ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[version] support transformers 449 (#6982)

Browse Source 
* support transformers 449

* fix mm plugin

Former-commit-id: e9118a9df0839d24f6ddff5a0b55ef101a1d3d22
This commit is contained in:
hoshi-hiyouga
2025-02-18 17:05:40 +08:00
committed by GitHub
parent be33ef67fb
commit 1d675a287d
7 changed files with 16 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/llamafactory/data/collator.py
View File

@@ -187,8 +187,6 @@ class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
mm_inputs["cross_attention_mask"] = F.pad(cross_attention_mask, (0, 0, 0, 0, 0, seq_len - orig_len))
features.update(mm_inputs)
if isinstance(features.get("pixel_values"), list): # for pixtral inputs
features = features.data # use default_collate() instead of BatchEncoding.to()
if "image_bound" in features: # for minicpmv inputs
bsz, seq_length = features["input_ids"].shape

19
src/llamafactory/data/mm_plugin.py
View File

@@ -380,10 +380,8 @@ class LlavaNextPlugin(BasePlugin):
num_image_tokens = 0
messages = deepcopy(messages)
mm_inputs = self._get_mm_inputs(images, videos, audios, processor)
if "image_sizes" in mm_inputs:
image_sizes = iter(mm_inputs["image_sizes"])
if "pixel_values" in mm_inputs:
image_sizes = iter(mm_inputs["image_sizes"].tolist())
height, width = get_image_size(to_numpy_array(mm_inputs["pixel_values"][0][0]))
for message in messages:
@@ -439,7 +437,7 @@ class LlavaNextVideoPlugin(BasePlugin):
messages = deepcopy(messages)
mm_inputs = self._get_mm_inputs(images, videos, audios, processor)
if "pixel_values" in mm_inputs:
image_sizes = iter(mm_inputs["image_sizes"])
image_sizes = iter(mm_inputs["image_sizes"].tolist())
height, width = get_image_size(to_numpy_array(mm_inputs["pixel_values"][0][0]))
for message in messages:
content = message["content"]
@@ -916,16 +914,14 @@ class PixtralPlugin(BasePlugin):
num_image_tokens = 0
messages = deepcopy(messages)
mm_inputs = self._get_mm_inputs(images, videos, audios, processor)
image_input_sizes = mm_inputs.get("image_sizes", None)
if "pixel_values" in mm_inputs:
image_sizes = iter(mm_inputs["image_sizes"].tolist())
for message in messages:
content = message["content"]
while IMAGE_PLACEHOLDER in content:
if image_input_sizes is None:
raise ValueError("Cannot get image input sizes.")
if self.expand_mm_tokens:
image_size = image_input_sizes[0][num_image_tokens]
height, width = image_size
height, width = next(image_sizes)
num_height_tokens = height // patch_size
num_width_tokens = width // patch_size
replace_tokens = [[image_token] * num_width_tokens + [image_break_token]] * num_height_tokens
@@ -959,9 +955,6 @@ class PixtralPlugin(BasePlugin):
) -> Dict[str, Union[List[int], "torch.Tensor"]]:
self._validate_input(images, videos, audios)
mm_inputs = self._get_mm_inputs(images, videos, audios, processor)
if mm_inputs.get("pixel_values"):
mm_inputs["pixel_values"] = mm_inputs["pixel_values"][0]
mm_inputs.pop("image_sizes", None)
return mm_inputs