bots-garden/houlapublic Fork
⑂ 1dfc968
Commits
⬇ Clone ▾
git clone https://git.rickub.com/bots-garden/houla.git
git clone ssh://git@rickub.com/bots-garden/houla.git

Host key fingerprint (ed25519): SHA256:iycHnxEyq0Q7uyVpB7JlznP0G7JrTPXLYRcAU5CSLhc — verify it before your first connect.

Download ZIP

🛟 Updated.Unverified

k33g committed 2026-11-21T09:11:52+01:00 Browse files
1dfc968
added .gitignore +8 -0
new file mode 100644
@@ -0,0 +1,8 @@
1+set-hf-token.md
2+
3+# Downloaded GGUF models (scripts/olala.sh, default OLALA_MODELS_DIR)
4+/models/
5+*.gguf
6+**/.cache/huggingface/
7+.claude
8+.memory
new file mode 100644
@@ -0,0 +1,8 @@
1+set-hf-token.md
2+
3+# Downloaded GGUF models (scripts/olala.sh, default OLALA_MODELS_DIR)
4+/models/
5+*.gguf
6+**/.cache/huggingface/
7+.claude
8+.memory
added .tickets/config.yaml +2 -0
new file mode 100644
@@ -0,0 +1,2 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+version: 1
new file mode 100644
@@ -0,0 +1,2 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+version: 1
added .tickets/epics.yaml +6 -0
new file mode 100644
@@ -0,0 +1,6 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+epics:
3+ - title: Olala in sbx
4+ state: open
5+ color: "#5319e7"
6+ description: Run OVHai's Olala-v1.0-mini-7A1B (GGUF, llama.cpp fork, CPU only) in a Docker sandbox and serve it to the host
new file mode 100644
@@ -0,0 +1,6 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+epics:
3+ - title: Olala in sbx
4+ state: open
5+ color: "#5319e7"
6+ description: Run OVHai's Olala-v1.0-mini-7A1B (GGUF, llama.cpp fork, CPU only) in a Docker sandbox and serve it to the host
added .tickets/issues/0001-initialize.yaml +36 -0
new file mode 100644
@@ -0,0 +1,36 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 1
3+title: initialize
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:18:39.078Z
9+updatedAt: 2026-11-21T07:52:58.000Z
10+labels: []
11+body: |
12+ OVH released a new model: https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B
13+
14+ To run it in GGUF format we need a fork of llama.cpp, explained here: https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF
15+ (the links to the fork are on that page, along with how to build and run it)
16+
17+ Everything happens in an sbx sandbox, so we use the build without GPU (CPU only).
18+
19+ Download the model: Olala-v1.0-mini-7A1B-q8_0-safe2.gguf
20+
21+ then serve it so it can be reached from outside.
22+
23+ Create a sbxenv.yaml file so this is easy to repeat: it must handle the model being already downloaded or not, expose a port, and optionally allow downloading and serving another model.
24+
25+ And give the curl commands to test it.
26+tasks: []
27+comments:
28+ - id: 1
29+ author:
30+ name: k33g
31+ email: ph.charriere@gmail.com
32+ createdAt: 2026-11-21T06:37:13.000Z
33+ body: |
34+ Split into #2 to #8 (epic ❖ Olala in sbx). Done in sandbox `houla-houla`: fork built, model downloaded, llama-server running and answering on port 8080 (~14 tokens/s, aarch64 CPU).
35+ Still to do on the host: publish the port (`sbx ports houla-houla --publish 8080:8080`) and try `sbx env create sbxenv.houla.yaml` (#6). Kit: #7.
36+ See [docs/olala-sbx.en.md](docs/olala-sbx.en.md).
new file mode 100644
@@ -0,0 +1,36 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 1
3+title: initialize
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:18:39.078Z
9+updatedAt: 2026-11-21T07:52:58.000Z
10+labels: []
11+body: |
12+ OVH released a new model: https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B
13+
14+ To run it in GGUF format we need a fork of llama.cpp, explained here: https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF
15+ (the links to the fork are on that page, along with how to build and run it)
16+
17+ Everything happens in an sbx sandbox, so we use the build without GPU (CPU only).
18+
19+ Download the model: Olala-v1.0-mini-7A1B-q8_0-safe2.gguf
20+
21+ then serve it so it can be reached from outside.
22+
23+ Create a sbxenv.yaml file so this is easy to repeat: it must handle the model being already downloaded or not, expose a port, and optionally allow downloading and serving another model.
24+
25+ And give the curl commands to test it.
26+tasks: []
27+comments:
28+ - id: 1
29+ author:
30+ name: k33g
31+ email: ph.charriere@gmail.com
32+ createdAt: 2026-11-21T06:37:13.000Z
33+ body: |
34+ Split into #2 to #8 (epic ❖ Olala in sbx). Done in sandbox `houla-houla`: fork built, model downloaded, llama-server running and answering on port 8080 (~14 tokens/s, aarch64 CPU).
35+ Still to do on the host: publish the port (`sbx ports houla-houla --publish 8080:8080`) and try `sbx env create sbxenv.houla.yaml` (#6). Kit: #7.
36+ See [docs/olala-sbx.en.md](docs/olala-sbx.en.md).
added .tickets/issues/0002-build-the-ovhai-llama-cpp-fork-cpu-only.yaml +60 -0
new file mode 100644
@@ -0,0 +1,60 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 2
3+title: Build the OVHai llama.cpp fork (CPU only)
4+state: closed
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T06:37:13.000Z
10+labels:
11+ - kind::chore
12+ - priority::high
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ Upstream llama.cpp does not know the `olala` architecture: the fork https://github.com/OVHai-LLM/llama.cpp is required
18+ (see [the GGUF model card](https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF)).
19+
20+ CPU only build (the sandbox has no GPU; it is **aarch64**, 18 CPUs, 31 GB RAM — the card benchmarks x86 CPUs).
21+
22+ ```bash
23+ cmake -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_USE_PREBUILT_UI=OFF
24+ cmake --build build -j --target llama-server llama-cli
25+ ```
26+
27+ ## Acceptance criteria
28+ - `llama-server` and `llama-cli` built in `~/olala/llama.cpp-olala/build/bin`.
29+tasks:
30+ - id: 1
31+ title: Install build-essential, cmake, libssl-dev
32+ state: closed
33+ priority: high
34+ author:
35+ name: k33g
36+ email: ph.charriere@gmail.com
37+ createdAt: 2026-11-21T06:35:00.000Z
38+ updatedAt: 2026-11-21T06:35:00.000Z
39+ closedAt: 2026-11-21T06:35:00.000Z
40+ - id: 2
41+ title: Clone the fork (depth 1, commit 55c22dc)
42+ state: closed
43+ priority: high
44+ author:
45+ name: k33g
46+ email: ph.charriere@gmail.com
47+ createdAt: 2026-11-21T06:35:00.000Z
48+ updatedAt: 2026-11-21T06:35:00.000Z
49+ closedAt: 2026-11-21T06:35:00.000Z
50+ - id: 3
51+ title: Build llama-server and llama-cli
52+ state: closed
53+ priority: high
54+ author:
55+ name: k33g
56+ email: ph.charriere@gmail.com
57+ createdAt: 2026-11-21T06:35:00.000Z
58+ updatedAt: 2026-11-21T06:35:00.000Z
59+ closedAt: 2026-11-21T06:35:00.000Z
60+comments: []
new file mode 100644
@@ -0,0 +1,60 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 2
3+title: Build the OVHai llama.cpp fork (CPU only)
4+state: closed
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T06:37:13.000Z
10+labels:
11+ - kind::chore
12+ - priority::high
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ Upstream llama.cpp does not know the `olala` architecture: the fork https://github.com/OVHai-LLM/llama.cpp is required
18+ (see [the GGUF model card](https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF)).
19+
20+ CPU only build (the sandbox has no GPU; it is **aarch64**, 18 CPUs, 31 GB RAM — the card benchmarks x86 CPUs).
21+
22+ ```bash
23+ cmake -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_USE_PREBUILT_UI=OFF
24+ cmake --build build -j --target llama-server llama-cli
25+ ```
26+
27+ ## Acceptance criteria
28+ - `llama-server` and `llama-cli` built in `~/olala/llama.cpp-olala/build/bin`.
29+tasks:
30+ - id: 1
31+ title: Install build-essential, cmake, libssl-dev
32+ state: closed
33+ priority: high
34+ author:
35+ name: k33g
36+ email: ph.charriere@gmail.com
37+ createdAt: 2026-11-21T06:35:00.000Z
38+ updatedAt: 2026-11-21T06:35:00.000Z
39+ closedAt: 2026-11-21T06:35:00.000Z
40+ - id: 2
41+ title: Clone the fork (depth 1, commit 55c22dc)
42+ state: closed
43+ priority: high
44+ author:
45+ name: k33g
46+ email: ph.charriere@gmail.com
47+ createdAt: 2026-11-21T06:35:00.000Z
48+ updatedAt: 2026-11-21T06:35:00.000Z
49+ closedAt: 2026-11-21T06:35:00.000Z
50+ - id: 3
51+ title: Build llama-server and llama-cli
52+ state: closed
53+ priority: high
54+ author:
55+ name: k33g
56+ email: ph.charriere@gmail.com
57+ createdAt: 2026-11-21T06:35:00.000Z
58+ updatedAt: 2026-11-21T06:35:00.000Z
59+ closedAt: 2026-11-21T06:35:00.000Z
60+comments: []
added .tickets/issues/0003-download-olala-v1-0-mini-7a1b-q8-0-safe2-gguf.yaml +51 -0
new file mode 100644
@@ -0,0 +1,51 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 3
3+title: Download Olala-v1.0-mini-7A1B-q8_0-safe2.gguf
4+state: closed
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T06:37:13.000Z
10+labels:
11+ - kind::chore
12+ - priority::high
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ Download the 8-bit file (6.9 GB, recommended for CPUs) from `OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF` with `hf download`.
18+
19+ The model is stored in `~/models` inside the sandbox (root overlay: 19 GB free). The workspace is on kDrive: storing 7 GB there would sync it to the cloud, so it is not the default.
20+
21+ ## Acceptance criteria
22+ - The file is in `~/models`; a second run does not download it again.
23+tasks:
24+ - id: 1
25+ title: Download with hf download
26+ state: closed
27+ priority: high
28+ author:
29+ name: k33g
30+ email: ph.charriere@gmail.com
31+ createdAt: 2026-11-21T06:35:00.000Z
32+ updatedAt: 2026-11-21T06:37:13.000Z
33+ closedAt: 2026-11-21T06:37:13.000Z
34+ - id: 2
35+ title: Skip the download when the file already exists
36+ state: closed
37+ priority: medium
38+ author:
39+ name: k33g
40+ email: ph.charriere@gmail.com
41+ createdAt: 2026-11-21T06:35:00.000Z
42+ updatedAt: 2026-11-21T06:37:13.000Z
43+ closedAt: 2026-11-21T06:37:13.000Z
44+comments:
45+ - id: 1
46+ author:
47+ name: k33g
48+ email: ph.charriere@gmail.com
49+ createdAt: 2026-11-21T06:37:13.000Z
50+ body: |
51+ Downloaded in about 2 minutes; second run of `scripts/olala.sh download` skips it (checked).
new file mode 100644
@@ -0,0 +1,51 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 3
3+title: Download Olala-v1.0-mini-7A1B-q8_0-safe2.gguf
4+state: closed
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T06:37:13.000Z
10+labels:
11+ - kind::chore
12+ - priority::high
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ Download the 8-bit file (6.9 GB, recommended for CPUs) from `OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF` with `hf download`.
18+
19+ The model is stored in `~/models` inside the sandbox (root overlay: 19 GB free). The workspace is on kDrive: storing 7 GB there would sync it to the cloud, so it is not the default.
20+
21+ ## Acceptance criteria
22+ - The file is in `~/models`; a second run does not download it again.
23+tasks:
24+ - id: 1
25+ title: Download with hf download
26+ state: closed
27+ priority: high
28+ author:
29+ name: k33g
30+ email: ph.charriere@gmail.com
31+ createdAt: 2026-11-21T06:35:00.000Z
32+ updatedAt: 2026-11-21T06:37:13.000Z
33+ closedAt: 2026-11-21T06:37:13.000Z
34+ - id: 2
35+ title: Skip the download when the file already exists
36+ state: closed
37+ priority: medium
38+ author:
39+ name: k33g
40+ email: ph.charriere@gmail.com
41+ createdAt: 2026-11-21T06:35:00.000Z
42+ updatedAt: 2026-11-21T06:37:13.000Z
43+ closedAt: 2026-11-21T06:37:13.000Z
44+comments:
45+ - id: 1
46+ author:
47+ name: k33g
48+ email: ph.charriere@gmail.com
49+ createdAt: 2026-11-21T06:37:13.000Z
50+ body: |
51+ Downloaded in about 2 minutes; second run of `scripts/olala.sh download` skips it (checked).
added .tickets/issues/0004-serve-the-model-with-llama-server-and-expose-the-port.yaml +67 -0
new file mode 100644
@@ -0,0 +1,67 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 4
3+title: Serve the model with llama-server and expose the port
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T07:08:59.000Z
10+labels:
11+ - kind::feature
12+ - priority::high
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ Run `llama-server` (OpenAI-compatible API) bound on `0.0.0.0:8080` so the port can be published to the host.
18+
19+ CPU flags from the model card: `-t <threads> -fa 1 -nr -ctk q8_0 -ctv q8_0 -c 32768 -np 4`, sampling `--temp 0.8 --top-p 0.9 --top-k 40`.
20+
21+ On the host, for the current sandbox: `sbx ports houla-houla --publish 8080:8080`.
22+
23+ ## Acceptance criteria
24+ - `curl http://localhost:8080/health` answers from the host.
25+ - A chat completion answers.
26+tasks:
27+ - id: 1
28+ title: Start llama-server in the background, wait for /health
29+ state: closed
30+ priority: high
31+ author:
32+ name: k33g
33+ email: ph.charriere@gmail.com
34+ createdAt: 2026-11-21T06:35:00.000Z
35+ updatedAt: 2026-11-21T06:37:13.000Z
36+ closedAt: 2026-11-21T06:37:13.000Z
37+ - id: 2
38+ title: Publish port 8080 on the host (user, host side)
39+ state: open
40+ priority: high
41+ author:
42+ name: k33g
43+ email: ph.charriere@gmail.com
44+ createdAt: 2026-11-21T06:35:00.000Z
45+ - id: 3
46+ title: Test /v1/chat/completions from the host
47+ state: open
48+ priority: high
49+ author:
50+ name: k33g
51+ email: ph.charriere@gmail.com
52+ createdAt: 2026-11-21T06:35:00.000Z
53+comments:
54+ - id: 1
55+ author:
56+ name: k33g
57+ email: ph.charriere@gmail.com
58+ createdAt: 2026-11-21T06:37:13.000Z
59+ body: |
60+ Inside the sandbox: `/health`, `/v1/models` and `/v1/chat/completions` (normal and streaming) answer. Not yet checked from the host: needs `sbx ports houla-houla --publish 8080:8080` on the host.
61+ - id: 2
62+ author:
63+ name: k33g
64+ email: ph.charriere@gmail.com
65+ createdAt: 2026-11-21T07:08:59.000Z
66+ body: |
67+ Measured (q8_0-safe2, prompt ~20 tokens, 150 generated): with the model on the workspace mount, mmap gives ~1.7 tok/s; `--load-mode none` (model read into RAM) gives 35–44 tok/s with 16 threads. 18 threads (= every CPU): ~26 tok/s. Script defaults changed accordingly (`OLALA_LOAD_MODE=none`, threads = nproc − 2). Note: this fork has no `--no-mmap`; use `--load-mode`.
new file mode 100644
@@ -0,0 +1,67 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 4
3+title: Serve the model with llama-server and expose the port
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T07:08:59.000Z
10+labels:
11+ - kind::feature
12+ - priority::high
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ Run `llama-server` (OpenAI-compatible API) bound on `0.0.0.0:8080` so the port can be published to the host.
18+
19+ CPU flags from the model card: `-t <threads> -fa 1 -nr -ctk q8_0 -ctv q8_0 -c 32768 -np 4`, sampling `--temp 0.8 --top-p 0.9 --top-k 40`.
20+
21+ On the host, for the current sandbox: `sbx ports houla-houla --publish 8080:8080`.
22+
23+ ## Acceptance criteria
24+ - `curl http://localhost:8080/health` answers from the host.
25+ - A chat completion answers.
