flint-iq8r / README.md
StillDeadcode's picture
Upload README.md with huggingface_hub
77f74a3 verified
|
Raw History Blame Contribute Delete
1.84 kB
metadata
license: apache-2.0
base_model: RecursiveSI/flint
base_model_relation: quantized
pipeline_tag: text-generation
tags:
  - radiance

Recursive Flint, IQ8R, for radiance

flint-iq8r.rad (37.9 GiB) is RecursiveSI/flint converted for the radiance inference engine, version 2.1.0 or later. It is a quick conversion for trying the model: round to nearest with a scale search, no calibration data.

What is in it

  • IQ8R trunk: the attention projections, the two dense layers and the shared expert's gate/up as int8 codes with one u8 scale per (row, 256 columns) under one bf16 scale a row, each 128 columns rotated by a Walsh-Hadamard transform before rounding.
  • The 256 routed experts of every layer and the shared expert's down projection as int8 on the same rotated grid, one bf16 scale per (row, 128 columns).
  • The lm_head as plain int8 with a bf16 scale per (row, 128 columns).
  • The router, the norms, the attention sinks and the embedding in bf16.

Converted with the recipe the engine ships:

rad-convert <flint checkpoint> --add data/flint-zero-bias.safetensors \
    --recipe data/recipes/flint-iq8r.recipe \
    --chat-template data/flint-chat-template.jinja -o flint-iq8r.rad

Serving

Two 32 GB cards:

models:
  flint:
    model: /models/flint-iq8r.rad
    tp: 2
    kv-cache-dtype: i8

What was checked

On two Radeon AI PRO R9700 at tp: 2 with the int8 KV cache: the container loads, answers short factual and coding prompts correctly, finds a string hidden in a 4.5k-token prompt, and decodes at 6.8 ms a step for one stream (no drafter). Not measured: KL divergence against the bf16 model, and any benchmark.

License

Apache-2.0, as the source model; its LICENSE file is included.