Add cu130-torch213 dependency group (torch 2.13 on CUDA 13.0) (#134)

## What

Adds `cu130-torch213` and `cu130-torch213-train` dependency groups
pinning **torch 2.13.0+cu130** (with matching torchvision 0.28.0,
torchcodec 0.14.0, torchao 0.17.0, triton 3.7.1, `natten
0.21.6+cu130.torch213` on aarch64, and the cu13 NVIDIA runtime libs),
plus a `conflicts` entry making it mutually exclusive with the other
backend groups. `uv.lock` is regenerated for the new closure.

## Why the code change

torchvision 0.28 removed `torchvision.io.read_video`, so
`cosmos_framework/inference/vision.py` switches to a torchcodec-based
THWC uint8 decoder. It's forced onto CPU so an active default-CUDA
context during generation doesn't route torchcodec's internal
frame-index tensor to CUDA (which would raise `NotImplementedError`).

## Docs

Bumps the recommended NGC base image to
`nvcr.io/nvidia/pytorch:26.06-py3` in `README.md` and `docs/setup.md` to
match the torch 2.13 line.

## Verification

On 4× GB200 (aarch64):

- `uv sync --all-extras --group=cu130-torch213-train` resolves clean;
`torch.__version__ == 2.13.0+cu130`, `torch.version.cuda == 13.0`.
- Cosmos3-Nano `throughput` inference over **all 13**
`inputs/omni/*.json` samples (t2i, t2v, i2v, v2v, t2vs, i2vs, and the 6
action modes) produced valid outputs: videos decode with real pixel
variation (std 44–76, well above the degenerate floor), audio modes
carry an audio track, and all action arrays are non-empty and finite.

### Known limitations (in code comments)
- x86_64 has no torch2.13 wheels yet for flash-attn / TE / natten —
those are commented out / left on the torch2.10 build; this line is
validated on **aarch64**.
- torch 2.13.0+cu130 hard-pins cuDNN to 9.20.0.48; the cuDNN attention
backend wants ≥9.22 (else it falls back to NATTEN). Override at runtime
if needed.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: lfengad <liangf@nvidia.com>

