[Misc] Set block size at initialization & Fix test_model_runner (#4705)

This commit is contained in:
Woosuk Kwon
2024-05-09 09:04:59 -07:00
committed by GitHub
parent 190bc838e1
commit 0ee535b294
5 changed files with 64 additions and 104 deletions
+9 -12
View File
@@ -4,8 +4,9 @@ import torch
from torch import nn
from vllm.attention import AttentionMetadata, get_attn_backend
from vllm.config import (DeviceConfig, LoadConfig, LoRAConfig, ModelConfig,
ParallelConfig, SchedulerConfig, VisionLanguageConfig)
from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
ModelConfig, ParallelConfig, SchedulerConfig,
VisionLanguageConfig)
from vllm.distributed import broadcast_tensor_dict
from vllm.logger import init_logger
from vllm.model_executor import SamplingMetadata
@@ -26,6 +27,7 @@ class CPUModelRunner:
parallel_config: ParallelConfig,
scheduler_config: SchedulerConfig,
device_config: DeviceConfig,
cache_config: CacheConfig,
load_config: LoadConfig,
lora_config: Optional[LoRAConfig],
vision_language_config: Optional[VisionLanguageConfig],
@@ -39,27 +41,22 @@ class CPUModelRunner:
self.scheduler_config = scheduler_config
# Currently, CPU worker doesn't support chunked prefill.
assert self.scheduler_config.chunked_prefill_enabled is False
self.device_config = device_config
self.cache_config = cache_config
self.lora_config = lora_config
self.vision_language_config = vision_language_config
self.load_config = load_config
self.is_driver_worker = is_driver_worker
# model_config can be None in tests/samplers/test_sampler.py.
# FIXME(woosuk): This is a hack to make the tests work. Refactor this.
self.sliding_window = (model_config.get_sliding_window()
if model_config is not None else None)
self.device_config = (device_config
if device_config is not None else DeviceConfig())
self.device = self.device_config.device
self.kv_cache_dtype = kv_cache_dtype
self.attn_backend = get_attn_backend(
self.model_config.dtype if model_config is not None else None)
self.sliding_window = model_config.get_sliding_window()
self.block_size = cache_config.block_size
self.attn_backend = get_attn_backend(self.model_config.dtype)
# Lazy initialization.
self.model: nn.Module # Set after init_Model
self.block_size: int # Set after initial profiling.
def load_model(self) -> None:
self.model = get_model(