sglang 0.4.3.post2__py3-none-any.whl → 0.4.3.post4__py3-none-any.whl

sglang/srt/layers/quantization/fp8.py

@@ -51,6 +51,10 @@ ACTIVATION_SCHEMES = ["static", "dynamic"]
 
 is_hip_ = is_hip()
 
+if is_hip_:
+    from aiter.fused_moe_bf16_asm import asm_moe
+    from aiter.ops.shuffle import shuffle_weight
+
 logger = logging.getLogger(__name__)
 
 
@@ -533,6 +537,20 @@ class Fp8MoEMethod:
             )
             layer.register_parameter("w13_weight_scale", w13_weight_scale)
             layer.register_parameter("w2_weight_scale", w2_weight_scale)
+
+            if is_hip_ and get_bool_env_var("CK_MOE"):
+                # ROCm - using column scaling, duplicate scaling numbers in case per tensor scaling
+                w13_weight_scale1 = torch.nn.Parameter(
+                    torch.ones(num_experts, 2 * intermediate_size, dtype=torch.float32),
+                    requires_grad=False,
+                )
+                w2_weight_scale1 = torch.nn.Parameter(
+                    torch.ones(num_experts, hidden_size, dtype=torch.float32),
+                    requires_grad=False,
+                )
+                layer.register_parameter("w13_weight_scale1", w13_weight_scale1)
+                layer.register_parameter("w2_weight_scale1", w2_weight_scale1)
+
         # Add the quantization method used (per tensor/grouped/channel)
         # to ensure the weight scales are loaded in properly
         extra_weight_attrs.update(
@@ -602,6 +620,15 @@ class Fp8MoEMethod:
                     w2_weight_scale, requires_grad=False
                 )
                 layer.w2_input_scale = None
+
+                if get_bool_env_var("CK_MOE"):
+                    # Pre-shuffle weights
+                    layer.w13_weight.data = shuffle_weight(
+                        layer.w13_weight.contiguous(), (16, 16)
+                    )
+                    layer.w2_weight.data = shuffle_weight(
+                        layer.w2_weight.contiguous(), (16, 16)
+                    )
             return
         # If checkpoint is fp16 or bfloat16, quantize in place.
         if not self.quant_config.is_checkpoint_fp8_serialized:
@@ -640,6 +667,9 @@ class Fp8MoEMethod:
                         requires_grad=False,
                     )
                     torch.cuda.empty_cache()
+                    # ROCm (CK_MOE): using column-wise scaling
+                    layer.w13_weight_scale1 *= layer.w13_weight_scale.unsqueeze(-1)
+                    layer.w2_weight_scale1 *= layer.w2_weight_scale.unsqueeze(-1)
                 elif get_bool_env_var("MOE_PADDING"):
                     # If ROCm, apply weight padding (min. Mem channel contention) only if set
                     layer.w13_weight = torch.nn.Parameter(
@@ -744,6 +774,9 @@ class Fp8MoEMethod:
                         requires_grad=False,
                     )
                     torch.cuda.empty_cache()
+                    # ROCm (CK_MOE): using column-wise scaling
+                    layer.w13_weight_scale1 *= layer.w13_weight_scale.unsqueeze(-1)
+                    layer.w2_weight_scale1 *= layer.w2_weight_scale.unsqueeze(-1)
                 elif get_bool_env_var("MOE_PADDING"):
                     # If ROCm, apply weight padding (min. Mem channel contention) only if set
                     layer.w13_weight = torch.nn.Parameter(
@@ -771,6 +804,8 @@ class Fp8MoEMethod:
         custom_routing_function: Optional[Callable] = None,
         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.topk import select_experts
@@ -788,33 +823,38 @@ class Fp8MoEMethod:
             correction_bias=correction_bias,
         )
 