Author: 3 people

README.md

@@ -57,7 +57,7 @@ uv sync --all-extras --group=cu130-train
 source .venv/bin/activate && export LD_LIBRARY_PATH=
 ```
 
-If you are starting from the recommended NGC image (`nvcr.io/nvidia/pytorch:25.09-py3`), see the [one-shot quickstart](./docs/setup.md#quickstart-from-the-recommended-base-image).
+If you are starting from the recommended NGC image (`nvcr.io/nvidia/pytorch:26.06-py3`), see the [one-shot quickstart](./docs/setup.md#quickstart-from-the-recommended-base-image).
 
 ## Training
 

cosmos_framework/inference/inference.py

@@ -14,7 +14,6 @@
 import cattrs.preconf.json
 import safetensors.torch
 import torch
-import torchvision.io
 from PIL import Image
 from qwen_vl_utils.vision_process import smart_nframes
 from torch.utils._pytree import tree_map_only
@@ -46,6 +45,7 @@
 from cosmos_framework.inference.vision import (
     build_conditioned_video_batch,
     build_image_edit_batch,
+    decode_video_thwc_uint8,
     load_conditioning_image,
     load_conditioning_video,
     load_prompt_upsampling_image,
@@ -574,17 +574,16 @@ def _get_prompt_sample_data(sample_args: OmniSampleArgs, model: OmniMoTModel, *,
 def _decode_reasoner_video(vision_path: str, video_fps: float | None) -> dict[str, Any]:
     """Decode a local video file into the frame-list payload the Qwen3-VL processor expects.
 
-    Returns ``{"frames": [PIL.Image, ...], "fps": float}``. Uses the same
-    ``torchvision.io.read_video`` decode the rest of the inference path relies on
-    (no ``decord`` dependency), then uniformly samples frames toward ``video_fps``
-    (default 2.0) via Qwen's ``smart_nframes``. The repo ``Qwen3VLProcessor`` runs
-    with ``do_sample_frames=False``, so it consumes this pre-sampled frame list
-    as-is and handles its own per-frame resize."""
-    frames, _, info = torchvision.io.read_video(str(vision_path), pts_unit="sec")  # [T,H,W,C] uint8
+    Returns ``{"frames": [PIL.Image, ...], "fps": float}``. Uses the same TorchCodec
+    decode the rest of the inference path relies on (no ``decord`` dependency), then
+    uniformly samples frames toward ``video_fps`` (default 2.0) via Qwen's
+    ``smart_nframes``. The repo ``Qwen3VLProcessor`` runs with ``do_sample_frames=False``,
+    so it consumes this pre-sampled frame list as-is and handles its own per-frame resize."""
+    frames, src_fps = decode_video_thwc_uint8(Path(vision_path))  # [T,H,W,C] uint8
     total_frames = int(frames.shape[0])
     if total_frames == 0:
         raise ValueError(f"Decoded zero frames from reasoner video: {vision_path}")
-    src_fps = float(info.get("video_fps") or 0.0) or 1.0
+    src_fps = src_fps or 1.0
     target_fps = video_fps if video_fps is not None else 2.0
     nframes = smart_nframes({"fps": target_fps}, total_frames=total_frames, video_fps=src_fps)
     idx = torch.linspace(0, total_frames - 1, nframes).round().long().tolist()

cosmos_framework/inference/vision.py

@@ -7,7 +7,6 @@
 
 import numpy as np
 import torch
-import torchvision.io
 import torchvision.transforms.functional as TF
 from PIL import Image
 
@@ -76,6 +75,26 @@ def load_conditioning_image(image_path: Path, target_h: int, target_w: int) -> t
     return img_tensor.unsqueeze(1)  # [3,1,target_h,target_w]
 
 
+def decode_video_thwc_uint8(path: Path) -> tuple[torch.Tensor, float]:
+    """Read all frames of a video as a uint8 [T, H, W, C] tensor plus fps.
+
+    TorchCodec replacement for ``torchvision.io.read_video``; mirrors its frame layout so
+    callers are unchanged. As of torchvision 0.28 the public wheel no longer binds that
+    symbol — ``torchvision/io/__init__.py`` imports it from the Meta-internal
+    ``pytorch.vision.fb.io.video`` under ``except ImportError: pass``, so ``import
+    torchvision.io`` still succeeds but calling ``read_video`` raises ``AttributeError``.
+    """
+    from cosmos_framework.utils.generator.torchcodec_video import decode_frames_nhwc_uint8, probe_video
+
+    # torchcodec's core ops are CPU-only. Force CPU tensor creation so an active default-CUDA
+    # device context (torch.set_default_device during generation) doesn't route torchcodec's
+    # internal frame-index tensor to CUDA and raise NotImplementedError.
+    with torch.device("cpu"):
+        num_frames = probe_video(path).num_frames
+        frames_nhwc, meta = decode_frames_nhwc_uint8(path, list(range(num_frames)))
+    return torch.from_numpy(frames_nhwc), float(meta.average_fps)
+
+
 def load_conditioning_video(
     video_path: Path,
     target_h: int,
@@ -88,7 +107,7 @@ def load_conditioning_video(
 
     ``keep`` selects which ``max_frames`` to take when the input is longer.
     """
-    frames, _, _ = torchvision.io.read_video(str(video_path), pts_unit="sec")
+    frames, _ = decode_video_thwc_uint8(video_path)
     frames = frames[-max_frames:] if keep == "last" else frames[:max_frames]  # [T,H,W,3]
     frames_tchw = frames.permute(0, 3, 1, 2).float()  # [T,3,H,W]
     frames_resized = _resize_and_center_crop(frames_tchw, target_h, target_w)  # [T,3,target_h,target_w]
@@ -185,9 +204,8 @@ def read_media_frames(path: Path, max_frames: int) -> tuple[torch.Tensor, float]
         return frames, 1.0
     if ext not in _VIDEO_EXTENSIONS:
         raise ValueError(f"Unsupported media extension: {ext}")
-    frames, _, info = torchvision.io.read_video(str(path), pts_unit="sec")
+    frames, fps = decode_video_thwc_uint8(path)
     frames = frames[:max_frames].permute(0, 3, 1, 2).permute(1, 0, 2, 3)
-    fps = float(info.get("video_fps", 24.0))
     return frames, fps
 
 

docs/setup.md

@@ -46,7 +46,7 @@ ______________________________________________________________________
 For CUDA 13 builds, the [NVIDIA NGC PyTorch container](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch) is the recommended starting point — it bundles PyTorch + CUDA 13 + cuDNN + NCCL tuned for NVIDIA hardware, plus Apex, TransformerEngine, and Megatron utilities that training infra users commonly need.
 
