mrq/bitsandbytes-rocm
1
0
Fork 1
Code Issues 1 Pull Requests Packages Projects Releases Wiki Activity

fix a bug in quantize_no_absmax and dequantize_no_absmax with multiple gpus

Browse Source
This commit is contained in:
Ji Lin 2023-03-22 22:14:57 -04:00
parent 49a04253fb
commit b6383ba116
1 changed files with 4 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
bitsandbytes/functional.py
View File

@ -656,9 +656,11 @@ def quantize_no_absmax(A: Tensor, code: Tensor, out: Tensor = None) -> Tensor:
torch.Tensor: torch.Tensor:
Quantized 8-bit tensor. Quantized 8-bit tensor.
''' '''
prev_device = pre_call(A.device)
if out is None: out = torch.zeros_like(A, dtype=torch.uint8) if out is None: out = torch.zeros_like(A, dtype=torch.uint8)
is_on_gpu([A, out]) is_on_gpu([A, out])
lib.cquantize(get_ptr(code), get_ptr(A), get_ptr(out), ct.c_int(A.numel())) lib.cquantize(get_ptr(code), get_ptr(A), get_ptr(out), ct.c_int(A.numel()))
post_call(prev_device)
return out return out
@ -683,9 +685,11 @@ def dequantize_no_absmax(A: Tensor, code: Tensor, out: Tensor = None) -> Tensor:
torch.Tensor: torch.Tensor:
32-bit output tensor. 32-bit output tensor.
''' '''
prev_device = pre_call(A.device)
if out is None: out = torch.zeros_like(A, dtype=torch.float32) if out is None: out = torch.zeros_like(A, dtype=torch.float32)
is_on_gpu([code, A, out]) is_on_gpu([code, A, out])
lib.cdequantize(get_ptr(code), get_ptr(A), get_ptr(out), ct.c_int(A.numel())) lib.cdequantize(get_ptr(code), get_ptr(A), get_ptr(out), ct.c_int(A.numel()))
post_call(prev_device)
return out return out