WaveCut/Edge0-ngram-table-FP8-NInfer-v3

Edge0 n-gram table, NInfer v3 artifact

The n-gram table of the per-layer embedding (PLE) of Edge0's models, as a NInfer v3 artifact for NInfer-all, the master branch of iamwavecut/ninfer-all. Every NInfer conversion of an Edge0 release published without its table reads its rows from this file:

model	artifact
Edge0-145B-NVFP4	WaveCut/Edge0-145B-NVFP4-NInfer-v3
Edge0-145B-q3k	WaveCut/Edge0-145B-q3k-NInfer-v3
Edge0-120B-NVFP4	WaveCut/Edge0-120B-NVFP4-NInfer-v3
Edge0-120B-q3k	WaveCut/Edge0-120B-q3k-NInfer-v3

Stock NInfer builds cannot read it; NInfer-all reads it from commit 72d1d8e04 on.

What is inside

The table holds 320,001,536 rows of 160 values (51.2B parameters), 16 hash heads over the 2- and 3-token n-grams of the context; every token reads the 16 rows its n-grams address. The rows are the FP8 E4M3 codes of the releases' ple-fp8.gguf (the same file in all four), copied byte for byte, as fp8_e4m3fn_row_fp16 rows: 160 codes and the release's global scale, 209 × 2^-20, which is exact in FP16.

One file, Edge0-ngram-table-FP8-ninfer-v3.ninfer, of 51,945,146,104 bytes (48.38 GiB), next to its conversion report, SHA256SUMS, and Edge0-120B-q3k's LICENSE, NOTICE, release-scope.json and third-party/Qwen-LICENSE.txt. Its ngram component records the hash constants the rows were written for, their format and the SHA-256 of the rows, 57496326706760c564e80ed401f6b26a7d6b340cc05966502675c8dc01ef8fd9. A model names the table it needs by those. The file also carries the broad hot-row profile of the IQ4_NL table as ngram.resources.hot_profile: 26,223,284 rows ranked by use, counted from 3,711,509 tokens of text for the same hash. It was added on 2026-11-13 without changing the rows or their SHA-256, so every model that names this table reads the new file. NInfer-all from 57395b20b on reads any table of a model's hash constants and geometry, in any row format, and warns when it is not the one the model names: Flash-Next's models read this table in place of their IQ4_NL table (the GSQ-RCO Q2_0 release: perplexity 2.531971 with it, 2.542153 with its own), and Edge0's models read that IQ4_NL table in place of this one. Source: revision 51a60a20 of Edge0/Edge0-120B-q3k. Conversion command:

python3 -m tools.convert --model Qwen3.8-Flash-Next \
  --recipe edge0_pack --source pack=Edge0-120B-q3k \
  --components ngram --device cpu --name edge0-ngram-table-fp8 --out Edge0-ngram-table-FP8-ninfer-v3.ninfer

Running

Download it next to a model and pass it with --ngram-table:

hf download WaveCut/Edge0-ngram-table-FP8-NInfer-v3 Edge0-ngram-table-FP8-ninfer-v3.ninfer --local-dir models
hf download WaveCut/Edge0-120B-q3k-NInfer-v3 Edge0-120B-q3k-ninfer-v3.ninfer --local-dir models
ninfer-serve models/Edge0-120B-q3k-ninfer-v3.ninfer \
  --ngram-table models/Edge0-ngram-table-FP8-ninfer-v3.ninfer --model-id edge0-120b --max-context 32768

NInfer reads 16 rows a token from the file; --ngram-residency ram loads all of it into RAM, and --ngram-residency ram-hot keeps the rows the embedded profile ranks first in RAM (as many as --ngram-ram-mib holds, 4096 MiB by default) and reads the rest from the file. On an A40 the GSQ-RCO Q2_0 model of Qwen3.8-Flash-Next, bound to this table, scored the first 40 KB of WikiText (9,234 tokens) at perplexity 2.531971 with ram-hot and with the default disk.

Changes from the release and license

WaveCut repacked the release's files into NInfer's container for NInfer-all, with the converter in tools/convert/edge0.py. The table's FP8 codes are stored byte for byte, with the release's global scale written as FP16 into every row; the hot-row profile is the IQ4_NL table's, stored byte for byte. No weight was retrained, and no other block was re-quantized.

The Edge0-specific contributions that the release's release-scope.json designates are under the Edge0 Community Model License 1.0, reproduced in LICENSE with Edge0's NOTICE. The Qwen materials stay under the Qwen Community License 1.0, reproduced unchanged in third-party/Qwen-LICENSE.txt; the Edge0 license does not replace it. Commercial model offerings need a separate license from Edge0 (its Section 4), and Qwen's conditions apply as well. These files imply no endorsement by Edge0 or Qwen.