Convert formatting to use ruff instead of yapf + isort (#26247)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -11,7 +11,7 @@ __all__ = ["QuarkScheme"]
|
||||
|
||||
class QuarkScheme(ABC):
|
||||
"""
|
||||
Abstract class used to describe the weight creation and forward pass
|
||||
Abstract class used to describe the weight creation and forward pass
|
||||
of different quantization schemes supported by Quark.
|
||||
"""
|
||||
|
||||
@@ -26,20 +26,21 @@ class QuarkScheme(ABC):
|
||||
@abstractmethod
|
||||
def create_weights(self, *args, **kwargs):
|
||||
"""
|
||||
Weight creation for the particular scheme. Inputs to this function
|
||||
Weight creation for the particular scheme. Inputs to this function
|
||||
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor]):
|
||||
def apply_weights(
|
||||
self, layer: torch.nn.Module, x: torch.Tensor, bias: Optional[torch.Tensor]
|
||||
):
|
||||
"""
|
||||
Run the forward pass for the particular scheme. This is where
|
||||
Run the forward pass for the particular scheme. This is where
|
||||
scheme-specific dequant/quant steps/kernels should be applied.
|
||||
|
||||
:param layer: torch.nn.Module with the registered weights and
|
||||
other parameters relevant to the particular scheme.
|
||||
:param layer: torch.nn.Module with the registered weights and
|
||||
other parameters relevant to the particular scheme.
|
||||
:param x: input to the layer
|
||||
:param bias: bias parameter
|
||||
|
||||
|
||||
@@ -10,17 +10,21 @@ import torch.nn.functional as F
|
||||
from vllm import envs
|
||||
from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
|
||||
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
|
||||
OCP_MX_BLOCK_SIZE, dequant_mxfp4, quant_dequant_mxfp4)
|
||||
from vllm.model_executor.parameter import (GroupQuantScaleParameter,
|
||||
PackedvLLMParameter)
|
||||
OCP_MX_BLOCK_SIZE,
|
||||
dequant_mxfp4,
|
||||
quant_dequant_mxfp4,
|
||||
)
|
||||
from vllm.model_executor.parameter import GroupQuantScaleParameter, PackedvLLMParameter
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
|
||||
@cache
|
||||
def is_rocm_aiter_fp4_asm_gemm_enabled() -> bool:
|
||||
return current_platform.is_rocm() \
|
||||
and envs.VLLM_ROCM_USE_AITER_FP4_ASM_GEMM \
|
||||
return (
|
||||
current_platform.is_rocm()
|
||||
and envs.VLLM_ROCM_USE_AITER_FP4_ASM_GEMM
|
||||
and envs.VLLM_ROCM_USE_AITER
|
||||
)
|
||||
|
||||
|
||||
try:
|
||||
@@ -29,6 +33,7 @@ try:
|
||||
from aiter.ops.triton.quant import dynamic_mxfp4_quant
|
||||
|
||||
from vllm.utils import direct_register_custom_op
|
||||
|
||||
if is_rocm_aiter_fp4_asm_gemm_enabled():
|
||||
from aiter import gemm_a4w4, per_1x32_f4_quant_hip
|
||||
|
||||
@@ -51,17 +56,13 @@ try:
|
||||
|
||||
# 32 alignment is enough for dim0 padding of output for
|
||||
# gemm_a4w4 kernel
|
||||
y = torch.empty((M + 31) // 32 * 32,
|
||||
weight.shape[0],
|
||||
device=x_q.device,
|
||||
dtype=out_dtype)
|
||||
y = torch.empty(
|
||||
(M + 31) // 32 * 32, weight.shape[0], device=x_q.device, dtype=out_dtype
|
||||
)
|
||||
|
||||
gemm_a4w4(x_q,
|
||||
weight,
|
||||
x_s,
|
||||
weight_scale.view(x_s.dtype),
|
||||
y,
|
||||
bpreshuffle=True)
|
||||
gemm_a4w4(
|
