jakeatx's picture
Update dataset card
13e59c0 verified
|
Raw History Blame Contribute Delete
9 kB
metadata
license: other
pretty_name: llamAmpere eval caches (Swift 1.5, SJ-KVaRN 4/4)
tags:
  - llama.cpp
  - llamAmpere
  - long-context
  - prefill-cache

llamAmpere eval caches for Swift 1.5 (IQ4_XS-M) at SJ-KVaRN 4/4

Why use this

About 6-7x faster long-context evals with everything on: eval cache + llamAmpere's MTP drafter + --parallel 2. For a 100K-token Qwen3.8-27B (Swift 1.5) prompt on one RTX 3090 Ti with a typical 2,000-token answer, that is about 153 s down to about 22 s; about 17x for 500-token answers and about 3x for 10,000-token answers. Prompt loading alone: 1.65 s vs 105.1 s fresh. This is arithmetic from measured parts, not one end-to-end run (release notes: https://github.com/JakeATX/llamAmpere/blob/main/docs/llamampere-v0.5.5/RELEASE_NOTES.md).

Long-context benchmarks send the same huge prompts every run, and the model has to read ("prefill") each one before it writes a word. At 100K tokens that is minutes per prompt. This dataset holds the finished result of that reading step for several well-known benchmarks, so your server can load it from disk instead of recomputing it. It is made for the --eval-cache option of llamAmpere v0.5.5 (see tools/eval-cache/README.md in that release).

Every folder was built with llamAmpere commit c305c699f (the v0.5.5 release code; eval-cache format 2) on RTX 3090 cards. A cache only works with the exact model file and the exact settings below. A server with different settings refuses to start and says which setting differs. Do not change the model, -ctk, -ctv, the drafter defaults, or the cache format.

Run it (copy and paste)

llama-server \
  --hf-repo jakeatx/ATX-Swift-1.5-Qwen3.8-27B-Uncensored-IQ4_XS-M-GGUF \
  --hf-file ATX-Swift-1.5-Qwen3.8-27B-Uncensored-IQ4_XS-M.gguf \
  -c 262144 -ngl 99 -fa on -ctk sjkvarn4 -ctv sjkvarn4 -b 4096 -ub 1024 --parallel 1 \
  --eval-cache ./evalcache/swift15-iq4xs-m/sjkvarn-44/longcodeqa/32K

(--parallel 1 and -b/-ub match how the folders were built. Keeping them gives byte-identical output at temperature 0. The multi-token-prediction drafter and the vocabulary map are on by default and are part of the key.)

Download one folder only

huggingface-cli download jakeatx/llamampere-evalcache --repo-type dataset \
  --include "swift15-iq4xs-m/sjkvarn-44/longcodeqa/32K/*" --local-dir ./evalcache

Replace the path with any folder in the table below. Each folder is self-contained: eval-cache.json (index), entries/ (the saved states), manifest.json (model hash, flags, checksums) and build-server.log. Check a download with python3 tools/eval-cache/llama-eval-cache.py verify --dir <folder>.

Your prompts must match exactly

A cache entry is used only when your request starts with the same tokens. Use the prompt files in prompts/ (the exact lines the folders were built from): one user message, no system message, reasoning_effort "medium", sent to /v1/chat/completions. Lines sharing a group share one saved prefix. Where a tier is split into parts, the part files are in prompts/<benchmark>/parts/.

What is in it (measured)

