[Core] Refactor model loading code (#4097)

This commit is contained in:
Antoni Baum
2024-04-16 11:34:39 -07:00
committed by GitHub
parent 05434764cd
commit 69e1d2fb69
67 changed files with 1054 additions and 963 deletions
+7 -8
View File
@@ -9,9 +9,8 @@ import torch.nn as nn
from vllm.attention import (AttentionMetadata, AttentionMetadataPerStage,
get_attn_backend)
from vllm.config import (DeviceConfig, LoRAConfig, ModelConfig, ParallelConfig,
SchedulerConfig, TensorizerConfig,
VisionLanguageConfig)
from vllm.config import (DeviceConfig, LoadConfig, LoRAConfig, ModelConfig,
ParallelConfig, SchedulerConfig, VisionLanguageConfig)
from vllm.distributed import broadcast_tensor_dict, with_pynccl_for_all_reduce
from vllm.distributed.device_communicators import (custom_all_reduce,
pynccl_utils)
@@ -108,17 +107,17 @@ class ModelRunner:
parallel_config: ParallelConfig,
scheduler_config: SchedulerConfig,
device_config: DeviceConfig,
load_config: LoadConfig,
lora_config: Optional[LoRAConfig],
kv_cache_dtype: Optional[str] = "auto",
is_driver_worker: bool = False,
vision_language_config: Optional[VisionLanguageConfig] = None,
tensorizer_config: Optional[TensorizerConfig] = None,
):
self.model_config = model_config
self.parallel_config = parallel_config
self.scheduler_config = scheduler_config
self.lora_config = lora_config
self.tensorizer_config = tensorizer_config
self.load_config = load_config
self.is_driver_worker = is_driver_worker
# model_config can be None in tests/samplers/test_sampler.py.
@@ -156,13 +155,13 @@ class ModelRunner:
def load_model(self) -> None:
with CudaMemoryProfiler() as m:
self.model = get_model(
self.model_config,
self.device_config,
model_config=self.model_config,
device_config=self.device_config,
load_config=self.load_config,
lora_config=self.lora_config,
vision_language_config=self.vision_language_config,
parallel_config=self.parallel_config,
scheduler_config=self.scheduler_config,
tensorizer_config=self.tensorizer_config,
)
self.model_memory_usage = m.consumed_memory