sglang 0.4.4.post2__py3-none-any.whl → 0.4.4.post4__py3-none-any.whl

sglang/srt/layers/quantization/moe_wna16.py

@@ -0,0 +1,501 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/quantization/moe_wna16.py
+
+import logging
+from typing import Any, Callable, Dict, List, Optional
+
+import torch
+
+from sglang.srt.distributed import get_tensor_model_parallel_rank
+from sglang.srt.distributed.parallel_state import get_tp_group
+from sglang.srt.layers.linear import LinearBase, UnquantizedLinearMethod
+from sglang.srt.layers.quantization.awq import AWQConfig
+from sglang.srt.layers.quantization.base_config import (
+    QuantizationConfig,
+    QuantizeMethodBase,
+)
+from sglang.srt.layers.quantization.gptq import GPTQConfig, GPTQMarlinConfig
+from sglang.srt.utils import get_device_capability, set_weight_attrs
+
+logger = logging.getLogger(__name__)
+
+
+class MoeWNA16Config(QuantizationConfig):
+    """Config class for MOE WNA16 (W8A16/W4A16) quantization."""
+
+    def __init__(
+        self,
+        linear_quant_method: str,
+        weight_bits: int,
+        group_size: int,
+        has_zp: bool,
+        lm_head_quantized: bool,
+        modules_to_not_convert: Optional[List[str]],
+        full_config: Dict[str, Any],
+    ) -> None:
+        super().__init__()
+        self.weight_bits = weight_bits
+        self.group_size = group_size
+        self.has_zp = has_zp
+        self.bit8_pack_factor = 8 // self.weight_bits
+        self.lm_head_quantized = lm_head_quantized
+        self.linear_quant_method = linear_quant_method
+        self.full_config = full_config
+        self.use_marlin = False
+        # Avoid circular import
+
+        if self.linear_quant_method == "gptq":
+            self.use_marlin = GPTQMarlinConfig.is_gptq_marlin_compatible(full_config)
+        elif self.linear_quant_method == "awq":
+            capability_tuple = get_device_capability()
+            device_capability = (
+                -1
+                if capability_tuple is None
+                else capability_tuple[0] * 10 + capability_tuple[1]
+            )
+            awq_min_capability = AWQConfig.get_min_capability()
+            if device_capability < awq_min_capability:
+                raise ValueError(
+                    "The quantization method moe_wna16 + awq is not supported "
+                    "for the current GPU. "
+                    f"Minimum capability: {awq_min_capability}. "
+                    f"Current capability: {device_capability}."
+                )
+        else:
+            raise ValueError("moe_wna16 only support gptq and awq.")
+
+        if modules_to_not_convert is None:
+            self.modules_to_not_convert = []
+        else:
+            self.modules_to_not_convert = modules_to_not_convert
+
+    @classmethod
+    def get_name(cls) -> str:
+        return "moe_wna16"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.bfloat16, torch.half]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 70
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return ["quantize_config.json"]
+
+    def get_scaled_act_names(self) -> List[str]:
+        raise NotImplementedError
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "MoeWNA16Config":
+        quant_method = cls.get_from_keys(config, ["quant_method"])
+        weight_bits = cls.get_from_keys(config, ["bits"])
+        group_size = cls.get_from_keys(config, ["group_size"])
+        lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False)
+        if quant_method == "gptq":
+            has_zp = not cls.get_from_keys(config, ["sym"])
+            modules_to_not_convert = []
+        elif quant_method == "awq":
+            has_zp = cls.get_from_keys(config, ["zero_point"])
+            modules_to_not_convert = cls.get_from_keys_or(
+                config, ["modules_to_not_convert"], None
+            )
+        else:
+            raise ValueError("moe_wna16 only support gptq and awq.")
+
+        return cls(
+            quant_method,
+            weight_bits,
+            group_size,
+            has_zp,
+            lm_head_quantized,
+            modules_to_not_convert,
+            config,
+        )
+
+    @classmethod
+    def override_quantization_method(cls, hf_quant_cfg, user_quant) -> Optional[str]:
+        can_convert = cls.is_moe_wna16_compatible(hf_quant_cfg)
+        if can_convert and user_quant == "moe_wna16":
+            return cls.get_name()
+        return None
+
+    @classmethod
+    def is_moe_wna16_compatible(cls, quant_config: Dict[str, Any]):
+        # Extract data from quant config.
+        quant_method = quant_config.get("quant_method", "").lower()
+        num_bits = quant_config.get("bits")
+        desc_act = quant_config.get("desc_act")
+
+        capability_tuple = get_device_capability()
+        device_capability = (
+            -1
+            if capability_tuple is None
+            else capability_tuple[0] * 10 + capability_tuple[1]
+        )
+        # Avoid circular import
+        awq_min_capability = AWQConfig.get_min_capability()
+
+        gptq_compatible = quant_method == "gptq" and not desc_act and num_bits in [4, 8]
+        awq_compatible = (
+            quant_method == "awq"
+            and num_bits == 4
+            and device_capability >= awq_min_capability
+        )
+
+        return gptq_compatible or awq_compatible
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional["QuantizeMethodBase"]:
+        # avoid circular import
+        from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
+
+        if is_layer_skipped_quant(prefix, self.modules_to_not_convert):
+            return UnquantizedLinearMethod()
+        elif isinstance(layer, LinearBase):
+
+            if self.linear_quant_method == "gptq":
+                if self.use_marlin:
+                    return GPTQMarlinConfig.from_config(
+                        self.full_config
+                    ).get_quant_method(layer, prefix)
+                else:
+                    return GPTQConfig.from_config(self.full_config).get_quant_method(
+                        layer, prefix
+                    )
+            elif self.linear_quant_method == "awq":
+                return AWQConfig.from_config(self.full_config).get_quant_method(
+                    layer, prefix
+                )
+            else:
+                raise ValueError("moe_wna16 only support gptq and awq.")
+        elif isinstance(layer, FusedMoE):
+            return MoeWNA16Method(self)
+        return None
+
+
+def is_layer_skipped_quant(prefix: str, modules_to_not_convert: List[str]):
+    return any(module_name in prefix for module_name in modules_to_not_convert)
+
+
+class MoeWNA16Method:
+    """Linear method for MOE WNA16 (W8A16/W4A16) quantization.
+
+    Args:
+        quant_config: The MOE WNA16 (W8A16/W4A16) quantization config.
+    """
+
+    def __new__(cls, *args, **kwargs):
+        # avoid circular import
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoEMethodBase
+
+        if not hasattr(cls, "_initialized"):
+            original_init = cls.__init__
+            new_cls = type(
+                cls.__name__,
+                (FusedMoEMethodBase,),
+                {
+                    "__init__": original_init,
+                    **{k: v for k, v in cls.__dict__.items() if k != "__dict__"},
+                },
+            )
+            obj = super(new_cls, new_cls).__new__(new_cls)
+            obj.__init__(*args, **kwargs)
+            return obj
+        return super().__new__(cls)
+
+    def __init__(self, quant_config: MoeWNA16Config):
+        self.quant_config = quant_config
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
+
+        layer.quant_config = self.quant_config
+        bit8_pack_factor = self.quant_config.bit8_pack_factor
+        group_size = self.quant_config.group_size
+        group_size_div_factor = 1
+
+        # make intermediate_size and hidden_size diviable by group_size
+        # we reduce the group size to ensure that
+        # and we would repeat the loaded_weight later
+        while intermediate_size_per_partition % group_size or hidden_size % group_size:
+            group_size = group_size // 2
+            group_size_div_factor *= 2
+            assert group_size >= 32
+        layer.group_size = group_size
+        layer.group_size_div_factor = group_size_div_factor
+
+        strategy = FusedMoeWeightScaleSupported.GROUP.value
+        extra_weight_attrs.update({"quant_method": strategy, "is_transposed": False})
+
+        assert "weight_loader" in extra_weight_attrs
+        weight_loader = extra_weight_attrs["weight_loader"]
+        wrapped_weight_loader = MoeWNA16Method.get_weight_loader(layer, weight_loader)
+        extra_weight_attrs["weight_loader"] = wrapped_weight_loader
+
+        # Fused gate_up_proj (column parallel)
+        w13_qweight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                hidden_size // bit8_pack_factor,
+                dtype=torch.uint8,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_qweight", w13_qweight)
+        set_weight_attrs(w13_qweight, extra_weight_attrs)
+
+        # down_proj (row parallel)
+        w2_qweight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                intermediate_size_per_partition // bit8_pack_factor,