-        if is_hip_ and get_bool_env_var("CK_MOE"):
-            import ater
-            from ater.fused_moe import fused_experts_ck
-
-            assert activation == "silu", f"{activation=} is not supported."
-
-            return fused_experts_ck(
-                x,
-                layer.w13_weight,
-                layer.w2_weight,
-                topk_weights=topk_weights,
-                topk_ids=topk_ids,
-                use_fp8_w8a8=True,
-                w1_scale=(
-                    layer.w13_weight_scale_inv
-                    if self.block_quant
-                    else layer.w13_weight_scale
-                ),
-                w2_scale=(
-                    layer.w2_weight_scale_inv
-                    if self.block_quant
-                    else layer.w2_weight_scale
-                ),
-                a1_scale=layer.w13_input_scale,
-                a2_scale=layer.w2_input_scale,
-            )
-
+        if is_hip_ and get_bool_env_var("CK_MOE") and activation == "silu":
+            # TODO(CK_MOE): FP8 or FP8 block_quant only supports 'silu' for the time-being.
+            assert not no_combine, f"{no_combine=} is not supported."
+            if self.block_quant:
+                return asm_moe(
+                    x,
+                    layer.w13_weight,
+                    layer.w2_weight,
+                    topk_weights,
+                    topk_ids,
+                    layer.w13_weight_scale_inv,
+                    layer.w2_weight_scale_inv,
+                    None,
+                    None,
+                    False,
+                    None,
+                    block_shape=tuple(self.quant_config.weight_block_size),
+                    expert_mask=None,
+                )
+            else:
+                return asm_moe(
+                    x,
+                    layer.w13_weight,
+                    layer.w2_weight,
+                    topk_weights,
+                    topk_ids,
+                    layer.w13_weight_scale1,
+                    layer.w2_weight_scale1,
+                    None,
+                    None,
+                    False,
+                )
         else:
             # Expert fusion with FP8 quantization
             return fused_experts(
@@ -823,7 +863,7 @@ class Fp8MoEMethod:
                 layer.w2_weight,
                 topk_weights=topk_weights,
                 topk_ids=topk_ids,
-                inplace=True,
+                inplace=inplace and not no_combine,
                 activation=activation,
                 use_fp8_w8a8=True,
                 w1_scale=(
@@ -839,6 +879,7 @@ class Fp8MoEMethod:
                 a1_scale=layer.w13_input_scale,
                 a2_scale=layer.w2_input_scale,
                 block_shape=self.quant_config.weight_block_size,
+                no_combine=no_combine,
             )
 
 

sglang/srt/layers/quantization/fp8_utils.py

@@ -1,3 +1,4 @@
+import os
 from typing import List, Optional, Tuple
 
 import torch
@@ -7,9 +8,12 @@ from sglang.srt.layers.quantization.fp8_kernel import (
     per_token_group_quant_fp8,
     w8a8_block_fp8_matmul,
 )
-from sglang.srt.utils import is_hip
+from sglang.srt.utils import get_bool_env_var, is_hip
 
 is_hip_ = is_hip()
+if is_hip_ and get_bool_env_var("CK_MOE"):
+    from aiter import gemm_a8w8_blockscale
+
 _is_cuda = torch.cuda.is_available() and torch.version.cuda
 if _is_cuda:
     from sgl_kernel import fp8_blockwise_scaled_mm
@@ -40,6 +44,8 @@ def normalize_e4m3fn_to_e4m3fnuz(
 
 
 def cutlass_block_fp8_supported() -> bool:
+    if os.environ.get("SUPPORT_CUTLASS_BLOCK_FP8") is None:
+        return False
     if _is_cuda:
         major, minor = torch.cuda.get_device_capability()
         sm_version = major * 10 + minor
@@ -75,6 +81,16 @@ def apply_w8a8_block_fp8_linear(
         output = fp8_blockwise_scaled_mm(
             q_input, weight.T, x_scale, weight_scale.T, out_dtype=input.dtype
         )
+    elif is_hip_ and get_bool_env_var("CK_MOE"):
+        q_input, x_scale = per_token_group_quant_fp8(
+            input_2d, block_size[1], column_major_scales=False
+        )
+        output = torch.zeros(
+            [q_input.shape[0], weight.shape[0]],
+            dtype=input.dtype,
+            device=q_input.device,
+        )
+        gemm_a8w8_blockscale(q_input, weight, x_scale, weight_scale, output)
     else:
         q_input, x_scale = per_token_group_quant_fp8(
             input_2d, block_size[1], column_major_scales=False

