vllm-npu 0.4.2__py3-none-any.whl

vllm/model_executor/layers/quantization/__init__.py

@@ -0,0 +1,35 @@
+from typing import Dict, Type
+
+from vllm.model_executor.layers.quantization.aqlm import AQLMConfig
+from vllm.model_executor.layers.quantization.awq import AWQConfig
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
+from vllm.model_executor.layers.quantization.fp8 import Fp8Config
+from vllm.model_executor.layers.quantization.gptq import GPTQConfig
+from vllm.model_executor.layers.quantization.gptq_marlin import (
+    GPTQMarlinConfig)
+from vllm.model_executor.layers.quantization.marlin import MarlinConfig
+from vllm.model_executor.layers.quantization.squeezellm import SqueezeLLMConfig
+
+QUANTIZATION_METHODS: Dict[str, Type[QuantizationConfig]] = {
+    "aqlm": AQLMConfig,
+    "awq": AWQConfig,
+    "fp8": Fp8Config,
+    "gptq": GPTQConfig,
+    "squeezellm": SqueezeLLMConfig,
+    "gptq_marlin": GPTQMarlinConfig,
+    "marlin": MarlinConfig,
+}
+
+
+def get_quantization_config(quantization: str) -> Type[QuantizationConfig]:
+    if quantization not in QUANTIZATION_METHODS:
+        raise ValueError(f"Invalid quantization method: {quantization}")
+    return QUANTIZATION_METHODS[quantization]
+
+
+__all__ = [
+    "QuantizationConfig",
+    "get_quantization_config",
+    "QUANTIZATION_METHODS",
+]

