adv-optm 1.2.dev8__py3-none-any.whl → 1.2.dev9__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Potentially problematic release.
This version of adv-optm might be problematic. Click here for more details.
- adv_optm/__init__.py +1 -1
- adv_optm/optim/AdaMuon_adv.py +4 -2
- {adv_optm-1.2.dev8.dist-info → adv_optm-1.2.dev9.dist-info}/METADATA +1 -1
- {adv_optm-1.2.dev8.dist-info → adv_optm-1.2.dev9.dist-info}/RECORD +7 -7
- {adv_optm-1.2.dev8.dist-info → adv_optm-1.2.dev9.dist-info}/WHEEL +0 -0
- {adv_optm-1.2.dev8.dist-info → adv_optm-1.2.dev9.dist-info}/licenses/LICENSE +0 -0
- {adv_optm-1.2.dev8.dist-info → adv_optm-1.2.dev9.dist-info}/top_level.txt +0 -0
adv_optm/__init__.py
CHANGED
adv_optm/optim/AdaMuon_adv.py
CHANGED
|
@@ -327,7 +327,8 @@ class AdaMuon_adv(torch.optim.Optimizer):
|
|
|
327
327
|
# RMS-aligned rescaling
|
|
328
328
|
rms_target = group['rms_target']
|
|
329
329
|
num_elements = update.numel()
|
|
330
|
-
|
|
330
|
+
# Add eps to prevent division by zero
|
|
331
|
+
scaling_factor = rms_target * (num_elements ** 0.5) / (update.norm() + group['eps'])
|
|
331
332
|
|
|
332
333
|
update.mul_(scaling_factor)
|
|
333
334
|
update = update.view(p.shape).mul_(group['lr'])
|
|
@@ -422,7 +423,8 @@ class AdaMuon_adv(torch.optim.Optimizer):
|
|
|
422
423
|
# RMS-aligned rescaling
|
|
423
424
|
rms_target = group['rms_target']
|
|
424
425
|
num_elements = update.numel()
|
|
425
|
-
|
|
426
|
+
# Add eps to prevent division by zero
|
|
427
|
+
scaling_factor = rms_target * (num_elements ** 0.5) / (update.norm() + group['eps'])
|
|
426
428
|
|
|
427
429
|
update.mul_(scaling_factor)
|
|
428
430
|
del num_elements, scaling_factor
|
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
adv_optm/__init__.py,sha256=
|
|
2
|
-
adv_optm/optim/AdaMuon_adv.py,sha256=
|
|
1
|
+
adv_optm/__init__.py,sha256=TzvKgGTLkK0_XANeZzhURcSO9xmtUi-H9_C7tV3rXn4,379
|
|
2
|
+
adv_optm/optim/AdaMuon_adv.py,sha256=yr1oJV339Zv7D8n148O1FJJAgdOsH8NZDZTKlcDOyu0,21181
|
|
3
3
|
adv_optm/optim/AdamW_adv.py,sha256=7IvdD1rqYeHZwQCZU9X0H7x87MCKcHQ5M68GLuMCkvE,17702
|
|
4
4
|
adv_optm/optim/Adopt_adv.py,sha256=C2FsEZGvCk9q4YNKAj0qIxdZ5AfPlda-1lIpSX0a1nE,21256
|
|
5
5
|
adv_optm/optim/Lion_Prodigy_adv.py,sha256=LEA3UYJpPeFnmxeniLNv1u2LKKj4ufx3Bq_MLw-nWXk,14617
|
|
@@ -17,8 +17,8 @@ adv_optm/util/Newton_Schulz.py,sha256=wJ_sKRaGVIsOofQ737my4ng494qX_pfgOqlDDmYtnC
|
|
|
17
17
|
adv_optm/util/One_Bit_Boolean.py,sha256=Wat49esdwohuN-OHOFMW8D0aOQgV9cP5Rl8z6yfmpos,1068
|
|
18
18
|
adv_optm/util/OrthoGrad.py,sha256=NzInuBQGy_Ja__M1R9XbvqVaQ0fhGbtGgFE9YON7B3I,707
|
|
19
19
|
adv_optm/util/__init__.py,sha256=jAaUfaAjFrTJ6-Q915ezAbq0efRbpYjriW2OdeCbSzo,433
|
|
20
|
-
adv_optm-1.2.
|
|
21
|
-
adv_optm-1.2.
|
|
22
|
-
adv_optm-1.2.
|
|
23
|
-
adv_optm-1.2.
|
|
24
|
-
adv_optm-1.2.
|
|
20
|
+
adv_optm-1.2.dev9.dist-info/licenses/LICENSE,sha256=HrhfyXIkWY2tGFK11kg7vPCqhgh5DcxleloqdhrpyMY,11558
|
|
21
|
+
adv_optm-1.2.dev9.dist-info/METADATA,sha256=GmAYWjZdfgvg9QbzyiV2PUNmzQFgJz8AjaY5F0x7Nv8,14022
|
|
22
|
+
adv_optm-1.2.dev9.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
|
|
23
|
+
adv_optm-1.2.dev9.dist-info/top_level.txt,sha256=iNfBIIzu-lPrQ7jyC56WBCcbkRwitM2nJ15-MRQ_6fg,9
|
|
24
|
+
adv_optm-1.2.dev9.dist-info/RECORD,,
|
|
File without changes
|
|
File without changes
|
|
File without changes
|