[Fix] Compatibility of window attention and cuda graph (#1090)

python/sglang/srt/layers/radix_attention.py

@@ -34,6 +34,7 @@ class RadixAttention(nn.Module):
         scaling: float,
         num_kv_heads: int,
         layer_id: int,
+        reuse: bool = False,
         sliding_window_size: int = -1,
         logit_cap: int = -1,
         v_head_dim: int = -1,
@@ -47,6 +48,7 @@ class RadixAttention(nn.Module):
         self.v_head_dim = v_head_dim if v_head_dim != -1 else head_dim
         self.scaling = scaling
         self.layer_id = layer_id
+        self.reuse = reuse
         self.sliding_window_size = sliding_window_size
 
         if (
@@ -127,8 +129,9 @@ class RadixAttention(nn.Module):
             if isinstance(prefill_wrapper_paged, list):
                 prefill_wrapper_paged = prefill_wrapper_paged[1]
 
-        if not input_metadata.flashinfer_use_ragged:
-            self.store_kv_cache(k, v, input_metadata)
+        if not input_metadata.flashinfer_use_ragged or self.reuse:
+            if not self.reuse:
+                self.store_kv_cache(k, v, input_metadata)
 
             o = prefill_wrapper_paged.forward(
                 q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
@@ -179,7 +182,8 @@ class RadixAttention(nn.Module):
             if isinstance(decode_wrapper, list):
                 decode_wrapper = decode_wrapper[1]
 
-        self.store_kv_cache(k, v, input_metadata)
+        if not self.reuse:
+            self.store_kv_cache(k, v, input_metadata)
 
         o = decode_wrapper.forward(
             q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
@@ -191,8 +195,10 @@ class RadixAttention(nn.Module):
         return o.view(-1, self.tp_q_head_num * self.head_dim)
 
     def forward(self, q, k, v, input_metadata: InputMetadata):
-        k = k.view(-1, self.tp_k_head_num, self.qk_head_dim)
-        v = v.view(-1, self.tp_v_head_num, self.v_head_dim)
+        if k is not None:
+            assert v is not None
+            k = k.view(-1, self.tp_k_head_num, self.qk_head_dim)
+            v = v.view(-1, self.tp_v_head_num, self.v_head_dim)
 
         if input_metadata.forward_mode == ForwardMode.EXTEND:
             return self.extend_forward(q, k, v, input_metadata)

python/sglang/srt/model_executor/cuda_graph_runner.py

@@ -107,9 +107,6 @@ class CudaGraphRunner:
         )
 
         # FlashInfer inputs
-        self.flashinfer_workspace_buffer = (
-            self.model_runner.flashinfer_workspace_buffers[0]
-        )
         self.flashinfer_kv_indptr = torch.zeros(
             (self.max_bs + 1,), dtype=torch.int32, device="cuda"
         )
@@ -121,6 +118,23 @@ class CudaGraphRunner:
         self.flashinfer_kv_last_page_len = torch.ones(
             (self.max_bs,), dtype=torch.int32, device="cuda"
         )
+        if model_runner.sliding_window_size is None:
+            self.flashinfer_workspace_buffer = (
+                self.model_runner.flashinfer_workspace_buffers[0]
+            )
+        else:
+            self.flashinfer_workspace_buffers = [
+                self.model_runner.flashinfer_workspace_buffers[0],
+                self.model_runner.flashinfer_workspace_buffers[2],
+            ]
+            self.flashinfer_kv_indptr = [
+                self.flashinfer_kv_indptr,
+                self.flashinfer_kv_indptr.clone(),
+            ]
+            self.flashinfer_kv_indices = [
+                self.flashinfer_kv_indices,
+                self.flashinfer_kv_indices.clone(),
+            ]
 
         self.compile_bs = [1, 2, 4, 8, 16, 24, 32] if use_torch_compile else []
 
