Choose a path
Kiln documentation
Start with a product workflow below. Use search or open the reference library when you need an exact field, operating contract, benchmark protocol, or engineering guarantee.
Product guides
Start with a workflow
Each guide begins with a working action, states the relevant limits, and links to exact reference material when you need it.
01 · RunQuickstartInstall Kiln, start Qwen3.5-4B, and verify the first response.Run Kiln →02 · TrainGRPO GuideStart Kiln normally, submit a scored group, and inspect the adapter and receipt.Train an adapter →03 · EvaluateOpenEnv TrainingTrain directly against stateful OpenEnv environments with seed-matched episodes, native GRPO, ECHO, and replayable receipts.Evaluate a change →04 · IntegrateEvals GuideGrade outputs, compare base and adapters on shared seeds, and feed reviewed judgments back to training.Integrate the API →
Core documentation
Use these shorter guides for product scope, configuration, security, performance, and reasoning limits.
Start here
Kiln overviewDecide whether Kiln fits your workload, then choose an installation path and first workflow.Configuration referenceConfigure typed TOML and mechanically derived environment overrides; understand precedence, validation, and backend-owned policy such as SFT loss routing.Security policyFind supported versions, private vulnerability-reporting instructions, and deployment trust boundaries.
Serving
Training and evals
Reference libraryField-level schemas, operating contracts, qualification evidence, and maintainer policy 51 documents
Start here
Complete configuration referenceLook up every typed field, canonical environment override, retired name, validation rule, and provenance boundary.Configuration schemaInspect the machine-readable defaults, constraints, environment mappings, migrations, and cross-field rules for kiln.toml.HTTP API contractLook up every production HTTP operation, transport, request and response schema, handler owner, error, and trust boundary.ChangelogSee user-visible changes and compatibility notes for each release.
Serving
Inference API schemaLook up inference requests, responses, streaming events, thinking limits, timing, batch behavior, and rollout provenance.Observability API schemaInterpret health, readiness, resolved configuration, model diagnostics, decode statistics, recent requests, and cache responses.Architecture deep diveTrace runtime policy, request scheduling, accelerator ownership, capability-based routes, training, evaluation, and safe failure.HF next-token request schemaDefine the source, model, rendered prompt, continuation prefix, and candidates for an independent Hugging Face next-token comparison.ROCm/HF next-token result schemaInterpret the bound source, process, model, full logits, retained artifact, and candidate attribution for one ROCm oracle run.ROCm/HF path-attribution result schemaCompare eager and retained HIP-graph logits and greedy selections against one bound Hugging Face reference state.ROCm/HF layer-attribution result schemaLocate numerical divergence after embedding, each transformer layer, and final RMSNorm while preserving source, process, model, and implementation identity.Serving benchmark launch schemaDefine direct argv launch, working and log directories, readiness checks, startup and shutdown bounds, and accepted exits for an owned benchmark server.Serving benchmark protocolBind a workload, own the server process, measure wall-clock performance, interpret each metric, and retain comparable evidence.Serving profilesUse the complete stable product by default, drained maintenance when needed, and experimental only for backend qualification.Latency observabilityChoose the correct request, stream, rolling, or Prometheus view; interpret TTFT, ITL, phases, and stalls without mixing metric populations.Thinking budget schemaLook up request, server-default, effective, outcome, recent-record, source, and trigger fields for thinking budgets.
