xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-15 04:45:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Kernel] Add env variable to force flashinfer backend to enable tensor cores (#9497)

Browse Source 
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
Co-authored-by: Chih-Chieh Yang <chih.chieh.yang@ibm.com>
Co-authored-by: Cody Yu <hao.yu.cody@gmail.com>
This commit is contained in:
Thomas Parnell 2024-10-19 02:55:48 +02:00 committed by GitHub
parent d11bf435a0
commit 0c9a5258f9
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 11 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/attention/backends/flashinfer.py
View File

@ -17,6 +17,7 @@ except ImportError:
import torch import torch
import vllm.envs as envs
from vllm import _custom_ops as ops from vllm import _custom_ops as ops
from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl, from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
AttentionMetadata, AttentionMetadata,
@ -124,7 +125,8 @@ class FlashInferState(AttentionState):
self.runner.parallel_config)) self.runner.parallel_config))
num_kv_heads = self.runner.model_config.get_num_kv_heads( num_kv_heads = self.runner.model_config.get_num_kv_heads(
self.runner.parallel_config) self.runner.parallel_config)
use_tensor_cores = num_qo_heads // num_kv_heads > 4 use_tensor_cores = envs.VLLM_FLASHINFER_FORCE_TENSOR_CORES or (
num_qo_heads // num_kv_heads > 4)
self._decode_wrapper = BatchDecodeWithPagedKVCacheWrapper( self._decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
self._get_workspace_buffer(), self._get_workspace_buffer(),
"NHD", "NHD",
@ -183,7 +185,8 @@ class FlashInferState(AttentionState):
self.runner.parallel_config)) self.runner.parallel_config))
num_kv_heads = self.runner.model_config.get_num_kv_heads( num_kv_heads = self.runner.model_config.get_num_kv_heads(
self.runner.parallel_config) self.runner.parallel_config)
use_tensor_cores = num_qo_heads // num_kv_heads > 4 use_tensor_cores = envs.VLLM_FLASHINFER_FORCE_TENSOR_CORES or (
num_qo_heads // num_kv_heads > 4)
self._graph_decode_wrapper = \ self._graph_decode_wrapper = \
CUDAGraphBatchDecodeWithPagedKVCacheWrapper( CUDAGraphBatchDecodeWithPagedKVCacheWrapper(
self._graph_decode_workspace_buffer, _indptr_buffer, self._graph_decode_workspace_buffer, _indptr_buffer,

6
vllm/envs.py
View File

@ -32,6 +32,7 @@ if TYPE_CHECKING:
VLLM_ATTENTION_BACKEND: Optional[str] = None VLLM_ATTENTION_BACKEND: Optional[str] = None
VLLM_USE_FLASHINFER_SAMPLER: bool = False VLLM_USE_FLASHINFER_SAMPLER: bool = False
VLLM_USE_FLASHINFER_REJECTION_SAMPLER: bool = False VLLM_USE_FLASHINFER_REJECTION_SAMPLER: bool = False
VLLM_FLASHINFER_FORCE_TENSOR_CORES: bool = False
VLLM_PP_LAYER_PARTITION: Optional[str] = None VLLM_PP_LAYER_PARTITION: Optional[str] = None
VLLM_CPU_KVCACHE_SPACE: int = 0 VLLM_CPU_KVCACHE_SPACE: int = 0
VLLM_CPU_OMP_THREADS_BIND: str = "" VLLM_CPU_OMP_THREADS_BIND: str = ""
@ -286,6 +287,11 @@ environment_variables: Dict[str, Callable[[], Any]] = {
"VLLM_USE_FLASHINFER_SAMPLER": "VLLM_USE_FLASHINFER_SAMPLER":
lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_SAMPLER", "0"))), lambda: bool(int(os.getenv("VLLM_USE_FLASHINFER_SAMPLER", "0"))),
# If set, vllm will force flashinfer to use tensor cores;
# otherwise will use heuristic based on model architecture.
"VLLM_FLASHINFER_FORCE_TENSOR_CORES":
lambda: bool(int(os.getenv("VLLM_FLASHINFER_FORCE_TENSOR_CORES", "0"))),
# Pipeline stage partition strategy # Pipeline stage partition strategy
"VLLM_PP_LAYER_PARTITION": "VLLM_PP_LAYER_PARTITION":
lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None), lambda: os.getenv("VLLM_PP_LAYER_PARTITION", None),