Ollama: Tightening Up the Serving Layer
Today's activity centers on server reliability and correctness — fixing a goroutine leak on client disconnect, correcting GPU device filtering on mixed hardware, and easing an outdated parallel-request restriction — alongside smaller API compatibility and documentation fixes.
Duration: PT2M40S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-07-12T13:01:04Z
- Audio duration: PT2M40S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning, it's July 12th, 2026, and this is Ollama.
The clearest theme today is server-side reliability, with three separate fixes touching how Ollama handles requests under real-world conditions.
First, PR 17135 addresses a bug where streaming goroutines in the generate, push, and chat handlers could block forever if a client disconnected mid-stream. The underlying Gin streaming mechanism stops reading from the channel once it detects the client is gone, but nothing was draining the channel afterward —…
Second, PR 17137 fixes GPU discovery on mixed NVIDIA and AMD systems. Setting the ROCm visibility variables to negative one was supposed to disable AMD devices entirely, and the setting was being read correctly, but it wasn't actually being applied when the scheduler built its list of candidate devices. So disabled…
Third, PR 17144 relaxes a hardcoded restriction that forced qwen35 and qwen35moe models to run single-threaded. That limit was a safety measure against a known llama-dot-cpp crash under parallel load. The author confirms that crash was fixed upstream back in March, and Ollama's vendored llama-dot-cpp already…
A second, smaller theme is API correctness. PR 17145 adds…
Nearby episodes from Ollama
- Thinking Model Output Is Getting a Real Fix
- Stability Sweep Across Cloud, GPU, and Agent Tools
- Qwen3.5 Gets Untangled, and Small Bugs With Big Blast Radius
- Model Behavior Bugs and Blob Security Hardening
- Launch Gets Smarter, Model Support Gets Wider
- Cleaning Up the Scheduler's Edge Cases
- Weekly Recap - Rebuilding the Model Pipeline and Tightening the Guardrails
- Truth in Reporting