autofl-nvflare
Skill1k starsChanged 14 days agoApache-2.0View on GitHub
Help coding agents work on an NVFlare-based Auto-FL harness that follows an autoresearch-style loop. Use when the user wants to create, edit, debug, or extend program.md, task folders such as tasks/cifar10/ and tasks/vlm_med/, task-local job.py, client.py, model.py, shared custom_aggregators.py, mutation policies, results.tsv logging, or coding-agent prompts for a bounded federated-learning research loop. This skill is specifically for NVFlare harness work where the Client API loop, DIFF upload contract, and NUM_STEPS_CURRENT_ROUND metadata must stay intact unless the user explicitly asks for a protocol upgrade.
Trending
Less talk lately.about the whole repository
Who's talking about NVIDIA/NVFlare
- Kaggle/kaggle-sdk-python“kagglesdk 0.1.31 and 0.1.32 import a missing competitions.legacy package”GitHub · Jun 30, 2026 · 1 reactions
- londonaicentre/FLIP“docs: call the NVIDIA framework FLARE, its short official name”GitHub · Oct 17, 2026
- NVIDIA/SkillEvaluator“[BUG]: Generated BENCHMARK.md does not faithfully explain a passing gate and splits canonical case IDs”GitHub · Oct 9, 2026
- Denimworld12/Hospital-Chain“feat: run hospital nodes on separate machines with provisioning kits and a cross-address verifier”GitHub · Oct 2, 2026
- ai4os/ai4os-nvflare“Update NVFlare version to 2.8”GitHub · Jul 16, 2026
The biggest 5 of 25 mentions linking to the repository that our crawl found. Each line is their own headline.
See all 25
Talked about on GitHub (25 times).
On GitHub · 25
- Jun 30, 2026kagglesdk 0.1.31 and 0.1.32 import a missing competitions.legacy package — Kaggle/kaggle-sdk-python1 reaction · 2 comments
- Oct 17, 2026docs: call the NVIDIA framework FLARE, its short official name — londonaicentre/FLIP0 reactions · 1 comment
- Oct 9, 2026[BUG]: Generated BENCHMARK.md does not faithfully explain a passing gate and splits canonical case IDs — NVIDIA/SkillEvaluatorsays it tested it0 reactions · 1 comment
- Oct 2, 2026feat: run hospital nodes on separate machines with provisioning kits and a cross-address verifier — Denimworld12/Hospital-Chain0 reactions
- Jul 16, 2026Update NVFlare version to 2.8 — ai4os/ai4os-nvflare0 reactions
20 more on GitHub
- Jul 6, 2026Upgrade NVFlare backbone 2.7.2 → 2.8.0 (deferred until a clean 2.7.2 run) — KatherLab/MediSwarm0 reactions · 3 comments
- Jun 26, 2026Upgrade NVFLARE backend 2.7.2 → 2.8.0 — londonaicentre/FLIP0 reactions
- Jun 22, 2026Add step-ca ephemeral admin certs — pcnudde/NVFlare0 reactions · 1 comment
- May 7, 2026feat(privacy): federated learning architecture-ready (Horizon Europe / EU4Health alignment) — Yanstart/VarunaPoC0 reactions · 1 comment
- Jan 7, 2026T3: Model fitting (deploying GWAS tool / model on every client) [Enamul, ++] — collaborativebioinformatics/FedGen0 reactions
- Sep 10, 2025[Feature]: From Vertical FL feasibility (#363) to first example — londonaicentre/FLIP0 reactions · 1 comment
- Jun 23, 2025Secure xgboost example — RhinoHealth/user-resources0 reactions · 3 comments
- May 14, 2025NVFlare backbone upgrade for better connection performance — KatherLab/MediSwarm0 reactions · 2 comments
- Mar 24, 2025NVFlare Integration — KCL-BMEIS/UltraFlwr0 reactions
- Jul 14, 2024Try to fix experiment already exist issue in `MLFlowHandler` — Project-MONAI/MONAI0 reactions · 11 comments
- Sep 13, 2023Add stats_sender to MonaiAlgo for FL stats — Project-MONAI/MONAI0 reactions · 1 comment
- Aug 9, 2023【CS-part3】New submissions for Wed, 9 Aug 23 — Yukeaaa/arxiv-daily0 reactions
- Jul 12, 2023New learner cifar10 — yanchengnv/NVFlare0 reactions
- May 6, 2023Bump nvflare from 2.0.10 to 2.3.0 in /federated_learning — NVIDIA/fsi-samples0 reactions · 1 comment
- Mar 22, 2023[WIP] add `FLStatsHandler` — Project-MONAI/MONAI0 reactions · 3 comments
- Mar 21, 2023Try NVFlare cifar10 example — KatherLab/swarm-learning-hpe0 reactions
- Mar 21, 2023Apply the new data split strategies — KatherLab/swarm-learning-hpe0 reactions
- Dec 6, 2022How to use bst.eval_set() and bst.update() with xgboost_ray — ray-project/xgboost_ray0 reactions · 3 comments
- May 18, 2022713 Update nvflare example for 2.1 — Project-MONAI/tutorials0 reactions · 2 comments
- Mar 22, 2022Bump nvflare from 2.0.2 to 2.0.16 in /federated_learning/nvflare/nvflare_spleen_example — Project-MONAI/tutorials0 reactions · 4 comments
In this repository16 other files it gives its agents.
