ilotoki_thu/alpaca_lora_4bit
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

fix bug

Browse Source
This commit is contained in:
John Smith
2023-03-28 22:05:18 +08:00
parent f26615fc0c
commit 234004ceb5
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
matmul_utils_4bit.py
View File

@@ -108,9 +108,9 @@ def matmul4bit(x, qweight, scales, zeros, groupsize=-1):
if np.prod(x.shape[:-1]) > auto_switch_thd: if np.prod(x.shape[:-1]) > auto_switch_thd:
output = _matmul4bit_v1_recons(x, qweight, scales, zeros) output = _matmul4bit_v1_recons(x, qweight, scales, zeros)
else: else:
output = _matmul4bit_v1(x, qweight, scales, zeros) output = _matmul4bit_v1(x, qweight, scales.float(), zeros)
else: else:
output = _matmul4bit_v1(x, qweight, scales, zeros) output = _matmul4bit_v1(x, qweight, scales.float(), zeros)
else: else:
# use v2 # use v2
if use_new: if use_new:
@@ -118,9 +118,9 @@ def matmul4bit(x, qweight, scales, zeros, groupsize=-1):
if np.prod(x.shape[:-1]) > auto_switch_thd: if np.prod(x.shape[:-1]) > auto_switch_thd:
output = _matmul4bit_v2_recons(x, qweight, scales, zeros, groupsize) output = _matmul4bit_v2_recons(x, qweight, scales, zeros, groupsize)
else: else:
output = _matmul4bit_v2(x, qweight, scales, zeros, groupsize) output = _matmul4bit_v2(x, qweight, scales.float(), zeros, groupsize)
else: else:
output = _matmul4bit_v2(x, qweight, scales, zeros, groupsize) output = _matmul4bit_v2(x, qweight, scales.float(), zeros, groupsize)
return output return output