Dev (Qwen 3.5 recipe) - #22

Merged

haizhongzheng merged 18 commits into

mainfrom

Jul 13, 2026

Commits on Jun 22, 2026

Add Qwen3.5-4B math RL recipes (full + delta) + Qwen3.5 enablement

haizhongzheng
and
claude
committed
Make Qwen3 dense recipes run on transformers-5.8 / torch-2.11 stack

haizhongzheng
and
claude
committed
hygiene: drop dead is_qwen3_5_model; document qwen3_5 model routing

haizhongzheng
and
claude
committed
docs: add Qwen3.5-4B math recipe README (validated stack + eval)

haizhongzheng
and
claude
committed
deps: pin transformers 5.8.1 (validated stack) + kernels 0.14

haizhongzheng
and
claude
committed
deps: pin sglang to the validated git build (qwen3_5 + transformers 5.8.1)

haizhongzheng
and
claude
committed
style: align comments with repo conventions; refresh sglang install note

haizhongzheng
and
claude
committed
deps: match astraflow35 -- pin flash-linear-attention 0.5.0; networkx 3.6.1

haizhongzheng
and
claude
committed

Commits on Jun 23, 2026

recipes: default dense Qwen3 to kernels-hub FlashAttention-2

haizhongzheng
and
claude
committed

Commits on Jun 24, 2026

deps: pin sglang to validated 0.5.13.post1 release (supersedes git build)

haizhongzheng
and
claude
committed

Commits on Jun 25, 2026

docs: note Qwen3.5-4B re-validation on pinned sglang 0.5.13.post1

haizhongzheng
and
claude
committed

Commits on Jul 10, 2026

recipes: set chunked_prefill_size=8192 for Qwen3.5-4B (H100 rollout fix)

haizhongzheng
and
claude
committed

Commits on Jul 12, 2026

Commits on Jul 13, 2026