sglang 0.4.9__py3-none-any.whl → 0.4.9.post2__py3-none-any.whl

sglang/srt/layers/quantization/modelopt_quant.py

@@ -26,6 +26,7 @@ from sglang.srt.layers.quantization.kv_cache import BaseKVCacheMethod
 from sglang.srt.layers.quantization.utils import (
     convert_to_channelwise,
     is_layer_skipped,
+    per_tensor_dequantize,
     requantize_with_max_scale,
 )
 from sglang.srt.layers.radix_attention import RadixAttention
@@ -110,7 +111,12 @@ class ModelOptFp8Config(QuantizationConfig):
         self, layer: torch.nn.Module, prefix: str
     ) -> Optional["QuantizeMethodBase"]:
         if self.exclude_modules and any(
-            module in prefix for module in self.exclude_modules
+            module in prefix
+            or (
+                prefix.startswith("language_model.")
+                and module in prefix.removeprefix("language_model.")
+            )
+            for module in self.exclude_modules
         ):
             return None
 
@@ -119,6 +125,12 @@ class ModelOptFp8Config(QuantizationConfig):
         if self.kv_cache_quant_method and isinstance(layer, RadixAttention):
             return ModelOptFp8KVCacheMethod(self)
 
+        # Add MoE support
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+
+        if isinstance(layer, FusedMoE):
+            return ModelOptFp8MoEMethod(self)
+
         return None
 
     def get_scaled_act_names(self) -> List[str]:
@@ -234,6 +246,237 @@ class ModelOptFp8KVCacheMethod(BaseKVCacheMethod):
         super().__init__(quant_config)
 
