[Submodule] Change FlashInfer to import (#156)

.gitmodules

@@ -1,3 +0,0 @@
-[submodule "3rdparty/flashinfer"]
-	path = 3rdparty/flashinfer
-	url = https://github.com/flashinfer-ai/flashinfer.git

docs/flashinfer.md

@@ -5,13 +5,15 @@ It can be used in SGLang runtime to accelerate attention computation.
 
 ### Install flashinfer
 
-Note: The compilation can take a very long time.
+You can install flashinfer via pip as follows for CUDA 12.1.
 
 ```bash
-git submodule update --init --recursive
+pip install flashinfer -i https://flashinfer.ai/whl/cu121/
-pip install 3rdparty/flashinfer/python
 ```
 
+You can look for other CUDA versions in https://github.com/flashinfer-ai/flashinfer?tab=readme-ov-file#installation. If there is no desire version for your environment,
+please build it from source (the compilation takes a long time).
+
 ### Run a Server With Flashinfer Mode
 
 Add `--model-mode flashinfer` argument to enable flashinfer when launching a server.

python/sglang/srt/layers/radix_attention.py

@@ -98,12 +98,7 @@ class RadixAttention(nn.Module):
 
         o = input_metadata.prefill_wrapper.forward(
             q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
-            input_metadata.qo_indptr,
             input_metadata.token_to_kv_pool.kv_data[self.layer_id],
-            input_metadata.kv_indptr,
-            input_metadata.kv_indices,
-            input_metadata.kv_last_page_len,
-            allow_fp16_qk_reduction=True,
         )
 
         return o.view(-1, self.tp_q_head_num * self.head_dim)
@@ -114,9 +109,6 @@ class RadixAttention(nn.Module):
         o = input_metadata.decode_wrapper.forward(
             q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
             input_metadata.token_to_kv_pool.kv_data[self.layer_id],
-            input_metadata.kv_indptr,
-            input_metadata.kv_indices,
-            input_metadata.kv_last_page_len,
         )
 
         return o.view(-1, self.tp_q_head_num * self.head_dim)

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

@@ -90,6 +90,11 @@ class InputMetadata:
     decode_wrapper = None
 
     def init_flashinfer_args(self, tp_size):
+        from flashinfer import (
+            BatchDecodeWithPagedKVCacheWrapper,
+            BatchPrefillWithPagedKVCacheWrapper,
+        )
+
         self.kv_indptr = torch.zeros(
             (self.batch_size + 1,), dtype=torch.int32, device="cuda"
         )
@@ -107,11 +112,7 @@ class InputMetadata:
             (self.batch_size,), dtype=torch.int32, device="cuda"
         )
 
-        from flashinfer.ops import (
+        workspace_buffer = torch.empty(32 * 1024 * 1024, dtype=torch.int8, device="cuda")
-            BatchDecodeWithPagedKVCacheWrapper,
-            BatchPrefillWithPagedKVCacheWrapper,
-        )
-
         if (
             self.forward_mode == ForwardMode.PREFILL
             or self.forward_mode == ForwardMode.EXTEND
@@ -120,19 +121,21 @@ class InputMetadata:
                 (self.batch_size + 1,), dtype=torch.int32, device="cuda"
             )
             self.qo_indptr[1:] = torch.cumsum(self.extend_seq_lens, dim=0)
-            self.prefill_wrapper = BatchPrefillWithPagedKVCacheWrapper()
+            self.prefill_wrapper = BatchPrefillWithPagedKVCacheWrapper(workspace_buffer, "NHD")
             self.prefill_wrapper.begin_forward(
                 self.qo_indptr,
-                self.batch_size,
+                self.kv_indptr,
+                self.kv_indices,
+                self.kv_last_page_len,
                 self.model_runner.model_config.num_attention_heads // tp_size,
                 self.model_runner.model_config.num_key_value_heads // tp_size,
             )
         else:
-            self.decode_wrapper = BatchDecodeWithPagedKVCacheWrapper()
+            self.decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(workspace_buffer, "NHD")
             self.decode_wrapper.begin_forward(
                 self.kv_indptr,
+                self.kv_indices,
                 self.kv_last_page_len,
-                self.batch_size,
                 self.model_runner.model_config.num_attention_heads // tp_size,
                 self.model_runner.model_config.num_key_value_heads // tp_size,
                 self.model_runner.model_config.head_dim,