antirez-recipe quants
MoE quants following the antirez recipe: routed experts quantized hard, everything else protected.
| Tensor class | Quant | Notes |
|---|---|---|
blk.*.ffn_gate_exps, blk.*.ffn_up_exps (ffn_gate_up_exps if fused) |
GU_TYPE (imatrix) | routed-expert gate/up |
blk.*.ffn_down_exps |
DOWN_TYPE (imatrix) | routed-expert down (K-quant for quality) |
blk.*.ffn_{gate,up,down}_shexp, all attn_*, output.weight |
Q8_0 |
shared experts / attention / head |
token_embd.weight |
F16 |
embeddings |
router (ffn_gate_inp), norms, sinks, biases |
source precision (BF16/F32) |
never quantized by llama.cpp |
Base type Q8_0 + --tensor-type overrides, imatrix calibration
(see SPEC per-model imatrix links). Runs on upstream llama.cpp — no fork or
patch needed.
Validation per file: gguf-dump tensor audit (experts = GU/DOWN, protection intact),
llama-cli smoke load, llama-perplexity spot ppl + --kl-divergence vs BF16 logits.
Files
| File | Size | SHA-256 |
|---|---|---|
occamy-1.0-Q2-IQ2_XXS-Q2_K.gguf |
12052349184 | 923faddcac1f0d4268509f073a2ce0e01919b9183299ede48abc99c647167752 |
Imatrix: https://huggingface.co/mradermacher/occamy-1.0-i1-GGUF
Recipe
Antirez recipe on occamy-ai/occamy-1.0 (BF16 source: bartowski/occamy-ai_occamy-1.0-GGUF):
ffn_gate_exps/ffn_up_exps-> IQ2_XXSffn_down_exps-> Q2_K- attention + shared experts + output -> Q8_0
- token embeddings -> F16
- imatrix: mradermacher/occamy-1.0-i1-GGUF
~11.4 GB, 2.78 BPW (66.1 GiB -> 11.5 GiB).
Tokenizer note: the bartowski BF16 carries tokenizer.ggml.pre = qwen35, which Accio's TOKENIZER.md documents as wrong for occamy; this build has qwen2 baked in at quantize time via --override-kv tokenizer.ggml.pre=str:qwen2 (no post-hoc copy).
Validation (llama.cpp, qwen35moe arch)
- 12-case hard-smoke eval (temp 0, ctx 15360, model's own embedded chat template): 7/12 PASS; ~27 t/s generation (AMD Vulkan/UMA); fails cases 2826, 10360, 10567, 9487, 1716. Passes both big-code cases.
- Exact output tokens (own tokenizer) over the 7 solved: 56,142 = 8,020 tok/solved.
- KL divergence vs a same-tokenizer Q8_0 reference built from the same BF16 with the same pre-tokenizer correction (ds4 short prompts, ctx 60, 2 chunks, CPU): per-chunk KL 0.150 / 0.141 (avg 0.146), delta-p RMS 10-12%, same-token top-1 96.6 / 91.4%.