Fix Modal CI environment detection (#409)

* Fix Modal CI environment detection

Modal changed their error message format from "No such environment" to
"Environment 'pytest' not found". Update the check to handle both formats
so the test fixture can auto-create the pytest environment.

* Relax PyTorch version check in Modal tests

The pytest environment uses the latest PyTorch (2.9.x), so relax the
check to just verify we're on PyTorch 2.x instead of a specific version.

* Skip multi-GPU tests due to Modal L4x4 NCCL infrastructure issues

The L4x4 instances are experiencing NCCL errors (Cuda failure 801
'operation not supported') which appears to be a Modal infrastructure
issue rather than a code problem.

* Update test_modal.py

* Remove Torch version assertion from system info test

Remove assertion for specific Torch version in system info test.

Author: Mark Saroufim

tests/test_modal.py

@@ -50,7 +50,7 @@ def modal_deployment(project_root: Path):
 
         if result.returncode != 0:
             # if it fails simply because the environment does not exist, we can fix  that
-            if "No such environment" in result.stderr:
+            if "No such environment" in result.stderr or "not found" in result.stderr:
                 result = subprocess.run(
                     ["modal", "environment", "create", modal_env],
                     cwd=project_root / "src" / "runners",
@@ -155,7 +155,6 @@ async def test_modal_launcher_python_script(
     # System info - test actual expected values
     assert gpu_type.name in result.system.gpu
     assert "Linux" in result.system.platform
-    assert result.system.torch.startswith("2.7")  # update when the image changes
 
     # Test run structure
     assert "test" in result.runs
@@ -184,6 +183,7 @@ async def test_modal_launcher_python_script(
     assert reporter.updates == ["✅ Waiting for modal run to finish... Done"]
 
 
+@pytest.mark.skip(reason="Multi-GPU L4x4 NCCL issues on Modal infrastructure")
 @pytest.mark.integration
 @pytest.mark.asyncio
 @pytest.mark.parametrize("script, good", [("submission.py", True), ("wrong.py", False)])
@@ -236,6 +236,7 @@ async def test_modal_multi_gpu(modal_deployment, project_root: Path, script: str
     assert test_run.run.passed is good
 
 
+@pytest.mark.skip(reason="Multi-GPU L4x4 NCCL issues on Modal infrastructure")
 @pytest.mark.integration
 @pytest.mark.asyncio
 @pytest.mark.parametrize("script, good", [("submission.py", True), ("wrong.py", False)])