MossYume int8 (ConvRot)

Unofficial int8 quantisation of the diffusion model of duongve/MossYume (a fine-tune of OpenMOSS-Team/MOSS-Image-Base). All credit for the model goes to its author; this repository only re-encodes the weights. License: Apache-2.0, as the original.

File diffusion_models/MossYume_v01_int8_convrot.safetensors (5,413,698,934 bytes, sha256 66e48d87ebea81aa7742b900c7e1f623ecd11b82345f6c91f50161c9faa4d159)
Source diffusion_models/MossYume_v01.safetensors from duongve/MossYume (10,164,655,808 bytes, sha256 8807f1ea05853ac04676101957322640fe28593f65e1ef8bf70f191e82747b27)
Format ComfyUI single file in the layout of the Comfy-Org moss_image_*_int8_convrot.safetensors files

Method

Written with scripts/export/comfy_int8.py of KohakuMossTrainer (commit 78416a2; the int8 ConvRot export was added in 851e246), on CPU, no calibration data: kogine run scripts/export/comfy_int8.py --set WEIGHTS=MossYume_v01.safetensors --set OUT=... --set DEVICE=cpu --set GROUP=256 --set SAVE_DTYPE=bf16.

  • ConvRot: the input dimension of each weight is rotated in blocks of 256 by a normalised regular Hadamard matrix (W R), which spreads weight outliers inside each block, then quantised to int8 with one symmetric fp32 scale per output row (max|row| / 127).
  • int8 (168 tensors): layers.N.proj.{qkv,o,ffn_up_gate,ffn_down}.weight of all 42 blocks, stored as weight (int8), weight_scale (fp32 [N,1]) and comfy_quant ({"format": "int8_tensorwise", "convrot": true, "convrot_groupsize": 256}).
  • bf16 (everything else, bit-identical to the source): text refiner, time / video embedders, modulation, norm gains, final linear. rope.bands stays fp32. Each block also has a comfy_attention.config hint tensor.

Usage

Use it like the Comfy-Org MOSS-Image int8 ConvRot diffusion model, in a ComfyUI build / loader that reads this format (KohakuMoss reads it as well). The text encoder (Qwen3.5-9B) and VAE (Flux.2 VAE) are not included: take text_encoders/qwen3.5_9b_bf16.safetensors and vae/flux2-vae.safetensors from duongve/MossYume (byte-identical to the Comfy-Org MOSS-Image files).

Sampling settings recommended by the original author: CFG 3-5, 30 steps, sampler er_sde, scheduler simple; native resolutions 1024-2048.

Measured agreement with the original (what was and was not tested)

  • Weights: the int8 weights, de-quantised and rotated back, were compared with the source bf16 weights for all 168 quantised layers: relative Frobenius error 0.85 % to 0.98 % (median 0.86 %; o projections 0.86-0.98 %, qkv 0.85 %, ffn_up_gate 0.85 %, ffn_down 0.90-0.91 %); worst layer layers.41.proj.o (0.98 %); largest single-element error 1.4e-2 (o). No non-finite values. All other tensors are identical to the source.
  • Forward pass (CPU, bf16 compute with the int8 layers de-quantised to bf16, upstream KohakuMoss backbone), one fixed-seed random 256x256 latent at sigma 0.9 / 0.5 / 0.1, conditioned on one real prompt embedding from the stock Qwen3.5-9B: velocity cosine int8 vs original MossYume 0.99993 / 0.99933 / 0.99915 (relative L2 1.2 % / 3.7 % / 4.1 %, no non-finite values). Residual-stream statistics are unchanged by quantisation: largest stream value 67,072 (int8) vs 67,072 (original), mean per-block RMS 98.65 vs 98.79, top-32 channel energy share 0.919 vs 0.919. With a fixed random text tensor instead of a real embedding the cosines were 0.9985 / 0.9959 / 0.9973.
  • Not tested: image quality (no image was generated), real image latents, more than one prompt, int8 GEMM kernels (the test de-quantises), ComfyUI loading of this exact file. It is a numerical agreement check on one input, not an image-quality test.
  • Source model is not stream-healed (same test, same inputs): largest residual-stream value 67,072 (MossYume) vs 58,112 (MOSS-Image-Base) vs 3,104 (KBlueLeaf/MOSS-Image-heal); top-32 channel energy share 0.919 / 0.877 / 0.767; channel 1356 dominates every block of MossYume and of the base. The weights agree: 344 weight elements above 64x their tensor RMS (base 349, healed 0). This quantisation does not change that.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for SumomoLee/MossYume-int8

Finetuned
duongve/MossYume
Quantized
(1)
this model