mirror of
https://github.com/theroyallab/tabbyAPI.git
synced 2026-05-11 16:30:16 +00:00
20 lines
498 B
Python
20 lines
498 B
Python
import torch
|
|
|
|
|
|
def hardware_supports_flash_attn(gpu_device_list: list[int]):
|
|
"""
|
|
Check whether all GPUs in list support FA2
|
|
|
|
Compute capability < 8 is not supported by FA2
|
|
AMD is also unsupported until ROCm updates its FA2 fork
|
|
"""
|
|
|
|
min_compute_capability = min(
|
|
torch.cuda.get_device_capability(device=device_idx)[0] for device_idx in gpu_device_list
|
|
)
|
|
|
|
if torch.version.hip or min_compute_capability < 8:
|
|
return False
|
|
else:
|
|
return True
|