Skip to content

Update the longrope inv_freq buffer in place #50480

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
hmellor merged 1 commit into from
Nov 9, 2026
+25 −3
Merged

Update the longrope inv_freq buffer in place #50480

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
3 changes: 2 additions & 1 deletion src/transformers/modeling_rope_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,7 +77,8 @@ def longrope_frequency_update(self, position_ids, device, layer_type=None):

# `seq_len` is a scalar, so this picks the long or short set for all positions, like an if/else but traceable
inv_freq = torch.where(seq_len > original_max_position_embeddings, long_inv_freq, original_inv_freq)
self.register_buffer(f"{prefix}inv_freq", inv_freq, persistent=False)
# Update in place so the buffer keeps its address, which CUDA graphs replay from
getattr(self, f"{prefix}inv_freq").copy_(inv_freq)

@vasqu vasqu Nov 9, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

wondering whether we need something similar for dynamic freq updates?

@zucchini-nlp zucchini-nlp Nov 9, 2026 •
edited
Loading

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

it is still not compile friendly overall (input-dependent seq_len > max_len, jiqing tried but it led to computational overhead at each forward step

@hmellor hmellor Nov 9, 2026

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

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

This was motivated by vLLM, but ultimately I now must choose either long or short depending on how vLLM is configured. This is how vLLM treats log rope for its native models too


def dynamic_frequency_update(self, position_ids, device, layer_type=None):
"""
Expand Down
25 changes: 23 additions & 2 deletions tests/models/phi3/test_modeling_phi3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -132,8 +132,7 @@ def test_longrope_generation_crossing_boundary(self, sliding_window):
expected = model(prefix, use_cache=False).logits[:, -1]
torch.testing.assert_close(logits, expected, rtol=1e-4, atol=1e-4)

@parameterized.expand([(8,), (16,)])
def test_longrope_compiled_matches_eager(self, seq_len):
def _get_longrope_config(self):
config = self.model_tester.get_config()
config.original_max_position_embeddings = 8
config.max_position_embeddings = 32
Expand All @@ -146,6 +145,11 @@ def test_longrope_compiled_matches_eager(self, seq_len):
"original_max_position_embeddings": 8,
"factor": 4.0,
}
return config

@parameterized.expand([(8,), (16,)])
def test_longrope_compiled_matches_eager(self, seq_len):
config = self._get_longrope_config()
model = Phi3ForCausalLM(config).to(torch_device).eval()
input_ids = torch.arange(1, seq_len + 1, device=torch_device).unsqueeze(0)

Expand All @@ -154,6 +158,23 @@ def test_longrope_compiled_matches_eager(self, seq_len):
compiled = torch.compile(model, fullgraph=True)(input_ids, use_cache=False).logits
torch.testing.assert_close(compiled, expected, rtol=1e-4, atol=1e-4)

def test_longrope_updates_inv_freq_in_place(self):
# CUDA graphs replay from the buffer's address, so switching frequency sets must not replace it
config = self._get_longrope_config()
rotary_emb = Phi3ForCausalLM(config).to(torch_device).model.rotary_emb
inv_freq = rotary_emb.inv_freq
short_inv_freq = rotary_emb.original_inv_freq.clone()
x = torch.zeros(1, device=torch_device)

with torch.no_grad():
rotary_emb(x, torch.arange(16, device=torch_device)[None])
self.assertIs(rotary_emb.inv_freq, inv_freq)
torch.testing.assert_close(inv_freq, rotary_emb.long_inv_freq)

rotary_emb(x, torch.arange(8, device=torch_device)[None])
self.assertIs(rotary_emb.inv_freq, inv_freq)
torch.testing.assert_close(inv_freq, short_inv_freq)


@slow
@require_torch
Expand Down
Loading