Fix weight sync selector for frozen speculative drafts

miles/backends/megatron_utils/update_weight/common.py

@@ -406,9 +406,20 @@ def collect_named_tensors_for_weight_transfer(
             yield name, tensor
 
 
-def begin_weight_update(rollout_engines: Sequence[ActorHandle]):
-    """Open a weight-update session on all rollout engines (restore packed weights)."""
-    ray.get([engine.begin_weight_update.remote() for engine in rollout_engines])
+def begin_weight_update(rollout_engines: Sequence[ActorHandle], selector: str = "all"):
+    """Open a weight-update session on the selected rollout engines (restore packed weights)."""
+    ray.get([engine.begin_weight_update.remote(selector=selector) for engine in rollout_engines])
+
+
+def weight_update_selector(args) -> str:
+    """Exclude the draft only when the trainer provably has no MTP block to send it."""
+    if (
+        getattr(args, "sglang_speculative_algorithm", None)
+        and not getattr(args, "mtp_num_layers", None)
+        and getattr(args, "megatron_to_hf_mode", "raw") != "bridge"
+    ):
+        return "target"
+    return "all"
 
 
 def end_weight_update(rollout_engines: Sequence[ActorHandle]):

miles/backends/megatron_utils/update_weight/update_weight_from_distributed/broadcast.py

@@ -115,6 +115,7 @@ def _update_weight_implementation(
             self.weight_version,
             self.rollout_engines,
             converted_named_tensors,
+            selector=self._weight_update_selector,
         )
         ray.get(refs)
         converted_named_tensors.clear()
@@ -258,6 +259,7 @@ def update_weights_from_distributed(
     weight_version: int,
     rollout_engines: Sequence[ActorHandle],
     converted_named_tensors: Sequence[tuple[str, torch.Tensor]],
+    selector: str = "all",
 ) -> list[ObjectRef]:
     """
     Send metadata (Ray), broadcast tensors (NCCL rank 0 → engines).
@@ -267,6 +269,7 @@ def update_weights_from_distributed(
             names=[name for name, _ in converted_named_tensors],
             dtypes=[param.dtype for _, param in converted_named_tensors],
             shapes=[param.shape for _, param in converted_named_tensors],
+            selector=selector,
             group_name=group_name,
             weight_version=str(weight_version),
         )

miles/backends/megatron_utils/update_weight/update_weight_from_distributed/mixin.py

@@ -21,6 +21,7 @@
     end_weight_update,
     get_atomic_update_groups,
     get_named_value_update_units,
+    weight_update_selector,
 )
 from ..hf_weight_iterator_base import HfWeightIteratorBase
 
@@ -306,13 +307,14 @@ def _send_one_multi_lora_adapter(self, adapter) -> None:
 
     def _pause_and_prepare_engines(self) -> None:
         """Pause rollout engines, flush cache, and open the weight-update session."""
+        self._weight_update_selector = weight_update_selector(self.args)
         if dist.get_rank() == 0:
             mode = self.args.pause_generation_mode
             ray.get([engine.pause_generation.remote(mode=mode) for engine in self.rollout_engines])
             if mode != "in_place":
                 ray.get([engine.flush_cache.remote() for engine in self.rollout_engines])
 
-            begin_weight_update(self.rollout_engines)
+            begin_weight_update(self.rollout_engines, self._weight_update_selector)
 
     def _finalize_and_resume_engines(self) -> None:
         """Close the weight-update session and resume rollout engines."""

miles/backends/megatron_utils/update_weight/update_weight_from_tensor.py

@@ -20,7 +20,7 @@
 from miles.utils.lora import LORA_ADAPTER_NAME
 
 from ..sglang import FlattenedTensorBucket, MultiprocessingSerializer
-from .common import _check_weight_sync_results, begin_weight_update, end_weight_update
+from .common import _check_weight_sync_results, begin_weight_update, end_weight_update, weight_update_selector
 from .hf_weight_iterator_base import HfWeightIteratorBase
 from .update_weight_from_distributed.broadcast import (
     connect_rollout_engines_from_distributed,
@@ -215,7 +215,7 @@ def update_weights(self) -> None:
             ray.get([engine.pause_generation.remote(mode=mode) for engine in self.rollout_engines])
             ray.get([engine.flush_cache.remote() for engine in self.rollout_engines])
             if not skip_base_sync:
-                begin_weight_update(self.rollout_engines)
+                begin_weight_update(self.rollout_engines, weight_update_selector(self.args))
         dist.barrier(group=get_gloo_group())
 
