Commit 07d1278

committed

Refactor CUTLASS kernels to always take scales that reside on the GPU

1 parent 4238bc8 commit 07d1278Copy full SHA for 07d1278

5 files changed

+507

-180

lines changed

csrc/quantization/cutlass_w8a8
vllm/model_executor/layers/quantization/compressed_tensors/schemes
- compressed_tensors_w8a8_statictensor.py

5 files changed

+507

-180

lines changed

Comments

(0)