feat(profiler): fall back to local save_path for cloud ckpt_path; reject empty activities

Two robustness fixes from PR review:

- Profiler.__init__ detects a cloud save_path (s3://, gs://, gcs:// via the
  existing is_cloud_path helper) and falls back to ./profiler_traces with a
  warning. save_path commonly defaults to {ckpt_path}/profiler_traces, and
  ckpt_path can be a cloud URI -- which torch.profiler can't write to, so the
  trace would otherwise be silently lost.
- TorchProfilerConfig.validate rejects an empty `activities` list. It passed the
  membership check vacuously, but torch.profiler.profile(activities=[]) records
  nothing -- now it fails fast at startup.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: dyurk-lila

skyrl/backends/skyrl_train/utils/profiler.py

@@ -4,6 +4,8 @@
 import torch.distributed
 from loguru import logger
 
+from skyrl.backends.skyrl_train.utils.io.io import is_cloud_path
+
 # Map config activity strings to torch ProfilerActivity members.
 _ACTIVITY_MAP = {
     "cpu": torch.profiler.ProfilerActivity.CPU,
@@ -61,6 +63,16 @@ def __init__(self, config, default_save_path: str = None):
             return
         self.config = config
         self.save_path = config.save_path or default_save_path or "./profiler_traces"
+        # torch.profiler writes traces with the local filesystem only. ``save_path``
+        # commonly defaults to ``{ckpt_path}/profiler_traces``, and ckpt_path can be a
+        # cloud URI (s3://, gs://) -- which torch can't write to, so the trace would be
+        # silently lost. Fall back to a local dir so profiling still produces output.
+        if is_cloud_path(self.save_path):
+            logger.warning(
+                f"[Profiler] cloud save_path {self.save_path!r} is not writable by torch.profiler; "
+                f"falling back to local './profiler_traces'."
+            )
+            self.save_path = "./profiler_traces"
         self.ranks = list(config.ranks)
         self.export_type = getattr(config, "export_type", "chrome_trace")
         self.rank = torch.distributed.get_rank() if torch.distributed.is_initialized() else 0

skyrl/train/config/config.py

@@ -216,6 +216,10 @@ def validate(self) -> None:
             return
         if not self.ranks:
             raise ValueError("`torch_profiler_config.ranks` must be non-empty when profiling is enabled.")
+        # An empty `activities` passes the membership check below vacuously, but
+        # `torch.profiler.profile(activities=[])` records nothing -- fail fast instead.
+        if not self.activities:
+            raise ValueError("`torch_profiler_config.activities` must be non-empty when profiling is enabled.")
         bad_activities = [a for a in self.activities if a.lower() not in TORCH_PROFILER_ACTIVITIES]
         if bad_activities:
             raise ValueError(

tests/backends/skyrl_train/utils/test_profiler.py

@@ -108,6 +108,14 @@ def test_default_save_path_used_when_none(tmp_path):
     assert prof.save_path == default
 
 
+def test_cloud_save_path_falls_back_to_local(tmp_path):
+    # ckpt_path (and thus the derived default save_path) can be a cloud URI, which
+    # torch.profiler can't write to. The wrapper must fall back to a local dir so
+    # the trace isn't silently lost.
+    prof = Profiler(_ProfCfg(save_path="s3://my-bucket/run/profiler_traces"), default_save_path=str(tmp_path))
+    assert prof.save_path == "./profiler_traces"
+
+
 def test_kernel_summary_none_when_disabled(tmp_path):
     prof = Profiler(_ProfCfg(enable=False), default_save_path=str(tmp_path))
     assert prof.get_kernel_summary() is None

tests/train/test_config.py

@@ -287,6 +287,12 @@ def test_unknown_activity_rejected(self):
         with pytest.raises(ValueError, match=r"activities"):
             self._cfg(activities=["cpu", "gpu"]).validate()
 
+    def test_empty_activities_rejected(self):
+        # An empty list profiles nothing; the membership check would pass it
+        # vacuously, so it needs its own guard.
+        with pytest.raises(ValueError, match=r"activities.*non-empty"):
+            self._cfg(activities=[]).validate()
+
     def test_activities_case_insensitive(self):
         self._cfg(activities=["CPU", "CUDA"]).validate()  # must not raise