Fix CI and open telemetry API usage (meta-pytorch#306)

Summary:
Pull Request resolved: meta-pytorch#306

TorchFT CI is installing the latest `opentelemetry-sdk`, which caused breakages in some of the APIs in otel.py. This PR updates `otel.py` to use the correct APIs.

This PR also fixes TorchFT lint (pyre check was failing).

This should resolve the failing TorchFT CI failures:
https://github.com/meta-pytorch/torchft/actions/runs/19983176027/job/57313269930

This is also causing downstream CI failures in in torchtitan: https://github.com/pytorch/torchtitan/actions/workflows/integration_test_8gpu_torchft.yaml

Pull Request resolved: meta-pytorch#305

Reviewed By: tianyu-l

Differential Revision: D89743414

Pulled By: H-Huang

fbshipit-source-id: c066fe535b332ba94b918f5b684595c8de8b6740

Author: H-Huang

pyproject.toml

@@ -14,9 +14,9 @@ dynamic = ["version"]
 readme = "README.md"
 dependencies = [
     "torch>=2.7",
-    "opentelemetry-exporter-otlp-proto-http>=1.37.0",
-    "opentelemetry-sdk>=1.37.0",
-    "opentelemetry-api>=1.37.0",
+    "opentelemetry-exporter-otlp-proto-http>=1.39.0",
+    "opentelemetry-sdk>=1.39.0",
+    "opentelemetry-api>=1.39.0",
 ]
 
 [project.urls]

torchft/optim.py

@@ -59,5 +59,5 @@ def param_groups(self) -> List[Dict[str, Any]]:
         return self.optim.param_groups
 
     @property
-    def state(self) -> Mapping[torch.Tensor, Any]:  # pyre-fixme[3]
+    def state(self) -> Mapping[torch.Tensor, object]:
         return self.optim.state

torchft/otel.py

@@ -8,37 +8,53 @@
 import logging
 import os
 import time
-from typing import List, Sequence
+from typing import Any, List, Sequence, TYPE_CHECKING
 
 from opentelemetry._logs import set_logger_provider
 
 from opentelemetry.exporter.otlp.proto.http._log_exporter import OTLPLogExporter
 from opentelemetry.sdk._logs import LoggerProvider, LoggingHandler
-from opentelemetry.sdk._logs._internal import LogData
-from opentelemetry.sdk._logs.export import (
-    BatchLogRecordProcessor,
-    ConsoleLogExporter,
-    LogExporter,
-)
+from opentelemetry.sdk._logs.export import BatchLogRecordProcessor
 from opentelemetry.sdk.resources import Resource
 
+# These types are available in opentelemetry-sdk but Pyre's type stubs
+# don't include them. We import them at runtime and provide type aliases for
+# static type checking.
+if TYPE_CHECKING:
+    # pyre-fixme[33]: Aliasing to Any is prohibited. opentelemetry-sdk lacks type stubs.
+    ReadableLogRecord = Any
+    # pyre-fixme[33]: Aliasing to Any is prohibited. opentelemetry-sdk lacks type stubs.
+    LogRecordExporter = Any
+    # pyre-fixme[33]: Aliasing to Any is prohibited. opentelemetry-sdk lacks type stubs.
+    LogRecordExportResult = Any
+    # pyre-fixme[33]: Aliasing to Any is prohibited. opentelemetry-sdk lacks type stubs.
+    ConsoleLogRecordExporter = Any
+else:
+    from opentelemetry.sdk._logs import ReadableLogRecord
+    from opentelemetry.sdk._logs.export import (
+        ConsoleLogRecordExporter,
+        LogRecordExporter,
+        LogRecordExportResult,
+    )
+
 _LOGGER_PROVIDER: dict[str, LoggerProvider] = {}
 # Path to the file containing OTEL resource attributes
 TORCHFT_OTEL_RESOURCE_ATTRIBUTES_JSON = "TORCHFT_OTEL_RESOURCE_ATTRIBUTES_JSON"
 
 
-class TeeLogExporter(LogExporter):
+class TeeLogExporter(LogRecordExporter):
     """Exporter that writes to multiple exporters."""
 
     def __init__(
         self,
-        exporters: List[LogExporter],
+        exporters: List[LogRecordExporter],
     ) -> None:
         self._exporters = exporters
 
-    def export(self, batch: Sequence[LogData]) -> None:
+    def export(self, batch: Sequence[ReadableLogRecord]) -> LogRecordExportResult:
         for e in self._exporters:
             e.export(batch)
+        return LogRecordExportResult.SUCCESS
 
     def shutdown(self) -> None:
         for e in self._exporters:
@@ -49,8 +65,6 @@ def setup_logger(name: str) -> None:
     if os.environ.get("TORCHFT_USE_OTEL", "false") == "false":
         return
 
-    global _LOGGER_PROVIDER
-
     if name in _LOGGER_PROVIDER:
         return
 
@@ -70,7 +84,7 @@ def setup_logger(name: str) -> None:
 
     exporter = TeeLogExporter(
         exporters=[
-            ConsoleLogExporter(),
+            ConsoleLogRecordExporter(),
             OTLPLogExporter(
                 timeout=5,
             ),

torchft/process_group_test.py

@@ -79,7 +79,7 @@ def _test_pg(
     ]
     tensor_list = [torch.empty_like(input_tensor)]
 
-    def check_tensors(arg: Any) -> None:  # pyre-ignore[2]
+    def check_tensors(arg: object) -> None:
         """Recursively check tensors for expected shape and dtype."""
         if isinstance(arg, torch.Tensor):
             assert arg.dtype == dtype, f"Output dtype mismatch: {arg.dtype} != {dtype}"
@@ -740,7 +740,10 @@ def test_functional_collectives(self) -> None:
 
         self.assertEqual(pg.group_name, str(dist.get_pg_count() - 1))
 
-        self.assertIs(_resolve_process_group(pg.group_name), pg)
+        self.assertIs(
+            _resolve_process_group(pg.group_name),  # pyre-ignore[6]: GroupName vs str
+            pg,
+        )
 
         try:
             t = torch.zeros(10)