sglang 0.4.8.post1__py3-none-any.whl → 0.4.9.post1__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

@@ -131,7 +131,7 @@ class MoeWNA16Config(QuantizationConfig):
         capability_tuple = get_device_capability()
         device_capability = (
             -1
-            if capability_tuple is None
+            if all(capability is None for capability in capability_tuple)
             else capability_tuple[0] * 10 + capability_tuple[1]
         )
         # Avoid circular import

sglang/srt/layers/quantization/quant_utils.py

@@ -0,0 +1,166 @@
+# SPDX-License-Identifier: Apache-2.0
+# Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/quantization/utils/quant_utils.py
+
+from typing import Optional
+
+import numpy
+import torch
+from sgl_kernel.scalar_type import ScalarType
+
+
+def get_pack_factor(num_bits):
+    assert 32 % num_bits == 0, f"Unsupported num_bits = {num_bits}"
+    return 32 // num_bits
+
+
+def pack_cols(
+    q_w: torch.Tensor,
+    num_bits: int,
+    size_k: int,
+    size_n: int,
+):
+    assert q_w.shape == (size_k, size_n)
+
+    pack_factor = get_pack_factor(num_bits)
+    assert size_n % pack_factor == 0
+
+    orig_device = q_w.device
+
+    q_w = q_w.cpu().numpy().astype(numpy.uint32)
+
+    q_res = numpy.zeros((size_k, size_n // pack_factor), dtype=numpy.uint32)
+
+    for i in range(pack_factor):
+        q_res |= q_w[:, i::pack_factor] << num_bits * i
+
+    q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device)
+    q_res = q_res.contiguous()
+
+    return q_res
+
+
+def unpack_cols(
+    packed_q_w: torch.Tensor,
+    num_bits: int,
+    size_k: int,
+    size_n: int,
+):
+    pack_factor = get_pack_factor(num_bits)
+    assert size_n % pack_factor == 0
+    assert packed_q_w.shape == (
+        size_k,
+        size_n // pack_factor,
+    ), "packed_q_w.shape = {} size_k = {}, size_n = {} pack_Factor = {}".format(
+        packed_q_w.shape, size_k, size_n, pack_factor
+    )
+
+    orig_device = packed_q_w.device
+
+    packed_q_w_cpu = packed_q_w.cpu().numpy().astype(numpy.uint32)
+    q_res = numpy.zeros((size_k, size_n), dtype=numpy.uint32)
+
+    mask = (1 << num_bits) - 1
+    for i in range(pack_factor):
+        vals = packed_q_w_cpu & mask
+        packed_q_w_cpu >>= num_bits
+        q_res[:, i::pack_factor] = vals
+
+    q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device)
+    q_res = q_res.contiguous()
+
+    return q_res
+
+
+def quantize_weights(
+    w: torch.Tensor,
+    quant_type: ScalarType,
+    group_size: Optional[int],
+    zero_points: bool = False,
+    ref_zero_points_after_scales: bool = False,
+):
+    assert (
+        quant_type.is_integer()
+    ), "Floating point quantization may work but has not been tested"
+    assert not zero_points or group_size is not None, (
+        "to have group zero points, group_size must be provided "
+        "(-1 group_size is channelwise)"
+    )
+
+    orig_device = w.device
+    orig_type = w.dtype
+    size_k, size_n = w.shape
+
+    assert w.is_floating_point(), "w must be float"
+
+    if group_size == -1:
+        group_size = size_k
+
+    # Reshape to [groupsize, -1]
+    if group_size is not None and group_size < size_k:
+        w = w.reshape((-1, group_size, size_n))
+        w = w.permute(1, 0, 2)
+        w = w.reshape((group_size, -1))
+
+    # Compute scale for each group
+    max_val = torch.max(w, 0, keepdim=True).values
+    min_val = torch.min(w, 0, keepdim=True).values
+
+    max_q_val = quant_type.max()
+    min_q_val = quant_type.min()
+
+    w_s = torch.Tensor([1.0]).to(w.device)  # unscaled case
+    maybe_w_zp = None
+    if group_size is not None:
+        if zero_points:
+            assert not quant_type.is_signed() and quant_type.max() > 0
+            w_s = (max_val - min_val).clamp(min=1e-5) / quant_type.max()
+            maybe_w_zp = (
+                torch.round(torch.abs(min_val / w_s)).clamp(min_q_val, max_q_val).int()
+            )
+        else:
+            # If the bias is such that there are no possible negative/positive
+            #  values, set the max value to inf to avoid divide by 0
+            w_s = torch.max(
+                abs(max_val / (max_q_val if max_q_val != 0 else torch.inf)),
+                abs(min_val / (min_q_val if min_q_val != 0 else torch.inf)),
+            )
+
+    # Quantize
+    w_q = torch.round(w / w_s).int() + (maybe_w_zp if zero_points else 0)
+    w_q = torch.clamp(w_q, min_q_val, max_q_val)
+
+    # Compute ref (dequantized)
+    # For some kernels (namely Machete) the zero-points are applied after the
+    # scales are applied, for this case computing the reference in similar way
+    # allows us to use tighter error tolerances in our unit tests.
+    if ref_zero_points_after_scales and maybe_w_zp is not None:
+        w_ref = w_q.to(orig_type) * w_s - maybe_w_zp.to(orig_type) * w_s
+    else:
+        w_ref = (w_q - (maybe_w_zp if zero_points else 0)).to(orig_type) * w_s
+
+    if quant_type.has_bias():
+        w_q += quant_type.bias
+
+    # Restore original shapes
+    if group_size is not None and group_size < size_k:
+
+        def reshape_w(w):
+            w = w.reshape((group_size, -1, size_n))
+            w = w.permute(1, 0, 2)
+            w = w.reshape((size_k, size_n)).contiguous()
+            return w
+
+        w_q = reshape_w(w_q)
+        w_ref = reshape_w(w_ref)
+        w_s = w_s.reshape((-1, size_n)).contiguous()
+
+    if maybe_w_zp is not None:
+        maybe_w_zp = maybe_w_zp.reshape((-1, size_n)).contiguous()
+        maybe_w_zp = maybe_w_zp.to(device=orig_device)
+
+    return (
+        w_ref.to(device=orig_device),
+        w_q.to(device=orig_device),
+        w_s if group_size is not None else None,
+        maybe_w_zp,
+    )