[Cherry-Pick][BugFix] Fix batch_size derivation and relax shape checks in SM90 flash_mask_attn (#7216)

Author: xiaoxiaohehe001

custom_ops/gpu_ops/flash_mask_attn/flash_mask_attn.cu

@@ -54,7 +54,6 @@ void DispatchFlashAttentionMask(const paddle::Tensor& q_input,
   PADDLE_ENFORCE(k_token_num == v_input.dims()[0], "Unmatched shape");
   PADDLE_ENFORCE(head_dim == 128, "Unmatched shape");
   PADDLE_ENFORCE(batch_size > 0, "Unmatched shape");
-  PADDLE_ENFORCE(batch_size == seq_len_encoder.dims()[0], "Unmatched shape");
   PADDLE_ENFORCE(batch_size == cu_seq_k.dims()[0] - 1, "Unmatched shape");
 
   constexpr int kBlockM = 128;

fastdeploy/model_executor/layers/attention/flash_mask_attn_backend.py

@@ -307,7 +307,7 @@ def forward_mixed(
                     q,
                     k,
                     v,
-                    forward_meta.cu_seqlens_q,
+                    forward_meta.cu_seqlens_q[: forward_meta.attn_cu_seqlens_k.shape[0]],
                     forward_meta.attn_cu_seqlens_k,
                     forward_meta.seq_lens_encoder,
                     res_encoder,