         megatron_local_weights = self.weights_getter()
@@ -269,6 +269,7 @@ def _send_base_params(self, hf_named_tensors) -> tuple[list[ObjectRef], Any]:
             ipc_engine=self._ipc_engine,
             ipc_gather_src=self._ipc_gather_src,
             ipc_gather_group=self._ipc_gather_group,
+            selector=weight_update_selector(self.args),
             weight_version=self.weight_version,
         )
         if self.use_distribute and self._is_distributed_src_rank:
@@ -278,6 +279,7 @@ def _send_base_params(self, hf_named_tensors) -> tuple[list[ObjectRef], Any]:
                 self.weight_version,
                 self.distributed_rollout_engines,
                 hf_named_tensors,
+                selector=weight_update_selector(self.args),
             )
             if refs_distributed:
                 refs = (refs or []) + refs_distributed
@@ -297,6 +299,7 @@ def _send_lora_params(self, hf_named_tensors) -> tuple[list[ObjectRef], Any]:
                 ipc_engine=self._ipc_engine,
                 ipc_gather_src=self._ipc_gather_src,
                 ipc_gather_group=self._ipc_gather_group,
+                selector=weight_update_selector(self.args),
                 lora_config=self._lora_config,
                 lora_name=LORA_ADAPTER_NAME,
                 lora_loaded=self._lora_loaded,
@@ -317,6 +320,7 @@ def _send_to_colocated_engine(
     lora_name: str | None = None,
     lora_loaded: bool = False,
     check_equal: bool = False,
+    selector: str = "all",
 ) -> tuple[list[ObjectRef], Any]:
     # Placeholder ranks (GPU slots reserved but no engine) have no gather group.
     # gather_object is only collective among group members, so we skip entirely.
@@ -394,6 +398,7 @@ def _send_to_colocated_engine(
                     "serialized_named_tensors": [tensors[i] for tensors in serialized_named_tensors],
                     "load_format": "flattened_bucket",
                     "weight_version": str(weight_version),
+                    "selector": selector,
                 }
                 refs.append(ipc_engine.update_weights_from_tensor.remote(**kwargs))
 

miles/backends/sglang_utils/sglang_engine.py

@@ -319,6 +319,7 @@ def update_weights_from_tensor(
         load_format: str | None = None,
         flush_cache: bool = False,
         weight_version: str | None = None,
+        selector: str = "all",
     ):
         """
         Update model weights from tensor data. The HTTP server will only post meta data, and the real weights will be copied directly from GPUs.
@@ -330,6 +331,7 @@ def update_weights_from_tensor(
             "serialized_named_tensors": serialized_named_tensors,
             "load_format": load_format,
             "flush_cache": flush_cache,
+            "selector": selector,
         }
         if weight_version is not None:
             payload["weight_version"] = weight_version
@@ -584,14 +586,22 @@ def destroy_weights_update_group(self, group_name):
             pass
 
     def update_weights_from_distributed(
-        self, names, dtypes, shapes, group_name, flush_cache=False, weight_version: str | None = None
+        self,
+        names,
+        dtypes,
+        shapes,
+        group_name,
+        flush_cache=False,
+        weight_version: str | None = None,
+        selector: str = "all",
     ):
         payload = {
             "names": names,
             "dtypes": [str(dtype).replace("torch.", "") for dtype in dtypes],
             "shapes": shapes,
             "group_name": group_name,
             "flush_cache": flush_cache,
+            "selector": selector,
         }
         if weight_version is not None:
             payload["weight_version"] = weight_version
@@ -613,9 +623,9 @@ def continue_generation(self):
         response.raise_for_status()
         return response
 
-    def begin_weight_update(self):
+    def begin_weight_update(self, selector: str = "all"):
         """Open a weight-update session on the engine (restores packed weights for loading)."""
-        return self._make_request("begin_weight_update", {})
+        return self._make_request("begin_weight_update", {"selector": selector})
 
     def end_weight_update(self):
         """Close the weight-update session (post-load + quant post-process on the full model)."""