A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 07:30 UTC
[Bug] Fix AttributeError: 'ColumnParallelLinear' object has no attribute `weight_scale_inv` (#30823)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Dec 17, 2025 at 05:00 UTC
f284d7bd0c55f929fa7912936b1d247089679191
1 file changed
+5
-2
vllm/model_executor/layers/quantization/utils/fp8_utils.py
+5
-2
@@ -1437,14 +1437,17 @@ def maybe_post_process_fp8_weight_block(layer: torch.nn.Module):
1437
layer.orig_dtype, layer.weight
1438
)
1439
if should_use_deepgemm:
1440
+ scale_attr = (
1441
+ "weight_scale_inv" if hasattr(layer, "weight_scale_inv") else "weight_scale"
1442
+ )
1443
dg_weight, dg_weight_scale = deepgemm_post_process_fp8_weight_block(
1444
wq=layer.weight.data,
1442
- ws=layer.weight_scale_inv.data,
1445
+ ws=getattr(layer, scale_attr).data,
1446
quant_block_shape=tuple(layer.weight_block_size),
1447
use_e8m0=is_deep_gemm_e8m0_used(),
1448
)
1449
replace_parameter(layer, "weight", dg_weight)
1447
- replace_parameter(layer, "weight_scale_inv", dg_weight_scale)
1450
+ replace_parameter(layer, scale_attr, dg_weight_scale)
1451
1452
1453
def expert_weight_is_col_major(x: torch.Tensor) -> bool: