sglang 0.4.5.post1__py3-none-any.whl → 0.4.5.post3__py3-none-any.whl

sglang/srt/layers/moe/topk.py

@@ -13,7 +13,6 @@
 # ==============================================================================
 
 import math
-import os
 from typing import Callable, Optional
 
 import torch
@@ -29,6 +28,10 @@ _is_hip = is_hip()
 if _is_cuda:
     from sgl_kernel import moe_fused_gate
 
+if _is_cuda or _is_hip:
+    from sgl_kernel import topk_softmax
+
+
 expert_distribution_recorder = ExpertDistributionRecorder()
 
 
@@ -59,11 +62,6 @@ def fused_topk(
     topk: int,
     renormalize: bool,
 ):
-    if _is_cuda or _is_hip:
-        from sgl_kernel import topk_softmax
-    else:
-        from vllm import _custom_ops as vllm_ops
-
     assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
 
     M, _ = hidden_states.shape
@@ -76,20 +74,12 @@ def fused_topk(
         M, topk, dtype=torch.int32, device=hidden_states.device
     )
 
-    if _is_cuda or _is_hip:
-        topk_softmax(
-            topk_weights,
-            topk_ids,
-            token_expert_indicies,
-            gating_output.float(),
-        )
-    else:
-        vllm_ops.topk_softmax(
-            topk_weights,
-            topk_ids,
-            token_expert_indicies,
-            gating_output.float(),
-        )
+    topk_softmax(
+        topk_weights,
+        topk_ids,
+        token_expert_indicies,
+        gating_output.float(),
+    )
     del token_expert_indicies
 
     if renormalize:
@@ -108,6 +98,7 @@ def grouped_topk(
     num_expert_group: int = 0,
     topk_group: int = 0,
     n_share_experts_fusion: int = 0,
+    routed_scaling_factor: Optional[float] = None,
 ):
     assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
 
@@ -137,9 +128,7 @@ def grouped_topk(
             dtype=topk_ids.dtype,
             device=topk_ids.device,
         )
-        topk_weights[:, -1] = (
-            topk_weights[:, :-1].sum(dim=-1) / 2.5
-        )  # 2.5 is the routed_scaling_factor.
+        topk_weights[:, -1] = topk_weights[:, :-1].sum(dim=-1) / routed_scaling_factor
 
     if renormalize:
         topk_weights_sum = (
@@ -161,6 +150,7 @@ def biased_grouped_topk_impl(
     num_expert_group: int = 0,
     topk_group: int = 0,
     n_share_experts_fusion: int = 0,
+    routed_scaling_factor: Optional[float] = None,
 ):
     assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch"
 
@@ -197,9 +187,7 @@ def biased_grouped_topk_impl(
             dtype=topk_ids.dtype,
             device=topk_ids.device,
         )
-        topk_weights[:, -1] = (
-            topk_weights[:, :-1].sum(dim=-1) / 2.5
-        )  # 2.5 is the routed_scaling_factor.
+        topk_weights[:, -1] = topk_weights[:, :-1].sum(dim=-1) / routed_scaling_factor
 
     if renormalize:
         topk_weights_sum = (
@@ -226,11 +214,16 @@ def biased_grouped_topk(
     topk_group: int = 0,
     compiled: bool = True,
     n_share_experts_fusion: int = 0,
+    routed_scaling_factor: Optional[float] = None,
 ):
+    assert (
+        routed_scaling_factor is not None
+    ), "routed_scaling_factor is required for biased_grouped_topk"
     # TODO: moe_fused_gate kernel is not supported for n_share_experts_fusion > 0 now.
     if (
         _is_cuda
-        and n_share_experts_fusion == 0
+        and gating_output.shape[1] // num_expert_group
+        <= 32  # moe_fused_gate kernel ensure that num_experts/num_expert_group does not exceed MAX_VPT=32 now. And when kernel can handle MAX_VPT > 32, we can remove this assertion.
         and is_power_of_two(correction_bias.shape[0])
     ):
         return moe_fused_gate(
@@ -239,6 +232,8 @@ def biased_grouped_topk(
             num_expert_group,
             topk_group,
             topk,
+            n_share_experts_fusion,
+            routed_scaling_factor,
         )
     else:
         biased_grouped_topk_fn = (
@@ -257,6 +252,7 @@ def biased_grouped_topk(
             num_expert_group,
             topk_group,
             n_share_experts_fusion=n_share_experts_fusion,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
 
