[Misc] Update GPTQ to use vLLMParameters (#7976)

This commit is contained in:
Dipika Sikka
2024-09-03 17:21:44 -04:00
committed by GitHub
parent dc0b6066ab
commit 2188a60c7e
6 changed files with 93 additions and 62 deletions
+58 -45
View File
@@ -11,7 +11,11 @@ from vllm.model_executor.layers.linear import LinearBase, LinearMethodBase
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig)
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.model_executor.utils import set_weight_attrs
from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
GroupQuantScaleParameter,
PackedColumnParameter,
PackedvLLMParameter,
RowvLLMParameter)
class GPTQConfig(QuantizationConfig):
@@ -108,6 +112,7 @@ class GPTQLinearMethod(LinearMethodBase):
**extra_weight_attrs,
):
del output_size # Unused.
weight_loader = extra_weight_attrs.get("weight_loader")
if input_size_per_partition % self.quant_config.group_size != 0:
raise ValueError(
"The input size is not aligned with the quantized "
@@ -138,73 +143,81 @@ class GPTQLinearMethod(LinearMethodBase):
scale_and_zero_size = input_size_per_partition // group_size
scale_and_zero_input_dim = 0
qweight = Parameter(
torch.empty(
qweight = PackedvLLMParameter(
data=torch.empty(
input_size_per_partition // self.quant_config.pack_factor,
output_size_per_partition,
dtype=torch.int32,
),
requires_grad=False,
)
set_weight_attrs(
qweight, {
"input_dim": 0,
"output_dim": 1,
"packed_dim": 0,
"pack_factor": self.quant_config.pack_factor,
})
g_idx = Parameter(
torch.tensor(
[
i // self.quant_config.group_size
for i in range(input_size_per_partition)
],
dtype=torch.int32,
),
requires_grad=False,
)
# Ignore warning from fused linear layers such as QKVParallelLinear.
set_weight_attrs(g_idx, {"input_dim": 0, "ignore_warning": True})
qzeros = Parameter(
input_dim=0,
output_dim=1,
packed_dim=0,
packed_factor=self.quant_config.pack_factor,
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(
scale_and_zero_size,
output_size_per_partition // self.quant_config.pack_factor,
dtype=torch.int32,
),
requires_grad=False,
)
set_weight_attrs(
qzeros, {
"input_dim": scale_and_zero_input_dim,
"output_dim": 1,
"packed_dim": 1,
"pack_factor": self.quant_config.pack_factor,
})
scales = Parameter(
"weight_loader":
weight_loader
}
weight_scale_args = {
"data":
torch.empty(
scale_and_zero_size,
output_size_per_partition,
dtype=params_dtype,
),
requires_grad=False,
)
set_weight_attrs(scales, {
"input_dim": scale_and_zero_input_dim,
"output_dim": 1,
})
"weight_loader":
weight_loader
}
if scale_and_zero_input_dim is None:
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)
else:
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)
layer.register_parameter("qweight", qweight)
set_weight_attrs(qweight, extra_weight_attrs)
layer.register_parameter("g_idx", g_idx)
set_weight_attrs(g_idx, 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)
layer.exllama_state = exllama_state
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
# for torch.compile
layer.qweight = Parameter(layer.qweight.data, requires_grad=False)
layer.qzeros = Parameter(layer.qzeros.data, requires_grad=False)
layer.qweight = Parameter(layer.qweight.data, requires_grad=False)
layer.g_idx = Parameter(layer.g_idx.data, requires_grad=False)
# exllama needs to shuffle the weight after the weight is loaded
# here we do the shuffle on first forward pass
if layer.exllama_state == ExllamaState.UNINITIALIZED: