add IQ2_KL and IQ4_KSS

README.md

@@ -35,6 +35,60 @@ These two are just test quants for baseline perplexity comparison and not availa
 
 *NOTE*: The first split file is much smaller on purpose to only contain metadata, its fine!
 
+## IQ4_KSS 194.058 GiB (4.206 BPW)
+PPL over 580 chunks for n_ctx=512 = 3.5102 +/- 0.01896
+
+<details>
+
+<summary>👈 Secret Recipe</summary>
+
+```bash
+#!/usr/bin/env bash
+
+custom="
+# 60 Repeating Layers [0-59]
+
+## Gated Attention/Delta Net [Blended 0-59]
+blk\..*\.attn_gate\.weight=q8_0
+blk\..*\.attn_qkv\.weight=q8_0
+blk\..*\.attn_output\.weight=q8_0
+blk\..*\.attn_q\.weight=q8_0
+blk\..*\.attn_k\.weight=q8_0
+blk\..*\.attn_v\.weight=q8_0
+blk\..*\.ssm_alpha\.weight=q8_0
+blk\..*\.ssm_beta\.weight=q8_0
+blk\..*\.ssm_out\.weight=q8_0
+
+# Shared Expert Layers [0-59]
+blk\..*\.ffn_down_shexp\.weight=q8_0
+blk\..*\.ffn_(gate|up)_shexp\.weight=q8_0
+
+# Routed Experts Layers [0-59]
+blk\..*\.ffn_down_exps\.weight=iq4_ks
+blk\..*\.ffn_(gate|up)_exps\.weight=iq4_kss
+
+# Non-Repeating Layers
+token_embd\.weight=iq6_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/Qwen3.5-397B-A17B-GGUF/imatrix-Qwen3.5-397B-A17B-BF16.dat \
+    /mnt/data/models/ubergarm/Qwen3.5-397B-A17B-GGUF/Qwen3.5-397B-A17B-BF16-00001-of-00017.gguf \
+    /mnt/data/models/ubergarm/Qwen3.5-397B-A17B-GGUF/Qwen3.5-397B-A17B-IQ4_KSS.gguf \
+    IQ4_KSS \
+    128
+```
+
+</details>
+
 ## Q3_K 179.97 GiB (3.90 BPW)
 PPL over 580 chunks for n_ctx=512 = 3.5409 +/- 0.01924
 
@@ -62,6 +116,60 @@ This is a *custom* mainline llama.cpp compatible MoE optimized mix similar to Ae
 
 </details>
 
+## IQ2_KL 138.142 GiB (2.994 BPW)
+PPL over 580 chunks for n_ctx=512 = 3.6536 +/- 0.02000
+
+<details>
+
+<summary>👈 Secret Recipe</summary>
+
+```bash
+#!/usr/bin/env bash
+
+custom="
+# 60 Repeating Layers [0-59]
+
+## Gated Attention/Delta Net [Blended 0-59]
+blk\..*\.attn_gate\.weight=q8_0
+blk\..*\.attn_qkv\.weight=q8_0
+blk\..*\.attn_output\.weight=q8_0
+blk\..*\.attn_q\.weight=q8_0
+blk\..*\.attn_k\.weight=q8_0
+blk\..*\.attn_v\.weight=q8_0
+blk\..*\.ssm_alpha\.weight=q8_0
+blk\..*\.ssm_beta\.weight=q8_0
+blk\..*\.ssm_out\.weight=q8_0
+
+# Shared Expert Layers [0-59]
+blk\..*\.ffn_down_shexp\.weight=q8_0
+blk\..*\.ffn_(gate|up)_shexp\.weight=q8_0
+
+# Routed Experts Layers [0-59]
+blk\..*\.ffn_down_exps\.weight=iq3_ks
+blk\..*\.ffn_(gate|up)_exps\.weight=iq2_kl
+
+# 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/Qwen3.5-397B-A17B-GGUF/imatrix-Qwen3.5-397B-A17B-BF16.dat \
+    /mnt/data/models/ubergarm/Qwen3.5-397B-A17B-GGUF/Qwen3.5-397B-A17B-BF16-00001-of-00017.gguf \
+    /mnt/data/models/ubergarm/Qwen3.5-397B-A17B-GGUF/Qwen3.5-397B-A17B-IQ2_KL.gguf \
+    IQ2_KL \
+    128
+```
+
+</details>
+
 ## smol-IQ2_XS 113.41 GiB (2.46 BPW)
 PPL over 580 chunks for n_ctx=512 = 3.8717 +/- 0.02131