26+tasks:
27+ - id: 1
28+ title: Start llama-server in the background, wait for /health
29+ state: closed
30+ priority: high
31+ author:
32+ name: k33g
33+ email: ph.charriere@gmail.com
34+ createdAt: 2026-11-21T06:35:00.000Z
35+ updatedAt: 2026-11-21T06:37:13.000Z
36+ closedAt: 2026-11-21T06:37:13.000Z
37+ - id: 2
38+ title: Publish port 8080 on the host (user, host side)
39+ state: open
40+ priority: high
41+ author:
42+ name: k33g
43+ email: ph.charriere@gmail.com
44+ createdAt: 2026-11-21T06:35:00.000Z
45+ - id: 3
46+ title: Test /v1/chat/completions from the host
47+ state: open
48+ priority: high
49+ author:
50+ name: k33g
51+ email: ph.charriere@gmail.com
52+ createdAt: 2026-11-21T06:35:00.000Z
53+comments:
54+ - id: 1
55+ author:
56+ name: k33g
57+ email: ph.charriere@gmail.com
58+ createdAt: 2026-11-21T06:37:13.000Z
59+ body: |
60+ Inside the sandbox: `/health`, `/v1/models` and `/v1/chat/completions` (normal and streaming) answer. Not yet checked from the host: needs `sbx ports houla-houla --publish 8080:8080` on the host.
61+ - id: 2
62+ author:
63+ name: k33g
64+ email: ph.charriere@gmail.com
65+ createdAt: 2026-11-21T07:08:59.000Z
66+ body: |
67+ Measured (q8_0-safe2, prompt ~20 tokens, 150 generated): with the model on the workspace mount, mmap gives ~1.7 tok/s; `--load-mode none` (model read into RAM) gives 35–44 tok/s with 16 threads. 18 threads (= every CPU): ~26 tok/s. Script defaults changed accordingly (`OLALA_LOAD_MODE=none`, threads = nproc − 2). Note: this fork has no `--no-mmap`; use `--load-mode`.
added .tickets/issues/0005-idempotent-script-scripts-olala-sh.yaml +66 -0
new file mode 100644
@@ -0,0 +1,66 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 5
3+title: Idempotent script scripts/olala.sh
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T07:08:59.000Z
10+labels:
11+ - kind::feature
12+ - priority::medium
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ One script, run inside the sandbox, with subcommands `setup | download | serve | up | stop | restart | status | logs`.
18+ Every step is skipped when its result exists (binary built, model present, server running).
19+
20+ Configured through env vars so another model can be downloaded and served: `OLALA_MODEL_REPO`, `OLALA_MODEL_FILE`, `OLALA_MODELS_DIR`, `OLALA_PORT`, `OLALA_THREADS`, `OLALA_CTX`, `OLALA_PARALLEL`, `OLALA_EXTRA_ARGS`.
21+
22+ File: [scripts/olala.sh](scripts/olala.sh)
23+tasks:
24+ - id: 1
25+ title: Write the script
26+ state: closed
27+ priority: medium
28+ author:
29+ name: k33g
30+ email: ph.charriere@gmail.com
31+ createdAt: 2026-11-21T06:35:00.000Z
32+ updatedAt: 2026-11-21T06:35:00.000Z
33+ closedAt: 2026-11-21T06:35:00.000Z
34+ - id: 2
35+ title: Run it end to end (second run = all skipped)
36+ state: closed
37+ priority: medium
38+ author:
39+ name: k33g
40+ email: ph.charriere@gmail.com
41+ createdAt: 2026-11-21T06:35:00.000Z
42+ updatedAt: 2026-11-21T06:37:13.000Z
43+ closedAt: 2026-11-21T06:37:13.000Z
44+ - id: 3
45+ title: Serve another model through env vars
46+ state: open
47+ priority: low
48+ author:
49+ name: k33g
50+ email: ph.charriere@gmail.com
51+ createdAt: 2026-11-21T06:35:00.000Z
52+comments:
53+ - id: 1
54+ author:
55+ name: k33g
56+ email: ph.charriere@gmail.com
57+ createdAt: 2026-11-21T06:37:13.000Z
58+ body: |
59+ `up` run twice: the second run skips the three steps. Added `OLALA_ALIAS` (model name exposed by the API, default: file name without `.gguf`).
60+ - id: 2
61+ author:
62+ name: k33g
63+ email: ph.charriere@gmail.com
64+ createdAt: 2026-11-21T07:08:59.000Z
65+ body: |
66+ Fixes: pid file was wrong when `setsid` forks (now written by the process itself before `exec`); `stop` waits for the process to exit; uses `llama-server` from PATH (kit) before building; models dir default `./models` (relative to the repo root); `OLALA_LOAD_MODE`, `OLALA_SERVER_BIN`. Write resets the exec bit: keep `chmod +x`.
new file mode 100644
@@ -0,0 +1,66 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 5
3+title: Idempotent script scripts/olala.sh
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T07:08:59.000Z
10+labels:
11+ - kind::feature
12+ - priority::medium
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ One script, run inside the sandbox, with subcommands `setup | download | serve | up | stop | restart | status | logs`.
18+ Every step is skipped when its result exists (binary built, model present, server running).
19+
20+ Configured through env vars so another model can be downloaded and served: `OLALA_MODEL_REPO`, `OLALA_MODEL_FILE`, `OLALA_MODELS_DIR`, `OLALA_PORT`, `OLALA_THREADS`, `OLALA_CTX`, `OLALA_PARALLEL`, `OLALA_EXTRA_ARGS`.
21+
22+ File: [scripts/olala.sh](scripts/olala.sh)
23+tasks:
24+ - id: 1
25+ title: Write the script
26+ state: closed
27+ priority: medium
28+ author:
29+ name: k33g
30+ email: ph.charriere@gmail.com
31+ createdAt: 2026-11-21T06:35:00.000Z
32+ updatedAt: 2026-11-21T06:35:00.000Z
33+ closedAt: 2026-11-21T06:35:00.000Z
34+ - id: 2
35+ title: Run it end to end (second run = all skipped)
36+ state: closed
37+ priority: medium
38+ author:
39+ name: k33g
40+ email: ph.charriere@gmail.com
41+ createdAt: 2026-11-21T06:35:00.000Z
42+ updatedAt: 2026-11-21T06:37:13.000Z
43+ closedAt: 2026-11-21T06:37:13.000Z
44+ - id: 3
45+ title: Serve another model through env vars
46+ state: open
47+ priority: low
48+ author:
49+ name: k33g
50+ email: ph.charriere@gmail.com
51+ createdAt: 2026-11-21T06:35:00.000Z
52+comments:
53+ - id: 1
54+ author:
55+ name: k33g
56+ email: ph.charriere@gmail.com
57+ createdAt: 2026-11-21T06:37:13.000Z
58+ body: |
59+ `up` run twice: the second run skips the three steps. Added `OLALA_ALIAS` (model name exposed by the API, default: file name without `.gguf`).
60+ - id: 2
61+ author:
62+ name: k33g
63+ email: ph.charriere@gmail.com
64+ createdAt: 2026-11-21T07:08:59.000Z
65+ body: |
66+ Fixes: pid file was wrong when `setsid` forks (now written by the process itself before `exec`); `stop` waits for the process to exit; uses `llama-server` from PATH (kit) before building; models dir default `./models` (relative to the repo root); `OLALA_LOAD_MODE`, `OLALA_SERVER_BIN`. Write resets the exec bit: keep `chmod +x`.
added .tickets/issues/0006-environment-file-sbxenv-yaml.yaml +72 -0
new file mode 100644
@@ -0,0 +1,72 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 6
3+title: Environment file sbxenv.yaml
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T07:52:58.000Z
10+labels:
11+ - kind::feature
12+ - priority::medium
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ An sbx environment file to recreate the whole setup in one command: `sbx env run sbxenv.yaml`.
18+
19+ - `args`: model repo, model file, host port, threads → `env` of the sandbox.
20+ - `ports`: sandbox 8080 → host port.
21+ - `lifecycle.postCreate` (runs on the host): `sbx env exec … -- scripts/olala.sh up`.
22+ - Restart after a re-attach (postCreate does not run then): `sbx env exec sbxenv.yaml -- scripts/olala.sh serve`.
23+
24+ Traps: env files are mounted read-only in the sandbox once used; lifecycle commands do not receive the file's `env` values; the `sbx env exec` syntax inside a lifecycle hook is not verified yet.
