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

sglang/srt/layers/quantization/awq.py

@@ -0,0 +1,200 @@
+# SPDX-License-Identifier: Apache-2.0
+import logging
+from typing import Any, Dict, List, Optional
+
+import torch
+from sgl_kernel import awq_dequantize
+
+from sglang.srt.layers.linear import (
+    LinearBase,
+    LinearMethodBase,
+    UnquantizedLinearMethod,
+)
+from sglang.srt.layers.parameter import GroupQuantScaleParameter, PackedvLLMParameter
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+
+logger = logging.getLogger(__name__)
+
+
+def is_layer_skipped_awq(prefix: str, modules_to_not_convert: List[str]):
+    return any(module_name in prefix for module_name in modules_to_not_convert)
+
+
+class AWQConfig(QuantizationConfig):
+    """Config class for AWQ.
+
+    Reference: https://arxiv.org/abs/2306.00978
+    """
+
+    def __init__(
+        self,
+        weight_bits: int,
+        group_size: int,
+        zero_point: bool,
+        modules_to_not_convert: Optional[List[str]] = None,
+    ) -> None:
+        super().__init__()
+        self.weight_bits = weight_bits
+        self.group_size = group_size
+        self.zero_point = zero_point
+        self.modules_to_not_convert = modules_to_not_convert or []
+
+        if self.weight_bits != 4:
+            raise ValueError(
+                "Currently, only 4-bit weight quantization is supported for "
+                f"AWQ, but got {self.weight_bits} bits."
+            )
+        self.pack_factor = 32 // self.weight_bits
+
+    def __repr__(self) -> str:
+        return (
+            f"AWQConfig(weight_bits={self.weight_bits}, "
+            f"group_size={self.group_size}, "
+            f"zero_point={self.zero_point}, "
+            f"modules_to_not_convert={self.modules_to_not_convert})"
+        )
+
+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+    def get_name(self) -> str:
+        return "awq"
+
+    def get_supported_act_dtypes(self) -> List[torch.dtype]:
+        return [torch.half]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # The AWQ kernel only supports Turing or newer GPUs.
+        return 75
+
+    @staticmethod
+    def get_config_filenames() -> List[str]:
+        return [
+            "quant_config.json",  # E.g., casperhansen/vicuna-7b-v1.5-awq
+            # E.g., abhinavkulkarni/mosaicml-mpt-7b-instruct-w4-g128-awq
+            "quantize_config.json",
+        ]
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "AWQConfig":
+        weight_bits = cls.get_from_keys(config, ["w_bit", "bits"])
+        group_size = cls.get_from_keys(config, ["q_group_size", "group_size"])
+        zero_point = cls.get_from_keys(config, ["zero_point"])
+        modules_to_not_convert = cls.get_from_keys_or(
+            config, ["modules_to_not_convert"], None
+        )
+        return cls(weight_bits, group_size, zero_point, modules_to_not_convert)
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional["LinearMethodBase"]:
+
+        if isinstance(layer, LinearBase):
+            if is_layer_skipped_awq(prefix, self.modules_to_not_convert):
+                return UnquantizedLinearMethod()
+            return AWQLinearMethod(self)
+        return None
+
+
+class AWQLinearMethod(LinearMethodBase):
+    """Linear method for AWQ.
+
+    Args:
+        quant_config: The AWQ quantization config.
+    """
+
+    def __init__(self, quant_config: AWQConfig):
+        self.quant_config = quant_config
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: List[int],
+        input_size: int,
+        output_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        if input_size_per_partition % self.quant_config.group_size != 0:
+            raise ValueError(
+                "The input size is not aligned with the quantized "
+                "weight shape. This can be caused by too large "
+                "tensor parallel size."
+            )
+
+        output_size_per_partition = sum(output_partition_sizes)
+        if output_size_per_partition % self.quant_config.pack_factor != 0:
+            raise ValueError(
+                "The output size is not aligned with the quantized "
+                "weight shape. This can be caused by too large "
+                "tensor parallel size."
+            )
+
+        weight_loader = extra_weight_attrs.get("weight_loader")
+        qweight = PackedvLLMParameter(
+            data=torch.empty(
+                input_size_per_partition,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            output_dim=1,
+            packed_dim=1,
+            packed_factor=self.quant_config.pack_factor,
+            weight_loader=weight_loader,
+        )
+
+        qzeros = PackedvLLMParameter(
+            data=torch.empty(
+                input_size_per_partition // self.quant_config.group_size,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            output_dim=1,
+            packed_dim=1,
+            packed_factor=self.quant_config.pack_factor,
+            weight_loader=weight_loader,
+        )
+
+        scales = GroupQuantScaleParameter(
+            data=torch.empty(
+                input_size_per_partition // self.quant_config.group_size,
+                output_size_per_partition,
+                dtype=params_dtype,
+            ),
+            input_dim=0,
+            output_dim=1,
+            weight_loader=weight_loader,
+        )
+
+        layer.register_parameter("qweight", qweight)
+        layer.register_parameter("qzeros", qzeros)
+        layer.register_parameter("scales", scales)
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        layer.qweight = torch.nn.Parameter(layer.qweight.data, requires_grad=False)
+        layer.qzeros = torch.nn.Parameter(layer.qzeros.data, requires_grad=False)
+        layer.scales = torch.nn.Parameter(layer.scales.data, requires_grad=False)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        qweight = layer.qweight
+        scales = layer.scales
+        qzeros = layer.qzeros
+        pack_factor = self.quant_config.pack_factor
+        out_shape = x.shape[:-1] + (qweight.shape[-1] * pack_factor,)
+        reshaped_x = x.reshape(-1, x.shape[-1])
+
+        out = awq_dequantize(qweight, scales, qzeros)
+        out = torch.matmul(reshaped_x, out)
+
+        if bias is not None:
+            out.add_(bias)
+        return out.reshape(out_shape)

