ollama - 💡(How to fix) Fix [Performance Regression] Significant drop in prompt evaluation speed and increased load time in `0.30.0-RC15` on qwen3:0.6b

What is the issue?

Description:

I have observed a major performance regression in ollama 0.30.0-RC15 compared to ollama 0.23.4 when running a small model (qwen3:0.6b).

While the token generation speed (eval rate) remains relatively stable, the prompt evaluation rate (prefill phase) has dropped by approximately 75% (from ~1326 tokens/s to ~322 tokens/s). Additionally, the model loading time has nearly doubled.

Environment:

• Ollama Version (Old): 0.23.4
• Ollama Version (New): 0.30.0-RC15
• Model: qwen3:0.6b
• Hardware: NVIDIA RTX 3090 (24Gb VRAM)
• Software: Debian 13 ; Cuda 13.2

Performance Comparison:

Version 0.23.4 (Baseline):

• Load duration: 965.56 ms
• Prompt eval rate: 1326.78 tokens/s
• Eval rate (Generation): ~467.7 tokens/s (268 tokens / 572.95 ms)
• Total duration: 1.632 s

Version 0.30.0-RC15 (Current):

• Load duration: 1733.69 ms (+79% increase)
• Prompt eval rate: 322.22 tokens/s (-75.7% decrease)
• Eval rate (Generation): ~474.18 tokens/s (Stable/Slight increase)
• Total duration: 2.284 s

Summary of Regression:

Analysis/Observations:

The regression is specifically located in the prompt processing (prefill) phase. The system takes much longer to process the input tokens. This could be due to changes in how the KV cache is allocated, changes in the graph computation in llama.cpp integration, or new overheads in the prompt parsing logic in the newer Ollama release candidate.

The fact that the generation speed (eval rate) remains stable suggests that the kernel efficiency for decoding is intact, but the "prefill" mechanism is heavily impacted.

Relevant log output

OS

Linux

GPU

AMD, Nvidia

CPU

AMD

Ollama version

0.30.0-RC15