biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Restore support for larger block sizes (#11259)

Browse Source 
Signed-off-by: Konrad Zawora <kzawora@habana.ai>
This commit is contained in:
Konrad Zawora
2024-12-18 01:39:07 +01:00
committed by GitHub
parent bf8717ebae
commit 866fa4550d
2 changed files with 8 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/config.py
View File

@@ -917,6 +917,10 @@ class CacheConfig:
raise ValueError( raise ValueError(
"GPU memory utilization must be less than 1.0. Got " "GPU memory utilization must be less than 1.0. Got "
f"{self.gpu_memory_utilization}.") f"{self.gpu_memory_utilization}.")
if (current_platform.is_cuda() and self.block_size is not None
and self.block_size > 32):
raise ValueError("CUDA Paged Attention kernel only supports "
f"block sizes up to 32. Got {self.block_size}.")
def _verify_cache_dtype(self) -> None: def _verify_cache_dtype(self) -> None:
if self.cache_dtype == "auto": if self.cache_dtype == "auto":

6
vllm/engine/arg_utils.py
View File

@@ -424,10 +424,12 @@ class EngineArgs:
parser.add_argument('--block-size', parser.add_argument('--block-size',
type=int, type=int,
default=EngineArgs.block_size, default=EngineArgs.block_size,
choices=[8, 16, 32], choices=[8, 16, 32, 64, 128],
help='Token block size for contiguous chunks of ' help='Token block size for contiguous chunks of '
'tokens. This is ignored on neuron devices and ' 'tokens. This is ignored on neuron devices and '
'set to max-model-len') 'set to max-model-len. On CUDA devices, '
'only block sizes up to 32 are supported. '
'On HPU devices, block size defaults to 128.')
parser.add_argument( parser.add_argument(
"--enable-prefix-caching", "--enable-prefix-caching",