Qwen3.6-27B-FP8 — fni8 (int8/W4A8 dp4a, Volta sm_70)
⚠️ EXPERIMENTAL — do not rely on this checkpoint.
It was converted before the hybrid-config-drop converter fix: the hybrid (Gated-DeltaNet) portion of the config was dropped during conversion, so the file may not load or run correctly. A corrected reconversion is pending. Even once reconverted, full decode additionally requires the Track-2 DeltaNet int8 kernel (currently a fp16 linear-attention fallback).
Qwen3.5/3.6-family hybrid (Gated-DeltaNet) model, quantized from Qwen/Qwen3.6-27B-FP8. Repackaged to the .fni8 resident format (~25.6 GB (int4)) for the fni8 W8A8/W4A8 DP4A kernels on NVIDIA Volta (sm_70) — Tesla V100 / CMP 100-210.
Status
See the experimental warning above. Not validated end-to-end.
Format
- Weights: int8 per-row (W8A8) or int4 per-group (W4A8), fp32 scales, resident dp4a VRAM layout.
- Why dp4a: sm_70 has no int8 tensor cores, so the matmul contraction runs on the
__dp4aCUDA-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.