- AGENTS.mdAGENTS.md
- examples/tutorials/self-paced-training/AGENTS.mdAGENTS.md
- research/auto-fl-research/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- research/auto-fl-research/CLAUDE.mdCLAUDE.md
- autofl-nvflare-reportSkill
- nvflare-autofl-reportSkill
- nvflare-autoflSkill
- nvflare-convert-huggingfaceSkill
- nvflare-convert-lightningSkill
- nvflare-convert-pytorchSkill
- nvflare-diagnose-jobSkill
- nvflare-fed-statsSkill
- nvflare-orientSkill
- nvflare-sharedSkill
- nvflare-example-skillSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
15 rules. About 0 are common, 3 are distinctive, and 12 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
12
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 8
Use this skill to keep edits to the Auto-FL NVFlare starter coherent, safe, and aligned with an autoresearch-style operating model.
- Very specific: line 12
When the target repo includes program.md, read it first and treat it as the general control plane.
- Very specific: line 14
Use the active task's mutation_schema.yaml for bounded mutation details only when program.md or the active task profile points you there, …
- Very specific: line 59
Do not change model architecture outside registered model_arch variants or the active max_model_params budget.
- Distinctive: line 66
do not create virtual environments or install dependencies unless the user explicitly asks.
- Very specific: line 68
run the active task profile's static checks and syntax validation, with TASK_DIR set to the active task
- Very specific: line 69
run the client contract validator against the active task's client.py, not a stale root-level path
- Very specific: line 70
run the active task profile's smoke command if the prepared environment has nvflare.
- Very specific: line 71
follow the active task profile's local hardware and candidate-width rules.
- Very specific: line 73
use unique RUN_LOG and job --name values for each candidate.
- Very specific: line 75
after every completed batch, update reviewed results.tsv statuses before launching the next batch: promote the selected survivor to keep, …
- Very specific: line 76
commit that ledger on the active autoresearch/ branch after baseline and completed runs/checkpoints, and commit surviving code changes as …
- Very specific: line 77
if a candidate implements a paper-derived method, include a compact source ref in the results.tsv description field and fuller citation …
- Distinctive: line 79
after setup and baseline, continue launching same-budget candidate batches until manually interrupted.
- Distinctive: line 85
Read these when relevant:
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
autofl-nvflare
Use this skill to keep edits to the Auto-FL NVFlare starter coherent, safe, and aligned with an autoresearch-style operating model.
Entry point
When the target repo includes program.md, read it first and treat it as the general control plane. Then read the active task profile; use tasks/cifar10/profile.md when the human does not specify another profile.
Use the active task's mutation_schema.yaml for bounded mutation details only when program.md or the active task profile points you there, or when choosing a mutation axis. Use AGENTS.md / CLAUDE.md only as thin local guardrails.
Core rules
Preserve these invariants unless the user explicitly asks for a protocol change:
flare.init()while flare.is_running():input_model = flare.receive()flare.send(output_model)model.load_state_dict(input_model.params, strict=True)compute_model_diff(model, global_model)output_model.params_type == ParamsType.DIFFoutput_model.meta["NUM_STEPS_CURRENT_ROUND"]- the optional
flare.is_evaluate()branch - the same selected
model_archon server and clients for a run - the active
max_model_paramscap for architecture campaigns
Preferred mutation order
- Client-local changes in the active task's
client.py- optimizer family
- scheduler settings
- local epochs, fixed local training steps, batch size, workers
- weight decay
- gradient clipping
- label smoothing
- FedProx local loss
- extra scalar metrics
- Aggregation changes in shared
tasks/shared/custom_aggregators.py- weighted aggregation refinements
- FedAvg/FedOpt-style DIFF aggregation that stays inside the existing FLModel contract
- explicit SCAFFOLD control-variate metadata when the user has opted into that protocol mode
- clipping / robust aggregation
- median or trimmed-mean style logic
- Recipe changes in the active task's
job.py- rounds
- clients
cross_site_evallaunch_external_processclient_memory_gc_rounds
- Registered architecture changes in the active task's
model.py- named
model_archvariants - parameter-count checks through
max_model_params - no new dependencies
- named
Do not change model architecture outside registered model_arch variants or the active max_model_params budget. Do not add server-coupled protocol fields outside an explicitly requested protocol mode.
