silvann_tq_D8E4_glm_5.3_flash

GLM 5.3 Flash by Z.AI, packed for the SiLVANN engine (the Eval-Apply GPU Interpreter and its neural network library): every weight matrix rotated along its input axis with a blocked Hadamard-and-signs rotation and quantised with a per-row scale (TurboQuant); norms, routers, the hyper-connection maps and other small tensors kept lossless; the embedding and head at 8 bits. D8E4: the dense part (attention, the shared experts, the dense MLPs) at 8 bits, the routed experts at 4. 598.5 GiB of raw weights in 153.8 GiB.

The format is silvann-packed-v3 — pack.json describes the codec and the per-tensor width policy, layers/NNN/ holds each layer's records and global.pkl the embedding, head, final norm and the vision tower. It is read by SiLVANN's runtime, not by transformers. The tokenizer, chat template and configs are the original model's.

All credit for the model belongs to Z.AI; this repository only changes how its weights are stored. The original license (MIT) applies and is included.

D8E4 or D4E4

This pack's routed experts are the D4E4 pack's, byte for byte; only the dense part is wider, packed again from the original weights. Measured on WikiText-2's test split (16 windows of 512 tokens): D8E4 perplexity 3.018, 0.3584 bits per byte; D4E4 3.095, 0.3666 — 2.5% for 3 GiB more. Both run with the same settings.