You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
- Added support for saving and loading DeepSpeed checkpoints through `Fabric.save/load()` ([#16452](https://github.com/Lightning-AI/lightning/pull/16452))
16
-
17
-
18
14
- Added support for automatically calling `set_epoch` on the `dataloader.batch_sampler.sampler` ([#16841](https://github.com/Lightning-AI/lightning/pull/16841))
19
-
20
-
21
15
- Added support for writing logs to remote file systems with the `CSVLogger` ([#16880](https://github.com/Lightning-AI/lightning/pull/16880))
22
-
23
-
24
16
- Added support for frozen dataclasses in the optimizer state ([#16656](https://github.com/Lightning-AI/lightning/pull/16656))
25
-
26
-
27
17
- Added `lightning.fabric.is_wrapped` to check whether a module, optimizer, or dataloader was already wrapped by Fabric ([#16953](https://github.com/Lightning-AI/lightning/pull/16953))
28
18
29
-
30
19
### Changed
31
20
32
21
- Fabric now chooses `accelerator="auto", strategy="auto", devices="auto"` as defaults ([#16842](https://github.com/Lightning-AI/lightning/pull/16842))
33
-
34
-
35
22
- Checkpoint saving and loading redesign ([#16434](https://github.com/Lightning-AI/lightning/pull/16434))
36
23
* Changed the method signatrue of `Fabric.save` and `Fabric.load`
37
24
* Changed the method signature of `Strategy.save_checkpoint` and `Fabric.load_checkpoint`
38
25
*`Fabric.save` accepts a state that can contain model and optimizer references
39
26
*`Fabric.load` can now load state in-place onto models and optimizers
40
27
*`Fabric.load` returns a dictionary of objects that weren't loaded into the state
41
28
*`Strategy.save_checkpoint` and `Fabric.load_checkpoint` are now responsible for accessing the state of the model and optimizers
42
-
43
-
44
29
-`DataParallelStrategy.get_module_state_dict()` and `DDPStrategy.get_module_state_dict()` now correctly extracts the state dict without keys prefixed with 'module' ([#16487](https://github.com/Lightning-AI/lightning/pull/16487))
*`strategy="fsdp_full_shard_offload"` is now `strategy="fsdp_cpu_offload"`
48
32
*`lightning.fabric.plugins.precision.native_amp` is now `lightning.fabric.plugins.precision.amp`
49
-
50
-
51
33
- Enabled all shorthand strategy names that can be supported in the CLI ([#16485](https://github.com/Lightning-AI/lightning/pull/16485))
52
-
53
34
- Renamed `strategy='tpu_spawn'` to `strategy='xla'` and `strategy='tpu_spawn_debug'` to `strategy='xla_debug'` ([#16781](https://github.com/Lightning-AI/lightning/pull/16781))
54
-
55
-
56
35
- Changed arguments for precision settings (from [64|32|16|bf16] to ["64-true"|"32-true"|"16-mixed"|"bf16-mixed"]) ([#16767](https://github.com/Lightning-AI/lightning/pull/16767))
57
-
58
36
- The selection `Fabric(strategy="ddp_spawn", ...)` no longer falls back to "ddp" when a cluster environment gets detected ([#16780](https://github.com/Lightning-AI/lightning/pull/16780))
59
-
60
37
- Renamed `setup_dataloaders(replace_sampler=...)` to `setup_dataloaders(use_distributed_sampler=...)` ([#16829](https://github.com/Lightning-AI/lightning/pull/16829))
61
38
62
-
### Deprecated
63
-
64
-
-
65
-
66
-
67
39
### Removed
68
40
69
41
- Removed support for PyTorch 1.10 ([#16492](https://github.com/Lightning-AI/lightning/pull/16492))
70
-
71
-
72
42
- Removed support for Python 3.7 ([#16579](https://github.com/Lightning-AI/lightning/pull/16579))
73
43
74
-
75
44
### Fixed
76
45
77
46
- Fixed issue where the wrapped dataloader `iter()` would be called twice ([#16841](https://github.com/Lightning-AI/lightning/pull/16841))
0 commit comments