 
+class ModelOptFp8MoEMethod:
+    """MoE method for ModelOpt FP8.
+    Supports loading FP8 checkpoints with static weight scale and activation scale.
+
+    Args:
+        quant_config: The ModelOpt quantization config.
+    """
+
+    def __new__(cls, *args, **kwargs):
+        """
+        Dynamic class composition pattern.
+
+        This allows us to effectively "inject" FusedMoEMethodBase as a parent class
+        at runtime while avoiding circular import issues.
+        """
+        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: ModelOptFp8Config):
+        self.quant_config = quant_config
+        self.cutlass_fp8_supported = cutlass_fp8_supported()
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
+
+        # Use FP8 dtype if checkpoint is serialized, otherwise use the default dtype
+        weight_dtype = (
+            torch.float8_e4m3fn
+            if self.quant_config.is_checkpoint_fp8_serialized
+            else params_dtype
+        )
+        weight_loader = extra_weight_attrs.get("weight_loader")
+
+        w13_weight = ModelWeightParameter(
+            data=torch.empty(
+                num_experts, 2 * intermediate_size, hidden_size, dtype=weight_dtype
+            ),
+            input_dim=2,
+            output_dim=1,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("w13_weight", w13_weight)
+
+        w2_weight = ModelWeightParameter(
+            data=torch.empty(
+                num_experts, hidden_size, intermediate_size, dtype=weight_dtype
+            ),
+            input_dim=2,
+            output_dim=1,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("w2_weight", w2_weight)
+
+        if self.quant_config.is_checkpoint_fp8_serialized:
+            # WEIGHT SCALES - Per-tensor scaling for ModelOpts
+            # Allocate 2 scales for w1 and w3 respectively.
+            # They will be combined to a single scale after weight loading.
+            w13_weight_scale = PerTensorScaleParameter(
+                data=torch.full(
+                    (num_experts, 2),
+                    torch.finfo(torch.float32).min,
+                    dtype=torch.float32,
+                ),
+                weight_loader=weight_loader,
+            )
+            w2_weight_scale = PerTensorScaleParameter(
+                data=torch.full(
+                    (num_experts,), torch.finfo(torch.float32).min, dtype=torch.float32
+                ),
+                weight_loader=weight_loader,
+            )
+            layer.register_parameter("w13_weight_scale", w13_weight_scale)
+            layer.register_parameter("w2_weight_scale", w2_weight_scale)
+
+            # Set weight loader attributes for scales
+            extra_weight_attrs.update(
+                {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
+            )
+
+            # INPUT SCALES - Per-tensor scaling for ModelOpt
+            w13_input_scale = PerTensorScaleParameter(
+                data=torch.full((num_experts,), 1.0, dtype=torch.float32),
+                weight_loader=weight_loader,
+            )
+            w2_input_scale = PerTensorScaleParameter(
+                data=torch.full((num_experts,), 1.0, dtype=torch.float32),
+                weight_loader=weight_loader,
+            )
+            layer.register_parameter("w13_input_scale", w13_input_scale)
+            layer.register_parameter("w2_input_scale", w2_input_scale)
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        """Process FP8 MoE weights after loading from serialized checkpoint.
+
+        Only supports pre-quantized checkpoints with FP8 weights and scales.
+        """
+
+        layer.w13_weight = Parameter(layer.w13_weight.data, requires_grad=False)
+        layer.w2_weight = Parameter(layer.w2_weight.data, requires_grad=False)
+
+        # Handle scale parameters
+        if hasattr(layer, "w13_weight_scale") and layer.w13_weight_scale is not None:
+            # Fp8 moe kernel needs single weight scale for w13 per expert.
+            # We take the max of the w1 and w3 scales then dequant and requant each expert.
+            if layer.w13_weight_scale.dim() == 2:  # Shape: (num_experts, 2)
+                from sglang.srt.layers.quantization.fp8_kernel import scaled_fp8_quant
+
+                # Get the maximum scale across w1 and w3 for each expert
+                max_w13_scales = layer.w13_weight_scale.max(dim=1).values
+
+                # Requantize each expert's weights using the combined scale
+                # w13_weight has shape (num_experts, 2 * intermediate_size, hidden_size)
+                # where the first intermediate_size rows are w1, the next are w3
+                intermediate_size = layer.w13_weight.shape[1] // 2
+                for expert_id in range(layer.w13_weight.shape[0]):
+                    start = 0
+                    for shard_id in range(2):  # w1 and w3
+                        # Dequantize using the original scale for this shard
+                        dq_weight = per_tensor_dequantize(
+                            layer.w13_weight[expert_id][
+                                start : start + intermediate_size, :
+                            ],
+                            layer.w13_weight_scale[expert_id][shard_id],
+                        )
+                        # Requantize using the combined max scale
+                        (
+                            layer.w13_weight[expert_id][
+                                start : start + intermediate_size, :
+                            ],
+                            _,
+                        ) = scaled_fp8_quant(dq_weight, max_w13_scales[expert_id])
+
+                        start += intermediate_size
+
+                # Update the scale parameter to be per-expert instead of per-shard
+                layer.w13_weight_scale = Parameter(max_w13_scales, requires_grad=False)
+            else:
+                layer.w13_weight_scale = Parameter(
+                    layer.w13_weight_scale.data, requires_grad=False
+                )
+
+        if hasattr(layer, "w2_weight_scale") and layer.w2_weight_scale is not None:
+            layer.w2_weight_scale = Parameter(
+                layer.w2_weight_scale.data, requires_grad=False
+            )
+        if hasattr(layer, "w13_input_scale") and layer.w13_input_scale is not None:
+            layer.w13_input_scale = Parameter(
+                layer.w13_input_scale.max(), requires_grad=False
+            )
+        if hasattr(layer, "w2_input_scale") and layer.w2_input_scale is not None:
+            layer.w2_input_scale = Parameter(
+                layer.w2_input_scale.max(), requires_grad=False
+            )
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        num_fused_shared_experts: Optional[int] = None,
+        custom_routing_function: Optional[Callable] = None,
+        correction_bias: Optional[torch.Tensor] = None,
+        activation: str = "silu",
+        apply_router_weight_on_input: bool = False,
+        inplace: bool = True,
+        no_combine: bool = False,
+        routed_scaling_factor: Optional[float] = None,
+    ) -> torch.Tensor:
+        from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_experts
+        from sglang.srt.layers.moe.topk import select_experts
+
+        # Expert selection
+        topk_weights, topk_ids = select_experts(
+            hidden_states=x,
+            router_logits=router_logits,
+            use_grouped_topk=use_grouped_topk,
+            top_k=top_k,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            num_fused_shared_experts=num_fused_shared_experts,
+            custom_routing_function=custom_routing_function,
+            correction_bias=correction_bias,
+            routed_scaling_factor=routed_scaling_factor,
+        )
+
+        return fused_experts(
+            x,
+            layer.w13_weight,
+            layer.w2_weight,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            inplace=inplace,
+            activation=activation,
+            use_fp8_w8a8=True,
+            per_channel_quant=False,  # ModelOpt uses per-tensor quantization
+            w1_scale=layer.w13_weight_scale,
+            w2_scale=layer.w2_weight_scale,
+            a1_scale=layer.w13_input_scale,
+            a2_scale=layer.w2_input_scale,
+            no_combine=no_combine,
+        )
+
+
 class ModelOptFp4Config(QuantizationConfig):
     """Config class for FP4."""
 

