Cosmos3-ConvRot

Cosmos3 quantized transformers

Weight-only quantized transformers of the NVIDIA Cosmos3 models, for use with ComfyUI-Cosmos3. Each file is one quantized transformer; take the VAE, tokenizer(s) and config.json from the matching official nvidia/Cosmos3-* repo.

Two formats:

• int8 — weight-only INT8, per-output-channel scale + a group-wise Hadamard rotation (ConvRot).
• int4 — MLP weights INT4 (GPTQ-calibrated with a ConvRot Hadamard rotation, packed as AWQ W4A16); attention kept INT8; the excluded layers (below) stay bf16. Larger quantization error than int8 (at a fixed seed the output diverges from bf16 more than int8 does); about a third of the bf16 size.

Both are weight-only: activations stay bf16, so this lowers memory/download size, not compute speed. int4 runs at about bf16 speed — the per-forward dequantize and un-rotate add a little.

Files and sizes

File names: Cosmos3-<name>-int8-convrot.safetensors and Cosmos3-<name>-int4-convrot.safetensors. int4 and int8 are provided for every model.

Prompting note (Edge): Cosmos3-Edge is trained on JSON-structured prompts and is less robust to plain text than the larger Nano/Super. Plain text usually works, but on some detailed scenes (notably reflective surfaces) it can produce flare/pulsation artifacts; wrapping the text as {"temporal_caption": "<your prompt>"} avoids them. This is a base-model property, not a quantization effect (it shows in bf16 too).

Runtime footprint

With ComfyUI's dynamic VRAM the transformer is streamed from host RAM, so the GPU holds only the activations. Measured at 832×480, 93 frames, at the minimum VRAM budget (maximum streaming):

Min VRAM is the activation floor (set by resolution × frame count, not the weight format). RAM is the peak host memory — larger than the file on disk (staging + overhead), and bf16 peaks near twice the weight size. RAM and VRAM trade off: giving the GPU more VRAM holds more weights on-card and lowers the RAM figure. The Super-family int4 checkpoints fit a 64 GB host (≈63 GB); int8 needs a little more (≈67 GB).

Usage

1. Download the official nvidia/Cosmos3-<name> into ComfyUI/models/cosmos3/<name>/.
2. In its transformer/ folder, delete the bf16 shards and *.index.json, then put the quantized file there renamed to diffusion_pytorch_model.safetensors. Keep the official config.json, vae/, text_tokenizer/, sound_tokenizer/.
3. Load with the Cosmos3 Loader (weight_dtype = default); the loader reads the format from the checkpoint metadata. Requires comfy-kitchen (int8 from ComfyUI >= 0.27) and the latest ComfyUI-Cosmos3 (int4 needs the ConvRot-aware loader).

How these were quantized

Reproducible in method, not bit-for-bit (the calibration set and RNG vary per run). Both formats keep the same escape set in bf16: proj_in, proj_out, time_embedder, audio_proj, modality_embed, the embeddings, and all norms and biases.

int8 — every linear

Weight-only symmetric INT8, per-output-channel scale (weight_scale, float32, [out, 1]), with a group-wise Hadamard rotation (ConvRot, group 256) applied before quantization and undone at load. comfy_quant tag int8_tensorwise, convrot=true. No calibration: at 8-bit the per-channel scale and the rotation keep the round-to-nearest error small — error feedback (GPTQ) is only needed at int4. Produced with convert_to_quant:

ctq -i transformer_bf16.safetensors -o out_int8_convrot.safetensors \
    --int8 --scaling_mode row --simple --convrot --convrot-group-size 256 \
    --comfy_quant --save-quant-metadata --cosmos3 --device cuda --low-memory

int4 — MLP → int4, attention → int8

Round-to-nearest INT4 — even with the ConvRot rotation — leaves visible artifacts on these models, so the MLP path uses GPTQ error compensation on real activations. Steps:

1. Capture activations. Run genuine denoising at 832×480, 93 frames with the model's normal sampler (t2v / base-i2v: 35 steps, cfg 6, uni_pc_bh2; 4-step model: 4 steps, cfg 1, euler) over ≈4 prompts, with a forward pre-hook on every target linear. Keep a reservoir of up to 4096 rows per layer (random replacement beyond that). The understanding tower sees the text prefill; the generation tower sees every denoising step.
2. Rotate (ConvRot). Apply a Sylvester block-Hadamard — symmetric, orthonormal, group 32 — to both the weight and the captured activations of each MLP linear.
3. GPTQ. Quantize the rotated weight to symmetric INT4 (codes −8…7, per-(row, group-32) scale). Hessian H = XᵀX · 2/N from the rotated activations; diagonal damping raised through {0.01, 0.03, 0.1, 0.3, 1, 3} × mean(diag) until the Cholesky factors; columns processed in blocks of 128 with per-column error feedback into the not-yet-quantized columns; plain round-to-nearest only if damping never succeeds.
4. Attention → int8 (row-wise, [out, 1] scale). INT4 on attention produces visible artifacts.
5. Pack the INT4 codes into comfy-kitchen's AWQ W4A16 layout; the loader un-rotates each group at dequant.

Layer counts follow the checkpoint's config.json — e.g. Cosmos3-Super-Image2Video packs 384 MLP linears (INT4) + 512 attention linears (INT8).

Execution. Both formats run as an explicit dequantize-then-bf16 matmul (unpack the 4-bit weights to bf16, un-rotate, multiply in bf16). This is weight-only regardless: there is no int4-weight × bf16-activation tensor-core op on any GPU (Hopper or Blackwell — Blackwell's FP4 cores are for NVFP4, both operands 4-bit, not this W4A16 layout), so a W4A16 kernel would dequantize internally too, with no compute speedup either way. We dequantize explicitly instead of calling comfy-kitchen's gemv_awq_w4a16, which is non-deterministic (atomic accumulation jitters the video frame-to-frame) and numerically off on these shapes.

License

Derived from NVIDIA Cosmos3 checkpoints; the OpenMDW-1.1 license applies (same as the upstream nvidia/Cosmos3-* models).