# DeepSeek-V4-Flash MXFP4 Experts + INT8 Dense

Experimental DeepSeek-V4-Flash MXFP4 experts + INT8 dense checkpoint for Ampere
serving in the AppMana vLLM fork.

Source checkpoint: `deepseek-ai/DeepSeek-V4-Flash@fd53f944496234770ba80e15004f9b6d269a71f5`

Conversion:

```bash
CUDA_VISIBLE_DEVICES=1 .venv/bin/python tools/ampere/dsv4_requant_checkpoint.py \
  --src /home/administrator/inference/.cache/huggingface/models--deepseek-ai--DeepSeek-V4-Flash/snapshots/fd53f944496234770ba80e15004f9b6d269a71f5 \
  --dst /home/administrator/inference/deepseek-v4-flash-dsv4-mxfp4-int8-channel-vllm \
  --device cuda:0 \
  --dense-int8-strategy channel \
  --expert-format mxfp4 \
  --num-output-shards 72 \
  --overwrite
```

Expected canary values:

- `quantization_config.quant_method`: `dsv4_mxfp4_int8`
- `quantization_config.format`: `mxfp4_int8_packed`
- Dense linear weights: signed INT8, channelwise scales
- Routed experts: native MXFP4 with native E8M0 scales
- `expert_dtype`: `fp4`

This artifact is still under active development and should be evaluated for
quality before production use.