+                dtype=torch.uint8,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_qweight", w2_qweight)
+        set_weight_attrs(w2_qweight, extra_weight_attrs)
+
+        w13_scales = torch.nn.Parameter(
+            torch.zeros(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                hidden_size // group_size,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_scales", w13_scales)
+        set_weight_attrs(w13_scales, extra_weight_attrs)
+
+        w2_scales = torch.nn.Parameter(
+            torch.zeros(
+                num_experts,
+                hidden_size,
+                intermediate_size_per_partition // group_size,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_scales", w2_scales)
+        set_weight_attrs(w2_scales, extra_weight_attrs)
+
+        if self.quant_config.has_zp:
+            w13_qzeros = torch.nn.Parameter(
+                torch.zeros(
+                    num_experts,
+                    2 * intermediate_size_per_partition // bit8_pack_factor,
+                    hidden_size // group_size,
+                    dtype=torch.uint8,
+                ),
+                requires_grad=False,
+            )
+            layer.register_parameter("w13_qzeros", w13_qzeros)
+            set_weight_attrs(w13_qzeros, extra_weight_attrs)
+
+            w2_qzeros = torch.nn.Parameter(
+                torch.zeros(
+                    num_experts,
+                    hidden_size // bit8_pack_factor,
+                    intermediate_size_per_partition // group_size,
+                    dtype=torch.uint8,
+                ),
+                requires_grad=False,
+            )
+            layer.register_parameter("w2_qzeros", w2_qzeros)
+            set_weight_attrs(w2_qzeros, extra_weight_attrs)
+
+        if self.quant_config.linear_quant_method == "gptq":
+            # some param are unused, but we need to init them in order to
+            # load weights
+            invalid_param_keys = ["w13_g_idx", "w2_g_idx"]
+            if not self.quant_config.has_zp:
+                invalid_param_keys += ["w13_qzeros", "w2_qzeros"]
+            for key in invalid_param_keys:
+                param = torch.nn.Parameter(
+                    torch.empty((0,), dtype=torch.int32), requires_grad=False
+                )
+                layer.register_parameter(key, param)
+                set_weight_attrs(param, extra_weight_attrs)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool = False,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        custom_routing_function: Optional[Callable] = None,
+        correction_bias: Optional[torch.Tensor] = None,
+        activation: str = "silu",
+        inplace: bool = True,
+        no_combine: bool = False,
+    ) -> torch.Tensor:
+        # avoid circular import
+        from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_experts
+        from sglang.srt.layers.moe.topk import select_experts
+
+        assert activation == "silu", "Only SiLU activation is supported."
+        topk_weights, topk_ids = select_experts(
+            hidden_states=x,
+            router_logits=router_logits,
+            top_k=top_k,
+            use_grouped_topk=use_grouped_topk,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            custom_routing_function=custom_routing_function,
+            correction_bias=correction_bias,
+        )
+
+        weight_bits = self.quant_config.weight_bits
+        has_zp = self.quant_config.has_zp
+
+        return fused_experts(
+            x,
+            layer.w13_qweight,
+            layer.w2_qweight,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            inplace=inplace,
+            use_int4_w4a16=weight_bits == 4,
+            use_int8_w8a16=weight_bits == 8,
+            w1_scale=layer.w13_scales,
+            w2_scale=layer.w2_scales,
+            w1_zp=layer.w13_qzeros if has_zp else None,
+            w2_zp=layer.w2_qzeros if has_zp else None,
+            block_shape=[0, layer.group_size],
+            no_combine=no_combine,
+        )
+
+    @staticmethod
+    def get_weight_loader(layer, weight_loader):
+
+        def convert_awq_tensor(tensor, tensor_type):
+            # convert awq qweight/qzeros to a standard format (assume int4)
+            # qweight: (k, n // pack_factor_bit32) -> (n, k // pack_factor_bit8)
+            # qzeros: (k // group_size, n // pack_factor_bit32) ->
+            #         (n // pack_factor_bit8, k // group_size)
+            # pack_factor_bit32 = 32 // weight_bits
+            # pack_factor_bit8 = 8 // weight_bits
+
+            # 0. suppose origin shape (a, b), dtype int32
