[Fix] Window attention compatible with RadixAttention and chunked prefill (#1112)

python/sglang/srt/layers/radix_attention.py

@@ -15,6 +15,8 @@ limitations under the License.
 
 """Radix attention."""
 
+from typing import Optional
+
 import torch
 from flashinfer.cascade import merge_state
 from torch import nn
@@ -34,8 +36,7 @@ class RadixAttention(nn.Module):
         scaling: float,
         num_kv_heads: int,
         layer_id: int,
-        reuse: bool = False,
-        sliding_window_size: int = -1,
+        sliding_window_size: Optional[int] = None,
         logit_cap: int = -1,
         v_head_dim: int = -1,
     ):
@@ -48,8 +49,7 @@ 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
+        self.sliding_window_size = sliding_window_size if sliding_window_size else -1
 
         if (
             not global_server_args_dict.get("disable_flashinfer", False)
@@ -118,16 +118,16 @@ class RadixAttention(nn.Module):
 
     def extend_forward_flashinfer(self, q, k, v, input_metadata: InputMetadata):
         # using two wrappers is unnecessary in the current PR, but are prepared for future PRs
-        prefill_wrapper_ragged = input_metadata.flashinfer_prefill_wrapper_ragged
         prefill_wrapper_paged = input_metadata.flashinfer_prefill_wrapper_paged
-        if self.sliding_window_size != -1 or self.reuse:
+        if self.sliding_window_size != -1:
             prefill_wrapper_paged = prefill_wrapper_paged[0]
         else:
             if isinstance(prefill_wrapper_paged, list):
                 prefill_wrapper_paged = prefill_wrapper_paged[1]
 
-        if not input_metadata.flashinfer_use_ragged or self.reuse:
-            if not self.reuse:
+        if not input_metadata.flashinfer_use_ragged:
+            if k is not None:
+                assert v is not None
                 self.store_kv_cache(k, v, input_metadata)
 
             o = prefill_wrapper_paged.forward(
@@ -139,21 +139,20 @@ class RadixAttention(nn.Module):
                 logits_soft_cap=self.logit_cap,
             )
         else:
-            o1, s1 = prefill_wrapper_ragged.forward_return_lse(
-                q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
-                k.contiguous().view(-1, self.tp_k_head_num, self.head_dim),
-                v.contiguous().view(-1, self.tp_v_head_num, self.head_dim),
-                causal=True,
-                sm_scale=self.scaling,
-                window_left=self.sliding_window_size,
-                logits_soft_cap=self.logit_cap,
+            o1, s1 = (
+                input_metadata.flashinfer_prefill_wrapper_ragged.forward_return_lse(
+                    q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
+                    k.contiguous().view(-1, self.tp_k_head_num, self.head_dim),
+                    v.contiguous().view(-1, self.tp_v_head_num, self.head_dim),
+                    causal=True,
+                    sm_scale=self.scaling,
+                    logits_soft_cap=self.logit_cap,
+                )
             )
 
             if input_metadata.extend_no_prefix:
                 o = o1
             else:
-                # TODO window attention + radix attention will come up in next PR
-                assert self.sliding_window_size == -1
                 o2, s2 = prefill_wrapper_paged.forward_return_lse(
                     q.contiguous().view(-1, self.tp_q_head_num, self.head_dim),
                     input_metadata.token_to_kv_pool.get_kv_buffer(self.layer_id),
@@ -179,7 +178,8 @@ class RadixAttention(nn.Module):
             if isinstance(decode_wrapper, list):
                 decode_wrapper = decode_wrapper[1]
 