25+tasks:
26+ - id: 1
27+ title: Write sbxenv.yaml
28+ state: closed
29+ priority: medium
30+ author:
31+ name: k33g
32+ email: ph.charriere@gmail.com
33+ createdAt: 2026-11-21T06:35:00.000Z
34+ updatedAt: 2026-11-21T07:52:58.000Z
35+ closedAt: 2026-11-21T06:37:13.000Z
36+ - id: 2
37+ title: Try it on the host (sbx env create)
38+ state: open
39+ priority: medium
40+ author:
41+ name: k33g
42+ email: ph.charriere@gmail.com
43+ createdAt: 2026-11-21T06:35:00.000Z
44+comments:
45+ - id: 1
46+ author:
47+ name: k33g
48+ email: ph.charriere@gmail.com
49+ createdAt: 2026-11-21T06:37:13.000Z
50+ body: |
51+ Written: [sbxenv.houla.yaml](sbxenv.houla.yaml). Not tried with sbx (no sbx in the sandbox). Unverified: `sbx env exec` from a postCreate hook, `--env-arg` substitution in `ports.host`, server surviving the end of the `sbx env exec` session.
52+ - id: 2
53+ author:
54+ name: k33g
55+ email: ph.charriere@gmail.com
56+ createdAt: 2026-11-21T07:23:50.000Z
57+ body: |
58+ Second host run: the kit built, the sandbox was created, then postCreate failed: `no sbxenv.yaml found at …/houla/--name/sbxenv.yaml`. `sbx env exec` reads every word after the first path, up to `--`, as another path: flags must come first. Now `sbx env exec --name "$SBX_SANDBOX_NAME" --no-input "$SBX_ENV_FILE" -- …`.
59+ - id: 3
60+ author:
61+ name: k33g
62+ email: ph.charriere@gmail.com
63+ createdAt: 2026-11-21T07:40:26.000Z
64+ body: |
65+ Third host run: postCreate OK (`ready`), but curl got nothing: the next `sbx env exec` printed `sandbox houla-olala — started` and `olala.sh status` said `not running`. Per the sbx docs (usage), a sandbox created with `sbx create` stops when no session keeps it running; `sbx run -d --name <name>` makes it detached (permanent) so it keeps running. Added it as the first postCreate command. Not verified yet.
66+ - id: 4
67+ author:
68+ name: k33g
69+ email: ph.charriere@gmail.com
70+ createdAt: 2026-11-21T07:52:58.000Z
71+ body: |
72+ Renamed: this file is now [sbxenv.yaml](sbxenv.yaml) (was `sbxenv.houla.yaml`); the former `sbxenv.yaml` (dev sandbox `houla-houla`) is now [sbxenv.dev.yaml](sbxenv.dev.yaml). Older comments keep the old name. The last step, `mv sbxenv.houla.yaml sbxenv.yaml`, has to be run on the host (read-only mount inside the sandbox).
new file mode 100644
@@ -0,0 +1,72 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 6
3+title: Environment file sbxenv.yaml
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T07:52:58.000Z
10+labels:
11+ - kind::feature
12+ - priority::medium
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ An sbx environment file to recreate the whole setup in one command: `sbx env run sbxenv.yaml`.
18+
19+ - `args`: model repo, model file, host port, threads → `env` of the sandbox.
20+ - `ports`: sandbox 8080 → host port.
21+ - `lifecycle.postCreate` (runs on the host): `sbx env exec … -- scripts/olala.sh up`.
22+ - Restart after a re-attach (postCreate does not run then): `sbx env exec sbxenv.yaml -- scripts/olala.sh serve`.
23+
24+ Traps: env files are mounted read-only in the sandbox once used; lifecycle commands do not receive the file's `env` values; the `sbx env exec` syntax inside a lifecycle hook is not verified yet.
25+tasks:
26+ - id: 1
27+ title: Write sbxenv.yaml
28+ state: closed
29+ priority: medium
30+ author:
31+ name: k33g
32+ email: ph.charriere@gmail.com
33+ createdAt: 2026-11-21T06:35:00.000Z
34+ updatedAt: 2026-11-21T07:52:58.000Z
35+ closedAt: 2026-11-21T06:37:13.000Z
36+ - id: 2
37+ title: Try it on the host (sbx env create)
38+ state: open
39+ priority: medium
40+ author:
41+ name: k33g
42+ email: ph.charriere@gmail.com
43+ createdAt: 2026-11-21T06:35:00.000Z
44+comments:
45+ - id: 1
46+ author:
47+ name: k33g
48+ email: ph.charriere@gmail.com
49+ createdAt: 2026-11-21T06:37:13.000Z
50+ body: |
51+ Written: [sbxenv.houla.yaml](sbxenv.houla.yaml). Not tried with sbx (no sbx in the sandbox). Unverified: `sbx env exec` from a postCreate hook, `--env-arg` substitution in `ports.host`, server surviving the end of the `sbx env exec` session.
52+ - id: 2
53+ author:
54+ name: k33g
55+ email: ph.charriere@gmail.com
56+ createdAt: 2026-11-21T07:23:50.000Z
57+ body: |
58+ Second host run: the kit built, the sandbox was created, then postCreate failed: `no sbxenv.yaml found at …/houla/--name/sbxenv.yaml`. `sbx env exec` reads every word after the first path, up to `--`, as another path: flags must come first. Now `sbx env exec --name "$SBX_SANDBOX_NAME" --no-input "$SBX_ENV_FILE" -- …`.
59+ - id: 3
60+ author:
61+ name: k33g
62+ email: ph.charriere@gmail.com
63+ createdAt: 2026-11-21T07:40:26.000Z
64+ body: |
65+ Third host run: postCreate OK (`ready`), but curl got nothing: the next `sbx env exec` printed `sandbox houla-olala — started` and `olala.sh status` said `not running`. Per the sbx docs (usage), a sandbox created with `sbx create` stops when no session keeps it running; `sbx run -d --name <name>` makes it detached (permanent) so it keeps running. Added it as the first postCreate command. Not verified yet.
66+ - id: 4
67+ author:
68+ name: k33g
69+ email: ph.charriere@gmail.com
70+ createdAt: 2026-11-21T07:52:58.000Z
71+ body: |
72+ Renamed: this file is now [sbxenv.yaml](sbxenv.yaml) (was `sbxenv.houla.yaml`); the former `sbxenv.yaml` (dev sandbox `houla-houla`) is now [sbxenv.dev.yaml](sbxenv.dev.yaml). Older comments keep the old name. The last step, `mv sbxenv.houla.yaml sbxenv.yaml`, has to be run on the host (read-only mount inside the sandbox).
added .tickets/issues/0007-package-the-llama-cpp-fork-as-an-sbx-kit.yaml +83 -0
new file mode 100644
@@ -0,0 +1,83 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 7
3+title: Package the llama.cpp fork as an sbx kit
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T07:52:58.000Z
10+labels:
11+ - kind::feature
12+ - priority::medium
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ Instead of building the fork at every sandbox creation (several minutes), ship it as a **kit v3 mixin** in this project (e.g. `kits/llama-cpp-olala/`), on the model of `go-toolchain` in mini-me:
18+
19+ - `llama-cpp-olala.yaml`: `schemaVersion: "3"`, `kind: mixin`, arg `ref` (fork commit, `buildArg`), `provides`, `network-policy@1` allowing `huggingface.co`, `*.huggingface.co`, `*.hf.co`.
20+ - `llama-cpp-olala.dockerfile`: build stage (debian trixie, cmake, CPU only, `-DLLAMA_USE_PREBUILT_UI=OFF`, multi-arch `TARGETARCH`), then `FROM scratch` overlay with `llama-server`, `llama-cli` and their `.so` (or a static build `-DBUILD_SHARED_LIBS=OFF`).
21+ - Reference it from `sbxenv.yaml` `kits:` (local path first, then Docker Hub).
22+
23+ ## Acceptance criteria
24+ - A new sandbox has `llama-server` on the PATH without building anything.
25+ - `scripts/olala.sh setup` detects it and skips the build.
