Skip to content

Latest commit

 

History

History
33 lines (28 loc) · 2.49 KB

File metadata and controls

33 lines (28 loc) · 2.49 KB

Choosing Formats — Full Reference

Back to README

Ten format families are available. The safetensors formats run through convert_to_quant / comfy-kitchen; the GGUF formats run through ggufy.

Format Best Use Notes
FP8 RTX 40/50-series quality baseline Good default for video model compression
NVFP4 Blackwell VRAM savings More aggressive; sensitive layers rescued to FP8 where local tables exist
NVFP4 HQ MiniMax H3 VRAM + quality balance Mixed NVFP4 profile: same --nvfp4 --comfy_quant command as NVFP4, but a verified 30-layer plan (27× attn.out_proj + 3× mlp.fc2 at specific blocks) stays at source BF16 while the other 170 main-matrix layers stay NVFP4-packed. MiniMax H3 only
MXFP8 Blackwell microscaling Pure MXFP8; requires SM >= 10.0 (Blackwell). Use Hybrid for Ada compatibility
Hybrid MXFP8 Ada + Blackwell compatibility Two-pass: MXFP8 quantize then hybrid conversion with tensorwise FP8 fallback
INT4 ConvRot Maximum compression w4a4 ConvRot via comfy-kitchen TensorCore layout. Supports LTX-2.3, WAN 2.2, Krea 2, MiniMax H3. Requires BF16/FP16 source (refuses lossy sources)
W4A8 MiniMax H3 reference low-bit asym_w4a8_int8 ConvRot via comfy-kitchen AsymW4A8Int8Layout (packed INT8 + 16-value codebook + FP8 group scales, ConvRot group 256). MiniMax H3 only; packs the heavy linears, preserves structural layers. Requires BF16/FP16 source, Simple strategy
INT8 Tensor-wise Safer INT8 path Recommended INT8 choice for broad ComfyUI compatibility
INT8 Row-wise ConvRot Runtime Experimental/runtime-specific INT8 Requires inference code that reads ConvRot metadata and rotates activations
GGUF Q formats llama.cpp-style deployment Uses ggufy plus sensitivity maps for verified video tensors

Practical notes

  • If an INT8 model produces pixel clutter, first try INT8 Tensor-wise. ConvRot row-wise INT8 only works correctly when the runtime implements the matching activation rotation.
  • INT4 ConvRot and W4A8 require Simple strategy, BF16/FP16 source, and comfy-kitchen[cublas] (W4A8 needs the AsymW4A8Int8Layout build, installed via the unpinned default-branch comfy-kitchen).
  • MXFP8 requires SM >= 10.0 (Blackwell); use Hybrid MXFP8 for Ada compatibility.
  • W4A8 and NVFP4 HQ are MiniMax H3 only. NVFP4 HQ is a quality variant of NVFP4 — the same packed NVFP4 layout plus a verified per-block BF16 retention plan (30 heavy linears kept at source precision).