423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-30 02:30:35 +08:00
Code Issues Packages Projects Releases Wiki Activity

[misc] fix packing and eval plot (#7623)

Browse Source
This commit is contained in:
hoshi-hiyouga
2025-04-07 18:20:57 +08:00
committed by GitHub
parent c0079ab9fd
commit fb46193364
70 changed files with 288 additions and 194 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/llamafactory/data/processor/supervised.py
View File

@@ -164,28 +164,28 @@ class PackedSupervisedDatasetProcessor(SupervisedDatasetProcessor):
model_inputs = defaultdict(list)
knapsacks = greedy_knapsack(lengths, self.data_args.cutoff_len)
for knapsack in knapsacks:
packed_input_ids, packed_attention_masks, packed_labels = [], [], []
packed_images, packed_videos, packed_audios, packed_position_ids = [], [], [], []
packed_input_ids, packed_attention_masks, packed_position_ids, packed_labels = [], [], [], []
packed_images, packed_videos, packed_audios = [], [], []
for i, length in enumerate(knapsack):
index = length2indexes[length].pop()
packed_input_ids += batch_input_ids[index]
packed_position_ids += list(range(len(batch_input_ids[index]))) # NOTE: pad_to_multiple_of ignore this
packed_labels += batch_labels[index]
packed_images += batch_images[index]
packed_videos += batch_videos[index]
packed_audios += batch_audios[index]
if self.data_args.neat_packing:
packed_attention_masks += [i + 1] * len(batch_input_ids[index]) # start from 1
packed_position_ids += list(range(len(batch_input_ids[index])))
else:
packed_attention_masks += [1] * len(batch_input_ids[index])
if len(packed_input_ids) < self.data_args.cutoff_len + 1: # avoid flash_attn drops attn mask
pad_length = self.data_args.cutoff_len - len(packed_input_ids) + 1
packed_input_ids += [self.tokenizer.pad_token_id] * pad_length
packed_position_ids += [0] * pad_length
packed_labels += [IGNORE_INDEX] * pad_length
if self.data_args.neat_packing:
packed_attention_masks += [0] * pad_length
packed_position_ids += [0] * pad_length
else:
packed_attention_masks += [1] * pad_length # more efficient flash_attn
@@ -194,10 +194,10 @@ class PackedSupervisedDatasetProcessor(SupervisedDatasetProcessor):
model_inputs["input_ids"].append(packed_input_ids)
model_inputs["attention_mask"].append(packed_attention_masks)
model_inputs["position_ids"].append(packed_position_ids)
model_inputs["labels"].append(packed_labels)
model_inputs["images"].append(packed_images or None)
model_inputs["videos"].append(packed_videos or None)
model_inputs["audios"].append(packed_audios or None)
model_inputs["position_ids"].append(packed_position_ids or None)
return model_inputs