Repository navigation

All pull requests

Labels (25)Milestones (0)New pull request

Open

(103)

Closed

(2,329)

Fix env_var_path_add to use os.pathsep for Windows PYTHONPATH
#4527kevin9327 opened Nov 12, 2026
Fix typo: unecessary → unnecessary in example comment
#4526ninglinLiu opened Nov 12, 20262 of 5 tasks
Compile checkpointed blocks across graph breaks in compile_regions
#4524qgallouedec opened Nov 11, 2026Member
1
Account for requested upcasts when planning model memory
#4523josepha-mayo opened Nov 11, 20262 tasks done
Allow opt-out of linear-attention refusal and make disable_in_eval configurable in sequence parallelism (#4476)
#4522Rajan-vish opened Nov 11, 2026
[FSDP2] Avoid capturing model state dict during prepare when cpu_ram_efficient_loading is False (#4511)
#4521Rajan-vish opened Nov 10, 2026
docs: point the Kornia Trainer link at a published docs version
#4520mrhy7176 opened Nov 10, 20263 of 5 tasks
Allow preparing a model without an optimizer under FSDP2
#4519qgallouedec opened Nov 9, 2026Member
4
Fix layerwise casting of MultiheadAttention
#4518remote-controlled-man opened Nov 9, 20263 of 4 tasks
1
Fix nvidia-smi output parsing on Windows
#4517emo-ocean opened Nov 9, 20262 of 4 tasks
Refuse merging FSDP weights into a checkpoint dir that gets removed
#4516assass1nlike opened Nov 9, 20264 of 5 tasks
chore(deps): bump the actions group across 1 directory with 4 updates
#4515dependabot[bot] opened Nov 9, 2026Bot
1
Fix scheduler association with prepared optimizers
#4513remote-controlled-man opened Nov 9, 20263 of 4 tasks
1
Preserve loaded tied weights when restoring initialized models
#4512remote-controlled-man opened Nov 8, 20262 of 3 tasks
1
data_loader: do not mutate sampler-owned batches in BatchSamplerShard padding
#4509davrens101 opened Nov 8, 2026
2
Allow variadic arguments in find_executable_batch_size
#4508Yushu-Li opened Nov 7, 2026
Stop BatchSamplerShard padding from mutating sampler-owned batches
#4507inchang-ing opened Nov 6, 2026
Avoid mutating sampler-owned batches during distributed padding
#4506davrens101 opened Nov 6, 20265 tasks done
Persist partial gradients in save_state for mid-accumulation checkpoints
#4504inchang-ing opened Nov 6, 2026
9
Continue automatic checkpoint numbering after load_state()
#4502ely2ba opened Nov 6, 20262 of 3 tasks
Fix metric gathering with sized samplers
#4501gorkadabo1 opened Nov 5, 20263 of 5 tasks
1
docs: fix crashing fp8 example (kwarg_handlers -> kwargs_handlers)
#4500shaikn6 opened Nov 5, 2026
Don't force CPU in write_basic_config on Apple Silicon
#4498udsy19 opened Nov 3, 20262 tasks done
1
Fix scheduler stepping too fast with tensor/context parallelism
#4497akanyaani opened Nov 3, 20262 of 5 tasks
Handle unknown batch size in sharded dataloader metadata
#4491Aurantium-47 opened Nov 2, 20263 of 4 tasks

Navigation Menu

Repository navigation

All pull requests

Fix env_var_path_add to use os.pathsep for Windows PYTHONPATH

Fix typo: unecessary → unnecessary in example comment

Compile checkpointed blocks across graph breaks in `compile_regions`

Account for requested upcasts when planning model memory

Allow opt-out of linear-attention refusal and make disable_in_eval configurable in sequence parallelism (#4476)

[FSDP2] Avoid capturing model state dict during prepare when cpu_ram_efficient_loading is False (#4511)

docs: point the Kornia Trainer link at a published docs version

Allow preparing a model without an optimizer under FSDP2

Fix layerwise casting of MultiheadAttention

Fix nvidia-smi output parsing on Windows

Refuse merging FSDP weights into a checkpoint dir that gets removed

chore(deps): bump the actions group across 1 directory with 4 updates

Fix scheduler association with prepared optimizers

Preserve loaded tied weights when restoring initialized models

data_loader: do not mutate sampler-owned batches in BatchSamplerShard padding

Allow variadic arguments in find_executable_batch_size

Stop BatchSamplerShard padding from mutating sampler-owned batches

Avoid mutating sampler-owned batches during distributed padding

Persist partial gradients in save_state for mid-accumulation checkpoints

Continue automatic checkpoint numbering after load_state()

Fix metric gathering with sized samplers

docs: fix crashing fp8 example (kwarg_handlers -> kwargs_handlers)

Don't force CPU in write_basic_config on Apple Silicon

Fix scheduler stepping too fast with tensor/context parallelism

Handle unknown batch size in sharded dataloader metadata

Pull requests