[diffusion] feat: use LocalAttention for mistral3 encoder (#28176)

python/sglang/multimodal_gen/runtime/models/encoders/mistral_3.py

@@ -27,7 +27,6 @@ from transformers.masking_utils import (
     create_sliding_window_causal_mask,
 )
 from transformers.modeling_outputs import BaseModelOutputWithPast
-from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
 from transformers.models.mistral3.modeling_mistral3 import (
     Mistral3CausalLMOutputWithPast,
     Mistral3ModelOutputWithPast,
@@ -37,13 +36,13 @@ from transformers.models.mistral.modeling_mistral import (
     MistralRMSNorm,
     MistralRotaryEmbedding,
     apply_rotary_pos_emb,
-    eager_attention_forward,
 )
 
 from sglang.multimodal_gen.runtime.distributed import (
     get_tp_world_size,
     model_parallel_is_initialized,
 )
+from sglang.multimodal_gen.runtime.layers.attention import LocalAttention
 from sglang.multimodal_gen.runtime.layers.linear import (
     ColumnParallelLinear,
     RowParallelLinear,
@@ -52,7 +51,10 @@ from sglang.multimodal_gen.runtime.loader.weight_utils import default_weight_loa
 from sglang.multimodal_gen.runtime.managers.memory_managers.layerwise_offload import (
     LayerwiseOffloadableModuleMixin,
 )
-from sglang.multimodal_gen.runtime.platforms import current_platform
+from sglang.multimodal_gen.runtime.platforms import (
+    AttentionBackendEnum,
+    current_platform,
+)
 from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
 
 logger = init_logger(__name__)
@@ -107,6 +109,23 @@ def _make_row_linear(
     return nn.Linear(in_features, out_features, bias=bias)
 
 
+def _can_use_unmasked_causal_attention(
+    attention_mask: Optional[torch.Tensor],
+    config: MistralConfig,
+    past_key_values: Optional[Cache],
+) -> bool:
+    if (
+        getattr(config, "sliding_window", None) is not None
+        or past_key_values is not None
+    ):
+        return False
+    if attention_mask is None:
+        return True
+    if attention_mask.dim() != 2:
+        return False
+    return bool(torch.all(attention_mask > 0).item())
+
+
 def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     """
     This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep).
@@ -183,6 +202,18 @@ class MistralAttention(nn.Module):
             use_tensor_parallel=self.use_tensor_parallel,
         )
         self.is_causal = True
+        self.attn = LocalAttention(
+            self.num_heads,
+            self.head_dim,
+            self.num_key_value_heads,
+            softmax_scale=self.scaling,
+            causal=True,
+            supported_attention_backends={
+                AttentionBackendEnum.FA,
+                AttentionBackendEnum.TORCH_SDPA,
+            },
+            allow_cudnn_sdp=True,
+        )
 
     def forward(
         self,
@@ -224,32 +255,15 @@ class MistralAttention(nn.Module):
                 key_states, value_states, self.layer_idx, cache_kwargs
             )
 
-        attn_implementation = getattr(self.config, "_attn_implementation", None)
+        attn_output = self.attn(
-        attention_interface = eager_attention_forward
+            query_states.transpose(1, 2),
-        if attn_implementation and attn_implementation != "eager":
+            key_states.transpose(1, 2),
-            if hasattr(ALL_ATTENTION_FUNCTIONS, "get_interface"):
+            value_states.transpose(1, 2),
-                attention_interface = ALL_ATTENTION_FUNCTIONS.get_interface(
+            attn_mask=attention_mask,
-                    attn_implementation, eager_attention_forward
         )
-            else:
-                attention_interface = ALL_ATTENTION_FUNCTIONS[attn_implementation]
-        attn_output, attn_weights = attention_interface(
-            self,
-            query_states,
-            key_states,
-            value_states,
-            attention_mask,
-            dropout=0.0,
-            scaling=self.scaling,
-            sliding_window=getattr(
-                self.config, "sliding_window", None
-            ),  # main diff with Llama
-            **kwargs,
-        )
-
         attn_output = attn_output.reshape(*input_shape, -1).contiguous()
         attn_output = _linear_output(self.o_proj, attn_output)
-        return attn_output, attn_weights
+        return attn_output, None
 
 
 class MistralTPMLP(nn.Module):
@@ -388,6 +402,11 @@ class MistralModel(MistralPreTrainedModel):
 
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
+        if _can_use_unmasked_causal_attention(
+            attention_mask, self.config, past_key_values
+        ):
+            causal_mask = None
+        else:
             mask_function = (
                 create_causal_mask
                 if getattr(self.config, "sliding_window", None) is None
@@ -506,7 +525,7 @@ class Mistral3ForConditionalGeneration(nn.Module, LayerwiseOffloadableModuleMixi
         "^language_model.lm_head": "lm_head",
     }
     _tied_weights_keys = ["lm_head.weight"]
-    uses_sglang_forward_context = False
+    uses_sglang_forward_context = True
     layerwise_offload_dit_group_enabled = False
     layer_names = ["model.language_model.layers"]