Merged
ci(gen_ops_soc): 950/960dt 编译obj分组配置——新增matmul_all_reduce等5项，mega_moe 20→40、grad 5→10 #14790
景明Add on 12 days ago
ci(gen_ops_soc): 950/960dt 编译obj分组配置——新增matmul_all_reduce等5项，mega_moe 20→40、grad 5→10 #14790
Merged
Merged fromgitcode_lijd/ops-transformerfix_2intocann/ops-transformermaster
Total: 1 files changed+11-3
| @@ -52,15 +52,23 @@ group_op_dict = { | |||
| 52 | "ascend950": { | 52 | "ascend950": { |
| 53 | "fused_infer_attention_score": 20, | 53 | "fused_infer_attention_score": 20, |
| 54 | "flash_attention_score": 5, | 54 | "flash_attention_score": 5, |
| 55 | - "flash_attention_score_grad": 5, | 55 | + "flash_attention_score_grad": 10, |
| 56 | - "mega_moe": 20, | 56 | + "mega_moe": 40, |
| 57 | + "matmul_all_reduce": 40, | ||
| 58 | + "allto_all_matmul": 20, | ||
| 59 | + "matmul_reduce_scatter_v2": 20, | ||
| 57 | }, | 60 | }, |
| 58 | "ascend350": { | 61 | "ascend350": { |
| 59 | "flash_attention_score_grad": 5, | 62 | "flash_attention_score_grad": 5, |
| 60 | }, | 63 | }, |
| 61 | "ascend960dt": { | 64 | "ascend960dt": { |
| 62 | "fused_infer_attention_score": 20, | 65 | "fused_infer_attention_score": 20, |
| 63 | - "flash_attention_score_grad": 5, | 66 | + "flash_attention_score": 5, |
| 67 | + "flash_attention_score_grad": 10, | ||
| 68 | + "mega_moe": 40, | ||
| 69 | + "matmul_all_reduce": 40, | ||
| 70 | + "allto_all_matmul": 20, | ||
| 71 | + "matmul_reduce_scatter_v2": 20, | ||
| 64 | }, | 72 | }, |
| 65 | } | 73 | } |
| 66 | 74 | ||