[fsdp, sglang] fix: Using Agreesive Empty Cache instead (volcengine#3136)

### What does this PR do?

> Add **concise** overview of what this PR aims to achieve or
accomplish. Reference related GitHub issues and PRs that help with the
review.

### Checklist Before Starting

- [ ] Search for similar PRs. Paste at least one query link here: ...
- [ ] Format the PR title as `[{modules}] {type}: {description}` (This
will be checked by the CI)
- `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`,
`trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`,
`ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`,
`env`, `tool`, `ckpt`, `doc`, `data`
- If this PR involves multiple modules, separate them with `,` like
`[megatron, fsdp, doc]`
  - `{type}` is in `feat`, `fix`, `refactor`, `chore`, `test`
- If this PR breaks any API (CLI arguments, config, function signature,
etc.), add `[BREAKING]` to the beginning of the title.
  - Example: `[BREAKING][fsdp, megatron] feat: dynamic batching`

### Test

> For changes that can not be tested by CI (e.g., algorithm
implementation, new model support), validate by experiment(s) and show
results like training curve plots, evaluation results, etc.

### API and Usage Example

> Demonstrate how the API changes if any, and provide usage example(s)
if possible.

```python
# Add code snippet or script demonstrating how to use this
```

### Design & Code Changes

> Demonstrate the high-level design if this PR is complex, and list the
specific changes.

### Checklist Before Submitting

> [!IMPORTANT]
> Please check all the following items before requesting a review,
otherwise the reviewer might deprioritize this PR for review.

- [ ] Read the [Contribute
Guide](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md).
- [ ] Apply [pre-commit
checks](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md#code-linting-and-formatting):
`pre-commit install && pre-commit run --all-files --show-diff-on-failure
--color=always`
- [ ] Add / Update [the
documentation](https://github.com/volcengine/verl/tree/main/docs).
- [ ] Add unit or end-to-end test(s) to [the CI
workflow](https://github.com/volcengine/verl/tree/main/.github/workflows)
to cover all the code. If not feasible, explain why: ...
- [ ] Once your PR is ready for CI, send a message in [the `ci-request`
channel](https://verl-project.slack.com/archives/C091TCESWB1) in [the
`verl` Slack
workspace](https://join.slack.com/t/verl-project/shared_invite/zt-3855yhg8g-CTkqXu~hKojPCmo7k_yXTQ).
(If not accessible, please try [the Feishu group
(飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).)

---------

Co-authored-by: zhaochenyang20 <[email protected]>

Author: 2 people

examples/grpo_trainer/run_qwen2_5_vl-7b-sglang.sh

@@ -41,6 +41,10 @@ python3 -m verl.trainer.main_ppo \
     trainer.logger='["console","wandb"]' \
     trainer.project_name='verl_grpo_example_geo3k' \
     trainer.experiment_name='qwen2_5_vl_7b_function_rm' \
+    actor_rollout_ref.ref.strategy=fsdp2 \
+    actor_rollout_ref.actor.strategy=fsdp2 \
+    critic.strategy=fsdp2 \
+    reward_model.strategy=fsdp2 \
     trainer.n_gpus_per_node=8 \
     trainer.nnodes=1 \
     trainer.save_freq=20 \

examples/sglang_multiturn/run_qwen2.5-3b_gsm8k_multiturn.sh

@@ -55,6 +55,10 @@ python3 -m verl.trainer.main_ppo \
     trainer.save_freq=-1 \
     trainer.test_freq=20 \
     trainer.val_before_train=True \
+    actor_rollout_ref.ref.strategy=fsdp2 \
+    actor_rollout_ref.actor.strategy=fsdp2 \
+    critic.strategy=fsdp2 \
+    reward_model.strategy=fsdp2 \
     data.train_files=$HOME/data/gsm8k/train.parquet \
     data.val_files=$HOME/data/gsm8k/test.parquet \
     actor_rollout_ref.rollout.multi_turn.tool_config_path="$PROJECT_DIR/examples/sglang_multiturn/config/tool_config/gsm8k_tool_config.yaml" \

