delete huge objectref manually after ray.get()

xtuner/v1/ray/base/accelerator.py

@@ -196,6 +196,25 @@ def device_visible_env_name(self):
         else:
             raise ValueError(f"Unsupported accelerator type: {self.accelerator}")
 
+    def get_logical_local_rank(self) -> int:
+        """Resolve the assigned accelerator id to the logical local rank.
+
+        Ray reports accelerator ids in the physical numbering space. Torch selects devices from the current visible-
+        device list, which is indexed logically from zero after applying visibility masks.
+        """
+        accelerator_id = str(ray.get_runtime_context().get_accelerator_ids()[self.accelerator][0])
+        visible_devices = os.environ.get(self.device_visible_env_name)
+        if visible_devices is None:
+            return int(accelerator_id)
+
+        visible_device_ids = [device_id.strip() for device_id in visible_devices.split(",") if device_id.strip()]
+        if accelerator_id not in visible_device_ids:
+            raise ValueError(
+                f"Assigned accelerator id {accelerator_id} is not present in "
+                f"{self.device_visible_env_name}={visible_devices}."
+            )
+        return visible_device_ids.index(accelerator_id)
+
     def setup_distributed(self, rank: int, master_addr: str, master_port: int, world_size: int):
         """Set up the distributed environment for the worker.
 
@@ -215,8 +234,7 @@ def setup_distributed(self, rank: int, master_addr: str, master_port: int, world
         os.environ["MASTER_PORT"] = str(master_port)
         os.environ["RANK"] = str(rank)
         os.environ["WORLD_SIZE"] = str(world_size)
-        os.environ["LOCAL_RANK"] = str(ray.get_runtime_context().get_accelerator_ids()[self.accelerator][0])
-
+        os.environ["LOCAL_RANK"] = str(self.get_logical_local_rank())
         # backend 参数是指定通信后端，不是从环境变量获取
         # - 'nccl': NVIDIA GPU 间通信（推荐用于 GPU）
         # - 'gloo': CPU 通信或跨平台

xtuner/v1/ray/environment/single_turn_env.py

@@ -94,14 +94,14 @@ async def generate(  # type: ignore[override]
         if self.rollout_controller:
             response_future = []
             for sample in group_data_items:
-                sample.data.extra_info["root_id"] = sample.uid.root_id
-                sample.data.extra_info["action_id"] = sample.uid.action_id
+                rollout_extra_info = dict(sample.data.extra_info)
+                rollout_extra_info["root_id"] = sample.uid.root_id
+                rollout_extra_info["action_id"] = sample.uid.action_id
                 update_sample_params = sample_params
 
                 if "partial_rollout_input_ids" in sample.env.rollout.extra_info:
                     input_ids_length = len(sample.data.input_ids) if sample.data.input_ids is not None else 0
                     current_partial_length = len(sample.env.rollout.extra_info["partial_rollout_input_ids"])
-                    rollout_extra_info = copy.deepcopy(sample.data.extra_info)
                     rollout_extra_info["partial_rollout_input_ids"] = sample.env.rollout.extra_info[
                         "partial_rollout_input_ids"
                     ]
@@ -113,8 +113,6 @@ async def generate(  # type: ignore[override]
                     self.logger.debug(
                         f"root_id: {sample.uid.root_id}, action_id {sample.uid.action_id} pass current_partial_length {current_partial_length}, input_ids_length {input_ids_length} to rollout and set max_tokens to {update_sample_params.max_tokens}"
                     )
-                else:
-                    rollout_extra_info = sample.data.extra_info
 
                 if "routed_experts" in sample.env.rollout.extra_info:
                     rollout_extra_info["routed_experts"] = sample.env.rollout.extra_info["routed_experts"]
@@ -126,6 +124,8 @@ async def generate(  # type: ignore[override]
                     extra_params=extra_params,
                     extra_info=rollout_extra_info,
                 )
+                del rollout_extra_info
+
                 response_future.append(fut)
             try:
                 rollout_responses = await asyncio.wait_for(

xtuner/v1/ray/rollout/worker.py

@@ -573,6 +573,7 @@ async def _handle_non_stream_response(
 
                             data = base64.b64decode(routed_experts)
                             routed_experts = ray.cloudpickle.loads(data)
+                            del data
                         else:
                             routed_experts = torch.tensor(routed_experts)  # n,layer,expert
                             routed_experts = ray.put(routed_experts)
@@ -586,13 +587,14 @@ async def _handle_non_stream_response(
                             routed_experts = ray.cloudpickle.loads(data)
                             cur_routed_experts = await routed_experts  # n,layer,expert
                             ray.internal.free(routed_experts, local_only=False)
+                            del data
                         else:
                             routed_experts = torch.tensor(routed_experts)  # n,layer,expert
                             cur_routed_experts = routed_experts
 
                         history_routed_experts = await input_extra_info["routed_experts"]  # n, layer, expert
                         ray.internal.free(input_extra_info["routed_experts"], local_only=False)
-                        del input_extra_info["routed_experts"]
+                        del input_extra_info
 
                         assert (history_routed_experts.shape[0] - 1) > 0 and history_routed_experts.shape[
                             0
@@ -613,6 +615,8 @@ async def _handle_non_stream_response(
                             f"Experts(exist={history_routed_experts.shape}, init_cur={init_cur_roued_experts}, cur={cur_routed_experts.shape}, concat={concat_routed_experts.shape})"
                         )
                         extra_info["routed_experts"] = ray.put(concat_routed_experts)
+                        del history_routed_experts
+                        del cur_routed_experts
                     else:
                         assert finish_reason == "abort", (
                             f"routed_experts is None, but finish_reason is {finish_reason}, expected abort. response: {response}"

xtuner/v1/ray/utils.py

@@ -5,6 +5,7 @@
 from typing import TYPE_CHECKING, Callable, Coroutine, List, Optional, cast
 
 import ray
+from ray import ObjectRef
 
 
 if TYPE_CHECKING:
@@ -208,3 +209,13 @@ def create_task(
     for callback in done_callbacks:
         task.add_done_callback(callback)
     return task
+
+
+def free_object_refs(refs: List[ObjectRef]) -> None:
+    valid_refs = [ref for ref in refs if isinstance(ref, ObjectRef)]
+    if not valid_refs:
+        return
+    try:
+        ray._private.internal_api.free(valid_refs, local_only=False)
+    except Exception:
+        ray.internal.free(valid_refs, local_only=False)

xtuner/v1/rl/base/controller.py

@@ -260,7 +260,10 @@ def fit(self, data_batches: list[ColateItem], pack_max_length: int, rollout_idx:
                     rollout_idx=rollout_idx,
                 )
             )
-        log_infos = ray.get(handles, timeout=TRAIN_RAY_GET_TIMEOUT)
+        try:
+            log_infos = ray.get(handles, timeout=TRAIN_RAY_GET_TIMEOUT)
+        finally:
+            del packed_data_batches
         return log_infos
 
     @ray_method

xtuner/v1/rl/base/worker.py

@@ -35,6 +35,7 @@
 from xtuner.v1.model.utils.misc import ModelForwardExtraLogInfo
 from xtuner.v1.ray.base import SingleAcceleratorWorker
 from xtuner.v1.ray.config import RolloutConfig
+from xtuner.v1.ray.utils import free_object_refs
 from xtuner.v1.rl.base.loss import BaseRLLossContext
 from xtuner.v1.rl.utils import gather_logprobs
 from xtuner.v1.train.trainer import LoadCheckpointConfig
@@ -483,8 +484,12 @@ def fit(self, data_batches: list[WorkerInputItem], rollout_idx: int) -> WorkerLo
                     assert isinstance(pixel_values, list), (
                         f"pixel_values should be list of tensor, got {type(pixel_values)}"
                     )
-                    pixel_values = [ray.get(pixel_obf) for pixel_obf in pixel_values]
-                    pixel_values = torch.cat(pixel_values, dim=0)
+                    pixel_value_refs = list(pixel_values)
+                    try:
+                        pixel_values = torch.cat(ray.get(pixel_value_refs), dim=0)
+                    finally:
+                        free_object_refs(pixel_value_refs)
+
                     seq_ctx.pixel_values = pixel_values
 
             rollout_routed_experts = seq_ctx.rollout_routed_experts

xtuner/v1/train/rl_trainer.py

@@ -118,7 +118,7 @@ def get_train_seq_ctx(
 ):
     seq_ctx = SequenceContext.from_input_ids((input_ids,), device="cpu")
     if multimodal_train_info and len(multimodal_train_info) > 0:
-        position_ids = multimodal_train_info.get("position_ids")  # (1,n) or (3,1,n)
+        position_ids = multimodal_train_info.pop("position_ids")  # (1,n) or (3,1,n)
         if position_ids is not None and len(position_ids.shape) == 3:
             # qwen3vl 需要特殊处理，其余的不需要额外处理
             max_value = position_ids.max(dim=-1).values  # (3,1)
@@ -128,8 +128,9 @@ def get_train_seq_ctx(
             position_ids = torch.cat([position_ids, response_position_ids], dim=-1)
             seq_ctx.position_ids = position_ids  # type: ignore[assignment]
             assert position_ids.size(-1) == input_ids.size(-1)
-        seq_ctx.pixel_values = multimodal_train_info.get("pixel_values")
-        seq_ctx.image_grid_thw = multimodal_train_info.get("image_grid_thw")
+        seq_ctx.pixel_values = multimodal_train_info.pop("pixel_values")
+        seq_ctx.image_grid_thw = multimodal_train_info.pop("image_grid_thw")
+        del multimodal_train_info
     return seq_ctx
 
 
@@ -623,6 +624,8 @@ def fit(self):
                 # 1. Rollout to generate experience
                 rollout_info = self._rollout_step(rollout_idx, step_timer_dict)
 
+                train_log_info = {}
+                eval_log_info = {}
                 if not self._debug_rollout:
                     # 2. Train on the generated experience
                     train_log_info = self._train_step(