||||
x_q, weight, x_s, weight_scale.view(x_s.dtype), y, bpreshuffle=True
|
||||
)
|
||||
return y[:M]
|
||||
else:
|
||||
if x_scales is None:
|
||||
@@ -69,10 +70,9 @@ try:
|
||||
else:
|
||||
x_q = x
|
||||
x_s = x_scales
|
||||
y = torch.empty(x_q.shape[0],
|
||||
weight.shape[0],
|
||||
device=x_q.device,
|
||||
dtype=out_dtype)
|
||||
y = torch.empty(
|
||||
x_q.shape[0], weight.shape[0], device=x_q.device, dtype=out_dtype
|
||||
)
|
||||
|
||||
gemm_afp4wfp4(x_q, weight, x_s, weight_scale.T, out_dtype, y)
|
||||
return y
|
||||
@@ -85,9 +85,9 @@ try:
|
||||
rocm_use_aiter_fp4_asm_gemm: bool = False,
|
||||
out_dtype: Optional[torch.dtype] = torch.bfloat16,
|
||||
) -> torch.Tensor:
|
||||
return torch.empty((*x.shape[:-1], weight.shape[0]),
|
||||
dtype=out_dtype,
|
||||
device=x.device)
|
||||
return torch.empty(
|
||||
(*x.shape[:-1], weight.shape[0]), dtype=out_dtype, device=x.device
|
||||
)
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="gemm_with_dynamic_quant",
|
||||
@@ -104,46 +104,45 @@ __all__ = ["QuarkW4A4MXFP4"]
|
||||
|
||||
|
||||
class QuarkW4A4MXFP4(QuarkScheme):
|
||||
|
||||
def __init__(self, weight_quant_spec: dict[str, Any],
|
||||
input_quant_spec: dict[str, Any]):
|
||||
def __init__(
|
||||
self, weight_quant_spec: dict[str, Any], input_quant_spec: dict[str, Any]
|
||||
):
|
||||
self.out_dtype = torch.get_default_dtype()
|
||||
self.qscheme = "per_group"
|
||||
self.weight_quant_spec = weight_quant_spec
|
||||
self.input_quant_spec = input_quant_spec
|
||||
self.emulate = not current_platform.supports_mx()
|
||||
self.rocm_use_aiter_fp4_asm_gemm = is_rocm_aiter_fp4_asm_gemm_enabled()
|
||||
if not self.emulate and (dynamic_mxfp4_quant is None
|
||||
or gemm_afp4wfp4 is None):
|
||||
if not self.emulate and (dynamic_mxfp4_quant is None or gemm_afp4wfp4 is None):
|
||||
# Currently need these kernels if not emulating
|
||||
raise NotImplementedError(
|
||||
f"{self.__class__.__name__} requires AITER to be installed "
|
||||
"for non-emulation mode! Please refer to "
|
||||
"https://github.com/ROCm/aiter for installation details.")
|
||||
"https://github.com/ROCm/aiter for installation details."
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def get_min_capability(cls) -> int:
|
||||
return 70
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
layer.weight = torch.nn.Parameter(layer.weight.data,
|
||||
requires_grad=False)
|
||||
layer.weight = torch.nn.Parameter(layer.weight.data, requires_grad=False)
|
||||
|
||||
if self.emulate:
|
||||
layer.weight_scale = torch.nn.Parameter(layer.weight_scale.data,
|
||||
requires_grad=False)
|
||||
layer.weight_scale = torch.nn.Parameter(
|
||||
layer.weight_scale.data, requires_grad=False
|
||||
)
|
||||
try:
|
||||
from quark.torch.export.nn.modules import realquantizer
|
||||
from quark.torch.quantization.config.config import (
|
||||
QuantizationSpec)
|
||||
from quark.torch.quantization.config.config import QuantizationSpec
|
||||
except ImportError as err:
|
||||
raise ImportError(
|
||||
"The package `amd-quark` is required to use AMD Quark "
|
||||
"MX-FP4 models. Please install it with `pip install "
|
||||
"amd-quark`.") from err
|
||||
"amd-quark`."
