[Qwen3VL] fix device mismatch error for FSDP2 training #41536
Conversation
|
[For maintainers] Suggested jobs to run (before merge) run-slow: qwen3_omni_moe, qwen3_vl, qwen3_vl_moe |
For FSDP2, parameters might be on a meta device, and the weight.device attribute may
not accurately reflect where the actual computation will happen during forward passes.
```log
File "transformers/models/qwen3_vl_moe/modeling_qwen3_vl_moe.py", line 776, in forward
pos_embeds = self.fast_pos_embed_interpolate(grid_thw)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "transformers/models/qwen3_vl_moe/modeling_qwen3_vl_moe.py", line 745, in fast_pos_embed_interpolate
pos_embeds = self.pos_embed(idx_tensor) * weight_tensor[:, :, None]
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "torch/nn/modules/module.py", line 1879, in _call_impl
return inner()
^^^^^^^
File "torch/nn/modules/module.py", line 1827, in inner
result = forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "torch/nn/modules/sparse.py", line 192, in forward
return F.embedding(
^^^^^^^^^^^^
File "torch/nn/functional.py", line 2546, in embedding
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Expected all tensors to be on the same device, but got index is on cpu, different from other tensors on cuda:0 (when checking argument in method wrapper_CUDA__index_select)
```
volcengine/verl#3686 (comment)
Signed-off-by: Hollow Man <[email protected]>
| ) | ||
| idx_tensor = torch.tensor(idx_list, dtype=torch.long, device=device) | ||
| weight_tensor = torch.tensor(weight_list, dtype=self.pos_embed.weight.dtype, device=device) | ||
| pos_embeds = self.pos_embed(idx_tensor) * weight_tensor[:, :, None] |
There was a problem hiding this comment.
i think it is possible that the embedded positions and the weight tensors are on different devices, if the input grid was not on the same device as the positional embedding weight
There was a problem hiding this comment.
Yeah, since the device for grid_thw is pretty much dependent on the user-side implementation (passed as a parameter for the forward method), I think it's better to take the device of grid_thw for unifying the device of idx_tensor and weight_tensor, so that user-side implementation can have more control over this and to guarantee nothing can go wrong here. User-side code can ensure the input grid is on the same device as the positional embedding weight, as the size of grid_thw is small, so there shouldn't be too much overhead. This is tested on verl and worked fine.
2 participants
What does this PR do?
For FSDP2, parameters might be on a meta device, and the weight.device attribute may not accurately reflect where the actual computation will happen during forward passes.
volcengine/verl#3686 (comment)
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.
@yonigozlan @molbap @ArthurZucker @Cyrilvallez @zucchini-nlp