cjw/vllm-r9700-container
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

x

Browse Source
This commit is contained in:
SZCJW
2026-03-10 23:57:37 +08:00
parent a48423d817
commit 65ccd278c9
4 changed files with 63 additions and 312 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Dockerfile
+6 -18
View File
@@ -15,7 +15,7 @@ ARG ROCM_MAJOR_VER=7
ARG GFX=gfx120X-all
RUN set -euo pipefail; \
BASE="https://therock-nightly-tarball.s3.amazonaws.com"; \
PREFIX="therock-dist-linux-${GFX}-${ROCM_MAJOR_VER}"; \
PREFIX="therock-dist-linux-${GFX}-${ROCM_MAJOR_VER}-${ROCM_MAJOR_VER}"; \
KEY="$(curl -s "${BASE}?list-type=2&prefix=${PREFIX}" \
| tr '<' '\n' \
| grep -o "therock-dist-linux-${GFX}-${ROCM_MAJOR_VER}\..*\.tar\.gz" \
@@ -80,7 +80,7 @@ RUN echo "import sys, re" > patch_vllm.py && \
echo "p = Path('vllm/platforms/__init__.py')" >> patch_vllm.py && \
echo "txt = p.read_text()" >> patch_vllm.py && \
echo "txt = txt.replace('import amdsmi', '# import amdsmi')" >> patch_vllm.py && \
echo "txt = re.sub(r'is_rocm = .*', 'is_rocm = True', txt)" >> patch_vllm.py && \
echo "txt = re.sub(r'import is_rocm = .*', 'is_rocm = True', txt)" >> patch_vllm.py && \
echo "txt = re.sub(r'if len\(amdsmi\.amdsmi_get_processor_handles\(\)\) > 0:', 'if True:', txt)" >> patch_vllm.py && \
echo "txt = txt.replace('amdsmi.amdsmi_init()', 'pass')" >> patch_vllm.py && \
echo "txt = txt.replace('amdsmi.amdsmi_shut_down()', 'pass')" >> patch_vllm.py && \
@@ -145,21 +145,9 @@ RUN chmod -R a+rwX /opt && \
rm -rf /root/.cache/pip || true && \
dnf clean all && rm -rf /var/cache/dnf/*
# Create vLLM configuration directory
RUN mkdir -p /etc/vllm
# Create default YAML configuration file
RUN printf '# model_config.yaml - 多模型配置\n# 默认启动的模型\ndefault: "deepseek_r1_distill_qwen_32b_awq"\n\n# 模型配置\nmodels:\n deepseek_r1_distill_qwen_14b:\n path: "/app/vllm/models/DeepSeek-R1-Distill-Qwen-14B"\n name: "DeepSeek-R1-Distill-Qwen-14B"\n max_model_len: 8192\n gpu_memory_utilization: 0.9\n port: 2001\n dtype: "float16"\n quantization: "awq"\n tensor_parallel_size: 1\n enforce_eager: true\n api_key: "sk-14b-20240101-abcdef123456"\n \n deepseek_r1_distill_qwen_32b_awq:\n path: "/app/vllm/models/DeepSeek-R1-Distill-Qwen-32B-AWQ"\n name: "DeepSeek-R1-Distill-Qwen-32B-AWQ"\n \n # 模型性能参数\n max_model_len: 32768\n gpu_memory_utilization: 0.95\n enforce_eager: true\n max_num_seqs: 2\n max_num_batched_tokens: 1024\n block_size: 16\n tensor_parallel_size: 1\n swap_space: 0\n \n # 新增：采样参数默认值\n sampling_defaults:\n temperature: 0.6\n max_tokens: 4096\n top_p: 0.9\n frequency_penalty: 0.0\n presence_penalty: 0.0\n stop:\n - "用户:"
- "助手:"
- "###"
- "问题:"
- "回答:"
\n # 其他配置\n dtype: "auto"\n quantization: "awq"\n port: 2001\n api_key: "sk-32b-20240101-ghijk789012"\n \n glm_4_7_flash_awq:\n path: "/app/vllm/models/GLM-4.7-Flash-AWQ"\n name: "GLM-4.7-Flash-AWQ"\n \n # 模型性能参数\n max_model_len: 32768\n gpu_memory_utilization: 0.9\n enforce_eager: false\n max_num_seqs: 2\n max_num_batched_tokens: 1024\n block_size: 16\n tensor_parallel_size: 1\n swap_space: 0\n \n # 新增：采样参数默认值\n sampling_defaults:\n temperature: 0.6\n max_tokens: 4096\n top_p: 0.9\n frequency_penalty: 0.0\n presence_penalty: 0.0\n stop:\n - "用户:"
- "助手:"
- "###"
- "问题:"
- "回答:"
\n # 其他配置\n dtype: "auto"\n quantization: "awq"\n port: 2001\n api_key: "sk-32b-20240101-ghijk789012"\n \n qwen3_vl_32b_instruct_awq:\n path: "/app/vllm/models/Qwen3-VL-32B-Instruct-AWQ"\n name: "Qwen3-VL-32B-Instruct-AWQ"\n max_model_len: 32768\n gpu_memory_utilization: 0.7\n port: 2001\n dtype: "auto"\n quantization: "awq"\n tensor_parallel_size: 1\n enforce_eager: true\n api_key: "sk-glm-20240101-lmnop345678"\n\n# 服务器通用设置\nserver:\n host: "0.0.0.0"\n log_level: "info"\n # 全局管理员密钥（拥有所有模型的访问权限）\n admin_key: "sk-admin-20240101-xyz789"\n # 允许的请求头名称（支持多个，按顺序检查）\n api_key_headers: ["Authorization", "X-API-Key", "api-key"]\n # 是否允许通过查询参数传递密钥\n allow_query_param: true\n # 查询参数名称\n api_key_param: "api_key"' > /etc/vllm/model_config.yaml
# Create vLLM configuration directory and model directory
RUN mkdir -p /etc/vllm && \
mkdir -p /model
# Copy necessary scripts
COPY scripts/01-rocm-envs.sh /etc/profile.d/01-rocm-envs.sh
@@ -169,4 +157,4 @@ COPY scripts/start_vllm.py /usr/local/bin/start-vllm
RUN chmod 0644 /etc/profile.d/*.sh && chmod +x /usr/local/bin/start-vllm
RUN printf 'ulimit -S -c 0\n' > /etc/profile.d/90-nocoredump.sh && chmod 0644 /etc/profile.d/90-nocoredump.sh
CMD ["/bin/bash"]
CMD ["/bin/bash"]