chore: reenable py313

[zewenli98]

Description

Reenabled Python 3.13 builds since TensorRT has 3.13 wheels available now for TensorRT 10.9.

Type of change

• New feature (non-breaking change which adds functionality)

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

[pytorch-bot]

No ciflow labels are configured for this repo.
For information on how to enable CIFlow bot see this wiki

[HolyWu]

I think you don't need to install rust. Instead, you should try to bump the version of transformers in tests/py/requirements.txt. transformers 4.40.2 depended on tokenizers 0.19.x, but tokenizers 0.19.x didn't have wheels for cp313. That's why pip downloaded the source archive of tokenizers and then compiled unsuccessfully.

[zewenli98]

The updated transformers doesn't match the results of BERT, compared with old transformers and TRT, so we fixed the version here. Any idea?

[HolyWu]

The test_bert_base_uncased in test_models and test_models_export was flawed. The inputs were defined as torch.randint(0, 1, ...), which returned a meaningless tensor filled with all zeros because the value of the second arg high is exclusive. To get a useful tensor the high arg should be at least 2, like https://pytorch.org/TensorRT/tutorials/_rendered_examples/dynamo/torch_compile_transformers_example.html.

[peri044]

LGTM

[peri044]

modelopt doesn't support 3.13 yet ? So, quantization also won't work right in Python 3.13?

[zewenli98]

right, I think we've already directly skip quantization tests like

TensorRT/tests/py/dynamo/models/test_models_export.py

Lines 201 to 220 in d3eb817

	@unittest.skipIf(
	not importlib.util.find_spec("modelopt"),
	"ModelOpt is required to run this test",
	)
	@pytest.mark.unit
	def test_base_fp8(ir):
	import modelopt.torch.quantization as mtq
	from modelopt.torch.quantization.utils import export_torch_mode
	class SimpleNetwork(torch.nn.Module):
	def __init__(self):
	super(SimpleNetwork, self).__init__()
	self.linear1 = torch.nn.Linear(in_features=10, out_features=5)
	self.linear2 = torch.nn.Linear(in_features=5, out_features=1)
	def forward(self, x):
	x = self.linear1(x)
	x = torch.nn.ReLU()(x)
	x = self.linear2(x)
	return x

Users may not be aware of it though

[narendasan]

Maybe we can make this a bit more generic (like for any feature in the FeatureSet)

[narendasan]

Think it mostly looks good, I think we should think about making the requires decorator a bit more generic and unify needs_torch_tensorrt_runtime and needs_refit, maybe something like needs_feature. Maybe something for after 2.7

[zewenli98]

Think it mostly looks good, I think we should think about making the requires decorator a bit more generic and unify needs_torch_tensorrt_runtime and needs_refit, maybe something like needs_feature. Maybe something for after 2.7

sure, will think about it for the next release.