Skip to content

Count files that leave the lookup table out in full - #321

Merged
Andyyyy64 merged 1 commit into
v0.6.10from
fix/lookup-table-files
Nov 19, 2026
Merged

Andyyyy64 merged 1 commit into
v0.6.10from
fix/lookup-table-files

Conversation

@Andyyyy64

Copy link
Copy Markdown
Owner

Qwen3.8-Flash-Next's n-gram table (51.2B parameters, 28.8 GB at IQ4_NL) stays on disk, so whichllm takes it out of a file's resident size. It took it out of every file, up to 90% of the file, but some repos ship the table apart from the model.

  • agentionai's Gyro repos publish -noPLE builds that leave the table out, and a file that holds the table alone, Qwen3.8-Flash-Next-PLE-IQ4_NL-GyroS.gguf. The index picked the -noPLE builds, so Gyro-S's 29.7 GB file needed 3.0 GB of VRAM and Gyro-M's 37.6 GB file 8.8 GB. The table file ranked as an IQ4_NL build of Gyro with 2.9 GB resident, ninth on a 16 GB RTX 5080 with 32 GB of RAM at 888 tok/s.
  • CYBER-FROST has the table as the last part of a split file in an ngram/ folder, and Tokspur and a DeepSeek V4.1 CoderPlus build had 90% of their files taken out the same way.

What changes

  • lazy_lookup_bytes takes nothing out of a file named noPLE, or out of a file the table at 4.5 bits would fill past 90%. Such a file stays resident whole. Before, the table came out up to that 90%.
  • The index skips a file of the table alone, named -PLE- or in an ngram/ folder, the way it skips DeepSeek V4.1's Engram file.

Across today's index this changes 12 of the 90 variants of models with a lazy table, all from the repos above. The full Qwen3.8-Flash-Next and DeepSeek V4.1 repos keep their sizes. The index change reaches every version once main has it, since the index is built from main.

Left for later

  • Gyro's files use agentionai's own quant types, which upstream llama.cpp cannot load yet. whichllm reads them as TQ1_0 to TQ3_0 from the filenames and names llama.cpp for them.
  • Gyro's parameter count already leaves the table out, and the table comes out of it again for the active share, so the speed estimate for its rows reads about 1.7 times too much per token.

Tests

Five new tests cover the -noPLE files, a file too small to hold the table and the table files in a repo listing. They fail on the current code. All 2,046 tests pass, and ruff is clean.

Qwen3.8-Flash-Next's n-gram table came out of every file, up to 90% of
it. agentionai's Gyro -noPLE builds leave the table out, so Gyro-S's
29.7 GB file needed 3.0 GB, and the file of the table alone ranked as a
model at 888 tok/s on a 16 GB GPU. A file named noPLE, or one the table
would nearly fill, now stays resident whole, and the index skips a file
of the table alone.
@Andyyyy64
Andyyyy64 merged commit e046a03 into v0.6.10 Nov 19, 2026
6 checks passed
@Andyyyy64
Andyyyy64 deleted the fix/lookup-table-files branch November 19, 2026 13:41
Andyyyy64 added a commit that referenced this pull request Nov 19, 2026
Bring #321 into main before 0.6.10
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.

1 participant