Repository navigation

All pull requests

Fix load_kwargs for PyTorch model checkpoints
#4415mottopanikeiku opened Oct 17, 20263 of 4 tasks
Fix disk offload for module extra state
#4413AaravMehta-07 opened Oct 16, 20263 of 4 tasks
Fix gather_for_metrics remainder when a sized sampler selects a subset
#4412Eryx-Z opened Oct 16, 2026
9
[FSDP2] Full QLoRA support: PEFT adapter exclusion, LoRA-param upcast, detect helpers
#4408Asthenia0412 opened Oct 15, 2026
11
Check trained results and requested precision
#4407Titus-von-Koeller opened Oct 15, 2026Contributor
1
Preserve zero gradient clipping in DeepSpeed configuration
#4406Titus-von-Koeller opened Oct 14, 2026Contributor
1
Restore minimum-Torch CI compatibility and CPU operation coverage
#4405Titus-von-Koeller opened Oct 14, 2026Contributor
1
Keep a split module's own buffers in infer_auto_device_map
#4404Kayvan-Zahiri opened Oct 14, 20262 tasks done
Fix pad across processes keep dtype
#4400arthurdjn opened Oct 13, 20263 of 5 tasks
2
Fix send to device skip keys for mappings with .to method
#4398arthurdjn opened Oct 13, 20263 of 5 tasks
Fix TorchXLA dynamo backend config option
#4396chenshj73 opened Oct 13, 2026
Fix split-batch shard length for uneven tails
#4395AaravMehta-07 opened Oct 13, 2026
1
Fix total batch size metadata for dispatched dataloaders
#4394kocean9-freedom opened Oct 13, 20262 of 5 tasks
chore(deps): bump the actions group with 4 updates
#4391dependabot[bot] opened Oct 12, 2026Bot
1
Fix FSDP2 cpu_ram_efficient_loading for models with persistent buffers
#4390li12345678990 opened Oct 11, 2026
5
Preserve shared modules in regional compilation
#4389AaravMehta-07 opened Oct 10, 20262 of 3 tasks
1
Compile checkpointed repeated blocks across graph breaks
#4388zhechen-cs opened Oct 10, 20262 of 4 tasks
Preserve RNG state during FSDP2 activation checkpointing
#4387Michael-RDev opened Oct 10, 2026
Allow clip_grad_value_ under FSDP2
#4386hjinnkim opened Oct 10, 2026Contributor
4
Restore environment variables when patched keys differ only in case
#4385ktz03 opened Oct 10, 20263 of 5 tasks
Preserve empty tensors when saving safetensors
#4383ktz03 opened Oct 9, 20263 of 5 tasks
1
Fix PYTHONPATH separator on Windows
#4380mrhimanshu opened Oct 8, 20263 tasks done
Handle empty tensors in disk offload
#4379Afloat16 opened Oct 8, 20262 of 4 tasks
Preserve dictionary inputs in split_between_processes
#4378aalquwayfili opened Oct 8, 20262 of 3 tasks
Add BF16/FP8 training benchmarks and refresh the TE container
#4377sbhavani opened Oct 7, 2026

Navigation Menu

Repository navigation

All pull requests

Fix load_kwargs for PyTorch model checkpoints

Fix disk offload for module extra state

Fix gather_for_metrics remainder when a sized sampler selects a subset

[FSDP2] Full QLoRA support: PEFT adapter exclusion, LoRA-param upcast, detect helpers

Check trained results and requested precision

Preserve zero gradient clipping in DeepSpeed configuration

Restore minimum-Torch CI compatibility and CPU operation coverage

Keep a split module's own buffers in infer_auto_device_map

Fix pad across processes keep dtype

Fix send to device skip keys for mappings with `.to` method

Fix TorchXLA dynamo backend config option

Fix split-batch shard length for uneven tails

Fix total batch size metadata for dispatched dataloaders

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

Fix FSDP2 cpu_ram_efficient_loading for models with persistent buffers

Preserve shared modules in regional compilation

Compile checkpointed repeated blocks across graph breaks

Preserve RNG state during FSDP2 activation checkpointing

Allow clip_grad_value_ under FSDP2

Restore environment variables when patched keys differ only in case

Preserve empty tensors when saving safetensors

Fix PYTHONPATH separator on Windows

Handle empty tensors in disk offload

Preserve dictionary inputs in split_between_processes

Add BF16/FP8 training benchmarks and refresh the TE container

Pull requests