[PYTHON] Bugfix on FP32 blocksparse matmul
This commit is contained in:
@@ -362,7 +362,7 @@ class matmul:
|
||||
return self.lut_cache[key]
|
||||
# C look-up table
|
||||
layout, block = self.layout, self.block
|
||||
step = 8 if dtype == torch.float32 else 16
|
||||
step = 16
|
||||
if self.mode == 'sdd':
|
||||
c_lut, c_num_locks, c_width, c_packs = _matmul.make_sdd_lut(layout, block, dtype, device)
|
||||
elif self.mode == 'dsd':
|
||||
|
Reference in New Issue
Block a user