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
224 lines
4.8 KiB
JSON
224 lines
4.8 KiB
JSON
{
|
|
"tag": "bfloat16/None/eager=False",
|
|
"wall_step_ms": 14.345,
|
|
"tps": 69.71,
|
|
"profiled_total_ms_per_step": 13.87,
|
|
"buckets_ms_per_step": {
|
|
"gemm": 8.263,
|
|
"framework": 3.017,
|
|
"other_op": 1.581,
|
|
"attn": 1.009
|
|
},
|
|
"top": [
|
|
{
|
|
"op": "aten::_weight_int4pack_mm_for_cpu",
|
|
"bucket": "gemm",
|
|
"ms_per_step": 7.7116,
|
|
"calls_per_step": 113.0
|
|
},
|
|
{
|
|
"op": "## Call CompiledFxGraph None ##",
|
|
"bucket": "framework",
|
|
"ms_per_step": 2.2035,
|
|
"calls_per_step": 1.0
|
|
},
|
|
{
|
|
"op": "vllm::cpu_int4_linear",
|
|
"bucket": "gemm",
|
|
"ms_per_step": 0.5513,
|
|
"calls_per_step": 113.0
|
|
},
|
|
{
|
|
"op": "vllm::unified_attention_with_output",
|
|
"bucket": "attn",
|
|
"ms_per_step": 0.4727,
|
|
"calls_per_step": 28.0
|
|
},
|
|
{
|
|
"op": "vllm::unified_kv_cache_update",
|
|
"bucket": "framework",
|
|
"ms_per_step": 0.4491,
|
|
"calls_per_step": 28.0
|
|
},
|
|
{
|
|
"op": "_C::cpu_attention_with_kv_cache",
|
|
"bucket": "attn",
|
|
"ms_per_step": 0.3842,
|
|
"calls_per_step": 28.0
|
|
},
|
|
{
|
|
"op": "aten::copy_",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.2688,
|
|
"calls_per_step": 75.0
|
|
},
|
|
{
|
|
"op": "aten::argmax",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.2479,
|
|
"calls_per_step": 1.0
|
|
},
|
|
{
|
|
"op": "vllm::spark_head_gate",
|
|
"bucket": "framework",
|
|
"ms_per_step": 0.2182,
|
|
"calls_per_step": 28.0
|
|
},
|
|
{
|
|
"op": "_C::cpu_attn_reshape_and_cache",
|
|
"bucket": "attn",
|
|
"ms_per_step": 0.1523,
|
|
"calls_per_step": 28.0
|
|
},
|
|
{
|
|
"op": "aten::slice",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.152,
|
|
"calls_per_step": 230.0
|
|
},
|
|
{
|
|
"op": "Pregraph bytecode",
|
|
"bucket": "framework",
|
|
"ms_per_step": 0.1399,
|
|
"calls_per_step": 1.0
|
|
},
|
|
{
|
|
"op": "aten::mul",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.1302,
|
|
"calls_per_step": 28.0
|
|
},
|
|
{
|
|
"op": "aten::sum",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0956,
|
|
"calls_per_step": 28.0
|
|
},
|
|
{
|
|
"op": "aten::as_strided",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0952,
|
|
"calls_per_step": 290.0
|
|
},
|
|
{
|
|
"op": "aten::narrow",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0844,
|
|
"calls_per_step": 112.0
|
|
},
|
|
{
|
|
"op": "aten::_to_copy",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0794,
|
|
"calls_per_step": 61.0
|
|
},
|
|
{
|
|
"op": "aten::chunk",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0631,
|
|
"calls_per_step": 56.0
|
|
},
|
|
{
|
|
"op": "aten::split",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0547,
|
|
"calls_per_step": 56.0
|
|
},
|
|
{
|
|
"op": "aten::view",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0531,
|
|
"calls_per_step": 62.0
|
|
},
|
|
{
|
|
"op": "aten::empty",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0456,
|
|
"calls_per_step": 145.0
|
|
},
|
|
{
|
|
"op": "aten::to",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0335,
|
|
"calls_per_step": 64.0
|
|
},
|
|
{
|
|
"op": "aten::empty_strided",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0316,
|
|
"calls_per_step": 61.0
|
|
},
|
|
{
|
|
"op": "aten::fill_",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0244,
|
|
"calls_per_step": 40.0
|
|
},
|
|
{
|
|
"op": "aten::index",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0217,
|
|
"calls_per_step": 2.0
|
|
},
|
|
{
|
|
"op": "aten::unsqueeze",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0184,
|
|
"calls_per_step": 29.0
|
|
},
|
|
{
|
|
"op": "_C::compute_slot_mapping_kernel_impl",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.018,
|
|
"calls_per_step": 4.0
|
|
},
|
|
{
|
|
"op": "_C::get_scheduler_metadata",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0161,
|
|
"calls_per_step": 4.0
|
|
},
|
|
{
|
|
"op": "aten::new_empty",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0149,
|
|
"calls_per_step": 28.0
|
|
},
|
|
{
|
|
"op": "aten::add",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0095,
|
|
"calls_per_step": 3.0
|
|
},
|
|
{
|
|
"op": "AOTDispatcher Runtime Wrapper Prologue",
|
|
"bucket": "framework",
|
|
"ms_per_step": 0.0063,
|
|
"calls_per_step": 1.0
|
|
},
|
|
{
|
|
"op": "aten::sub",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.005,
|
|
"calls_per_step": 1.0
|
|
},
|
|
{
|
|
"op": "aten::lt",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.004,
|
|
"calls_per_step": 1.0
|
|
},
|
|
{
|
|
"op": "aten::index_select",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0034,
|
|
"calls_per_step": 1.0
|
|
},
|
|
{
|
|
"op": "aten::reshape",
|
|
"bucket": "other_op",
|
|
"ms_per_step": 0.0024,
|
|
"calls_per_step": 4.0
|
|
}
|
|
]
|
|
} |