5000t/s 本地部署DeepSeek-V4-Flash-0731*4卡 CMP-170HX 破解卡 总显存256GB pp方式运行
4卡支持10M上下文Cache prefil 5000 t/s /71token
https://github.com/344303947/deepseek-v4-cmp170hx
3卡&4K 支持1M上下文 3M KV Cache 5000 t/s 50 token/s
https://github.com/344303947/vllm/tree/dsv4fpp3
需要加170HX 群的 加w_x:su009486 备注加群
# 3. 环境变量
export CUDA_HOME=/usr/local/cuda
export PATH=/usr/local/cuda/bin:$PATH
export CUDA_VISIBLE_DEVICES=0,1,2,3
export HF_HUB_OFFLINE=1
export VLLM_WORKER_MULTIPROC_METHOD=spawn
export VLLM_PP_LAYER_PARTITION=12,12,12,7
export DSV4_LOGITS_ROW_CHUNK=64
# 4. 前台启动（Ctrl+C 停止）
python -m vllm.entrypoints.openai.api_server \
--model /models/models/DeepSeek-V4-Flash-0731 \
--served-model-name /DeepSeek-V4-Flash-0731 \
--api-key sk_344303 \
--port 9070 \
--tensor-parallel-size 1 \
--pipeline-parallel-size 4 \
--kv-cache-dtype fp8 \
--block-size 256 \
--max-model-len 1048576 \
--max-num-batched-tokens 2048 \
--max-num-seqs 8 \
--gpu-memory-utilization 0.85 \
--enable-prefix-caching \
--no-enable-flashinfer-autotune \
--tokenizer-mode deepseek_v4 \
--tool-call-parser deepseek_v4 \
--enable-auto-tool-choice \
--reasoning-parser deepseek_v4 \
--speculative-config '{"method":"dspark","num_speculative_tokens":5}'
