tiny Mistral4 with fp8 + nvfp4 layers

ac53c7a verified 2 days ago

259 Bytes

---
library_name: transformers
---
Tiny random Mistral4, pre-quantized with a per-layer plan: fp8 (block 16x16) on attention and routed experts, nvfp4 on shared experts. `dense_reference.pt` holds the bf16 weights before quantization; `make.py` rebuilds it.

	---
	library_name: transformers
	---

	Tiny random Mistral4, pre-quantized with a per-layer plan: fp8 (block 16x16) on attention and routed experts, nvfp4 on shared experts. `dense_reference.pt` holds the bf16 weights before quantization; `make.py` rebuilds it.