423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-16 20:00:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix mixed mm inputs and rlhf-v

Browse Source
This commit is contained in:
hiyouga
2024-09-01 20:52:47 +08:00
parent b5063b4144
commit 9967ccb3ae
20 changed files with 306 additions and 277 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
src/llamafactory/data/processors/unsupervised.py
View File

@@ -21,6 +21,7 @@ from .processor_utils import infer_seqlen
if TYPE_CHECKING:
from PIL.Image import Image
from transformers import PreTrainedTokenizer, ProcessorMixin
from ...hparams import DataArguments
@@ -35,25 +36,30 @@ def _encode_unsupervised_example(
response: Sequence[Dict[str, str]],
system: Optional[str],
tools: Optional[str],
images: Sequence["Image"],
template: "Template",
tokenizer: "PreTrainedTokenizer",
processor: Optional["ProcessorMixin"],
cutoff_len: int,
) -> Tuple[List[int], List[int]]:
) -> Tuple[List[int], List[int], Dict[str, Any]]:
if len(response) == 1:
messages = prompt + response
else:
messages = prompt + [{"role": Role.ASSISTANT.value, "content": ""}]
messages = template.mm_plugin.process_messages(messages, images, processor)
input_ids, labels = template.encode_oneturn(tokenizer, messages, system, tools)
if template.efficient_eos:
labels += [tokenizer.eos_token_id]
input_ids, _ = template.mm_plugin.process_token_ids(input_ids, None, tokenizer, processor)
input_ids, _ = template.mm_plugin.process_token_ids(input_ids, None, images, tokenizer, processor)
source_len, target_len = infer_seqlen(len(input_ids), len(labels), cutoff_len)
input_ids = input_ids[:source_len]
labels = labels[:target_len]
return input_ids, labels
extra_inputs = template.mm_plugin.get_mm_inputs(
images=images, feature_seqlens={"token_type_ids": len(input_ids)}, processor=processor
)
return input_ids, labels, extra_inputs
def preprocess_unsupervised_dataset(
@@ -70,12 +76,12 @@ def preprocess_unsupervised_dataset(
logger.warning("Dropped invalid example: {}".format(examples["prompt"][i] + examples["response"][i]))
continue
prompt = template.mm_plugin.process_messages(examples["prompt"][i], examples["images"][i], processor)
input_ids, labels = _encode_unsupervised_example(
prompt=prompt,
input_ids, labels, extra_inputs = _encode_unsupervised_example(
prompt=examples["prompt"][i],
response=examples["response"][i],
system=examples["system"][i],
tools=examples["tools"][i],
images=examples["images"][i],
template=template,
tokenizer=tokenizer,
processor=processor,
@@ -84,12 +90,8 @@ def preprocess_unsupervised_dataset(
model_inputs["input_ids"].append(input_ids)
model_inputs["attention_mask"].append([1] * len(input_ids))
model_inputs["labels"].append(labels)
template.mm_plugin.process_model_inputs(
model_inputs=model_inputs,
images=examples["images"][i],
feature_seqlens={"token_type_ids": len(input_ids)},
processor=processor,
)
for key, value in extra_inputs.items():
model_inputs[key].append(value)
return model_inputs