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

add more llava-next series template

Browse Source 
Former-commit-id: 93f64f2aebf41582d39aa8a2c6059e562ca694b0
This commit is contained in:
BUAADreamer
2024-09-29 21:29:29 +08:00
parent 4d43317aa1
commit ab5d042cd3
6 changed files with 141 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

57
src/llamafactory/data/template.py
View File

@@ -771,6 +771,52 @@ _register_template(
)
_register_template(
name="llava_next_llama3",
format_user=StringFormatter(
slots=[
(
"<|start_header_id|>user<|end_header_id|>\n\n{{content}}<|eot_id|>"
"<|start_header_id|>assistant<|end_header_id|>\n\n"
)
]
),
format_system=StringFormatter(slots=["<|start_header_id|>system<|end_header_id|>\n\n{{content}}<|eot_id|>"]),
format_observation=StringFormatter(
slots=[
(
"<|start_header_id|>tool<|end_header_id|>\n\n{{content}}<|eot_id|>"
"<|start_header_id|>assistant<|end_header_id|>\n\n"
)
]
),
format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
stop_words=["<|eot_id|>"],
replace_eos=True,
replace_jinja_template=False,
mm_plugin=get_mm_plugin(name="llava_next", image_token="<image>"),
)
_register_template(
name="llava_next_mistral",
format_user=StringFormatter(slots=["[INST] {{content}} [/INST]"]),
format_prefix=EmptyFormatter(slots=[{"bos_token"}]),
mm_plugin=get_mm_plugin(name="llava_next", image_token="<image>"),
)
_register_template(
name="llava_next_yi",
format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
format_separator=EmptyFormatter(slots=["\n"]),
stop_words=["<|im_end|>"],
replace_eos=True,
mm_plugin=get_mm_plugin(name="llava_next", image_token="<image>"),
)
_register_template(
name="llava_next_video",
format_user=StringFormatter(slots=["USER: {{content}} ASSISTANT:"]),
@@ -782,6 +828,17 @@ _register_template(
)
_register_template(
name="llava_next_video_yi",
format_user=StringFormatter(slots=["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]),
format_system=StringFormatter(slots=["<|im_start|>system\n{{content}}<|im_end|>\n"]),
format_separator=EmptyFormatter(slots=["\n"]),
stop_words=["<|im_end|>"],
replace_eos=True,
mm_plugin=get_mm_plugin(name="llava_next_video", image_token="<image>", video_token="<video>"),
)
_register_template(
name="mistral",
format_user=StringFormatter(slots=["[INST] {{content}} [/INST]"]),

57
src/llamafactory/extras/constants.py
View File

@@ -843,14 +843,47 @@ register_model_group(
"LLaVA-NeXT-13B-Chat": {
DownloadSource.DEFAULT: "llava-hf/llava-v1.6-vicuna-13b-hf",
},
"LLaVA-NeXT-34B-Chat": {
DownloadSource.DEFAULT: "llava-hf/llava-v1.6-34b-hf",
},
template="llava_next",
vision=True,
)
register_model_group(
models={
"LLaVA-NeXT-LLaMA3-8B-Chat": {
DownloadSource.DEFAULT: "llava-hf/llama3-llava-next-8b-hf",
},
"LLaVA-NeXT-LLaMA3-72B-Chat": {
DownloadSource.DEFAULT: "llava-hf/llava-next-72b-hf",
},
"LLaVA-NeXT-LLaMA3-110B-Chat": {
DownloadSource.DEFAULT: "llava-hf/llava-next-110b-hf",
},
},
template="llava_next_llama3",
vision=True,
)
register_model_group(
models={
"LLaVA-NeXT-Mistral-7B-Chat": {
DownloadSource.DEFAULT: "llava-hf/llava-v1.6-mistral-7b-hf",
},
},
template="llava_next",
template="llava_next_mistral",
vision=True,
)
register_model_group(
models={
"LLaVA-NeXT-Yi-34B-Chat": {
DownloadSource.DEFAULT: "llava-hf/llava-v1.6-34b-hf",
},
},
template="llava_next_yi",
vision=True,
)
@@ -860,20 +893,28 @@ register_model_group(
"LLaVA-NeXT-Video-7B-Chat": {
DownloadSource.DEFAULT: "llava-hf/LLaVA-NeXT-Video-7B-hf",
},
"LLaVA-NeXT-Video-34B-Chat": {
DownloadSource.DEFAULT: "llava-hf/LLaVA-NeXT-Video-34B-hf",
},
"LLaVA-NeXT-Video-7B-32k-Chat": {
DownloadSource.DEFAULT: "llava-hf/LLaVA-NeXT-Video-7B-32K-hf",
},
"LLaVA-NeXT-Video-7B-DPO": {
DownloadSource.DEFAULT: "llava-hf/LLaVA-NeXT-Video-7B-DPO-hf",
},
"LLaVA-NeXT-Video-34B-DPO": {
},
template="llava_next_video",
vision=True,
)
register_model_group(
models={
"LLaVA-NeXT-Video-Yi-34B-Chat": {
DownloadSource.DEFAULT: "llava-hf/LLaVA-NeXT-Video-34B-hf",
},
"LLaVA-NeXT-Video-Yi-34B-DPO": {
DownloadSource.DEFAULT: "llava-hf/LLaVA-NeXT-Video-34B-DPO-hf",
},
},
template="llava_next_video",
template="llava_next_video_yi",
vision=True,
)

2
src/llamafactory/model/model_utils/misc.py
View File

@@ -34,7 +34,7 @@ def find_all_linear_modules(model: "PreTrainedModel", freeze_vision_tower: bool)
forbidden_modules.add("output_layer")
elif model_type == "internlm2":
forbidden_modules.add("output")
elif model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
elif model_type in ["llava", "paligemma", "video_llava"] or "llava_next" in model_type:
forbidden_modules.add("multi_modal_projector")
elif model_type == "qwen2_vl":
forbidden_modules.add("merger")

13
src/llamafactory/model/model_utils/visual.py
View File

@@ -92,7 +92,7 @@ def autocast_projector_dtype(model: "PreTrainedModel", model_args: "ModelArgumen
if getattr(model, "quantization_method", None):
model_type = getattr(model.config, "model_type", None)
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
if model_type in ["llava", "paligemma", "video_llava"] or "llava_next" in model_type:
mm_projector: "torch.nn.Module" = getattr(model, "multi_modal_projector")
elif model_type == "qwen2_vl":
mm_projector: "torch.nn.Module" = getattr(getattr(model, "visual"), "merger")
@@ -108,12 +108,7 @@ def configure_visual_model(config: "PretrainedConfig") -> None:
Patches VLMs before loading them.
"""
model_type = getattr(config, "model_type", None)
if model_type in [
"llava",
"llava_next",
"llava_next_video",
"video_llava",
]: # required for ds zero3 and valuehead models
if model_type in ["llava", "video_llava"] or "llava_next" in model_type: # required for ds zero3 and valuehead models
setattr(config, "hidden_size", getattr(config.text_config, "hidden_size", None))
if getattr(config, "is_yi_vl_derived_model", None):
@@ -127,7 +122,7 @@ def get_forbidden_modules(config: "PretrainedConfig", finetuning_args: "Finetuni
"""
model_type = getattr(config, "model_type", None)
forbidden_modules = set()
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
if model_type in ["llava", "paligemma", "video_llava"] or "llava_next" in model_type:
if finetuning_args.freeze_vision_tower:
forbidden_modules.add("vision_tower")
@@ -185,7 +180,7 @@ def patch_target_modules(
"""
model_type = getattr(config, "model_type", None)
if finetuning_args.freeze_vision_tower:
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
if model_type in ["llava", "paligemma", "video_llava"] or "llava_next" in model_type:
return "^(?!.*vision_tower).*(?:{}).*".format("|".join(target_modules))
elif model_type == "qwen2_vl":
return "^(?!.*visual).*(?:{}).*".format("|".join(target_modules))