MLOPS & MODELS

Your models, managed like infrastructure.

BETA

GPU fleet, model registry, training runs, gated promotion, canary serving, drift and cost — on the servers, containers, clusters and storage Talos already manages. Talos records where your weights are and what it observed about them. It never holds the bytes.

Install itSee the security model
NVIDIA / ROCm·PyTorch · ONNX · GGUF·vLLM · Triton · KServe·DVC · LakeFS·MLflow · W&B · Hugging Face

THE MANUAL WAY

What running models without a control plane looks like.

THE SPREADSHEET

Which weights are in production, which dataset version produced them, which commit, which run, who approved it. Four people have four answers, and the authoritative one is a tab somebody last edited in March.

THE IDLE ACCELERATOR

Three A100s have been under five per cent for a fortnight, held by a run whose owner left the company. Nothing inventories them, so nothing notices. It is the most boring expensive problem in ML.

THE GREEN TICK THAT MEANT NOTHING

A model was promoted because an evaluation passed. The evaluation ran on different bytes, under a threshold somebody loosened afterwards. The tick was real and it proved nothing.

WHAT TALOS DOES

Six things that stop living in a notebook.

GPU fleet, discovered not declared

A read-only pass over the SSH path that already exists finds every accelerator, its driver, its CUDA version and its UUID — the identity that survives a reseat. It installs nothing.

A registry of pointers, not weights

Talos records an artifact URI and a SHA-256 checksum. Gigabytes of weights never transit the backend and never enter the database — the bytes stay in your storage.

Training runs as tracked jobs

Launch on a GPU server, a Kubernetes Job or a hypervisor VM. Every run records the GPU-hours it consumed and the image digest that makes it reproducible, never a mutable tag.

Promotion gated on the exact artifact

A version cannot enter production unless a required suite passed on that checksum, under the thresholds in force today. Loosening a threshold does not inherit last month’s pass.

Canary on the proxy you already run

Shifting five per cent of inference traffic to a new revision is a weight change on an upstream that already exists. Rollback is one write, not a redeploy.

Cost and idle findings, with their evidence

Observed GPU-hours against an immutable rate history, so a figure quoted last month reproduces today. Every finding carries its window, its row counts and the table each number came from.

talos.internal/mlops/gpu
Servers
Kubernetes
Frappe
Docker
Databases
Pipelines
Monitoring
MLOps
GPU fleet2 hosts · 8 accelerators · 640 GB
gpu-01 · 4× A100-SXM4OBSERVED
320 GB · driver 550.90.07 · CUDA 12.4
gpu-02 · 4× A100-SXM4OBSERVED
320 GB · driver 550.90.07 · CUDA 12.4
run r_8fj2 · support-classifierSUCCEEDED
7.4 GPU-hours · sha256:9f2c41a8…
gpu-02 · devices 1–3IDLE
under 5% for 72h · held by run r_2b71
$ ml gpu discover gpu-01
→ 4 devices · 320 GB · CUDA 12.4 · read-only
✓ artifact sha256:9f2c41a8… verified
! 3 devices under 5% for 72h — held by run r_2b71
✓ gate: holdout-accuracy passed on this checksum

HOW IT RUNS

Every action is a tracked job.

Requestyou, API or schedule
Queuegated if required
the Sentinelbackground worker
AdapterSSH · Ansible · Terraform · API
Your infrastructurelogs stream back live

WHAT MAKES THIS DIFFERENT

A stage is an instruction. A revision is what is actually deployed.

Promoting a model version to production changes one database column. It deploys nothing, and Talos refuses to let the word imply traffic — only an endpoint revision changes what a user experiences. The same distinction runs through the whole type: a checksum is an observation, a registered version is a claim, and an evaluation bound to that exact checksum is the only thing that turns one into the other. Talos keeps its own ledger so it works with none of MLflow, W&B or Hugging Face installed, and reads from them when they are.

Weights, datasets and images are referenced, never stored. Talos holds a URI, an immutable revision and a digest — the bytes stay in your systems.
Serving, training and traffic shifting ride the Docker, Kubernetes and proxy machinery that already exists rather than growing a fourth deployer.
Recommendations are advisory and carry their evidence, including whether there was enough of it. A GPU with no samples is a different finding from a GPU reading zero.
coverage · what shipped
✓
GPU fleet
discovery · utilisation · attribution · idle findings · cost
BUILT
✓
Registry
models · versions · lineage · artifact stores · checksums
BUILT
✓
Training
SSH · Kubernetes Job · sweeps · GPU-hours · image digest
BUILT
✓
Serving
endpoints · revisions · canary · rollback · health
BUILT
✓
Evidence
evaluations · promotion gate · drift · recommendations
BUILT
◐
Cloud GPU provisioning
AWS path built; the live test is written and has not been run
UNPROVEN

RELATED CAPABILITIES

Kubernetes

Training Jobs and serving Deployments on clusters you already manage.

explore →

Servers

The GPU boxes are ordinary managed hosts, discovered over SSH.

explore →

Pipelines & Gates

Model promotion uses the same approval machinery as everything else.

explore →

Point it at one GPU box.

Discovery is read-only. It installs nothing, changes no configuration, and tells you what is actually in the machine.

Install in one commandSee every capability →