fix: validate cached ONNX before reuse to prevent empty-graph TensorRT build crash

[forkni]

Summary

• Fixes a crash during TensorRT engine compilation: Exception: Acceleration has failed: 'NoneType' object has no attribute 'graph'
• Root cause: an interrupted/failed ONNX export (OOM, process kill, crash mid-write) can leave a syntactically-valid-but-empty protobuf (0 nodes, no/invalid opset) at the cached onnx_path. The export cache check in EngineBuilder.build() was existence-only (os.path.exists(onnx_path)), so the corrupt file was silently reused instead of being re-exported.
• That empty graph flows into Optimizer.fold_constants() → polygraphy fold_constants, whose ORT symbolic shape inference refuses opset < 7 and returns None; polygraphy then calls get_num_nodes(None), raising the opaque NoneType error deep in a third-party dependency.
• Confirmed via runtime log from an affected machine: Found cached model: ...vae_decoder.engine.onnx immediately followed by VAE decoder: original .. 0 nodes, 0 tensors, 0 inputs, 0 outputs.

Changes

• src/streamdiffusion/acceleration/tensorrt/builder.py: adds _onnx_cache_valid(path) — checks file size > 0, ≥1 graph node, opset ≥ 7 (loaded structure-only, no external weight data, so it's cheap even for >2GB SDXL models). The export cache-hit check now requires this in addition to os.path.exists; on a stale/corrupt hit it logs a warning, deletes the bad file, and falls through to re-export.
• src/streamdiffusion/acceleration/tensorrt/models/models.py: BaseModel.optimize() now fails fast with a clear RuntimeError if the incoming graph has 0 nodes, instead of letting it reach polygraphy and surface as an unreadable NoneType crash.
• Generic to every engine type: all five entry points (compile_vae_encoder/decoder/safety_checker/unet/controlnet) route through EngineBuilder.build(), and every model subclasses BaseModel.

Test plan

• Both files pass ast.parse syntax check
• Verified via code-search that no equivalent ONNX-content validation existed anywhere in the codebase prior to this change
• Confirmed all 5 compile_* entry points route through EngineBuilder.build() where the new guard lives
• End-to-end: on a machine with the CUDA/TensorRT stack, truncate a cached engine's .onnx to an empty-but-valid graph and confirm the build now logs Cached ONNX invalid, re-exporting and completes successfully (not runnable in this dev environment — no GPU/TensorRT here)

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.