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 14, 22:35 UTC
[Misc] Expand SUPPORTED_HIDDEN_SIZES for DeepEP low-latency kernels (#21818)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Jul 31, 2025 at 11:41 UTC
0f7919fca05d7cf60b773da26d898b72bc07a089
1 file changed
+1
-1
vllm/model_executor/layers/fused_moe/deepep_ll_prepare_finalize.py
+1
-1
@@ -40,7 +40,7 @@ class DeepEPLLPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
40
41
# DeepEP low-latency kernels are compiled only for certain
42
# specific hidden sizes.
43
- SUPPORTED_HIDDEN_SIZES = [2048, 2560, 4096, 5120, 7168]
43
+ SUPPORTED_HIDDEN_SIZES = [2048, 2560, 4096, 5120, 6144, 7168]
44
45
def __init__(self,
46
buffer: deep_ep.Buffer,