Benchmark Folder Entries Questions served Tokens cached Size (GB) Build GPU-minutes KB per token Break-even download speed (MB/s) Identity check
AA-LCR swift15-iq4xs-m/sjkvarn-44/aa-lcr/all30 30 100 3,103,744 107.7 70.3 34.7 26 2/2 byte-identical at temperature 0 (downloaded back from HF, checksums OK); request time 24.3 s with cache vs 289.2 s without
LongCodeQA swift15-iq4xs-m/sjkvarn-44/longcodeqa/32K 27 113 427,008 22.1 8.5 51.7 44 2/2 byte-identical at temperature 0 (downloaded from HF); request time 9.5 s with cache vs 28.2 s without
LongCodeQA swift15-iq4xs-m/sjkvarn-44/longcodeqa/64K 16 76 798,720 30.7 15.1 38.5 34 not run on this part (file sizes and SHA-256 checked at upload; box verify passed)
LongCodeQA swift15-iq4xs-m/sjkvarn-44/longcodeqa/128K 20 92 1,985,536 69.1 43.0 34.8 27 not run on this part (file sizes and SHA-256 checked at upload; box verify passed)
MRCR swift15-iq4xs-m/sjkvarn-44/mrcr/8needle-64k-128k-part-a 19 19 1,720,320 62.8 32.7 36.5 32 2/2 byte-identical at temperature 0 (downloaded back from HF, checksums OK); request time 8.8 s with cache vs 279.7 s without
MRCR swift15-iq4xs-m/sjkvarn-44/mrcr/8needle-64k-128k-part-b 14 14 1,604,608 57.0 33.1 35.5 29 not run on this part (file sizes and SHA-256 checked at upload; box verify passed)
MRCR swift15-iq4xs-m/sjkvarn-44/mrcr/8needle-64k-128k-part-c 17 17 1,634,304 59.3 32.1 36.3 31 not run on this part (file sizes and SHA-256 checked at upload; box verify passed)
RULER swift15-iq4xs-m/sjkvarn-44/ruler/128k-part-a 10 10 1,300,480 46.4 30.9 35.7 25 2/2 byte-identical at temperature 0 (downloaded back from HF, checksums OK); request time 15.7 s with cache vs 514.7 s without
RULER swift15-iq4xs-m/sjkvarn-44/ruler/128k-part-b 10 10 1,294,336 45.8 30.1 35.3 25 not run on this part (file sizes and SHA-256 checked at upload; box verify passed)
RULER swift15-iq4xs-m/sjkvarn-44/ruler/128k-part-c 10 10 1,300,480 46.4 29.9 35.7 26 not run on this part (file sizes and SHA-256 checked at upload; box verify passed)
RULER swift15-iq4xs-m/sjkvarn-44/ruler/128k-part-d 9 9 1,138,688 40.4 26.2 35.5 26 not run on this part (file sizes and SHA-256 checked at upload; box verify passed)
LongBench v2 swift15-iq4xs-m/sjkvarn-44/longbench-v2/196k-245k-part-a 7 7 1,539,072 51.7 68.5 33.6 13 2/2 byte-identical at temperature 0 (downloaded back from HF, checksums OK); request time 19.3 s with cache vs 1223.9 s without
LongBench v2 swift15-iq4xs-m/sjkvarn-44/longbench-v2/196k-245k-part-b 5 5 1,105,920 37.3 33.0 33.7 19 1/1 byte-identical at temperature 0 on the build box before upload (1 prompt, not re-run after download); request time 9.5 s with cache vs 427.5 s without
LongBench v2 swift15-iq4xs-m/sjkvarn-44/longbench-v2/196k-245k-part-c 5 5 1,120,256 37.6 34.1 33.5 18 1/1 byte-identical at temperature 0 on the build box before upload (1 prompt, not re-run after download); request time 10.8 s with cache vs 432.4 s without
LongBench v2 swift15-iq4xs-m/sjkvarn-44/longbench-v2/196k-245k-part-d 5 5 1,130,496 38.0 34.2 33.6 19 1/1 byte-identical at temperature 0 on the build box before upload (1 prompt, not re-run after download); request time 7.7 s with cache vs 410.6 s without
LongBench v2 swift15-iq4xs-m/sjkvarn-44/longbench-v2/196k-245k-part-e 5 5 1,084,416 36.7 31.8 33.9 19 1/1 byte-identical at temperature 0 on the build box before upload (1 prompt); downloaded back from HF and checksums verified (5 entries, 0 problems); request time 6.6 s with cache vs 312.1 s without

Total published: 209 entries, 497 questions, 22,288,384 tokens cached, 789 GB, 553 build GPU-minutes.

Is downloading faster than recomputing?

An entry stores the model state for the cached prefix: the size per cached token is in the table (it is larger for shorter prefixes). The break-even column multiplies that size by the prefill speed seen while building (tokens cached divided by build time on an RTX 3090). A download beats recomputing only when your link is faster than that. Below that speed, recomputing is quicker. Once a folder is on a local disk, a restore from NVMe is far faster than a prefill. These caches pay off for repeated runs of the same eval, or for a fast link.

Sources and terms

Benchmark Source Terms
AA-LCR ArtificialAnalysis/AA-LCR v1.1 (revision 9a77ef56b717), documents as extracted text Questions Apache-2.0; documents are text of publicly available documents, "We do not claim copyright or place any license over this data."
LongCodeQA Steefano/LCB (revision 989d5eff750d), prompt field used verbatim Dataset Apache-2.0; each repository inside is under its own GitHub license
MRCR openai/mrcr (revision f4c69fae7cf8), 8-needle rows MIT
RULER NVIDIA/RULER commit c3f5e3b4, --random_seed 42 --chunk_amount 1 --model_template_type base --prepare_for_ns --max_seq_length 131072 --num_samples 3 Code Apache-2.0. Prompt text is NOT published here (essay haystack is not redistributable; QA tasks are CC BY-SA). Only prompts/ruler/rendered_prompt_hashes.json (SHA-256 of each user message) and the render scripts in render_scripts/ are provided, so you can regenerate the prompts and check them against the hashes.
LongBench v2 zai-org/LongBench-v2 (revision 2b48e494f2c7), THUDM 0shot template Dataset Apache-2.0, code MIT

The cached states are derived from the benchmark texts above and from the model; they follow the terms of the sources.