LOCAL LLM BENCHMARK
========================================================================
ID:              20260822-110735
Timestamp:       2026-08-22T11:07:35-07:00
Model:           qwen3-coder:30b
Context:         4096
Output tokens:   256
CPU threads:     8
Measured runs:   3

BENCHMARK SESSION
------------------------------------------------------------------------
Pre-scan:        0.054 s
Cold phase:      38.015 s
Warmup phase:    27.259 s
Measured phase:  77.967 s
Post-scan:       0.161 s
Benchmark phase: 143.743 s
Total session:   143.958 s

WARM PERFORMANCE
------------------------------------------------------------------------
Generation:      10.02 tok/s
Prompt eval:     344.17 tok/s
Average total:   25986.56 ms
Average wall:    25.990 s
Average load:    194.55 ms

COLD PERFORMANCE
------------------------------------------------------------------------
Generation:      9.44 tok/s
Prompt eval:     37.97 tok/s
Load time:       9621.72 ms
Total:           38008.17 ms
Wall:            38.015 s

WARMUP DETAILS
------------------------------------------------------------------------
Warmup 1: gen=9.57 tok/s | prompt=333.23 tok/s | wall=27.259s

RUN DETAILS
------------------------------------------------------------------------
Run 1: gen=9.80 tok/s | prompt=333.26 tok/s | wall=26.545s
Run 2: gen=10.18 tok/s | prompt=343.50 tok/s | wall=25.583s
Run 3: gen=10.08 tok/s | prompt=355.75 tok/s | wall=25.838s

RECOMMENDATIONS
------------------------------------------------------------------------
[HIGH] Swap is already in use
  1.39 GB of swap is currently used. LLM inference can become dramatically slower when model or KV-cache pages spill out of RAM. Reduce context size, close memory-heavy applications, or choose a smaller/stronger quantization.