Skip to content

UPSTREAM PR #17579: Add PagedAttention support (experimental, CUDA only) - #352

Open
loci-dev wants to merge 1 commit into
mainfrom
upstream-PR17579-branch_ericcurtin-add-pagedattention
Open

UPSTREAM PR #17579: Add PagedAttention support (experimental, CUDA only)#352
loci-dev wants to merge 1 commit into
mainfrom
upstream-PR17579-branch_ericcurtin-add-pagedattention

Add PagedAttention support (experimental, CUDA only)

19466fb
Select commit
Loading
Failed to load commit list.
LOCI Review / Performance Review #352 succeeded Dec 1, 2025 in 28m 22s

Performance varied across binaries, overall acceptable

2 binaries improved · 6 binaries unchanged · 8 binaries stable ~ within threshold · 0 binaries degraded ~ beyond threshold

Binary Δ % Response Δ % Throughput Performance (based on response time)
build.bin.libggml-base.so 0.3 0.13 stable
build.bin.libggml-cpu.so 0.09 -0.15 stable
build.bin.libggml.so 0 0 unchanged
build.bin.libllama.so 0.52 2.57 stable
build.bin.libmtmd.so 0 0 unchanged
build.bin.llama-bench 0.01 0.04 stable
build.bin.llama-cvector-generator 0.27 -0.38 stable
build.bin.llama-gemma3-cli 0 0 unchanged
build.bin.llama-gguf-split -0.06 -0.08 improved
build.bin.llama-llava-cli 0 0 unchanged
build.bin.llama-minicpmv-cli 0 0 unchanged
build.bin.llama-quantize 0.01 0.07 stable
build.bin.llama-qwen2vl-cli 0 0 unchanged
build.bin.llama-run -0.37 -0.1 improved
build.bin.llama-tokenize 0.01 -0.02 stable
build.bin.llama-tts 0.39 -0.19 stable

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.