|
||||
) from err
|
||||
|
||||
weight_quant_spec = QuantizationSpec.from_dict(
|
||||
self.weight_quant_spec)
|
||||
weight_quant_spec = QuantizationSpec.from_dict(self.weight_quant_spec)
|
||||
|
||||
weight_quantizer = realquantizer.get_real_quantizer(
|
||||
qspec=weight_quant_spec,
|
||||
@@ -170,29 +169,34 @@ class QuarkW4A4MXFP4(QuarkScheme):
|
||||
weight_scale_shuffle = layer.weight_scale.data
|
||||
sm, sn = weight_scale_shuffle.shape
|
||||
weight_scale_shuffle = weight_scale_shuffle.view(
|
||||
sm // 32, 2, 16, sn // 8, 2, 4, 1)
|
||||
sm // 32, 2, 16, sn // 8, 2, 4, 1
|
||||
)
|
||||
weight_scale_shuffle = weight_scale_shuffle.permute(
|
||||
0, 3, 5, 2, 4, 1, 6).contiguous()
|
||||
0, 3, 5, 2, 4, 1, 6
|
||||
).contiguous()
|
||||
weight_scale_shuffle = weight_scale_shuffle.view(sm, sn)
|
||||
layer.weight_scale = torch.nn.Parameter(weight_scale_shuffle,
|
||||
requires_grad=False)
|
||||
layer.weight_scale = torch.nn.Parameter(
|
||||
weight_scale_shuffle, requires_grad=False
|
||||
)
|
||||
|
||||
# shuffle weight
|
||||
weight_shuffle = layer.weight.data
|
||||
weight_shuffle = shuffle_weight(weight_shuffle,
|
||||
layout=(16, 16))
|
||||
layer.weight = torch.nn.Parameter(weight_shuffle,
|
||||
requires_grad=False)
|
||||
weight_shuffle = shuffle_weight(weight_shuffle, layout=(16, 16))
|
||||
layer.weight = torch.nn.Parameter(weight_shuffle, requires_grad=False)
|
||||
else:
|
||||
layer.weight_scale = torch.nn.Parameter(
|
||||
layer.weight_scale.data.T.contiguous(),
|
||||
requires_grad=False)
|
||||
layer.weight_scale.data.T.contiguous(), requires_grad=False
|
||||
)
|
||||
|
||||
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):
|
||||
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,
|
||||
):
|
||||
output_size_per_partition = sum(output_partition_sizes)
|
||||
layer.logical_widths = output_partition_sizes
|
||||
|
||||
@@ -224,16 +228,21 @@ class QuarkW4A4MXFP4(QuarkScheme):
|
||||
)
|
||||
layer.register_parameter("weight_scale", weight_scale)
|
||||
|
||||
def apply_weights(self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None) -> torch.Tensor:
|
||||
|
||||
def apply_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
if self.emulate:
|
||||
dq_w = dequant_mxfp4(layer.weight, layer.weight_scale, x.dtype)
|
||||
x = quant_dequant_mxfp4(x)
|
||||
return F.linear(x, dq_w, bias)
|
||||
else:
|
||||
return torch.ops.vllm.gemm_with_dynamic_quant(
|
||||
x, layer.weight, layer.weight_scale,
|
||||
self.rocm_use_aiter_fp4_asm_gemm, self.out_dtype)
|
||||
x,
|
||||
layer.weight,
|
||||
layer.weight_scale,
|
||||
self.rocm_use_aiter_fp4_asm_gemm,
|
||||
self.out_dtype,
|
||||
)
|
||||
|
||||
@@ -7,37 +7,43 @@ import torch
|
||||
from torch.nn import Parameter
|
||||
|
||||
from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
||||
GroupShape)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
|
||||
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
|
||||
Fp8LinearOp, normalize_e4m3fn_to_e4m3fnuz, requantize_with_max_scale)
|
||||
from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
|
||||
ModelWeightParameter,
|
||||
PerTensorScaleParameter)
|
||||
Fp8LinearOp,
|
||||
normalize_e4m3fn_to_e4m3fnuz,
|
||||
requantize_with_max_scale,
|
||||
)
|
||||
from vllm.model_executor.parameter import (
|
||||
ChannelQuantScaleParameter,
|
||||
ModelWeightParameter,
|
||||
PerTensorScaleParameter,
|
||||
)
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
__all__ = ["QuarkW8A8Fp8"]
|
||||
|
||||
|
||||
class QuarkW8A8Fp8(QuarkScheme):
|
||||
|
||||
def __init__(self, weight_config: dict[str, Any],
|
||||
input_config: Optional[dict[str, Any]]):
|
||||
def __init__(
|
||||
self, weight_config: dict[str, Any], input_config: Optional[dict[str, Any]]
|
||||
):
|
||||
self.weight_qscheme = cast(str, weight_config.get("qscheme"))
|
||||
self.is_static_input_scheme: bool = False
|
||||
self.input_qscheme: Optional[str] = None
|
||||
if input_config is not None:
|
||||
self.is_static_input_scheme = not cast(
|
||||
bool, input_config.get("is_dynamic"))
|
||||
self.is_static_input_scheme = not cast(bool, input_config.get("is_dynamic"))
|
||||
self.input_qscheme = cast(str, input_config.get("qscheme"))
