[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 -20
View File
@@ -22,8 +22,8 @@ from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
from vllm.model_executor.model_loader import get_model
from vllm.model_executor.models import supports_lora, supports_multimodal
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
MultiModalKwargs, MultiModalPlaceholderMap)
from vllm.multimodal import (BatchedTensorInputs, MultiModalKwargs,
MultiModalPlaceholderMap)
from vllm.sequence import (IntermediateTensors, SequenceData,
SequenceGroupMetadata)
from vllm.worker.model_runner_base import (
@@ -154,7 +154,6 @@ class ModelInputForCPUBuilder(ModelRunnerInputBuilderBase[ModelInputForCPU]):
self.sliding_window = self.runner.sliding_window
self.block_size = self.runner.block_size
self.device = self.runner.device
self.multi_modal_input_mapper = self.runner.multi_modal_input_mapper
self.enable_lora = self.runner.lora_config is not None
if self.runner.attn_backend is not None:
# spec decode (e.g. Medusa) does not have atten backend
@@ -359,22 +358,14 @@ class ModelInputForCPUBuilder(ModelRunnerInputBuilderBase[ModelInputForCPU]):
computed_len = seq_data.get_num_computed_tokens()
seq_len = self.input_data.seq_lens[-1]
# 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.
mm_data, placeholder_maps = MultiModalPlaceholderMap.from_seq_group(
mm_kwargs, placeholder_maps = MultiModalPlaceholderMap.from_seq_group(
seq_group_metadata, range(computed_len, seq_len))
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,
)
# special processing for mrope position deltas.
if self.runner.model_config.uses_mrope:
assert not self.chunked_prefill, \
@@ -480,12 +471,6 @@ class CPUModelRunnerBase(ModelRunnerBase[TModelInputForCPU]):
use_mla=self.model_config.use_mla,
) if needs_attn_backend else None
# Multi-modal data support
self.mm_registry = MULTIMODAL_REGISTRY
self.multi_modal_input_mapper = self.mm_registry \
.create_input_mapper(self.model_config)
self.mm_registry.init_mm_limits_per_prompt(self.model_config)
# Lazy initialization.
self.model: nn.Module # Set after init_Model
# Set after load_model.
+2
View File
@@ -100,6 +100,8 @@ class EncoderDecoderModelRunner(GPUModelRunnerBase[EncoderDecoderModelInput]):
vllm_config=vllm_config,
kv_cache_dtype=kv_cache_dtype,
is_driver_worker=is_driver_worker,
input_registry=input_registry,
mm_registry=mm_registry,
)
# Crash for unsupported encoder/scenarios
+3 -9
View File
@@ -45,8 +45,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding)
from vllm.model_executor.model_loader import get_model
from vllm.model_executor.sampling_metadata import SequenceGroupToSample
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
MultiModalKwargs)
from vllm.multimodal import BatchedTensorInputs, MultiModalKwargs
from vllm.sampling_params import SamplingParams
from vllm.sequence import (CompletionSequenceGroupOutput, IntermediateTensors,
Logprob, SequenceData, SequenceGroupMetadata,
@@ -545,10 +544,6 @@ class HPUModelRunnerBase(ModelRunnerBase[TModelInputForHPU]):
]
gc.set_threshold(*requested_gc_thrs)
# Multi-modal data support
self.multi_modal_input_mapper = MULTIMODAL_REGISTRY \
.create_input_mapper(self.model_config)
self.skip_warmup = os.environ.get('VLLM_SKIP_WARMUP',
'false').lower() == 'true'
@@ -731,9 +726,8 @@ class HPUModelRunnerBase(ModelRunnerBase[TModelInputForHPU]):
# is always the first token in the sequence.
input_positions.append(list(range(context_len, seq_len)))
mm_data = seq_group_metadata.multi_modal_data
if mm_data:
mm_kwargs = self.multi_modal_input_mapper(mm_data)
mm_kwargs = seq_group_metadata.multi_modal_data
if mm_kwargs:
multi_modal_kwargs_list.append(mm_kwargs)
if seq_group_metadata.block_tables is None:
+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),
+3 -17
View File
@@ -15,8 +15,7 @@ from vllm.logger import init_logger
from vllm.model_executor import SamplingMetadata
from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.model_executor.model_loader.neuron import get_neuron_model
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
MultiModalKwargs)
from vllm.multimodal import BatchedTensorInputs, MultiModalKwargs
from vllm.sequence import IntermediateTensors, SequenceGroupMetadata
from vllm.utils import is_pin_memory_available, make_tensor_with_pad
from vllm.worker.model_runner_base import ModelRunnerBase, ModelRunnerInputBase
@@ -69,11 +68,6 @@ class NeuronModelRunner(ModelRunnerBase[ModelInputForNeuron]):
self.device = self.device_config.device
self.pin_memory = is_pin_memory_available()
# Multi-modal data support
self.mm_registry = MULTIMODAL_REGISTRY
self.multi_modal_input_mapper = self.mm_registry \
.create_input_mapper(self.model_config)
# Lazy initialization.
self.model: nn.Module # initialize after load_model.
@@ -149,16 +143,8 @@ class NeuronModelRunner(ModelRunnerBase[ModelInputForNeuron]):
assert len(block_table) == 1
input_block_ids.append(block_table[0])
mm_data = seq_group_metadata.multi_modal_data
if mm_data:
if self.mm_registry.has_processor(self.model_config):
mm_kwargs = mm_data
else:
mm_kwargs = self.multi_modal_input_mapper(
mm_data,
seq_group_metadata.mm_processor_kwargs,
)
mm_kwargs = seq_group_metadata.multi_modal_data
if mm_kwargs:
multi_modal_kwargs_list.append(mm_kwargs)
max_seq_len = max(seq_lens)
+2 -14
View File
@@ -188,20 +188,11 @@ class ModelInputForXPUBuilder(ModelRunnerInputBuilderBase[ModelInputForXPU]):
input_positions.extend(list(positions_range))
if seq_group_metadata.multi_modal_data:
# NOTE: mm_data only includes the subset of multi-modal items
# NOTE: mm_kwargs only includes the subset of multi-modal items
# that intersect with the current prefill positions.
mm_data, placeholder_maps = MultiModalPlaceholderMap \
mm_kwargs, placeholder_maps = MultiModalPlaceholderMap \
.from_seq_group(seq_group_metadata, positions_range)
if self.runner.mm_registry.has_processor(
self.runner.model_config):
mm_kwargs = mm_data
else:
mm_kwargs = self.runner.multi_modal_input_mapper(
mm_data,
seq_group_metadata.mm_processor_kwargs,
)
multi_modal_kwargs_list.append(mm_kwargs)
for modality, placeholder_map in placeholder_maps.items():
@@ -404,9 +395,6 @@ class XPUModelRunner(ModelRunnerBase[ModelInputForXPUWithSamplingMetadata]):
# 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 init_Model