Re-enable the 80 char line width limit (#3305)

This commit is contained in:
Zhuohan Li
2024-03-10 19:49:14 -07:00
committed by GitHub
parent 4b59f00e91
commit 2f8844ba08
67 changed files with 557 additions and 528 deletions
@@ -31,8 +31,8 @@ class GPTQConfig(QuantizationConfig):
self.pack_factor = Fraction(32, self.weight_bits)
if self.weight_bits not in [2, 3, 4, 8]:
raise ValueError(
"Currently, only 2/3/4/8-bit weight quantization is supported for "
f"GPTQ, but got {self.weight_bits} bits.")
"Currently, only 2/3/4/8-bit weight quantization is "
f"supported for GPTQ, but got {self.weight_bits} bits.")
def __repr__(self) -> str:
return (f"GPTQConfig(weight_bits={self.weight_bits}, "
@@ -101,7 +101,8 @@ class GPTQLinearMethod(LinearMethodBase):
"The input size is not aligned with the quantized "
"weight shape. This can be caused by too large "
"tensor parallel size.")
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 "
@@ -114,7 +115,8 @@ 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