26+tasks:
27+ - id: 1
28+ title: Write the kit descriptor and Dockerfile
29+ state: closed
30+ priority: medium
31+ author:
32+ name: k33g
33+ email: ph.charriere@gmail.com
34+ createdAt: 2026-11-21T06:35:00.000Z
35+ updatedAt: 2026-11-21T07:08:59.000Z
36+ closedAt: 2026-11-21T07:08:59.000Z
37+ - id: 2
38+ title: Build statically (or copy the .so) and check llama-server runs
39+ state: closed
40+ priority: medium
41+ author:
42+ name: k33g
43+ email: ph.charriere@gmail.com
44+ createdAt: 2026-11-21T06:35:00.000Z
45+ updatedAt: 2026-11-21T07:08:59.000Z
46+ closedAt: 2026-11-21T07:08:59.000Z
47+ - id: 3
48+ title: Use the kit from sbxenv.yaml
49+ state: closed
50+ priority: medium
51+ author:
52+ name: k33g
53+ email: ph.charriere@gmail.com
54+ createdAt: 2026-11-21T06:35:00.000Z
55+ updatedAt: 2026-11-21T07:52:58.000Z
56+ closedAt: 2026-11-21T07:08:59.000Z
57+ - id: 4
58+ title: Publish it on Docker Hub (optional)
59+ state: open
60+ priority: low
61+ author:
62+ name: k33g
63+ email: ph.charriere@gmail.com
64+ createdAt: 2026-11-21T06:35:00.000Z
65+comments:
66+ - id: 1
67+ author:
68+ name: k33g
69+ email: ph.charriere@gmail.com
70+ createdAt: 2026-11-21T07:08:59.000Z
71+ body: |
72+ Done: [kits/llama-cpp-olala/](kits/llama-cpp-olala/llama-cpp-olala.yaml) — v3 mixin, arg `ref` (full SHA of the fork), portable build (`GGML_NATIVE=OFF`, `GGML_BACKEND_DL=ON`, `GGML_CPU_ALL_VARIANTS=ON`, `LLAMA_OPENSSL=OFF`, RPATH `$ORIGIN`), installed in `/opt/llama-cpp-olala` + links in `/usr/local/bin`, `libgomp.so.1` shipped. Network policy: Hugging Face hosts; agent context file.
73+
74+ Tested: `docker buildx build` of the Dockerfile (1m22, 26 MB overlay, arm64); `llama-server --version` in a bare `debian:trixie-slim`; installed in this sandbox and served by `scripts/olala.sh` (35–44 tok/s, 16 threads, `--load-mode none`).
75+ Not tested: sbx building the kit itself (`sbx env create sbxenv.houla.yaml`), amd64.
76+ - id: 2
77+ author:
78+ name: k33g
79+ email: ph.charriere@gmail.com
80+ createdAt: 2026-11-21T07:13:05.000Z
81+ body: |
82+ First `sbx env create sbxenv.houla.yaml` on the host: the kit frontend rejected `provides: ["llama-cpp-olala@<sha>"]` (`invalid version`). A provides version must be `X.Y.Z`: now `llama-cpp-olala@0.5.0` (what the fork reports); the commit stays pinned by `ref`.
83+ Checked with the same frontend: `docker buildx build -f kits/llama-cpp-olala/llama-cpp-olala.yaml .` passes (overlay exported).
new file mode 100644
@@ -0,0 +1,83 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 7
3+title: Package the llama.cpp fork as an sbx kit
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T07:52:58.000Z
10+labels:
11+ - kind::feature
12+ - priority::medium
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ Instead of building the fork at every sandbox creation (several minutes), ship it as a **kit v3 mixin** in this project (e.g. `kits/llama-cpp-olala/`), on the model of `go-toolchain` in mini-me:
18+
19+ - `llama-cpp-olala.yaml`: `schemaVersion: "3"`, `kind: mixin`, arg `ref` (fork commit, `buildArg`), `provides`, `network-policy@1` allowing `huggingface.co`, `*.huggingface.co`, `*.hf.co`.
20+ - `llama-cpp-olala.dockerfile`: build stage (debian trixie, cmake, CPU only, `-DLLAMA_USE_PREBUILT_UI=OFF`, multi-arch `TARGETARCH`), then `FROM scratch` overlay with `llama-server`, `llama-cli` and their `.so` (or a static build `-DBUILD_SHARED_LIBS=OFF`).
21+ - Reference it from `sbxenv.yaml` `kits:` (local path first, then Docker Hub).
22+
23+ ## Acceptance criteria
24+ - A new sandbox has `llama-server` on the PATH without building anything.
25+ - `scripts/olala.sh setup` detects it and skips the build.
26+tasks:
27+ - id: 1
28+ title: Write the kit descriptor and Dockerfile
29+ state: closed
30+ priority: medium
31+ author:
32+ name: k33g
33+ email: ph.charriere@gmail.com
34+ createdAt: 2026-11-21T06:35:00.000Z
35+ updatedAt: 2026-11-21T07:08:59.000Z
36+ closedAt: 2026-11-21T07:08:59.000Z
37+ - id: 2
38+ title: Build statically (or copy the .so) and check llama-server runs
39+ state: closed
40+ priority: medium
41+ author:
42+ name: k33g
43+ email: ph.charriere@gmail.com
44+ createdAt: 2026-11-21T06:35:00.000Z
45+ updatedAt: 2026-11-21T07:08:59.000Z
46+ closedAt: 2026-11-21T07:08:59.000Z
47+ - id: 3
48+ title: Use the kit from sbxenv.yaml
49+ state: closed
50+ priority: medium
51+ author:
52+ name: k33g
53+ email: ph.charriere@gmail.com
54+ createdAt: 2026-11-21T06:35:00.000Z
55+ updatedAt: 2026-11-21T07:52:58.000Z
56+ closedAt: 2026-11-21T07:08:59.000Z
57+ - id: 4
58+ title: Publish it on Docker Hub (optional)
59+ state: open
60+ priority: low
61+ author:
62+ name: k33g
63+ email: ph.charriere@gmail.com
64+ createdAt: 2026-11-21T06:35:00.000Z
65+comments:
66+ - id: 1
67+ author:
68+ name: k33g
69+ email: ph.charriere@gmail.com
70+ createdAt: 2026-11-21T07:08:59.000Z
71+ body: |
72+ Done: [kits/llama-cpp-olala/](kits/llama-cpp-olala/llama-cpp-olala.yaml) — v3 mixin, arg `ref` (full SHA of the fork), portable build (`GGML_NATIVE=OFF`, `GGML_BACKEND_DL=ON`, `GGML_CPU_ALL_VARIANTS=ON`, `LLAMA_OPENSSL=OFF`, RPATH `$ORIGIN`), installed in `/opt/llama-cpp-olala` + links in `/usr/local/bin`, `libgomp.so.1` shipped. Network policy: Hugging Face hosts; agent context file.
73+
74+ Tested: `docker buildx build` of the Dockerfile (1m22, 26 MB overlay, arm64); `llama-server --version` in a bare `debian:trixie-slim`; installed in this sandbox and served by `scripts/olala.sh` (35–44 tok/s, 16 threads, `--load-mode none`).
75+ Not tested: sbx building the kit itself (`sbx env create sbxenv.houla.yaml`), amd64.
76+ - id: 2
77+ author:
78+ name: k33g
79+ email: ph.charriere@gmail.com
80+ createdAt: 2026-11-21T07:13:05.000Z
81+ body: |
82+ First `sbx env create sbxenv.houla.yaml` on the host: the kit frontend rejected `provides: ["llama-cpp-olala@<sha>"]` (`invalid version`). A provides version must be `X.Y.Z`: now `llama-cpp-olala@0.5.0` (what the fork reports); the commit stays pinned by `ref`.
83+ Checked with the same frontend: `docker buildx build -f kits/llama-cpp-olala/llama-cpp-olala.yaml .` passes (overlay exported).
added .tickets/issues/0008-document-the-setup-and-the-curl-test-commands.yaml +61 -0
new file mode 100644
@@ -0,0 +1,61 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 8
3+title: Document the setup and the curl test commands
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T07:08:59.000Z
10+labels:
11+ - documentation
12+ - priority::medium
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ How-to (FR + EN) in `docs/`: create the environment, check the server, curl commands (`/health`, `/v1/models`, `/v1/chat/completions` with `reasoning_effort`, streaming), change model, restart after re-attach.
