A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[V0 Deprecation] Remove V0-only methods in multi-modal registry (#25362)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Sep 22, 2025 at 13:58 UTC
6d0b827cbd0510173f6a9e77549d917828e80c76
2 files changed
+1
-32
tests/models/multimodal/generation/test_qwen2_vl.py
-1
@@ -209,7 +209,6 @@ def batch_make_video_embeddings(
209
return visual(pixel_values_on_device,
210
grid_thw=video_grid_thw_on_device).cpu()
211
212
- # V1 Test: this calls a V0 internal.
212
video_embeds = torch.concat(llm.apply_model(get_image_embeds))
213
214
# split into original batches
vllm/multimodal/registry.py
+1
-31
@@ -12,8 +12,7 @@ from vllm.transformers_utils.tokenizer import (AnyTokenizer,
12
cached_tokenizer_from_config)
13
from vllm.utils import ClassRegistry
14
15
-from .cache import (BaseMultiModalProcessorCache,
16
- processor_only_cache_from_config)
15
+from .cache import BaseMultiModalProcessorCache
16
from .processing import BaseMultiModalProcessor, BaseProcessingInfo
17
from .profiling import (BaseDummyInputsBuilder, DummyDecoderData,
18
DummyEncoderData, MultiModalProfiler)
@@ -176,35 +175,6 @@ class MultiModalRegistry:
175
if mm_limits[key] > 0
176
}
177
179
- # TODO: Remove once V0 is gone
180
- def get_max_tokens_by_modality(
181
- self,
182
- model_config: "ModelConfig",
183
- ) -> Mapping[str, int]:
184
- """
185
- Get the maximum number of tokens from each modality
186
- for profiling the memory usage of a model.
187
- """
188
- cache = processor_only_cache_from_config(model_config, self)
189
- mm_limits = self.get_mm_limits_per_prompt(model_config, cache=cache)
190
- max_tokens_per_item = self.get_max_tokens_per_item_by_modality(
191
- model_config,
192
- cache=cache,
193
- )
194
-
195
- return {
196
- key: mm_limits[key] * max_tokens_per_mm_item
197
- for key, max_tokens_per_mm_item in max_tokens_per_item.items()
198
- }
199
-
200
- # TODO: Remove once V0 is gone
201
- def get_max_multimodal_tokens(self, model_config: "ModelConfig") -> int:
202
- """
203
- Get the maximum number of multi-modal tokens
204
- for profiling the memory usage of a model.
205
- """
206
- return sum(self.get_max_tokens_by_modality(model_config).values())
207
-
178
def get_mm_limits_per_prompt(
179
self,
180
model_config: "ModelConfig",