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

[forkni]

Problem

Loading StreamDiffusionTD with sd-turbo + fp8: true + acceleration: tensorrt crashes during TensorRT engine build:

FileNotFoundError: [Errno 2] No such file or directory:
'...\stabilityai\sd-turbo--tiny_vae-True--min_batch-1--max_batch-4--use_cached_attn-False--fi-False--fp8v3--sbatch1--batch-2--cachef4--optlvl4--mode-img2img--trt10.16.1.11--cc120--res-512x512\unet.engine.onnx'
→ Exception: Acceleration has failed

Root cause

EngineManager.get_engine_path encodes every UNet build flag verbatim into the on-disk directory name. With a realistic engine_dir and a config using fp8 + static batch + pin_cache_frames + optlvl + resolution, the generated directory was 246 chars; the derived unet.engine.onnx path was 263 and unet.engine.opt.onnx was 267 — both over Windows' 260-char MAX_PATH.

Because the directory fit but the file didn't, mkdir silently succeeded while torch.onnx.export → open(onnx_path, "wb") raised a bare FileNotFoundError. wrapper.py's OOM-only fallback doesn't catch this, so it re-raises as Acceleration has failed.

Fix

For EngineType.UNET only, hash the fully-assembled flag prefix (sha1, first 12 hex chars) — reusing the existing _lora_signature hashing idiom — keeping model name, fp8, and resolution human-readable in the folder name.

• Cache-key precision is unchanged: identical configs still hash identically, differing configs still diverge (sha1, no realistic collision).
• Path drops from ~267 chars to ~132.
• VAE/ControlNet directory naming is untouched — existing VAE/ControlNet engines are not invalidated. UNet engines rebuild once under their new hashed folder name.
• Adds a belt-and-suspenders warning if a derived path would still approach MAX_PATH (e.g. an unusually deep user engine_dir), so a future added flag can't silently regress this again.

Testing

Adds tests/unit/test_engine_path_length.py, which reproduces the exact crash-log config and asserts:

• the derived .onnx / .opt.onnx paths stay under MAX_PATH (red before the fix, green after)
• the path is deterministic across repeated calls (cache stability)
• distinct configs (e.g. differing static_batch_size) still produce distinct directories (no collisions)

pytest tests/unit/test_engine_path_length.py -v

3 passed.

🤖 Generated with Claude Code