feat: Fix CPU offloading + add options for FSDP offload and expandable segments#122

Closed

yfw wants to merge 27 commits intomainfrom

yifu/cpu_offload

Commits on Apr 2, 2025

WIP commit
yfw
committed
fix: Remove reference of tokenizer from generation backend (#75) (#82)

parthchadha
authored and
yfw
committed
feat: unit test metric tracking (#40)

terrykong
authored and
yfw
committed
fix: unit test error when coverage wasn't specified (#88)

terrykong
authored and
yfw
committed
ci: temporarily disable CI on main since PRs must be up to date before merge (#91)

terrykong
authored and
yfw
committed
Cleanup + expandable_segments
yfw
committed
More cleanup
yfw
committed
cpu offload
yfw
committed
Make everything configurable
yfw
committed
expandable_segments and fsdp offload configs
yfw
committed
Make port configurable in ray.sub
yfw
committed
fix: error out early if ray cluster does not have resources (#89)

parthchadha
authored and
yfw
committed
ci: skip functional until more capacity available and/or tests speed up (#94)

terrykong
authored and
yfw
committed
feat: evaluation implement (#16)

yuki-97
authored and
yfw
committed
fix: gradient should be averaged instead of summed across mbs (#86)

parthchadha
authored and
yfw
committed
fix: Use separate step_metric for GPU Monitoring (#92)

yfw
and
terrykong
committed
feat: Update sft config to use single GPU (#90)

authored and
yfw
committed
fix: Grammar nit (#98)

SahilJain314
authored and
yfw
committed
feat: add capability to set min/max eps separately as proposed in the… (#95)

authored and
yfw
committed
fix: change format messages to out of place (#77)

authored and
yfw
committed
fix: correct version and use setuptools.dynamic metadata for version/readme (#104)

terrykong
authored and
yfw
committed
fix: remove usage of vllm to get device uuid and instead use nvidia-m… (#105)

parthchadha
authored and
yfw
committed
fix: Change optional-dependencies to dependency-groups (#81)

hemildesai
authored and
yfw
committed
feat: Add support for hydra style overrides (#80)

hemildesai
authored and
yfw
committed
fix: Do not initialize reference model for sft (#71)

authored and
yfw
committed
fix: change grpo default to use 64 prompts per step and 32 generation… (#111)

parthchadha
authored and
yfw
committed
feat: use cuda_graph by default for vllm (#116)

parthchadha
authored and
yfw
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: Fix CPU offloading + add options for FSDP offload and expandable segments#122

feat: Fix CPU offloading + add options for FSDP offload and expandable segments#122
yfw wants to merge 27 commits intomainfrom
yifu/cpu_offload

Commits on Apr 2, 2025

WIP commit

fix: Remove reference of tokenizer from generation backend (#75) (#82)

feat: unit test metric tracking (#40)

fix: unit test error when coverage wasn't specified (#88)

ci: temporarily disable CI on main since PRs must be up to date before merge (#91)

Cleanup + expandable_segments

More cleanup

cpu offload

Make everything configurable

expandable_segments and fsdp offload configs

Make port configurable in ray.sub

fix: error out early if ray cluster does not have resources (#89)

ci: skip functional until more capacity available and/or tests speed up (#94)

feat: evaluation implement (#16)

fix: gradient should be averaged instead of summed across mbs (#86)

fix: Use separate step_metric for GPU Monitoring (#92)

feat: Update sft config to use single GPU (#90)

fix: Grammar nit (#98)

feat: add capability to set min/max eps separately as proposed in the… (#95)

fix: change format messages to out of place (#77)

fix: correct version and use setuptools.dynamic metadata for version/readme (#104)

fix: remove usage of vllm to get device uuid and instead use nvidia-m… (#105)

fix: Change optional-dependencies to dependency-groups (#81)

feat: Add support for hydra style overrides (#80)

fix: Do not initialize reference model for sft (#71)

fix: change grpo default to use 64 prompts per step and 32 generation… (#111)

feat: use cuda_graph by default for vllm (#116)