firefox + llama.cpp == very good prose.
chore(readme) quickstart llaamaa
+24
+24
README.md
+24
README.md
···
4
4
5
5
6
6
7
+
## quickstart
8
+
9
+
```bash
10
+
env | sort -u | grep -iP '^llama.*'
11
+
LLAMA_ARG_CPU_MOE=true
12
+
LLAMA_ARG_CTX_CHECKPOINTS=3
13
+
LLAMA_ARG_DIO=true
14
+
LLAMA_ARG_KV_UNIFIED=true
15
+
LLAMA_ARG_PERF=false
16
+
LLAMA_ARG_SWA_FULL=true
17
+
LLAMA_LOG_FILE=/tmp/llamacpp.log
18
+
LLAMA_LOG_VERBOSITY=3
19
+
20
+
llama-server -hf unsloth/gemma-4-E2B-it-GGUF:Q4_K_S \
21
+
-ngl 99 \
22
+
--ubatch-size 512 --batch-size 2048 \
23
+
--ctx-size 2048 \
24
+
--cache-ram 0 \
25
+
--reasoning-budget 1024 \
26
+
--threads 8 \
27
+
--fit off \
28
+
--device CUDA0
29
+
```
30
+
7
31
## disclaimer
8
32
9
33
- coded with an LLM, adjusted, refactored, verified by hand.