CompressedGemma commited on
Commit
0e9c7e9
·
verified ·
1 Parent(s): 75e8d10

Upload 3 files

Browse files
Files changed (1) hide show
  1. segmentation.py +4 -2
segmentation.py CHANGED
@@ -239,8 +239,10 @@ def q3k_proxy(xg: np.ndarray, wg: np.ndarray | None) -> np.ndarray:
239
  wu = np.ones_like(u)
240
  else:
241
  wu = np.asarray(wg, dtype=np.float64).reshape(n * (B // 16), 16)
242
- scale, xmin, codes = make_qkx2_batch(u, wu, nmax=7)
243
- deq = scale[:, None] * codes + xmin[:, None]
 
 
244
  return deq.reshape(n, B)
245
 
246
 
 
239
  wu = np.ones_like(u)
240
  else:
241
  wu = np.asarray(wg, dtype=np.float64).reshape(n * (B // 16), 16)
242
+ # make_qkx2_batch returns (scale, -min, codes) with the GGML positive-dmin
243
+ # convention; reconstruction is scale * code - (-min).
244
+ scale, neg_min, codes = make_qkx2_batch(u, wu, nmax=7)
245
+ deq = scale[:, None] * codes - neg_min[:, None]
246
  return deq.reshape(n, B)
247
 
248