feat(trtllm): support agentic training for SWE rollout#3130

Open

hchings wants to merge 21 commits into

NVIDIA-NeMo:mainfrom

hchings:trtllm-agentic-swe

Commits on Jul 24, 2026

add trtllm agentic scripts

hchings
authored and
joyang-nv
committed
4b smoke e2e for trtllm (agentic): prompt_token_ids + max_input_len fixes, agentic Dockerfile, 4b smoke script

hchings
authored and
joyang-nv
committed
30B 16 nodes trtllm swe

hchings
authored and
joyang-nv
committed
fix multiturn, fix env

hchings
authored and
joyang-nv
committed
fix empty gen token case

hchings
authored and
joyang-nv
committed
refactor

hchings
authored and
joyang-nv
committed
rebase fix

hchings
authored and
joyang-nv
committed
cleanup

hchings
authored and
joyang-nv
committed
refactor into utils, cleanup

hchings
authored and
joyang-nv
committed
reconcile trtllm http-server piping post-rebase

hchings
authored and
joyang-nv
committed
address comments

hchings
authored and
joyang-nv
committed
address comments, pre-commit

hchings
authored and
joyang-nv
committed
multinode rollout for non colocated, allow quantized ckpt

hchings
authored and
joyang-nv
committed
cross-node rollout + tool parser refactor and fixes

hchings
authored and
joyang-nv
committed
qwen3.5b refit

hchings
authored and
joyang-nv
committed
address comments (reasoning parsers, eos handling, tests)

hchings
authored and
joyang-nv
committed
trtllm: nsys profiling of the generation (rollout) GPU workers

Superjomn
authored and
joyang-nv
committed
docs: nsys profiling for TRT-LLM generation (rollout) workers

Superjomn
authored and
joyang-nv
committed
fix reasoning parsing

hchings
authored and
joyang-nv
committed
fix: satisfy lint checks
joyang-nv
committed
fix: include OpenAI server utils in Pyrefly
joyang-nv
committed