UPSTREAM PR #18137: common : restore grammar-based rejection sampling #607
LOCI Review / Performance Review #607
succeeded
Dec 17, 2025 in 38m 54s
Performance varied across binaries, overall acceptable
3 binaries improved · 9 binaries unchanged · 4 binaries stable ~ within threshold · 0 binaries degraded ~ beyond threshold
| Binary | Δ % Response | Δ % Throughput | Performance (based on response time) |
|---|---|---|---|
build.bin.libggml-base.so |
0 | 0 | unchanged |
build.bin.libggml-cpu.so |
0 | 0 | unchanged |
build.bin.libggml.so |
0 | 0 | unchanged |
build.bin.libllama.so |
-0.03 | 0 | improved |
build.bin.libmtmd.so |
0 | 0 | unchanged |
build.bin.llama-bench |
0.12 | 0.06 | stable |
build.bin.llama-cvector-generator |
-0.64 | 0.05 | improved |
build.bin.llama-gemma3-cli |
0 | 0 | unchanged |
build.bin.llama-gguf-split |
0.26 | 0.04 | stable |
build.bin.llama-llava-cli |
0 | 0 | unchanged |
build.bin.llama-minicpmv-cli |
0 | 0 | unchanged |
build.bin.llama-quantize |
0.26 | 0.1 | stable |
build.bin.llama-qwen2vl-cli |
0 | 0 | unchanged |
build.bin.llama-run |
0 | 0.07 | unchanged |
build.bin.llama-tokenize |
0.26 | 0.09 | stable |
build.bin.llama-tts |
-0.07 | 0.01 | improved |
Performance threshold: 30%
Default configuration used.
Note: Performance status is evaluated only from Δ% Response. Throughput is displayed for reference.
Explore the complete analysis inside the Version Insights.
Loading