[fix]q_nope shape

Author: WANDY666

lightllm/models/deepseek2/layer_infer/transformer_layer_infer.py

@@ -567,7 +567,7 @@ def _token_gqa_decode_attention_mtp(
             q=q_rope.reshape(-1, self.tp_q_head_num_ * self.mtp_size, self.qk_rope_head_dim),
             k_cache=k_rope,
             v_cache=kv_nope,
-            qv=q_nope.reshape(-1, self.tp_q_head_num_ * self.mtp_size, self.qk_nope_head_dim),
+            qv=q_nope.reshape(-1, self.tp_q_head_num_ * self.mtp_size, self.kv_lora_rank),
             page_table=infer_state.page_table[self.mtp_size - 1 :: self.mtp_size],
             cache_seqlens=infer_state.b_seq_len[self.mtp_size - 1 :: self.mtp_size],
             cu_seqlens_q=infer_state.cu_seqlens_q,

lightllm/models/deepseek2/model.py

@@ -69,7 +69,7 @@ def __init__(self, kvargs):
         return
 
     def _init_inferstate_cls(self):
-        if get_env_start_args().enable_fa3:
+        if get_env_start_args().enable_fa3 or get_env_start_args().enable_fa3_mtp:
             self.infer_state_class = Deepseek2FlashAttentionStateInfo
         elif self.enable_flashinfer:
             self.infer_state_class = Deepseek2FlashInferStateInfo