Follow-up fixes for flashinfer 0.0.5 (#556)

python/sglang/srt/managers/controller/model_runner.py

@@ -67,7 +67,7 @@ class InputMetadata:
     flashinfer_prefill_wrapper: "BatchPrefillWithPagedKVCacheWrapper" = None
     flashinfer_decode_wrapper: "BatchDecodeWithPagedKVCacheWrapper" = None
 
-    def init_flashinfer_args(self, num_attention_heads, num_key_value_heads, head_dim):
+    def init_flashinfer_args(self, num_qo_heads, num_kv_heads, head_dim):
         self.kv_indptr = torch.zeros(
             (self.batch_size + 1,), dtype=torch.int32, device="cuda"
         )
@@ -102,8 +102,8 @@ class InputMetadata:
                 self.kv_indptr,
                 self.kv_indices,
                 self.kv_last_page_len,
-                num_attention_heads,
-                num_key_value_heads,
+                num_qo_heads,
+                num_kv_heads,
                 head_dim,
                 1
             )
@@ -113,8 +113,8 @@ class InputMetadata:
                 self.kv_indptr,
                 self.kv_indices,
                 self.kv_last_page_len,
-                num_attention_heads,
-                num_key_value_heads,
+                num_qo_heads,
+                num_kv_heads,
                 head_dim,
                 1,
                 pos_encoding_mode="NONE",
@@ -203,7 +203,7 @@ class InputMetadata:
         if global_server_args_dict.get("enable_flashinfer", False):
             ret.init_flashinfer_args(
                 model_runner.model_config.num_attention_heads // tp_size,
-                model_runner.model_config.num_key_value_heads // tp_size,
+                model_runner.model_config.get_num_kv_heads(tp_size),
                 model_runner.model_config.head_dim
             )
 
@@ -350,6 +350,15 @@ class ModelRunner:
                 BatchPrefillWithPagedKVCacheWrapper,
                 BatchDecodeWithPagedKVCacheWrapper,
             )
+            from flashinfer.decode import _grouped_size_compiled_for_decode_kernels
+
+            if not _grouped_size_compiled_for_decode_kernels(
+                self.model_config.num_attention_heads // self.tp_size,
+                self.model_config.get_num_kv_heads(self.tp_size)):
+                use_tensor_cores = True
+            else:
+                use_tensor_cores = False
+
             workspace_buffer = torch.empty(
                 32 * 1024 * 1024, dtype=torch.int8, device="cuda"
             )
@@ -357,8 +366,10 @@ class ModelRunner:
                 workspace_buffer, "NHD"
             )
             self.flashinfer_decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
-                workspace_buffer, "NHD"
+                workspace_buffer, "NHD", use_tensor_cores=use_tensor_cores
             )
+        else:
+            self.flashinfer_prefill_wrapper = self.flashinfer_decode_wrapper = None
 
     @torch.inference_mode()
     def forward_prefill(self, batch: Batch):