ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Fix dequant global buffer dtype reuse across mixed precision

Browse source
This commit is contained in:
Daniel Han-Chen 2026-02-11 12:16:47 +00:00
commit 631b36db15
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
unsloth/kernels/utils.py
View file

@ -388,7 +388,7 @@ if DEVICE_TYPE == "xpu" and HAS_XPU_STREAM:
global ABSMAX_BUFFERS
WEIGHT_BUFFER = WEIGHT_BUFFERS[device_index]
ABSMAX_BUFFER = ABSMAX_BUFFERS[device_index]
if WEIGHT_BUFFER is None:
if WEIGHT_BUFFER is None or WEIGHT_BUFFER.dtype != dtype:
WEIGHT_BUFFERS[device_index] = WEIGHT_BUFFER = torch_empty(
size, dtype = dtype, device = device, requires_grad = False
)
@ -498,7 +498,7 @@ elif DEVICE_TYPE in ("cuda", "hip") and HAS_CUDA_STREAM:
global ABSMAX_BUFFERS
WEIGHT_BUFFER = WEIGHT_BUFFERS[device_index]
ABSMAX_BUFFER = ABSMAX_BUFFERS[device_index]
if WEIGHT_BUFFER is None:
if WEIGHT_BUFFER is None or WEIGHT_BUFFER.dtype != dtype:
WEIGHT_BUFFERS[device_index] = WEIGHT_BUFFER = torch_empty(
size, dtype = dtype, device = device, requires_grad = False
)