@@ -150,7 +150,7 @@ def test_cp_with_flash_attention(dtype, model, qkv_format, cp_comm_type):
150150 "cp_1_4" : ModelConfig (
151151 2 , 4096 , 12 , 128 , attn_bias_type = "post_scale_bias" , bias_shape = "bhss"
152152 ), # MHA
153- "cp_1_5" : ModelConfig (2 , 4096 , 12 , 128 , attn_mask_type = "causal" , window_size = (512 , 0 )), # MHA
153+ "cp_1_5" : ModelConfig (2 , 4096 , 12 , 128 , attn_mask_type = "causal" , window_size = (512 , 512 )), # MHA
154154 "cp_2_0" : ModelConfig (2 , 4096 , 12 , 128 , num_gqa_groups = 2 , attn_mask_type = "causal" ), # GQA
155155 "cp_2_1" : ModelConfig (2 , 4096 , 12 , 128 , num_gqa_groups = 2 ), # GQA
156156 "cp_2_2" : ModelConfig (
@@ -187,7 +187,7 @@ def test_cp_with_flash_attention(dtype, model, qkv_format, cp_comm_type):
187187 2 , 4096 , 12 , 128 , num_gqa_groups = 2 , attn_bias_type = "post_scale_bias"
188188 ), # GQA
189189 "cp_2_6" : ModelConfig (
190- 2 , 4096 , 12 , 128 , num_gqa_groups = 2 , attn_mask_type = "causal" , window_size = (512 , 0 )
190+ 2 , 4096 , 12 , 128 , num_gqa_groups = 2 , attn_mask_type = "causal" , window_size = (512 , 512 )
191191 ), # GQA
192192 "cp_3_0" : ModelConfig (2 , 4096 , 12 , 128 , attn_mask_type = "causal" , head_dim_v = 64 ), # MLA
193193 "cp_3_1" : ModelConfig (2 , 4096 , 12 , 128 , head_dim_v = 64 ), # MLA
0 commit comments