Kailash ML 1.0.0 — MANDATORY for ALL ML training/inference/feature/drift/AutoML/RL work. Engine-first km.* surface (train/register/serve/track/diagnose/watch/dashboard/seed/reproduce/resume/lineage/rl_train/engine_info/list_engines/autolog) + 18 engines. Raw sklearn / pytorch / numpy training loops BLOCKED.
Kailash ML 1.0.0 — MANDATORY for ALL ML training/inference/feature/drift/AutoML/RL work. Engine-first km.* surface (train/register/serve/track/diagnose/watch/dashboard/seed/reproduce/resume/lineage/rl_train/engine_info/list_engines/autolog) + 18 engines. Raw sklearn / pytorch / numpy training loops BLOCKED.
Kailash ML 1.0.0 — Classical / Deep Learning / RL Lifecycle
Production ML lifecycle framework built on Kailash Core SDK — engine-first km.* verb surface, 18-engine discovery registry, polars-native, ONNX-default serialisation, Agent Tool Discovery for Kaizen integration, wave-released with 6 sibling packages.
1.0.0 Engine-First Surface (Canonical)
Single entry: import kailash_ml as km. Zero-arg construction. 14 lifecycle verbs + 2 discovery verbs grouped in __all__:
import kailash_ml as km
asyncwith km.track("demo") as run: # Group 1 lifecycle
result = await km.train(df, target="y") # Group 1 lifecycle
registered = await km.register(result, name="demo") # Group 1 lifecycle
server = await km.serve("demo@production") # Group 1 lifecycle# $ kailash-ml-dashboard (separate shell) # Group 1 lifecycle
km.diagnose(model) # Group 1 — DLDiagnostics / RAGDiagnostics / RLDiagnostics
km.watch(model, reference_df) # Group 1 — DriftMonitor
km.seed(42); await km.reproduce(run_id) # Group 1 — reproducibilityawait km.resume(run_id) # Group 1 — checkpoint resume
graph = await km.lineage("demo@v1", tenant_id=None) # Group 1 — LineageGraph; ambient tenant via get_current_tenant_id()await km.rl_train(env, policy) # Group 1 — RL
km.autolog() # Group 1 — sklearn/lgb/Lightning/torch auto-logging
info = km.engine_info("TrainingPipeline") # Group 6 Engine Discovery (agents MUST use this, not imports)
engines = km.list_engines() # Group 6 — 18-engine catalog per §E1.1
Quick Start fingerprint (pinned, regression-tested via ml-engines-v2 §16.3):
c962060cf467cc732df355ec9e1212cfb0d7534a3eed4480b511adad5a9ceb00
21 Canonical Specs
Authoritative domain truth lives in specs/. Read the spec before touching the code:
Legacy v0.x material below retained as internal-implementation reference only. The canonical user surface is the engine-first km.* verbs above; the spec files under specs/ml-*.md are the authority.
Install Matrix
pip install kailash-ml # Core: polars, numpy, scipy, sklearn, lightgbm, xgboost, plotly, onnx
pip install kailash-ml[dl] # + PyTorch, Lightning, transformers, timm
pip install kailash-ml[dl-gpu] # + onnxruntime-gpu
pip install kailash-ml[rl] # + Stable-Baselines3, Gymnasium
pip install kailash-ml[agents] # + kailash-kaizen (agent integration)
# NOTE: [xgb] is a no-op alias — xgboost is now a base dep (xgboost>=2.0 ships
# with CUDA built in and auto-detects GPU at runtime, CPU fallback otherwise).
pip install kailash-ml[catboost] # + CatBoost
pip install kailash-ml[explain] # + SHAP (model explainability)
pip install kailash-ml[imbalance] # + imbalanced-learn (SMOTE, ADASYN)
pip install kailash-ml[stats] # + statsmodels
pip install kailash-ml[full] # Everything (CPU)
pip install kailash-ml[all-gpu] # Everything (GPU)
from kailash_ml import ExperimentTracker
tracker = ExperimentTracker(conn)
await tracker.initialize()
asyncwith tracker.run("hyperopt-sweep") as parent:
for params in param_grid:
asyncwith tracker.run("trial", parent_run_id=parent.run_id) as child:
await child.log_params(params)
Decision Tree: kailash-ml vs kailash-align vs kailash-kaizen
You Want To...
Use
Train sklearn/LightGBM/XGBoost models
kailash-ml
Manage feature pipelines
kailash-ml
Monitor model drift
kailash-ml
Export models to ONNX
kailash-ml
Fine-tune an LLM (LoRA, DPO, RLHF)
kailash-align
Serve a fine-tuned LLM via Ollama
kailash-align
Build an AI agent with tools
kailash-kaizen
Add agent intelligence to ML engines
kailash-ml[agents] (uses Kaizen under the hood)
Train RL policies (Gymnasium)
kailash-ml[rl]
Polars-Native Rule (ABSOLUTE)
Every engine accepts and returns polars.DataFrame. Conversion to numpy/pandas/LightGBM Dataset happens ONLY in interop.py at sklearn/framework boundaries.
# DO: Work in polars throughout
df = pl.read_csv("data.csv")
await fs.ingest("features", schema, df)
# DO NOT: Convert to pandas first
df_pd = pd.read_csv("data.csv") # WRONG — polars is the native format
Interop Conversion Table
All conversions live in interop.py. Import from there only.
SQL placement: Zero raw SQL outside _feature_sql.py — all queries go through that module
Model classes: Dynamic model imports validated via validate_model_class() against ALLOWED_MODEL_PREFIXES (sklearn., lightgbm., xgboost., catboost., kailash_ml., torch., lightning.)
Financial fields: math.isfinite() on all cost/budget fields (NaN/Inf bypass comparisons)
Table prefix: Regex-validated in constructor (^[a-zA-Z_][a-zA-Z0-9_]*$)
Bounded collections: Audit trails, cost logs, trial history use deque(maxlen=N)
