File size: 3,752 Bytes
9835d54
 
 
 
 
 
 
 
8f83938
 
9835d54
 
 
 
 
 
 
 
8bd5e50
 
 
 
9835d54
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
quantized_by: anikifoss
pipeline_tag: text-generation
base_model: Qwen/Qwen3-Coder-480B-A35B-Instruct
license: apache-2.0
base_model_relation: quantized
tags:
- conversational
- gguf
- no_imatrix
---

# Model Card

High quality quantization of **Qwen3-Coder-480B-A35B-Instruct** 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/Qwen3-Coder-480B-A35B-Instruct-HQ4_K \
    --model /mnt/data/Models/anikifoss/Qwen3-Coder-480B-A35B-Instruct-HQ4_K/Qwen3-Coder-480B-A35B-Instruct-HQ4_K-00001-of-00007.gguf \
    --no-mmap -rtr \
    --temp 0.5 --top-k 0 --top-p 1.0 --min-p 0.1 --repeat-penalty 1.0 \
    --ctx-size 51000 \
    -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/Qwen3-Coder-480B-A35B-Instruct-HQ4_K \
    --model /mnt/data/Models/anikifoss/Qwen3-Coder-480B-A35B-Instruct-HQ4_K/Qwen3-Coder-480B-A35B-Instruct-HQ4_K-00001-of-00007.gguf \
    --no-mmap \
    --temp 0.5 --top-k 0 --top-p 1.0 --min-p 0.1 --repeat-penalty 1.0 \
    --ctx-size 51000 \
    -ctk f16 -ctv f16 \
    -fa \
    -b 1024 -ub 1024 \
    --n-gpu-layers 99 \
    --override-tensor exps=CPU \
    --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="
# Token embedding and output tensors
output\.weight=bf16
output_norm\.weight=f32
token_embd\.weight=bf16

blk\.[0-9]\.attn_k\.weight=q8_0
blk\.[0-9]\.attn_k_norm\.weight=f32
blk\.[0-9]\.attn_norm\.weight=f32
blk\.[0-9]\.attn_output\.weight=q8_0
blk\.[0-9]\.attn_q\.weight=q8_0
blk\.[0-9]\.attn_q_norm\.weight=f32
blk\.[0-9]\.attn_v\.weight=q8_0
blk\.[0-9]\.ffn_down_exps\.weight=q6_K
blk\.[0-9]\.ffn_gate_exps\.weight=q4_K
blk\.[0-9]\.ffn_up_exps\.weight=q4_K
blk\.[0-9]\.ffn_gate_inp\.weight=f32
blk\.[0-9]\.ffn_norm\.weight=f32
blk\.[1-5][0-9]\.attn_k\.weight=q8_0
blk\.[1-5][0-9]\.attn_k_norm\.weight=f32
blk\.[1-5][0-9]\.attn_norm\.weight=f32
blk\.[1-5][0-9]\.attn_output\.weight=q8_0
blk\.[1-5][0-9]\.attn_q\.weight=q8_0
blk\.[1-5][0-9]\.attn_q_norm\.weight=f32
blk\.[1-5][0-9]\.attn_v\.weight=q8_0
blk\.[1-5][0-9]\.ffn_down_exps\.weight=q6_K
blk\.[1-5][0-9]\.ffn_gate_exps\.weight=q4_K
blk\.[1-5][0-9]\.ffn_up_exps\.weight=q4_K
blk\.[1-5][0-9]\.ffn_gate_inp\.weight=f32
blk\.[1-5][0-9]\.ffn_norm\.weight=f32
blk\.6[0-1]\.attn_k\.weight=q8_0
blk\.6[0-1]\.attn_k_norm\.weight=f32
blk\.6[0-1]\.attn_norm\.weight=f32
blk\.6[0-1]\.attn_output\.weight=q8_0
blk\.6[0-1]\.attn_q\.weight=q8_0
blk\.6[0-1]\.attn_q_norm\.weight=f32
blk\.6[0-1]\.attn_v\.weight=q8_0
blk\.6[0-1]\.ffn_down_exps\.weight=q6_K
blk\.6[0-1]\.ffn_gate_exps\.weight=q4_K
blk\.6[0-1]\.ffn_up_exps\.weight=q4_K
blk\.6[0-1]\.ffn_gate_inp\.weight=f32
blk\.6[0-1]\.ffn_norm\.weight=f32
"

custom=$(
  echo "$custom" | grep -v '^#' | \
  sed -Ez 's:\n+:,:g;s:,$::;s:^,::'
)

echo "Running with: -custom-q $custom"

TARGET_MODEL="Qwen3-Coder-480B-A35B-Instruct-HQ4_K"
mkdir -p ~/Env/models/anikifoss/$TARGET_MODEL
./build/bin/llama-quantize \
    --custom-q "$custom" \
    /mnt/data/Models/Qwen/Qwen3-Coder-480B-A35B-Instruct-GGUF/Qwen3-Coder-480B-A35B-Instruct-BF16-00001-of-00021.gguf \
    ~/Env/models/anikifoss/$TARGET_MODEL/$TARGET_MODEL.gguf \
    Q4_K \
    32
```