Description

Every runtime check in execute_engine says what went wrong and none says where. On a model
with 14 engines that is all the caller gets — and because a refused engine writes none of its
outputs, what usually surfaces first is an unrelated node reading one of them.

Diagnostics only, no behavior change. Each check now carries the engine name and the binding
index and name, the input-shape check reports the shape TensorRT refused next to the one the
engine declares, and both allocations are caught and re-thrown naming the engine and the size
it asked for, appending to the allocator's message rather than replacing it. TORCHTRT_CHECK
only evaluates its message on failure, so the success path is unchanged.

Before:

Expected input tensors to have device cuda, found device cpu
Expected input tensors to have type Float, found type int
Error while setting the input shape
Error while setting the tensor address for shape inputs
Failed to bind tensor address for x
Error while setting the output tensor address
The shapes of the inputs: [x] cannot be inferred. This could happen if ...
<the allocator's own error, uncaught, for the output tensors>
<the allocator's own error, uncaught, for the activation workspace>

After:

Input 0 ("x") of engine _run_on_acc_0_engine must be on a CUDA device, found device cpu
Input 0 ("x") of engine _run_on_acc_0_engine expects type Float, found type int
Error while setting the input shape for input 0 ("x") of engine _run_on_acc_0_engine. TensorRT refused shape [8, 17], the engine declares [8, 16]
Error while setting the tensor address for shape input 0 ("x") of engine _run_on_acc_0_engine
Failed to bind the tensor address for input 0 ("x") of engine _run_on_acc_0_engine
Error while setting the tensor address for output "output0" of engine _run_on_acc_0_engine
The shapes of the inputs: [x] of engine _run_on_acc_0_engine cannot be inferred. This could happen if ...
Failed to allocate output "output0" of engine _run_on_acc_0_engine with shape [8, 16] and dtype Float.
Failed to allocate the 134217728-byte activation workspace of engine _run_on_acc_0_engine.

torch.ops.tensorrt.set_logging_level(4) does not help, because the LOG_DEBUG that names the
binding runs after the check that throws. TensorRT does report the shapes on the
setInputShape path, but only to its own logger, never into the exception.

Issues

Closes #4678

Type of change

• Bug fix (non-breaking change which fixes an issue)

Testing

• adds a new unit test tests/py/dynamo/runtime/test_engine_failure_diagnostics.py
• existing unit tests pass

Checklist:

• My code follows the style guidelines of this project (You can use the linters)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas and hacks
• I have made corresponding changes to the documentation
• I have added tests to verify my fix or my feature
• New and existing unit tests pass locally with my changes
• I have added the relevant labels to my PR in so that relevant reviewers are notified