fix eb5 mtp(mix) (#6800)

fastdeploy/model_executor/forward_meta.py

@@ -264,6 +264,10 @@ class XPUForwardMeta(ForwardMeta):
     # for pd_disaggregation
     kv_signal_sender: Optional[paddle.Tensor] = None
 
+    hidden_states: Optional[paddle.Tensor] = None
+
+    is_draft: bool = False
+
     def copy_from(self, other: "XPUForwardMeta", skip_keys: Optional[list] = None):
         """
         Synchronize attributes from another XPUForwardMeta object

fastdeploy/model_executor/layers/backends/xpu/attention.py

@@ -190,7 +190,8 @@ class XPUAttentionBackend(AttentionBackend):
         else:
             q_norm_weight = None
             k_norm_weight = None
-
+        # draft model not use rope3d now
+        use_rope3d = self.rope_3d and not forward_meta.is_draft
         res = block_attn(
             qkv,
             forward_meta.caches[2 * layer.layer_id],
@@ -229,7 +230,7 @@ class XPUAttentionBackend(AttentionBackend):
             metadata.kv_signal_data_list[layer.layer_id],
             forward_meta.kv_signal_sender,
             layer.use_neox_rotary_style,
-            self.rope_3d,
+            use_rope3d,
         )
 
         return res

fastdeploy/model_executor/layers/sample/sampler.py

@@ -1068,7 +1068,7 @@ class SpeculativeSampler(nn.Layer):
             sampling_metadata.top_p,
             sampling_metadata.top_k,
             sampling_metadata.seed,
-            share_inputs["seq_lens_this_time"],
+            paddle.reshape(share_inputs["seq_lens_this_time"], shape=[-1]),
             paddle.reshape(share_inputs["seq_lens_encoder"], shape=[-1]),
         )
         _, sampled_token_ids = top_k_top_p_sampling(probs, top_p=top_p, top_k=top_k, topp_seed=topp_seed)