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:
@@ -14,14 +14,19 @@ from vllm import _custom_ops as ops
|
||||
from vllm.model_executor.layers.fused_moe.layer import FusedMoE
|
||||
from vllm.model_executor.layers.linear import LinearMethodBase
|
||||
from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig, QuantizeMethodBase)
|
||||
QuantizationConfig,
|
||||
QuantizeMethodBase,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.gptq_utils import (
|
||||
get_linear_quant_method)
|
||||
from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
|
||||
GroupQuantScaleParameter,
|
||||
PackedColumnParameter,
|
||||
PackedvLLMParameter,
|
||||
RowvLLMParameter)
|
||||
get_linear_quant_method,
|
||||
)
|
||||
from vllm.model_executor.parameter import (
|
||||
ChannelQuantScaleParameter,
|
||||
GroupQuantScaleParameter,
|
||||
PackedColumnParameter,
|
||||
PackedvLLMParameter,
|
||||
RowvLLMParameter,
|
||||
)
|
||||
from vllm.transformers_utils.config import get_safetensors_params_metadata
|
||||
from vllm.utils import is_list_of
|
||||
|
||||
@@ -81,7 +86,8 @@ class GPTQConfig(QuantizationConfig):
|
||||
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.")
|
||||
f"supported for GPTQ, but got {self.weight_bits} bits."
|
||||
)
|
||||
|
||||
self.modules_in_block_to_quantize = modules_in_block_to_quantize or []
|
||||
|
||||
@@ -123,14 +129,22 @@ class GPTQConfig(QuantizationConfig):
|
||||
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)
|
||||
autoround_version = cls.get_from_keys_or(config, ["autoround_version"],
|
||||
default="")
|
||||
lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False)
|
||||
autoround_version = cls.get_from_keys_or(
|
||||
config, ["autoround_version"], default=""
|
||||
)
|
||||
modules_in_block_to_quantize = cls.get_from_keys_or(
|
||||
config, ["modules_in_block_to_quantize"], default=None)
|
||||
return cls(weight_bits, group_size, desc_act, lm_head_quantized,
|
||||
dynamic, autoround_version, modules_in_block_to_quantize)
|
||||
config, ["modules_in_block_to_quantize"], default=None
|
||||
)
|
||||
return cls(
|
||||
weight_bits,
|
||||
group_size,
|
||||
desc_act,
|
||||
lm_head_quantized,
|
||||
dynamic,
|
||||
autoround_version,
|
||||
modules_in_block_to_quantize,
|
||||
)
|
||||
|
||||
def get_quant_method(
|
||||
self, layer: torch.nn.Module, prefix: str
|
||||
@@ -146,43 +160,40 @@ class GPTQConfig(QuantizationConfig):
|
||||
"sym": True, # GPTQ typically uses symmetric quantization
|
||||
"lm_head": False,
|
||||
}
|
||||
return MoeWNA16Config.from_config(config).get_quant_method(
|
||||
layer, prefix)
|
||||
return MoeWNA16Config.from_config(config).get_quant_method(layer, prefix)
|
||||
|
||||
return get_linear_quant_method(self, layer, prefix, GPTQLinearMethod)
|
||||
|
||||
def apply_vllm_mapper(self, hf_to_vllm_mapper):
|
||||
if self.modules_in_block_to_quantize is not None:
|
||||
self.modules_in_block_to_quantize = hf_to_vllm_mapper.apply_list(
|
||||
self.modules_in_block_to_quantize)
|
||||
self.modules_in_block_to_quantize
|
||||
)
|
||||
|
||||
def maybe_update_config(self,
|
||||
model_name: str,
|
||||
revision: Optional[str] = None):
|
||||
def maybe_update_config(self, model_name: str, revision: Optional[str] = None):
