DeepSeek-V4-Flash-FP8
FP8 (block-wise e4m3 + f32 scale, 128×128 block) repackaging of deepseek-ai/DeepSeek-V4-Flash-0731.
What this is
The source checkpoint mixes three storage formats:
- MXFP4 (FP4 packed in int8 + E8M0 scale) — routed experts only
- MXFP8 (F8_E4M3 + E8M0 scale) — attention, shared experts, MTP
- BF16 / F32 — norms, embeddings, gates
This repo unifies everything into a single block-wise FP8 layout so the model can run through one FP8 kernel path.
Conversion
| Tensor | Source | Target |
|---|---|---|
ffn.experts.*.{w1,w2,w3} |
int8 (FP4) + E8M0 scale | F8_E4M3 + F32 scale |
attn.*, shared_experts.*, mtp.*.e_proj |
F8_E4M3 + E8M0 scale | F8_E4M3 + F32 scale (weight unchanged) |
attn.wo_a.weight |
F8_E4M3 + E8M0 scale | BF16 (scale absorbed) |
attn.wo_a.scale |
E8M0 | dropped |
| norms, embed, gate, etc. | BF16 / F32 | unchanged |