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

[forkni]

Summary

A user hit a crash during TensorRT engine build:

AttributeError: 'NoneType' object has no attribute 'graph'
  ... polygraphy/backend/onnx/util.py, in get_num_nodes
      return _get_num_graph_nodes(model.graph)
Exception: Acceleration has failed: 'NoneType' object has no attribute 'graph'

Root cause: a prior build was interrupted mid-torch.onnx.export() (OOM, crash, manual
kill). torch.onnx.export() writes directly to the final onnx_path with no temp-file-then-
rename, so the interrupted run left a syntactically-valid-but-empty (or truncated) protobuf on
disk. EngineBuilder.build()'s cache check is existence-only (os.path.exists(onnx_path)), so
the next run silently reused the corrupt file instead of re-exporting. That empty/0-node graph
then reached Optimizer.fold_constants() → polygraphy's fold_constants → ONNX Runtime's
symbolic shape inference, which refuses opset < 7 graphs and returns None — crashing
get_num_nodes(None) with the opaque AttributeError above, several layers away from the
actual cause.

Fix (two layers, so every model type is covered)

• builder.py: add _onnx_cache_valid(path) — loads the ONNX structure-only (no external
  weight data, so it's cheap even for large models), and rejects it if the file is empty, has
  0 graph nodes, or has no opset ≥ 7. Gate the export cache-hit on this check instead of on
  os.path.exists() alone; if a cached file fails validation, delete it and re-export instead
  of silently reusing (or crashing on) garbage.
• models.py: guard Optimizer.fold_constants() itself against a 0-node graph, rather than
  only guarding BaseModel.optimize(). CLIP.optimize() overrides BaseModel.optimize()
  entirely and calls fold_constants() directly, so a guard placed only in
  BaseModel.optimize() wouldn't catch a corrupt CLIP ONNX. Placing it in fold_constants()
  covers CLIP, UNet, VAE encoder, and VAE decoder from the one shared call site.

With this fix, an interrupted-export leftover is detected and transparently re-exported instead
of crashing on the next run.

Test plan

• ast.parse() on both modified files (syntax check in this sandbox)
• Verified this exact fix (validated on our downstream fork forkni/StreamDiffusion,
  merged in fix: validate cached ONNX before reuse to prevent empty-graph TensorRT build crash forkni/StreamDiffusion#16 and fix: core-pipeline crash fixes + effective-python audit (9 commits) #17) resolves the reported crash
• End-to-end GPU rebuild (truncate a real cached .onnx, rerun build, confirm re-export +
  no crash) — requires a CUDA/TensorRT machine, not runnable in this environment; the
  reporting user's original repro trace is included above as evidence of the failure mode.

🤖 Generated with Claude Code