rename fused_get_rope.cu (#3752)

* rename fused_get_rope.cu

* fix

* fix typos

* fix

* fix

custom_ops/gpu_ops/cpp_extensions.cc

@@ -1023,7 +1023,7 @@ PYBIND11_MODULE(fastdeploy_ops, m) {
   m.def("moe_expert_ffn", &MoeExpertFFNFunc, "moe export ffn function");
 
   /**
-   * moe/fused_moe/moe_ffn_wint2.cu
+   * moe/fused_moe/moe_expert_ffn_wint2.cu
    * moe_expert_ffn_wint2
    */
   m.def("moe_expert_ffn_wint2", &MoeExpertFFNWint2Func, "moe export ffn wint2 function");

custom_ops/setup_ops.py

@@ -204,7 +204,7 @@ if paddle.is_compiled_with_rocm():
         "gpu_ops/get_output_msg_with_topk.cc",
         "gpu_ops/save_output_msg_with_topk.cc",
         "gpu_ops/transfer_output.cc",
-        "gpu_ops/set_value_by_flags.cu",
+        "gpu_ops/set_value_by_flags_and_idx.cu",
         "gpu_ops/token_penalty_multi_scores.cu",
         "gpu_ops/stop_generation.cu",
         "gpu_ops/stop_generation_multi_ends.cu",
@@ -223,7 +223,7 @@ if paddle.is_compiled_with_rocm():
         "gpu_ops/speculate_decoding/speculate_get_output_padding_offset.cu",
         "gpu_ops/speculate_decoding/speculate_get_seq_lens_output.cu",
         "gpu_ops/speculate_decoding/speculate_save_output.cc",
-        "gpu_ops/speculate_decoding/speculate_set_value_by_flags.cu",
+        "gpu_ops/speculate_decoding/speculate_set_value_by_flags_and_idx.cu",
         "gpu_ops/speculate_decoding/speculate_step.cu",
         "gpu_ops/speculate_decoding/speculate_step_system_cache.cu",
         "gpu_ops/speculate_decoding/speculate_update_v3.cu",
@@ -261,7 +261,7 @@ elif paddle.is_compiled_with_cuda():
         "gpu_ops/save_output_msg_with_topk.cc",
         "gpu_ops/transfer_output.cc",
         "gpu_ops/set_mask_value.cu",
-        "gpu_ops/set_value_by_flags.cu",
+        "gpu_ops/set_value_by_flags_and_idx.cu",
         "gpu_ops/ngram_mask.cu",
         "gpu_ops/gather_idx.cu",
         "gpu_ops/get_output_ep.cc",
@@ -276,7 +276,7 @@ elif paddle.is_compiled_with_cuda():
         "gpu_ops/recover_decode_task.cu",
         "gpu_ops/step.cu",
         "gpu_ops/step_reschedule.cu",
-        "gpu_ops/fused_get_rope.cu",
+        "gpu_ops/fused_get_rotary_embedding.cu",
         "gpu_ops/get_padding_offset.cu",
         "gpu_ops/update_inputs.cu",
         "gpu_ops/update_inputs_beam.cu",
@@ -560,7 +560,7 @@ elif paddle.is_compiled_with_custom_device("iluvatar_gpu"):
                 "gpu_ops/save_output_msg_with_topk.cc",
                 "gpu_ops/transfer_output.cc",
                 "gpu_ops/get_padding_offset.cu",
-                "gpu_ops/set_value_by_flags.cu",
+                "gpu_ops/set_value_by_flags_and_idx.cu",
                 "gpu_ops/rebuild_padding.cu",
                 "gpu_ops/update_inputs.cu",
                 "gpu_ops/stop_generation_multi_ends.cu",
@@ -609,7 +609,7 @@ elif paddle.device.is_compiled_with_custom_device("metax_gpu"):
         "gpu_ops/transfer_output.cc",
         "gpu_ops/save_with_output.cc",
         "gpu_ops/set_mask_value.cu",
-        "gpu_ops/set_value_by_flags.cu",
+        "gpu_ops/set_value_by_flags_and_idx.cu",
         "gpu_ops/ngram_mask.cu",
         "gpu_ops/gather_idx.cu",
         "gpu_ops/get_output_ep.cc",
@@ -618,7 +618,7 @@ elif paddle.device.is_compiled_with_custom_device("metax_gpu"):
         "gpu_ops/stop_generation.cu",
         "gpu_ops/stop_generation_multi_ends.cu",
         "gpu_ops/set_flags.cu",
-        "gpu_ops/fused_get_rope.cu",
+        "gpu_ops/fused_get_rotary_embedding.cu",
         "gpu_ops/get_padding_offset.cu",
         "gpu_ops/update_inputs.cu",
         "gpu_ops/update_inputs_beam.cu",

