[NPU] FIX CMB illusion of garbled characters acc problems, in prefix cache mtp scenarios. (#31659)

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

@@ -265,13 +265,22 @@ class AscendHybridLinearAttnBackend(HybridLinearAttnBackend):
         )
         last_steps = last_correct_step_indices.to(torch.int64)  # [N]
 
-        move_intermediate_cache(
+        # NPU: skip intermediate_ssm copy when accept_lens == 1.
-            ssm_states,
+        # The state at step 0 is the just-computed recurrent state, which
-            intermediate_state_cache,
+        # is already correct in the model's ssm buffer.  The intermediate
-            dst_indices_tensor,
+        # cache path exists for CUDA's per-step gather; on NPU the fused
-            src_indices_tensor,
+        # kernel may produce slightly different bfloat16 rounding that
-            last_steps,
+        # accumulates over thousands of decode steps and drifts into
-        )
+        # garbage output.
+        all_step0 = (last_steps == 0).all().item()
+        if not all_step0:
+            move_intermediate_cache(
+                ssm_states,
+                intermediate_state_cache,
+                dst_indices_tensor,
+                src_indices_tensor,
+                last_steps,
+            )
 
         draft_token_num = intermediate_state_cache.shape[2]
         if mamba_track_indices is not None: