wesleysimplicio commited on
Commit
df67e5e
·
verified ·
1 Parent(s): c94a4e2

docs: sync deploy/convert_gguf.sh with updated benchmark numbers

Browse files
Files changed (1) hide show
  1. deploy/convert_gguf.sh +41 -0
deploy/convert_gguf.sh ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env bash
2
+ # ==============================================================================
3
+ # Pipeline de Conversão e Quantização GGUF para Ollama (Simplicio 27B)
4
+ # ==============================================================================
5
+ set -e
6
+
7
+ MODEL_DIR=${1:-"./checkpoint-final"}
8
+ OUTPUT_NAME=${2:-"simplicio-27b"}
9
+
10
+ echo "=== 1. Clonando llama.cpp (se necessário) ==="
11
+ if [ ! -d "llama.cpp" ]; then
12
+ git clone https://github.com/ggerganov/llama.cpp
13
+ cd llama.cpp
14
+ cmake -B build
15
+ cmake --build build --config Release -j
16
+ pip install -r requirements.txt
17
+ cd ..
18
+ fi
19
+
20
+ echo "=== 2. Convertendo Hugging Face para GGUF (FP16 / BF16) ==="
21
+ python3 llama.cpp/convert_hf_to_gguf.py "$MODEL_DIR" \
22
+ --outfile "${OUTPUT_NAME}-bf16.gguf" \
23
+ --outtype bf16
24
+
25
+ echo "=== 3. Gerando Quantização Recomendada (Q4_K_M ~16.8 GB) ==="
26
+ ./llama.cpp/build/bin/llama-quantize \
27
+ "${OUTPUT_NAME}-bf16.gguf" \
28
+ "${OUTPUT_NAME}-Q4_K_M.gguf" \
29
+ Q4_K_M
30
+
31
+ echo "=== 4. Gerando Quantização Alta Fidelidade (Q8_0 ~28.5 GB) ==="
32
+ ./llama.cpp/build/bin/llama-quantize \
33
+ "${OUTPUT_NAME}-bf16.gguf" \
34
+ "${OUTPUT_NAME}-Q8_0.gguf" \
35
+ Q8_0
36
+
37
+ echo "=== 5. Criando Modelo no Ollama Local ==="
38
+ ollama create wesleysimplicio/simplicio-27b -f Modelfile
39
+
40
+ echo "=== Concluído! Para testar execute: ==="
41
+ echo "ollama run wesleysimplicio/simplicio-27b"