Commit History
-
Replace legacy exports with optimized W4A32 RTN g32 ONNX
Verified signaturePublish FP32 graph activations/KV cache, symmetric W4 RTN group-size 32 backbone, GroupQueryAttention shared KV buffers, and SkipLayerNormalization fusions. Embeddings, conditioner, and LM head remain FP32.7d1a2cc
-
Add model card, licenses, configs, and ONNX Runtime helper
Verified signatureee4cb7d