Fix torch test meshes, run torch tests in CI

[dlwh]

previously we were skipping these in CI because of slowness. But let's juts fork off yet another test to run them.

[Copilot]

Pull Request Overview

This PR adds infrastructure for running PyTorch-dependent tests in CI and ensures proper JAX mesh context management across model roundtrip tests.

• Adds a torch pytest marker to the skip_if_no_torch decorator to enable selective test execution
• Introduces use_test_mesh() context manager usage across multiple model roundtrip tests
• Creates a new GitHub Actions workflow for running torch-specific tests
• Removes a test that depends on unavailable models

Reviewed Changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/test_utils.py	Adds torch marker to skip_if_no_torch decorator for test filtering
tests/test_qwen3.py	Imports and applies use_test_mesh context manager to roundtrip test
tests/test_qwen2.py	Imports and applies use_test_mesh context manager to roundtrip test
tests/test_olmo.py	Imports and applies use_test_mesh context manager to roundtrip test
tests/test_mixtral.py	Refactors to use use_test_mesh at tempdir level and removes redundant nested context
tests/test_mistral.py	Imports and applies use_test_mesh context manager to roundtrip test
tests/test_lora.py	Removes PEFT integration test and applies use_test_mesh to remaining tests
tests/test_llama3.py	Imports and applies use_test_mesh context manager to roundtrip test
tests/test_llama.py	Imports use_test_mesh and applies it to roundtrip test
tests/test_hf_gpt2_serialize.py	Wraps model loading and testing logic in use_test_mesh context
tests/test_gemma.py	Imports and applies use_test_mesh context manager to roundtrip tests
pyproject.toml	Registers torch as a pytest marker for test filtering
.github/workflows/run_torch_tests.yaml	New workflow for running torch-specific tests in CI

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Installing PyTorch after running uv sync may cause version conflicts or reinstall dependencies. The torch version constraint (>=2.7.0) in pyproject.toml torch_test extra may not be respected. Consider including torch in the uv sync command or explicitly specifying the version here to match the constraint.

Suggested change

	run: uv pip install --index-url https://download.pytorch.org/whl/cpu torch
	run: uv pip install --index-url https://download.pytorch.org/whl/cpu "torch>=2.7.0"

[Copilot]

Pull Request Overview

Copilot reviewed 13 out of 13 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The nested marker application pytest.mark.torch(pytest.mark.skipif(...)(f)) is incorrect. Pytest markers should be composed using stacking, not nesting. The correct approach is to apply both markers separately by stacking them or using a marker list. This current implementation will result in pytest.mark.torch being called with a wrapped function, which is not the intended API usage.

Suggested change

	return pytest.mark.torch(pytest.mark.skipif(not has_torch(), reason="torch not installed")(f))
	f = pytest.mark.torch(f)
	f = pytest.mark.skipif(not has_torch(), reason="torch not installed")(f)
	return f