12 Commits (4363ff3e41003ea11e19e71604351b97bb801c44)

Author	SHA1	Message	Date
Jiarui Fang	9214d1fe28	[Gemini] chunk init using runtime visited param order (#2115)	2022-12-12 18:06:16 +08:00
Jiarui Fang	e5aa8333e4	[NFC] update chunk manager API (#2119)	2022-12-12 16:57:22 +08:00
Jiarui Fang	e99edfcb51	[NFC] polish comments for Chunk class (#2116)	2022-12-12 15:39:31 +08:00
Jiarui Fang	8afc001f4f	[Gemini] chunk init use OrderedParamGenerator (#2110)	2022-12-11 21:41:13 +08:00
HELSON	63fbba3c19	[zero] add L2 gradient clipping for ZeRO (#2112) ... * [zero] add L2 gradient clipping * [testing] add MlpModel * [zero] add unit test for grad clipping * fix atol	2022-12-09 18:09:17 +08:00
Jiarui Fang	61f31c3cf0	[Gemini] NFC, polish search_chunk_configuration (#2107)	2022-12-09 15:00:39 +08:00
HELSON	c6a1a62636	[hotfix] fix zero's incompatibility with checkpoint in torch-1.12 (#1786) ... * [hotfix] fix zero's incompatibility with checkpoint in torch-1.12 * [zero] add cpu shard init * [zero] add tiny example test * [colo_tensor] fix bugs for torch-1.11	2022-11-02 16:11:34 +08:00
HELSON	f69f9bf223	[zero] add chunk init function for users (#1729) ... * add chunk manager init function * fix unit tests * add comment * add flush=True	2022-10-18 16:31:22 +08:00
HELSON	1468e4bcfc	[zero] add constant placement policy (#1705) ... * fixes memory leak when paramter is in fp16 in ZeroDDP init. * bans chunk releasement in CUDA. Only when a chunk is about to offload, it is allowed to release. * adds a constant placement policy. With it, users can allocate a reserved caching memory space for parameters.	2022-10-14 17:53:16 +08:00
HELSON	b28991dd0a	[feature] A new ZeRO implementation (#1644)	2022-10-09 09:18:51 +08:00
Jiarui Fang	c5d39215f6	Revert "[feature] new zero implementation (#1623)" (#1643) ... This reverts commit 5be118f405.	2022-09-26 10:06:03 +08:00
HELSON	5be118f405	[feature] new zero implementation (#1623)	2022-09-24 19:58:18 +08:00