 ```dockerfile
-FROM nvcr.io/nvidia/pytorch:25.09-py3
+FROM nvcr.io/nvidia/pytorch:26.06-py3
 ```
 
 For CUDA 12.8 builds, pin to an earlier NGC tag (e.g. `nvcr.io/nvidia/pytorch:25.06-py3`) that still ships CUDA 12.
@@ -70,7 +70,7 @@ The two supported install paths are the recommended base image and the Docker co
 
 ### Quickstart: From the Recommended Base Image
 
-If you started from the [recommended base image](#recommended-base-image) (`nvcr.io/nvidia/pytorch:25.09-py3`), the following commands set up the full environment in one go. Run them **from the root of this repository** (i.e. inside the `Cosmos/` directory you just cloned):
+If you started from the [recommended base image](#recommended-base-image) (`nvcr.io/nvidia/pytorch:26.06-py3`), the following commands set up the full environment in one go. Run them **from the root of this repository** (i.e. inside the `Cosmos/` directory you just cloned):
 
 ```shell
 apt-get update

pyproject.toml

@@ -233,6 +233,46 @@ cu130-train = [
   "transformer-engine==2.12.0+cu130.torch210",
   "triton==3.6.0",
 ]
+# torch 2.13 on CUDA 13.0. Same layout as cu130 but pinned to torch 2.13.
+cu130-torch213 = [
+  #"flash-attn-3-nv==1.0.3+cu130.torch210; platform_machine == 'x86_64'", # TODO: torch213 build (cu130.torch210 wheels are cp313-only + torch210 ABI)
+  #"flash-attn==2.7.4.post1+cu130.torch210; platform_machine == 'x86_64'", # TODO: torch213 build (cu130.torch210 wheels are cp313-only + torch210 ABI)
+  # torch2.13 build (natten 0.21.6, sm 80/86/89/90/100/103/110/120/121) — matches this group's torch/ABI.
+  # Both arches now ship a torch213 wheel, so no platform split is needed.
+  "natten==0.21.6+cu130.torch213",
+  "torch==2.13.0+cu130",
+  "torchcodec==0.14.0+cu130", # https://github.com/meta-pytorch/torchcodec/releases
+  "torchvision==0.28.0+cu130", # TODO: release matching torch 2.13 — https://github.com/pytorch/vision/releases
+  # Dependencies determined from 'uv pip compile --group cu130-torch213' — these are the exact
+  # ==-pins that torch==2.13.0+cu130 (download.pytorch.org) declares. NOTE: this is the CUDA 13.0/13.1
+  # line the public wheel ships, NOT the CUDA 13.3 libs in nvcr.io/nvidia/pytorch:26.06.
+  # Issue: https://github.com/astral-sh/uv/issues/14237
+  "nvidia-cublas==13.1.1.3",
+  "nvidia-cuda-cupti==13.0.85",
+  "nvidia-cuda-nvrtc==13.0.88",
+  "nvidia-cuda-runtime==13.0.96",
+  # NOTE: torch 2.13.0+cu130 hard-pins this to ==9.20.0.48. The framework's cuDNN attention backend
+  # needs >= 9.22 (else it falls back to NATTEN, which is fine). To use the cuDNN backend, override at
+  # runtime: `uv pip install nvidia-cudnn-cu13==9.25.0.15` (can't pin here without a global override).
+  "nvidia-cudnn-cu13==9.20.0.48",
+  "nvidia-cufft==12.0.0.61",
+  "nvidia-cufile==1.15.1.6",
+  "nvidia-curand==10.4.0.35",
+  "nvidia-cusolver==12.0.4.66",
+  "nvidia-cusparse==12.6.3.3",
+  "nvidia-cusparselt-cu13==0.8.1",
+  "nvidia-nccl-cu13==2.29.7",
+  "nvidia-npp==13.0.0.50",
+  "nvidia-nvjitlink==13.3.33",
+  "nvidia-nvshmem-cu13==3.4.5",
+  "nvidia-nvtx==13.0.85",
+]
+cu130-torch213-train = [
+  {include-group = "cu130-torch213"},
+  "torchao==0.17.0+cu130; platform_machine == 'x86_64'", # 0.18.0+cu130 not published; no aarch64 build; https://github.com/pytorch/ao/issues/2919
+  "transformer-engine==2.12.0+cu130.torch210", # TODO: torch213 build
+  "triton==3.7.1", # torch==2.13.0+cu130 pins triton==3.7.1
+]
 # LIBERO simulator dependencies for the closed-loop eval client.
 # Mirrors packages/cosmos-policy
 # `libero` group; `libero` from PyPI declares `robosuite` transitively so we
@@ -274,10 +314,13 @@ repository = "https://github.com/NVIDIA/cosmos-framework"
 required-version = ">=0.11.3"
 conflicts = [
   [
+    {group = "vllm"},
     {group = "cu128"},
     {group = "cu128-train"},
     {group = "cu130"},
     {group = "cu130-train"},
+    {group = "cu130-torch213"},
+    {group = "cu130-torch213-train"},
   ],
 ]
 override-dependencies = [