NVIDIA CMD Model Loader

Every workflow in the ComfyUI-NVIDIA-CMD pack starts here, and this is the node that's going to hurt. NVIDIACMDModelLoader is the one that loads NVIDIA's CMD student pipeline into memory - it's slow, it's the only part that can genuinely fail, and it's where nearly all your setup mistakes surface. Once it's green, the rest of the pack is easy.

CMD is NVIDIA's causal few-step image-to-video model: a distilled student of Cosmos-Predict2.5-2B, built on the Cosmos stack. The official stack expects Linux and FlashAttention. This adapter from hiroki-abe-58 (the same person behind the ComfyUI-Win-Blackwell stack) is what runs the student on Windows Native, no WSL, using PyTorch SDPA on Blackwell instead of a flash-attn wheel. This node is the front door to all of that.

How it works

The loader builds the official CausalInferencePipeline from your nv-tlabs/cmd checkout, then does a bunch of scoped patching: it swaps torch.compile for an identity function, keeps Cosmos-Reason1 (the 7B text encoder) off the GPU in the BALANCED preset, and caps the KV cache to the checkpoint's local_attn_size so long rollouts don't balloon. Load once per ComfyUI session and reuse the cmd_model across many generations - the model stays loaded and the CMD patches persist in the process.

Crucially, this node never downloads weights. If something's missing you get a FileNotFoundError that tells you exactly what to place and where. There is no silent auto-fetch.

The inputs that matter

Only two need your attention. The rest you can leave on defaults:

• checkpoint - picks which CMD student you get. Default is chunk1_short (plain 24-latent I2V). Use chunk1_long for the long-video node and chunk1_camera for camera control. It must match the node you wire the model into.
• memory_preset - BALANCED (default) keeps Reason1 on CPU after text encode and the VAE on GPU; FULL keeps everything on the card (big-VRAM only); LOW_VRAM offloads more. On an RTX 5090 32GB, BALANCED peaks around 22–24 GiB during generation.

attention defaults to sdpa - that's the whole point of the pack, leave it. dtype is bfloat16 and device is cuda; both are locked to a single choice. The optional checkpoint_path and model_root are overrides you only touch if your models live somewhere unusual. Output is a single cmd_model that feeds into NVIDIACMDImageToVideo or NVIDIACMDLongVideo.

Install

Clone into custom_nodes, install the adapter's own deps (it does not want flash-attn, Transformer Engine, or the official CMD requirements), and clone the official repo it calls:

cd <ComfyUI>\custom_nodes
git clone https://github.com/hiroki-abe-58/ComfyUI-NVIDIA-CMD ComfyUI-NVIDIA-CMD
cd ComfyUI-NVIDIA-CMD
.\<ComfyUI-python> -m pip install -r requirements.txt
git clone https://github.com/nv-tlabs/cmd.git third_party\cmd

Then the manual part - weights go into <ComfyUI>/models/nvidia_cmd/:

transformer/chunk1_short_t24_l21.safetensors   (plus the long/camera chunks you use)
text_encoder/                                  nvidia/Cosmos-Reason1-7B
vae/tokenizer.pth                              nvidia/Cosmos-Predict2.5-2B (gated), or Wan2.1_VAE.pth

The official hf download commands are in the README. Cosmos-Predict2.5's tokenizer.pth is gated, so accept the license on Hugging Face first - or grab the Wan2.1 VAE fallback, which uses the same mean/std.

Gotchas

• License first. The CMD code and student weights are NVIDIA OneWay Noncommercial - research and education only. The adapter itself is Apache-2.0, but using it doesn't grant you a commercial license to the models. Read the NOTICE.
• The load is a minutes-long affair. It's constructing and loading a 2B DiT plus a 7B text encoder. Be patient, don't kill it.
• Linux, RTX 40-series, and non-portable ComfyUI are untested by the author. This pack is a Windows/Blackwell bet.
• If the node refuses with a "CMD file not found" error, you're missing weights - recheck the layout above. If it complains about a config, third_party\cmd isn't where the loader expects it.

Get past this node once, and the rest of the pack is genuinely pleasant.