@@ -271,10 +267,9 @@ def select_experts(
     custom_routing_function: Optional[Callable] = None,
     correction_bias: Optional[torch.Tensor] = None,
     torch_native: bool = False,
+    routed_scaling_factor: Optional[float] = None,
 ):
-    n_share_experts_fusion = 0
-    if global_server_args_dict["n_share_experts_fusion"] is not None:
-        n_share_experts_fusion = global_server_args_dict["n_share_experts_fusion"]
+    n_share_experts_fusion = global_server_args_dict["n_share_experts_fusion"]
     # DeekSeek V2/V3/R1 serices models uses grouped_top_k
     if use_grouped_topk:
         assert topk_group is not None
@@ -288,6 +283,7 @@ def select_experts(
                 num_expert_group=num_expert_group,
                 topk_group=topk_group,
                 n_share_experts_fusion=n_share_experts_fusion,
+                routed_scaling_factor=routed_scaling_factor,
             )
         else:
             topk_weights, topk_ids = biased_grouped_topk(
@@ -299,6 +295,7 @@ def select_experts(
                 num_expert_group=num_expert_group,
                 topk_group=topk_group,
                 n_share_experts_fusion=n_share_experts_fusion,
+                routed_scaling_factor=routed_scaling_factor,
             )
     elif torch_native and custom_routing_function is None:
         topk_weights, topk_ids = fused_topk_native(

sglang/srt/layers/parameter.py

@@ -7,8 +7,6 @@ from typing import Callable, Optional, Union
 import torch
 from torch.nn import Parameter
 
-from sglang.srt.distributed import get_tensor_model_parallel_rank
-
 __all__ = [
     "BasevLLMParameter",
     "PackedvLLMParameter",

sglang/srt/layers/quantization/__init__.py

@@ -290,6 +290,7 @@ def monkey_patch_moe_apply(class_obj: "FusedMoEMethodBase"):
         apply_router_weight_on_input: bool = False,
         inplace: bool = True,
         no_combine: bool = False,
+        routed_scaling_factor: Optional[float] = None,
     ):
         assert activation == "silu"
         assert inplace and not no_combine

sglang/srt/layers/quantization/blockwise_int8.py

@@ -373,6 +373,7 @@ class BlockInt8MoEMethod:
         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
@@ -388,6 +389,7 @@ class BlockInt8MoEMethod:
             num_expert_group=num_expert_group,
             custom_routing_function=custom_routing_function,
             correction_bias=correction_bias,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
         # Expert fusion with INT8 quantization

sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py

@@ -1,4 +1,4 @@
-# Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization/compressed_tensors
+# Adapted from https://github.com/vllm-project/vllm/tree/v0.8.2/vllm/model_executor/layers/quantization/compressed_tensors
 # SPDX-License-Identifier: Apache-2.0
 
 import logging
@@ -33,13 +33,20 @@ from sglang.srt.layers.quantization.compressed_tensors.compressed_tensors_moe im
 from sglang.srt.layers.quantization.compressed_tensors.schemes import (
     CompressedTensorsScheme,
     CompressedTensorsW8A8Fp8,
+    CompressedTensorsW8A16Fp8,
 )
 from sglang.srt.layers.quantization.compressed_tensors.utils import (
     find_matched_target,
     is_activation_quantization_format,
     should_ignore_layer,
 )
-from sglang.srt.layers.quantization.kv_cache import BaseKVCacheMethod
+
+try:
+    import vllm
+
+    VLLM_AVAILABLE = True
+except ImportError:
+    VLLM_AVAILABLE = False
 
 logger = logging.getLogger(__name__)
 

sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -1,22 +1,16 @@
-# Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization/compressed_tensors
+# Adapted from https://github.com/vllm-project/vllm/tree/v0.8.2/vllm/model_executor/layers/quantization/compressed_tensors
 # SPDX-License-Identifier: Apache-2.0
 
 import enum
 import logging
 from enum import Enum
-from typing import TYPE_CHECKING, Callable, List, Optional
+from typing import Callable, List, Optional
 
 import torch
 from compressed_tensors import CompressionFormat
 from compressed_tensors.quantization import QuantizationStrategy
 