|
||||
|
||||
per_token = (not self.is_static_input_scheme
|
||||
and self.input_qscheme == "per_channel")
|
||||
self.act_quant_group_shape = GroupShape.PER_TOKEN \
|
||||
if per_token else GroupShape.PER_TENSOR
|
||||
per_token = (
|
||||
not self.is_static_input_scheme and self.input_qscheme == "per_channel"
|
||||
)
|
||||
self.act_quant_group_shape = (
|
||||
GroupShape.PER_TOKEN if per_token else GroupShape.PER_TENSOR
|
||||
)
|
||||
self.fp8_linear = Fp8LinearOp(
|
||||
act_quant_static=self.is_static_input_scheme,
|
||||
act_quant_group_shape=self.act_quant_group_shape)
|
||||
act_quant_group_shape=self.act_quant_group_shape,
|
||||
)
|
||||
self.out_dtype = torch.get_default_dtype()
|
||||
|
||||
@classmethod
|
||||
@@ -51,14 +57,14 @@ class QuarkW8A8Fp8(QuarkScheme):
|
||||
# requantize so we can always run per tensor
|
||||
if self.weight_qscheme == "per_tensor":
|
||||
if current_platform.is_fp8_fnuz():
|
||||
input_scale = getattr(layer, 'input_scale', None)
|
||||
input_scale = getattr(layer, "input_scale", None)
|
||||
weight, max_w_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
|
||||
weight=layer.weight,
|
||||
weight_scale=layer.weight_scale,
|
||||
input_scale=input_scale)
|
||||
input_scale=input_scale,
|
||||
)
|
||||
if input_scale is not None:
|
||||
layer.input_scale = Parameter(input_scale,
|
||||
requires_grad=False)
|
||||
layer.input_scale = Parameter(input_scale, requires_grad=False)
|
||||
else:
|
||||
max_w_scale = layer.weight_scale
|
||||
weight = layer.weight
|
||||
@@ -77,15 +83,14 @@ class QuarkW8A8Fp8(QuarkScheme):
|
||||
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)
|
||||
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)
|
||||
layer.input_scale = Parameter(input_scale, requires_grad=False)
|
||||
else:
|
||||
weight_scale = layer.weight_scale.data
|
||||
if self.act_quant_group_shape == GroupShape.PER_TOKEN:
|
||||
@@ -95,32 +100,37 @@ class QuarkW8A8Fp8(QuarkScheme):
|
||||
layer.weight_scale = Parameter(weight_scale, requires_grad=False)
|
||||
|
||||
else:
|
||||
raise ValueError(
|
||||
f"Unknown quantization scheme {self.weight_qscheme}")
|
||||
raise ValueError(f"Unknown quantization scheme {self.weight_qscheme}")
|
||||
|
||||
# INPUT SCALE
|
||||
if self.is_static_input_scheme:
|
||||
layer.input_scale = Parameter(layer.input_scale.max(),
|
||||
requires_grad=False)
|
||||
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):
|
||||
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,
|
||||
):
|
||||
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)
|
||||
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
|
||||
@@ -128,15 +138,16 @@ class QuarkW8A8Fp8(QuarkScheme):
|
||||
# the newly added parameters
|
||||
if self.weight_qscheme == "per_channel":
|
||||
weight_scale = ChannelQuantScaleParameter(
|
||||
data=torch.empty((sum(output_partition_sizes)),
|
||||
dtype=torch.float32),
|
||||
data=torch.empty((sum(output_partition_sizes)), dtype=torch.float32),
|
||||
output_dim=0,
|
||||
weight_loader=weight_loader)
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
else:
|
||||
assert self.weight_qscheme == "per_tensor"
|
||||
weight_scale = PerTensorScaleParameter(data=torch.empty(
|
||||
len(output_partition_sizes), dtype=torch.float32),
|
||||
weight_loader=weight_loader)
|
||||
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
|
||||
@@ -144,20 +155,24 @@ class QuarkW8A8Fp8(QuarkScheme):
|
||||
|
||||
# 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 = 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)
|
||||
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,
|
||||
)
|
||||
|
||||
@@ -7,12 +7,16 @@ import torch
|
||||
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.quantization.kernels.scaled_mm import (
|
||||
ScaledMMLinearLayerConfig, choose_scaled_mm_linear_kernel)
|
||||
ScaledMMLinearLayerConfig,
|
||||
choose_scaled_mm_linear_kernel,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
|
||||
from vllm.model_executor.parameter import (BasevLLMParameter,
|
||||
ChannelQuantScaleParameter,
|
||||
ModelWeightParameter,
|
||||
PerTensorScaleParameter)
|
||||
from vllm.model_executor.parameter import (
|
||||
