Fix TensorRT engine cache shape collisions

[Aziz Muminov (AzizMuminov)]

Description

Fixes #32254.

Include graph input shapes in the TensorRT model/cache identifier so otherwise identical graphs with different static input shapes do not reuse an incompatible engine.

The fingerprint uses rank, ordered static dimensions, and a marker for dynamic dimensions. It uses methods exposed by the shared-provider bridge and deliberately ignores type/dimension denotations and symbolic dimension names. No execution kernels, graph partitioning decisions, public APIs, or provider options are changed.

Add an in-memory Add-model regression test covering 16 shapes (including the reported 3x3 versus 4x4 case), repeated-model ID stability, and annotation-only changes. Existing model-path tests remain unchanged.

Existing TensorRT cache IDs change for shaped inputs, so cached engines may be rebuilt once after adopting this change.

Validation

• Lint and git diff --check pass for the two changed C++ files.
• A local isolated C++ harness compiled the actual TRTGenerateId function and repository MurmurHash implementation with lightweight graph stand-ins. It reproduced the collision on the original code and passed 16 shape/rank separation cases, repeated-input stability, directory-independent IDs, and byte-loaded versus file-loaded ID separation with the revised code. This is not a provider build or inference test.
• The TensorRT provider and new regression test have not been built/run locally: this macOS host has no CUDA/TensorRT toolchain. TensorRT CI validation is still required.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟢 Approval recommended

The targeted cache-key correction is consistent with the reported failure and has focused regression coverage.

Pull request overview

Prevents incompatible TensorRT engine-cache reuse by incorporating graph input type and shape metadata into cache identifiers.

Changes:

• Hashes serialized input type metadata when generating TensorRT model IDs.
• Tests that identical graphs with different static shapes produce different IDs.

File summaries

File	Description
onnxruntime/core/providers/tensorrt/tensorrt_execution_provider_utils.h	Adds input type and shape metadata to cache hashing.
onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc	Adds static-shape collision regression coverage.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Aziz Muminov (AzizMuminov)]

Hi maintainers, could someone please authorize the pending fork CI runs, especially Linux TensorRT CI and Windows GPU TensorRT CI, and review this fix for #32254 when convenient? The current head is 3034618. CLA is green, but the 32 GitHub Actions workflows are still marked action_required. Local lint and isolated cache-hash checks pass; the full TensorRT build and regression test still need CI validation. Thank you!