jajmangold/DeepSeek-V4-Flash-fni8

DeepSeek-V4-Flash — fni8 (int8/W4A8 dp4a, Volta sm_70)

DeepSeek V4-Flash (MLA) model, quantized from deepseek-ai/DeepSeek-V4-Flash. Repackaged to the .fni8 resident format (~150 GB (int8)) for the fni8 W8A8/W4A8 DP4A kernels on NVIDIA Volta (sm_70) — Tesla V100 / CMP 100-210.

Status

Not a working int8 model yet. This checkpoint runs only via the fp16 MLA-decompress fallback path — the int8 latent-absorb (MLA) kernel is still pending, so you do not get the int8 compute win here. Storage is ~150 GB. Treat as an early/experimental checkpoint. Performance is fleet-specific. All fni8 speedups are measured on the CMP 100-210 mining-card fleet, where the fp16 tensor cores are firmware-gimped. These numbers do not transfer to a real Tesla V100 (whose fp16 tensor cores would beat dp4a).

Format

• Weights: int8 per-row (W8A8), fp32 scales, resident dp4a VRAM layout.
• Why dp4a: sm_70 has no int8 tensor cores, so the matmul contraction runs on the __dp4a CUDA-core intrinsic. On the CMP 100-210 fleet (whose fp16 tensor cores are firmware-limited) dp4a is the fast path, not a compromise.

How to run

fni8-serve is the LLM runtime (load_fni8_state_dict(<file>) into an LLMEngine; the architecture is read from the file). ComfyUI-fni8 is for diffusion DiTs only and does not load this model.

Limitations

• Quantization is lossy: int8 (and especially int4) outputs differ from the fp16/bf16 parent, and the difference varies by task.
• Capabilities, biases, and risks of the parent model carry over — see the parent card.
• This is a derivative quantization, not a relicense; the parent model's license and acceptable uses apply.

---

Part of the fni8 stack: kernels · LLM serving · ComfyUI DiTs.