fix: fix gpt oss export + bump mbridge

3rdparty/Megatron-Bridge-workspace/setup.py

@@ -56,7 +56,7 @@
     "flash-linear-attention",
     "timm",
     "open-clip-torch>=3.2.0",
-    "mlflow>=3.5.0",
+    "mlflow>=3.9.0",
     "comet-ml>=3.50.0",
     "torch>=2.6.0",
 ]

3rdparty/Megatron-LM-workspace/setup.py

@@ -51,7 +51,7 @@
     # TODO(https://github.com/NVIDIA-NeMo/RL/issues/2111): upgrade to core_cu13 when we move to CUDA 13 base container
     "transformer-engine[pytorch,core_cu12]",
     # VCS dependency - must match pyproject.toml [tool.uv.sources]
-    "nvidia-resiliency-ext @ git+https://github.com/NVIDIA/nvidia-resiliency-ext.git@63154570cea17f8805a7fd15cc3b8cc2919ba575",
+    "nvidia-resiliency-ext @ git+https://github.com/NVIDIA/nvidia-resiliency-ext.git@15a851565a4ce846c04431ecb0cf09903ab4837e",
     "tqdm",
     "einops~=0.8",
     "tensorstore~=0.1,!=0.1.46,!=0.1.72",

nemo_rl/models/generation/vllm/vllm_backend.py

@@ -37,6 +37,20 @@
     )
 
 
+def fix_gpt_oss_export_transpose(key: str, weight: torch.Tensor) -> torch.Tensor:
+    """Apply GPT-OSS down_proj transpose fix to the weight.
+
+    This is a workaround for the issue that the down_proj layout is not the same across different frameworks.
+        - HF needs [in, out] layout.
+        - Megatron needs [in, out] layout.
+        - vLLM needs [out, in] layout.
+    See https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/3271 for more details.
+    """
+    if key.endswith("mlp.experts.down_proj"):
+        weight = weight.transpose(-2, -1).contiguous()
+    return weight
+
+
 class VllmInternalWorkerExtension:
     def init_collective(
         self,
@@ -199,20 +213,30 @@ def update_weights_via_ipc_zmq(self) -> bool:
                     shape, dtype = self.state_dict_info[key]  # pyrefly
                     if isinstance(shape, list):
                         shape = torch.Size(shape)
+
+                    # Get the weight from the buffer
                     size_in_bytes = dtype.itemsize * shape.numel()
-                    weights.append(
-                        (
-                            key,
-                            buffer[offset : offset + size_in_bytes]
-                            .view(dtype=dtype)
-                            .view(shape),
-                        )
+                    weight = (
+                        buffer[offset : offset + size_in_bytes]
+                        .view(dtype=dtype)
+                        .view(shape)
                     )
+                    # apply gpt-oss transpose fix
+                    if (
+                        "GptOssForCausalLM"
+                        in self.model_runner.vllm_config.model_config.architectures
+                    ):
+                        weight = fix_gpt_oss_export_transpose(key, weight)
+                    weights.append((key, weight))
+
+                    # Move offset to the next weight
                     aligned_size = calculate_aligned_size(size_in_bytes)
                     offset += aligned_size
+
                 assert offset == used_bytes, (
                     "Offset is not equal to used bytes, usually indicate inaccurate info like keys or cached dtype in state_dict_info"
                 )
+
                 # Load weights into the model
                 from nemo_rl.models.generation.vllm.quantization import fp8
 
@@ -276,6 +300,15 @@ def _load_model_weights(weights, model_runner):
             """
             from nemo_rl.models.generation.vllm.quantization import fp8
 
+            # apply gpt-oss transpose fix
+            if (
+                "GptOssForCausalLM"
+                in self.model_runner.vllm_config.model_config.architectures
+            ):
+                for idx, (key, weight) in enumerate(weights):
+                    weight = fix_gpt_oss_export_transpose(key, weight)
+                    weights[idx] = (key, weight)
+
             policy_weights, draft_weights = self._split_policy_and_draft_weights(
                 weights
             )

nemo_rl/models/megatron/setup.py

@@ -142,29 +142,6 @@ def destroy_parallel_state():
     except ImportError:
         pass
 
-    # Reset the third global async_calls instance in base strategy module
-    try:
-        import megatron.core.dist_checkpointing.strategies.base as base_strategy
-        from megatron.core.dist_checkpointing.strategies.async_utils import (
-            AsyncCallsQueue,
-        )
-
-        # Clean up and reset the global async_calls in base strategy
-        old_call_idx = getattr(base_strategy.async_calls, "call_idx", None)
-        num_unfinalized = base_strategy.async_calls.get_num_unfinalized_calls()
-        if num_unfinalized > 0:
-            print(
-                f"[WARNING] Resetting base strategy async_calls with {num_unfinalized} unfinalized calls"
-            )
-        try:
-            base_strategy.async_calls.close()
-        except:
-            pass
-        base_strategy.async_calls = AsyncCallsQueue()
-        print(f"[DEBUG] Reset base strategy async_calls (old call_idx: {old_call_idx})")
-    except ImportError:
-        pass
-
 
 def setup_distributed() -> None:
     """Handle NCCL settings, dtype mapping, and basic config setup."""

pyproject.toml

@@ -497,7 +497,7 @@ requires-dist = [
   "flash-linear-attention",
   "timm",
   "open-clip-torch>=3.2.0",
-  "mlflow>=3.5.0",
+  "mlflow>=3.9.0",
   "comet-ml>=3.50.0",
   "torch>=2.6.0",
 ]