nodes/silveroxides-ComfyUI-ModelUtils
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

perf: use svd_lowrank for fixed rank resize (10x faster)

Browse Source
This commit is contained in:
silveroxides
2026-01-03 20:14:25 +01:00
parent 3a71ad9952
commit 75d1f0f5c9
1 changed files with 30 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
nodes/lora_resize.py
+30 -14
View File
@@ -205,13 +205,21 @@ def _extract_linear(
"""Extract LoRA from merged linear weight."""
out_size, in_size = weight.shape
U, S, Vh = linalg.svd(weight, full_matrices=False)
new_rank, new_alpha, stats = _compute_resize(S, max_rank, dynamic_method, dynamic_param, scale)
U = U[:, :new_rank]
S = S[:new_rank]
Vh = Vh[:new_rank, :]
if dynamic_method is None:
# Fixed rank - use svd_lowrank for 10x speedup
rank = min(max_rank, min(out_size, in_size) - 1)
U, S, Vh = torch.svd_lowrank(weight, q=rank, niter=2)
Vh = Vh.T # svd_lowrank returns V, not Vh
new_rank = rank
new_alpha = float(scale * new_rank)
stats = {"sum_retained": 1.0, "fro_retained": 1.0} # Not computed for lowrank
else:
# Dynamic methods need full SVD to compute rank from all singular values
U, S, Vh = linalg.svd(weight, full_matrices=False)
new_rank, new_alpha, stats = _compute_resize(S, max_rank, dynamic_method, dynamic_param, scale)
U = U[:, :new_rank]
S = S[:new_rank]
Vh = Vh[:new_rank, :]
lora_up = U @ torch.diag(S)
lora_down = Vh
@@ -236,13 +244,21 @@ def _extract_conv(
out_ch, in_ch, kh, kw = weight.shape
mat = weight.reshape(out_ch, -1)
U, S, Vh = linalg.svd(mat, full_matrices=False)
new_rank, new_alpha, stats = _compute_resize(S, max_rank, dynamic_method, dynamic_param, scale)
U = U[:, :new_rank]
S = S[:new_rank]
Vh = Vh[:new_rank, :]
if dynamic_method is None:
# Fixed rank - use svd_lowrank for 10x speedup
rank = min(max_rank, min(mat.shape) - 1)
U, S, Vh = torch.svd_lowrank(mat, q=rank, niter=2)
Vh = Vh.T # svd_lowrank returns V, not Vh
new_rank = rank
new_alpha = float(scale * new_rank)
stats = {"sum_retained": 1.0, "fro_retained": 1.0} # Not computed for lowrank
else:
# Dynamic methods need full SVD to compute rank from all singular values
U, S, Vh = linalg.svd(mat, full_matrices=False)
new_rank, new_alpha, stats = _compute_resize(S, max_rank, dynamic_method, dynamic_param, scale)
U = U[:, :new_rank]
S = S[:new_rank]
Vh = Vh[:new_rank, :]
lora_up = (U @ torch.diag(S)).reshape(out_ch, new_rank, 1, 1)
lora_down = Vh.reshape(new_rank, in_ch, kh, kw)