Skip to content

Add cosmos-nvfp4 0.1.0 (cu130.torch210, sm100/sm103) to v1.5.0 index#59

Merged
lfengad merged 1 commit into
nvidia-cosmos:mainfrom
Aaronhuang-778:add-cosmos-nvfp4
Jul 1, 2026
Merged

Add cosmos-nvfp4 0.1.0 (cu130.torch210, sm100/sm103) to v1.5.0 index#59
lfengad merged 1 commit into
nvidia-cosmos:mainfrom
Aaronhuang-778:add-cosmos-nvfp4

Conversation

@Aaronhuang-778

@Aaronhuang-778 Aaronhuang-778 commented Jun 30, 2026

Copy link
Copy Markdown
Collaborator

Adds cosmos-nvfp4 to the v1.5.0 index — the cosmos3 NVFP4 FP4-GEMM + activation-quantize CUDA
kernels packaged as a pre-built wheel (per the discussion with @Liang Feng about pulling these out
of the cosmos3 inference package so it stays free of on-the-fly native builds).

The two aarch64 wheels are already uploaded to the v1.5.0 release (cu130 / torch 2.10 / py3.13),
following the existing local-version naming convention:

  • cosmos_nvfp4-0.1.0+cu130.torch210.sm100 — GB200 / B200 (sm_100a); validated: imports + registers torch.ops.cosmos3_fp4.*
  • cosmos_nvfp4-0.1.0+cu130.torch210.sm103 — GB300 / B300 (sm_103a); experimental: compiles + correct SASS, not yet run on GB300

This PR adds the matching PEP 503 index entry (docs/v1.5.0/cosmos-nvfp4/index.html + one line in the
top-level index). The URLs resolve against the uploaded release assets (HTTP 200), so it's ready to
merge. Apache-2.0, built on NVIDIA CUTLASS (BSD-3-Clause).

@lfengad lfengad enabled auto-merge (squash) June 30, 2026 13:30
@lfengad lfengad disabled auto-merge July 1, 2026 04:09
@lfengad lfengad merged commit e289b1b into nvidia-cosmos:main Jul 1, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants