Skip to content

Commit 3419f6d

Browse files
committed
add test for triton mla
1 parent 4e6720c commit 3419f6d

3 files changed

Lines changed: 944 additions & 5 deletions

File tree

fastdeploy/model_executor/models/deepseek_v3.py

Lines changed: 1 addition & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -489,11 +489,7 @@ def forward(
489489
attn_out = attn_out * ((F.softsign(gate_out) + 1.0) / 2.0)
490490
else:
491491
raise NotImplementedError(f"{gated_attn_act} not implemented")
492-
if attn_out is None:
493-
attn_out = paddle.zeros(
494-
[hidden_states.shape[0], self.num_attention_heads_tp * self.v_head_dim],
495-
dtype=hidden_states.dtype,
496-
)
492+
497493
output = self.o_proj(attn_out)
498494
return output
499495

0 commit comments

Comments
 (0)