Add IQ1_KT (with iq4_nl ffn_down_exps lmao)

README.md

@@ -347,6 +347,66 @@ numactl -N 0 -m 0 \
 
 </details>
 
+## IQ1_KT 36.039 GiB (2.802 BPW)
+Final estimate: PPL = 5.8214 +/- 0.03767
+
+<details>
+
+<summary>👈 Secret Recipe</summary>
+
+```bash
+#!/usr/bin/env bash
+
+custom="
+# 47 Repeating Layers [0-46]
+# Note: All ffn_down.* layers are not divisible by 256 so have limited quantization options.
+
+# Attention
+blk\..*\.attn_q.*=iq4_kt
+blk\..*\.attn_k.*=iq4_kt
+blk\..*\.attn_v.*=iq4_kt
+blk\..*\.attn_output.*=iq4_kt
+
+# First 1 Dense Layers [0]
+blk\..*\.ffn_down\.weight=iq4_nl
+blk\..*\.ffn_(gate|up)\.weight=iq4_kt
+
+# Shared Expert Layers [1-46]
+blk\..*\.ffn_down_shexp\.weight=iq4_nl
+blk\..*\.ffn_(gate|up)_shexp\.weight=iq4_kt
+
+# Routed Experts Layers [1-46]
+blk\..*\.ffn_down_exps\.weight=iq4_nl
+blk\..*\.ffn_(gate|up)_exps\.weight=iq1_kt
+
+# NextN MTP Layer [46]
+blk\..*\.nextn\.embed_tokens\.weight=iq4_kt
+blk\..*\.nextn\.shared_head_head\.weight=iq4_kt
+blk\..*\.nextn\.eh_proj\.weight=q8_0
+
+# Non-Repeating Layers
+token_embd\.weight=iq4_k
+output\.weight=iq6_k
+"
+
+custom=$(
+  echo "$custom" | grep -v '^#' | \
+  sed -Ez 's:\n+:,:g;s:,$::;s:^,::'
+)
+
+numactl -N 1 -m 1 \
+./build/bin/llama-quantize \
+    --custom-q "$custom" \
+    --imatrix /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/imatrix-GLM-4.5-Air-BF16.dat \
+    /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/GLM-4.5-Air-128x9.4B-BF16-00001-of-00005.gguf \
+    /mnt/raid/models/ubergarm/GLM-4.5-Air-GGUF/GLM-4.5-Air-IQ1_KT.gguf \
+    IQ1_KT \
+    192
+```
+
+</details>
+
+
 ## Quick Start
 If you want to disable thinking, add `/nothink` (correct, no underscore) at the *end* of your prompt.