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

[V1] Clarify input processing and multimodal feature caching logic (#13211)

Browse Source
This commit is contained in:
Roger Wang
2025-02-13 03:43:24 -08:00
committed by GitHub
parent 578087e56c
commit fdcf64d3c6
4 changed files with 46 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/v1/worker/gpu_model_runner.py
View File

@@ -27,7 +27,7 @@ from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
from vllm.v1.attention.backends.flash_attn import (FlashAttentionBackend,
FlashAttentionMetadata)
from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
from vllm.v1.engine.mm_input_mapper import MMInputMapperClient
from vllm.v1.engine.mm_input_cache import MMInputCacheClient
from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheConfig,
KVCacheSpec)
from vllm.v1.outputs import LogprobsTensors, ModelRunnerOutput
@@ -95,9 +95,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
self.mm_registry = MULTIMODAL_REGISTRY
self.uses_mrope = model_config.uses_mrope
# NOTE: Initialized input mapper is only used for processing dummy
# NOTE: Initialized client is only used for processing dummy
# multimodal data into multimodal kwargs for GPU memory profiling.
self.mm_input_mapper_profiling = MMInputMapperClient(self.model_config)
# Only applicable to multimodal models with legacy input mapper.
self.mm_input_mapper_profiling = MMInputCacheClient(self.model_config)
self.mm_input_mapper_profiling.use_cache = False
encoder_compute_budget, encoder_cache_size = compute_encoder_budget(