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

@@ -23,7 +23,6 @@ from sglang.srt.layers.linear import (
     LinearMethodBase,
     UnquantizedLinearMethod,
 )
-from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.quantization.base_config import (
     QuantizationConfig,
     QuantizeMethodBase,
@@ -123,6 +122,8 @@ class CompressedTensorsConfig(QuantizationConfig):
                 return UnquantizedLinearMethod()
             layer.scheme = scheme
             return CompressedTensorsLinearMethod(self)
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+
         if isinstance(layer, FusedMoE):
             return CompressedTensorsMoEMethod.get_moe_method(self)
         return None

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

@@ -4,18 +4,19 @@
 import enum
 import logging
 from enum import Enum
-from typing import Callable, List, Optional
+from typing import TYPE_CHECKING, Callable, List, Optional
 
 import torch
 from compressed_tensors import CompressionFormat
 from compressed_tensors.quantization import QuantizationStrategy
 
-from sglang.srt.layers.moe.fused_moe_triton import (
-    FusedMoE,
-    FusedMoEMethodBase,
-    FusedMoeWeightScaleSupported,
-)
-from sglang.srt.layers.moe.topk import select_experts
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.fused_moe_triton import (
+        FusedMoE,
+        FusedMoEMethodBase,
+        FusedMoeWeightScaleSupported,
+    )
+
 from sglang.srt.layers.quantization.fp8_utils import normalize_e4m3fn_to_e4m3fnuz
 from sglang.srt.layers.quantization.utils import (
     all_close_1d,
@@ -55,7 +56,13 @@ __all__ = [
 ]
 
 
-class CompressedTensorsMoEMethod(FusedMoEMethodBase):
+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)
 
     @staticmethod
     def get_moe_method(
@@ -85,6 +92,11 @@ 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(
@@ -112,6 +124,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
         params_dtype: torch.dtype,
         **extra_weight_attrs,
     ):
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
 
         params_dtype = torch.float8_e4m3fn
 
@@ -270,8 +283,11 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
         scoring_func: str = "softmax",
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
+        inplace: bool = True,
+        no_combine: bool = False,
     ) -> torch.Tensor:
-        from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_experts
+        from sglang.srt.layers.moe.fused_moe_triton import fused_experts
+        from sglang.srt.layers.moe.topk import select_experts
 
         topk_weights, topk_ids = select_experts(
             hidden_states=x,
@@ -291,7 +307,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
             layer.w2_weight,
             topk_weights=topk_weights,
             topk_ids=topk_ids,
-            inplace=True,
+            inplace=inplace,
             activation=activation,
             use_fp8_w8a8=True,
             w1_scale=layer.w13_weight_scale,
@@ -306,6 +322,11 @@ 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.
@@ -617,6 +638,9 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
     ) -> 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(

sglang/srt/layers/quantization/fp8_kernel.py

@@ -24,6 +24,7 @@ import triton.language as tl
 
 from sglang.srt.utils import (
     direct_register_custom_op,
+    get_bool_env_var,
     get_device_core_count,
     get_device_name,
     get_device_sm,
@@ -43,7 +44,7 @@ if _is_cuda:
     from sgl_kernel import sgl_per_token_group_quant_fp8, sgl_per_token_quant_fp8
 
     sm_version = get_device_sm()
-    if sm_version >= 90 and int(os.getenv("SGL_ENABLE_JIT_DEEPGEMM", "1")):
+    if sm_version >= 90 and get_bool_env_var("SGL_ENABLE_JIT_DEEPGEMM", default="true"):
         _enable_jit_deepgemm = True
 