18+tasks:
19+ - id: 1
20+ title: curl commands tested from the host
21+ state: open
22+ priority: medium
23+ author:
24+ name: k33g
25+ email: ph.charriere@gmail.com
26+ createdAt: 2026-11-21T06:35:00.000Z
27+ - id: 2
28+ title: How-to FR
29+ state: closed
30+ priority: medium
31+ author:
32+ name: k33g
33+ email: ph.charriere@gmail.com
34+ createdAt: 2026-11-21T06:35:00.000Z
35+ updatedAt: 2026-11-21T06:37:13.000Z
36+ closedAt: 2026-11-21T06:37:13.000Z
37+ - id: 3
38+ title: How-to EN
39+ state: closed
40+ priority: medium
41+ author:
42+ name: k33g
43+ email: ph.charriere@gmail.com
44+ createdAt: 2026-11-21T06:35:00.000Z
45+ updatedAt: 2026-11-21T06:37:13.000Z
46+ closedAt: 2026-11-21T06:37:13.000Z
47+comments:
48+ - id: 1
49+ author:
50+ name: k33g
51+ email: ph.charriere@gmail.com
52+ createdAt: 2026-11-21T06:37:13.000Z
53+ body: |
54+ Written: [docs/olala-sbx.md](docs/olala-sbx.md) (FR) and [docs/olala-sbx.en.md](docs/olala-sbx.en.md) (EN). curl commands tested inside the sandbox only.
55+ - id: 2
56+ author:
57+ name: k33g
58+ email: ph.charriere@gmail.com
59+ createdAt: 2026-11-21T07:08:59.000Z
60+ body: |
61+ Docs updated for the kit, `./models`, `models_dir`, load mode and threads.
new file mode 100644
@@ -0,0 +1,61 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 8
3+title: Document the setup and the curl test commands
4+state: open
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T06:35:00.000Z
9+updatedAt: 2026-11-21T07:08:59.000Z
10+labels:
11+ - documentation
12+ - priority::medium
13+epic: Olala in sbx
14+related:
15+ - 1
16+body: |
17+ How-to (FR + EN) in `docs/`: create the environment, check the server, curl commands (`/health`, `/v1/models`, `/v1/chat/completions` with `reasoning_effort`, streaming), change model, restart after re-attach.
18+tasks:
19+ - id: 1
20+ title: curl commands tested from the host
21+ state: open
22+ priority: medium
23+ author:
24+ name: k33g
25+ email: ph.charriere@gmail.com
26+ createdAt: 2026-11-21T06:35:00.000Z
27+ - id: 2
28+ title: How-to FR
29+ state: closed
30+ priority: medium
31+ author:
32+ name: k33g
33+ email: ph.charriere@gmail.com
34+ createdAt: 2026-11-21T06:35:00.000Z
35+ updatedAt: 2026-11-21T06:37:13.000Z
36+ closedAt: 2026-11-21T06:37:13.000Z
37+ - id: 3
38+ title: How-to EN
39+ state: closed
40+ priority: medium
41+ author:
42+ name: k33g
43+ email: ph.charriere@gmail.com
44+ createdAt: 2026-11-21T06:35:00.000Z
45+ updatedAt: 2026-11-21T06:37:13.000Z
46+ closedAt: 2026-11-21T06:37:13.000Z
47+comments:
48+ - id: 1
49+ author:
50+ name: k33g
51+ email: ph.charriere@gmail.com
52+ createdAt: 2026-11-21T06:37:13.000Z
53+ body: |
54+ Written: [docs/olala-sbx.md](docs/olala-sbx.md) (FR) and [docs/olala-sbx.en.md](docs/olala-sbx.en.md) (EN). curl commands tested inside the sandbox only.
55+ - id: 2
56+ author:
57+ name: k33g
58+ email: ph.charriere@gmail.com
59+ createdAt: 2026-11-21T07:08:59.000Z
60+ body: |
61+ Docs updated for the kit, `./models`, `models_dir`, load mode and threads.
added .tickets/issues/0009-store-models-in-a-configurable-repository-folder.yaml +65 -0
new file mode 100644
@@ -0,0 +1,65 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 9
3+title: Store models in a configurable repository folder (default ./models)
4+state: closed
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T07:08:59.000Z
9+updatedAt: 2026-11-21T07:52:58.000Z
10+labels:
11+ - kind::feature
12+ - priority::high
13+epic: Olala in sbx
14+related:
15+ - 3
16+ - 5
17+body: |
18+ Store downloaded models in a folder of the repository, so dropping the sandbox does not mean downloading them again.
19+
20+ - Default: `./models` (relative to the repo root), configurable with `OLALA_MODELS_DIR` / the `models_dir` arg of [sbxenv.yaml](sbxenv.yaml); absolute paths accepted.
21+ - `.gitignore`: `/models/`, `*.gguf`, `**/.cache/huggingface/` (the `hf download --local-dir` cache).
22+ - The workspace is on kDrive: the models are synced by kDrive unless `models/` is excluded there.
23+ - Serve with `--load-mode none`: mmap over the workspace mount is ~20x slower (see #4).
24+tasks:
25+ - id: 1
26+ title: Default OLALA_MODELS_DIR to ./models, relative to the repo root
27+ state: closed
28+ priority: high
29+ author:
30+ name: k33g
31+ email: ph.charriere@gmail.com
32+ createdAt: 2026-11-21T07:08:59.000Z
33+ updatedAt: 2026-11-21T07:08:59.000Z
34+ closedAt: 2026-11-21T07:08:59.000Z
35+ - id: 2
36+ title: models_dir arg in sbxenv.yaml
37+ state: closed
38+ priority: high
39+ author:
40+ name: k33g
41+ email: ph.charriere@gmail.com
42+ createdAt: 2026-11-21T07:08:59.000Z
43+ updatedAt: 2026-11-21T07:52:58.000Z
44+ closedAt: 2026-11-21T07:08:59.000Z
45+ - id: 3
46+ title: Update .gitignore
47+ state: closed
48+ priority: high
49+ author:
50+ name: k33g
51+ email: ph.charriere@gmail.com
52+ createdAt: 2026-11-21T07:08:59.000Z
53+ updatedAt: 2026-11-21T07:08:59.000Z
54+ closedAt: 2026-11-21T07:08:59.000Z
55+ - id: 4
56+ title: Move the downloaded model to ./models
57+ state: closed
58+ priority: high
59+ author:
60+ name: k33g
61+ email: ph.charriere@gmail.com
62+ createdAt: 2026-11-21T07:08:59.000Z
63+ updatedAt: 2026-11-21T07:08:59.000Z
64+ closedAt: 2026-11-21T07:08:59.000Z
65+comments: []
new file mode 100644
@@ -0,0 +1,65 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+id: 9
3+title: Store models in a configurable repository folder (default ./models)
4+state: closed
5+author:
6+ name: k33g
7+ email: ph.charriere@gmail.com
8+createdAt: 2026-11-21T07:08:59.000Z
9+updatedAt: 2026-11-21T07:52:58.000Z
10+labels:
11+ - kind::feature
12+ - priority::high
13+epic: Olala in sbx
14+related:
15+ - 3
16+ - 5
17+body: |
18+ Store downloaded models in a folder of the repository, so dropping the sandbox does not mean downloading them again.
19+
20+ - Default: `./models` (relative to the repo root), configurable with `OLALA_MODELS_DIR` / the `models_dir` arg of [sbxenv.yaml](sbxenv.yaml); absolute paths accepted.
21+ - `.gitignore`: `/models/`, `*.gguf`, `**/.cache/huggingface/` (the `hf download --local-dir` cache).
22+ - The workspace is on kDrive: the models are synced by kDrive unless `models/` is excluded there.
23+ - Serve with `--load-mode none`: mmap over the workspace mount is ~20x slower (see #4).
