sglang 0.4.6.post2__py3-none-any.whl → 0.4.6.post4__py3-none-any.whl

sglang/srt/models/llama4.py

@@ -30,9 +30,9 @@ from sglang.srt.distributed import (
 from sglang.srt.layers.dp_attention import (
     dp_gather_partial,
     dp_scatter,
-    get_attention_dp_size,
     get_attention_tp_rank,
     get_attention_tp_size,
+    get_local_attention_dp_size,
 )
 from sglang.srt.layers.layernorm import RMSNorm
 from sglang.srt.layers.linear import (
@@ -46,7 +46,11 @@ from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import get_rope
 from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding
 from sglang.srt.managers.schedule_batch import global_server_args_dict
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
+from sglang.srt.model_executor.forward_batch_info import (
+    ForwardBatch,
+    ForwardMode,
+    PPProxyTensors,
+)
 from sglang.srt.models.llama import LlamaForCausalLM, LlamaMLP
 from sglang.srt.utils import add_prefix, fast_topk, get_compiler_backend, make_layers
 
@@ -81,6 +85,7 @@ class Llama4MoE(nn.Module):
         super().__init__()
         self.tp_size = get_tensor_model_parallel_world_size()
         self.top_k = config.num_experts_per_tok
+        self.device_module = torch.get_device_module()
 
         intermediate_size_moe = config.intermediate_size
         self.router = ReplicatedLinear(
@@ -113,7 +118,25 @@ class Llama4MoE(nn.Module):
             reduce_results=False,  # We need to do scatter before reduce
         )
 
-    def forward(self, hidden_states):
+    def forward(self, hidden_states, forward_batch: ForwardBatch):
+        shared_out, routed_out = self._forward_core(
+            hidden_states, forward_batch.forward_mode
+        )
+
+        out_aD = routed_out + shared_out
+
+        if self.tp_size > 1:
+            out_aD = tensor_model_parallel_all_reduce(out_aD)
+
+        return out_aD
+
+    def _forward_core(self, hidden_states, forward_mode: ForwardMode):
+        if hidden_states.shape[0] < 4:
+            return self._forward_core_shared_routed_overlap(hidden_states)
+        else:
+            return self._forward_core_normal(hidden_states)
+
+    def _forward_core_normal(self, hidden_states):
         # router_scores: [num_tokens, num_experts]
         router_logits, _ = self.router(hidden_states)
         shared_out = self.shared_expert(hidden_states)
@@ -121,12 +144,35 @@ class Llama4MoE(nn.Module):
             hidden_states=hidden_states,
             router_logits=router_logits,
         )
-        out_aD = routed_out + shared_out
+        return shared_out, routed_out
 
-        if self.tp_size > 1:
-            out_aD = tensor_model_parallel_all_reduce(out_aD)
+    def _forward_core_shared_routed_overlap(self, hidden_states):
+        alt_stream = _get_or_create_alt_stream(self.device_module)
 
-        return out_aD
+        alt_stream.wait_stream(self.device_module.current_stream())
+
+        shared_out = self.shared_expert(hidden_states)
+
+        with self.device_module.stream(alt_stream):
+            # router_scores: [num_tokens, num_experts]
+            router_logits, _ = self.router(hidden_states)
+            routed_out = self.experts(
+                hidden_states=hidden_states,
+                router_logits=router_logits,
+            )
+        self.device_module.current_stream().wait_stream(alt_stream)
+
+        return shared_out, routed_out
+
+
+_alt_stream = None
+
+
+def _get_or_create_alt_stream(device_module):
+    global _alt_stream
+    if _alt_stream is None:
+        _alt_stream = device_module.Stream()
+    return _alt_stream
 
 
 class Llama4Attention(nn.Module):
@@ -152,7 +198,6 @@ class Llama4Attention(nn.Module):
         self.use_rope = int((layer_id + 1) % 4 != 0)
         self.use_qk_norm = config.use_qk_norm and self.use_rope
 
-        self.dp_size = get_attention_dp_size()
         attn_tp_rank = get_attention_tp_rank()
         attn_tp_size = get_attention_tp_size()
 
@@ -296,7 +341,7 @@ class Llama4DecoderLayer(nn.Module):
         rope_theta = config.rope_theta
         rope_scaling = config.rope_scaling
         max_position_embeddings = config.max_position_embeddings
-        self.dp_size = get_attention_dp_size()
+        self.local_dp_size = get_local_attention_dp_size()
         self.attn_tp_size = get_attention_tp_size()
         self.attn_tp_rank = get_attention_tp_rank()
 
@@ -359,7 +404,7 @@ class Llama4DecoderLayer(nn.Module):
         # Gather
         if get_tensor_model_parallel_world_size() > 1:
             # all gather and all reduce
-            if self.dp_size != 1:
+            if self.local_dp_size != 1:
                 if self.attn_tp_rank == 0:
                     hidden_states += residual
                 hidden_states, local_hidden_states = (
@@ -380,11 +425,11 @@ class Llama4DecoderLayer(nn.Module):
             )
 
         # Fully Connected
-        hidden_states = self.feed_forward(hidden_states)
+        hidden_states = self.feed_forward(hidden_states, forward_batch)
 
-        # TODO(ch-wan): ues reduce-scatter in MLP to avoid this scatter
+        # TODO(ch-wan): use reduce-scatter in MLP to avoid this scatter
         # Scatter
-        if self.dp_size != 1:
+        if self.local_dp_size != 1:
             # important: forward batch.gathered_buffer is used both after scatter and after gather.
             # be careful about this!
             hidden_states, global_hidden_states = (

sglang/srt/models/llava.py

@@ -15,7 +15,8 @@
 
 import math
 import re
-from typing import Iterable, List, Optional, Tuple
+from functools import lru_cache
+from typing import Dict, Iterable, List, Optional, Tuple, Type, Union
 
 import numpy as np
 import torch
@@ -28,10 +29,18 @@ from transformers import (
     Qwen2Config,
     SiglipVisionModel,
 )
+from transformers.models.auto.modeling_auto import AutoModel, AutoModelForCausalLM
 from transformers.models.llava.modeling_llava import LlavaMultiModalProjector
 
+# leave till last and symbol only in case circular import
+import sglang.srt.models as sgl_models
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
-from sglang.srt.managers.schedule_batch import Modality, MultimodalInputs
+from sglang.srt.managers.mm_utils import general_mm_embed_routine
+from sglang.srt.managers.schedule_batch import (
+    Modality,
+    MultimodalDataItem,
+    MultimodalInputs,
+)
 from sglang.srt.mm_utils import (
     get_anyres_image_grid_shape,
     unpad_image,
@@ -42,7 +51,7 @@ from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.llama import LlamaForCausalLM
 from sglang.srt.models.mistral import MistralForCausalLM
 from sglang.srt.models.qwen2 import Qwen2ForCausalLM
-from sglang.srt.utils import add_prefix, flatten_nested_list
+from sglang.srt.utils import add_prefix, flatten_nested_list, logger
 
 
 class LlavaBaseForCausalLM(nn.Module):
@@ -114,7 +123,16 @@ class LlavaBaseForCausalLM(nn.Module):
         image_inputs.image_offsets = offset_list
         return input_ids
 
-    def encode_images(self, pixel_values: torch.Tensor) -> torch.Tensor:
+    def encode_images(
+        self, pixel_values: Union[torch.Tensor, List[torch.Tensor]]
+    ) -> torch.Tensor:
+        """
+        encode images by vision tower and multimodal projector
+        Args:
+            pixel_values: torch.Tensor or List[torch.Tensor]: each tensor for an input image
+        Returns:
+            torch.Tensor: encoded image features from the input image; if multiple, flattened by seq_len axis
+        """
         image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
         # NOTE: This is not memory efficient. (output_hidden_states=True) will save all the hidden stated.
 
@@ -583,4 +601,229 @@ class LlavaMistralForCausalLM(LlavaBaseForCausalLM):
             )
 
 
-EntryClass = [LlavaLlamaForCausalLM, LlavaQwenForCausalLM, LlavaMistralForCausalLM]
+class LlavaForConditionalGeneration(LlavaBaseForCausalLM):
+    """
+    An adaptor class to enable support for multiple mmlm such as mistral-community/pixtral-12b
+    It follows the structure of (vision_tower, multi_modal_projector, language_model)
+
+    Once a model config is loaded, text_config and vision_config will be extracted, and
+    LlavaForConditionalGeneration will load the language_model and vision_tower models
+    according to config.
+    """
+
+    MULTIMODAL_PROJECTOR_TYPE = LlavaMultiModalProjector
+
+    def pad_input_ids(self, input_ids: List[int], image_inputs: MultimodalInputs):
+        if hasattr(self.vision_tower, "pad_input_ids"):
+            return self.vision_tower.pad_input_ids(input_ids, image_inputs)
+        else:
+            return super().pad_input_ids(input_ids, image_inputs)
+
+    def _get_sgl_model_cls(self, config, auto_model_type: Type[AutoModel] = AutoModel):
+        """
+        Get the SGLang model implementation class according to config.
+
+        Args:
+            config: The config object of the model.
+            auto_model_type: The type of the auto model.
+
+        Returns:
+            The SGLang model implementation class.
+        """
+        config_cls_name = config.__class__.__name__
+        arch_name_mapping = self._config_cls_name_to_arch_name_mapping(auto_model_type)
+        if arch := arch_name_mapping.get(config_cls_name):
+            if isinstance(arch, tuple):
+                arch = arch[0]
+                logger.warning(
+                    f"Multiple {auto_model_type.__name__} models found for submodule config `{config_cls_name}`, defaulting to [0]: {arch.__name__}"
+                )
+            try:
+                return sgl_models.registry.ModelRegistry.resolve_model_cls(arch)[0]
+            except Exception as e:
+                raise ValueError(
+                    f"{auto_model_type.__name__} found a corresponding model `{arch}` for config class `{config_cls_name}`, but failed to load it from SGLang ModelRegistry. \n{e}"
+                )
+        else:
+            raise ValueError(
+                f"{auto_model_type.__name__} cannot find a corresponding model for config class `{config_cls_name}`"
+            )
+
+    @lru_cache
+    def _config_cls_name_to_arch_name_mapping(
+        self, auto_model_type: Type[AutoModel]
+    ) -> Dict[str, str]:
+        mapping = {}
+        for config_cls, archs in auto_model_type._model_mapping.items():
+            if isinstance(archs, tuple):
+                mapping[config_cls.__name__] = tuple(arch.__name__ for arch in archs)
+            else:
+                mapping[config_cls.__name__] = archs.__name__
+        return mapping
+
+    def __init__(
+        self,
+        config: LlavaConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        assert hasattr(config, "text_config")
+        assert hasattr(config, "vision_config")
+        self.config = config
+        self.text_config = config.text_config
+        self.vision_config = config.vision_config
+
+        if not hasattr(self.config, "vocab_size"):
+            self.config.vocab_size = self.config.text_config.vocab_size
+        if not hasattr(self.config, "image_aspect_ratio"):
+            self.config.image_aspect_ratio = "anyres"
+        if not hasattr(self.config, "image_grid_pinpoints"):
+            # from transformers.models.llava_onevision.configuration_llava_onevision import LlavaOnevisionConfig
+            # self.config.image_grid_pinpoints = LlavaOnevisionConfig().image_grid_pinpoints
+            self.config.image_grid_pinpoints = [
+                [96, 96],
+                [224, 224],
+                [384, 384],
+                [512, 512],
+                [768, 768],
+                [1024, 1024],
+            ]
+        if not hasattr(self.config, "mm_patch_merge_type"):
+            self.config.mm_patch_merge_type = "flat"
+        if not hasattr(self.config, "image_token_index"):
+            self.config.image_token_index = 10
+        if not hasattr(self.config, "projector_hidden_act"):
+            self.config.projector_hidden_act = "gelu"
+
+        self.vision_feature_layer = getattr(config, "vision_feature_layer", -1)
+        self.vision_feature_select_strategy = getattr(
+            config, "vision_feature_select_strategy", "full"
+        )
+        self.image_size = self.config.vision_config.image_size
+        self.patch_size = self.config.vision_config.patch_size
+
+        self.mm_patch_merge_type = config.mm_patch_merge_type
+        self.image_aspect_ratio = config.image_aspect_ratio
+        self.image_grid_pinpoints = config.image_grid_pinpoints
+
+        self.image_feature_len = int((self.image_size // self.patch_size) ** 2)
+
+        self.multi_modal_projector = self.MULTIMODAL_PROJECTOR_TYPE(config)
+
+        language_model_cls = self._get_sgl_model_cls(
+            config.text_config, AutoModelForCausalLM
+        )
+        vision_model_cls = self._get_sgl_model_cls(config.vision_config, AutoModel)
+        self.language_model = language_model_cls(
+            config.text_config,
+            quant_config=quant_config,
+            prefix=add_prefix("language_model", prefix),
+        )
+        self.vision_tower = vision_model_cls(
+            config.vision_config,
+            quant_config=quant_config,
+            prefix=add_prefix("vision_tower", prefix),
+        )
+
+        if "unpad" in getattr(config, "mm_patch_merge_type", ""):
+            self.language_model.model.image_newline = nn.Parameter(
+                torch.empty(config.text_config.hidden_size, dtype=torch.float16)
+            )
+
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        """Extract features from image inputs.
+
+        Args:
+            items: List of MultimodalDataItem objects containing image data
+                Note that an item can be either "image" or "multi-images"
+
+        Returns:
+            torch.Tensor: features from image inputs, concatenated
+        """
+        features = []
+        for item in items:
+            # in each item, we assume pixel_values is always batched
+            pixel_values, image_sizes = item.pixel_values, item.image_sizes
+            image_outputs = self.vision_tower(
+                pixel_values, image_sizes, output_hidden_states=True
+            )
+            selected_image_feature = image_outputs.hidden_states[
+                self.vision_feature_layer
+            ]
+
+            if self.vision_feature_select_strategy in ["default", "patch"]:
+                selected_image_feature = selected_image_feature[:, 1:]
+            elif self.vision_feature_select_strategy == "full":
+                selected_image_feature = selected_image_feature
+            else:
+                raise ValueError(
+                    f"Unexpected select feature: {self.vision_feature_select_strategy}"
+                )
+            features.append(
+                self.multi_modal_projector(selected_image_feature.squeeze(0))
+            )
+        ret = torch.cat(features, dim=0)
+        return ret
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        get_embedding: bool = False,
+    ):
+        hidden_states = general_mm_embed_routine(
+            input_ids=input_ids,
+            forward_batch=forward_batch,
+            get_embedding=get_embedding,
+            language_model=self.language_model,
+            image_data_embedding_func=self.get_image_feature,
+            placeholder_tokens=None,  # using mm_item.pad_value
+            positions=positions,
+        )
+
+        return hidden_states
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        """Load weights for LlavaForConditionalGeneration.
+
+        Unlike the base class implementation, this one doesn't need to handle
+        weight name remapping as the weights are already properly structured with
+        'language_model' and 'vision_tower' prefixes in the safetensors files.
+        """
+        if (
+            self.vision_feature_select_strategy == "patch"
+            or self.vision_feature_select_strategy == "full"
+        ):
+            pass
+        elif self.vision_feature_select_strategy == "cls_patch":
+            self.image_feature_len += 1
+        else:
+            raise ValueError(
+                f"Unexpected select feature: {self.vision_feature_select_strategy}"
+            )
+
+        # Create dictionaries for direct parameter loading
+        params_dict = dict(self.named_parameters())
+
+        # Load weights directly without remapping
+        for name, loaded_weight in weights:
+            for part in ("language_model", "vision_tower"):
+                if name.startswith(part):
+                    name = name[len(part + ".") :]
+                    getattr(self, part).load_weights([(name, loaded_weight)])
+                    break
+            else:
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+
+EntryClass = [
+    LlavaLlamaForCausalLM,
+    LlavaQwenForCausalLM,
+    LlavaMistralForCausalLM,
+    LlavaForConditionalGeneration,
+]