-        if not self.reuse:
+        if k is not None:
+            assert v is not None
             self.store_kv_cache(k, v, input_metadata)
 
         o = decode_wrapper.forward(

python/sglang/srt/model_executor/forward_batch_info.py

@@ -194,6 +194,7 @@ class InputMetadata:
             if (
                 forward_mode != ForwardMode.DECODE
                 and int(torch.sum(ret.seq_lens)) > 4096
+                and model_runner.sliding_window_size is None
             ):
                 flashinfer_use_ragged = True
             ret.init_flashinfer_handlers(
@@ -322,22 +323,25 @@ def update_flashinfer_indices(
                 1,
             )
     else:
+        # window attention use paged only
         kv_last_page_len = torch.ones((batch_size,), dtype=torch.int32, device="cuda")
         for wrapper_id in range(2):
-            if flashinfer_use_ragged and wrapper_id == 1:
-                # full attention use ragged+paged
-                paged_kernel_lens = prefix_lens
+            if wrapper_id == 0:
+                if forward_mode == ForwardMode.DECODE:
+                    paged_kernel_lens = torch.minimum(
+                        seq_lens, torch.tensor(model_runner.sliding_window_size + 1)
+                    )
+                else:
+                    paged_kernel_lens = torch.minimum(
+                        seq_lens,
+                        torch.tensor(model_runner.sliding_window_size)
+                        + seq_lens
+                        - prefix_lens,
+                    )
             else:
-                # window attention use paged only
                 paged_kernel_lens = seq_lens
 
-            if wrapper_id == 0 and forward_mode == ForwardMode.DECODE:
-                paged_kernel_lens = torch.minimum(
-                    paged_kernel_lens, torch.tensor(model_runner.sliding_window_size)
-                )
-                kv_start_idx = seq_lens - paged_kernel_lens
-            else:
-                kv_start_idx = torch.zeros(batch_size, dtype=torch.int32, device="cuda")
+            kv_start_idx = seq_lens - paged_kernel_lens
 
             kv_indptr = torch.zeros((batch_size + 1,), dtype=torch.int32, device="cuda")
             kv_indptr[1:] = torch.cumsum(paged_kernel_lens, dim=0)
@@ -376,17 +380,6 @@ def update_flashinfer_indices(
                 )
                 qo_indptr[1:] = torch.cumsum(seq_lens - prefix_lens, dim=0)
 
-                if flashinfer_use_ragged and wrapper_id == 1:
-                    model_runner.flashinfer_prefill_wrapper_ragged.end_forward()
-                    model_runner.flashinfer_prefill_wrapper_ragged.begin_forward(
-                        qo_indptr,
-                        qo_indptr,
-                        num_qo_heads,
-                        num_kv_heads,
-                        head_dim,
-                    )
-
-                # cached part
                 model_runner.flashinfer_prefill_wrapper_paged[wrapper_id].end_forward()
                 model_runner.flashinfer_prefill_wrapper_paged[wrapper_id].begin_forward(
                     qo_indptr,

python/sglang/srt/model_executor/model_runner.py

@@ -334,11 +334,7 @@ class ModelRunner:
                 dtype=torch.uint8,
                 device="cuda",
             )
-            self.flashinfer_prefill_wrapper_ragged = (
-                BatchPrefillWithRaggedKVCacheWrapper(
-                    self.flashinfer_workspace_buffer, "NHD"
-                )
-            )
+            self.flashinfer_prefill_wrapper_ragged = None
             self.flashinfer_prefill_wrapper_paged = []
             self.flashinfer_decode_wrapper = []
             for i in range(2):

python/sglang/srt/models/gemma2.py

@@ -213,7 +213,7 @@ class Gemma2Attention(nn.Module):
             self.scaling,
             num_kv_heads=self.num_kv_heads,
             layer_id=layer_idx,
-            sliding_window_size=get_window_size(config) if use_sliding_window else -1,
+            sliding_window_size=get_window_size(config) if use_sliding_window else None,
             logit_cap=self.config.attn_logit_softcapping,
         )
 

python/sglang/srt/server_args.py

@@ -450,16 +450,8 @@ class ServerArgs:
             self.dp_size > 1 and self.node_rank is not None
         ), "multi-node data parallel is not supported"
         if "gemma-2" in self.model_path.lower():
-            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
+            logger.info(f"When using sliding window in gemma-2, turn on flashinfer.")
             self.disable_flashinfer = False
-            # FIXME: compatibility with chunked prefill
-            self.chunked_prefill_size = None
 
 
 @dataclasses.dataclass