Skip to content

fix(smooth): update convert smooth weight for hy_v3 in transformer @gavingavin99 #338

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

Closed
gavingavin99 wants to merge 2 commits into from
Closed

fix(smooth): update convert smooth weight for hy_v3 in transformer @gavingavin99 #338

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7e99d29
fix(smooth) fix typos and change default smooth config
gavingavin99 Jun 9, 2026
a541d77
fix(smooth): update convert smooth weight for hy_v3 in transformer
gavingavin99 Jun 10, 2026
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
8 changes: 4 additions & 4 deletions configs/Hy3/ptq/fp8/Hy3_smooth.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,8 +27,8 @@ collect_moe: true
ema_momentum: 0.9

# ========== Phase 2: smooth flavours & fixed alpha ==========
smooth_qk: true
smooth_vo: true
smooth_qk: false
smooth_vo: false
smooth_down: true
alpha_qk: 0.6
alpha_vo: 0.5
Expand Down Expand Up @@ -56,5 +56,5 @@ alpha_smooth_search_mode: default
# per-tensor-act-first mode parameters (only effective when mode=per-tensor-act-first)
# alpha_act_mul_min: 0.1
# alpha_act_mul_max: 1.0
# alpha_smooth_min: 1e-6
# alpha_smooth_max: 1e6
# alpha_smooth_min: 0.000001
# alpha_smooth_max: 100000
8 changes: 4 additions & 4 deletions scripts/ptq/run_smooth_for_HY3.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,8 +33,6 @@ done
# -------- Environment Variables --------
# Allow function serialization for apply_model in vLLM v1 engine
export VLLM_ALLOW_INSECURE_SERIALIZATION=1
# Enable MoE expert statistics collection
export VLLM_MOE_COLLECT_STATS=1
# Force Ray to reload code (disable code caching)
export RAY_DEDUP_LOGS=0
# Force Python to not use bytecode cache
Expand All @@ -43,7 +41,6 @@ export PYTHONDONTWRITEBYTECODE=1

export MAX_NUM_BATCHED_TOKENS=32768
export VLLM_ENABLE_CHUNKED_PREFILL=1
export MOE_MODE=fused
export VLLM_ATTENTION_BACKEND=FLASHINFER
export ASYNC_SCHEDULING=1
export VLLM_ENABLE_PREFIX_CACHING=1
Expand All @@ -54,7 +51,6 @@ export PRECISIONMODE=HF
export VLLM_MOE_COLLECT_SMOOTH_STATS=1
export VLLM_MOE_COLLECT_ALPHA_SEARCH=1

export PYTHONPATH=/cfs_cloud_code/gavinlee/work/open_source_smooth/AngelSlim
# -------- Phase 1: Collect Smooth Stats + Alpha Search --------
if [ "$SKIP_CALIBRATE" = false ]; then
echo "========================================"
Expand All @@ -72,6 +68,10 @@ if [ "$SKIP_CONVERT" = false ]; then
python3 tools/smooth/convert_smooth_weights.py -c "$CONFIG"
fi

# revert
unset VLLM_MOE_COLLECT_SMOOTH_STATS
unset VLLM_MOE_COLLECT_ALPHA_SEARCH

echo "========================================"
echo "Done."
echo "========================================"
Loading
Loading