nemo
Router for trlX's optional NVIDIA NeMo/Megatron backend: when to choose it, how configs and checkpoints fit together, and what remains unverified without NeMo/Apex.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- nemo
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Router for trlX's optional NVIDIA NeMo/Megatron backend: when to choose it, how configs and checkpoints fit together, and what remains unverified without NeMo/Apex.
- disable-model-invocation
- true
02 资源文件 3
- 参考文档references/configuration.md6.6 KB
- 参考文档references/troubleshooting.md5.4 KB
- 参考文档references/workflows.md7.3 KB
03 SKILL.md 正文
nemo
Use this sub-skill only when the task explicitly depends on the trlX NeMo / Megatron backend, .nemo checkpoints, or the trlX NeMo wrappers.
Route here when the request mentions
NeMoPPOTrainer,NeMoILQLTrainer, orNeMoSFTTrainerPPOGPT,ILQLGPT, orSFTGPTdefault_nemo_20b_config,default_nemo_2b_config, ordefault_nemo_1_3b_config- NeMo YAMLs, tensor/pipeline parallel sizing, rank-sharded checkpoints, or LLaMA-to-NeMo conversion
- checkpoint loading, NeMo-style inference, or
pretrained_model/megatron_cfgwiring - the optional NeMo / Apex / Megatron backend and its install prerequisites
Route elsewhere when the task is really about
- standard
trlx.trainPPO / ILQL / SFT / RFT on Accelerate, reward functions, samples/rewards, PEFT, data shapes, or Ray sweeps →../training/SKILL.md - generic NeMo usage that is not tied to trlX wrappers or configs
- direct NeMo/Apex installation, large-model conversion, or distributed launch execution
Backend status in this inspection
- CUDA torch smoke passed in the minimum env.
- NeMo / Apex / Megatron were not installed in that env.
- The installed trlX package registers the NeMo trainer names, but they resolve to dummy import-error stubs until NeMo is installed.
- Treat all NeMo guidance here as source-backed and intentionally unverified for the local backend.
What this sub-skill provides
references/workflows.mdfor setup, conversion, training, inference, and default selection.references/configuration.mdfor NeMo YAML shape and trlx-to-NeMo config mapping.references/troubleshooting.mdfor import, Apex/CUDA, parallelism, checkpoint, tokenizer, and memory failures.
Provenance used to distill this guidance
Source evidence was read from the NeMo-related trainer/model files, NeMo defaults, and docs/examples in the repo, including docs/source/installation.rst, docs/source/api.rst, trlx/models/README.md, examples/llama_nemo/README.md, trlx/data/default_configs.py, trlx/trainer/nemo_*.py, trlx/models/modeling_nemo_*.py, and configs/nemo_configs/*.yaml.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).