upload smol-IQ4_KSS and post all perplexity data

README.md

@@ -147,6 +147,60 @@ numactl -N ${SOCKET} -m ${SOCKET} \
 
 </details>
 
+## smol-IQ4_KSS 94.080 GiB (4.103 BPW)
+PPL over 561 chunks for n_ctx=512 = 2.5705 +/- 0.01211
+
+<details>
+
+<summary>👈 Secret Recipe</summary>
+
+```bash
+#!/usr/bin/env bash
+
+custom="
+# 45 Repeating Layers [0-44]
+
+# Attention [0-44] GPU
+blk\..*\.attn_gate.*=iq6_k
+blk\..*\.attn_q.*=iq6_k
+blk\..*\.attn_k.*=iq6_k
+blk\..*\.attn_v.*=iq6_k
+blk\..*\.attn_output.*=iq6_k
+
+# First 3 Dense Layers [0-2] GPU
+blk\..*\.ffn_down\.weight=iq6_k
+blk\..*\.ffn_(gate|up)\.weight=iq6_k
+
+# Shared Expert Layers [3-44] GPU
+blk\..*\.ffn_down_shexp\.weight=iq6_k
+blk\..*\.ffn_(gate|up)_shexp\.weight=iq6_k
+
+# Routed Experts Layers [3-44] CPU
+blk\..*\.ffn_down_exps\.weight=iq4_kss
+blk\..*\.ffn_(gate|up)_exps\.weight=iq4_kss
+
+# 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/Step-3.5-Flash-GGUF/imatrix-Step-3.5-Flash-BF16.dat \
+    /mnt/data/models/ubergarm/Step-3.5-Flash-GGUF/Step-3.5-Flash-288x7.4B-BF16-00001-of-00009.gguf \
+    /mnt/data/models/ubergarm/Step-3.5-Flash-GGUF/Step-3.5-Flash-smol-IQ4_KSS.gguf \
+    IQ4_KSS \
+    128
+```
+
+</details>
+
 ## smol-IQ3_KS 75.934 GiB (3.312 BPW)
 PPL over 561 chunks for n_ctx=512 = 2.7856 +/- 0.01365