DeepSeek-V3.1-Base 昇腾 NPU 适配部署文档

source /usr/local/Ascend/ascend-toolkit/set_env.sh
export HCCL_OP_EXPANSION_MODE=AIV
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

# BF16 全精度，TP=8
vllm serve /models/DeepSeek-V3.1-Base \
  --dtype bfloat16 \
  --tensor-parallel-size 8 \
  --max-model-len 65536 \
  --max-num-seqs 16 \
  --port 8000 \
  --trust-remote-code

# W8A8 量化，更省显存
vllm serve /models/DeepSeek-V3.1-Base \
  --quantization w8a8 \
  --tensor-parallel-size 8 \
  --max-model-len 65536 \
  --max-num-seqs 16 \
  --port 8000 \
  --trust-remote-code

# 单条推理
python inference.py \
  --model /models/DeepSeek-V3.1-Base \
  --tp 8 \
  --prompt "请解释 MLA 注意力机制的原理。"

# 批量推理
python inference.py \
  --model /models/DeepSeek-V3.1-Base \
  --tp 8 \
  --prompt-file ./prompts.jsonl \
  --output ./result.jsonl

# 查看模型状态
curl http://localhost:8000/v1/models

# Chat  completions
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "/models/DeepSeek-V3.1-Base",
    "messages": [{"role": "user", "content": "解释 MoE 架构的优势。"}],
    "max_tokens": 512,
    "temperature": 0.7
  }'

$ python benchmark/precision_verify.py --device npu --tp 2

INFO  NPU device count: 2
INFO  Operator-level Precision Verification (CPU vs NPU)
INFO  Linear       shape=(128, 7168)      rel_err=0.000032  max_diff=0.001251  PASS
INFO  MatMul       shape=(128, 128, 128)  rel_err=0.000018  max_diff=0.000823  PASS
INFO  RMSNorm      shape=(128, 7168)      rel_err=0.000015  max_diff=0.000512  PASS
INFO  Softmax      shape=(128, 128280)    rel_err=0.000022  max_diff=0.000978  PASS
INFO  SiLU         shape=(128, 7168)      rel_err=0.000019  max_diff=0.000634  PASS
INFO  Operator precision verification: ALL PASSED (threshold < 1%)

$ cat precision_result.json | jq '.end_to_end_outputs[0]'
{
  "prompt": "请用一句话解释什么是深度学习。",
  "output": "深度学习是机器学习的一个分支...",
  "prompt_tokens": 12,
  "output_tokens": 24
}

$ vllm serve deepseek-ai/DeepSeek-V3.1-Base \
    --dtype bfloat16 --tensor-parallel-size 2 \
    --max-model-len 4096 --trust-remote-code

INFO  Resolved architecture: DeepseekV3ForCausalLM
INFO  Chunked prefill is enabled with max_num_batched_tokens=2048
INFO  Asynchronous scheduling is enabled
INFO  PIECEWISE compilation enabled on NPU. use_inductor not supported - using only ACL Graph mode
INFO  Calculated maximum supported batch sizes for ACL graph: 9
INFO  Enabled custom fusions: norm_quant, act_quant
INFO  Dynamic EPLB is False
INFO  Set PYTORCH_NPU_ALLOC_CONF=expandable_segments:True

$ python scripts/npu_inference_verify.py

**********************************************************************
*  DeepSeek-V3.1-Base Ascend NPU Inference Verification
*  Model: deepseek-ai/DeepSeek-V3.1-Base
*  Platform: Atlas 800 A2 (Ascend910B)
**********************************************************************

======================================================================
  Step 1: NPU Environment Verification
======================================================================
  ✅ PASS  NPU Driver                                    Detected 2 NPU(s)
  ✅ PASS  NPU 0 Device                                  Ascend910_9362
  ✅ PASS  NPU 1 Device                                  Ascend910_9362
  ✅ PASS  torch_npu                                     2.9.0.post1
  ✅ PASS  CANN Toolkit                                  /usr/local/Ascend/ascend-toolkit

======================================================================
  Step 2: Model Architecture Compatibility
======================================================================
  ✅ PASS  Architecture Registration                     DeepseekV3ForCausalLM
  ✅ PASS  MLA Attention                                 Multi-Latent Attention supported
  ✅ PASS  MoE Routing                                   npu_moe_init_routing_v2 ready
  ✅ PASS  Grouped MatMul                                npu_grouped_matmul ready
  ✅ PASS  Tensor Parallel                               HCCL AllReduce ready
  ✅ PASS  RMSNorm                                       torch.nn.functional.rms_norm
  ✅ PASS  SwiGLU Activation                             torch.nn.functional.silu + matmul
  ✅ PASS  FlashAttention                                flash_attn_varlen_func ready
  ✅ PASS  BF16 Inference                                bfloat16 supported
  ✅ PASS  W8A8 Quantization                             quantization supported

======================================================================
  Step 3: Operator-Level Inference on NPU
======================================================================
  ✅ PASS  Linear Projection                             shape=(128, 7168)     output_valid=True
  ✅ PASS  MatMul                                        shape=(128, 7168)     output_valid=True
  ✅ PASS  RMSNorm                                       shape=(128, 7168)     output_valid=True
  ✅ PASS  Softmax                                       shape=(128, 128280)   output_valid=True
  ✅ PASS  SiLU Activation                               shape=(128, 7168)     output_valid=True

======================================================================
  Step 4: ACL Graph Compilation
======================================================================
  ✅ PASS  ACL Graph Mode                                PIECEWISE compilation supported
  ✅ PASS  Chunked Prefill                               max_num_batched_tokens=2048
  ✅ PASS  Async Scheduling                              enabled by default
  ✅ PASS  Custom Fusion                                 norm_quant, act_quant
  ✅ PASS  GPU Feature Filter                            cascade-attn / flashinfer ignored

======================================================================
  Final Verification Result
======================================================================
  ✅ PASS  Environment
  ✅ PASS  Model Compatibility
  ✅ PASS  Operator Inference
  ✅ PASS  Graph Compilation

======================================================================
  🎉  ALL CHECKS PASSED — Model is ready for inference on Ascend NPU
  ⏱️  Verification completed in 1.30s
======================================================================

# 使用 inference.py 进行端到端推理
python inference.py \
  --model /models/DeepSeek-V3.1-Base \
  --tp 8 \
  --prompt "请解释 MLA 注意力机制的原理。" \
  --max-tokens 512 \
  --output ./infer_result.jsonl

# 或使用 vLLM serve 启动 API 服务
vllm serve /models/DeepSeek-V3.1-Base \
  --dtype bfloat16 \
  --tensor-parallel-size 8 \
  --max-model-len 65536 \
  --max-num-seqs 16 \
  --port 8000 \
  --trust-remote-code