24+tasks:
25+ - id: 1
26+ title: Default OLALA_MODELS_DIR to ./models, relative to the repo root
27+ state: closed
28+ priority: high
29+ author:
30+ name: k33g
31+ email: ph.charriere@gmail.com
32+ createdAt: 2026-11-21T07:08:59.000Z
33+ updatedAt: 2026-11-21T07:08:59.000Z
34+ closedAt: 2026-11-21T07:08:59.000Z
35+ - id: 2
36+ title: models_dir arg in sbxenv.yaml
37+ state: closed
38+ priority: high
39+ author:
40+ name: k33g
41+ email: ph.charriere@gmail.com
42+ createdAt: 2026-11-21T07:08:59.000Z
43+ updatedAt: 2026-11-21T07:52:58.000Z
44+ closedAt: 2026-11-21T07:08:59.000Z
45+ - id: 3
46+ title: Update .gitignore
47+ state: closed
48+ priority: high
49+ author:
50+ name: k33g
51+ email: ph.charriere@gmail.com
52+ createdAt: 2026-11-21T07:08:59.000Z
53+ updatedAt: 2026-11-21T07:08:59.000Z
54+ closedAt: 2026-11-21T07:08:59.000Z
55+ - id: 4
56+ title: Move the downloaded model to ./models
57+ state: closed
58+ priority: high
59+ author:
60+ name: k33g
61+ email: ph.charriere@gmail.com
62+ createdAt: 2026-11-21T07:08:59.000Z
63+ updatedAt: 2026-11-21T07:08:59.000Z
64+ closedAt: 2026-11-21T07:08:59.000Z
65+comments: []
added .tickets/labels.yaml +38 -0
new file mode 100644
@@ -0,0 +1,38 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+labels:
3+ - name: bug
4+ color: "#d73a4a"
5+ description: Something is not working
6+ - name: documentation
7+ color: "#0075ca"
8+ description: Improvements or additions to documentation
9+ - name: enhancement
10+ color: "#a2eeef"
11+ description: New feature or request
12+ - name: question
13+ color: "#d876e3"
14+ description: Further information is requested
15+ - name: wontfix
16+ color: "#ffffff"
17+ description: This will not be worked on
18+ - name: priority::low
19+ color: "#c5def5"
20+ description: Can wait
21+ - name: priority::medium
22+ color: "#fbca04"
23+ description: Normal priority
24+ - name: priority::high
25+ color: "#d93f0b"
26+ description: Should be picked up next
27+ - name: priority::urgent
28+ color: "#b60205"
29+ description: Drop everything
30+ - name: kind::bug
31+ color: "#d73a4a"
32+ description: Something is broken
33+ - name: kind::feature
34+ color: "#0e8a16"
35+ description: Something new
36+ - name: kind::chore
37+ color: "#bfdadc"
38+ description: Maintenance, no visible change
new file mode 100644
@@ -0,0 +1,38 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+labels:
3+ - name: bug
4+ color: "#d73a4a"
5+ description: Something is not working
6+ - name: documentation
7+ color: "#0075ca"
8+ description: Improvements or additions to documentation
9+ - name: enhancement
10+ color: "#a2eeef"
11+ description: New feature or request
12+ - name: question
13+ color: "#d876e3"
14+ description: Further information is requested
15+ - name: wontfix
16+ color: "#ffffff"
17+ description: This will not be worked on
18+ - name: priority::low
19+ color: "#c5def5"
20+ description: Can wait
21+ - name: priority::medium
22+ color: "#fbca04"
23+ description: Normal priority
24+ - name: priority::high
25+ color: "#d93f0b"
26+ description: Should be picked up next
27+ - name: priority::urgent
28+ color: "#b60205"
29+ description: Drop everything
30+ - name: kind::bug
31+ color: "#d73a4a"
32+ description: Something is broken
33+ - name: kind::feature
34+ color: "#0e8a16"
35+ description: Something new
36+ - name: kind::chore
37+ color: "#bfdadc"
38+ description: Maintenance, no visible change
added .tickets/milestones.yaml +2 -0
new file mode 100644
@@ -0,0 +1,2 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+milestones: []
new file mode 100644
@@ -0,0 +1,2 @@
1+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2+milestones: []
added .vscode/extensions.json +9 -0
new file mode 100644
@@ -0,0 +1,9 @@
1+{
2+ "recommendations": [
3+ "ms-azuretools.vscode-docker",
4+ "pkief.material-icon-theme",
5+ "pkief.material-product-icons",
6+ "aaron-bond.better-comments",
7+ "bierner.markdown-mermaid",
8+ ]
9+}
\ No newline at end of file
new file mode 100644
@@ -0,0 +1,9 @@
1+{
2+ "recommendations": [
3+ "ms-azuretools.vscode-docker",
4+ "pkief.material-icon-theme",
5+ "pkief.material-product-icons",
6+ "aaron-bond.better-comments",
7+ "bierner.markdown-mermaid",
8+ ]
9+}
\ No newline at end of file\ No newline at end of file
added .vscode/settings.json +63 -0
new file mode 100644
@@ -0,0 +1,63 @@
1+{
2+ "workbench.iconTheme": "material-icon-theme",
3+ "workbench.colorTheme": "Light Modern",
4+ "editor.fontSize": 14,
5+ "terminal.integrated.fontSize": 14,
6+ "editor.insertSpaces": true,
7+ "editor.tabSize": 4,
8+ "editor.detectIndentation": true,
9+ "files.autoSave": "afterDelay",
10+ "files.autoSaveDelay": 1000,
11+ // "editor.defaultFormatter": "esbenp.prettier-vscode",
12+ "editor.formatOnSave": true,
13+ "workbench.tree.indent": 20,
14+ //"workbench.activityBar.location": "top",
15+ "workbench.editor.showTabs": "multiple",
16+ "window.zoomLevel": 0.0,
17+ "[markdown]": {
18+ "editor.unicodeHighlight.ambiguousCharacters": false,
19+ "editor.unicodeHighlight.invisibleCharacters": false,
20+ "diffEditor.ignoreTrimWhitespace": false,
21+ "editor.fontWeight": "normal",
22+ "editor.fontFamily": "'Droid Sans Mono', 'monospace', monospace",
23+ "editor.fontSize": 14,
24+ "editor.wordWrap": "on",
25+ "editor.quickSuggestions": {
26+ "comments": "off",
27+ "strings": "off",
28+ "other": "off"
29+ }
30+ },
31+ "markdown.preview.fontSize": 14,
32+ // "workbench.editorAssociations": {
33+ // "*.md": "vscode.markdown.preview.editor"
34+ // },
35+ "markdown.marp.html": "all",
36+ "[dockerfile]": {
37+ "editor.fontSize": 14
38+ },
39+ "[dockercompose]": {
40+ "editor.fontSize": 14
41+ },
42+ "[json]": {
43+ "editor.fontSize": 14
44+ },
45+ "[yaml]": {
46+ "editor.fontSize": 14
47+ },
48+ "[go]": {
49+ "editor.fontSize": 14,
50+ "editor.codeActionsOnSave": {
51+ "source.organizeImports": "explicit"
52+ }
53+ },
54+ "go.lintTool": "golangci-lint",
55+ "go.lintOnSave": "package",
56+ "go.formatTool": "goimports",
57+ "go.useLanguageServer": true,
58+ "gopls": {
59+ "ui.semanticTokens": true,
60+ "ui.completion.usePlaceholders": true
61+ },
62+ "issuespec.groupBy": "milestone"
63+}
\ No newline at end of file
new file mode 100644
@@ -0,0 +1,63 @@
1+{
2+ "workbench.iconTheme": "material-icon-theme",
3+ "workbench.colorTheme": "Light Modern",
4+ "editor.fontSize": 14,
5+ "terminal.integrated.fontSize": 14,
6+ "editor.insertSpaces": true,
7+ "editor.tabSize": 4,
8+ "editor.detectIndentation": true,
9+ "files.autoSave": "afterDelay",
10+ "files.autoSaveDelay": 1000,
11+ // "editor.defaultFormatter": "esbenp.prettier-vscode",
12+ "editor.formatOnSave": true,
13+ "workbench.tree.indent": 20,
14+ //"workbench.activityBar.location": "top",
15+ "workbench.editor.showTabs": "multiple",
16+ "window.zoomLevel": 0.0,
17+ "[markdown]": {
18+ "editor.unicodeHighlight.ambiguousCharacters": false,
19+ "editor.unicodeHighlight.invisibleCharacters": false,
20+ "diffEditor.ignoreTrimWhitespace": false,
21+ "editor.fontWeight": "normal",
22+ "editor.fontFamily": "'Droid Sans Mono', 'monospace', monospace",
23+ "editor.fontSize": 14,
24+ "editor.wordWrap": "on",
25+ "editor.quickSuggestions": {
26+ "comments": "off",
27+ "strings": "off",
28+ "other": "off"
29+ }
30+ },
31+ "markdown.preview.fontSize": 14,
32+ // "workbench.editorAssociations": {
33+ // "*.md": "vscode.markdown.preview.editor"
34+ // },
35+ "markdown.marp.html": "all",
36+ "[dockerfile]": {
37+ "editor.fontSize": 14
38+ },
39+ "[dockercompose]": {
40+ "editor.fontSize": 14
41+ },
42+ "[json]": {
43+ "editor.fontSize": 14
44+ },
45+ "[yaml]": {
46+ "editor.fontSize": 14
47+ },
48+ "[go]": {
49+ "editor.fontSize": 14,
50+ "editor.codeActionsOnSave": {
51+ "source.organizeImports": "explicit"
52+ }
53+ },
54+ "go.lintTool": "golangci-lint",
55+ "go.lintOnSave": "package",
56+ "go.formatTool": "goimports",
57+ "go.useLanguageServer": true,
58+ "gopls": {
59+ "ui.semanticTokens": true,
60+ "ui.completion.usePlaceholders": true
61+ },
62+ "issuespec.groupBy": "milestone"
63+}
\ No newline at end of file\ No newline at end of file
added README.fr.md +110 -0
new file mode 100644
@@ -0,0 +1,110 @@
1+# Houla ... Houla
2+
3+🇬🇧 [English version](README.md)
4+
5+## 1. C'est quoi ce projet ?
