tzhouamandClaude Opus 5
9264ff81ab
The packaging commit shipped the code and a summary of the numbers. It left out
everything the numbers are measured *against*, which is most of what makes them
checkable.
Added:
baselines/ llama.cpp runs (thread sweeps, NUMA sweeps, the
interleaved A/B passes behind the headline decode
figure) and the quality references: each build
against its own bf16 over 12 greedy prompts.
benchmarks/.../results/ 102 artifacts backing the published figures, each
with a provenance block naming the model,
quantization, thread set and build SHAs.
docs/engines/ the engine analyses -- llama.cpp, ExecuTorch, MNN,
ncnn, vLLM, vLLM-Omni -- and the cross-engine and
device-simulation comparisons.
docs/analysis/ the full study documents the summaries distil.
docs/edge/comparisons.md where this fork sits against llama.cpp and why:
decode is llama.cpp's, prefill is ours by ~2x, and
since neither engine exceeds 39% of the measured
222 GB/s ceiling the decode gap is the execution
model rather than the kernels.
docs/edge/environment.md pins and machine, and the correction below.
On "the modified vllm": there isn't one, and the doc says so plainly because it
is the obvious assumption. The vLLM checkout is clean upstream (0 commits
ahead, 0 uncommitted) and the installed wheel is stock 0.28.0+cpu. This fork
changes vLLM's CPU *behaviour* through vLLM's own extension points at import
time -- three CPU patches in vllm_omni/patch.py, the quantization-config
registry, the IR op registry, platform hooks -- specifically so the changes
survive a wheel upgrade rather than being lost on it. One of those patches is
why "KV sizing is not a memory lever" looked true for months: CpuPlatform
overwrote kv_cache_memory_bytes unconditionally, so the setting had no effect.
There is no patch series to apply; pin the version and let patch.py do the rest.
.gitignore gains explicit exceptions: upstream's blanket `*.json` and `results/`
rules were silently eating the measurement artifacts, which in this fork are the
substance rather than build output. Binary captures stay ignored.
No vLLM or llama.cpp source is redistributed -- only output produced with them.
NOTICE records both.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L1ax5wYLA6XbqEL1i8L9gZ
1.4 KiB
Baselines
Measurements of other engines and of unquantized references, kept so the comparisons in docs/edge/comparisons.md can be checked rather than taken on trust.
llama_cpp/
llama.cpp is the reference for CPU inference of a quantized model. Pinned commit and build flags are in docs/edge/environment.md; no llama.cpp source is redistributed here, only output produced with it.
| file | what |
|---|---|
llamacpp_q4km_t{4,8,16,32}.json |
Q4_K_M across thread counts |
llamacpp_grid_t*.json, llamacpp_node1_t*.json |
thread and NUMA-node sweeps |
ab_llamacpp_*.json |
interleaved A/B against vLLM on the same cores |
il_llamacpp_p{1,2,3}.json |
the interleaved passes behind the headline decode figure |
cmp_llamacpp.py |
the comparison driver |
Both engines are pinned with taskset to the same physical cores and run
strictly one at a time; vLLM's prefix caching is disabled, because llama-bench
has no such cache and leaving it on would not be a comparison.
Quality references
quality_*.json and quality_cmp_*.json hold the greedy-divergence comparison:
each build against its own bf16 output over 12 prompts, plus the
cross-engine roll-up. Read the caveats in
docs/edge/comparisons.md before quoting these —
the bit budgets are not matched, and the metric saturates.