 

sglang/srt/layers/quantization/fp8_utils.py

@@ -457,12 +457,9 @@ class Fp8LinearOp:
                 qinput, x_scale = sgl_scaled_fp8_quant(
                     input_2d,
                     input_scale,
+                    num_token_padding=self.output_padding,
                     use_per_token_if_dynamic=use_per_token_if_dynamic,
                 )
-                if self.output_padding:
-                    pad_size = max(self.output_padding - qinput.shape[0], 0)
-                    if pad_size > 0:
-                        qinput = torch.nn.functional.pad(qinput, (0, 0, 0, pad_size))
             else:
                 qinput, x_scale = ops.scaled_fp8_quant(
                     input_2d,

sglang/srt/layers/quantization/gptq.py

@@ -11,12 +11,29 @@ from sglang.srt.utils import is_cuda
 _is_cuda = is_cuda()
 
 try:
-    import vllm
+    from vllm.model_executor.layers.quantization.base_config import QuantizeMethodBase
+    from vllm.model_executor.layers.quantization.gptq import GPTQLinearMethod
+    from vllm.model_executor.layers.quantization.gptq_marlin import (
+        GPTQMarlinLinearMethod,
+        GPTQMarlinMoEMethod,
+    )
+    from vllm.model_executor.layers.quantization.marlin import MarlinLinearMethod
+    from vllm.model_executor.layers.quantization.utils.marlin_utils import (
+        check_marlin_supported,
+    )
+    from vllm.scalar_type import scalar_types
 
     VLLM_AVAILABLE = True
 except ImportError:
     VLLM_AVAILABLE = False
 
+    GPTQLinearMethod = MarlinLinearMethod = QuantizeMethodBase = Any
+
+    class scalar_types:
+        uint4b8 = "uint4b8"
+        uint8b128 = "uint8b128"
+
+
 logger = logging.getLogger(__name__)
 
 
@@ -117,12 +134,8 @@ class GPTQConfig(QuantizationConfig):
 
     def get_quant_method(
         self, layer: torch.nn.Module, prefix: str
-    ) -> Optional["GPTQLinearMethod"]:
-        if not VLLM_AVAILABLE:
-            raise ImportError("vllm is not installed")
-
-        from vllm.model_executor.layers.quantization.gptq import GPTQLinearMethod
-
+    ) -> Optional[GPTQLinearMethod]:
+        # Delay the import to avoid circular dependency
         from sglang.srt.layers.quantization import get_linear_quant_method
 
         return get_linear_quant_method(self, layer, prefix, GPTQLinearMethod)
@@ -131,16 +144,11 @@ class GPTQConfig(QuantizationConfig):
 class GPTQMarlinConfig(QuantizationConfig):
     """Config class for GPTQ Marlin"""
 
-    if VLLM_AVAILABLE:
-        from vllm.scalar_type import scalar_types
-
-        # (num_bits, is_sym) -> quant_type
-        TYPE_MAP = {
-            (4, True): scalar_types.uint4b8,
-            (8, True): scalar_types.uint8b128,
-        }
-    else:
-        raise ImportError("vllm is not installed")
+    # (num_bits, is_sym) -> quant_type
+    TYPE_MAP = {
+        (4, True): scalar_types.uint4b8,
+        (8, True): scalar_types.uint8b128,
+    }
 
     def __init__(
         self,
@@ -197,6 +205,7 @@ class GPTQMarlinConfig(QuantizationConfig):
                 "Unsupported quantization config: " f"bits={weight_bits}, sym={is_sym}"
             )
 
+        # (num_bits, is_sym) -> quant_type
         self.quant_type = self.TYPE_MAP[(weight_bits, is_sym)]
 
     def __repr__(self) -> str:
@@ -278,15 +287,8 @@ class GPTQMarlinConfig(QuantizationConfig):
 
     def get_quant_method(
         self, layer: torch.nn.Module, prefix: str
-    ) -> Optional["QuantizeMethodBase"]:
-        if not VLLM_AVAILABLE:
-            raise ImportError("vllm is not installed")
-
-        from vllm.model_executor.layers.quantization.gptq_marlin import (
-            GPTQMarlinLinearMethod,
-            GPTQMarlinMoEMethod,
-        )
-
+    ) -> Optional[QuantizeMethodBase]:
+        # Delay the import to avoid circular dependency
         from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
         from sglang.srt.layers.quantization import get_linear_quant_method
 
@@ -304,19 +306,12 @@ class GPTQMarlinConfig(QuantizationConfig):
 
     @classmethod
     def is_gptq_marlin_compatible(cls, quant_config: Dict[str, Any]):
-        if not VLLM_AVAILABLE:
-            return False
-
         quant_method = quant_config.get("quant_method", "").lower()
         num_bits = quant_config.get("bits")
         group_size = quant_config.get("group_size")
         sym = quant_config.get("sym")
         desc_act = quant_config.get("desc_act")
 
-        from vllm.model_executor.layers.quantization.utils.marlin_utils import (
-            check_marlin_supported,
-        )
-
         if not _is_cuda:
             return False
 
@@ -427,13 +422,8 @@ class MarlinConfig(QuantizationConfig):
 
     def get_quant_method(
         self, layer: torch.nn.Module, prefix: str
-    ) -> Optional["MarlinLinearMethod"]:
-        if not VLLM_AVAILABLE:
-            raise ImportError("vllm is not installed")
-
-        from vllm.model_executor.layers.quantization.marlin import MarlinLinearMethod
-
-        # Delay import to avoid circular dependency
+    ) -> Optional[MarlinLinearMethod]:
+        # Delay the import to avoid circular dependency
         from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
 
         if isinstance(layer, LinearBase) or (