Skip to content

Correctly drop tokens in SwitchTransformer #37123

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Apr 10, 2025
Merged

Correctly drop tokens in SwitchTransformer #37123

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 2 additions & 4 deletions src/transformers/models/switch_transformers/modeling_switch_transformers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -301,10 +301,8 @@ def forward(self, hidden_states):
router_mask, router_probs, router_logits = self.router(hidden_states)
expert_index = torch.argmax(router_mask, dim=-1)

# The routers introduced might not always map all the tokens, to a router, which means that some hidden states
# can be unchanged from one layer to another. That is why the hidden states are cloned before updating only the selected ones.

next_states = hidden_states.clone()
# If a token gets dropped, we just set it to zero such that it does not get updated.
next_states = torch.zeros(hidden_states.shape, device=hidden_states.device, dtype=hidden_states.dtype)

router_mask = router_mask.bool()
batch_size, seq_len, num_experts = router_mask.shape
Expand Down
14 changes: 14 additions & 0 deletions tests/models/switch_transformers/test_modeling_switch_transformers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,7 @@
SwitchTransformersEncoderModel,
SwitchTransformersForConditionalGeneration,
SwitchTransformersModel,
SwitchTransformersSparseMLP,
SwitchTransformersTop1Router,
)
from transformers.models.switch_transformers.modeling_switch_transformers import (
Expand Down Expand Up @@ -1133,3 +1134,16 @@ def test_small_batch_generate(self):

for i in range(0, BATCH_SIZE, 2):
self.assertEqual(batch_output[i], batch_output[i + 1])


@require_torch
class SwitchTransformersSparseMLPTests(unittest.TestCase):
def test_token_dropping(self):
r"""
This test checks if the token dropping actually drops tokens.
"""
config = SwitchTransformersConfig(expert_capacity=0) # we drop everything
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Using SwitchTransformersConfig with defaults init a huge model , let's make it tiny. We can even move this under general model tests

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Note that this config is used to init a single Expert-MLP module with 8 experts, hidden size 768, and intermediate size of 2048 and not a full model. There is no attention or multiple layers. In my case, this test ran super fast on CPU.

I could adjust the shapes to the ones chosen in SwitchTransformersModelTester but I don't think, it will have much impact.

I am new to HF testing and this code part. iIf I move this test to general model tests, I assume, I would have to initialize and run a whole model. In this case, I could not easily assert the result of the module, there would be some embedding, attentions, residual connection, and head that would influence the final result. The advantage of going with SwitchTransformersSparseMLP is that I know that the result needs to be all zeroes independent of the input (and input/model shapes).

moe = SwitchTransformersSparseMLP(config)
dropped_token_results = moe(torch.randn(2, 3, 768))[0]

assert (dropped_token_results == 0).all(), f"Some tokens not dropped: {dropped_token_results}."