--- license: mit base_model: inclusionAI/Ling-3.0-tiny base_model_relation: quantized pipeline_tag: text-generation tags: - radiance --- # Ling-3.0-tiny, IQ8R, for radiance `ling-3.0-tiny-iq8r.rad` (7.9 GiB) is [inclusionAI/Ling-3.0-tiny](https://huggingface.co/inclusionAI/Ling-3.0-tiny) converted for the [radiance](https://codeberg.org/StillDeadcode/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 KDA projections (fused q|k|v, forget, gate, output), the latent attention's output, the dense layer's feed-forward 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 128 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). - In bf16: the KDA beta projection, the latent attention's low-rank projections and gate, the convolutions, the router, the embedding and the lm_head. Converted with the recipe the engine ships: ``` rad-convert --recipe data/recipes/ling-3.0-tiny-iq8r.recipe \ -o ling-3.0-tiny-iq8r.rad ``` ## Serving One card, or two: ```yaml models: ling-3.0-tiny: model: /models/ling-3.0-tiny-iq8r.rad tp: 1 ``` ## What was checked On Radeon AI PRO R9700 cards: the container loads at `tp: 1` and `tp: 2`, answers short factual and coding prompts correctly, finds a string hidden in a 4.8k-token prompt, and decodes one stream at 4.3 ms a step on one card (230 tokens/s) and 3.7 ms on two (270 tokens/s). The model has no draft head. Not measured: KL divergence against the bf16 model, and any benchmark. ## License MIT, as the source model.