uploading IQ2_KS

README.md

@@ -173,16 +173,47 @@ numactl -N ${SOCKET} -m ${SOCKET} \
 
 </details>
 
-## IQ2_KS TODO
-TODO
+## IQ2_KS 69.800 GiB (2.622 BPW)
+PPL over 552 chunks for n_ctx=512 = 9.6827 +/- 0.07972
 
-I'm hoping this IQ2_KS/IQ3_KS mix will run on 96GB VRAM full offload with enough context for agentic use.
 <details>
 
 <summary>👈 Secret Recipe</summary>
 
 ```bash
-echo TODO
+#!/usr/bin/env bash
+
+custom="
+# 61 Repeating Layers [0-61]
+
+# Attention [0-61] GPU
+blk\..*\.attn_q.*=q8_0
+blk\..*\.attn_k.*=q8_0
+blk\..*\.attn_v.*=q8_0
+blk\..*\.attn_output.*=q8_0
+
+# Routed Experts Layers [0-61] CPU
+blk\..*\.ffn_down_exps\.weight=iq3_ks
+blk\..*\.ffn_(gate|up)_exps\.weight=iq2_ks
+
+# 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 ${SOCKET} -m ${SOCKET} \
+./build/bin/llama-quantize \
+    --custom-q "$custom" \
+    --imatrix /mnt/data/models/ubergarm/MiniMax-M2.5-GGUF/imatrix-MiniMax-M2.5-BF16.dat \
+    /mnt/data/models/ubergarm/MiniMax-M2.5-GGUF/MiniMax-M2.5-256x4.9B-BF16-00001-of-00010.gguf \
+    /mnt/data/models/ubergarm/MiniMax-M2.5-GGUF/MiniMax-M2.5-IQ2_KS.gguf \
+    IQ2_KS \
+    128
 ```
 
 </details>