[Bugfix] Fix broadcasting logic for multi_modal_kwargs (#6836)

This commit is contained in:
Cyrus Leung
2024-07-31 10:38:45 +08:00
committed by GitHub
parent da1f7cc12a
commit f230cc2ca6
16 changed files with 254 additions and 211 deletions
+7 -7
View File
@@ -4,8 +4,8 @@ import time
import warnings
import weakref
from dataclasses import dataclass
from typing import (TYPE_CHECKING, Any, Dict, List, Mapping, Optional, Set,
Tuple, Type, TypeVar, Union)
from typing import (TYPE_CHECKING, Any, Dict, List, Optional, Set, Tuple, Type,
TypeVar, Union)
import numpy as np
import torch
@@ -40,7 +40,7 @@ from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
from vllm.model_executor.models.interfaces import (supports_lora,
supports_vision)
from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensors,
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
MultiModalInputs)
from vllm.prompt_adapter.layers import PromptAdapterMapping
from vllm.prompt_adapter.request import PromptAdapterRequest
@@ -94,7 +94,7 @@ class ModelInputForGPU(ModelRunnerInputBase):
attn_metadata: Optional["AttentionMetadata"] = None
prompt_adapter_mapping: Optional[PromptAdapterMapping] = None
prompt_adapter_requests: Optional[Set[PromptAdapterRequest]] = None
multi_modal_kwargs: Optional[Mapping[str, BatchedTensors]] = None
multi_modal_kwargs: Optional[BatchedTensorInputs] = None
request_ids_to_seq_ids: Optional[Dict[str, List[int]]] = None
finished_requests_ids: Optional[List[str]] = None
virtual_engine: int = 0
@@ -608,8 +608,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
data.multi_modal_inputs for data in self.inter_data_list
if data.multi_modal_inputs is not None
]
multi_modal_kwargs = MultiModalInputs.batch(multi_modal_inputs_list,
device=self.runner.device)
multi_modal_kwargs = MultiModalInputs.batch(multi_modal_inputs_list)
return self.model_input_cls(
input_tokens=input_tokens_tensor,
@@ -1361,7 +1360,8 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
kv_caches=kv_caches,
attn_metadata=model_input.attn_metadata,
intermediate_tensors=intermediate_tensors,
**multi_modal_kwargs,
**MultiModalInputs.as_kwargs(multi_modal_kwargs,
device=self.device),
**seqlen_agnostic_kwargs)
# Compute the logits in the last pipeline stage.