trending/miles
1
0
Fork 0
mirror of https://github.com/radixark/miles.git synced 2026-10-02 07:14:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(convert): restore auto pipeline parallelism when ETP is unset (#3653)

Browse Source 
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Jiajun Li
2026-09-23 23:19:39 -07:00
committed by GitHub
co-authored by Claude Opus 5.5
parent a20947e65c
commit 2fbcd7601d
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/convert_hf_to_torch_dist.py
+6 -4
View File
@@ -74,13 +74,15 @@ def get_args():
args.pipeline_model_parallel_size == 1
and args.tensor_model_parallel_size == 1
and args.context_parallel_size == 1
and args.expert_model_parallel_size == 1
and args.expert_tensor_parallel_size == 1
and world_size > 1
# ETP defaults to None (= TP) until validate_args resolves it.
and (args.expert_tensor_parallel_size or args.tensor_model_parallel_size) == 1
# Each pipeline stage must hold whole EP groups, so auto PP only fills the ranks EP leaves.
and world_size % args.expert_model_parallel_size == 0
and world_size > args.expert_model_parallel_size
and not os.environ.get("CONVERT_KEEP_PP1")
)
if auto_pipeline_parallel:
pp_size = world_size
pp_size = world_size // args.expert_model_parallel_size
while True:
args.pipeline_model_parallel_size = pp_size
args.decoder_last_pipeline_num_layers = args.num_layers - ceildiv(