423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-23 14:22:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix system prompt and tests

Browse Source 
Former-commit-id: cfaa8e4890ad99ec1fb90d9550503d734b5c30b7
This commit is contained in:
fzc8578 2025-01-13 14:18:06 +08:00
parent bdded9d41a
commit d2afe0c63c
4 changed files with 17 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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

@ -157,7 +157,7 @@ class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
features["position_ids"] = [torch.arange(input_ids.size(0)).long() for input_ids in features["input_ids"]]
features["position_ids"] = pad_sequence(features["position_ids"], batch_first=True, padding_value=0)
new_features = {"data": features}
new_features.update(features)
new_features.update({"labels": features['labels']})
features = new_features
return features

7
src/llamafactory/data/mm_plugin.py
View File

@ -383,6 +383,7 @@ class CpmOPlugin(BasePlugin):
self._validate_input(images, videos)
image_bounds_list = []
valid_image_nums_ls = []
flag = False
for input_ids in batch_ids:
input_ids_ = torch.tensor(input_ids)
@ -394,6 +395,8 @@ class CpmOPlugin(BasePlugin):
image_start_tokens += 1
image_end_tokens = torch.where(end_cond)[0]
valid_image_nums = max(len(image_start_tokens), len(image_end_tokens))
if valid_image_nums > 0:
flag = True
valid_image_nums_ls.append(valid_image_nums)
image_bounds = torch.hstack(
[
@ -402,6 +405,10 @@ class CpmOPlugin(BasePlugin):
]
)
image_bounds_list.append(image_bounds)
if not flag and len(images)>0:
valid_image_nums_ls = [1 for _ in range(len(batch_ids))]
image_bounds_list = [torch.arange(64) for _ in range(len(batch_ids))]
mm_inputs = self._get_mm_inputs(images, videos, processor, valid_image_nums_ls=valid_image_nums_ls)
mm_inputs.update({"image_bound": image_bounds_list})

6
src/llamafactory/data/template.py
View File

@ -571,12 +571,6 @@ _register_template(
format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
format_assistant=StringFormatter(slots=["{{content}}<|im_end|>\n"]),
format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
format_function=FunctionFormatter(slots=["{{content}}<|im_end|>\n"], tool_format="qwen"),
format_observation=StringFormatter(
slots=["<|im_start|>user\n<tool_response>\n{{content}}\n</tool_response><|im_end|>\n<|im_start|>assistant\n"]
),
format_tools=ToolFormatter(tool_format="qwen"),
default_system="You are a helpful assistant.",
stop_words=["<|im_end|>"],
mm_plugin=get_mm_plugin(name="cpm_o", image_token="<image>", video_token="<video>"),
)

12
tests/data/test_mm_plugin.py
View File

@ -76,9 +76,15 @@ def _is_close(batch_a: Dict[str, Any], batch_b: Dict[str, Any]) -> None:
if isinstance(batch_a[key], torch.Tensor):
assert torch.allclose(batch_a[key], batch_b[key], rtol=1e-4, atol=1e-5)
elif isinstance(batch_a[key], list) and all(isinstance(item, torch.Tensor) for item in batch_a[key]):
assert len(batch_a[key]) == len(batch_b[key])
for tensor_a, tensor_b in zip(batch_a[key], batch_b[key]):
assert torch.allclose(tensor_a, tensor_b, rtol=1e-4, atol=1e-5)
assert len(batch_a[key]) == len(batch_b[key])
for tensor_a, tensor_b in zip(batch_a[key], batch_b[key]):
assert torch.allclose(tensor_a, tensor_b, rtol=1e-4, atol=1e-5)
elif isinstance(batch_a[key], list) and all(isinstance(item, list) for item in batch_a[key]) \
and len(batch_a[key])>0 and len(batch_a[key][0])>0 and isinstance(batch_a[key][0][0], torch.Tensor):
for item_a, item_b in zip(batch_a[key], batch_b[key]):
assert len(item_a) == len(item_a)
for tensor_a, tensor_b in zip(item_a, item_b):
assert torch.allclose(tensor_a, tensor_b, rtol=1e-4, atol=1e-5)
else:
assert batch_a[key] == batch_b[key]