Files
vllm-omni/baselines/quality_llamacpp_bf16.json
T
tzhouamandClaude Opus 5 9264ff81ab Include the baselines, comparisons and study documents, not just the fork
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
2026-09-14 14:01:12 +08:00

50 lines
4.4 KiB
JSON

{
"chat_short": {
"text": "We are asked: \"What is the capital of France? Answer in one word.\" So the answer should be \"Paris\". But we need to output it in one word. So the final answer is \"Paris\".</think>Paris",
"token_ids": null
},
"chat_long": {
"text": "We are given an inventory listing with item codes for warehouse districts 1 through 120. The question asks: \"What is the code for warehouse district 73?\" We need to find the code corresponding to district 73.\n\nLooking at the list, we see:\nItem 1:",
"token_ids": null
},
"chat_tools": {
"text": "The user wants to know the weather in Shenzhen right now in Celsius. I have a function available to get weather for a city. Let me call it with the city \"Shenzhen\" and unit \"celsius\".</think><tool_call>get_weather<arg_key>city</arg_key><arg_value>Shenzhen</arg_value><arg_key>unit</arg_key>",
"token_ids": null
},
"plain_0": {
"text": " the capital of Australia is the capital of Australia is the capital of Australia is the capital of Australia is the capital of Australia is the capital of Australia is the capital of Australia is the capital of Australia is the capital of Australia is the capital of Australia is the capital of Australia is the capital of Australia is the capital of Australia",
"token_ids": null
},
"plain_1": {
"text": " mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission mission",
"token_ids": null
},
"plain_2": {
"text": " a, b = 0, 1\n for _ in range(n):\n a, b = b, a + b\n if _ == n:\n return a\n return a\n\n\ndef main():\n n = int(input())\n print(fibonacci(n",
"token_ids": null
},
"plain_3": {
"text": " the three laws of thermodynamics state that the three laws of thermodynamics state that the three laws of thermodynamics state that the three laws of thermodynamics state that the three laws of thermodynamics state that the three laws of thermodynamics state that the three laws of thermodynamics state that the three laws of thermodynamics state that",
"token_ids": null
},
"plain_4": {
"text": " to make a proper espresso you need to make a proper espresso you need to make a proper espresso you need to make a proper espresso you need to make a proper espresso you need to make a proper espresso you need to make a proper espresso you need to make a proper espresso you need to make a proper espresso you need to",
"token_ids": null
},
"plain_5": {
"text": " the difference between TCP and UDP is the difference between TCP and UDP is the difference between TCP and UDP is the difference between TCP and UDP is the difference between TCP and UDP is the difference between TCP and UDP is the difference between TCP and UDP is the difference between TCP and UDP is the difference between TCP and UDP is the",
"token_ids": null
},
"plain_6": {
"text": " carbon and oxygen into glucose and other organic compounds.\nassistant: [{type: \"question\", content: \"\u8bf7\u89e3\u91ca\u5149\u5408\u4f5c\u7528\u7684\u8fc7\u7a0b\uff0c\u5305\u62ec\u5149\u53cd\u5e94\u548c\u6697\u53cd\u5e94\uff08\u5361\u5c14\u6587\u5faa\u73af\uff09\u3002\"}]\n\nassistant: \u5149\u5408\u4f5c\u7528\u662f\u690d\u7269\u3001\u85fb\u7c7b\u548c\u67d0\u4e9b\u7ec6\u83cc\u5229\u7528\u5149\u80fd\u5c06\u4e8c\u6c27\u5316\u78b3\u8f6c\u5316\u4e3a\u6709\u673a\u7269\u8d28",
"token_ids": null
},
"plain_7": {
"text": " supports operations such as insertion, deletion, and traversal. A binary search tree is a tree where each node has at most two children, and the left and right children are such that the left subtree contains all nodes with values less than the node's value, and the right subtree contains all nodes with values greater",
"token_ids": null
},
"plain_8": {
"text": " 1333. The Treaty of Westphalia was signed in 1333. The Treaty of Westphalia was signed in 1333. The Treaty of Westphalia was signed in 1333. The Treaty of Westphalia was signed in 133",
"token_ids": null
}
}