mirror of
https://github.com/deepseek-ai/DeepGEMM
synced 2025-04-29 23:40:21 +00:00
Minor fix
This commit is contained in:
parent
ed278eddd3
commit
bed67b234c
@ -105,8 +105,8 @@ def get_best_configs(m: int, n: int, k: int, num_groups: int, num_sms: int,
|
||||
# NOTES: less L2 cache usage and less GPU frequency drop
|
||||
num_waves = get_num_waves(best_block_m, best_block_n)
|
||||
num_min_sms = ceil_div(ceil_div(m, best_block_m) * ceil_div(n, best_block_n) * num_groups, num_waves)
|
||||
num_min_sms = ceil_div(max(num_min_sms, num_sms - 8), 2) * 2
|
||||
assert num_min_sms <= num_sms
|
||||
num_min_sms = ceil_div(max(num_min_sms, num_sms - 8), best_num_tma_multicast) * best_num_tma_multicast
|
||||
assert num_min_sms <= num_sms and is_tma_multicast_legal(n, best_block_n, best_num_tma_multicast, num_min_sms)
|
||||
|
||||
return num_min_sms, best_block_m, best_block_n, best_num_stages, best_num_tma_multicast, best_smem_size
|
||||
|
||||
|
Loading…
Reference in New Issue
Block a user