Feat/support encoder model (like bert)

python/sglang/srt/layers/attention/torch_native_backend.py

@@ -6,6 +6,7 @@
 from torch.nn.functional import scaled_dot_product_attention
 
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
+from sglang.srt.layers.radix_attention import AttentionType
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 if TYPE_CHECKING:
@@ -202,6 +203,10 @@ def forward_extend(
         q_ = q.view(-1, layer.tp_q_head_num, layer.qk_head_dim)
         o_ = o.view(-1, layer.tp_q_head_num, layer.v_head_dim)
 
+        causal = True
+        if layer.is_cross_attention or layer.attn_type == AttentionType.ENCODER_ONLY:
+            causal = False
+
         self._run_sdpa_forward_extend(
             q_,
             o_,
@@ -214,7 +219,7 @@ def forward_extend(
             forward_batch.extend_seq_lens,
             scaling=layer.scaling,
             enable_gqa=use_gqa,
-            causal=not layer.is_cross_attention,
+            causal=causal,
         )
         return o
 

python/sglang/srt/layers/attention/triton_backend.py

@@ -10,6 +10,7 @@
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
 from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_triton
 from sglang.srt.layers.dp_attention import get_attention_tp_size
+from sglang.srt.layers.radix_attention import AttentionType
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
 from sglang.srt.utils import get_bool_env_var, get_device_core_count
 
@@ -528,6 +529,10 @@ def forward_extend(
                 layer, forward_batch.out_cache_loc, k, v
             )
 
+        causal = True
+        if layer.attn_type == AttentionType.ENCODER_ONLY:
+            causal = False
+
         self.extend_attention_fwd(
             q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
             k.contiguous(),
@@ -539,6 +544,7 @@ def forward_extend(
             self.forward_metadata.kv_indptr,
             self.forward_metadata.kv_indices,
             self.forward_metadata.custom_mask,
+            causal,
             self.forward_metadata.mask_indptr,
             self.forward_metadata.max_extend_len,
             layer.scaling,

python/sglang/srt/layers/attention/triton_ops/extend_attention.py

@@ -74,6 +74,7 @@ def _fwd_kernel(
     BLOCK_M: tl.constexpr,
     BLOCK_N: tl.constexpr,
     USE_CUSTOM_MASK: tl.constexpr,
+    IS_CAUSAL: tl.constexpr,
     SKIP_PREFIX_CUSTOM_MASK: tl.constexpr,
     STORE_TRANSPOSE: tl.constexpr,
 ):
@@ -129,6 +130,7 @@ def _fwd_kernel(
     for start_n in range(0, cur_seq_len_prefix, BLOCK_N):
         start_n = tl.multiple_of(start_n, BLOCK_N)
         mask_n = (start_n + offs_n) < cur_seq_len_prefix
+
         offs_kv_loc = tl.load(
             kv_indices + cur_seq_kv_start_idx + start_n + offs_n, mask=mask_n, other=0
         )
@@ -196,7 +198,11 @@ def _fwd_kernel(
 
     # stage 2: compute the triangle part
 
-    cur_block_m_end = tl.minimum(cur_seq_len_extend, (cur_block_m + 1) * BLOCK_M)
+    cur_block_m_end = (
+        cur_seq_len_extend
+        if not IS_CAUSAL
+        else tl.minimum(cur_seq_len_extend, (cur_block_m + 1) * BLOCK_M)
+    )
     for start_n in range(0, cur_block_m_end, BLOCK_N):
         start_n = tl.multiple_of(start_n, BLOCK_N)
         mask_n = (start_n + offs_n) < cur_block_m_end
@@ -243,12 +249,15 @@ def _fwd_kernel(
             )
             custom_mask &= mask_m[:, None] & mask_n[None, :]
             qk = tl.where(custom_mask, qk, float("-inf"))
-        else:
+        elif IS_CAUSAL:
             mask_causual = (cur_block_m * BLOCK_M + offs_m[:, None]) >= (
                 start_n + offs_n[None, :]
             )
             mask_causual &= mask_m[:, None] & mask_n[None, :]
             qk = tl.where(mask_causual, qk, float("-inf"))
+        else:
+            mask_non_causal = mask_m[:, None] & mask_n[None, :]
+            qk = tl.where(mask_non_causal, qk, float("-inf"))
 
         n_e_max = tl.maximum(tl.max(qk, 1), e_max)
         re_scale = tl.exp(e_max - n_e_max)
@@ -299,6 +308,7 @@ def extend_attention_fwd(
     kv_indptr,
     kv_indices,
     custom_mask,
+    is_causal,
     mask_indptr,
     max_len_extend,
     sm_scale=None,
@@ -411,6 +421,7 @@ def extend_attention_fwd(
         Lq=Lq,
         Lv=Lv,
         USE_CUSTOM_MASK=USE_CUSTOM_MASK,
+        IS_CAUSAL=is_causal,
         SKIP_PREFIX_CUSTOM_MASK=SKIP_PREFIX_CUSTOM_MASK,
         STORE_TRANSPOSE=_is_hip,
         num_warps=num_warps,

python/sglang/srt/layers/radix_attention.py

@@ -13,6 +13,7 @@
 # ==============================================================================
 """Radix attention."""
 
+from enum import Enum
 from typing import Optional
 
 from torch import nn
@@ -22,6 +23,18 @@
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 
+class AttentionType(Enum):
+    """
+    Attention type.
+    Use string to be compatible with `torch.compile`.
+    """
+
+    # Decoder attention between previous layer Q/K/V
+    DECODER = "decoder"
+    # Encoder attention between previous layer Q/K/V
+    ENCODER_ONLY = "encoder_only"
+
+
 class RadixAttention(nn.Module):
     """
     The attention layer implementation.
@@ -39,6 +52,7 @@ def __init__(
         sliding_window_size: int = -1,
         is_cross_attention: bool = False,
         quant_config: Optional[QuantizationConfig] = None,
+        attn_type=AttentionType.DECODER,
         prefix: str = "",
         use_irope: bool = False,
     ):
@@ -64,6 +78,7 @@ def __init__(
             self.quant_method = quant_config.get_quant_method(self, prefix=prefix)
         if self.quant_method is not None:
             self.quant_method.create_weights(self)
+        self.attn_type = attn_type
 
     def forward(
         self,