[Bugfix][CI] fix typos (#34934)

Signed-off-by: 1195343015 <1195343015@qq.com>
Signed-off-by: Jiayi Yan <66017932+1195343015@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Jiayi Yan
2026-03-05 17:05:46 +00:00
committed by GitHub
co-authored by Harry Mellor
parent 8c760b6ab6
commit 6a895197fa
98 changed files with 227 additions and 366 deletions
@@ -191,7 +191,7 @@ class CompressedTensorsConfig(QuantizationConfig):
"""
Helper function to update target_scheme_map
since linear layers get fused into FusedMoE
targetting 'Linear' needs to also match
targeting 'Linear' needs to also match
FusedMoE modules.
"""
if (
@@ -2445,7 +2445,7 @@ class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
w2_scale=layer.w2_weight_scale, # group scale
g1_alphas=layer.w13_weight_chan_scale,
g2_alphas=layer.w2_weight_chan_scale,
per_act_token_quant=True, # always use dynamc per-token
per_act_token_quant=True, # always use dynamic per-token
per_out_ch_quant=True, # always use per-channel
)
@@ -261,7 +261,7 @@ class CPUAWQLinearMethod(LinearMethodBase):
zeros = pack_cols(zeros, bits, group_num, output_size).contiguous()
# make 16 output channel as a block and transpose to
# the make the block contigous
# the make the block contiguous
weight = pack_cols(weight, bits, input_size, output_size)
weight = (
weight.view(input_size, -1, 16 // pack_factor)
@@ -199,7 +199,7 @@ class TorchAOConfig(QuantizationConfig):
@classmethod
def from_config_dict_json(cls, config_dict_json: str) -> "TorchAOConfig":
"""Iniitalize class from a config_dict json string, got from
"""Initialize class from a config_dict json string, got from
torchao_config_object = some AOBaseConfig object
json.dumps(config_to_dict(torchao_config_object))
"""
@@ -255,7 +255,7 @@ def _flashinfer_fp8_blockscale_gemm_impl(
This batch-size-dependent selection is essential for maintaining model accuracy.
Benchmarks on GSM8K show a significant accuracy gap (88% vs 95%) for DeepSeek-V3.1
when using FlashInfer's DeepGEMM on M>=32. The M < 32 strategy fixes the accurracy
when using FlashInfer's DeepGEMM on M>=32. The M < 32 strategy fixes the accuracy
drop.
Args:
@@ -39,7 +39,7 @@ def query_machete_supported_group_sizes(act_type: torch.dtype) -> list[int]:
def check_machete_supports_shape(
in_features: int, out_featrues: int
in_features: int, out_features: int
) -> tuple[bool, str | None]:
if in_features % MACHETE_PREPACKED_BLOCK_SHAPE[0] != 0:
return (
@@ -47,7 +47,7 @@ def check_machete_supports_shape(
"Input features size must be divisible by "
f"{MACHETE_PREPACKED_BLOCK_SHAPE[0]}",
)
if out_featrues % MACHETE_PREPACKED_BLOCK_SHAPE[1] != 0:
if out_features % MACHETE_PREPACKED_BLOCK_SHAPE[1] != 0:
return (
False,
"Output features size must be divisible by "