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
```
|