BasevLLMParameter,
|
||||
ChannelQuantScaleParameter,
|
||||
ModelWeightParameter,
|
||||
PerTensorScaleParameter,
|
||||
)
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
@@ -20,8 +24,12 @@ logger = init_logger(__name__)
|
||||
class QuarkW8A8Int8(QuarkScheme):
|
||||
_kernel_backends_being_used: set[str] = set()
|
||||
|
||||
def __init__(self, qscheme: str, is_static_input_scheme: Optional[bool],
|
||||
input_symmetric: Optional[bool]):
|
||||
def __init__(
|
||||
self,
|
||||
qscheme: str,
|
||||
is_static_input_scheme: Optional[bool],
|
||||
input_symmetric: Optional[bool],
|
||||
):
|
||||
self.qscheme = qscheme
|
||||
self.is_static_input_scheme = is_static_input_scheme
|
||||
self.input_symmetric = input_symmetric
|
||||
@@ -31,92 +39,101 @@ class QuarkW8A8Int8(QuarkScheme):
|
||||
# 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):
|
||||
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.qscheme == "per_channel"),
|
||||
is_static_input_scheme=(self.is_static_input_scheme is True),
|
||||
input_symmetric=(self.input_symmetric is True))
|
||||
input_symmetric=(self.input_symmetric is True),
|
||||
)
|
||||
|
||||
kernel_type = choose_scaled_mm_linear_kernel(
|
||||
scaled_mm_linear_kernel_config)
|
||||
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 QuarkW8A8Int8", 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)
|
||||
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.qscheme == "per_channel":
|
||||
weight_scale = ChannelQuantScaleParameter(
|
||||
data=torch.empty((sum(output_partition_sizes)),
|
||||
dtype=torch.float32),
|
||||
data=torch.empty((sum(output_partition_sizes)), dtype=torch.float32),
|
||||
output_dim=0,
|
||||
weight_loader=weight_loader)
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
ChannelQuantZPParameter = ChannelQuantScaleParameter
|
||||
weight_zero_point = ChannelQuantZPParameter(
|
||||
data=torch.empty((sum(output_partition_sizes)),
|
||||
dtype=torch.int8),
|
||||
data=torch.empty((sum(output_partition_sizes)), dtype=torch.int8),
|
||||
output_dim=0,
|
||||
weight_loader=weight_loader)
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
else:
|
||||
assert self.qscheme == "per_tensor"
|
||||
weight_scale = PerTensorScaleParameter(data=torch.empty(
|
||||
len(output_partition_sizes), dtype=torch.float32),
|
||||
weight_loader=weight_loader)
|
||||
weight_scale = PerTensorScaleParameter(
|
||||
data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
PerTensorZPParameter = PerTensorScaleParameter
|
||||
weight_zero_point = PerTensorZPParameter(
|
||||
data=torch.empty(len(output_partition_sizes),
|
||||
dtype=torch.int8),
|
||||
weight_loader=weight_loader)
|
||||
data=torch.empty(len(output_partition_sizes), dtype=torch.int8),
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
layer.register_parameter("weight_scale", weight_scale)
|
||||
layer.register_parameter("weight_zero_point", weight_zero_point)
|
||||
|
||||
# INPUT SCALE
|
||||
if self.is_static_input_scheme:
|
||||
input_scale = BasevLLMParameter(data=torch.empty(
|
||||
1, dtype=torch.float32),
|
||||
weight_loader=weight_loader)
|
||||
input_scale = BasevLLMParameter(
|
||||
data=torch.empty(1, dtype=torch.float32), weight_loader=weight_loader
|
||||
)
|
||||
layer.register_parameter("input_scale", input_scale)
|
||||
|
||||
input_zero_point = BasevLLMParameter(data=torch.empty(
|
||||
1, dtype=torch.int8),
|
||||
weight_loader=weight_loader)
|
||||
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")
|
||||
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 quark 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:
|
||||
layer.register_parameter("weight_zero_point", None)
|
||||
delattr(layer, 'weight_zero_point')
|
||||
delattr(layer, "weight_zero_point")
|
||||
if self.input_symmetric:
|
||||
layer.register_parameter("input_zero_point", None)
|
||||
delattr(layer, 'input_zero_point')
|
||||
delattr(layer, "input_zero_point")
|
||||
|
||||
self.kernel.process_weights_after_loading(layer)
|
||||
|
||||
def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor]) -> torch.Tensor:
|
||||
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)
|
||||
|
||||
Reference in New Issue
Block a user