biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model] Use context managers for encoder- and LM-only mode (#32605)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-20 11:43:38 +08:00
committed by GitHub
parent 6c01ffb897
commit 4753f3bf69
21 changed files with 290 additions and 353 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
vllm/model_executor/models/qwen2_5_vl.py
View File

@@ -35,7 +35,7 @@ import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import BatchFeature, Qwen2ForCausalLM
from transformers import BatchFeature
from transformers.models.qwen2_5_vl import Qwen2_5_VLProcessor
from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import (
Qwen2_5_VLConfig,
@@ -1145,9 +1145,7 @@ class Qwen2_5_VLForConditionalGeneration(
multimodal_config.is_multimodal_pruning_enabled()
)
if multimodal_config.get_limit_per_prompt(
"image"
) or multimodal_config.get_limit_per_prompt("video"):
with self._mark_tower_model(vllm_config, {"image", "video"}):
self.visual = Qwen2_5_VisionTransformer(
vision_config=config.vision_config,
norm_eps=getattr(config, "rms_norm_eps", 1e-6),
@@ -1155,14 +1153,13 @@ class Qwen2_5_VLForConditionalGeneration(
prefix=maybe_prefix(prefix, "visual"),
multimodal_config=multimodal_config,
)
else:
self.visual = None
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "language_model"),
architectures=["Qwen2ForCausalLM"],
)
with self._mark_language_model(vllm_config):
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "language_model"),
architectures=["Qwen2ForCausalLM"],
)
self.make_empty_intermediate_tensors = (
self.language_model.make_empty_intermediate_tensors
@@ -1447,9 +1444,6 @@ class Qwen2_5_VLForConditionalGeneration(
)
return mm_input_by_modality
def get_language_model(self) -> torch.nn.Module:
return self.language_model
def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
mm_input_by_modality = self._parse_and_validate_multimodal_inputs(**kwargs)
if not mm_input_by_modality:
@@ -1516,10 +1510,7 @@ class Qwen2_5_VLForConditionalGeneration(
return self.language_model.compute_logits(hidden_states)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
skip_prefixes = []
if self.visual is None:
skip_prefixes.extend(["visual."])
loader = AutoWeightsLoader(self, skip_prefixes=skip_prefixes)
loader = AutoWeightsLoader(self)
return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
def get_mm_mapping(self) -> MultiModelKeys:
@@ -1550,11 +1541,3 @@ class Qwen2_5_VLForConditionalGeneration(
vision_config = hf_config.vision_config
merge_size = vision_config.spatial_merge_size
return num_vision_tokens // merge_size**2
@classmethod
def get_language_model_spec(cls) -> tuple[nn.Module | None, str | None]:
"""
Return the language model spec:
(language model class, language model attr)
"""
return Qwen2ForCausalLM, "language_model"