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:
Harry Mellor
2025-10-05 07:06:22 -07:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115205 additions and 94107 deletions
@@ -8,8 +8,10 @@ from torch.nn import Module
from torch.nn.parameter import Parameter
from vllm.model_executor.layers.linear import LinearBase, LinearMethodBase
from vllm.model_executor.layers.quantization import (QuantizationConfig,
QuantizationMethods)
from vllm.model_executor.layers.quantization import (
QuantizationConfig,
QuantizationMethods,
)
from vllm.model_executor.parameter import ModelWeightParameter
ACTIVATION_SCHEMES = ["none", "dynamic"]
@@ -24,8 +26,7 @@ class Int8TpuConfig(QuantizationConfig):
) -> None:
super().__init__()
if activation_scheme not in ACTIVATION_SCHEMES:
raise ValueError(
f"Unsupported activation scheme {activation_scheme}")
raise ValueError(f"Unsupported activation scheme {activation_scheme}")
self.activation_scheme = activation_scheme
def get_name(self) -> QuantizationMethods:
@@ -36,8 +37,7 @@ class Int8TpuConfig(QuantizationConfig):
@classmethod
def get_min_capability(cls) -> int:
raise NotImplementedError(
"This function should not be called with TPU Backend")
raise NotImplementedError("This function should not be called with TPU Backend")
@staticmethod
def get_config_filenames() -> list[str]:
@@ -48,50 +48,61 @@ class Int8TpuConfig(QuantizationConfig):
activation_scheme = cls.get_from_keys(config, ["activation_scheme"])
return cls(activation_scheme=activation_scheme)
def get_quant_method(self, layer: Module,
prefix: str) -> Optional["TPUInt8LinearMethod"]:
def get_quant_method(
self, layer: Module, prefix: str
) -> Optional["TPUInt8LinearMethod"]:
if isinstance(layer, LinearBase):
return TPUInt8LinearMethod(self)
return None
class TPUInt8LinearMethod(LinearMethodBase):
"""Int8 Linear method for TPU Quant. """
"""Int8 Linear method for TPU Quant."""
def __init__(self, quant_config: Int8TpuConfig):
self.quant_config = quant_config
self.quantize_activation = False
if self.quant_config.activation_scheme == 'dynamic':
if self.quant_config.activation_scheme == "dynamic":
self.quantize_activation = True
def create_weights(self, layer: Module, input_size_per_partition: int,
output_partition_sizes: list[int], input_size: int,
output_size: int, params_dtype: torch.dtype,
**extra_weight_attrs):
def create_weights(
self,
layer: Module,
input_size_per_partition: int,
output_partition_sizes: list[int],
input_size: int,
output_size: int,
params_dtype: torch.dtype,
**extra_weight_attrs,
):
weight_loader = extra_weight_attrs.get("weight_loader")
weight = ModelWeightParameter(data=torch.empty(
sum(output_partition_sizes),
input_size_per_partition,
dtype=params_dtype),
input_dim=1,
output_dim=0,
weight_loader=weight_loader)
weight = ModelWeightParameter(
data=torch.empty(
sum(output_partition_sizes),
input_size_per_partition,
dtype=params_dtype,
),
input_dim=1,
output_dim=0,
weight_loader=weight_loader,
)
layer.register_parameter("weight", weight)
def _quantize_weight(
self, weight: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
self, weight: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
weight_dtype = weight.dtype
weight = weight.cpu().to(torch.float32)
n_bit = 8
eps = 1e-5
max_int = 2**(n_bit - 1) - 1
min_int = -(2**(n_bit - 1))
max_int = 2 ** (n_bit - 1) - 1
min_int = -(2 ** (n_bit - 1))
max_val = weight.abs().amax(dim=-1, keepdim=True)
max_val = max_val.clamp(min=eps)
qscale = max_val / max_int
qweight = torch.clamp(torch.round(weight * (1.0 / qscale)), min_int,
max_int).to(torch.int8)
qweight = torch.clamp(
torch.round(weight * (1.0 / qscale)), min_int, max_int
).to(torch.int8)
qscale = qscale.squeeze().to(weight_dtype)
return qweight, qscale
@@ -104,21 +115,25 @@ class TPUInt8LinearMethod(LinearMethodBase):
layer.weight = Parameter(qweight, requires_grad=False)
layer.scale = Parameter(qscale, requires_grad=False)
def apply(self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: Optional[torch.Tensor] = None) -> torch.Tensor:
def apply(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: Optional[torch.Tensor] = None,
) -> torch.Tensor:
try:
import torch_xla.experimental.custom_kernel # noqa: F401
except ImportError as err:
raise ImportError(
"Please install torch_xla by following the instructions at "
"https://docs.vllm.ai/en/latest/getting_started/tpu-installation.html " # noqa: E501
"to run vLLM on TPU.") from err
"to run vLLM on TPU."
) from err
weight = layer.weight
scale = layer.scale
out = torch.ops.xla.quantized_matmul_int8(
x, weight, scale, quantize_activation=self.quantize_activation)
x, weight, scale, quantize_activation=self.quantize_activation
)
if bias is not None:
out = out + bias
return out