Skip to content

perf(qwen): project only the final token during trainside generation #442

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
CjhHa1 merged 2 commits into from
Sep 13, 2026
Merged

perf(qwen): project only the final token during trainside generation #442

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9072e7d
perf(qwen3): project only the final token during trainside generation
Yuanjia-Lee Sep 12, 2026
f16278b
perf(qwen): extend final-token projection to VL and 3.5
CjhHa1 Sep 13, 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
1 change: 1 addition & 0 deletions unirl/models/qwen3/ar.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -316,6 +316,7 @@ def autoregress(
past_key_values=model_kwargs.get("past_key_values"),
attention_mask=model_kwargs.get("attention_mask"),
use_cache=True,
logits_to_keep=1,
)
with torch.no_grad():
out = transformer(**model_inputs, return_dict=True)
Expand Down
2 changes: 2 additions & 0 deletions unirl/models/qwen3_5/ar.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -320,6 +320,7 @@ def autoregress(
"mm_token_type_ids": model_kwargs.get("mm_token_type_ids"),
"next_sequence_length": next_sequence_length,
"use_cache": True,
"logits_to_keep": 1,
}
if is_first_step:
if pv is not None:
Expand All @@ -336,6 +337,7 @@ def autoregress(
"input_ids": cur_input_ids,
"attention_mask": model_kwargs["attention_mask"],
"use_cache": False,
"logits_to_keep": 1,
}
if pv is not None:
model_inputs["pixel_values"] = pv
Expand Down
1 change: 1 addition & 0 deletions unirl/models/qwen_vl/ar.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -203,6 +203,7 @@ def autoregress(
"attention_mask": model_kwargs.get("attention_mask"),
"cache_position": model_kwargs.get("cache_position"),
"use_cache": True,
"logits_to_keep": 1,
}
if is_first_step:
if "pixel_values" in model_kwargs:
Expand Down
Loading