Skip to content

perf(qwen3-next): set expandable_segments on GB300 BF16/FP8_MX to fix OOM #3767

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
ko3n1g merged 1 commit into from
May 11, 2026
+7 −3
Merged

perf(qwen3-next): set expandable_segments on GB300 BF16/FP8_MX to fix OOM #3767

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
10 changes: 7 additions & 3 deletions scripts/performance/perf_plugins.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -284,13 +284,17 @@ def _set_model_specific_environment_variables(
model_family_name in ["qwen"]
and model_recipe_name in ["qwen3_next_80b_a3b"]
and train_task == "pretrain"
and gpu in ["h100"]
and compute_dtype == "fp8_cs"
and (
(gpu == "h100" and compute_dtype == "fp8_cs")
or (gpu == "gb300" and compute_dtype in ["bf16", "fp8_mx"])
)
):
# NCCL 2.29.7 increases memory pressure on H100, causing allocator
# fragmentation OOM. expandable_segments lets the allocator reclaim
# fragmented physical memory and avoids the OOM without disabling
# any NCCL algorithms.
# any NCCL algorithms. The GB300 BF16/FP8_MX path hits the same
# fragmentation pattern at MBS=4 under HybridEP + TE-scoped CUDA
# graphs (attn/moe_router/moe_preprocess), so the same fix applies.
executor.env_vars["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

if model_family_name in ["deepseek"]:
Expand Down
Loading