[Multimodal] Expose mm hash algothrim selection to cli args (#49686)

Signed-off-by: Isotr0py <Isotr0py@outlook.com>
This commit is contained in:
Isotr0py
2026-07-31 05:25:19 +00:00
committed by GitHub
parent 0f17394564
commit 1180b604f8
17 changed files with 234 additions and 61 deletions
+8 -1
View File
@@ -3,6 +3,7 @@
from collections.abc import Mapping
from dataclasses import dataclass, field
from vllm.config.multimodal import MMHasherAlgorithm
from vllm.inputs import MultiModalHashes
from ..hasher import MultiModalHasher
@@ -22,7 +23,11 @@ class ProcessorInputs:
hf_processor_mm_kwargs: Mapping[str, object] = field(default_factory=dict)
tokenization_kwargs: Mapping[str, object] = field(default_factory=dict)
def get_mm_hashes(self, model_id: str) -> MultiModalHashes:
def get_mm_hashes(
self,
model_id: str,
hash_algorithm: MMHasherAlgorithm,
) -> MultiModalHashes:
mm_data_items = self.mm_data_items
mm_uuid_items = self.mm_uuid_items or {}
hf_processor_mm_kwargs = self.hf_processor_mm_kwargs
@@ -49,6 +54,7 @@ class ProcessorInputs:
item = uuid_item if uuid_item is not None else item
hashes.append(
hasher.hash_kwargs(
hash_algorithm,
model_id=model_id,
**{modality: item},
**hf_processor_mm_kwargs,
@@ -61,6 +67,7 @@ class ProcessorInputs:
else:
mm_hashes[modality] = [
hasher.hash_kwargs(
hash_algorithm,
model_id=model_id,
**{modality: item},
**hf_processor_mm_kwargs,
+8 -2
View File
@@ -1422,7 +1422,10 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
# Use overrides if provided; fallback to data-dependent hashing.
with timing_ctx.record("get_mm_hashes"):
mm_hashes = inputs.get_mm_hashes(self.info.model_id)
mm_hashes = inputs.get_mm_hashes(
self.info.model_id,
self.info.ctx.get_mm_config().mm_hasher_algorithm,
)
mm_prompt_updates = self._get_mm_prompt_updates(
inputs.mm_data_items,
@@ -1454,7 +1457,10 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
return self._apply_hf_processor(inputs, timing_ctx)
with timing_ctx.record("get_mm_hashes"):
mm_hashes = inputs.get_mm_hashes(self.info.model_id)
mm_hashes = inputs.get_mm_hashes(
self.info.model_id,
self.info.ctx.get_mm_config().mm_hasher_algorithm,
)
with timing_ctx.record("get_cache_missing_items"):
mm_is_cached, mm_missing_data_items = self._get_cache_missing_items(