Feature/support longcat flash lite (#17838)

Co-authored-by: sunjiaqi11 <sunjiaqi11@meituan.com>
Co-authored-by: ispobock <ispobaoke@gmail.com>

python/sglang/srt/speculative/eagle_draft_cuda_graph_runner.py

@@ -105,7 +105,7 @@ class EAGLEDraftCudaGraphRunner:
         # Graph inputs
         with torch.device(model_runner.device):
             input_ids = torch.zeros((self.max_num_token,), dtype=torch.int64)
-            req_pool_indices = torch.zeros((self.max_bs,), dtype=torch.int32)
+            req_pool_indices = torch.zeros((self.max_bs,), dtype=torch.int64)
             out_cache_loc = torch.zeros(
                 (self.max_num_token * self.speculative_num_steps,),
                 dtype=self._cache_loc_dtype(),

python/sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py

@@ -110,7 +110,7 @@ class EAGLEDraftExtendCudaGraphRunner:
         # Graph inputs
         with torch.device(model_runner.device):
             input_ids = torch.zeros((self.max_num_token,), dtype=torch.int64)
-            req_pool_indices = torch.zeros((self.max_bs,), dtype=torch.int32)
+            req_pool_indices = torch.zeros((self.max_bs,), dtype=torch.int64)
             out_cache_loc = torch.ones(
                 (self.max_num_token,), dtype=self._cache_loc_dtype()
             )

python/sglang/srt/speculative/multi_layer_eagle_draft_extend_cuda_graph_runner.py

@@ -669,7 +669,7 @@ class MultiLayerEagleMultiStepDraftExtendCudaGraphRunner:
                 dtype=torch.int32,
             )
             self.cuda_graph_buffers["req_pool_indices"] = torch.zeros(
-                (self.max_bs,), dtype=torch.int32
+                (self.max_bs,), dtype=torch.int64
             )
             self.cuda_graph_buffers["accept_length"] = torch.full(
                 (self.max_bs,), 1, dtype=torch.int32