minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[diffusion][docs] Fix Cosmos3 model sizes (#32916)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2026-07-30 22:10:23 +08:00
committed by GitHub
parent 2e9c82b359
commit 7784ac8f91
3 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs_new/cookbook/diffusion/Cosmos/Cosmos3.mdx
+1 -1
View File
@@ -15,7 +15,7 @@ import { DiffusionModelTags } from '/src/snippets/diffusion/model-tags.jsx';
| Model | Status | Notes |
| --- | --- | --- |
| `nvidia/Cosmos3-Nano` | Supported | T2I, T2V, I2V, V2V, joint sound, and action |
| `nvidia/Cosmos3-Super` | Supported | T2I, T2V, I2V, and V2V; use multi-GPU for the 32B checkpoint |
| `nvidia/Cosmos3-Super` | Supported | T2I, T2V, I2V, and V2V; use multi-GPU for the 64B checkpoint |
| `nvidia/Cosmos3-Super-Text2Image` | Supported | T2I-specialized checkpoint |
| `nvidia/Cosmos3-Super-Image2Video` | Supported | I2V-specialized checkpoint |
| `nvidia/Cosmos3-Nano-Policy-DROID` | Supported | DROID policy action generation |
docs_new/docs/sglang-diffusion/compatibility_matrix.mdx
+2 -2
View File
@@ -612,8 +612,8 @@ Optimization columns are abbreviated to keep the matrix readable:
- `resident` usually provides the best latency/throughput but uses much more VRAM.
- Default is auto: `resident` on H200/high-memory CUDA GPUs, otherwise `original`.
- Deprecated compatibility: `snapshot` is accepted as an alias for `original` and may be removed after two release cycles.
5. Cosmos3 ships in two sizes — `nvidia/Cosmos3-Nano` (8B) and
`nvidia/Cosmos3-Super` (32B). Both share the same pipeline; the only
5. Cosmos3 ships in two sizes — `nvidia/Cosmos3-Nano` (16B) and
`nvidia/Cosmos3-Super` (64B). Both share the same pipeline; the only
difference is transformer depth and width, picked up from
`transformer/config.json` at load time. A single checkpoint serves T2V,
I2V (`--image-path`), and T2I (`--num-frames 1`).
python/sglang/multimodal_gen/registry.py
+1 -1
View File
@@ -1035,7 +1035,7 @@ def _register_configs():
# Cosmos3 — single checkpoint serves T2V, I2V, and T2I. Mode is dispatched
# per-request inside the pipeline from ``num_frames`` and ``image_path``.
# Both Nano (8B) and Super (32B) share the same pipeline; arch dimensions
# Both Nano (16B) and Super (64B) share the same pipeline; arch dimensions
# come from ``transformer/config.json`` via ``update_model_arch``.
register_configs(
sampling_param_cls=Cosmos3SamplingParams,