[Bugfix] Refactor composite weight loading logic (#8656)

This commit is contained in:
Isotr0py
2024-09-22 04:33:27 +00:00
committed by GitHub
parent d66ac62854
commit 13d88d4137
7 changed files with 70 additions and 61 deletions
+7 -13
View File
@@ -1,4 +1,3 @@
import itertools
from typing import (Iterable, List, Literal, Mapping, Optional, Tuple,
TypedDict, Union)
@@ -30,8 +29,8 @@ from .llava import LlavaMultiModalProjector
from .siglip import (SiglipVisionModel, dummy_image_for_siglip,
dummy_seq_data_for_siglip, get_siglip_image_feature_size,
get_siglip_patch_grid_length, input_processor_for_siglip)
from .utils import (filter_weights, flatten_bn, init_vllm_registered_model,
merge_multimodal_embeddings)
from .utils import (flatten_bn, group_weights_with_prefix,
init_vllm_registered_model, merge_multimodal_embeddings)
logger = init_logger(__name__)
@@ -637,25 +636,21 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal):
def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
# prepare weight iterators for components
vit_weights, mlp_weights, newline_weights, llm_weights = itertools.tee(
weights, 4)
weights_group = group_weights_with_prefix(weights)
# load vision encoder
vit_weights = filter_weights(vit_weights, "vision_tower")
self.vision_tower.load_weights(vit_weights)
self.vision_tower.load_weights(weights_group["vision_tower"])
# load mlp projector
mlp_weights = filter_weights(mlp_weights, "multi_modal_projector")
mlp_params_dict = dict(self.multi_modal_projector.named_parameters())
for name, loaded_weight in mlp_weights:
for name, loaded_weight in weights_group["multi_modal_projector"]:
param = mlp_params_dict[name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
weight_loader(param, loaded_weight)
# load newline
newline_weights = filter_weights(newline_weights, "image_newline")
for name, loaded_weight in newline_weights:
for name, loaded_weight in weights_group["image_newline"]:
assert name == ""
param = self.image_newline
weight_loader = getattr(param, "weight_loader",
@@ -663,5 +658,4 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal):
weight_loader(param, loaded_weight)
# load llm backbone
llm_weights = filter_weights(llm_weights, "language_model")
self.language_model.load_weights(llm_weights)
self.language_model.load_weights(weights_group["language_model"])