fastdeploy/model_executor/models/deepseek_v3.py

@@ -733,7 +733,7 @@ class DeepseekV3ForCausalLM(ModelForCasualLM):
         return hidden_states
 
     def clear_grpah_opt_backend(self):
-        """Clear graph optimization bakcend, the captured cuda graph will be cleaned"""
+        """Clear graph optimization backend, the captured cuda graph will be cleaned"""
         self.model.clear_grpah_opt_backend(fd_config=self.fd_config)
 
 

fastdeploy/model_executor/models/ernie4_5_moe.py

@@ -620,7 +620,7 @@ class Ernie4_5_MoeForCausalLM(ModelForCasualLM):
         return hidden_states
 
     def clear_grpah_opt_backend(self):
-        """Clear graph optimization bakcend, the captured cuda graph will be cleaned"""
+        """Clear graph optimization backend, the captured cuda graph will be cleaned"""
         self.ernie.clear_grpah_opt_backend(fd_config=self.fd_config)
 
 

fastdeploy/model_executor/models/ernie4_5_vl/ernie4_5_vl_moe.py

@@ -721,7 +721,7 @@ class Ernie4_5_VLMoeForConditionalGeneration(ModelForCasualLM):
         return hidden_states
 
     def clear_grpah_opt_backend(self):
-        """Clear graph optimization bakcend, the captured cuda graph will be cleaned"""
+        """Clear graph optimization backend, the captured cuda graph will be cleaned"""
         self.ernie.clear_grpah_opt_backend(fd_config=self.fd_config)
 
 

fastdeploy/model_executor/models/qwen2.py

@@ -395,7 +395,7 @@ class Qwen2ForCausalLM(ModelForCasualLM):
         return hidden_states
 
     def clear_grpah_opt_backend(self):
-        """Clear graph optimization bakcend, the captured cuda graph will be cleaned"""
+        """Clear graph optimization backend, the captured cuda graph will be cleaned"""
         self.qwen2.clear_grpah_opt_backend(fd_config=self.fd_config)
 
 

fastdeploy/model_executor/models/qwen3.py

@@ -331,7 +331,7 @@ class Qwen3ForCausalLM(ModelForCasualLM):
         return hidden_states
 
     def clear_grpah_opt_backend(self):
-        """Clear graph optimization bakcend, the captured cuda graph will be cleaned"""
+        """Clear graph optimization backend, the captured cuda graph will be cleaned"""
         self.model.clear_grpah_opt_backend(fd_config=self.fd_config)
 
 

fastdeploy/model_executor/models/qwen3moe.py

@@ -452,7 +452,7 @@ class Qwen3MoeForCausalLM(ModelForCasualLM):
         return hidden_states
 
     def clear_grpah_opt_backend(self):
-        """Clear graph optimization bakcend, the captured cuda graph will be cleaned"""
+        """Clear graph optimization backend, the captured cuda graph will be cleaned"""
         self.model.clear_grpah_opt_backend(fd_config=self.fd_config)
 
 

tests/graph_optimization/test_cuda_graph_dynamic_subgraph.py

@@ -170,14 +170,14 @@ class TestCUDAGrpahSubgraph(unittest.TestCase):
         input_tensor1 = paddle.ones([8])
         forward_meta1 = ForwardMeta(input_ids=input_tensor1, ids_remove_padding=input_tensor1, step_use_cudagraph=True)
 
