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 6, 08:15 UTC
[Bugfix] Reject channelwise quantization (group_size <= 0) in ExllamaLinearKernel (#37331)
Signed-off-by: Matthias Gehre <matthias.gehre@amd.com>
Matthias Gehre committed
Mar 20, 2026 at 16:31 UTC
56a62c310cc4840671949488c60c40df5e0e2f1f
1 file changed
+7
vllm/model_executor/kernels/linear/mixed_precision/exllama.py
+7
@@ -59,6 +59,13 @@ class ExllamaLinearKernel(MPLinearKernel):
59
f"{cls.SUPPORTED_QUANT_TYPES}",
60
)
61
62
+ if c.group_size <= 0:
63
+ return (
64
+ False,
65
+ f"Group size ({c.group_size}) must be positive, "
66
+ "Exllama does not support channelwise quantization",
67
+ )
68
+
69
if c.full_weight_shape[0] % c.group_size != 0:
70
return (
71
False,