[fsdp] fix: qwen3vlmoe with Monkey patch to fix a bug in transformers 4.57.x (verl-project#4402)

fix qwen3vlmoe with Monkey patch to fix a bug in transformers 4.57.3
where `Qwen3VLMoeTextSparseMoeBlock.forward`

incorrectly uses torch.zeros_like(router_logits) instead of
torch.zeros_like(hidden_states)
    when creating router_weights (line 148 in modeling_qwen3_vl_moe.py).

to close issue: verl-project#4401 RuntimeError: scatter(): Expected self.dtype to be
equal to src.dtype
### 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

- [x] Search for similar PRs. Paste at least one query link here: ...
verl-project#3686 huggingface/transformers#41418
- [ ] 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: qian.cheng <mgr.qiancheng@gmail.com>
Co-authored-by: wuxibin <wuxibin@bytedance.com>

Author: 3 people

verl/models/transformers/monkey_patch.py

@@ -356,14 +356,22 @@ def state_dict(self, *args, **kwargs):
             Qwen3VLMoeTextModel,
         )
 
-        from verl.models.transformers.qwen3_vl import forward_with_normal_backend, qwen3_vl_base_forward
+        from verl.models.transformers.qwen3_vl import (
+            forward_with_normal_backend,
+            patch_qwen3_vl_moe_sparse_moe_block_forward,
+            qwen3_vl_base_forward,
+        )
 
         Qwen3VLModel.forward = qwen3_vl_base_forward
         Qwen3VLMoeModel.forward = qwen3_vl_base_forward
         Qwen3VLForConditionalGeneration.forward = forward_with_normal_backend
         Qwen3VLMoeForConditionalGeneration.forward = forward_with_normal_backend
         print(f"Monkey patch {model.__class__.__name__} model forward")
 
+        # Step 1.5: patch Qwen3VLMoeTextSparseMoeBlock to fix transformers 4.57.3 bug
+        if model.config.model_type == "qwen3_vl_moe" and is_transformers_version_in_range(max_version="4.57.3"):
+            patch_qwen3_vl_moe_sparse_moe_block_forward()
+
         # Step 2: patch input for multimodal sequence parallelism
         if ulysses_sp_size > 1:
             patch_vlm_for_ulysses_input_slicing(Qwen3VLTextModel)

verl/models/transformers/qwen3_vl.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import functools
 import logging
 import os
 from dataclasses import dataclass
@@ -334,3 +335,41 @@ def forward_with_triton_backend(
         entropy=entropy,
         hidden_states=outputs.hidden_states,
     )
+
+
+def patch_qwen3_vl_moe_sparse_moe_block_forward():
+    """
+    Monkey patch to fix a bug in transformers 4.57.3 where Qwen3VLMoeTextSparseMoeBlock.forward
+    incorrectly uses torch.zeros_like(hidden_states) instead of torch.zeros_like(router_logits)
+    when creating router_weights (line 148 in modeling_qwen3_vl_moe.py).
+    
+    This is a minimal fix that only changes the problematic line while keeping the rest of the
+    original implementation intact.
+    """
+    try:
+        from transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe import Qwen3VLMoeTextSparseMoeBlock
+    except ImportError:
+        # Model not available, skip patching
+        return
+
+    # Store the original forward method for reference
+    original_forward = Qwen3VLMoeTextSparseMoeBlock.forward
+
+    @functools.wraps(original_forward)
+    def patched_forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size = hidden_states.shape[0]
+        hidden_states = hidden_states.reshape(-1, self.hidden_size)
+        router_logits = self.gate(hidden_states)
+        routing_weights = torch.nn.functional.softmax(router_logits, dim=-1, dtype=torch.float)
+        routing_weights, router_indices = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True)
+        # BUG FIX: Original code incorrectly uses hidden_states here, should use router_logits
+        routing_weights = routing_weights.to(router_logits.dtype)
+        router_weights = torch.zeros_like(router_logits).scatter_(1, router_indices, routing_weights)
+        hidden_states = hidden_states.reshape(batch_size, -1, self.hidden_size)
+        routed_out = self.experts(hidden_states, router_weights, router_indices)
+        return routed_out
+
+    # Apply the patch
+    Qwen3VLMoeTextSparseMoeBlock.forward = patched_forward
+    logger.info("Monkey patched Qwen3VLMoeTextSparseMoeBlock.forward to fix router_weights bug")