|
||||
if self.modules_in_block_to_quantize:
|
||||
if is_list_of(self.modules_in_block_to_quantize, list):
|
||||
# original modules_in_block_to_quantize: list[list[str]]
|
||||
# flatten original modules_in_block_to_quantize
|
||||
self.modules_in_block_to_quantize = [
|
||||
item for sublist in self.modules_in_block_to_quantize
|
||||
item
|
||||
for sublist in self.modules_in_block_to_quantize
|
||||
for item in sublist
|
||||
]
|
||||
return
|
||||
|
||||
unquant_dtypes = [torch.float16, torch.bfloat16, torch.float32]
|
||||
metadata = get_safetensors_params_metadata(model_name,
|
||||
revision=revision)
|
||||
metadata = get_safetensors_params_metadata(model_name, revision=revision)
|
||||
quant_layers: set[str] = {
|
||||
param_name.rsplit(".", 1)[0]
|
||||
for param_name, info in metadata.items()
|
||||
if (dtype := info.get('dtype', None))
|
||||
if (dtype := info.get("dtype", None))
|
||||
and _SAFETENSORS_TO_TORCH_DTYPE[dtype] not in unquant_dtypes
|
||||
}
|
||||
self.modules_in_block_to_quantize = list(quant_layers)
|
||||
|
||||
|
||||
class ExllamaState(Enum):
|
||||
|
||||
UNUSED = enum.auto()
|
||||
UNINITIALIZED = enum.auto()
|
||||
READY = enum.auto()
|
||||
@@ -214,14 +225,15 @@ class GPTQLinearMethod(LinearMethodBase):
|
||||
raise ValueError(
|
||||
"The input size is not aligned with the quantized "
|
||||
"weight shape. This can be caused by too large "
|
||||
"tensor parallel size.")
|
||||
"tensor parallel size."
|
||||
)
|
||||
output_size_per_partition = sum(output_partition_sizes)
|
||||
if (output_size_per_partition % self.quant_config.pack_factor.numerator
|
||||
!= 0):
|
||||
if output_size_per_partition % self.quant_config.pack_factor.numerator != 0:
|
||||
raise ValueError(
|
||||
"The output size is not aligned with the quantized "
|
||||
"weight shape. This can be caused by too large "
|
||||
"tensor parallel size.")
|
||||
"tensor parallel size."
|
||||
)
|
||||
|
||||
if self.quant_config.group_size != -1:
|
||||
group_size = self.quant_config.group_size
|
||||
@@ -230,8 +242,10 @@ class GPTQLinearMethod(LinearMethodBase):
|
||||
exllama_state = ExllamaState.UNINITIALIZED
|
||||
scale_and_zero_size = input_size // group_size
|
||||
scale_and_zero_input_dim = None
|
||||
if (input_size != input_size_per_partition
|
||||
and self.quant_config.group_size != -1):
|
||||
if (
|
||||
input_size != input_size_per_partition
|
||||
and self.quant_config.group_size != -1
|
||||
):
|
||||
# For act-order models, we cannot use Exllama for row parallel layer
|
||||
if self.quant_config.desc_act:
|
||||
exllama_state = ExllamaState.UNUSED
|
||||
@@ -250,56 +264,56 @@ class GPTQLinearMethod(LinearMethodBase):
|
||||
output_dim=1,
|
||||
packed_dim=0,
|
||||
packed_factor=self.quant_config.pack_factor,
|
||||
weight_loader=weight_loader)
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
|
||||
g_idx = RowvLLMParameter(data=torch.tensor(
|
||||
[
|
||||
i // self.quant_config.group_size
|
||||
for i in range(input_size_per_partition)
|
||||
],
|
||||
dtype=torch.int32,
|
||||
),
|
||||
input_dim=0,
|
||||
weight_loader=weight_loader)
|
||||
g_idx = RowvLLMParameter(
|
||||
data=torch.tensor(
|
||||
[
|
||||
