`torch_tensorrt.dynamo.compile` silently drops settings it does not know, so a typo changes the result with no warning

Open
  • Is something wrong?
#5,040 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 3 days

Nobody has claimed this yet.

Assessment

Difficulty
3/5
Estimated time
Half a day
Newbie friendliness
65/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
python
Domain
api

Research direction

Start in torch_tensorrt.dynamo.compile, which takes **kwargs and reads only a few known names, and in the torch_tensorrt.compile wrapper that forwards to it. Run the repro with the misspelled min_blok_size and confirm no _run_on_acc submodule appears. Done when an unknown keyword raises TypeError naming it, while the deprecated names that are still accepted keep their warnings.

Written by the indexing model from the issue text.

Description

What is wrong

torch_tensorrt.dynamo.compile and torch_tensorrt.compile accept any keyword. A setting name they do not know is dropped with no error and no warning. So a misspelled setting, a setting from a newer release, or a setting that only one code path reads, simply does nothing, and the user finds out much later, if at all.

What this blocks

No policy in pytorch/executorch#24353 is blocked by this today.
Earlier, lingbot_vla_v2 lost a setting this way: the export passed, and its program failed only later, when it was loaded without PyTorch.
A fix turns such mistakes into an error at compile time.

Repro
import torch
import torch_tensorrt


class Small(torch.nn.Module):
    def forward(self, x):
        return torch.relu(x * 2 + 1)


x = torch.randn(4, 8, device="cuda")
program = torch.export.export(Small().cuda(), (x,))

# "min_blok_size" is misspelled on purpose.
trt_gm = torch_tensorrt.dynamo.compile(program, arg_inputs=[x], min_blok_size=1)
print([name for name, _ in trt_gm.named_children()])
# No error and no warning. The default min_block_size (5) is used,
# so the 3-op graph stays in PyTorch: no _run_on_acc submodule.
Actual

No message mentions min_blok_size. The only related message is about the default min_block_size:

WARNING:torch_tensorrt.dynamo._compiler:3 supported operations detected in subgraph containing 3 computational nodes. Skipping this subgraph, since min_block_size was detected to be 5
['_guards_fn']

torch_tensorrt.compile(model, ir="dynamo", inputs=[x], min_blok_size=1) prints the same 2 lines.

Expected

An unknown setting raises TypeError (or at least warns), naming the setting. The few deprecated names that are still accepted on purpose can keep their deprecation warnings.

Versions

torch 2.15.0 (CUDA 13.2), torch_tensorrt 2.15.0, TensorRT 11.4.0.106, Linux x86_64, H100, driver 580.126.09.

What was tried

Reading compile(): it takes **kwargs and reads only debug, truncate_long_and_double, refit, make_refittable and enable_cross_compile_for_windows from it. Everything else is ignored. Related but different: #5030 (a known setting ignored) and #5031 (extra kwarg_inputs ignored).

Dominant language
Python
Stars
3k
Forks
421
Avg merge
4d 7h
Merged PRs (30d)
62

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from pytorch/TensorRT

All issues in pytorch/TensorRT

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.