vllm-cpu 0.8.5.post2__cp310-cp310-manylinux_2_17_x86_64.whl

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a16_fp8.py

@@ -0,0 +1,119 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Callable, List, Optional
+
+import torch
+from compressed_tensors.quantization import QuantizationStrategy
+
+from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsScheme)
+from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
+    apply_fp8_marlin_linear, prepare_fp8_layer_for_marlin)
+from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
+    convert_to_channelwise)
+from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
+                                           ModelWeightParameter,
+                                           PerTensorScaleParameter)
+
+__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
+
+    @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)

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py

@@ -0,0 +1,149 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Callable, List, Optional
+
+import torch
+from compressed_tensors.quantization import QuantizationStrategy
+from torch.nn import Parameter
+
+from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsScheme)
+from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
+    Fp8LinearOp, maybe_create_device_identity, normalize_e4m3fn_to_e4m3fnuz,
+    requantize_with_max_scale)
+from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
+                                           ModelWeightParameter,
+                                           PerTensorScaleParameter)
+from vllm.platforms import current_platform
+
+__all__ = ["CompressedTensorsW8A8Fp8"]
+
+
+class CompressedTensorsW8A8Fp8(CompressedTensorsScheme):
+
+    def __init__(self, strategy: str, is_static_input_scheme: bool):
+        self.strategy = strategy
+        self.out_dtype = torch.get_default_dtype()
+        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:
+        # lovelace and up
+        return 89
+
+    def process_weights_after_loading(self, layer) -> None:
+        # If per tensor, when we have a fused module (e.g. QKV) with per
+        # tensor scales (thus N scales being passed to the kernel),
+        # requantize so we can always run per tensor
+        if self.strategy == QuantizationStrategy.TENSOR:
+            max_w_scale, weight = requantize_with_max_scale(
+                weight=layer.weight,
+                weight_scale=layer.weight_scale,
+                logical_widths=layer.logical_widths,
+            )
+
+            if current_platform.is_fp8_fnuz():
+                input_scale = getattr(layer, 'input_scale', None)
+
+                weight, max_w_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
+                    weight=weight,
+                    weight_scale=max_w_scale,
+                    input_scale=input_scale)
+                if input_scale is not None:
+                    layer.input_scale = Parameter(input_scale,
+                                                  requires_grad=False)
+
+            layer.weight = Parameter(weight.t(), requires_grad=False)
+            layer.weight_scale = Parameter(max_w_scale, requires_grad=False)
+
+        # If channelwise, scales are already lined up, so just transpose.
+        elif self.strategy == QuantizationStrategy.CHANNEL:
+            weight = layer.weight
+
+            if current_platform.is_fp8_fnuz():
+                input_scale = getattr(layer, 'input_scale', None)
+
+                weight, weight_scale, input_scale = \
+                    normalize_e4m3fn_to_e4m3fnuz(
+                        weight=weight,
+                        weight_scale=layer.weight_scale,
+                        input_scale=input_scale)
+                if input_scale is not None:
+                    layer.input_scale = Parameter(input_scale,
+                                                  requires_grad=False)
+            else:
+                weight_scale = layer.weight_scale.data
+
+            layer.weight = Parameter(weight.t(), requires_grad=False)
+            # required by torch.compile to be torch.nn.Parameter
+            layer.weight_scale = Parameter(weight_scale, requires_grad=False)
+
+        else:
+            raise ValueError(f"Unknown quantization strategy {self.strategy}")
+
+        # INPUT SCALE
+        if self.is_static_input_scheme and hasattr(layer, 'input_scale'):
+            layer.input_scale = Parameter(layer.input_scale.max(),
+                                          requires_grad=False)
+        else:
+            layer.input_scale = None
+
+    def create_weights(self, layer: torch.nn.Module,
+                       output_partition_sizes: List[int],
+                       input_size_per_partition: int,
+                       params_dtype: torch.dtype, weight_loader: Callable,
+                       **kwargs):
+        maybe_create_device_identity()
+
+        output_size_per_partition = sum(output_partition_sizes)
+        layer.logical_widths = output_partition_sizes
+
+        # 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
+        # TODO: update create_xxx_parameter functions to return
+        # the newly added parameters
+        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)
+        else:
+            assert self.strategy == QuantizationStrategy.TENSOR
+            weight_scale = PerTensorScaleParameter(data=torch.empty(
+                len(output_partition_sizes), dtype=torch.float32),
+                                                   weight_loader=weight_loader)
+
+        # min requirement for fp8 kernels
+        weight_scale[:] = torch.finfo(torch.float32).min
+        layer.register_parameter("weight_scale", weight_scale)
+
+        # INPUT SCALE
+        if self.is_static_input_scheme:
+            input_scale = PerTensorScaleParameter(data=torch.empty(
+                len(output_partition_sizes), dtype=torch.float32),
+                                                  weight_loader=weight_loader)
+            input_scale[:] = torch.finfo(torch.float32).min
+            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 self.fp8_linear.apply(input=x,
+                                     weight=layer.weight,
+                                     weight_scale=layer.weight_scale,
+                                     out_dtype=self.out_dtype,
+                                     input_scale=layer.input_scale,
+                                     bias=bias)

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_int8.py

@@ -0,0 +1,110 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Callable, List, Optional, Set
+
+import torch
+from compressed_tensors.quantization import QuantizationStrategy
+
+from vllm.logger import init_logger
+from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsScheme)
+from vllm.model_executor.layers.quantization.kernels.scaled_mm import (
+    ScaledMMLinearLayerConfig, choose_scaled_mm_linear_kernel)
+from vllm.model_executor.parameter import (BasevLLMParameter,
+                                           ChannelQuantScaleParameter,
+                                           ModelWeightParameter,
+                                           PerTensorScaleParameter)
+
+logger = init_logger(__name__)
+
+
+class CompressedTensorsW8A8Int8(CompressedTensorsScheme):
+    _kernel_backends_being_used: Set[str] = set()
+
+    def __init__(self, strategy: str, is_static_input_scheme: bool,
+                 input_symmetric: bool):
+        self.strategy = strategy
+        self.is_static_input_scheme = is_static_input_scheme
+        self.input_symmetric = input_symmetric
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # turing and up
+        return 75
+
+    def create_weights(self, layer: torch.nn.Module,
+                       output_partition_sizes: List[int],
+                       input_size_per_partition: int,
+                       params_dtype: torch.dtype, weight_loader: Callable,
+                       **kwargs):
+        layer.logical_widths = output_partition_sizes
+
+        scaled_mm_linear_kernel_config = ScaledMMLinearLayerConfig(
+            is_channelwise=(self.strategy == QuantizationStrategy.CHANNEL),
+            is_static_input_scheme=self.is_static_input_scheme,
+            input_symmetric=self.input_symmetric)
+
+        kernel_type = choose_scaled_mm_linear_kernel(
+            scaled_mm_linear_kernel_config)
+
+        if kernel_type.__name__ not in self._kernel_backends_being_used:
+            logger.info("Using %s for CompressedTensorsW8A8Int8",
+                        kernel_type.__name__)
+            self._kernel_backends_being_used.add(kernel_type.__name__)
+
+        # WEIGHT
+        weight = ModelWeightParameter(data=torch.empty(
+            sum(output_partition_sizes),
+            input_size_per_partition,
+            dtype=torch.int8),
+                                      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)
+        else:
+            assert self.strategy == QuantizationStrategy.TENSOR
+            weight_scale = PerTensorScaleParameter(data=torch.empty(
+                len(output_partition_sizes), dtype=torch.float32),
+                                                   weight_loader=weight_loader)
+        layer.register_parameter("weight_scale", weight_scale)
+
+        # INPUT SCALE
+        if self.is_static_input_scheme:
+            input_scale = BasevLLMParameter(data=torch.empty(
+                1, dtype=torch.float32),
+                                            weight_loader=weight_loader)
+            layer.register_parameter("input_scale", input_scale)
+
+            if not self.input_symmetric:
+                # Note: compressed-tensors stores the zp using the same dtype
+                # as the weights
+                # AZP loaded as int8 but used as int32
+                input_zero_point = BasevLLMParameter(
+                    data=torch.empty(1, dtype=torch.int8),
+                    weight_loader=weight_loader)
+                layer.register_parameter("input_zero_point", input_zero_point)
+
+        self.kernel = kernel_type(c=scaled_mm_linear_kernel_config,
+                                  w_q_param_name="weight",
+                                  w_s_param_name="weight_scale",
+                                  i_s_param_name="input_scale",
+                                  i_zp_param_name="input_zero_point",
+                                  azp_adj_param_name="azp_adj")
+
+    # Checkpoints are serialized in compressed-tensors format, which is
+    # different from the format the kernel may want. Handle repacking here.
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        self.kernel.process_weights_after_loading(layer)
+
+    def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor,
+                      bias: Optional[torch.Tensor]) -> torch.Tensor:
+        return self.kernel.apply_weights(layer, x, bias)

vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py

@@ -0,0 +1,200 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Callable, List, Optional, Set
+
+import torch
+from compressed_tensors.quantization import ActivationOrdering
+
+from vllm.logger import init_logger
+from vllm.model_executor.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsScheme)
+from vllm.model_executor.layers.quantization.kernels.mixed_precision import (
+    MPLinearLayerConfig, choose_mp_linear_kernel)
+from vllm.model_executor.layers.quantization.utils.marlin_utils import (
+    marlin_repeat_scales_on_all_ranks)
+# yapf conflicts with isort for this block
+# yapf: disable
+from vllm.model_executor.parameter import (BasevLLMParameter,
+                                           ChannelQuantScaleParameter,
+                                           GroupQuantScaleParameter,
+                                           PackedColumnParameter,
+                                           PackedvLLMParameter,
+                                           RowvLLMParameter)
+# yapf: enable
+from vllm.scalar_type import scalar_types
+
+logger = init_logger(__name__)
+
+__all__ = ["CompressedTensorsWNA16"]
+WNA16_SUPPORTED_TYPES_MAP = {
+    4: scalar_types.uint4b8,
+    8: scalar_types.uint8b128
+}
+WNA16_ZP_SUPPORTED_TYPES_MAP = {4: scalar_types.uint4, 8: scalar_types.uint8}
+WNA16_SUPPORTED_BITS = list(WNA16_SUPPORTED_TYPES_MAP.keys())
+
+
+class CompressedTensorsWNA16(CompressedTensorsScheme):
+    _kernel_backends_being_used: Set[str] = set()
+
+    def __init__(self,
+                 strategy: str,
+                 num_bits: int,
+                 group_size: Optional[int] = None,
+                 symmetric: Optional[bool] = True,
+                 actorder: Optional[ActivationOrdering] = None):
+
+        self.pack_factor = 32 // num_bits
+        self.strategy = strategy
+        self.symmetric = symmetric
+        self.group_size = -1 if group_size is None else group_size
+        self.has_g_idx = actorder == ActivationOrdering.GROUP
+
+        if self.group_size == -1 and self.strategy != "channel":
+            raise ValueError("Marlin kernels require group quantization or "
+                             "channelwise quantization, but found no group "
+                             "size and strategy is not channelwise.")
+
+        if num_bits not in WNA16_SUPPORTED_TYPES_MAP:
+            raise ValueError(
+                f"Unsupported num_bits = {num_bits}. "
+                f"Supported num_bits = {WNA16_SUPPORTED_TYPES_MAP.keys()}")
+
+        self.quant_type = (WNA16_ZP_SUPPORTED_TYPES_MAP[num_bits]
+                           if not self.symmetric else
+                           WNA16_SUPPORTED_TYPES_MAP[num_bits])
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # ampere and up
+        return 80
+
+    def create_weights(self, layer: torch.nn.Module, output_size: int,
+                       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)
+
+        mp_linear_kernel_config = MPLinearLayerConfig(
+            full_weight_shape=(input_size, output_size),
+            partition_weight_shape=\
+                (input_size_per_partition, output_size_per_partition),
+            weight_type=self.quant_type,
+            act_type=params_dtype,
+            group_size=self.group_size,
+            zero_points=not self.symmetric,
+            has_g_idx=self.has_g_idx
+        )
+
+        kernel_type = choose_mp_linear_kernel(mp_linear_kernel_config)
+
+        if kernel_type.__name__ not in self._kernel_backends_being_used:
+            logger.info("Using %s for CompressedTensorsWNA16",
+                        kernel_type.__name__)
+            self._kernel_backends_being_used.add(kernel_type.__name__)
+
+        # If group_size is -1, we are in channelwise case.
+        group_size = self.group_size if self.group_size != -1 else input_size
+        row_parallel = (input_size != input_size_per_partition)
+        partition_scales = not marlin_repeat_scales_on_all_ranks(
+            self.has_g_idx, self.group_size, row_parallel)
+
+        scales_and_zp_size = input_size // group_size
+
+        if partition_scales:
+            assert input_size_per_partition % group_size == 0
+            scales_and_zp_size = input_size_per_partition // group_size
+
+        weight = PackedvLLMParameter(input_dim=1,
+                                     output_dim=0,
+                                     weight_loader=weight_loader,
+                                     packed_factor=self.pack_factor,
+                                     packed_dim=1,
+                                     data=torch.empty(
+                                         output_size_per_partition,
+                                         input_size_per_partition //
+                                         self.pack_factor,
+                                         dtype=torch.int32,
+                                     ))
+
+        weight_scale_args = {
+            "weight_loader":
+            weight_loader,
+            "data":
+            torch.empty(
+                output_size_per_partition,
+                scales_and_zp_size,
+                dtype=params_dtype,
+            )
+        }
+
+        zeros_args = {
+            "weight_loader":
+            weight_loader,
+            "data":
+            torch.zeros(
+                output_size_per_partition // self.pack_factor,
+                scales_and_zp_size,
+                dtype=torch.int32,
+            )
+        }
+
+        if not partition_scales:
+            weight_scale = ChannelQuantScaleParameter(output_dim=0,
+                                                      **weight_scale_args)
+
+            if not self.symmetric:
+                qzeros = PackedColumnParameter(output_dim=0,
+                                               packed_dim=0,
+                                               packed_factor=self.pack_factor,
+                                               **zeros_args)
+        else:
+            weight_scale = GroupQuantScaleParameter(output_dim=0,
+                                                    input_dim=1,
+                                                    **weight_scale_args)
+            if not self.symmetric:
+                qzeros = PackedvLLMParameter(input_dim=1,
+                                             output_dim=0,
+                                             packed_dim=0,
+                                             packed_factor=self.pack_factor,
+                                             **zeros_args)
+
+        # A 2D array defining the original shape of the weights
+        # before packing
+        weight_shape = BasevLLMParameter(data=torch.empty(2,
+                                                          dtype=torch.int64),
+                                         weight_loader=weight_loader)
+
+        layer.register_parameter("weight_packed", weight)
+        layer.register_parameter("weight_scale", weight_scale)
+        layer.register_parameter("weight_shape", weight_shape)
+
+        if not self.symmetric:
+            layer.register_parameter("weight_zero_point", qzeros)
+
+        # group index (for activation reordering)
+        if self.has_g_idx:
+            weight_g_idx = RowvLLMParameter(data=torch.empty(
+                input_size_per_partition,
+                dtype=torch.int32,
+            ),
+                                            input_dim=0,
+                                            weight_loader=weight_loader)
+            layer.register_parameter("weight_g_idx", weight_g_idx)
+
+        self.kernel = kernel_type(mp_linear_kernel_config,
+                                  w_q_param_name="weight_packed",
+                                  w_s_param_name="weight_scale",
+                                  w_zp_param_name="weight_zero_point",
+                                  w_gidx_param_name="weight_g_idx")
+
+    # Checkpoints are serialized in compressed-tensors format, which is
+    # different from the format the kernel may want. Handle repacking here.
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        self.kernel.process_weights_after_loading(layer)
+
+    def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor,
+                      bias: Optional[torch.Tensor]) -> torch.Tensor:
+        return self.kernel.apply_weights(layer, x, bias)