Training and evals
Native SFT profileUnderstand the fixed native online-LoRA update, backend-owned loss route, memory admission, receipt, and checkpoint identity.SFT ingestion and row identityDetermine which SFT rows are admitted, how content identity and deduplication work, and what failures and receipts record.SFT tokenization and lossFollow canonical chat rendering, assistant-only labeling, masking, truncation, and parity verification.Native training checkpointsUnderstand checkpoint contents, publication, cancellation, validation, resume, promotion, and recovery, including route-bound SFT planning identity.GRPO training guideBuild scored groups, submit and observe GRPO, inspect artifacts, resume safely, and audit the result.OpenEnv replay and recoveryVerify content-bound rollout artifacts, exactly replay environment exchanges, and understand protocol recovery, capacity, drift, and implementation-neutral conformance.OpenEnv rollout schemaLook up environment discovery identity, episode outcomes, recovery statistics, dataset/replay hashes, exact transcript, verification report, and summary receipt published by kiln openenv.ECHO guideUnderstand when ECHO applies, construct environment-feedback requests, choose an operating mode, and interpret diagnostics and limits.Dataset splits and train/eval separationCreate deterministic group-aware splits, prevent contamination, synthesize held-out data, and preserve training-data provenance.Eval guideCreate suites and scorers, grade and compare outputs, replay results, enforce post-training gates, and train judges.Eval and judgment API schemaLook up suite, scorer, eval, dataset, synthesis, and pairwise-judgment fields, validation, results, and errors.Training and agent control plane API schemaLook up native training, distillation, preflight, agent, trace, recipe, correction, job, and adapter-library operations and state transitions.
Interoperability
HF/TRL export and PEFT importExport immutable SFT or recorded-GRPO bundles, train with HF/TRL/PEFT, and import adapters only after resident identity checks pass.Artifact lifecycle API schemaLook up adapter discovery and mutation, immutable exports, verified imports, audit receipts, and teacher identity fields.OPD teacher JSONLCreate and validate an identity-bound offline teacher corpus for OPD training.Immutable vLLM teacher identityLaunch, register, and qualify an exact remote vLLM teacher without turning one machine's choices into portable defaults.
Integrity and artifacts
Adapter manifests and restoreMove and verify an adapter as a content-bound file set, then check compatibility before activation.Base-weight identity and compatibilityCompare exact model-shard content without confusing names, paths, or one matching hash with full resume compatibility.Execution identity and provenanceRead, compare, and preserve Kiln's startup-owned executable, runtime, model, configuration, precision, and kernel identity.Request-lineage verificationVerify the historical hashed request chain without mistaking it for training replay, output reproduction, or authenticated provenance.Training receiptsInterpret one SFT, GRPO, or OPD attempt without confusing a receipt with a checkpoint, quality result, or reproduction proof.Qualification workload contractDefine a deterministic, device-neutral local workload with bounded inputs, owned evidence, and explicit comparison rules.Qualification case-result contractInterpret command-owned case evidence without mistaking it for runner-normalized results, comparison policy, or the qualification verdict.Qualification receipt contractInterpret a run receipt’s source, environment, model, workload, case outcomes, evidence hashes, verdict authority, and trust limits.
Operations and development
Local hardware qualificationRun, validate, compare, and publish a reproducible hardware qualification with explicit pass criteria and failure triage.Legacy backend latency evidenceMaintain or retire the machine-bound legacy fixture pipeline without confusing its thresholds with qualification, support policy, or portable performance.Runtime environment inventoryFind every direct environment read and mutation, its owner and class, and the migration or test boundary around it.Repository artifact retentionDecide which evidence belongs in Git, which raw artifacts remain external, and how limits, exceptions, archival, and recovery work.Source-parsing test debtKeep source-text assertions out of correctness claims, understand the scanner’s exact scope, and replace any future debt with behavioral evidence.ContributingSet up development, scope a change, run the right checks, update documentation, and prepare the work for review.CI and local qualification policyKnow which inexpensive checks run in hosted CI, which accelerator qualifications stay local, why jobs skip, and what evidence to retain.Verification policyChoose the required behavioral, typed, structured, or hardware evidence; understand ownership, exceptions, and the zero-source-text-test rule.Release version policyUpdate version references from their canonical owners and use drift checks without rewriting historical evidence.Tensor substrate quickstartCreate a backend-portable tensor operation and verify its device, dtype, shape, numerical, and error contracts.Backend capability reportInterpret current CUDA, ROCm, Vulkan, Metal, and CPU coverage by capability, constraint, source owner, and report date.