wehub-resource-sync
44 lines
1.3 KiB
Python
44 lines
1.3 KiB
Python
from typing import *
|
|
|
|
CONV = 'flex_gemm'
|
|
DEBUG = False
|
|
ATTN = 'flash_attn'
|
|
|
|
def __from_env():
|
|
import os
|
|
|
|
global CONV
|
|
global DEBUG
|
|
global ATTN
|
|
|
|
env_sparse_conv_backend = os.environ.get('SPARSE_CONV_BACKEND')
|
|
env_sparse_debug = os.environ.get('SPARSE_DEBUG')
|
|
env_sparse_attn_backend = os.environ.get('SPARSE_ATTN_BACKEND')
|
|
if env_sparse_attn_backend is None:
|
|
env_sparse_attn_backend = os.environ.get('ATTN_BACKEND')
|
|
|
|
if env_sparse_conv_backend is not None and env_sparse_conv_backend in ['none', 'spconv', 'torchsparse', 'flex_gemm']:
|
|
CONV = env_sparse_conv_backend
|
|
if env_sparse_debug is not None:
|
|
DEBUG = env_sparse_debug == '1'
|
|
if env_sparse_attn_backend is not None and env_sparse_attn_backend in ['xformers', 'flash_attn', 'flash_attn_3', 'flash_attn_4', 'sdpa']:
|
|
ATTN = env_sparse_attn_backend
|
|
|
|
print(f"[SPARSE] Conv backend: {CONV}; Attention backend: {ATTN}")
|
|
|
|
|
|
__from_env()
|
|
|
|
|
|
def set_conv_backend(backend: Literal['none', 'spconv', 'torchsparse', 'flex_gemm']):
|
|
global CONV
|
|
CONV = backend
|
|
|
|
def set_debug(debug: bool):
|
|
global DEBUG
|
|
DEBUG = debug
|
|
|
|
def set_attn_backend(backend: Literal['xformers', 'flash_attn', 'flash_attn_3', 'flash_attn_4', 'sdpa']):
|
|
global ATTN
|
|
ATTN = backend
|