sglang/srt/models/minicpmv.py

@@ -197,7 +197,7 @@ class Idefics2EncoderLayer(nn.Module):
             use_qkv_parallel=True,
             quant_config=quant_config,
             dropout=config.attention_dropout,
-            use_context_forward=False,
+            qkv_backend="sdpa",
             softmax_in_single_precision=True,
             flatten_batch=False,
             prefix=add_prefix("self_attn", prefix),

sglang/srt/models/mixtral.py

@@ -16,13 +16,15 @@
 # https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/mixtral.py#L1
 """Inference-only Mixtral model."""
 
-from typing import Iterable, Optional, Tuple
+import logging
+from typing import Iterable, Optional, Tuple, Union
 
 import torch
 from torch import nn
 from transformers import MixtralConfig
 
 from sglang.srt.distributed import (
+    get_pp_group,
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_reduce,
 )
@@ -38,14 +40,17 @@ from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import get_rope
+from sglang.srt.layers.utils import PPMissingLayer, get_layer_id
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
 from sglang.srt.managers.schedule_batch import global_server_args_dict
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import default_weight_loader
-from sglang.srt.utils import add_prefix
+from sglang.srt.utils import add_prefix, make_layers
+
+logger = logging.getLogger(__name__)
 
 
 class MixtralMoE(nn.Module):
@@ -257,24 +262,32 @@ class MixtralModel(nn.Module):
         super().__init__()
         self.padding_idx = config.pad_token_id
         self.vocab_size = config.vocab_size
+        self.pp_group = get_pp_group()
 
