[CPU] add indices in chunk_gated_delta_rule (#29267)

python/sglang/srt/layers/attention/linear/gdn_backend.py

@@ -503,7 +503,7 @@ class GDNAttnBackend(MambaAttnBackendBase):
                 query_start_loc=query_start_loc,
             )
 
-            if (is_npu() or is_cpu()) and last_recurrent_state is not None:
+            if is_npu() and last_recurrent_state is not None:
                 last_recurrent_state = last_recurrent_state.to(
                     ssm_states.dtype, copy=False
                 )

python/sglang/srt/layers/attention/linear/kernels/gdn_triton.py

@@ -141,9 +141,10 @@ class TritonGDNKernel(LinearAttnKernelBase):
     ) -> tuple:
         recurrent_state = ssm_states
         recurrent_state_indices_args = {"initial_state_indices": cache_indices}
-        if is_npu() or is_cpu():
+        if is_npu():
             recurrent_state = ssm_states[cache_indices]
             recurrent_state_indices_args = {}
+
         return chunk_gated_delta_rule(
             q=q,
             k=k,

python/sglang/srt/model_executor/cpu_graph_runner.py

@@ -534,7 +534,8 @@ def register_fake_ops(tp_size: int):
         cu_seqlens,
         head_first,
         use_qk_l2norm_in_kernel,
-        eps,
+        initial_state_indices,
+        eps=1e-6,
     ):
         output = torch.empty_like(value)
         assert initial_state is not None

python/sglang/srt/models/qwen3_5.py

@@ -1023,7 +1023,7 @@ class Qwen3_5AttentionDecoderLayer(nn.Module):
         attn_output = self.attn(q, k, v, forward_batch)
 
         if self.attn_output_gate:
-            if not _is_npu:
+            if not (_is_npu or _is_cpu):
                 attn_output = fused_sigmoid_mul(attn_output, gate, inplace=True)
             else:
                 gate_val = gate.reshape(gate.shape[0], -1) if gate.ndim == 3 else gate