Use Transformers v5 WeightRenaming for Transformers modeling backend (#31545)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-03-13 20:49:08 +00:00
committed by GitHub
parent d0b402974f
commit 0005d2a3c9
7 changed files with 162 additions and 89 deletions
@@ -24,7 +24,6 @@ import torch
from vllm.config.utils import getattr_iter
from vllm.logger import init_logger
from vllm.model_executor.models.interfaces import SupportsMRoPE, SupportsMultiModal
from vllm.model_executor.models.utils import WeightsMapper
from vllm.multimodal import MultiModalKwargsItems
from vllm.multimodal.inputs import (
MultiModalDataDict,
@@ -273,30 +272,6 @@ class MultiModalProcessor(BaseMultiModalProcessor[MultiModalProcessingInfo]):
class MultiModalMixin(SupportsMultiModal, SupportsMRoPE):
supports_multimodal_raw_input_only = True
# Backwards compatibility for prev released models. State dicts back then
# had different formats and cannot be loaded with `AutoModel` mapping as is
hf_to_vllm_mapper = WeightsMapper(
orig_to_new_prefix={
"language_model.model": "model.language_model",
"text_model.model": "model.text_model",
"vision_tower": "model.vision_tower",
"vqmodel": "model.vqmodel",
"visual": "model.visual",
"vision_model": "model.vision_model",
"vision_embed_tokens": "model.vision_embed_tokens",
"image_newline": "model.image_newline",
"multi_modal_projector": "model.multi_modal_projector",
"text_model.lm_head": "lm_head",
"language_model.lm_head": "lm_head",
# Qwen models used "model" as the name for the language model.
# Therefore, we must map each of submodule explicitly to avoid
# conflicts with newer models that use "model.language_model".
"model.embed_tokens": "model.language_model.embed_tokens",
"model.layers": "model.language_model.layers",
"model.norm": "model.language_model.norm",
}
)
def __init__(self, *, vllm_config: "VllmConfig", prefix: str = ""):
# Skip SupportsMRoPE.__init__ and call the next class in MRO
super(SupportsMRoPE, self).__init__(vllm_config=vllm_config, prefix=prefix)