perf(hy_worldplay): longer video bench + HY model card

[wenqingw-nv]

Follow-up to #155 for #203: GB300 perf re-bench, the hy_worldplay model card, and two num_chunk≥4 fixes the larger GPU surfaced.

Changes

• _action.py: unwrap the torch.compile OptimizedModule in prefill_memory_kv_cache; run memory-engaged chunks eager under CUDA graphs (the data-dependent memory torch.cat can't replay in a graph captured pre-memory). Both reachable only at num_chunk≥4.
• bench.sh: default HY_VENDOR_SDPA=1 (vendor on cuDNN SDPA, matching native). run.sh: add torchvision to the vendor deps.
• models/hy_worldplay.rst (LingBot-style) + perf-0530.md + native sample videos; registered in the toctree; clean under sphinx-build -W.

Bench

Native vs wan/generate.py: num_chunk=8 / pose=w-31 / seed=0 / 704×1280, warmup-discard 5, DiT+VAE scope, both legs cuDNN SDPA + torch.compile, single GB300. Post-warmup medians (chunks 5–7):

stage	native	vendor	speedup
DiT (diffuse)	632 ms	1206 ms	1.91×
VAE decode	383 ms	372 ms	~parity
DiT + VAE / chunk	1015 ms	1578 ms	1.55×

Per data_local/ input:

image	DiT nat/ven	VAE nat/ven	ratio	mean |Δ|
1.png	630 / 1208	382 / 374	1.56×	40.8
2.png	632 / 1210	382 / 374	1.56×	31.1
6.jpeg	628 / 1208	381 / 374	1.57×	21.1
cat_surf.jpg	632 / 1207	381 / 375	1.56×	51.8
median	631 / 1208	382 / 374	1.56×	—

Native ~1.91× on DiT, parity on VAE, input-independent (native DiT 628–632 ms across inputs). Parity |Δ| is cumulative bf16 AR drift (chunk-0 ≈14 matches #155's 12.91, ramps to ~49 by chunk-7, no jump at memory engagement); highest on the off-aspect cat_surf (625×350 upscaled).

Native vs vendor video pairs

image	native	vendor
1.png	hy-worldplay-wan-i2v-5b.mp4	hy-worldplay-wan-i2v-5b.mp4
2.png	hy-worldplay-wan-i2v-5b.mp4	hy-worldplay-wan-i2v-5b.mp4
6.jpeg	hy-worldplay-wan-i2v-5b.mp4	hy-worldplay-wan-i2v-5b.mp4
cat_surf.jpg	hy-worldplay-wan-i2v-5b.mp4	hy-worldplay-wan-i2v-5b.mp4

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[wenqingw-nv]

/ok to test 1e7a2a3

[liruilong940607]

The visual results here in the native column is choppy -- this seems to indicate something is wrong with the current implementation. Needs revisit and fix.

Let's not add this to the public-facing doc until we are sure it is correctly implemented.

[liruilong940607]

Lets not check in videos into this repo. you can send me the videos on slack and I will host them.

[wenqingw-nv]

Done — removed the checked-in .mp4s from the repo; I'll send the clips over Slack for hosting.

[liruilong940607]

This slipped through my previous review. Why would we need prefill_memory_kv? The BlockKVCache in flashdreams should naturally work with CUDAGraph.

I traced it and it seems to lead to HyWorldPlayMemoryKVCache, which stored roped K and roped V. What's the reason we can't use BlockKVCache for that? And what's the reason it can't be compatible with cudagraph like other models?

[wenqingw-nv]

HyWorldPlayMemoryKVCache is separate because memory isn't a rolling window, it's a per-chunk re-selection of non-contiguous frames, so BlockKVCache doesn't map directly. The non-graphability is just a stopgap; since the selected set is always exactly 16, a fixed-size in-place buffer is graph-compatible. This path also causes the choppy output, where the memory keys sit at collapsed RoPE positions [0..16) regardless of identity, so on the first window slide a frame's encoding shifts and the scene jolts.

I'll rework it into a fixed-size, consistently-positioned in-place buffer to fix both, and numerically diff vs vendor's prefill first. Do you think we should: gather K/V from retained history or re-prefill each chunk (vendor re-prefills)? @liruilong940607

[greptile-apps]

Greptile Summary

This PR lands two correctness fixes for num_chunk≥4 HY-WorldPlay runs, adds the hy_worldplay model card, and re-benchmarks on GB300. The code changes are small and targeted, the documentation follows existing patterns, and the shell/lock changes are well-explained.

• _action.py: Extracts memory_engaged as a standalone flag (set on the instance before super().predict_flow() is called) so that the new _select_network override can force CUDAGraphWrapper.drain for any denoising step in a memory-engaged chunk, and unwraps _orig_mod in prefill_memory_kv_cache so the prefill reaches the raw HyWorldPlayWanDiTNetwork instead of the compiled proxy.
• bench.sh / run.sh: Defaults HY_VENDOR_SDPA=1 for an apples-to-apples attention-backend comparison and moves torchvision from the native package's lock to a vendor-only uv pip install step, with a comment explaining the rationale.
• Docs: Registers hy_worldplay.rst in the toctree, ships benchmark data (perf-0530.md), and links externally-hosted sample videos; passes sphinx-build -W.

Confidence Score: 4/5

Safe to merge; the two runtime fixes are narrowly scoped to the memory-engaged path and are guarded by the same condition they test against.

The _action.py changes are well-reasoned: _hy_memory_engaged is always written by predict_flow before _select_network reads it (via super().predict_flow()), the _orig_mod unwrap falls back gracefully when the network isn't compiled, and the drain call is gated behind _use_cuda_graph so the non-graph debug path is unaffected. The only open item is the duplicate "Project page" / "Official code" link in the model card.

docs/source/models/hy_worldplay.rst — both header link buttons point to the same GitHub URL and should be reconciled once a dedicated project page is available.

Important Files Changed

Filename	Overview
integrations/hy_worldplay/hy_worldplay/_action.py	Two targeted num_chunk≥4 fixes: unwrap _orig_mod for prefill under compile_network=True, and override _select_network to drain memory-engaged chunks eagerly to avoid CUDA graph replay faults. Logic is sound; _hy_memory_engaged is set immediately before super().predict_flow() so _select_network always sees a fresh value.
docs/source/models/hy_worldplay.rst	New LingBot-style model card for HY-WorldPlay; benchmark chart, installation guide, and sample videos all present. Both header link buttons currently share the same GitHub URL (see comment).
integrations/hy_worldplay/tests/parity_check/bench.sh	Defaults HY_VENDOR_SDPA=1 so the vendor leg uses cuDNN SDPA to match the native leg for an apples-to-apples comparison; existing override semantics preserved via ${VAR:-default}.
integrations/hy_worldplay/tests/parity_check/run.sh	Adds torchvision==0.26.* to the vendor-only uv pip install block; justified because the upstream hyvideo import requires it and the native plugin does not, so it was intentionally removed from pyproject.toml.
integrations/hy_worldplay/tests/parity_check/uv.lock	Removes torchvision from the native package's lock (now installed externally for vendor-only use), relaxes torch spec from >=2.11 to >=2.9 (resolved pin stays 2.11.0), bumps tyro floor from >=0.8 to >=1.0, and adds >=4.5 specifier to opencv-python-headless.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["predict_flow()"] --> B["Evaluate memory_engaged\n(cache populated + memory indices present)"]
    B --> C["self._hy_memory_engaged = memory_engaged"]
    C --> D{memory_engaged AND\nprefill not done\nfor this chunk?}
    D -- yes --> E["prefill_memory_kv_cache()\n(unwrap _orig_mod to reach raw network)"]
    E --> F["cache.prefill_completed_for_chunk = ar_idx"]
    D -- no --> G["super().predict_flow()"]
    F --> G
    G --> H["_select_network(uncond)"]
    H --> I{_use_cuda_graph AND\n_hy_memory_engaged?}
    I -- yes --> J["return network_call.drain\n(eager, no CUDA graph replay)"]
    I -- no --> K["super()._select_network()"]
    K --> L{filling phase?\nar_idx < capture_idx}
    L -- yes --> M["return wrapper.drain\n(filling / autotune)"]
    L -- no --> N["return wrapper\n(CUDA graph capture/replay)"]

Loading

_{Reviews (1): Last reviewed commit: "docs(hy_worldplay): un-check-in sample v..." | Re-trigger Greptile}

[greptile-apps]

Duplicate "Project page" and "Official code" URLs

Both link buttons resolve to the same https://github.com/Tencent-Hunyuan/HY-WorldPlay URL. If there is no separate landing/demo page then the "Project page" label is misleading; the redundant button should either be removed or its href updated to the canonical project page (e.g., a Hugging Face space or a paper link) once one is available. As-is, readers who click both buttons land on identical content with no way to reach a distinct artifact.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.