[Fix] Init mamba related memory pools with torch.zeros (#10400)

python/sglang/srt/mem_cache/memory_pool.py

@@ -127,7 +127,7 @@ class MambaPool:
         if speculative_num_draft_tokens is not None:
             # Cache intermediate SSM states per draft token during target verify
             # Shape: [num_layers, size + 1, speculative_num_draft_tokens, HV, K, V]
-            intermediate_ssm_state_cache = torch.empty(
+            intermediate_ssm_state_cache = torch.zeros(
                 size=(
                     num_mamba_layers,
                     size + 1,
@@ -141,7 +141,7 @@ class MambaPool:
             )
             # Cache intermediate conv windows (last K-1 inputs) per draft token during target verify
             # Shape: [num_layers, size + 1, speculative_num_draft_tokens, dim, K-1]
-            intermediate_conv_window_cache = torch.empty(
+            intermediate_conv_window_cache = torch.zeros(
                 size=(
                     num_mamba_layers,
                     size + 1,
@@ -240,7 +240,7 @@ class HybridReqToTokenPool(ReqToTokenPool):
         self.mamba_map = {layer_id: i for i, layer_id in enumerate(mamba_layers)}
 
         self.device = device
-        self.req_index_to_mamba_index_mapping: torch.Tensor = torch.empty(
+        self.req_index_to_mamba_index_mapping: torch.Tensor = torch.zeros(
             size, dtype=torch.int32, device=self.device
         )