fix-hopper-flashinfer

ktransformers/models/custom_modeling_deepseek_v2.py

@@ -50,7 +50,8 @@ class KDeepseekV2ForCausalLM(DeepseekV2PreTrainedModel):
         self.wrapper = flashinfer.mla.BatchMLAPagedAttentionWrapper(
             self.workspace_buffer, use_cuda_graph=use_cuda_graph,
             qo_indptr=self.qo_indptr_buf,kv_indptr=self.paged_kv_indptr_buf,
-            kv_indices=self.paged_kv_indices_buf,kv_len_arr=self.paged_kv_len_buf
+            kv_indices=self.paged_kv_indices_buf,kv_len_arr=self.paged_kv_len_buf,
+            backend = "fa2",
         )
 
     def batch_embeddings(self, batch: ForwardBatchInput, device="cuda:0"):

ktransformers/models/custom_modeling_deepseek_v3.py

@@ -54,7 +54,8 @@ class KDeepseekV3ForCausalLM(DeepseekV3PreTrainedModel):
             self.workspace_buffer, use_cuda_graph=use_cuda_graph,
             qo_indptr=self.qo_indptr_buf,kv_indptr=self.paged_kv_indptr_buf,
             kv_indices=self.paged_kv_indices_buf,kv_len_arr=self.paged_kv_len_buf,
-            bsz_tensor=self.bsz_tensor_buf
+            bsz_tensor=self.bsz_tensor_buf,
+            backend = "fa2",
         )
 
     def batch_embeddings(self, batch: ForwardBatchInput, device="cuda:0"):

ktransformers/operators/flashinfer_wrapper.py

@@ -100,7 +100,8 @@ class MLAWrapper():
             kv_indptr=self.kv_indptr_buf,
             kv_indices=self.kv_indices_buf,
             kv_len_arr=self.kv_len_arr_buf,
-            bsz_tensor=self.batch_size_tensor_buf
+            bsz_tensor=self.batch_size_tensor_buf,
+            backend = "fa2",
         )
         self.need_plan = True