[NPU] Add MTP support for GLM-4.7-Flash (#28516)

python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py

@@ -2004,13 +2004,46 @@ class AscendAttnBackend(AttentionBackend):
                 self.speculative_num_draft_tokens,
             )
 
+            if (
+                self.q_head_num_padding is not None
+                and self.q_head_num_padding > self.tp_q_head_num
+            ):
+                nope_padding = torch.empty(
+                    [
+                        q_nope.shape[0],
+                        self.q_head_num_padding - self.tp_q_head_num,
+                        self.kv_lora_rank,
+                    ],
+                    dtype=(
+                        self.model_dtype
+                        if self.model_dtype is not None
+                        else torch.bfloat16
+                    ),
+                    device=q_nope.device,
+                )
+                rope_padding = torch.empty(
+                    [
+                        q_rope.shape[0],
+                        self.q_head_num_padding - self.tp_q_head_num,
+                        self.qk_rope_head_dim,
+                    ],
+                    dtype=(
+                        self.model_dtype
+                        if self.model_dtype is not None
+                        else torch.bfloat16
+                    ),
+                    device=q_rope.device,
+                )
+                q_nope = torch.cat([q_nope, nope_padding], dim=1).contiguous()
+                q_rope = torch.cat([q_rope, rope_padding], dim=1).contiguous()
+
             workspace = torch_npu._npu_fused_infer_attention_score_get_max_workspace(
                 q_nope,
                 c_kv_cache,
                 c_kv_cache,
                 query_rope=q_rope,
                 key_rope=k_rope_cache,
-                num_heads=layer.tp_q_head_num,
+                num_heads=self.q_head_num_padding,
                 num_key_value_heads=layer.tp_k_head_num,
                 input_layout="TND",
                 scale=layer.scaling,
@@ -2031,7 +2064,7 @@ class AscendAttnBackend(AttentionBackend):
                 c_kv_cache,
                 query_rope=q_rope,
                 key_rope=k_rope_cache,
-                num_heads=layer.tp_q_head_num,
+                num_heads=self.q_head_num_padding,
                 num_key_value_heads=layer.tp_k_head_num,
                 input_layout="TND",
                 scale=layer.scaling,
@@ -2046,6 +2079,7 @@ class AscendAttnBackend(AttentionBackend):
                 workspace=workspace,
                 out=[attn_output, softmax_lse],
             )
+            attn_output = attn_output[:, : layer.tp_q_head_num, :]
             attn_output = attn_output.view(-1, layer.tp_q_head_num * layer.v_head_dim)
             if (
                 not self.graph_mode

python/sglang/srt/models/glm4_moe_lite.py

@@ -547,6 +547,8 @@ class Glm4MoeLiteDecoderLayer(nn.Module):
     ) -> None:
 
         super().__init__()
+        # Required for MTP: Glm4MoeLiteModelNextN bypasses Glm4MoeLiteForCausalLM.__init__
+        config.moe_layer_freq = 1
         self.hidden_size = config.hidden_size
         self.config = config
         rope_theta, rope_scaling = get_rope_config(config)