A high-throughput and memory-efficient inference and serving engine for LLMs
[XPU] Support int4 group_size=32 W4A16 MoE (#45136)
Signed-off-by: Marceli Fylcek <marceli.fylcek@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Marceli Fylcek committed
Jun 14, 2026 at 10:14 UTC
9548a1887fe14e553c5db2c2a76e59fa79fd3ef4
1 file changed
+5
-1
vllm/model_executor/layers/fused_moe/experts/xpu_moe.py
+5
-1
@@ -19,6 +19,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
19
kFp8Static128BlockSym,
20
kFp8StaticTensorSym,
21
kInt4Static,
22
+ kInt4Static32,
23
kMxfp4Static,
24
kMxfp8Dynamic,
25
kMxfp8Static,
@@ -302,7 +303,10 @@ class XPUExpertsWNA16(XPUExperts):
303
weight_key: QuantKey | None,
304
activation_key: QuantKey | None,
305
) -> bool:
305
- return (weight_key, activation_key) == (kInt4Static, None)
306
+ return (weight_key, activation_key) in (
307
+ (kInt4Static, None),
308
+ (kInt4Static32, None),
309
+ )
310
311
312
class XPUExpertsMxFp4(XPUExperts):