-if TYPE_CHECKING:
-    from sglang.srt.layers.moe.fused_moe_triton import (
-        FusedMoE,
-        FusedMoEMethodBase,
-        FusedMoeWeightScaleSupported,
-    )
-
+from sglang.srt.layers.quantization.fp8_kernel import scaled_fp8_quant
 from sglang.srt.layers.quantization.fp8_utils import normalize_e4m3fn_to_e4m3fnuz
 from sglang.srt.layers.quantization.utils import (
     all_close_1d,
@@ -29,10 +23,9 @@ from sglang.srt.utils import set_weight_attrs
 
 _is_cuda = is_cuda()
 
-if _is_cuda:
-    from sglang.srt.custom_op import scaled_fp8_quant as sgl_scaled_fp8_quant
-else:
+if not _is_cuda:
     from vllm import _custom_ops as vllm_ops
+    from vllm._custom_ops import scaled_fp8_quant
 
 try:
     import vllm
@@ -58,8 +51,6 @@ __all__ = [
 
 class CompressedTensorsMoEMethod:
     def __new__(cls, *args, **kwargs):
-        from sglang.srt.layers.moe.fused_moe_triton import FusedMoEMethodBase
-
         if cls is CompressedTensorsMoEMethod:
             return super().__new__(cls)
         return super().__new__(cls)
@@ -76,7 +67,7 @@ class CompressedTensorsMoEMethod:
         if quant_config._is_wNa16_group_channel(weight_quant, input_quant):
             if not VLLM_AVAILABLE:
                 raise ImportError(
-                    "vllm is not installed, to use CompressedTensorsWNA16MoEMethod, please install vllm"
+                    "vllm is not installed, to use CompressedTensorsWNA16MoEMethod, please install vllm."
                 )
             return CompressedTensorsWNA16MoEMethod(quant_config)
         elif quant_config._is_fp8_w8a8(weight_quant, input_quant):
@@ -92,11 +83,6 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
     def __init__(
         self, quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
     ):
-        from sglang.srt.layers.moe.fused_moe_triton import (
-            FusedMoEMethodBase,
-            FusedMoeWeightScaleSupported,
-        )
-
         self.quant_config = quant_config
         self.weight_quant = self.quant_config.target_scheme_map["Linear"].get("weights")
         self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
@@ -267,19 +253,11 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
                         layer.w13_weight[expert_id][start : start + shard_size, :],
                         layer.w13_weight_scale[expert_id][shard_id],
                     )
+                    (
+                        layer.w13_weight[expert_id][start : start + shard_size, :],
+                        _,
+                    ) = scaled_fp8_quant(dq_weight, max_w13_scales[expert_id])
 
-                    if _is_cuda:
-                        (
-                            layer.w13_weight[expert_id][start : start + shard_size, :],
-                            _,
-                        ) = sgl_scaled_fp8_quant(dq_weight, max_w13_scales[expert_id])
-                    else:
-                        (
-                            layer.w13_weight[expert_id][start : start + shard_size, :],
-                            _,
-                        ) = vllm_ops.scaled_fp8_quant(
-                            dq_weight, max_w13_scales[expert_id]
-                        )
                     start += shard_size
 
             layer.w13_weight_scale = torch.nn.Parameter(
@@ -305,6 +283,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
         inplace: bool = True,
         no_combine: bool = False,
         apply_router_weight_on_input: bool = False,
+        routed_scaling_factor: Optional[float] = None,
     ) -> torch.Tensor:
         from sglang.srt.layers.moe.fused_moe_triton import fused_experts
         from sglang.srt.layers.moe.topk import select_experts
@@ -319,6 +298,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
             num_expert_group=num_expert_group,
             custom_routing_function=custom_routing_function,
             correction_bias=correction_bias,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
         return fused_experts(
@@ -345,11 +325,6 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
     def __init__(
         self, quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
     ):
-        from sglang.srt.layers.moe.fused_moe_triton import (
-            FusedMoEMethodBase,
-            FusedMoeWeightScaleSupported,
-        )
-
         self.quant_config = quant_config
         # TODO: @dsikka: refactor this to use schemes as other kernels
         # are supported + check if the layer is being ignored.
@@ -609,7 +584,7 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
                 requires_grad=False,
             )
 
