fix: hash UNet engine cache key so ONNX export path stays under Windows MAX_PATH

[forkni]

Summary

Cherry-picks 153669b from fix/unet-engine-path-maxpath (dotsimulate upstream PR cumulo-autumn#29, still open/unmerged) onto SDTD_032_dev. Without this, an sd-turbo + fp8 + tensorrt build with a nested engine_dir produces a UNet cache-key directory name long enough that the derived unet.engine.onnx / .opt.onnx paths exceed Windows' 260-char MAX_PATH — mkdir succeeds but torch.onnx.export's open(path, "wb") raises a bare FileNotFoundError, surfaced to users as "Acceleration has failed".

The fix hashes the full UNet prefix (same idea as the existing _lora_signature LoRA-hash pattern) into a short, stable, collision-free directory name, and adds a warning if the derived path would still exceed MAX_PATH.

• src/streamdiffusion/acceleration/tensorrt/engine_manager.py — UNet-only: short_hash = sha1(canonical_prefix)[:12], prefix = f"{base_name}{fp8_tag}--h{short_hash}{res_tag}", plus a longest_derived >= 250 warning.
• tests/unit/test_engine_path_length.py — 3 regression tests. Adapted from the upstream commit: SDTD_032_dev's get_engine_path doesn't yet have the pin_cache_frames/cache_maxframes params (a 040-release-only feature), so _CRASH_UNET_KWARGS is filtered against inspect.signature(get_engine_path) before use — keeps the test portable across branches without hand-editing it per cherry-pick. The UNet hashing fix itself is unconditional and unaffected by this filtering.

Companion PR on SDTD_040_beta_release: #9.

Does not touch fix/unet-engine-path-maxpath or upstream PR cumulo-autumn#29 — that PR's merge decision is dotsimulate's; this only lands the fix on our fork's own branches.

Test plan

• python -m py_compile on both changed files
• pytest tests/unit/test_engine_path_length.py -v — 3 passed (validated against this branch's src/, not the stale editable install, via PYTHONPATH override)
• CI (claude-code-review) green before merge

🤖 Generated with Claude Code

[claude]

Claude finished @forkni's task in 3m 6s —— View job

---

Code Review

Tasks:

• Gather context (PR diff, CLAUDE.md files — none exist in this repo, so reviewed against repo style/domain conventions instead)
• Summarize changes
• Manual review pass (traced the new UNet hashing logic end-to-end)
• Independent bug-hunt pass (opus agent, diff-only)
• Validate flagged issues
• Post final review

Summary

get_engine_path (src/streamdiffusion/acceleration/tensorrt/engine_manager.py:216-230) now replaces the fully-spelled-out UNet prefix with base_name + fp8_tag + h{sha1[:12]} + res_tag before it's joined into the final engine path, and adds a >= 250-char warning (:238-244) as a second line of defense. tests/unit/test_engine_path_length.py adds 3 regression tests (MAX_PATH bound, determinism, no-collision) built on the crash report's exact flag combination, with the 040-only pin_cache_frames/cache_maxframes kwargs filtered against the live get_engine_path signature so the test stays portable across branches.

Findings

No issues found.

Traced through the logic in detail:

• canonical = prefix captures the prefix after every build-flag suffix (fp8, cached_attn, FI, controlnet, static-batch, optlvl, mode, TRT version/cc, resolution) has already been appended, so sha1(canonical)[:12] is derived from the full, unabbreviated config — collision-safety is preserved even though the human-readable prefix is discarded.
• The re-appended fp8_tag/res_tag in the new short prefix are cosmetic only; uniqueness is carried entirely by the hash, so there's no double-counting risk.
• resolution[0]/[1] access is guarded by resolution is not None; base_name, fp8, resolution are all in scope from the outer function; hashlib was already imported, cast was correctly added to the typing import.
• The cast(str, ...) on filename is a type-checker-only change (per the inline comment, _configs[engine_type]["filename"] is always a str at runtime) — no behavior change.
• New/changed lines stay within the 119-char line-length convention (the two pre-existing >119-char lines in this file are untouched by this diff).
• No hot-path/tensor, hardcoded-device, or TensorRT buffer-safety concerns — this is pure string/path logic in the cache-key builder, not the render loop.
• An independent bug-hunt pass (separate agent, diff-only) reached the same conclusion.

One thing worth calling out for the PR description/reviewers (not a defect, just an expected consequence): since the UNet cache-key directory name changes from the verbose form to the hashed form, previously-built UNet engines will be cache-missed and rebuilt once on first run after this lands — expected and unavoidable for this kind of fix, not something to change.