_TOKEN:-}" ]; then
+            echo "FIXED_TOKEN is not configured on runner"
+            exit 1
+          fi
+
+          curl --silent --show-error --fail-with-body --location --request POST 'https://modelhub.org.cn//adminApi/image-verify' \
+            --header "Xc-Token: ${FIXED_TOKEN}" \
+            --header 'Content-Type: application/json' \
+            --data-raw "{
+              \"framework\": \"${FRAMEWORK}\",
+              \"gpuType\": \"${GPU_TYPE}\",
+              \"imageUrl\": \"${IMAGE}\",
+              \"taskType\": \"${TASK_TYPE}\",
+              \"createBy\": \"${{ gitea.actor }}\",
+              \"repoUrl\": \"${{ gitea.server_url }}/${{ gitea.repository }}\",
+              \"tag\": \"${{ github.ref_name }}\"
+            }"
diff --git a/.gitea/workflows/task_info.env b/.gitea/workflows/task_info.env
new file mode 100644
index 0000000..aeee76c
--- /dev/null
+++ b/.gitea/workflows/task_info.env
@@ -0,0 +1,3 @@
+FRAMEWORK=vllm_tokenizer_patch
+GPU_TYPE=Iluvatar_bi-150
+TASK_TYPE=text-generation
diff --git a/Dockerfile b/Dockerfile
new file mode 100644
index 0000000..151df5f
--- /dev/null
+++ b/Dockerfile
@@ -0,0 +1,10 @@
+FROM harbor-contest.4pd.io/luopingyi/enginex-iluvatar-bi150/vllm:0.8.3
+
+COPY fix_tokenizer.py /opt/fix_tokenizer.py
+COPY vllm_wrapper.sh /opt/vllm_wrapper.sh
+
+RUN set -e; \
+    vllm_bin="$(command -v vllm)"; \
+    mv "$vllm_bin" "$(dirname "$vllm_bin")/vllm_real"; \
+    cp /opt/vllm_wrapper.sh "$vllm_bin"; \
+    chmod +x "$vllm_bin" /opt/fix_tokenizer.py
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..8054a03
--- /dev/null
+++ b/README.md
@@ -0,0 +1,26 @@
+# BI-150 TokenizersBackend 修复补丁
+
+基于 BI-150 vLLM 0.8.3 镜像，仅修复以下错误：
+
+```text
+ValueError: Tokenizer class TokenizersBackend does not exist or is not currently imported.
+```
+
+## 工作方式
+
+镜像构建时用 wrapper 接管 `vllm` 命令。执行 `vllm serve <model_dir>` 时：
+
+1. 读取模型的 `tokenizer_config.json`。
+2. 仅当 `tokenizer_class` 严格等于 `TokenizersBackend` 时，将 tokenizer 文件复制到 `/tmp/fixed_tokenizer`。
+3. 根据已有文件选择替代类：`tokenizer.json` 使用 `PreTrainedTokenizerFast`，`tokenizer.model` 使用 `LlamaTokenizer`，`vocab.json` 与 `merges.txt` 使用 `GPT2TokenizerFast`；无法判断时回退到 `PreTrainedTokenizerFast`。
+4. 向原始 vLLM 命令追加 `--tokenizer /tmp/fixed_tokenizer`。
+
+原模型目录不会被修改。其他 tokenizer 错误及其他 `tokenizer_class` 均不处理。
+
+## 构建
+
+```bash
+docker build -t iluvatar-bi150-vllm-tokenizer-fix:0.8.3 .
+```
+
+使用时仅替换原 BI-150 镜像名，`vllm serve` 参数保持不变。
diff --git a/fix_tokenizer.py b/fix_tokenizer.py
new file mode 100644
index 0000000..42538f5
--- /dev/null
+++ b/fix_tokenizer.py
@@ -0,0 +1,95 @@
+#!/usr/bin/env python3
+"""
+检测并修复 tokenizer_config.json 中的两类问题：
+1. tokenizer_class 在 transformers 中不存在（如 TokenizersBackend）
+2. extra_special_tokens 为 list 格式（transformers 要求 dict）
+
+若存在问题，将 tokenizer 文件复制到 /tmp/fixed_tokenizer/ 并修复，
+最后将修复目录路径输出到 stdout。若无需修复，输出为空。
+"""
+import os
+import sys
+import json
+import shutil
+
+MODEL_DIR = sys.argv[1] if len(sys.argv) > 1 else os.environ.get("MODEL_DIR", "/model")
+OUT_DIR = "/tmp/fixed_tokenizer"
+
+
+def main():
+    cfg_path = os.path.join(MODEL_DIR, "tokenizer_config.json")
+    if not os.path.exists(cfg_path):
+        return
+
+    with open(cfg_path) as f:
+        cfg = json.load(f)
+
+    fixes = []
+
+    # --- 检测 1：tokenizer_class 是否在 transformers 中存在 ---
+    tokenizer_class = cfg.get("tokenizer_class", "")
+    bad_tokenizer_class = False
+    if tokenizer_class:
+        import transformers
+        if getattr(transformers, tokenizer_class, None) is None:
+            bad_tokenizer_class = True
+            fixes.append(f"tokenizer_class '{tokenizer_class}' not found in transformers")
+
+    # --- 检测 2：extra_special_tokens 是否为 list 格式 ---
+    bad_extra_special_tokens = (
+        "extra_special_tokens" in cfg
+        and isinstance(cfg["extra_special_tokens"], list)
+    )
+    if bad_extra_special_tokens:
+        fixes.append("extra_special_tokens is a list, expected dict")
+
+    if not fixes:
+        return  # 无需修复
+
+    # 复制 tokenizer 文件到临时目录
+    os.makedirs(OUT_DIR, exist_ok=True)
+    for fname in [
+        "tokenizer.json",
+        "tokenizer_config.json",
+        "special_tokens_map.json",
+        "vocab.json",
+        "merges.txt",
+        "tokenizer.model",
+    ]:
+        src = os.path.join(MODEL_DIR, fname)
+        if os.path.exists(src):
+            shutil.copy(src, OUT_DIR)
+
+    # --- 修复 1：替换 tokenizer_class ---
+    if bad_tokenizer_class:
+        files = os.listdir(MODEL_DIR)
+        if "tokenizer.json" in files:
+            fixed_class = "PreTrainedTokenizerFast"
+        elif "tokenizer.model" in files:
+            fixed_class = "LlamaTokenizer"
+        elif "vocab.json" in files and "merges.txt" in files:
+            fixed_class = "GPT2TokenizerFast"
+        else:
+            fixed_class = "PreTrainedTokenizerFast"
+        cfg["tokenizer_class"] = fixed_class
+        print(
+            f"[fix_tokenizer] tokenizer_class: '{tokenizer_class}' → '{fixed_class}'",
+            file=sys.stderr,
+        )
+
+    # --- 修复 2：extra_special_tokens list → dict ---
+    if bad_extra_special_tokens:
+        orig_list = cfg["extra_special_tokens"]
+        cfg["extra_special_tokens"] = {token: token for token in orig_list}
+        print(
+            f"[fix_tokenizer] extra_special_tokens: list({len(orig_list)}) → dict",
+            file=sys.stderr,
+        )
+
+    with open(os.path.join(OUT_DIR, "tokenizer_config.json"), "w") as f:
+        json.dump(cfg, f, indent=2)
+
+    print(OUT_DIR)  # 输出修复目录，供 entrypoint.sh 捕获
+
+
+main()
diff --git a/vllm_wrapper.sh b/vllm_wrapper.sh
new file mode 100644
index 0000000..1f7c884
--- /dev/null
+++ b/vllm_wrapper.sh
@@ -0,0 +1,19 @@
+#!/bin/bash
+set -e
+
+VLLM_REAL="$(dirname "$(readlink -f "$0")")/vllm_real"
+
+# 仅拦截 vllm serve <model_dir>，其他命令原样透传。
+if [ "$1" = "serve" ] && [ -n "$2" ]; then
+    MODEL_DIR="$2"
+    shift 2
+
+    FIXED_DIR=$(python3 /opt/fix_tokenizer.py "$MODEL_DIR")
+    if [ -n "$FIXED_DIR" ]; then
+        exec "$VLLM_REAL" serve "$MODEL_DIR" --tokenizer "$FIXED_DIR" "$@"
+    fi
+
+    exec "$VLLM_REAL" serve "$MODEL_DIR" "$@"
+fi
+
+exec "$VLLM_REAL" "$@"