Ollama: Resource Leaks and GPU Placement Get a Clean-Up Pass
A cluster of fixes targets goroutine and memory leaks across the GGUF parser, recurrent model caches, and background conversion jobs, while a separate group of changes corrects how Ollama places models across multiple GPUs. Both threads point to the same lesson: edge cases in resource handling and hardware detection have been quietly costing users crashes, wasted VRAM, and poor performance.
Duration: PT2M43S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-07-15T13:00:40Z
- Audio duration: PT2M43S
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 July 15th, 2026, and here's what mattered in Ollama's codebase.
The dominant pattern today is resource leaks getting hunted down and fixed. Three separate contributors converged on the same problem: goroutines and file handles left dangling when GGUF file parsing fails partway through. PRs 17206 and 17184 both close a leak where a truncated model file causes Ollama's parser to…
The second theme is GPU placement accuracy. Multiple reports, including issue 17018, described multi-GPU setups where a smaller card got zero layers assigned, forcing output tensors onto the CPU despite plenty of free VRAM elsewhere. PR 17194 fixes this by ordering the GPU spread group by most free memory first.…
Smaller but notable: PR 17182 brings Gemma4's chat template in line with upstream fixes for tool-calling and multi-turn reasoning. And PR 17195 addresses repeated-output bugs in GLM-OCR by registering a missing end-of-generation token.
What to watch next: the MLX cache rework from Jesse Gross is laying groundwork for MTP speculative decoding ahead of the qwen3.6 model, so expect more runner-internal changes before that lands. And with several independent leak fixes converging this…
That's…
Nearby episodes from Ollama
- Cleaning Up Memory Planning and the Launch Experience
- Fixing What Defaults Got Wrong
- Weekly Recap - Model Correctness and Runtime Hardening
- Tightening Up the Serving Layer
- 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