sglang/srt/layers/quantization/gptq.py

@@ -0,0 +1,416 @@
+import logging
+from fractions import Fraction
+from typing import Any, Dict, List, Optional, Union
+
+import torch
+from vllm.scalar_type import scalar_types
+
+from sglang.srt.layers.linear import LinearBase
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+
+logger = logging.getLogger(__name__)
+
+
+class GPTQConfig(QuantizationConfig):
+    """Config class for GPTQ.
+
+    Reference: https://arxiv.org/abs/2210.17323
+    """
+
+    def __init__(
+        self,
+        weight_bits: int,
+        group_size: int,
+        desc_act: bool,
+        lm_head_quantized: bool,
+        dynamic: Dict[str, Dict[str, Union[int, bool]]],
+    ) -> None:
+        # GPTQModel use `dynamic` config property to allow per module
+        # quantization config so each module can be individually optimized.
+        # Format is Dict[str, Dict] where key is a regex string that can
+        # perform both positive ("+:" prefixed) or negative ("-:" prefixed)
+        # matching of a module.
+        # Default to positive match, override base quant config mode, if no
+        # prefix is used. Value is in dict format of field key and override
+        # value.
+        # Negative matching will skip quantization init for this module
+        # entirely:
+        # non-quantized inference. More details and quantization examples can be
+        # found at: https://github.com/ModelCloud/GPTQModel
+        # Example:
+        #  # last 1/2 of the layers 10-21 has 8bit vs 4bit for 0-9
+        #  # last 1/4 of the layers 16-21 has 8bit and group_size 64
+        # dynamic = {
+        #  #`.*\.` matches the layers_node prefix
+        #  # positive match layer 10-15
+        #  r"+:.*\.(?:1[0-5])\..*": {"bits": 8,},
+        #  # positive match layer 16-21
+        #  r"+:.*\.(?:1[6-9]|20|21)\..*": {"bits": 8, "group_size": 64,},
+        #  r"-:.*\.moe\..*": {}, # negative match (skip) all `moe` layers
+        # }
+        super().__init__()
+        self.dynamic = dynamic
+
+        self.weight_bits = weight_bits
+        self.group_size = group_size
+        self.desc_act = desc_act
+        self.lm_head_quantized = lm_head_quantized
+        self.pack_factor = Fraction(32, self.weight_bits)
+        if self.weight_bits not in [2, 3, 4, 8]:
+            raise ValueError(
+                "Currently, only 2/3/4/8-bit weight quantization is "
+                f"supported for GPTQ, but got {self.weight_bits} bits."
+            )
+
+    def __repr__(self) -> str:
+        return (
+            f"GPTQConfig(weight_bits={self.weight_bits}, "
+            f"group_size={self.group_size}, "
+            f"desc_act={self.desc_act}),"
+            f"lm_head_quantized={self.lm_head_quantized}), "
+            f"dynamic={self.dynamic}"
+        )
+
+    def get_scaled_act_names(self) -> List[str]:
+        """Returns the activation function names that should be post-scaled.
+
+        For now, this is only used by AWQ.
+        """
+        raise NotImplementedError
+
+    @classmethod
+    def get_name(cls) -> str:
+        return "gptq"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.half]
+
+    @classmethod
+    # Need to figure it out
+    def get_min_capability(cls) -> int:
+        return 60
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return ["quantize_config.json"]
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "GPTQConfig":
+        dynamic = cls.get_from_keys_or(config, ["dynamic"], default={})
+        dynamic = {} if dynamic is None else dynamic
+
+        weight_bits = cls.get_from_keys(config, ["bits"])
+        group_size = cls.get_from_keys(config, ["group_size"])
+        desc_act = cls.get_from_keys(config, ["desc_act"])
+        lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False)
+        return cls(weight_bits, group_size, desc_act, lm_head_quantized, dynamic)
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional["GPTQLinearMethod"]:
+        from vllm.model_executor.layers.quantization.gptq import GPTQLinearMethod
+
+        from sglang.srt.layers.quantization import get_linear_quant_method
+
+        return get_linear_quant_method(self, layer, prefix, GPTQLinearMethod)
+
+
+class GPTQMarlinConfig(QuantizationConfig):
+    """Config class for GPTQ Marlin"""
+
+    # (num_bits, is_sym) -> quant_type
+    TYPE_MAP = {
+        (4, True): scalar_types.uint4b8,
+        (8, True): scalar_types.uint8b128,
+    }
+
+    def __init__(
+        self,
+        weight_bits: int,
+        group_size: int,
+        desc_act: bool,
+        is_sym: bool,
+        lm_head_quantized: bool,
+        dynamic: Dict[str, Dict[str, Union[int, bool]]],
+        full_config: Dict[str, Any],
+    ) -> None:
+        super().__init__()
+        if desc_act and group_size == -1:
+            # In this case, act_order == True is the same as act_order == False
+            # (since we have only one group per output channel)
+            desc_act = False
+
+        # GPTQModel use `dynamic` config property to allow per module
+        # quantization config so each module can be individually optimized.
+        # Format is Dict[str, Dict] where key is a regex string that can
+        # perform both positive ("+:" prefixed) or negative ("-:" prefixed)
+        # matching of a module.
+        # Default to positive match, override base quant config mode, if no
+        # prefix is used. Value is in dict format of field key and override
+        # value.
+        # Negative matching will skip quantization init for this module
+        # entirely:
+        # non-quantized inference. More details and quantization examples can be
+        # found at: https://github.com/ModelCloud/GPTQModel
+        # Example:
+        #  # last 1/2 of the layers 10-21 has 8bit vs 4bit for 0-9
+        #  # last 1/4 of the layers 16-21 has 8bit and group_size 64
+        # dynamic = {
+        #  #`.*\.` matches the layers_node prefix
+        #  # positive match layer 10-15
+        #  r"+:.*\.(?:1[0-5])\..*": {"bits": 8,},
+        #  # positive match layer 16-21
+        #  r"+:.*\.(?:1[6-9]|20|21)\..*": {"bits": 8, "group_size": 64,},
+        #  r"-:.*\.moe\..*": {}, # negative match (skip) all `moe` layers
+        # }
+        self.dynamic = dynamic
+
+        self.weight_bits = weight_bits
+        self.is_sym = is_sym
+
+        self.pack_factor = 32 // weight_bits  # packed into int32
+        self.group_size = group_size
+        self.desc_act = desc_act
+        self.lm_head_quantized = lm_head_quantized
+        self.full_config = full_config
+
+        if (weight_bits, is_sym) not in self.TYPE_MAP:
+            raise ValueError(
+                "Unsupported quantization config: " f"bits={weight_bits}, sym={is_sym}"
+            )
+
+        self.quant_type = self.TYPE_MAP[(weight_bits, is_sym)]
+
+    def __repr__(self) -> str:
+        return (
+            f"GPTQMarlinConfig(quant_type={self.quant_type}, "
+            f"group_size={self.group_size}, "
+            f"desc_act={self.desc_act}, "
+            f"lm_head_quantized={self.lm_head_quantized}), "
+            f"dynamic={self.dynamic}"
+        )
+
+    def get_scaled_act_names(self) -> List[str]:
+        """Returns the activation function names that should be post-scaled.
+
+        For now, this is only used by AWQ.
+        """
+        raise NotImplementedError
+
+    @classmethod
+    def get_name(cls) -> str:
+        return "gptq_marlin"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.half, torch.bfloat16]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 80
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return ["quantize_config.json"]
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "GPTQMarlinConfig":
+        dynamic = cls.get_from_keys_or(config, ["dynamic"], default={})
+        dynamic = {} if dynamic is None else dynamic
+
+        weight_bits = cls.get_from_keys(config, ["bits"])
+        group_size = cls.get_from_keys(config, ["group_size"])
+        desc_act = cls.get_from_keys(config, ["desc_act"])
+        is_sym = cls.get_from_keys(config, ["sym"])
+        lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False)
+        return cls(
+            weight_bits,
+            group_size,
+            desc_act,
+            is_sym,
+            lm_head_quantized,
+            dynamic,
+            config,
+        )
+
+    @classmethod
+    def override_quantization_method(cls, hf_quant_cfg, user_quant) -> Optional[str]:
+        can_convert = cls.is_gptq_marlin_compatible(hf_quant_cfg)
+
+        is_valid_user_quant = (
+            user_quant is None or user_quant == "marlin" or user_quant == "gptq_marlin"
+        )
+
+        if can_convert and is_valid_user_quant:
+            msg = (
+                "The model is convertible to {} during runtime."
+                " Using {} kernel.".format(cls.get_name(), cls.get_name())
+            )
+            logger.info(msg)
+            return cls.get_name()
+
+        if can_convert and user_quant == "gptq":
+            logger.info(
+                "Detected that the model can run with gptq_marlin"
+                ", however you specified quantization=gptq explicitly,"
+                " so forcing gptq. Use quantization=gptq_marlin for"
+                " faster inference"
+            )
+        return None
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional["QuantizeMethodBase"]:
+        from vllm.model_executor.layers.quantization.gptq_marlin import (
+            GPTQMarlinLinearMethod,
+            GPTQMarlinMoEMethod,
+        )
+
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+        from sglang.srt.layers.quantization import get_linear_quant_method
+
+        if isinstance(layer, FusedMoE):
+            return GPTQMarlinMoEMethod(self)
+            # TODO: re-enable after SGLang syncs with vllm >= 0.7.3
+            # if layer.num_experts > 32:
+            #     # For MoEs with many experts the moe_wna16 kernel is faster
+            #     return MoeWNA16Config.from_config(self.full_config).get_quant_method(
+            #         layer, prefix
+            #     )
+            # else:
+            #     return GPTQMarlinMoEMethod(self)
+        return get_linear_quant_method(self, layer, prefix, GPTQMarlinLinearMethod)
+
+    @classmethod
+    def is_gptq_marlin_compatible(cls, quant_config: Dict[str, Any]):
+        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,
+        )
+        from vllm.platforms import current_platform
+
+        if not current_platform.is_cuda():
+            return False
+
+        if quant_method != "gptq":
+            return False
+
+        # Marlin conversion is only valid if required properties are found
+        if num_bits is None or group_size is None or sym is None or desc_act is None:
+            return False
+
+        if (num_bits, sym) not in cls.TYPE_MAP:
+            return False
+
+        return check_marlin_supported(
+            quant_type=cls.TYPE_MAP[(num_bits, sym)], group_size=group_size
+        )
+
+
+class MarlinConfig(QuantizationConfig):
+    """Config class for Marlin.
+
+    Reference: https://github.com/IST-DASLab/marlin/tree/master
+    """
+
+    def __init__(
+        self,
+        group_size: int,
+        lm_head_quantized: bool,
+    ) -> None:
+        # Group size for the quantization.
+        self.group_size = group_size
+        self.lm_head_quantized = lm_head_quantized
+        if self.group_size != 128 and self.group_size != -1:
+            raise ValueError(
+                "Currently, only group size 128 and -1 (channelwise) "
+                "is supported for Marlin, but got group_size of "
+                f"{self.group_size}"
+            )
+
+        # 4 Bits packed into 32 bit datatype.
+        self.pack_factor = 32 // 4
+
+        # Tile size used by marlin kernels.
+        self.tile_size = 16
+
+        # Min out_features dim
+        self.min_n_threads = 64
+
+        # Min in_features dim
+        self.min_k_threads = 128
+
+        # Max parallel problems to solve at once (improves large
+        # batch performance)
+        self.max_parallel = 16
+
+        # Permutation length used by the marlin kernels.
+        self.perm_len = 1024
+
+    def __repr__(self) -> str:
+        return (
+            f"MarlinConfig(group_size={self.group_size}, "
+            f"lm_head_quantized={self.lm_head_quantized})"
+        )
+
+    @classmethod
+    def get_name(cls) -> str:
+        return "marlin"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.half]
+
+    @classmethod
+    # Need to figure it out
+    def get_min_capability(cls) -> int:
+        return 80
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return ["quantize_config.json"]
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "MarlinConfig":
+        group_size = cls.get_from_keys(config, ["group_size"])
+        lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False)
+        return cls(group_size, lm_head_quantized)
+
+    @classmethod
+    def override_quantization_method(cls, hf_quant_cfg, user_quant) -> Optional[str]:
+        # compat: autogptq >=0.8.0 use checkpoint_format: str
+        # compat: autogptq <=0.7.1 is_marlin_format: bool
+        is_marlin_format = hf_quant_cfg.get(
+            "checkpoint_format"
+        ) == "marlin" or hf_quant_cfg.get("is_marlin_format", False)
+
+        is_valid_user_quant = (
+            user_quant is None or user_quant == "gptq" or user_quant == "marlin"
+        )
+
+        if is_marlin_format and is_valid_user_quant:
+            msg = "The model is serialized in {} format. Using {} kernel.".format(
+                cls.get_name(), cls.get_name()
+            )
+            logger.info(msg)
+            return cls.get_name()
+
+        return None
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional["MarlinLinearMethod"]:
+        from vllm.model_executor.layers.quantization.marlin import MarlinLinearMethod
+
+        if isinstance(layer, LinearBase) or (
+            isinstance(layer, ParallelLMHead) and self.lm_head_quantized
+        ):
+            return MarlinLinearMethod(self)
+        return None