423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2026-09-06 15:25:43 +08:00
Code Issues Packages Projects Releases Wiki Activity

[data] fix position ids after merging packed-mrope (#10783)

Browse Source 
Co-authored-by: haqishen <qishen.ha@nalalys.com>
This commit is contained in:
haqishen
2026-08-31 17:13:06 +09:00
committed by GitHub
parent 6369f90cee
commit d6bb97ddff
1 changed files with 10 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/llamafactory/data/collator.py
View File

@@ -317,6 +317,16 @@ class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
[features["attention_mask"], dummy_image_right_padding_attention_mask], dim=-1
)
# Mirror the non-FA2 packing fix (#10737) for the packed-mrope path. The merged position_ids
# is built from per-subseq sequence_boundaries, which end at cutoff_len, while
# `DataCollatorForSeq2Seq(pad_to_multiple_of=...)` right-pads input_ids/attention_mask past
# cutoff_len. Right-pad the trailing (masked) positions with 0 so the merged position_ids
# matches seq_len before validating. Works for both 2D and 3D (mrope) position_ids since the
# sequence axis is last, and is idempotent with the has_dummy_image cat above.
pad_len = seq_len - features["position_ids"].shape[-1]
if pad_len > 0:
features["position_ids"] = F.pad(features["position_ids"], (0, pad_len), value=0)
if features["position_ids"].shape != expected_position_ids_shape:
raise ValueError(
"Merged position_ids shape mismatch: "