Skip to content

Remove forward_backward batch cap from 256k Qwen3.8-27B definition #64

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
micahtyong merged 1 commit into from
Sep 23, 2026
Merged

Remove forward_backward batch cap from 256k Qwen3.8-27B definition #64

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 0 additions & 6 deletions src/lilo/providers/modal/definitions/qwen3_8_27b_miles_lora_256k.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,11 +47,6 @@
# all-to-all runs behind a straggler's recompute, so a rank can sit in one
# collective far longer than ten minutes without anything being wrong.
DISTRIBUTED_TIMEOUT_MINUTES = 120
# Coalescing a whole rollout's datums into one Miles call turns a step into a
# single multi-thousand-collective forward_backward across both nodes, where
# one desynchronized rank wedges every process group. One datum per call keeps
# the collective chains short; gradients still accumulate until optim_step.
MAX_FORWARD_BACKWARD_BATCH = 1

ROLLOUT_GPU_TYPE = "H200"
ROLLOUT_GPUS = 4
Expand Down Expand Up @@ -222,7 +217,6 @@ def run_trainer(
nproc=1,
max_models=max_models,
sampler_persistence_concurrency=8,
max_forward_backward_batch=MAX_FORWARD_BACKWARD_BATCH,
)


Expand Down
Loading