verl/workers/sharding_manager/fsdp_sglang.py

@@ -28,6 +28,7 @@
 from verl.protocol import all_gather_data_proto
 from verl.utils.device import get_device_id, get_torch_device, set_expandable_segments
 from verl.utils.fsdp_utils import fsdp_version, load_fsdp_model_to_gpu, offload_fsdp_model_to_cpu
+from verl.utils.memory_utils import aggressive_empty_cache
 from verl.utils.model import convert_weight_keys
 from verl.utils.profiler import GPUMemoryLogger, log_gpu_memory_usage, simple_timer
 from verl.utils.torch_functional import check_device_is_available
@@ -124,7 +125,7 @@ async def release_memory(self):
 
     @GPUMemoryLogger(role="FSDPSGLangShardingManager enter", logger=logger)
     async def wake_up(self):
-        get_torch_device().empty_cache()
+        aggressive_empty_cache(force_sync=True)
 
         log_gpu_memory_usage("Before state_dict() in sharding manager memory", logger=logger)
         if self.offload_param:
@@ -146,6 +147,8 @@ async def wake_up(self):
 
         # sglang need to set _set_allocator_settings to False
         logger.debug("fsdp sglang sharding_manager _set_allocator_settings to False")
+        # Note(chenyang): SGLang is using torch memory pool to manage memory
+        # which is incompatible with expandable segments
         set_expandable_segments(False)
 
         if self.device_mesh["infer_tp"].get_local_rank() == 0 and self.rollout_config.free_cache_engine:
@@ -161,7 +164,7 @@ async def wake_up(self):
         log_gpu_memory_usage("After sync model weights in sharding manager", logger=logger)
 
         del params
-        get_torch_device().empty_cache()
+        aggressive_empty_cache(force_sync=True)
         log_gpu_memory_usage("After del state_dict and empty_cache in sharding manager", logger=logger)
 
         if (
@@ -187,7 +190,7 @@ async def sleep(self):
         self.module.train()
 
         # add empty cache after each compute
-        get_torch_device().empty_cache()
+        aggressive_empty_cache(force_sync=True)
 
         # always set _set_allocator_settings to True when using sglang
         # it is required by fsdp2 to avoid oom

verl/workers/sharding_manager/megatron_sglang.py

@@ -34,6 +34,7 @@
     offload_megatron_model_to_cpu,
     per_tensor_generator,
 )
+from verl.utils.memory_utils import aggressive_empty_cache
 from verl.utils.profiler import GPUMemoryLogger, log_gpu_memory_usage, simple_timer
 from verl.workers.rollout.sglang_rollout.utils import get_named_tensor_buckets
 
@@ -163,6 +164,8 @@ async def release_memory(self):
 
     @GPUMemoryLogger(role="MegatronSGLangShardingManager enter", logger=logger)
     async def wake_up(self):
+        aggressive_empty_cache(force_sync=True)
+
         if self.offload_param:
             load_megatron_model_to_gpu(self.actor_module, load_grad=False)
         if self.bridge is not None:
@@ -178,7 +181,7 @@ async def wake_up(self):
         await self.update_weights(per_tensor_param)
         if self.offload_param:
             offload_megatron_model_to_cpu(self.actor_module)
-        get_torch_device().empty_cache()
+        aggressive_empty_cache(force_sync=True)
         # important: need to manually set the random states of each tp to be identical.
         if self.device_mesh is not None:
             self.torch_random_states = get_torch_device().get_rng_state()
@@ -194,7 +197,7 @@ async def sleep(self):
         for model in self.actor_module:
             model.train()
         # add empty cache after each compute
-        get_torch_device().empty_cache()
+        aggressive_empty_cache(force_sync=True)
 
         # restore random states
         if self.device_mesh is not None: