Use FlashAttention for multi_query_kv_attention (#4)

This commit is contained in:
Woosuk Kwon
2023-03-01 21:13:08 -08:00
committed by GitHub
parent 0deacbce6e
commit 3e9f991d6a
4 changed files with 107 additions and 33 deletions
+5 -2
View File
@@ -9,10 +9,12 @@ parser = argparse.ArgumentParser(description='CacheFlow server')
parser.add_argument('--model', type=str, default='facebook/opt-125m', help='model name')
parser.add_argument('--num-nodes', type=int, default=1, help='number of nodes')
parser.add_argument('--num-workers', type=int, default=1, help='number of workers per node')
parser.add_argument('--block-size', type=int, default=8, help='token block size')
parser.add_argument('--block-size', type=int, default=8, choices=[8, 16], help='token block size')
# TODO(woosuk): Add an analytical model to determine the maximum number of GPU/CPU blocks.
parser.add_argument('--num-gpu-blocks', type=int, default=1024, help='number of GPU blocks (per GPU)')
parser.add_argument('--num-cpu-blocks', type=int, default=256, help='number of CPU blocks (per GPU)')
parser.add_argument('--num-cpu-blocks', type=int, default=32, help='number of CPU blocks (per GPU)')
# NOTE(woosuk): If FlashAttention is used, the float data type is not supported.
parser.add_argument('--dtype', type=str, default='half', choices=['half', 'float'], help='data type')
args = parser.parse_args()
@@ -27,6 +29,7 @@ def main():
block_size=args.block_size,
num_gpu_blocks=args.num_gpu_blocks,
num_cpu_blocks=args.num_cpu_blocks,
dtype=args.dtype,
)
controllers.append(controller)