@@ -171,15 +185,32 @@ class CudaGraphRunner:
             use_tensor_cores = True
         else:
             use_tensor_cores = False
-        flashinfer_decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
-            self.flashinfer_workspace_buffer,
-            "NHD",
-            use_cuda_graph=True,
-            use_tensor_cores=use_tensor_cores,
-            paged_kv_indptr_buffer=self.flashinfer_kv_indptr[: bs + 1],
-            paged_kv_indices_buffer=self.flashinfer_kv_indices,
-            paged_kv_last_page_len_buffer=self.flashinfer_kv_last_page_len[:bs],
-        )
+        if self.model_runner.sliding_window_size is None:
+            flashinfer_decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
+                self.flashinfer_workspace_buffer,
+                "NHD",
+                use_cuda_graph=True,
+                use_tensor_cores=use_tensor_cores,
+                paged_kv_indptr_buffer=self.flashinfer_kv_indptr[: bs + 1],
+                paged_kv_indices_buffer=self.flashinfer_kv_indices,
+                paged_kv_last_page_len_buffer=self.flashinfer_kv_last_page_len[:bs],
+            )
+        else:
+            flashinfer_decode_wrapper = []
+            for i in range(2):
+                flashinfer_decode_wrapper.append(
+                    BatchDecodeWithPagedKVCacheWrapper(
+                        self.flashinfer_workspace_buffers[i],
+                        "NHD",
+                        use_cuda_graph=True,
+                        use_tensor_cores=use_tensor_cores,
+                        paged_kv_indptr_buffer=self.flashinfer_kv_indptr[i][: bs + 1],
+                        paged_kv_indices_buffer=self.flashinfer_kv_indices[i],
+                        paged_kv_last_page_len_buffer=self.flashinfer_kv_last_page_len[
+                            :bs
+                        ],
+                    )
+                )
         update_flashinfer_indices(
             ForwardMode.DECODE,
             self.model_runner,

python/sglang/srt/model_executor/forward_batch_info.py

@@ -154,7 +154,6 @@ class InputMetadata:
         model_runner: "ModelRunner",
         batch: ScheduleBatch,
         forward_mode: ForwardMode,
-        sliding_window_size: Optional[int] = None,
     ):
         ret = cls(
             forward_mode=forward_mode,
@@ -198,7 +197,7 @@ class InputMetadata:
             ):
                 flashinfer_use_ragged = True
             ret.init_flashinfer_handlers(
-                model_runner, prefix_lens, flashinfer_use_ragged, sliding_window_size
+                model_runner, prefix_lens, flashinfer_use_ragged
             )
 
         return ret
@@ -221,7 +220,6 @@ class InputMetadata:
         model_runner,
         prefix_lens,
         flashinfer_use_ragged,
-        sliding_window_size=None,
     ):
         update_flashinfer_indices(
             self.forward_mode,
@@ -230,7 +228,6 @@ class InputMetadata:
             self.seq_lens,
             prefix_lens,
             flashinfer_use_ragged=flashinfer_use_ragged,
-            sliding_window_size=sliding_window_size,
         )
 
         (
@@ -254,7 +251,6 @@ def update_flashinfer_indices(
     prefix_lens,
     flashinfer_decode_wrapper=None,
     flashinfer_use_ragged=False,
-    sliding_window_size=None,
 ):
     """Init auxiliary variables for FlashInfer attention backend."""
     num_qo_heads = model_runner.model_config.num_attention_heads // model_runner.tp_size
@@ -262,7 +258,7 @@ def update_flashinfer_indices(
     head_dim = model_runner.model_config.head_dim
     batch_size = len(req_pool_indices)
 