sglang/srt/layers/quantization/moe_wna16.py

@@ -116,8 +116,7 @@ class MoeWNA16Config(QuantizationConfig):
 
     @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":
+        if user_quant == "moe_wna16" and cls.is_moe_wna16_compatible(hf_quant_cfg):
             return cls.get_name()
         return None
 

sglang/srt/layers/quantization/w4afp8.py

@@ -0,0 +1,264 @@
+import logging
+from typing import Any, Dict, List, Optional
+
+import torch
+from torch.nn import Module
+from torch.nn.parameter import Parameter
+
+from sglang.srt.layers.linear import LinearBase, UnquantizedLinearMethod
+from sglang.srt.layers.quantization.base_config import (
+    QuantizationConfig,
+    QuantizeMethodBase,
+)
+from sglang.srt.layers.quantization.fp8 import Fp8LinearMethod
+from sglang.srt.layers.quantization.utils import is_layer_skipped
+from sglang.srt.utils import set_weight_attrs
+
+ACTIVATION_SCHEMES = ["static", "dynamic"]
+
+logger = logging.getLogger(__name__)
+
+
+class W4AFp8Config(QuantizationConfig):
+    """Config class for MIXED_PRECISION W4AFp8."""
+
+    def __init__(
+        self,
+        is_checkpoint_fp8_serialized: bool = True,
+        is_checkpoint_w4afp8_serialized: bool = True,
+        linear_activation_scheme: str = "dynamic",
+        moe_activation_scheme: str = "static",
+        ignored_layers: Optional[List[str]] = None,
+        weight_block_size: Optional[List[int]] = None,
+        group_size: int = 128,
+    ) -> None:
+        super().__init__()
+        self.is_checkpoint_fp8_serialized = is_checkpoint_fp8_serialized
+        self.is_checkpoint_w4afp8_serialized = is_checkpoint_w4afp8_serialized
+        if is_checkpoint_w4afp8_serialized:
+            logger.warning("Detected w4afp8 checkpoint. Please note that")
+        if moe_activation_scheme not in ACTIVATION_SCHEMES:
+            raise ValueError(f"Unsupported activation scheme {moe_activation_scheme}")
+        self.linear_activation_scheme = linear_activation_scheme
+        self.moe_activation_scheme = moe_activation_scheme
+        self.ignored_layers = ignored_layers or []
+        self.weight_block_size = [128, 128]
+        self.group_size = group_size
+
+    @classmethod
+    def get_name(cls) -> str:
+        return "w4afp8"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.bfloat16, torch.float8_e4m3fn]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 90
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return []
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "W4AFp8Config":
+        quant_method = cls.get_from_keys(config, ["quant_method"])
+        is_checkpoint_fp8_serialized = "fp8" in quant_method
+        is_checkpoint_w4afp8_serialized = "w4afp8" in quant_method
+        linear_activation_scheme = "dynamic"
+        moe_activation_scheme = "static"
+        weight_block_size = [128, 128]
+        return cls(
+            is_checkpoint_fp8_serialized=is_checkpoint_fp8_serialized,
+            is_checkpoint_w4afp8_serialized=is_checkpoint_w4afp8_serialized,
+            linear_activation_scheme=linear_activation_scheme,
+            moe_activation_scheme=moe_activation_scheme,
+            weight_block_size=weight_block_size,
+        )
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional["QuantizeMethodBase"]:
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+
+        if isinstance(layer, LinearBase):
+            if is_layer_skipped(prefix, self.ignored_layers):
+                return UnquantizedLinearMethod()
+            return Fp8LinearMethod(self)
+        elif isinstance(layer, FusedMoE):
+            return W4AFp8MoEMethod(self)
+        return None
+
+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+
+class W4AFp8MoEMethod:
+
+    def __init__(self, quant_config: W4AFp8Config):
+        self.quant_config = quant_config
+
+    def create_weights(
+        self,
+        layer: Module,
+        num_experts_per_partition: int,
+        hidden_size: int,
+        intermediate_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        assert "weight_loader" in extra_weight_attrs
+
+        # Fused gate_up_proj (column parallel)
+        w13_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts_per_partition,
+                intermediate_size * 2,
+                hidden_size // 2,
+                dtype=torch.int8,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight", w13_weight)
+        set_weight_attrs(w13_weight, extra_weight_attrs)
+
+        # down_proj (row parallel)
+        w2_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts_per_partition,
+                hidden_size,
+                intermediate_size // 2,
+                dtype=torch.int8,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        w13_weight_scale = torch.nn.Parameter(
