[fsdp] fix: replicate ref compute_log_prob (disable calculate_entropy ...) in LoRA (#4675)

### What does this PR do?

The calculated entropy is discarded, as only `old_log_probs` is used to
create `ref_log_prob`. This is inefficient. The original logic for
`compute_ref_log_prob` for full-parameter fine-tuning does not calculate
entropy.

### Checklist Before Starting

- [X] Search for similar PRs. Paste at least one query link here: ...
- [X] 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`, `cfg`, `reward`
- 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.

- [X] Read the [Contribute
Guide](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md).
- [X] 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`
- [X] Add / Update [the
documentation](https://github.com/volcengine/verl/tree/main/docs).
- [X] 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: ...
- [X] 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).)

Signed-off-by: Hollow Man <[email protected]>

Author: HollowMan6

verl/workers/fsdp_workers.py

@@ -979,16 +979,20 @@ def compute_log_prob(self, data: DataProto):
         is_lora = data.meta_info.pop("is_lora", False)
         adapter_ctx = self.actor.actor_module.disable_adapter() if is_lora else nullcontext()
         # we should always recompute old_log_probs when it is HybridEngine
-        data.meta_info["micro_batch_size"] = self.config.rollout.log_prob_micro_batch_size_per_gpu
-        data.meta_info["max_token_len"] = self.config.rollout.log_prob_max_token_len_per_gpu
-        data.meta_info["use_dynamic_bsz"] = self.config.rollout.log_prob_use_dynamic_bsz
+        config_source = self.config.ref if is_lora else self.config.rollout
+        data.meta_info["micro_batch_size"] = config_source.log_prob_micro_batch_size_per_gpu
+        data.meta_info["max_token_len"] = config_source.log_prob_max_token_len_per_gpu
+        data.meta_info["use_dynamic_bsz"] = config_source.log_prob_use_dynamic_bsz
         data.meta_info["temperature"] = self.config.rollout.temperature
         # perform recompute log_prob
         with self.ulysses_sharding_manager:
             with adapter_ctx:
-                output, entropys = self.actor.compute_log_prob(data=data, calculate_entropy=True)
+                output, entropys = self.actor.compute_log_prob(data=data, calculate_entropy=not is_lora)
+            tensors = {"ref_log_prob": output} if is_lora else {"old_log_probs": output}
+            if not is_lora:
+                tensors["entropys"] = entropys
             output = DataProto.from_dict(
-                tensors={"old_log_probs": output, "entropys": entropys},
+                tensors=tensors,
                 meta_info={"temperature": self.config.rollout.temperature},
             )
 
@@ -1011,10 +1015,7 @@ def compute_ref_log_prob(self, data: DataProto):
         if self._is_lora:
             # if _is_lora, actor without lora applied is the ref
             data.meta_info["is_lora"] = True
-            data = self.compute_log_prob(data)
-            # this old_log_probs is in fact ref_log_prob
-            data = DataProto.from_dict(tensors={"ref_log_prob": data.batch["old_log_probs"]})
-            return data
+            return self.compute_log_prob(data)
         assert self._is_ref
         # else:
         # otherwise, the class have a standalone ref model