IlyasMoutawwakil/tiny-random-Mistral4-FP8-static

tiny-random-Mistral4-FP8-static

A tiny random model for testing, shrunk from `mistralai/Mistral-Small-4-119B-2603`: the same architecture, quantization config and checkpoint layout at test sizes. Its key patterns, dtypes and tensor ranks match the real checkpoint's (scripts/extract_layout.py).

static per-tensor FP8 on every decoder linear and the stacked experts: e4m3, BF16 weight_scale_inv (0-dim, (E, 1, 1) on the experts) and calibrated BF16 activation_scale (0-dim, (E,) on the experts); the MLA attention and shared experts included, router, vision tower and projector bf16.

reference/ holds the same weights dequantized to bf16, under the unquantized model's keys: the reference to compare logits against, so a test measures what the load path and kernels add, not the quantization itself.

The weights are random; the outputs mean nothing. scripts/ rebuilds it from the real checkpoint's config.json.