Ollama: Hardening the Runner Scheduler

Several pull requests today focus on making the model runner and scheduler more resilient under real-world constraints, from cache persistence to container CPU limits, while a batch of smaller fixes cleans up build packaging and documentation.

Duration: PT2M35S

Episode overview

This episode is a short developer briefing from Ollama.

It explains recent repository work in plain language.

  • Show: Ollama
  • Published: 2026-08-24T13:01:41Z
  • Audio duration: PT2M35S

Transcript excerpt

This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.

Good day, it's August 24th, and this is your Ollama briefing.

The clearest signal today is a push to make the runner and scheduler layer more reliable under pressure. Three separate pull requests attack this from different angles. PR seventeen-nine-five-three adds experimental prefill cache persistence, so a model reload doesn't force a full prompt recomputation - useful for…

Second theme: build and packaging cleanup on the MLX side. PR seventeen-nine-five-six deduplicates CUDA runtime payloads between the standard CUDA build and the MLX CUDA build, fixing path discovery so both in-process and subprocess loading find the right libraries. A companion PR, seventeen-nine-five-five, bumps…

Smaller items round out the day: PR seventeen-nine-six-five auto-detects the ornith and qwen35 renderer and parser, fixing malformed grammar generation when tools and a JSON schema were combined under native chat mode. PR seventeen-nine-five-nine adds an optional auto mode preference in app settings. And PR…

What to remember: if you're running Ollama in containers, the cgroup-aware threading fix directly affects throughput. And the exited-runner refusal logic in PR seventeen-nine-five-four is worth…

That…

Nearby episodes from Ollama

  1. Claude Desktop Integration Overhaul
  2. Weekly Recap - Desktop Apps Grow Up, Inference Gets More Reliable
  3. Visibility Into What the Model Actually Sees
  4. Fixing the Long-Prompt Timeout Trap
  5. Claude Desktop Gets Deep App Integration
  6. Closing the Gaps on Hung Requests and Broken Installs
  7. Closing the Gap on Model Metadata Overhead
  8. Silent Failures Get Loud