Skip to content

bench: existing RoPE vs YaRN at 512K and 1M, with setup guide - #1692

Closed
CC-David-CC wants to merge 2 commits into
Niko1221:mainfrom
CC-David-CC:bench/rope-yarn4-long-context
Closed

CC-David-CC wants to merge 2 commits into
Niko1221:mainfrom
CC-David-CC:bench/rope-yarn4-long-context

Conversation

@CC-David-CC

Copy link
Copy Markdown
Contributor

Recall and prefill comparison

Summary

Measures Strata's existing YaRN on untouched main fb58e0db; no inference changes or converter dependency.

ISTA IQ3_XXS, FP16 KV, RTX PRO 6000 Blackwell. One fresh three-code retrieval prompt per condition:

Configuration Correct pairs Prefill Decode
Ordinary RoPE, 512K 3/3 95.7 s 103.6 tok/s
Ordinary RoPE, 1M 0/3 235.5 s 82.4 tok/s
YaRN 4x, 1M 3/3 232.8 s 82.7 tok/s

The 1M input hashes and executable match. Ordinary RoPE retained the names/numbers but swapped their associations; exact answers are shown in the figure. Replies were only 22 tokens.

Use existing YaRN

Linux: ./setup.sh --setup --context 1048576 --rope-scaling yarn --rope-scale 4

Windows: START-HERE.bat --setup --context 1048576 --rope-scaling yarn --rope-scale 4

Stop the running server first. --setup saves these settings and starts the selected model; later launches use ./setup.sh or START-HERE.bat.

Includes engine flags, a 128-token API example, output headroom and cache compatibility guidance in the setup and reproduction guide.

Scope

This is a small reproducible probe, not a validated 512K switching cutoff or general 1M quality guarantee. Retain normal defaults within the declared 262K context; beyond that, follow setup's existing experimental YaRN policy. Fresh YaRN is not cache migration.

Adds raw measurements/hashes, the standalone probe, PNG/SVG figure and plot source. No weights, saved sessions, or credentials. Related measurements: #348 and #781.

@Niko1221

Copy link
Copy Markdown
Owner

Thank you, in v0.1.42 with your two commits, with a row in bench/results/COMMUNITY.md. The 1M ordinary-RoPE result (0/3 pairs) against YaRN (3/3) is a clear data point.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants