[yao-matrix]

Context

This is a doc & code-comment-only PR. No code changes are included.

Follow-up to #262, which is now merged. Rebased the accumulated documentation/comment changes onto upstream. Latest documentation head: d4d8794f; merged upstream main eceea1be.

Summary

• Move broader documentation and CUDA-to-accelerator comment rewording out of [4/N] extend CUDA-specific graph API to pytorch's accelerator graph backend, to make graph supported on Intel GPU (XPU) #262, including wording in helpers, tests, and YAML configs. Matching environment-variable rows, essential explanations of new abstractions, and repository skill guidance (AGENTS.md and .claude/skills/) remain in [4/N] extend CUDA-specific graph API to pytorch's accelerator graph backend, to make graph supported on Intel GPU (XPU) #262.

• Use accelerator graph terminology and update graph API references in comments and docstrings, including main's new LTX-2.5 shape buckets and combined-walk documentation.

• Define CG / cg_ as "captured graph" in the shared resource module.

• Correct Pi05 capture documentation: batch size 1, eager capture before forward compilation, and the actual runner/static-buffer contract. Correct Z-Image's module-level DummyRowPool reference.

• Document accelerator graph APIs, model options, and environment variables. Runtime code, configuration values, and shell commands are unchanged.

• Correct Qwen3-Omni code predictor input shapes, persistent KV cache lifecycle, and Triton RoPE description; document V-JEPA2 rollout/MPC wrappers and declared KV resources; remove references to deleted engine/cache classes.

• Clarify Qwen3-TTS CodePredictor cache ownership, shared-engine dispatch and resource declarations, and the worker's generic device events on the execution stream.

• Finish the CUDA-to-accelerator graph wording in Kokoro's bucket documentation and Qwen3-Omni's static-output docstring.

Verification (a3a1c491)

• uv tool run ruff check . and git diff --check upstream/main: passed.
• Accelerator graph compile modes, ragged attention CPU/config checks, worker speculative loop counters, piecewise config signatures, and Qwen3-TTS model tests: 116 passed, 10 skipped. Skips cover optional dependencies and a CUDA-only fused kernel.
• Initial collection of mstar/model/cosmos3/tests/test_denoise_compile.py failed because this environment lacks diffusers; the other targeted modules were subsequently run successfully.
• Executable Python ASTs match latest main across all 108 differing Python files after excluding docstrings and normalizing empty abstract-method bodies. All 9 differing YAML files and 3 shell files retain identical values/command tokens.
• Resolved conflicts in 16 files, preserving main's graph capture recovery, ragged per-step ceilings, speculative counters, and model refactors. Removed implementations deleted by main and updated Qwen3-TTS prefill/capture/cache-ownership comments to match the shared code predictor.
• AGENTS.md and .claude/skills/ are identical to main.
• No full CPU suite, GPU benchmark, or Sphinx rebuild was run for this documentation-only merge.

AI assistance was used for merge resolution and verification.

Wording follow-up (d4d8794f)

• Updated four model-note references to accelerator graphs: OmniVoice backbone, Higgs Audio encoder, Qwen3-Omni encoders, and VoxCPM2 Talker decode. Only docs/models.rst changed.
• uv tool run ruff check . and git diff --check: passed. Runtime tests were not rerun for these four prose edits; the merge validation above remains recorded at a3a1c491.

[yao-matrix]

@NSagan271 and @merceod, pls help review, this PR only includes doc & code-comments changes, thx.

[yao-matrix]

@NSagan271 @merceod gentle ping, thx. The PR failure is for timeout, seems not relevant to this PR.

[yao-matrix]

@NSagan271 , gentle ping, thx

[yao-matrix]

@NSagan271 , gentle ping, thx

[yao-matrix]

@NSagan271 @merceod , gentle ping, thx