Repository navigation
Merged
Conversation
The card's MEMORY block (GPU used of total, host RSS) never reached the site: the index carried no memory, so a model larger than its card read the same on the page as one that fit. The index now copies gpu_mem and the RSS figures from the tape with memorySection's gates, and the page prints them in the card's GiB rounding. No D1 column: the page reads index_json, and a reindex repairs runs already published. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Under tape.MinPrefillPromptTokens the server's prefill rate is mostly the
fixed per-request cost; the card says so, the site printed it as a plain
figure on the run page, in the description and in the search. The index
now marks short_prompt (card.ShortPrompt, one owner), leaves the rate out
so the search column stays NULL, and carries the probe's own rate and
fixed cost. The run page shows the probe in the stat and a prefill row
("not measured: 33-token prompt · probe … on one stream · … fixed"), and
the description quotes the probe instead. The two Go-rounding helpers in
run.js are one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The site page showed less than the tape. Two gaps:
gpu_memandrss_*from the tape withmemorySection's gates; the page prints them in the card's GiB rounding.tape.MinPrefillPromptTokens(100) the prefill rate is mostly fixed cost. The card warns; the page, the description and the search printed it as a plain figure. The index now marksshort_prompt(card.ShortPrompt), omitsprefill_per_sec(so the search column is NULL), and carries the probe's rate and fixed cost. The page shows the probe instead.No D1 migration: the page reads
index_json. Published runs change after atoktape reindex.Evidence: FAIL-first for each new Go and JS test (Go build failed /
short_prompt = <nil>, prefill_per_sec = 206; JS 0 of 1 and 2 of 7 passed). After:./scripts/check.shOK,npm test49/49.Example runs: https://tape.midagedev.com/r/3xzgnp3srbrwzfruga4v (21.2 GB model on a 12 GB RTX 3060), https://tape.midagedev.com/r/bi7v3b74snetzbq45dsz (33-token prompt).
The implementation was drafted by an AI model (grok) and reviewed by the bloomery AI assistant.
🤖 Generated with Claude Code