+            # 1. convert to uint8, shape (a, b) -> (a, 4 * b)
+            size0 = tensor.size(0)
+            tensor = tensor.view(torch.uint8)
+
+            # 2. unpack to uint4 (only when weight_bits == 4)
+            #    shape (a, 4 * b) -> (a, 4 * b, 2)
+            shifter = torch.tensor([0, 4], dtype=torch.uint8, device=tensor.device)
+            tensor = (tensor[:, :, None] >> shifter) & 0xF
+
+            # 3. change order, see
+            # https://github.com/casper-hansen/AutoAWQ/blob/v0.2.8/awq/utils/quant_utils.py
+            # shape -> (a, 4 * b * pack_factor_bit8)
+            reverse_awq_pack_order = [0, 4, 1, 5, 2, 6, 3, 7]
+            tensor = tensor.view(-1, 8)[:, reverse_awq_pack_order]
+            tensor = tensor.view(size0, -1)
+
+            # 4. transpose, shape -> (4 * b * pack_factor_bit8, a)
+            tensor = tensor.T.contiguous()
+
+            # 5. repack (only when weight_bits == 4)
+            # qweight shape -> (4 * b * pack_factor_bit8, a // pack_factor_bit8)
+            # qzeros shape -> (4 * b, a)
+
+            if tensor_type == "qweight":
+                tensor = tensor[:, 1::2] * 16 + tensor[:, ::2]
+            elif tensor_type == "qzeros":
+                tensor = tensor[1::2, :] * 16 + tensor[::2, :]
+            return tensor
+
+        def convert_gptq_int4_qzeros(tensor):
+            tensor = tensor.view(torch.uint8)
+            shifter = torch.tensor([0, 4], dtype=torch.uint8, device=tensor.device)
+            tensor = (tensor[:, :, None] >> shifter) & 0xF
+            tensor = tensor + 1
+            tensor = tensor[:, :, 0] + tensor[:, :, 1] * 16
+            return tensor
+
+        def moe_wna16_weight_loader(
+            param: torch.nn.Parameter,
+            loaded_weight: torch.Tensor,
+            weight_name: str,
+            shard_id: str,
+            expert_id: int,
+        ):
+            if "g_idx" in weight_name:
+                return
+            if not layer.quant_config.has_zp and "qzeros" in weight_name:
+                return
+
+            device = get_tp_group().device
+            tp_rank = get_tensor_model_parallel_rank()
+            loaded_weight = loaded_weight.to(device)
+            shard_size = layer.intermediate_size_per_partition
+
+            # convert gptq and awq weight to a standard format
+            if layer.quant_config.linear_quant_method == "awq":
+                assert layer.quant_config.weight_bits == 4
+                if "weight" in weight_name:
+                    loaded_weight = convert_awq_tensor(loaded_weight, "qweight")
+                elif "zeros" in weight_name:
+                    loaded_weight = convert_awq_tensor(loaded_weight, "qzeros")
+                else:
+                    loaded_weight = loaded_weight.T
+            elif layer.quant_config.linear_quant_method == "gptq":
+                assert layer.quant_config.weight_bits in [4, 8]
+                if "weight" in weight_name:
+                    loaded_weight = loaded_weight.T.contiguous().view(torch.uint8)
+                elif "zeros" in weight_name:
+                    # add 1 to gptq qzeros to align with awq
+                    loaded_weight = loaded_weight.view(torch.uint8)
+                    if layer.quant_config.weight_bits == 4:
+                        loaded_weight = convert_gptq_int4_qzeros(loaded_weight).T
+                    else:
+                        loaded_weight = loaded_weight.T + 1
+                else:
+                    loaded_weight = loaded_weight.T
+
+            # repeat the qzeros/scales to fit new group size
+            if (
+                layer.group_size_div_factor > 1
+                and "qzeros" in weight_name
+                or "scales" in weight_name
+            ):
+                loaded_weight = loaded_weight.repeat_interleave(
+                    layer.group_size_div_factor, 1
+                )
+
+            if "w13_qzeros" in weight_name:
+                tensor = loaded_weight.view(layer.tp_size, -1, loaded_weight.size(1))[
+                    tp_rank
+                ]
+                if shard_id == "w1":
+                    param.data[expert_id, : shard_size // 2] = tensor
+                else:
+                    param.data[expert_id, shard_size // 2 :] = tensor
+            elif "w2_qzeros" in weight_name:
+                param.data[expert_id] = loaded_weight.view(
+                    loaded_weight.size(0), layer.tp_size, -1
+                )[:, tp_rank]
+            else:
+                weight_loader(param, loaded_weight, weight_name, shard_id, expert_id)
+
+        return moe_wna16_weight_loader

