INCModel3/DeepSeek-V4.1-Flash-W4A16-Engram-AutoRound

DeepSeek-V4.1-Flash — W4A16 (experts INT4 GPTQ + Engram INT4) — AutoRound

Post-training quantization of DeepSeek-V4.1-Flash produced with auto-round --model_free (RTN, iters=0):

Checkpoint size: 451.7 GB (vs 510.3 GB official). gsm8k (raw 5-shot, thinking off, n=1319): 93.93 / 94.01 vs baseline 92.87 — lossless within noise.

Inference (vLLM)

Requires: vLLM main (≥ PR #56201)

DSV41_ENGRAM_DTYPE=int4 NCCL_NVLS_ENABLE=0 \
vllm serve <this-model-dir> \
  --tensor-parallel-size 4 --max-model-len 8192 --language-model-only \
  --gpu-memory-utilization 0.90 --port 8100

Verified on 4×H200 (TP4). The INT4 expert GEMM runs through vLLM's built-in GPTQ-MoE path (MoeWNA16/Marlin) — no custom kernels; dense layers run as plain BF16.

Quantization recipe

# experts only -> auto_gptq INT4 g32 (everything else ignored / dequantized to bf16)
auto-round --model_name deepseek-ai/DeepSeek-V4.1-Flash --model_free \
  --scheme W4A16 --group_size 32 \
  --ignore_layers attn,shared_experts,engram,vision,aligner,main_proj,gate,norm,embed,head \
  --output_dir <A>

# engram tables -> int4 g32 sym + fp16 scale (custom RTN script), then splice into A
# (see producer notes: quant_engram_int4.py + build_b_view_experts.py)

Limitations

• —Serving today requires the dsv41-quant-plugin (not yet upstreamed to vLLM) and DSV41_ENGRAM_DTYPE=int4 for the engram lookup.
• —Not compatible with the official reference implementation's FP8 kernels (dense is BF16 here).

Provenance

• —auto-round 0.15.0 (editable copy with DeepSeek-V4.1 source-format patch), vLLM main nightly (0.1.1.dev39+g46d2b23ac), transformers-side loading not used (model_free).
• —Full production log & design docs: see producer's journal (aqa pipeline).