[Core] Remove legacy input mapper/processor from V0 (#15686)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-04-28 15:38:48 +08:00
committed by GitHub
parent 7fcc4223dc
commit aec9674dbe
21 changed files with 142 additions and 1337 deletions
+5 -17
View File
@@ -457,7 +457,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
self.enable_lora = self.runner.lora_config is not None
self.enable_prompt_adapter = (self.runner.prompt_adapter_config
is not None)
self.multi_modal_input_mapper = self.runner.multi_modal_input_mapper
# Attention metadata inputs.
if self.attn_backend is not None:
@@ -675,23 +674,15 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
def _compute_multi_modal_input(self, inter_data: InterDataForSeqGroup,
seq_group_metadata: SequenceGroupMetadata):
"""If multi-modal data is given, add it to the input."""
# NOTE: mm_data only includes the subset of multi-modal items that
# NOTE: mm_kwargs only includes the subset of multi-modal items that
# intersect with the current prefill positions.
positions = inter_data.input_positions[0]
mm_data, placeholder_maps = MultiModalPlaceholderMap.from_seq_group(
mm_kwargs, placeholder_maps = MultiModalPlaceholderMap.from_seq_group(
seq_group_metadata,
range(positions[0], positions[0] + len(positions)))
if not mm_data:
if not mm_kwargs:
return
if self.runner.mm_registry.has_processor(self.runner.model_config):
mm_kwargs = mm_data
else:
mm_kwargs = self.multi_modal_input_mapper(
mm_data,
seq_group_metadata.mm_processor_kwargs,
)
inter_data.multi_modal_kwargs = mm_kwargs
inter_data.multi_modal_placeholder_maps = placeholder_maps
@@ -1085,9 +1076,6 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
# Multi-modal data support
self.input_registry = input_registry
self.mm_registry = mm_registry
self.multi_modal_input_mapper = mm_registry \
.create_input_mapper(model_config)
self.mm_registry.init_mm_limits_per_prompt(self.model_config)
# Lazy initialization
self.model: nn.Module # Set after load_model
@@ -1327,8 +1315,8 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
dummy_data = self.input_registry \
.dummy_data_for_profiling(self.model_config,
seq_len,
self.mm_registry)
seq_len,
self.mm_registry)
seq = SequenceGroupMetadata(
request_id=str(group_id),