DeepSeek/DeepGEMM
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepGEMM synced 2025-06-26 23:15:49 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Minor fix

Browse Source
This commit is contained in:
Chenggang Zhao 2025-02-28 09:21:35 +08:00
parent fbec9e5eee
commit 6e10cba207
1 changed files with 1 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
deep_gemm/jit_kernels/gemm.py
View File

@ -79,12 +79,10 @@ def get_best_configs(m: int, n: int, k: int, num_groups: int, num_sms: int,
elif num_waves < best_num_waves:
success = True
elif num_waves == best_num_waves:
div_n = bool(128 % block_n)
best_div_n = bool(128 % best_block_n)
# Check last wave utilization
util = get_last_wave_util(block_m, block_n)
best_util = get_last_wave_util(best_block_m, best_block_n)
success = util > best_util or (util == best_util and (block_m > best_block_m or block_m == best_block_m and (div_n < best_div_n or div_n == best_div_n and block_n < best_block_n)))
success = util > best_util or (util == best_util and (block_m > best_block_m or (block_m == best_block_m and block_n < best_block_n)))
best_block_m, best_block_n = (block_m, block_n) if success else (best_block_m, best_block_n)
assert best_block_m is not None and best_block_n is not None