[Bugfix][Quantization][MoE] Normalise an unset group_size on the compressed-tensors WNA16 MoE path (#53163)

Signed-off-by: Artur Fierka <artur.fierka@intel.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Artur Fierka
2026-09-08 09:42:46 +08:00
committed by GitHub
co-authored by Claude
parent 06334fae2d
commit 9dbdf8e9af
3 changed files with 120 additions and 3 deletions
@@ -160,7 +160,11 @@ def _backend_incompatibility_reason(
WNA16MoEBackend.BATCHED_MARLIN,
):
if isinstance(quant_config, (AutoAWQConfig, AutoGPTQConfig, QuantizationArgs)):
group_size = quant_config.group_size
# QuantizationArgs leaves group_size unset for per-channel
# strategies; -1 is the in-tree encoding for that.
group_size = (
-1 if quant_config.group_size is None else quant_config.group_size
)
else:
return "Marlin not supported for this layer"
@@ -1509,7 +1513,9 @@ def convert_to_wna16_moe_kernel_format(
elif isinstance(quant_config, QuantizationArgs):
num_bits = quant_config.num_bits
pack_factor = 32 // quant_config.num_bits
group_size = quant_config.group_size
group_size = (
-1 if quant_config.group_size is None else quant_config.group_size
)
actorder = quant_config.actorder
else:
raise TypeError(
@@ -70,7 +70,11 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
self.num_bits = weight_quant.num_bits
self.packed_factor = Fraction(32, weight_quant.num_bits)
self.strategy = weight_quant.strategy
self.group_size = weight_quant.group_size
# Per-channel strategies leave group_size unset; -1 is the in-tree
# encoding for that, as in CompressedTensorsWNA16 and siblings.
self.group_size = (
-1 if weight_quant.group_size is None else weight_quant.group_size
)
self.actorder = weight_quant.actorder
# Extract quant_type and create weight key for oracle selection