compare to remote HEAD#1

Draft

CharlieFRuan wants to merge 38 commits intoremote-originfrom

Commits on Mar 19, 2026

feat: add EvolveGenerator training entrypoint

authored and
CharlieFRuan
committed
feat: modify max_advisor_context_iters

joyemang33
authored and
CharlieFRuan
committed
fix: add PYTHONPATH, n_samples_per_prompt=8, max_advisor_context_iters=10

authored and
CharlieFRuan
committed
fix: use 2 GPUs, Qwen3.5-9B local path, correct model name handling

joyemang33
authored and
CharlieFRuan
committed
fix: add vllm tool/attention flags to match serving setup

joyemang33
authored and
CharlieFRuan
committed
fix: change HTTP endpoint port to 8002 (8000 and 8001 occupied)

joyemang33
authored and
CharlieFRuan
committed
fix: use 4 GPUs

joyemang33
authored and
CharlieFRuan
committed
fix: set UV_CACHE_DIR to /data/qmang/uv_cache

joyemang33
authored and
CharlieFRuan
committed
fix: redirect all outputs/caches to /data/qmang

joyemang33
authored and
CharlieFRuan
committed
fix: use UV_PROJECT_ENVIRONMENT instead of symlink for venv on /data

joyemang33
authored and
CharlieFRuan
committed
update training script

joyemang33
authored and
CharlieFRuan
committed
go

joyemang33
authored and
CharlieFRuan
committed
111

joyemang33
authored and
CharlieFRuan
committed
good

joyemang33
authored and
CharlieFRuan
committed
feat: separate solver and advisor

joyemang33
authored and
CharlieFRuan
committed
clean up unused things
CharlieFRuan
committed
feat: add logging dir

joyemang33
authored and
CharlieFRuan
committed
remove jinja unused
CharlieFRuan
committed
Make script runnable with Qwen3 and non-QMang machine
CharlieFRuan
committed
train-train-train

joyemang33
authored and
CharlieFRuan
committed
feat:

joyemang33
authored and
CharlieFRuan
committed
update training script

joyemang33
authored and
CharlieFRuan
committed
update
CharlieFRuan
committed
runnable on b200 e2e

zwcolin
authored and
CharlieFRuan
committed
nonfrozen-training

joyemang33
authored and
CharlieFRuan
committed
feat: H100 e2e training config with frozen GPT-5 solver and jemalloc/pyarrow fixes

joyemang33
authored and
CharlieFRuan
committed
[WIP] Set up Qwen3.5
CharlieFRuan
committed
more changes for qwen3.5
CharlieFRuan
committed
remove claude debug things that broke weight sync
CharlieFRuan
committed

Commits on Mar 21, 2026

stepwise H100
joyemang33
committed

Commits on Mar 22, 2026

log prompt
joyemang33
committed

Commits on Mar 24, 2026

Commits on Mar 25, 2026

reduce context length in RL training
joyemang33
committed

Commits on Mar 26, 2026

finalize H100 recipe
joyemang33
committed

Commits on Mar 30, 2026

update recipe
joyemang33
committed