vllm/model_executor/layers/quantization/aqlm.py

@@ -0,0 +1,376 @@
+# Supports AQLM compression, see https://github.com/Vahe1994/AQLM
+# and https://arxiv.org/pdf/2401.06118.pdf
+
+import math
+from typing import Any, Dict, List, Optional
+
+import torch
+import torch.nn.functional as F
+from torch.nn.parameter import Parameter
+
+from vllm import _custom_ops as ops
+from vllm.model_executor.layers.linear import LinearBase, LinearMethodBase
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
+from vllm.model_executor.utils import set_weight_attrs
+
+
+def get_int_dtype(nbits: int) -> torch.dtype:
+    if nbits <= 8:
+        return torch.int8
+    if nbits <= 16:
+        return torch.int16
+    if nbits <= 32:
+        return torch.int32
+    if nbits <= 64:
+        return torch.int64
+    raise ValueError(f"No dtype available for {nbits}-bit codebooks")
+
+
+@torch.inference_mode()
+def unpack_int_data(data: torch.IntTensor, nbits: int) -> torch.IntTensor:
+    return data.to(torch.int64) % (2**nbits)
+
+
+def dequantize_weight(codes: torch.Tensor,
+                      codebooks: torch.Tensor,
+                      scales: Optional[torch.Tensor] = None) -> torch.Tensor:
+    """
+    Decode float weights from quantization codes. Differentiable.
+    :param codes: tensor of integer quantization codes, shape 
+        [*dims, num_out_groups, num_in_groups, num_codebooks]
+    :param codebooks: tensor of vectors for each quantization code, 
+        [num_codebooks, codebook_size, out_group_size, in_group_size]
+    :param scales: weight will be multiplied by this factor, must be 
+        broadcastble with 
+        [*dims, out_groups, num_in_groups, out_group_size, in_group_size]
+    :return: reconstructed weight tensor of shape 
+        [*dims, num_in_groups*group_size]
+    """
+    num_out_groups, num_in_groups, num_codebooks = codes.shape[-3:]
+    num_codebooks, codebook_size, out_group_size, in_group_size = \
+        codebooks.shape
+    out_features = num_out_groups * out_group_size
+    in_features = num_in_groups * in_group_size
+    codebook_offsets = torch.arange(
+        0, num_codebooks * codebook_size, codebook_size,
+        device=codes.device)  # shape: [num_codebooks]
+    reconstructed_weight_flat = F.embedding_bag(
+        codes.flatten(0, -2) + codebook_offsets,
+        codebooks.flatten(0, 1).flatten(-2, -1),
+        mode="sum"
+    )  # [prod(dims) * num_out_groups * num_in_groups, out_group_size
+    # * in_group_size]
+
+    reconstructed_weight_groupwise = reconstructed_weight_flat.view(
+        list(codes.shape[:-3]) +
+        [num_out_groups, num_in_groups, out_group_size, in_group_size])
+    if scales is not None:
+        reconstructed_weight_groupwise = reconstructed_weight_groupwise.mul(
+            scales)
+    return reconstructed_weight_groupwise.swapaxes(
+        -3, -2).reshape(list(codes.shape[:-3]) + [out_features, in_features])
+
+
+def dequantize_gemm(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+    dequantized_weight = dequantize_weight(
+        unpack_int_data(codes, codebooks.shape[1].bit_length() - 1),
+        codebooks,
+        scales,
+    )
+    return F.linear(input, dequantized_weight, bias)
+
+
+# Generic dequantization, slow but flexible.
+def generic_dequantize_gemm(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+    output_shape = input.shape[:-1] + (scales.shape[0], )
+    output = torch.empty(output_shape, dtype=input.dtype, device=input.device)
+    num_outputs = len(output_partition_sizes)
+
+    # break the inputs and codebooks apart then combine the outputs.
+    # Surprisingly (to me) this is faster than doing 3 de-quants and 1 big
+    # multiply at the end.
+    num_codebooks = codebooks.shape[0] // num_outputs
+    assert (scales.shape[0] == codes.shape[0])
+    assert (sum(output_partition_sizes) == scales.shape[0])
+    output_offset = 0
+    codebooks_offset = 0
+    for output_size in output_partition_sizes:
+        shard_output = dequantize_gemm(
+            input, codes.narrow(0, output_offset, output_size),
+            codebooks.narrow(0, codebooks_offset, num_codebooks),
+            scales.narrow(0, output_offset, output_size), None
+            if bias is None else bias.narrow(0, output_offset, output_size))
+
+        output_slice = output.narrow(-1, output_offset, output_size)
+        assert (output_slice.shape == shard_output.shape)
+        output_slice.copy_(shard_output)
+        output_offset += output_size
+        codebooks_offset += num_codebooks
+    return output
+
+
+# Optimized dequnantize/decompression kernels, supports 1x16 and 2x8
+# at 6 and 9 times faster than the generic version above, respectively.
+def optimized_dequantize_gemm(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
+
+    if bias is None:
+        # scaling the output is fastest, so we do that when possible.
+        output = F.linear(input, weights, bias)
+        orig_shape = output.shape
+        flattened_output = output.view(-1, output.size(-1))
+        f_scales = scales.view(-1, scales.shape[0])
+        b_scales = f_scales.expand(flattened_output.shape[0], -1)
+        flattened_output *= b_scales
+        return output.view(orig_shape)
+    else:
+        b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
+            -1, weights.shape[1])
+        weights *= b_scales
+        return F.linear(input, weights, bias)
+
+
+class AQLMConfig(QuantizationConfig):
+    """Config class for AQLM.
+
+    Reference: https://github.com/Vahe1994/AQLM
+    """
+
+    def __init__(
+        self,
+        in_group_size: int,
+        nbits_per_codebook: int,
+        num_codebooks: int,
+        out_group_size: int,
+    ) -> None:
+        self.in_group_size = in_group_size
+        self.nbits_per_codebook = nbits_per_codebook
+        self.num_codebooks = num_codebooks
+        self.out_group_size = out_group_size
+
+        # out_group_size > 1 is untested, and probably won't work as-is.
+        assert (self.out_group_size == 1)
+        self.pack_factor = (self.in_group_size * self.out_group_size)
+
+    def __repr__(self) -> str:
+        return (f"AQLMConfig(in_group_size={self.in_group_size}, "
+                f"nbits_per_codebook={self.nbits_per_codebook}, "
+                f"num_codebooks={self.num_codebooks}, "
+                f"out_group_size={self.out_group_size})")
+
+    @classmethod
+    def get_name(cls) -> str:
+        return "aqlm"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.half]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 70
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return []  # no extra configs.
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "AQLMConfig":
+        in_group_size = cls.get_from_keys(config, ["in_group_size"])
+        nbits_per_codebook = cls.get_from_keys(config, ["nbits_per_codebook"])
+        num_code_books = cls.get_from_keys(config, ["num_codebooks"])
+        out_group_size = cls.get_from_keys(config, ["out_group_size"])
+        return cls(in_group_size, nbits_per_codebook, num_code_books,
+                   out_group_size)
+
+    def get_quant_method(
+            self, layer: torch.nn.Module) -> Optional["AQLMLinearMethod"]:
+        if isinstance(layer, LinearBase):
+            return AQLMLinearMethod(self)
+        return None
+
+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+
+class AQLMLinearMethod(LinearMethodBase):
+    """Linear method for AQLM.
+
+    Args:
+        quant_config: The AQLM quantization config.
+    """
+
+    def __init__(self, quant_config: AQLMConfig):
+        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):
+        del output_size  # Unused.
+        del input_size  # Unused.
+
+        if params_dtype != torch.half:
+            raise ValueError("Only half is currently supported by aqlm")
+        if input_size_per_partition % self.quant_config.in_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.out_group_size != 0:
+            raise ValueError(
+                "The output size is not aligned with the quantized "
+                "weight shape. This can be caused by too large "
+                "tensor parallel size.")
+
+        codes = Parameter(
+            torch.empty(
+                # There could actually be two pack factors, one along input and
+                # one along output, but we don't currently support
+                # out_group_size, and only the one along output needs to be
+                # marked with "packed_dim" in order for QKVLinear to work.
+                output_size_per_partition,
+                input_size_per_partition // self.quant_config.pack_factor,
+                self.quant_config.num_codebooks,
+                dtype=get_int_dtype(self.quant_config.nbits_per_codebook),
+            ),
+            requires_grad=False,
+        )
+
+        set_weight_attrs(
+            codes,
+            {
+                "input_dim": 1,
+                "output_dim": 0,
+                "packed_dim": 1,
+                "pack_factor": self.quant_config.pack_factor,
+            },
+        )
+
+        codebooks = Parameter(
+            torch.empty(
+                self.quant_config.num_codebooks * len(output_partition_sizes),
+                2**self.quant_config.nbits_per_codebook,
+                self.quant_config.out_group_size,
+                self.quant_config.in_group_size,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        set_weight_attrs(
+            codebooks,
+            {
+                # metadata indicates fixed size concatenated along dim 0
+                "is_metadata":
+                True,
+                "output_partition_sizes":
+                torch.tensor(output_partition_sizes, device='cpu'),
+            },
+        )
+
+        scales = Parameter(
+            torch.empty(
+                (
+                    output_size_per_partition //
+                    self.quant_config.out_group_size,
+                    1,
+                    1,
+                    1,
+                ),
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        set_weight_attrs(
+            scales,
+            {
+                "output_dim": 0,
+                "packed_dim": 0,
+                "pack_factor": self.quant_config.out_group_size
+            },
+        )
+
+        layer.register_parameter("codes", codes)
+        set_weight_attrs(codes, extra_weight_attrs)
+        layer.register_parameter("codebooks", codebooks)
+        set_weight_attrs(codebooks, extra_weight_attrs)
+        layer.register_parameter("scales", scales)
+        set_weight_attrs(scales, extra_weight_attrs)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        codebooks = layer.codebooks
+        codes = layer.codes
+        scales = layer.scales
+        output_partition_sizes = getattr(codebooks, "output_partition_sizes",
+                                         None)
+
+        nbooks = codes.shape[2]
+        ingroups = codebooks.shape[3]
+        outgroups = codebooks.shape[2]
+        bits = codebooks.shape[1]
+
+        # We support these formats with dedicated gemm and decompression
+        # kernels.
+        if ingroups == 8 and outgroups == 1 and (
+            (bits == 256 and nbooks == 2) or (bits == 65536 and nbooks == 1)):
+
+            # thresholds determined by timings on an A6000, one GPU
+            use_gemv = math.prod(x.shape[:-1]) <= 6
+
+            return ops.aqlm_gemm(
+                x,
+                codes,
+                codebooks,
+                scales,
+                output_partition_sizes,
+                bias,
+            ) if use_gemv else optimized_dequantize_gemm(
+                x,
+                codes,
+                codebooks,
+                scales,
+                output_partition_sizes,
+                bias,
+            )
+
+        # fall back all unoptimized formats
+        return generic_dequantize_gemm(
+            x,
+            codes,
+            codebooks,
+            scales,
+            output_partition_sizes,
+            bias,
+        )

vllm/model_executor/layers/quantization/awq.py

@@ -0,0 +1,175 @@
+from typing import Any, Dict, List, Optional
+
+import torch
+from torch.nn.parameter import Parameter
+
+from vllm import _custom_ops as ops
+from vllm.model_executor.layers.linear import LinearBase, LinearMethodBase
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
+from vllm.model_executor.utils import set_weight_attrs
+
+
+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,
+    ) -> None:
+        self.weight_bits = weight_bits
+        self.group_size = group_size
+        self.zero_point = zero_point
+
+        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})")
+
+    def get_name(self) -> str:
+        return "awq"
+
+    def get_supported_act_dtypes(self) -> List[torch.dtype]:
+        return [torch.half]
+
+    def get_min_capability(self) -> 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"])
+        return cls(weight_bits, group_size, zero_point)
+
+    def get_quant_method(
+            self, layer: torch.nn.Module) -> Optional["AWQLinearMethod"]:
+        if isinstance(layer, LinearBase):
+            return AWQLinearMethod(self)
+        return None
+
+    def get_scaled_act_names(self) -> List[str]:
+        return ["gelu", "gelu_fast", "gelu_new", "gelu_pytorch_tanh"]
+
+
+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.")
+
+        qweight = Parameter(
+            torch.empty(
+                input_size_per_partition,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        set_weight_attrs(
+            qweight, {
+                "input_dim": 0,
+                "output_dim": 1,
+                "packed_dim": 1,
+                "pack_factor": self.quant_config.pack_factor,
+            })
+        qzeros = Parameter(
+            torch.empty(
+                input_size_per_partition // self.quant_config.group_size,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        set_weight_attrs(
+            qzeros, {
+                "input_dim": 0,
+                "output_dim": 1,
+                "packed_dim": 1,
+                "pack_factor": self.quant_config.pack_factor,
+            })
+        scales = Parameter(
+            torch.empty(
+                input_size_per_partition // self.quant_config.group_size,
+                output_size_per_partition,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        set_weight_attrs(scales, {
+            "input_dim": 0,
+            "output_dim": 1,
+        })
+
+        layer.register_parameter("qweight", qweight)
+        set_weight_attrs(qweight, extra_weight_attrs)
+        layer.register_parameter("qzeros", qzeros)
+        set_weight_attrs(qzeros, extra_weight_attrs)
+        layer.register_parameter("scales", scales)
+        set_weight_attrs(scales, extra_weight_attrs)
+
+    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])
+
+        # num_tokens >= threshold
+        FP16_MATMUL_HEURISTIC_CONDITION = x.shape[:-1].numel() >= 256
+
+        if FP16_MATMUL_HEURISTIC_CONDITION:
+            out = ops.awq_dequantize(qweight, scales, qzeros, 0, 0, 0)
+            out = torch.matmul(reshaped_x, out)
+        else:
+            out = ops.awq_gemm(reshaped_x, qweight, scales, qzeros,
+                               pack_factor)
+        if bias is not None:
+            out.add_(bias)
+        return out.reshape(out_shape)

vllm/model_executor/layers/quantization/base_config.py

@@ -0,0 +1,97 @@
+from abc import ABC, abstractmethod
+from typing import Any, Dict, List, Optional
+
+import torch
+from torch import nn
+
+
+class QuantizeMethodBase(ABC):
+    """Base class for different quantized methods."""
+
+    @abstractmethod
+    def create_weights(self, layer: torch.nn.Module, *weight_args,
+                       **extra_weight_attrs):
+        """Create weights for a layer.
+
+        The weights will be set as attributes of the layer."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def apply(self, layer: torch.nn.Module, *args, **kwargs) -> torch.Tensor:
+        """Apply the weights in layer to the input tensor.
+
+        Expects create_weights to have been called before on the layer."""
+        raise NotImplementedError
+
+    def process_weights_after_loading(self, layer: nn.Module) -> None:
+        """Process the weight after loading.
+
+        This can be used for example, to transpose weights for computation.
+        """
+        return
+
+
+class QuantizationConfig(ABC):
+    """Base class for quantization configs."""
+
+    @abstractmethod
+    def get_name(self) -> str:
+        """Name of the quantization method."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def get_supported_act_dtypes(self) -> List[torch.dtype]:
+        """List of supported activation dtypes."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def get_min_capability(self) -> int:
+        """Minimum GPU capability to support the quantization method.
+
+        E.g., 70 for Volta, 75 for Turing, 80 for Ampere.
+        This requirement is due to the custom CUDA kernels used by the
+        quantization method.
+        """
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_config_filenames() -> List[str]:
+        """List of filenames to search for in the model directory."""
+        raise NotImplementedError
+
+    @classmethod
+    @abstractmethod
+    def from_config(cls, config: Dict[str, Any]) -> "QuantizationConfig":
+        """Create a config class from the model's quantization config."""
+        raise NotImplementedError
+
+    @staticmethod
+    def get_from_keys(config: Dict[str, Any], keys: List[str]) -> Any:
+        """Get a value from the model's quantization config."""
+        for key in keys:
+            if key in config:
+                return config[key]
+        raise ValueError(f"Cannot find any of {keys} in the model's "
+                         "quantization config.")
+
+    @abstractmethod
+    def get_quant_method(
+            self, layer: torch.nn.Module) -> Optional[QuantizeMethodBase]:
+        """Get the quantize method to use for the quantized layer.
+        
+        Args:
+            layer: The layer for the quant method.
+        Returns:
+            The quantize method. None if the given layer doesn't support quant
+            method.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    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