From 9884957c07ce0a98fad34ea0055803594c459cb1 Mon Sep 17 00:00:00 2001
From: Hexq0210 <893781835@qq.com>
Date: Tue, 10 Mar 2026 20:03:26 +0800
Subject: [PATCH] [NPU] Bugfix  for qwen35 on NPU (#19756)

---
 python/sglang/srt/layers/attention/linear/gdn_backend.py        | 2 ++
 python/sglang/srt/model_executor/model_runner_kv_cache_mixin.py | 2 +-
 2 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/python/sglang/srt/layers/attention/linear/gdn_backend.py b/python/sglang/srt/layers/attention/linear/gdn_backend.py
index 8b7851dfa..d15ad8fe5 100644
--- a/python/sglang/srt/layers/attention/linear/gdn_backend.py
+++ b/python/sglang/srt/layers/attention/linear/gdn_backend.py
@@ -33,11 +33,13 @@ if is_cuda():
 
     causal_conv1d_fn = causal_conv1d_fn_cuda
 elif is_npu():
+    from sgl_kernel_npu.fla.fused_gdn_gating import fused_gdn_gating_npu
     from sgl_kernel_npu.mamba.causal_conv1d import (
         causal_conv1d_fn_npu,
         causal_conv1d_update_npu,
     )
 
+    fused_gdn_gating = fused_gdn_gating_npu
     causal_conv1d_fn = causal_conv1d_fn_npu
     causal_conv1d_update = causal_conv1d_update_npu
 elif is_cpu():
diff --git a/python/sglang/srt/model_executor/model_runner_kv_cache_mixin.py b/python/sglang/srt/model_executor/model_runner_kv_cache_mixin.py
index 577508b1f..588da8034 100644
--- a/python/sglang/srt/model_executor/model_runner_kv_cache_mixin.py
+++ b/python/sglang/srt/model_executor/model_runner_kv_cache_mixin.py
@@ -440,7 +440,7 @@ class ModelRunnerKVCacheMixin:
 
         # Initialize token_to_kv_pool
         is_nsa_model = is_deepseek_nsa(self.model_config.hf_config)
-        if self.server_args.attention_backend == "ascend":
+        if self.server_args.attention_backend == "ascend" and not self.mambaish_config:
             if self.use_mla_backend:
                 from sglang.srt.hardware_backend.npu.memory_pool_npu import (
                     NPUMLATokenToKVPool,