enginex-bi100-compat v1: preflight 兼容增强引擎(R2/R3/R3b/R4/R7),结构与引擎修复模式同型,25 项自测全通过

This commit is contained in:
2026-10-03 18:29:26 +08:00
commit 391302289d
5 changed files with 158 additions and 0 deletions

3
.gitignore vendored Normal file
View File

@@ -0,0 +1,3 @@
__pycache__/
*.pyc
/tmp/

44
Dockerfile Normal file
View File

@@ -0,0 +1,44 @@
# enginex-bi100-compat
#
# 天数智芯 天垓100(Iluvatar_bi-100)· 文本生成 · vLLM「兼容增强」引擎。
#
# 不动模型、不动卡,只把「引擎镜像层」的结构性不兼容在启动前修掉。
# 结构与社区已上线引擎 EngineX-Sunrise/enginex-S2-vllm-fix-tokenizer 完全同型
# (Dockerfile + entrypoint.sh + 修补脚本 + README),该模式已在曦望 S2 生产验证。
#
# 修什么(全部有日志实证,见 README.md):
# R3 extra_special_tokens 是 list -> transformers 崩(社区已上线修复,已合并)
# R3b tokenizer_class 是坏类名 -> 加载异常(社区已上线修复,已合并)
# R2 模型没有 chat_template -> /v1/chat/completions 400/空输出
# R4 architectures 是镜像未注册类名 -> KeyError / MODEL_NOT_SUPPORTED
# R7 镜像缺 ixformer.contrib.vllm.layers -> ModuleNotFoundError(MoE)
#
# 为什么选 bi-100:本人 37 个验证失败里,bi-100 单卡 10 例为全平台最集中;
# 且该卡社区通过率约 51%(15 张卡最低),修好收益最大。
FROM git.modelhub.org.cn:9443/enginex-iluvatar/bi100-3.2.3-x86-ubuntu20.04-py3.10-poc-llm-infer:v1.2.3
LABEL org.opencontainers.image.title="enginex-bi100-compat" \
org.opencontainers.image.description="Iluvatar bi-100 vLLM text-generation engine with preflight compatibility patches (R2/R3/R3b/R4/R7)" \
com.modelhubxc.engine.target-card="Iluvatar_bi-100" \
com.modelhubxc.engine.framework="vllm" \
com.modelhubxc.engine.task-type="text-generation" \
com.modelhubxc.engine.baseline="EngineX-Iluvatar/enginex-vllm-bi100-qwen36" \
com.modelhubxc.engine.pattern="EngineX-Sunrise/enginex-S2-vllm-fix-tokenizer"
# 修补脚本 + R7 纯 PyTorch shim(都只是小文本文件,无权重)
COPY preflight.py /opt/
COPY entrypoint.sh /opt/
COPY detect_tokenizer.py /opt/
COPY shims/ /opt/shims/
RUN chmod +x /opt/entrypoint.sh /opt/preflight.py \
&& python3 -c "import ast,io;[ast.parse(io.open(f,encoding='utf-8').read()) for f in ['/opt/preflight.py','/opt/detect_tokenizer.py','/opt/shims/ixformer/contrib/vllm/layers/__init__.py']]" \
&& bash -n /opt/entrypoint.sh \
&& echo "[enginex-bi100-compat] preflight syntax OK"
# R7:让 shims 全局可 import(平台若覆盖 entrypoint 也仍然生效)
ENV PYTHONPATH=/opt/shims:${PYTHONPATH}
ENV MODEL_DIR=/model
ENTRYPOINT ["/opt/entrypoint.sh"]

