[NPU] Fix DeepEP LL dispatch BF16 flag and skip triton kernel on NPU for Qwen3.5 (#23815)

Co-authored-by: iridiumine <iridiumine@users.noreply.github.com>

python/sglang/srt/layers/moe/token_dispatcher/deepep.py

@@ -620,9 +620,13 @@ class _DeepEPDispatcherImplLowLatency(_DeepEPDispatcherImplBase):
         input_global_scale = self.quant_config.get("input_global_scale", None)
         if input_global_scale is not None:
             use_nvfp4 = True
-        elif not get_moe_runner_backend().is_flashinfer_cutedsl():
+        elif not get_moe_runner_backend().is_flashinfer_cutedsl() and (
+            not _is_npu or not envs.SGLANG_DEEPEP_BF16_DISPATCH.get()
+        ):
             # flashinfer_cutedsl expects BF16 dispatch when NVFP4 dispatch is
             # off; its kernel quantizes to NVFP4 internally.
+            # SGLANG_DEEPEP_BF16_DISPATCH forces BF16 dispatch for NPU
+            # where INT8 input + BF16 weight GMM is not supported.
             use_fp8 = True
 
         # round_scale / use_ue8m0 are FP8-DeepGEMM specific; they cause DeepEP

python/sglang/srt/models/qwen3_5.py

@@ -464,7 +464,11 @@ class Qwen3_5GatedDeltaNet(nn.Module):
             hidden_states
         )
 
-        if self.num_v_heads // self.num_k_heads in [1, 2, 4] and not _is_cpu:
+        if (
+            self.num_v_heads // self.num_k_heads in [1, 2, 4]
+            and not _is_cpu
+            and not _is_npu
+        ):
             mixed_qkv, z, b, a = fused_qkvzba_split_reshape_cat_contiguous(
                 projected_states_qkvz,
                 projected_states_ba,
@@ -488,6 +492,8 @@ class Qwen3_5GatedDeltaNet(nn.Module):
             query, key, value, z, b, a = self.fix_query_key_value_ordering(
                 projected_states_qkvz, projected_states_ba
             )
+            b = b.contiguous()
+            a = a.contiguous()
 
             query, key, value = map(
                 lambda x: x.reshape(x.shape[0], -1), (query, key, value)