-    if sliding_window_size is None:
+    if model_runner.sliding_window_size is None:
         if flashinfer_use_ragged:
             paged_kernel_lens = prefix_lens
         else:
@@ -335,7 +331,7 @@ def update_flashinfer_indices(
 
             if wrapper_id == 0 and forward_mode == ForwardMode.DECODE:
                 paged_kernel_lens = torch.minimum(
-                    paged_kernel_lens, torch.tensor(sliding_window_size)
+                    paged_kernel_lens, torch.tensor(model_runner.sliding_window_size)
                 )
                 kv_start_idx = seq_lens - paged_kernel_lens
             else:

python/sglang/srt/model_executor/model_runner.py

@@ -187,6 +187,11 @@ class ModelRunner:
             scheduler_config=None,
             cache_config=None,
         )
+        self.sliding_window_size = (
+            self.model.get_window_size()
+            if hasattr(self.model, "get_window_size")
+            else None
+        )
         self.is_generation = is_generation_model(
             self.model_config.hf_config.architectures
         )
@@ -295,12 +300,6 @@ class ModelRunner:
         return c
 
     def init_flashinfer(self):
-        self.sliding_window_size = (
-            self.model.get_window_size()
-            if hasattr(self.model, "get_window_size")
-            else None
-        )
-
         if self.server_args.disable_flashinfer:
             assert (
                 self.sliding_window_size is None
@@ -339,7 +338,7 @@ class ModelRunner:
                 use_tensor_cores=use_tensor_cores,
             )
         else:
-            workspace_buffers = torch.empty(
+            self.flashinfer_workspace_buffers = torch.empty(
                 4,
                 global_config.flashinfer_workspace_size,
                 dtype=torch.uint8,
@@ -351,17 +350,17 @@ class ModelRunner:
             for i in range(2):
                 self.flashinfer_prefill_wrapper_ragged.append(
                     BatchPrefillWithRaggedKVCacheWrapper(
-                        workspace_buffers[2 * i + 0], "NHD"
+                        self.flashinfer_workspace_buffers[2 * i + 0], "NHD"
                     )
                 )
                 self.flashinfer_prefill_wrapper_paged.append(
                     BatchPrefillWithPagedKVCacheWrapper(
-                        workspace_buffers[2 * i + 1], "NHD"
+                        self.flashinfer_workspace_buffers[2 * i + 1], "NHD"
                     )
                 )
                 self.flashinfer_decode_wrapper.append(
                     BatchDecodeWithPagedKVCacheWrapper(
-                        workspace_buffers[2 * i + 0],
+                        self.flashinfer_workspace_buffers[2 * i + 0],
                         "NHD",
                         use_tensor_cores=use_tensor_cores,
                     )
@@ -404,7 +403,6 @@ class ModelRunner:
             self,
             batch,
             ForwardMode.DECODE,
-            sliding_window_size=self.sliding_window_size,
         )
 
         return self.model.forward(
@@ -417,7 +415,6 @@ class ModelRunner:
             self,
             batch,
             forward_mode=ForwardMode.EXTEND,
-            sliding_window_size=self.sliding_window_size,
         )
         return self.model.forward(
             batch.input_ids, input_metadata.positions, input_metadata
@@ -429,7 +426,6 @@ class ModelRunner:
             self,
             batch,
             forward_mode=ForwardMode.EXTEND,
-            sliding_window_size=self.sliding_window_size,
         )
         return self.model.forward(
             batch.input_ids,

python/sglang/srt/server_args.py

@@ -453,10 +453,12 @@ class ServerArgs:
             logger.info(
                 f"When using sliding window in gemma-2, disable radix_cache, regex_jump_forward, and turn on flashinfer."
             )
+            # FIXME: compatibility with radix attention
             self.disable_radix_cache = True
+            # FIXME: compatibility with jump forward
             self.disable_regex_jump_forward = True
             self.disable_flashinfer = False
-            self.disable_cuda_graph = True
+            # FIXME: compatibility with chunked prefill
             self.chunked_prefill_size = None
 
 

python/sglang/test/runners.py

@@ -36,7 +36,7 @@ DEFAULT_PROMPTS = [
 ]
 
 dirpath = os.path.dirname(__file__)
-with open(os.path.join(dirpath, "long_prompt"), "r") as f:
+with open(os.path.join(dirpath, "long_prompt.txt"), "r") as f:
     long_prompt = f.read()
 DEFAULT_PROMPTS.append(long_prompt)