Hf checkpoint conversion for distributed checkpoints#424

Open

BlueCrescent wants to merge 45 commits into

mainfrom

hf_checkpoint_conversion_for_fsdp2

Commits on Nov 27, 2025

feat(config): Added OmegaConf based serializer save_yaml_config_dict().
BlueCrescent
committed
feat(huggingface): Added conversion of distributed gpt2 checkpoints to huggingface format.
BlueCrescent
committed

Commits on Nov 28, 2025

chore: Merge branch 'fix_rotary_transform_deferred_init' into hf_checkpoint_conversion_for_fsdp2
BlueCrescent
committed
refactor: More robust parent directory path handling.

BlueCrescent
and
Copilot
authored
docs: better dcp to torch conversion docstring

BlueCrescent
and
Copilot
authored
fix: Added handling for missing directory.

BlueCrescent
and
Copilot
authored
fix: use Path instead of string

BlueCrescent
and
Copilot
authored
fix: use cpu device for dcp to torch converted checkpoints

BlueCrescent
and
Copilot
authored
fix: error handling if wrong model key is set in checkpoint conversion

BlueCrescent
and
Copilot
authored

Commits on Dec 2, 2025

feat(utility): Moved MultiProcessingCudaEnv from tests to modalities.
BlueCrescent
committed

Commits on Dec 8, 2025

Commits on Dec 9, 2025

Commits on Feb 8, 2026

fix(huggingface): For now, Huggingface version 5.0.0 is not tested.
BlueCrescent
committed

Commits on Feb 26, 2026

Commits on Feb 27, 2026

fix(huggingface): for comparison fsdp2 mixed precision <-> huggingface transformers it is not necessary anymore to retain buffers in fp32
BlueCrescent
committed

Commits on Mar 17, 2026

chore: Merge remote-tracking branch 'origin/main' into hf_checkpoint_conversion_for_fsdp2
BlueCrescent
committed

Commits on Mar 20, 2026

fix(huggingface): added missing experiments_root_path when loading checkpoint config
BlueCrescent
committed

Commits on Mar 24, 2026

feat(huggingface): Added conversion of GPT2 models using pytorch rms norm to transformers Llama models.
BlueCrescent
committed

Commits on Apr 4, 2026

chore: Merge remote-tracking branch 'origin/main' into hf_checkpoint_conversion_for_fsdp2
BlueCrescent
committed

Commits on Apr 7, 2026

feat: moved find_free_port from test to src to use it in dcp to huggingface conversion
BlueCrescent
committed

Commits on May 6, 2026

chore: Merge remote-tracking branch 'origin/main' into hf_checkpoint_conversion_for_fsdp2
BlueCrescent
committed

Commits on Jun 26, 2026

chore: Merge remote-tracking branch 'origin/main' into hf_checkpoint_conversion_for_fsdp2
BlueCrescent
committed

Commits on Jul 14, 2026

fix(huggingface): fixes to reflect recent code changes
BlueCrescent
committed