Charlie Doern 6494658a10 feat: add finetune_multi_device recipe with fsdp support ... the HF SFTTrainer supports distributed training using FSDP. Add a new recipe, `finetune_multi_device` which supports multi-GPU (cuda) training using FSDP and optionally LoRA. transformers hides _alot_ of their usage of FSDP behind the training args: a6b51e7341/src/transformers/training_args.py (L1535) you need to pass both `fsdp` and `fsdp_config` to get it to work properly. However, it seems many of the `fsdp_config` entries are silently ignored. The key things to get this working were: full_shard offload (cpu offload) transformer_layer_cls_to_wrap (model specific wrapping) cpu_ram_efficient_loading sharding_strategy limit_all_gathers sync_module_states backward_prefetch use_orig_params these can be seen both in `fsdp=` and `fsdp_config=` int he `SFTConfig` call. I have tested this with different model architectures with and without LoRA with success. the user can now toggle `recipe` in their provider config between `single` and `multi` to access the two different recipes. for debugging purposes NCCL logging settings can now be accessed via the provider config as well Signed-off-by: Charlie Doern <cdoern@redhat.com>		2025-06-12 13:33:33 -04:00
..
agents	refactor: unify stream and non-stream impls for responses (#2388)	2025-06-05 17:48:09 +02:00
datasetio	chore(refact): move paginate_records fn outside of datasetio (#2137)	2025-05-12 10:56:14 -07:00
eval	feat: implementation for agent/session list and describe (#1606)	2025-05-07 14:49:23 +02:00
files/localfs	feat: reference implementation for files API (#2330)	2025-06-02 21:54:24 -07:00
inference	feat: New OpenAI compat embeddings API (#2314)	2025-05-31 22:11:47 -07:00
ios/inference	chore: removed executorch submodule (#1265)	2025-02-25 21:57:21 -08:00
post_training	feat: add finetune_multi_device recipe with fsdp support	2025-06-12 13:33:33 -04:00
safety	feat: add cpu/cuda config for prompt guard (#2194)	2025-05-28 12:23:15 -07:00
scoring	chore: enable pyupgrade fixes (#1806)	2025-05-01 14:23:50 -07:00
telemetry	revert: "chore: Remove zero-width space characters from OTEL service" (#2331)	2025-06-02 14:21:35 -07:00
tool_runtime	feat: Enable ingestion of precomputed embeddings (#2317)	2025-05-31 04:03:37 -06:00
vector_io	feat: OpenAIVectorIOMixin for vector_stores common logic (#2427)	2025-06-11 15:40:57 -07:00
__init__.py	impls -> inline, adapters -> remote (#381)	2024-11-06 14:54:05 -08:00