-        self.embed_tokens = VocabParallelEmbedding(
-            config.vocab_size,
-            config.hidden_size,
-            prefix=add_prefix("embed_tokens", prefix),
-        )
-        self.layers = nn.ModuleList(
-            [
-                MixtralDecoderLayer(
-                    config,
-                    i,
-                    quant_config=quant_config,
-                    prefix=add_prefix(f"layers.{i}", prefix),
-                )
-                for i in range(config.num_hidden_layers)
-            ]
+        if self.pp_group.is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+                prefix=add_prefix("embed_tokens", prefix),
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        self.layers, self.start_layer, self.end_layer = make_layers(
+            config.num_hidden_layers,
+            lambda idx, prefix: MixtralDecoderLayer(
+                config=config, quant_config=quant_config, layer_id=idx, prefix=prefix
+            ),
+            pp_rank=self.pp_group.rank_in_group,
+            pp_size=self.pp_group.world_size,
+            prefix="layers",
+            return_tuple=True,
         )
-        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        if self.pp_group.is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer(return_tuple=True)
 
     def forward(
         self,
@@ -282,18 +295,35 @@ class MixtralModel(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
-    ) -> torch.Tensor:
-        if input_embeds is None:
-            hidden_states = self.embed_tokens(input_ids)
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> Union[torch.Tensor, PPProxyTensors]:
+        if self.pp_group.is_first_rank:
+            if input_embeds is None:
+                hidden_states = self.embed_tokens(input_ids)
+            else:
+                hidden_states = input_embeds
+            residual = None
         else:
-            hidden_states = input_embeds
-        residual = None
-        for i in range(len(self.layers)):
+            assert pp_proxy_tensors is not None
+            hidden_states = pp_proxy_tensors["hidden_states"]
+            residual = pp_proxy_tensors["residual"]
+
+        for i in range(self.start_layer, self.end_layer):
             layer = self.layers[i]
             hidden_states, residual = layer(
                 positions, hidden_states, forward_batch, residual
             )
-        hidden_states, _ = self.norm(hidden_states, residual)
+
+        if not self.pp_group.is_last_rank:
+            return PPProxyTensors(
+                {
+                    "hidden_states": hidden_states,
+                    "residual": residual,
+                }
+            )
+        else:
+            hidden_states, _ = self.norm(hidden_states, residual)
+
         return hidden_states
 
 
@@ -306,6 +336,7 @@ class MixtralForCausalLM(nn.Module):
         prefix: str = "",
     ) -> None:
         super().__init__()
+        self.pp_group = get_pp_group()
         self.config = config
         self.quant_config = quant_config
         self.model = MixtralModel(
@@ -322,12 +353,31 @@ class MixtralForCausalLM(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
     ) -> torch.Tensor:
-        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
-        return self.logits_processor(
-            input_ids, hidden_states, self.lm_head, forward_batch
+        hidden_states = self.model(
+            input_ids,
+            positions,
+            forward_batch,
+            input_embeds,
+            pp_proxy_tensors=pp_proxy_tensors,
         )
 
+        if self.pp_group.is_last_rank:
+            return self.logits_processor(
+                input_ids, hidden_states, self.lm_head, forward_batch
+            )
+        else:
+            return hidden_states
+
+    @property
+    def start_layer(self):
+        return self.model.start_layer
+
+    @property
+    def end_layer(self):
+        return self.model.end_layer
+
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
@@ -348,6 +398,17 @@ class MixtralForCausalLM(nn.Module):
 
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in weights:
+            layer_id = get_layer_id(name)
+            if (
+                layer_id is not None
+                and hasattr(self.model, "start_layer")
+                and (
+                    layer_id < self.model.start_layer
+                    or layer_id >= self.model.end_layer
+                )
+            ):
+                continue
+
             if "rotary_emb.inv_freq" in name:
                 continue
 
@@ -398,11 +459,14 @@ class MixtralForCausalLM(nn.Module):
                     if name is None:
                         continue
 
-                    param = params_dict[name]
-                    weight_loader = getattr(
-                        param, "weight_loader", default_weight_loader
-                    )
-                    weight_loader(param, loaded_weight)
+                    if name in params_dict.keys():
+                        param = params_dict[name]
+                        weight_loader = getattr(
+                            param, "weight_loader", default_weight_loader
+                        )
+                        weight_loader(param, loaded_weight)
+                    else:
+                        logger.warning(f"Parameter {name} not found in params_dict")
 
 
 EntryClass = MixtralForCausalLM

sglang/srt/models/mllama.py

@@ -203,7 +203,7 @@ class MllamaVisionEncoderLayer(nn.Module):
             use_qkv_parallel=True,
             quant_config=quant_config,
             dropout=0.0,
-            use_context_forward=False,
+            qkv_backend="sdpa",
             softmax_in_single_precision=False,
             flatten_batch=False,
             prefix=add_prefix("self_attn", prefix),