narendra747/lfm2.5-230m-nvfp4

LFM2.5-230M NVFP4

This is the deterministic NML NVFP4 recipe-v2 conversion of LiquidAI/LFM2.5-230M@40cb2ad3b3044d5a41eee083a6103c8b523afa45. Rank-2 embedding, attention, convolution-projection, and MLP weights use last-axis one-dimensional blocks of 16, low-nibble-first E2M1 payloads, positive E4M3FN block scales, and one F32 global factor per quantized parameter. RMSNorm parameters and depthwise convolution kernels remain BF16. The embedding and LM head share one physical NVFP4 parameter; the source checkpoint does not store a second lm_head.weight matrix.

Conversion is CPU-only; exact converter and dependency provenance is recorded in nml-artifact-manifest.json. Exact source hashes, tensor disposition, and conversion semantics are included in the repository. Recipe: nml-nvfp4-weight-v2.