A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 03:00 UTC
fix logging msg for block manager (#3701)
Simon Mo committed
Mar 28, 2024 at 16:29 UTC
4716a32dd4b413674c104b96882ea9965051c4ff
3 files changed
+5
-4
vllm/attention/selector.py
+3
-1
@@ -41,6 +41,8 @@ def _can_use_flash_attn(dtype: torch.dtype) -> bool:
41
try:
42
import flash_attn # noqa: F401
43
except ImportError:
44
- logger.info("flash_attn is not found.")
44
+ logger.info(
45
+ "Cannot use FlashAttention because the package is not found. "
46
+ "Please install it for better performance.")
47
return False
48
return True
vllm/core/block_manager_v1.py
+1
-2
@@ -230,13 +230,12 @@ class BlockSpaceManagerV1(BlockSpaceManager):
230
self.watermark_blocks = int(watermark * num_gpu_blocks)
231
232
if self.enable_caching:
233
- logger.info("enable automatic prefix caching")
233
+ logger.info("Automatic prefix caching is enabled.")
234
self.gpu_allocator = CachedBlockAllocator(Device.GPU, block_size,
235
num_gpu_blocks)
236
self.cpu_allocator = CachedBlockAllocator(Device.CPU, block_size,
237
num_cpu_blocks)
238
else:
239
- logger.info("disable automatic prefix caching")
239
self.gpu_allocator = UncachedBlockAllocator(
240
Device.GPU, block_size, num_gpu_blocks)
241
self.cpu_allocator = UncachedBlockAllocator(
vllm/model_executor/parallel_utils/pynccl_utils.py
+1
-1
@@ -10,7 +10,6 @@ logger = logging.getLogger(__name__)
10
try:
11
from vllm.model_executor.parallel_utils.pynccl import (NCCLCommunicator,
12
ncclGetVersion)
13
- logger.info(f"vLLM is using nccl=={ncclGetVersion()}")
13
except Exception as e:
14
# in non-NVIDIA environments, we can't import the nccl module
15
# e.g. when running on machines with AMD GPUs
@@ -40,6 +39,7 @@ def init_process_group(world_size: int, local_rank: int, rank: int,
39
init_method: str) -> None:
40
assert not is_initialized()
41
global comm
42
+ logger.info(f"vLLM is using nccl=={ncclGetVersion()}")
43
comm = NCCLCommunicator(init_method=init_method,
44
world_size=world_size,
45
local_rank=local_rank,