Skip to content

[CB] Ensure parallel decoding test passes using FA #43277

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
vasqu merged 2 commits into from
Jan 14, 2026
Merged

[CB] Ensure parallel decoding test passes using FA #43277

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
376bf21
Ensure parallel decoding matches using FA
remi-or Jan 14, 2026
0f36dc7
Merge branch 'main' into cb-flakyness
remi-or Jan 14, 2026
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: 3 additions & 3 deletions tests/generation/test_continuous_batching.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,6 +38,7 @@
from transformers.testing_utils import (
Expectations,
require_deterministic_for_xpu,
require_flash_attn,
require_torch_accelerator,
slow,
torch_device,
Expand Down Expand Up @@ -601,10 +602,9 @@ def test_block_sharing_with_hybrid_model(self) -> None:

return self._test_block_sharing(model_id, num_layer_groups, input_msg, expected_generated_tokens)

# The test always passes on H100 with torch 2.9, but only passed case 0 on A100 with torch 2.6 and fails on A100
# with torch 2.9. This might be due to a GPU diff, so test might be flaky on the CI which runs on A10.
@parameterized.expand([True, False])
@require_torch_accelerator
@require_flash_attn # otherwise the test can fail because attention bias has a very slight impact on SDPA and eager
Copy link
Copy Markdown
Contributor

@vasqu vasqu Jan 14, 2026

Choose a reason for hiding this comment

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

Thx for the comment

def test_num_return_sequences(self, allow_block_sharing: bool) -> None:
model_id = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
tokenizer = AutoTokenizer.from_pretrained(model_id, padding_side="left")
Expand All @@ -616,7 +616,7 @@ def test_num_return_sequences(self, allow_block_sharing: bool) -> None:
input_ids = [(x if isinstance(x, list) else x["input_ids"]) for x in tokenized]

# Generation with continuous batching
model = AutoModelForCausalLM.from_pretrained(model_id, attn_implementation="sdpa")
model = AutoModelForCausalLM.from_pretrained(model_id, attn_implementation="flash_attention_2")
model = model.to(torch_device).eval()
model.generation_config.max_new_tokens = 30
model.generation_config.do_sample = False
Expand Down