A technical overview of the Streaming Code World Model, our setup and inference results, current limitations, and the plan to scale longer rollouts. This is a report on completed experiments, not a live streaming demo.
↗ Source repository ↗ Streaming checkpointsPurpose, source code, model weights, and the inference approach.
TingtingLiao/streaming-code-world-model is a research codebase for streaming or chunk-wise world-model video generation built on the Code World Model stack.
minimax_h3_ref2va_bf16.safetensors from the locally installed CWM release assets.streaming_dmd_4step.safetensors.streaming_df.safetensors.The model runs on our VM. Cloudflare hosts this report page and MP4 files; it does not host the GPU model or provide inference by itself.
These players load the actual MP4s copied from the VM's inference output folders.
Single-chunk baseline confirming the end-to-end generation and decode path.
runs/first_inference/example_01_stream.mp4
Two-chunk distributed inference confirms continuation beyond the baseline.
runs/two_chunk_fsdp/example_01_stream.mp4
Six chunks completed with two GPUs and KV reuse. H.264 MP4 was produced successfully.
runs/six_chunk_fsdp/example_01_stream.mp4
If a player cannot load its video, check that the MP4 was copied into the matching videos/ directory before deployment.
Chunk count controls the length of the generated rollout.
The current repo configuration uses 17 video frames per chunk at 24 fps, with 5 latent time steps per chunk. Six chunks decode to approximately 4.25 seconds.
Later chunks can use recent generated history through the attention KV cache. The model's configured span controls how much recent context is retained.
| Chunks | Approx. duration | Result |
|---|---|---|
| 1 | 0.709 s | Completed |
| 2 | 1.417 s | Completed |
| 6 | 4.25 s | Completed |
| 24 | ~17 seconds target | OOM failure |
Progress from repository setup to a reproducible inference path.
Cloned the streaming repo, base CWM repo and Musubi Tuner; installed the streaming package in editable mode.
Used PyTorch SDPA fallback because the FlashAttention-3 interface was unavailable in this environment.
Completed 1-, 2- and 6-chunk runs, inspected MP4 properties, and added per-chunk GPU-memory diagnostics.
The inference path works, but long-rollout memory behavior needs investigation.
The failure occurred around chunk 6 in rtcwm/model/attention.py, while concatenating cached and current attention keys using torch.cat. The failed allocation was about 582 MiB with roughly 575 MiB free. Other users' processes were also consuming GPU memory during testing.
Hardware environment and the checks used to verify today's results.
| Operating system | Ubuntu 22.04.5 LTS |
|---|---|
| GPU hardware | 2 × NVIDIA RTX PRO 6000 Blackwell Server Edition |
| VRAM | ~97,887 MiB per GPU (about 96 GiB each) |
| System RAM | 354 GiB |
| NVIDIA driver | 580.159.03 |
| Driver-reported CUDA | 13.0 |
| Python environment | Python 3.10 |
| PyTorch | 2.9.1+cu128 |
| Distributed inference | torchrun, 2 processes; NCCL Socket workaround |
This is an engineering smoke test and scalability investigation—not yet a formal video-quality or real-time latency benchmark.
Prioritized steps to move from a successful short rollout to reliable longer generation.
nvidia-smi and coordinating a suitable GPU window.Local folders used in today's experiments.
| Streaming repo | /devwork/harsha/streaming-cwm/streaming-code-world-model |
|---|---|
| Base CWM assets | /devwork/harsha/streaming-cwm/code-world-model/release |
| Streaming checkpoint | ckpt/streaming_dmd_4step.safetensors |
| Successful 6-chunk MP4 | runs/six_chunk_fsdp/example_01_stream.mp4 |
| Memory-debug MP4 | runs/memory_debug_6/example_01_stream.mp4 |