+            torch.zeros(
+                num_experts_per_partition,
+                2 * intermediate_size,
+                hidden_size // self.quant_config.group_size,
+                dtype=torch.float32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_scale_inv", w13_weight_scale)
+        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
+
+        w2_weight_scale = torch.nn.Parameter(
+            torch.zeros(
+                num_experts_per_partition,
+                hidden_size,
+                intermediate_size // self.quant_config.group_size,
+                dtype=torch.float32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_scale_inv", w2_weight_scale)
+        set_weight_attrs(w2_weight_scale, extra_weight_attrs)
+
+        # Input scales
+        w13_input_scale = torch.nn.Parameter(
+            torch.ones((num_experts_per_partition, 2), dtype=torch.bfloat16),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_input_scale", w13_input_scale)
+        set_weight_attrs(w13_input_scale, extra_weight_attrs)
+
+        w2_input_scale = torch.nn.Parameter(
+            torch.ones(num_experts_per_partition, dtype=torch.bfloat16),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_input_scale", w2_input_scale)
+        set_weight_attrs(w2_input_scale, extra_weight_attrs)
+
+        # Pre-populate the strides
+        device = layer.w13_weight.device
+
+        self.a_strides1 = torch.full(
+            (num_experts_per_partition, 3),
+            hidden_size,
+            device=device,
+            dtype=torch.int64,
+        )
+        self.c_strides1 = torch.full(
+            (num_experts_per_partition, 3),
+            2 * intermediate_size,
+            device=device,
+            dtype=torch.int64,
+        )
+        self.a_strides2 = torch.full(
+            (num_experts_per_partition, 3),
+            intermediate_size,
+            device=device,
+            dtype=torch.int64,
+        )
+        self.c_strides2 = torch.full(
+            (num_experts_per_partition, 3),
+            hidden_size,
+            device=device,
+            dtype=torch.int64,
+        )
+        self.b_strides1 = self.a_strides1
+        self.s_strides13 = self.c_strides1
+        self.b_strides2 = self.a_strides2
+        self.s_strides2 = self.c_strides2
+
+        self.expert_offsets = torch.empty(
+            (num_experts_per_partition + 1), dtype=torch.int32, device=device
+        )
+        self.problem_sizes1 = torch.empty(
+            (num_experts_per_partition, 3), dtype=torch.int32, device=device
+        )
+        self.problem_sizes2 = torch.empty(
+            (num_experts_per_partition, 3), dtype=torch.int32, device=device
+        )
+
+        return
+
+    def _interleave_scales(self, scales: torch.Tensor) -> torch.Tensor:
+        """Interleave scales in groups of 4 similar to TRT-LLM implementation."""
+        s_shape = scales.shape
+        # Reshape to separate groups of 4
+        scales_interleaved = scales.reshape(
+            s_shape[0], s_shape[1], (s_shape[2] // 4), 4
+        )
+        # Permute dimensions to interleave
+        scales_interleaved = scales_interleaved.permute(0, 2, 1, 3)
+        # Reshape back to original dimensions but with interleaved values
+        scales_interleaved = scales_interleaved.reshape(
+            s_shape[0], s_shape[2] // 4, s_shape[1] * 4
+        )
+        return scales_interleaved.contiguous()
+
+    def process_weights_after_loading(self, layer: Module) -> None:
+        dtype = torch.bfloat16
+        device = layer.w2_weight.device
+
+        # Interleave w13_weight_scale (gate_up_proj)
+        w13_weight_scale = layer.w13_weight_scale_inv.to(dtype)
+        w13_weight_scale = self._interleave_scales(w13_weight_scale)
+        layer.w13_weight_scale_inv = Parameter(w13_weight_scale, requires_grad=False)
+
+        # Interleave w2_weight_scale (down_proj)
+        w2_weight_scale = layer.w2_weight_scale_inv.to(dtype)
+        w2_weight_scale = self._interleave_scales(w2_weight_scale)
+        layer.w2_weight_scale_inv = Parameter(w2_weight_scale, requires_grad=False)
+
+        # Process input scales
+        w13_input_scale_max = layer.w13_input_scale.max().to(dtype).item()
+        new_w13_input_scale = torch.tensor(
+            [w13_input_scale_max],
+            dtype=dtype,
+            device=device,
+        )
+        layer.w13_input_scale = Parameter(new_w13_input_scale, requires_grad=False)
+
+        w2_input_scale_max = layer.w2_input_scale.max().to(dtype).item()
+        new_w2_input_scale = torch.tensor(
+            [w2_input_scale_max], dtype=dtype, device=device
+        )
+        layer.w2_input_scale = Parameter(new_w2_input_scale, requires_grad=False)