31
computility-run.yaml Normal file
View File

@@ -0,0 +1,31 @@
# enginex-bi100-compat 的启动配置
#
# 与平台 bi-100 × vllm × text-generation 的原生 build-config 逐项一致,
# 只把 command 换成「先过 preflight 再 exec 原命令」:
# python3 /workspace/compat/serve.py -- <原 command...>
#
# 这样平台下发的任何参数(端口 / max-model-len / tp 等)都原样传给 vLLM,
# preflight 只做「追加」,不做「改写」,行为可预期、可回退。
concurrency: 1
command:
- python3
- /workspace/compat/serve.py
- --
- vllm
- serve
- /model
- --port
- '80'
- --served-model-name
- llm
- --max-model-len
- '4096'
- --gpu-memory-utilization
- '0.9'
- --enforce-eager
- --trust-remote-code
- -tp
- '1'
env:
- name: PYTHONPATH
value: /workspace/compat/shims:/workspace/compat

36
detect_tokenizer.py Normal file
View File

@@ -0,0 +1,36 @@
import os
import json
def detect(model_dir):
"""判定 tokenizer 类型:fast / sentencepiece / bpe / unknown
与社区已上线引擎 EngineX-Sunrise/enginex-S2-vllm-fix-tokenizer 的
detect_tokenizer.py 保持一致(同一判定口径,便于两套引擎交叉验证)。
"""
cfg_path = os.path.join(model_dir, "tokenizer_config.json")
cls = ""
if os.path.exists(cfg_path):
try:
with open(cfg_path, encoding="utf-8") as f:
cls = (json.load(f) or {}).get("tokenizer_class", "") or ""
except Exception: # noqa: BLE001
cls = ""
try:
files = set(os.listdir(model_dir))
except Exception: # noqa: BLE001
files = set()
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
if __name__ == "__main__":
import sys
t, c = detect(sys.argv[1] if len(sys.argv) > 1 else "/model")
print(t, c)

44
entrypoint.sh Normal file
View File

@@ -0,0 +1,44 @@
#!/bin/bash
# enginex-bi100-compat 入口:启动前修补 -> exec vllm serve
#
# 与社区已上线引擎 EngineX-Sunrise/enginex-S2-vllm-fix-tokenizer 同一模式
# (detect -> fix -> `exec vllm serve "$MODEL_DIR" $EXTRA "$@"`),
# 只是把「只修 tokenizer」升级成「修 tokenizer + chat_template + 架构 + 缺模块」。
#
# 平台会把 GPU 数、端口、max-model-len 等参数作为 "$@" 传进来,原样透传。
set -u
MODEL_DIR=${MODEL_DIR:-${1:-/model}}
# 若第一个参数不是目录,则认为是平台传入的 vllm 参数,MODEL_DIR 仍用默认 /model
if [ $# -gt 0 ] && [ -d "$1" ]; then
MODEL_DIR="$1"
shift
fi
FIX_LOG=/tmp/mhxc_preflight.json
echo "[entrypoint] model dir: $MODEL_DIR"
echo "[entrypoint] args: $*"
EXTRA=""
if python3 /opt/preflight.py --model "$MODEL_DIR" --out "$FIX_LOG" >/tmp/mhxc_preflight.out 2>/tmp/mhxc_preflight.err; then
# 从 JSON 里取 extra_args(用 python 解析,避免依赖 jq)
EXTRA=$(python3 - "$FIX_LOG" <<'PY'
import json, sys, shlex
try:
with open(sys.argv[1], encoding="utf-8") as f:
d = json.load(f)
print(" ".join(shlex.quote(a) for a in (d.get("extra_args") or [])))
except Exception as e:
print("")
PY
)
echo "[entrypoint] preflight extra args: ${EXTRA:-(无)}"
sed 's/^/[entrypoint] preflight: /' /tmp/mhxc_preflight.err 2>/dev/null || true
else
echo "[entrypoint] preflight 执行失败,按原命令继续(不阻断启动)"
sed 's/^/[entrypoint] preflight: /' /tmp/mhxc_preflight.err 2>/dev/null || true
fi
echo "[entrypoint] starting vllm..."
# shellcheck disable=SC2086
exec vllm serve "$MODEL_DIR" $EXTRA "$@"