i // self.quant_config.group_size
|
||||
for i in range(input_size_per_partition)
|
||||
],
|
||||
dtype=torch.int32,
|
||||
),
|
||||
input_dim=0,
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
qzeros_args = {
|
||||
"data":
|
||||
torch.empty(
|
||||
"data": torch.empty(
|
||||
scale_and_zero_size,
|
||||
output_size_per_partition // self.quant_config.pack_factor,
|
||||
dtype=torch.int32,
|
||||
),
|
||||
"weight_loader":
|
||||
weight_loader
|
||||
"weight_loader": weight_loader,
|
||||
}
|
||||
weight_scale_args = {
|
||||
"data":
|
||||
torch.empty(
|
||||
"data": torch.empty(
|
||||
scale_and_zero_size,
|
||||
output_size_per_partition,
|
||||
dtype=params_dtype,
|
||||
),
|
||||
"weight_loader":
|
||||
weight_loader
|
||||
"weight_loader": weight_loader,
|
||||
}
|
||||
if scale_and_zero_input_dim is None:
|
||||
scales = ChannelQuantScaleParameter(output_dim=1,
|
||||
**weight_scale_args)
|
||||
scales = ChannelQuantScaleParameter(output_dim=1, **weight_scale_args)
|
||||
qzeros = PackedColumnParameter(
|
||||
output_dim=1,
|
||||
packed_dim=1,
|
||||
packed_factor=self.quant_config.pack_factor,
|
||||
**qzeros_args)
|
||||
**qzeros_args,
|
||||
)
|
||||
|
||||
else:
|
||||
scales = GroupQuantScaleParameter(output_dim=1,
|
||||
input_dim=0,
|
||||
**weight_scale_args)
|
||||
scales = GroupQuantScaleParameter(
|
||||
output_dim=1, input_dim=0, **weight_scale_args
|
||||
)
|
||||
qzeros = PackedvLLMParameter(
|
||||
input_dim=0,
|
||||
output_dim=1,
|
||||
packed_dim=1,
|
||||
packed_factor=self.quant_config.pack_factor,
|
||||
**qzeros_args)
|
||||
**qzeros_args,
|
||||
)
|
||||
|
||||
layer.register_parameter("qweight", qweight)
|
||||
layer.register_parameter("g_idx", g_idx)
|
||||
@@ -321,24 +335,30 @@ class GPTQLinearMethod(LinearMethodBase):
|
||||
if self.quant_config.desc_act:
|
||||
layer.g_idx.data = torch.argsort(layer.g_idx).to(torch.int)
|
||||
else:
|
||||
layer.g_idx.data = torch.empty((0, ),
|
||||
dtype=torch.int,
|
||||
device=layer.g_idx.device)
|
||||
layer.g_idx.data = torch.empty(
|
||||
(0,), dtype=torch.int, device=layer.g_idx.device
|
||||
)
|
||||
layer.exllama_state = ExllamaState.READY
|
||||
ops.gptq_shuffle(layer.qweight, layer.g_idx,
|
||||
self.quant_config.weight_bits)
|
||||
ops.gptq_shuffle(layer.qweight, layer.g_idx, self.quant_config.weight_bits)
|
||||
|
||||
def apply(self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None) -> torch.Tensor:
|
||||
out_shape = x.shape[:-1] + (layer.qweight.shape[-1], )
|
||||
def apply(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
out_shape = x.shape[:-1] + (layer.qweight.shape[-1],)
|
||||
reshaped_x = x.reshape(-1, x.shape[-1])
|
||||
|
||||
output = ops.gptq_gemm(reshaped_x, layer.qweight, layer.qzeros,
|
||||
layer.scales, layer.g_idx,
|
||||
layer.exllama_state == ExllamaState.READY,
|
||||
self.quant_config.weight_bits)
|
||||
output = ops.gptq_gemm(
|
||||
reshaped_x,
|
||||
layer.qweight,
|
||||
layer.qzeros,
|
||||
layer.scales,
|
||||
layer.g_idx,
|
||||
layer.exllama_state == ExllamaState.READY,
|
||||
self.quant_config.weight_bits,
|
||||
)
|
||||
if bias is not None:
|
||||
output.add_(bias)
|
||||
return output.reshape(out_shape)
|
||||
|
||||
Reference in New Issue
Block a user