-        # Triger Capture
+        # Trigger Capture
         _ = test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
 
-        # Reaplay
+        # Replay
         _ = test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
         output1 = test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
 
-        # Corrent output
+        # Correct output
         output1_correct = test_model1.forward_correct(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
 
         assert (output1 == output1_correct).all()

tests/graph_optimization/test_cuda_graph_recapture.py

@@ -102,43 +102,43 @@ class TestCUDAGrpahRecapture(unittest.TestCase):
         input_tensor1 = paddle.ones([1, 32768])
         forward_meta1 = ForwardMeta(input_ids=input_tensor1, ids_remove_padding=input_tensor1, step_use_cudagraph=True)
 
-        # Corrent output
+        # Correct output
         self.output_correct = self.test_model1.forward_correct(
             ids_remove_padding=input_tensor1, forward_meta=forward_meta1
         )
 
-        # Capture and Destory
+        # Capture and Destroy
         self.capture_and_replay(input_tensor1, forward_meta1)
         self.recapture_and_replay(input_tensor1, forward_meta1)
 
     def capture_and_replay(self, input_tensor1, forward_meta1):
         """ """
-        # Triger Capture
+        # Trigger Capture
         print_gpu_memory_use(0, "before capture")
         output1 = self.test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
         print_gpu_memory_use(0, "after capture")
 
-        # Reaplay
+        # Replay
         output1 = self.test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
         assert (output1 == self.output_correct).all()
 
-        # Destory
+        # Destroy
         print_gpu_memory_use(0, "before destory")
         self.test_model1.clear_grpah_opt_backend()
         print_gpu_memory_use(0, "after destory")
 
     def recapture_and_replay(self, input_tensor1, forward_meta1):
         """ """
-        # Triger Capture
+        # Trigger Capture
         print_gpu_memory_use(0, "before recapture")
         output2 = self.test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
         print_gpu_memory_use(0, "after recapture")
 
-        # Reaplay
+        # Replay
         output2 = self.test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
         assert (output2 == self.output_correct).all()
 
-        # Destory
+        # Destroy
         print_gpu_memory_use(0, "before destory")
         self.test_model1.clear_grpah_opt_backend()
         print_gpu_memory_use(0, "after destory")

tests/graph_optimization/test_cuda_graph_spec_decode.py

@@ -117,14 +117,14 @@ class TestCUDAGrpahSpecDecode(unittest.TestCase):
         input_tensor1 = paddle.ones([1, 32768])
         forward_meta1 = ForwardMeta(input_ids=input_tensor1, ids_remove_padding=input_tensor1, step_use_cudagraph=True)
 
-        # Triger Capture
+        # Trigger Capture
         _ = test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
 
-        # Reaplay
+        # Replay
         _ = test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
         output1 = test_model1(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
 
-        # Corrent output
+        # Correct output
         output1_correct = test_model1.forward_correct(ids_remove_padding=input_tensor1, forward_meta=forward_meta1)
 
         assert (output1 == output1_correct).all()

tests/graph_optimization/test_static_graph_cuda_graph_split.py

@@ -104,14 +104,14 @@ class TestStaticGraphCUDAGraphSplit(unittest.TestCase):
         x = paddle.randint(32, shape=[1, 8])
         forward_meta1 = ForwardMeta(input_ids=x, ids_remove_padding=x, step_use_cudagraph=True)
 
-        # Triger Capture
+        # Trigger Capture
         _ = test_model1(x, forward_meta=forward_meta1)
 
-        # Reaplay
+        # Replay
         _ = test_model1(x, forward_meta=forward_meta1)
         output1 = test_model1(x, forward_meta=forward_meta1)
 
-        # Corrent output
+        # Correct output
         output1_correct = test_model1.forward_correct(x, forward_meta=forward_meta1)
 
         assert (output1 == output1_correct).all()