sp + ep training / tp + ep inference #46292
3outeille
merged
32 commits into
tp_param_level
from
sp_tp_ep_plan
[distributed] Add resolve_parallel_plan merge helper
5f4729be
[distributed] Wire resolve_parallel_plan into apply_tensor_parallel
24ab0dcc
[distributed] Use merged plan in tp_plan property and load path
0b346b82
[distributed] Drop intra-expert moe_tp_* from MoE SP plans
4bb9d29a
[distributed] Document SP+EP and TP+EP flag combinations
8bbad67c
refactor merging plans
036cd6df
merge tp_param_level branch to here
9d59c8e9
merge tp_param_level to here
5e0e10e0
add test sp_ep and tp_ep
5ab3238d
extend verify_tp_plan to verify_tp_sp_ep_plan
19748bed
Merge branch 'tp_param_level' into sp_tp_ep_plan
ffbb7484
Merge branch 'tp_param_level' into sp_tp_ep_plan
d6952330
add ep_plan to mixtral and olmoe
eac0c3a5
Merge branch 'sp_tp_ep_plan' of https://github.com/huggingface/transf…
df22ffc2
cleaning _accumulate_local_param_grad (#46394)
a7031748
Add select_parallel_plan and explicit combo plan config fields
6441fdf8
Add base_model_tp_ep_plan and base_model_sp_ep_plan for Mixtral and Q…
e6898beb
Add explicit tp_ep / sp_ep plans for remaining MoE models
049b5a3a
Remove resolve_parallel_plan and use explicit combo plan selection
1fcb21ad
Sync modular MoE configs and update expert parallelism docs
241215b4
Refactor select_parallel_plan flag lookup for readability
9ab0463d
Write explicit combo parallel plans in MoE configs and remove plan_utils
66c6b970
Add lm_head entries to _tp_ep_plan and _sp_ep_plan on CausalLM classes
707d2ce8
cleaning
aae5751a
cleaning
1fdacfdc
cleaning
c7b106df
cleaning
7031566d
linting
51e0750f
add verify tp and fsdp pla
9e32233b
revert doc
a34de4fd
cleaning
e12da425
check-repository-consistency
be3f5454
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub