File size: 4,347 Bytes
69d993b cb85180 69d993b 38c2ba8 69d993b 38c2ba8 69d993b 38c2ba8 69d993b 38c2ba8 69d993b 38c2ba8 69d993b 38c2ba8 69d993b 38c2ba8 69d993b 38c2ba8 69d993b 38c2ba8 69d993b 38c2ba8 69d993b 38c2ba8 69d993b 38c2ba8 69d993b 38c2ba8 69d993b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 |
---
quantized_by: anikifoss
pipeline_tag: text-generation
base_model: zai-org/GLM-4.5-Air
license: mit
base_model_relation: quantized
tags:
- conversational
---
# Model Card
High quality quantization of **GLM-4.5-Air** without using imatrix.
# Run
## ik_llama.cpp
See [this detailed guide](https://github.com/ikawrakow/ik_llama.cpp/discussions/258) on how to setup ik_llama and how to make custom quants.
```
./build/bin/llama-server \
--alias anikifoss/GLM-4.5-Air-HQ4_K \
--model /mnt/data/Models/anikifoss/GLM-4.5-Air-HQ4_K/GLM-4.5-Air-HQ4_K-00001-of-00002.gguf \
--no-mmap -rtr \
--temp 0.5 --top-k 0 --top-p 1.0 --min-p 0.1 --repeat-penalty 1.0 \
--ctx-size 116000 \
-ctk f16 -ctv f16 \
-fa \
-b 1024 -ub 1024 \
-fmoe \
--n-gpu-layers 99 \
--override-tensor exps=CPU \
--parallel 1 \
--threads 32 \
--threads-batch 64 \
--host 127.0.0.1 \
--port 8090
```
## llama.cpp
```
./build/bin/llama-server \
--alias anikifoss/GLM-4.5-Air-HQ4_K \
--model /mnt/data/Models/anikifoss/GLM-4.5-Air-HQ4_K/GLM-4.5-Air-HQ4_K-00001-of-00002.gguf \
--no-mmap \
--temp 0.5 --top-k 0 --top-p 1.0 --min-p 0.1 \
--ctx-size 116000 \
-ctk f16 -ctv f16 \
-fa \
--override-tensor exps=CPU \
-ngl 99 \
--jinja \
--parallel 1 \
--threads 32 \
--threads-batch 64 \
--host 127.0.0.1 \
--port 8090
```
## Quantization Recipe
Quantized with [ik_llama](https://github.com/ikawrakow/ik_llama.cpp), but should work with any GGUF compatible inference framework.
```bash
#!/usr/bin/env bash
custom="
blk\.0\.ffn_down\.weight=q8_0
blk\.0\.ffn_gate\.weight=q8_0
blk\.0\.ffn_up\.weight=q8_0
blk\.46\.nextn\.eh_proj\.weight=bf16
blk\.46\.nextn\.embed_tokens\.weight=q8_0
blk\.46\.nextn\.enorm\.weight=f32
blk\.46\.nextn\.hnorm\.weight=f32
blk\.46\.nextn\.shared_head_head\.weight=q8_0
blk\.46\.nextn\.shared_head_norm\.weight=f32
blk\.[0-9]\.attn_k\.bias=f32
blk\.[0-9]\.attn_k\.weight=q8_0
blk\.[0-9]\.attn_norm\.weight=f32
blk\.[0-9]\.attn_output\.weight=q8_0
blk\.[0-9]\.attn_q\.bias=f32
blk\.[0-9]\.attn_q\.weight=q8_0
blk\.[0-9]\.attn_v\.bias=f32
blk\.[0-9]\.attn_v\.weight=q8_0
blk\.[0-9]\.post_attention_norm\.weight=f32
blk\.[1-3][0-9]\.attn_k\.bias=f32
blk\.[1-3][0-9]\.attn_k\.weight=q8_0
blk\.[1-3][0-9]\.attn_norm\.weight=f32
blk\.[1-3][0-9]\.attn_output\.weight=q8_0
blk\.[1-3][0-9]\.attn_q\.bias=f32
blk\.[1-3][0-9]\.attn_q\.weight=q8_0
blk\.[1-3][0-9]\.attn_v\.bias=f32
blk\.[1-3][0-9]\.attn_v\.weight=q8_0
blk\.[1-3][0-9]\.post_attention_norm\.weight=f32
blk\.4[0-6]\.attn_k\.bias=f32
blk\.4[0-6]\.attn_k\.weight=q8_0
blk\.4[0-6]\.attn_norm\.weight=f32
blk\.4[0-6]\.attn_output\.weight=q8_0
blk\.4[0-6]\.attn_q\.bias=f32
blk\.4[0-6]\.attn_q\.weight=q8_0
blk\.4[0-6]\.attn_v\.bias=f32
blk\.4[0-6]\.attn_v\.weight=q8_0
blk\.4[0-6]\.post_attention_norm\.weight=f32
blk\.[1-9]\.exp_probs_b\.bias=f32
blk\.[1-9]\.ffn_down_exps\.weight=q8_0
blk\.[1-9]\.ffn_down_shexp\.weight=bf16
blk\.[1-9]\.ffn_gate_exps\.weight=q4_K
blk\.[1-9]\.ffn_gate_inp\.weight=f32
blk\.[1-9]\.ffn_gate_shexp\.weight=bf16
blk\.[1-9]\.ffn_up_exps\.weight=q4_K
blk\.[1-9]\.ffn_up_shexp\.weight=bf16
blk\.[1-3][0-9]\.exp_probs_b\.bias=f32
blk\.[1-3][0-9]\.ffn_down_exps\.weight=q8_0
blk\.[1-3][0-9]\.ffn_down_shexp\.weight=bf16
blk\.[1-3][0-9]\.ffn_gate_exps\.weight=q4_K
blk\.[1-3][0-9]\.ffn_gate_inp\.weight=f32
blk\.[1-3][0-9]\.ffn_gate_shexp\.weight=bf16
blk\.[1-3][0-9]\.ffn_up_exps\.weight=q4_K
blk\.[1-3][0-9]\.ffn_up_shexp\.weight=bf16
blk\.4[0-6]\.exp_probs_b\.bias=f32
blk\.4[0-6]\.ffn_down_exps\.weight=q8_0
blk\.4[0-6]\.ffn_down_shexp\.weight=bf16
blk\.4[0-6]\.ffn_gate_exps\.weight=q4_K
blk\.4[0-6]\.ffn_gate_inp\.weight=f32
blk\.4[0-6]\.ffn_gate_shexp\.weight=bf16
blk\.4[0-6]\.ffn_up_exps\.weight=q4_K
blk\.4[0-6]\.ffn_up_shexp\.weight=bf16
output\.weight=bf16
output_norm\.weight=f32
token_embd\.weight=bf16
"
custom=$(
echo "$custom" | grep -v '^#' | \
sed -Ez 's:\n+:,:g;s:,$::;s:^,::'
)
echo "Running with: -custom-q $custom"
TARGET_MODEL="GLM-4.5-Air-HQ4_K"
mkdir -p ~/Env/models/anikifoss/$TARGET_MODEL
./build/bin/llama-quantize \
--custom-q "$custom" \
/mnt/data/Models/zai-org/GLM-4.5-Air-GGUF/GLM-4.5-Air-128x9.4B-BF16-00001-of-00005.gguf \
~/Env/models/anikifoss/$TARGET_MODEL/$TARGET_MODEL.gguf \
Q4_K \
32
``` |