-        marlin_w13_qweight = ops.gptq_marlin_moe_repack(
+        marlin_w13_qweight = vllm_ops.gptq_marlin_moe_repack(
             layer.w13_weight_packed,
             layer.w13_g_idx_sort_indices,
             layer.w13_weight_packed.shape[1] * self.packed_factor,
@@ -617,7 +592,7 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
             self.num_bits,
         )
         replace_tensor("w13_weight_packed", marlin_w13_qweight)
-        marlin_w2_qweight = ops.gptq_marlin_moe_repack(
+        marlin_w2_qweight = vllm_ops.gptq_marlin_moe_repack(
             layer.w2_weight_packed,
             layer.w2_g_idx_sort_indices,
             layer.w2_weight_packed.shape[1] * self.packed_factor,
@@ -660,15 +635,11 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
         scoring_func: str = "softmax",
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
+        routed_scaling_factor: Optional[float] = None,
     ) -> torch.Tensor:
-        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
         from sglang.srt.layers.moe.topk import select_experts
 
         assert activation == "silu", "Only SiLU activation is supported."
-        if not VLLM_AVAILABLE:
-            raise ImportError(
-                "vllm is not installed, to use fused_marlin_moe, please install vllm"
-            )
         if expert_map is not None:
             raise NotImplementedError(
                 "Expert Parallelism is not supported for " "fused Marlin MoE method."
@@ -685,6 +656,7 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
             custom_routing_function=custom_routing_function,
             scoring_func=scoring_func,
             correction_bias=correction_bias,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
         return torch.ops.vllm.fused_marlin_moe(

sglang/srt/layers/quantization/compressed_tensors/schemes/__init__.py

@@ -2,8 +2,10 @@
 
 from .compressed_tensors_scheme import CompressedTensorsScheme
 from .compressed_tensors_w8a8_fp8 import CompressedTensorsW8A8Fp8
+from .compressed_tensors_w8a16_fp8 import CompressedTensorsW8A16Fp8
 
 __all__ = [
     "CompressedTensorsScheme",
     "CompressedTensorsW8A8Fp8",
+    "CompressedTensorsW8A16Fp8",
 ]

sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a16_fp8.py

@@ -0,0 +1,153 @@
+# Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization/compressed_tensors
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Callable, List, Optional
+
+import torch
+from compressed_tensors.quantization import QuantizationStrategy
+
+from sglang.srt.layers.parameter import (
+    ChannelQuantScaleParameter,
+    ModelWeightParameter,
+    PerTensorScaleParameter,
+)
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsScheme,
+)
+from sglang.srt.layers.quantization.utils import convert_to_channelwise
+
+try:
+    from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
+        apply_fp8_marlin_linear,
+        prepare_fp8_layer_for_marlin,
+    )
+
+    MARLIN_FP8_AVAILABLE = True
+except ImportError:
+    MARLIN_FP8_AVAILABLE = False
+
+    def apply_fp8_marlin_linear(*args, **kwargs):
+        raise ImportError("vllm is not installed")
+
+    def prepare_fp8_layer_for_marlin(*args, **kwargs):
+        raise ImportError("vllm is not installed")
+
+
+__all__ = ["CompressedTensorsW8A16Fp8"]
+
+SUPPORTED_STRATEGIES = [QuantizationStrategy.CHANNEL, QuantizationStrategy.TENSOR]
+
+
+class CompressedTensorsW8A16Fp8(CompressedTensorsScheme):
+    def __init__(self, strategy: str, is_static_input_scheme: bool):
+        self.strategy = strategy
+        self.is_static_input_scheme = is_static_input_scheme
+
+        if not MARLIN_FP8_AVAILABLE:
+            raise ImportError(
+                "vllm is not installed. To use CompressedTensorsW8A16Fp8, please install vllm"
+            )
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # ampere and up
+        return 80
+
+    # W8A8-Fp8 kernels support only per-tensor and per-channel cases.
+    # So if we have a fused module (QKV, MLP) with per tensor scales,
+    # we expand each scale to its shard's channels.
+    def process_weights_after_loading(self, layer) -> None:
+        if self.strategy == QuantizationStrategy.TENSOR:
+            ws_channelwise = convert_to_channelwise(
+                layer.weight_scale, layer.logical_widths
+            )
+            layer.weight_scale = torch.nn.Parameter(ws_channelwise, requires_grad=False)
+        else:
+            # required by torch.compile to be torch.nn.Parameter
+            layer.weight_scale = torch.nn.Parameter(
+                layer.weight_scale.data, requires_grad=False
+            )
+
+        # Weights must be transposed for marlin
+        layer.weight = torch.nn.Parameter(layer.weight.t(), requires_grad=False)
+
+        if self.is_static_input_scheme:
+            # required by torch.compile to be torch.nn.Parameter
+            layer.input_scale = torch.nn.Parameter(
+                layer.input_scale.data, requires_grad=False
+            )
+        prepare_fp8_layer_for_marlin(layer, strategy="channel")
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size: int,
+        output_partition_sizes: List[int],
+        input_size_per_partition: int,
+        params_dtype: torch.dtype,
+        weight_loader: Callable,
+        **kwargs,
+    ):
+        output_size_per_partition = sum(output_partition_sizes)
+        layer.logical_widths = output_partition_sizes
+        layer.input_size_per_partition = input_size_per_partition
+        layer.output_size_per_partition = output_size_per_partition
+        layer.orig_dtype = params_dtype
+
+        # WEIGHT
+        weight = ModelWeightParameter(
+            data=torch.empty(
+                output_size_per_partition,
+                input_size_per_partition,
+                dtype=torch.float8_e4m3fn,
+            ),
+            input_dim=1,
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("weight", weight)
+
+        # WEIGHT SCALE
+        if self.strategy == QuantizationStrategy.CHANNEL:
+            weight_scale = ChannelQuantScaleParameter(
+                data=torch.empty((sum(output_partition_sizes), 1), dtype=torch.float32),
+                output_dim=0,
+                weight_loader=weight_loader,
+            )
+        elif self.strategy == QuantizationStrategy.TENSOR:
+            weight_scale = PerTensorScaleParameter(
+                data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+                weight_loader=weight_loader,
+            )
+        else:
+            raise ValueError(
+                f"Unsupported weight strategy={self.strategy}, "
+                f"supported strategies are {SUPPORTED_STRATEGIES}"
+            )
+
+        weight_scale[:] = torch.finfo(torch.float32).min
+        layer.register_parameter("weight_scale", weight_scale)
+
+        # INPUT SCALE (to deal with converted checkpoints)
+        if self.is_static_input_scheme:
+            input_scale = PerTensorScaleParameter(
+                data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+                weight_loader=weight_loader,
+            )
+            layer.register_parameter("input_scale", input_scale)
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        return apply_fp8_marlin_linear(
+            input=x,
+            weight=layer.weight,
+            weight_scale=layer.weight_scale,
+            workspace=layer.workspace,
+            size_n=layer.output_size_per_partition,
+            size_k=layer.input_size_per_partition,
+            bias=bias,
+        )

sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py

@@ -16,8 +16,7 @@ from sglang.srt.layers.quantization.compressed_tensors.schemes import (
     CompressedTensorsScheme,
 )
 from sglang.srt.layers.quantization.fp8_utils import (
-    Fp8LinearOp,
-    maybe_create_device_identity,
+    apply_fp8_linear,
     normalize_e4m3fn_to_e4m3fnuz,
 )
 from sglang.srt.layers.quantization.utils import is_fp8_fnuz, requantize_with_max_scale
@@ -30,7 +29,6 @@ class CompressedTensorsW8A8Fp8(CompressedTensorsScheme):
     def __init__(self, strategy: str, is_static_input_scheme: bool):
         self.strategy = strategy
         self.is_static_input_scheme = is_static_input_scheme
-        self.fp8_linear = Fp8LinearOp(use_per_token_if_dynamic=True)
 
     @classmethod
     def get_min_capability(cls) -> int:
@@ -99,8 +97,6 @@ class CompressedTensorsW8A8Fp8(CompressedTensorsScheme):
         weight_loader: Callable,
         **kwargs,
     ):
-        maybe_create_device_identity()
-
         output_size_per_partition = sum(output_partition_sizes)
         layer.logical_widths = output_partition_sizes
 
@@ -152,11 +148,12 @@ class CompressedTensorsW8A8Fp8(CompressedTensorsScheme):
         x: torch.Tensor,
         bias: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
-
-        return self.fp8_linear.apply(
+        return apply_fp8_linear(
             input=x,
             weight=layer.weight,
             weight_scale=layer.weight_scale,
             input_scale=layer.input_scale,
             bias=bias,
+            use_per_token_if_dynamic=True,
+            compressed_tensor_quant=True,
         )