FedProx is compatible as a client-local loss term. FedOpt is compatible only when it is implemented inside the custom aggregator over already-received DIFFs. SCAFFOLD is available only as an explicit opt-in mode that uses FLModel.meta for scaffold_c_diff and scaffold_c_global.
Required workflow
After making edits:
- use the interpreter and dependency rules from the active task profile. For the default CIFAR-10 profile, set and use
PYTHON=.venv/bin/pythonby default, unless the human explicitly provides a differentPYTHONvalue; treat the selected value as authoritative, verify it withtest -x "$PYTHON"and"$PYTHON" -c "import sys; assert sys.version_info[:2] == (3, 12), sys.version; print(sys.executable)", and do not search for alternate interpreters with glob or discovery commands such asls /usr/bin/python*,ls /workspace/.venv*/bin/python*, orwhich python - do not create virtual environments or install dependencies unless the user explicitly asks; if the active profile's interpreter is missing or invalid and no override was provided, tell the user to rerun that profile's preflight instead of guessing
- when initializing a campaign, use a descriptive branch tag with the pattern
<node>-<campaign-topic>-YYYYMMDD, such ash100-fedavgm-20260430orh100-archsearch-20260430; runbash scripts/init_run.sh <tag>before validation, baseline, or candidates; verifygit branch --show-currentstarts withautoresearch/; never run experiments onmain,upstream/main, the starter branch, or a shared feature branch; never use date-only branch names - run the active task profile's static checks and syntax validation, with
TASK_DIRset to the active task - run the client contract validator against the active task's
client.py, not a stale root-level path - run the active task profile's smoke command if the prepared environment has
nvflare; for non-CIFAR tasks, pass the task-specificSMOKE_ARGSor usescripts/run_iteration.shwith the active task budget - follow the active task profile's local hardware and candidate-width rules. For the default CIFAR-10/H100 profile, launch up to
PARALLEL_CANDIDATES=4same-budget candidates concurrently on one local H100 when memory allows, and reduce the width if candidates hit CUDA OOM or host contention - use the active task profile's default candidate budget unless told otherwise. For the default CIFAR-10/H100 profile, that budget is 8 clients, 20 communication rounds, 4 local epochs,
local_train_steps=0, training batch size 64, eval batch size 1024, alpha 0.5, seed 0,model_arch=moderate_cnn,max_model_params=5000000, weighted aggregation, deterministic client training, final global evaluation on site-1, and a 1200-second timeout; local epochs orlocal_train_stepsmay be swept under that runtime cap, but do not vary both in the same narrow sweep - use unique
RUN_LOGand job--namevalues for each candidate; if the active profile requires one local GPU, pin each run withCUDA_VISIBLE_DEVICES=0instead of spreading candidates across devices - record the outcome in
results.tsv;run_iteration.shinitializes the header before launching logged runs, and successful runs are appended ascandidate, which means unreviewed, not kept - after every completed batch, update reviewed
results.tsvstatuses before launching the next batch: promote the selected survivor tokeep, mark reviewed non-survivors asdiscard, leave crashes ascrash, and leave only unresolved active rows ascandidate; preferscripts/finalize_batch_status.py --last "${PARALLEL_CANDIDATES:-4}" - commit that ledger on the active
autoresearch/branch after baseline and completed runs/checkpoints, and commit surviving code changes as soon as they are kept rather than carrying them uncommitted into the next batch - if a candidate implements a paper-derived method, include a compact source ref in the
results.tsvdescription field and fuller citation details intemplates/mutation_report.md - rank the completed batch against the ledger before deciding whether to keep, narrow, or revert; rank primarily by score, use runtime as a coarse secondary signal, and prefer the faster/simpler candidate when scores are within noise
- after setup and baseline, continue launching same-budget candidate batches until manually interrupted; do not ask whether to keep going
- after every finalized batch, run
scripts/plateau_watchdog.py results.tsv; if it printsrecommendation=literature, stop local jitter sweeps and run the Camyla-inspired literature loop fromprogram.md: time it withscripts/log_literature_review.py --start/--finish, generate diverse queries, triage primary papers, extract challenge cards, score contract-safe proposals intemplates/literature_loop.md, record theliteratureevent row inresults.tsv, and launch the top compatible candidate batch next; if it printsrecommendation=continue, do not log another literature row for a routine missed batch, and keep iterating locally unless repeated crashes share one root cause or no non-duplicate safe axis remains - report the mutation hypothesis, changed files, commands run, observed outcome, literature basis, run analysis, and next mutation
References
Read these when relevant:
references/mutation-schema.mdfor the allowed mutation surfacereferences/runbook.mdfor the recommended iteration loop and reporting formatreferences/provenance.mdfor acknowledgement and attribution guidance
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