sglang/srt/layers/quantization/utils.py

@@ -1,7 +1,7 @@
 # Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/quantization/utils/quant_utils.py
 
 from types import MappingProxyType
-from typing import List, Mapping, Tuple, Union
+from typing import List, Mapping, Optional, Tuple, Union
 
 import torch
 

sglang/srt/layers/quantization/w8a8_fp8.py

@@ -37,7 +37,7 @@ class W8A8Fp8Config(QuantizationConfig):
     Note:
     - For models without offline quantization, weights will be quantized during model loading
     - If CUTLASS is supported: Per-channel weight quantization is used
-    - If CUTLASS is not supported: Falls back to per-token weight quantization
+    - If CUTLASS is not supported: Falls back to per-tensor weight quantization
     """
 
     def __init__(self, is_checkpoint_fp8_serialized: bool = False):

sglang/srt/lora/backend/base_backend.py

@@ -5,7 +5,7 @@ import torch
 from sglang.srt.lora.utils import LoRABatchInfo
 
 
-def get_fuse_output_scaling_add_from_name(name: str) -> bool:
+def get_fuse_output_add_from_name(name: str) -> bool:
     mapping = {
         "triton": True,
         "flashinfer": False,
@@ -28,14 +28,14 @@ class BaseLoRABackend:
     Args:
         name: name of backend
         batch_info: information of current batch for use
-        fuse_output_scaling_add: if set to True, the output buffer for storing result will be passed in when doing lora_b forward,
-                                 and the operation of scaling and adding will be fused into kernel
+        fuse_output_add: if set to True, the output buffer for storing result will be passed in when doing lora_b forward,
+                                 and the operation of adding will be fused into kernel
     """
 
     def __init__(self, name: str, batch_info: LoRABatchInfo = None):
         self.name = name
         self.batch_info = batch_info
-        self.fuse_output_scaling_add = get_fuse_output_scaling_add_from_name(name)
+        self.fuse_output_add = get_fuse_output_add_from_name(name)
         self.fuse_stacked_lora_b = get_fuse_stacked_lora_b_from_name(name)
 
     def run_lora_a_sgemm(

sglang/srt/lora/backend/flashinfer_backend.py

@@ -37,13 +37,16 @@ class FlashInferLoRABackend(BaseLoRABackend):
         self, x: torch.Tensor, weights: torch.Tensor, *args, **kwargs
     ) -> torch.Tensor:
 
-        return self.segment_gemm.run(
-            x=x,
-            weights=weights,
-            batch_size=self.batch_info.bs,
-            weight_column_major=True,
-            seg_indptr=self.batch_info.seg_indptr,
-            weight_indices=self.batch_info.weight_indices,
+        return (
+            self.segment_gemm.run(
+                x=x,
+                weights=weights,
+                batch_size=self.batch_info.bs,
+                weight_column_major=True,
+                seg_indptr=self.batch_info.seg_indptr,
+                weight_indices=self.batch_info.weight_indices,
+            )
+            * self.batch_info.scalings[0]
         )
 
     def run_qkv_lora(
@@ -90,7 +93,7 @@ class FlashInferLoRABackend(BaseLoRABackend):
             weights=kv_lora_b[1],
         )
 
-        return lora_output
+        return lora_output * self.batch_info.scalings[0]
 
     def run_gate_up_lora(
         self,
@@ -125,4 +128,4 @@ class FlashInferLoRABackend(BaseLoRABackend):
             weights=gate_up_lora_b[1],
         )
 
-        return lora_output
+        return lora_output * self.batch_info.scalings[0]

sglang/srt/lora/backend/triton_backend.py

@@ -25,11 +25,10 @@ class TritonLoRABackend(BaseLoRABackend):
         x: torch.Tensor,
         weights: torch.Tensor,
         base_output: torch.Tensor = None,
-        scaling: float = 1.0,
         *args,
         **kwargs
     ) -> torch.Tensor:
-        return sgemm_lora_b_fwd(x, weights, self.batch_info, base_output, scaling)
+        return sgemm_lora_b_fwd(x, weights, self.batch_info, base_output)
 
     def run_qkv_lora(
         self,
@@ -39,7 +38,6 @@ class TritonLoRABackend(BaseLoRABackend):
         output_offset: torch.Tensor,
         max_qkv_out_dim: int,
         base_output: torch.Tensor = None,
-        scaling: float = 1.0,
         *args,
         **kwargs
     ) -> torch.Tensor:
@@ -49,7 +47,7 @@ class TritonLoRABackend(BaseLoRABackend):
         # qkv_lora_b: (num_lora, output_dim_q + 2 * output_dim_kv, r)
         assert isinstance(qkv_lora_b, torch.Tensor)
 
-        lora_a_output = sgemm_lora_a_fwd(x, qkv_lora_a, self.batch_info)
+        lora_a_output = sgemm_lora_a_fwd(x, qkv_lora_a, self.batch_info, stack_num=3)
         lora_output = qkv_lora_b_fwd(
             lora_a_output,
             qkv_lora_b,
@@ -57,7 +55,6 @@ class TritonLoRABackend(BaseLoRABackend):
             output_offset,
             max_qkv_out_dim,
             base_output,
-            scaling,
         )
         return lora_output
 
@@ -67,7 +64,6 @@ class TritonLoRABackend(BaseLoRABackend):
         gate_up_lora_a: torch.Tensor,
         gate_up_lora_b: torch.Tensor,
         base_output: torch.Tensor = None,
-        scaling: float = 1.0,
         *args,
         **kwargs
     ) -> torch.Tensor:
@@ -79,13 +75,14 @@ class TritonLoRABackend(BaseLoRABackend):
         output_dim = gate_up_lora_b.shape[-2] // 2
 
         # lora_a_output: (s, 2 * r)
-        lora_a_output = sgemm_lora_a_fwd(x, gate_up_lora_a, self.batch_info)
+        lora_a_output = sgemm_lora_a_fwd(
+            x, gate_up_lora_a, self.batch_info, stack_num=2
+        )
         lora_output = gate_up_lora_b_fwd(
             lora_a_output,
             gate_up_lora_b,
             self.batch_info,
             output_dim,
             base_output,
-            scaling,
         )
         return lora_output