6+
7+Houla fait tourner le modèle de langage **Olala-v1.0-mini-7A1B** (d'OVHai) sur ta machine, dans un sandbox Docker (`sbx`), et l'expose sous forme d'**API compatible OpenAI** sur `http://localhost:8080`.
8+
9+- Le modèle est servi par `llama-server`, issu du [fork llama.cpp d'OVHai](https://github.com/OVHai-LLM/llama.cpp) (CPU uniquement, pas besoin de GPU).
10+- Le fichier du modèle (GGUF, ~7 Go) est téléchargé une seule fois dans le dossier `./models`, puis réutilisé.
11+- Tout est décrit dans [`sbxenv.yaml`](sbxenv.yaml) : une seule commande crée le sandbox, télécharge le modèle et démarre le serveur.
12+
13+## 2. Créer le secret Hugging Face
14+
15+**Pourquoi ?** Le modèle est téléchargé depuis [Hugging Face](https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF). Avec un token, les téléchargements sont authentifiés (moins de limites de débit, et accès aux modèles restreints ou privés).
16+
17+Le vrai token n'entre jamais dans le sandbox : le sandbox ne voit qu'une valeur factice dans `HF_TOKEN`, et le proxy sbx la remplace par le vrai token quand une requête part vers `huggingface.co`.
18+
19+1. Crée un token **Read** sur <https://huggingface.co/settings/tokens>.
20+2. Sur la machine hôte, lance une fois (remplace `hf_xxx` par ton token) :
21+
22+```bash
23+printf '%s' "hf_xxx" | sbx secret set-custom \
24+ --host huggingface.co \
25+ --host '*.huggingface.co' \
26+ --env HF_TOKEN \
27+ --value-stdin
28+```
29+
30+Le secret est global (il s'applique à tous les sandboxes). Fais-le **avant** de créer le sandbox. Plus de détails : [docs/huggingface-token-sbx.fr.md](docs/huggingface-token-sbx.fr.md).
31+
32+> Ne mets jamais le token dans `sbxenv.yaml` : ce fichier est commité.
33+
34+## 3. Premier lancement
35+
36+Sur la machine hôte, à la racine du projet :
37+
38+```bash
39+sbx env create sbxenv.yaml
40+```
41+
42+Cela crée le sandbox `houla-olala`, télécharge le modèle (la première fois seulement, quelques minutes) et démarre le serveur sur le port 8080.
43+
44+Pour tester depuis l'hôte :
45+
46+```bash
47+# Le serveur répond ?
48+curl -s http://localhost:8080/health
49+
50+# Quel modèle est servi ?
51+curl -s http://localhost:8080/v1/models
52+
53+# Poser une question
54+curl -s http://localhost:8080/v1/chat/completions \
55+ -H 'Content-Type: application/json' \
56+ -d '{
57+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
58+ "messages": [{"role": "user", "content": "Explique le fonctionnement d'\''une pompe à chaleur en deux phrases."}],
59+ "reasoning_effort": "low"
60+ }' | jq -r '.choices[0].message.content'
61+
62+# Réponse en streaming
63+curl -sN http://localhost:8080/v1/chat/completions \
64+ -H 'Content-Type: application/json' \
65+ -d '{
66+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
67+ "stream": true,
68+ "reasoning_effort": "disabled",
69+ "messages": [{"role": "user", "content": "Dis bonjour"}]
70+ }'
71+```
72+
73+`reasoning_effort` peut valoir `disabled`, `low`, `medium` (par défaut) ou `high`.
74+
75+## 4. Relancer après avoir supprimé le sandbox
76+
77+```bash
78+sbx env rm sbxenv.yaml # supprime le sandbox (le modèle reste dans ./models)
79+sbx env create sbxenv.yaml # le recrée : pas de nouveau téléchargement, le serveur démarre directement
80+```
81+
82+Si le sandbox a seulement été **arrêté** (`sbx stop`, redémarrage de la machine), pas besoin de le recréer :
83+
84+```bash
85+sbx run -d --name houla-olala
86+sbx env exec --no-input sbxenv.yaml -- bash -lc "$PWD/scripts/olala.sh serve"
87+```
88+
89+## 5. Paramètres : port et modèle
90+
91+Passe-les avec `--env-arg` à la création du sandbox :
92+
93+```bash
94+# Un autre port sur l'hôte
95+sbx env create sbxenv.yaml --env-arg host_port=9090
96+
97+# Un autre fichier de modèle (du même dépôt Hugging Face)
98+sbx env create sbxenv.yaml --env-arg model_file=Olala-v1.0-mini-7A1B-q5_k_m.gguf
99+```
100+
101+| Argument | Défaut | Rôle |
102+| --- | --- | --- |
103+| `host_port` | `8080` | Port sur l'hôte |
104+| `model_repo` | `OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF` | Dépôt Hugging Face |
105+| `model_file` | `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` | Fichier GGUF à télécharger et servir |
106+
107+> Utilise les **mêmes** valeurs `--env-arg` pour toutes les commandes `sbx env` de ce sandbox (`rm`, `exec`…).
108+> Avec un autre port, remplace `8080` dans les URL curl. Avec un autre modèle, le nom `"model"` dans les requêtes est le nom du fichier sans `.gguf`.
109+
110+Autres paramètres (`models_dir`, `threads`, `ctx`) et détails : [docs/olala-sbx.fr.md](docs/olala-sbx.fr.md).
new file mode 100644
@@ -0,0 +1,110 @@
1+# Houla ... Houla
2+
3+🇬🇧 [English version](README.md)
4+
5+## 1. C'est quoi ce projet ?
6+
7+Houla fait tourner le modèle de langage **Olala-v1.0-mini-7A1B** (d'OVHai) sur ta machine, dans un sandbox Docker (`sbx`), et l'expose sous forme d'**API compatible OpenAI** sur `http://localhost:8080`.
8+
9+- Le modèle est servi par `llama-server`, issu du [fork llama.cpp d'OVHai](https://github.com/OVHai-LLM/llama.cpp) (CPU uniquement, pas besoin de GPU).
10+- Le fichier du modèle (GGUF, ~7 Go) est téléchargé une seule fois dans le dossier `./models`, puis réutilisé.
11+- Tout est décrit dans [`sbxenv.yaml`](sbxenv.yaml) : une seule commande crée le sandbox, télécharge le modèle et démarre le serveur.
12+
13+## 2. Créer le secret Hugging Face
14+
15+**Pourquoi ?** Le modèle est téléchargé depuis [Hugging Face](https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF). Avec un token, les téléchargements sont authentifiés (moins de limites de débit, et accès aux modèles restreints ou privés).
16+
17+Le vrai token n'entre jamais dans le sandbox : le sandbox ne voit qu'une valeur factice dans `HF_TOKEN`, et le proxy sbx la remplace par le vrai token quand une requête part vers `huggingface.co`.
18+
19+1. Crée un token **Read** sur <https://huggingface.co/settings/tokens>.
20+2. Sur la machine hôte, lance une fois (remplace `hf_xxx` par ton token) :
21+
22+```bash
23+printf '%s' "hf_xxx" | sbx secret set-custom \
24+ --host huggingface.co \
25+ --host '*.huggingface.co' \
26+ --env HF_TOKEN \
27+ --value-stdin
28+```
29+
30+Le secret est global (il s'applique à tous les sandboxes). Fais-le **avant** de créer le sandbox. Plus de détails : [docs/huggingface-token-sbx.fr.md](docs/huggingface-token-sbx.fr.md).
31+
32+> Ne mets jamais le token dans `sbxenv.yaml` : ce fichier est commité.
33+
34+## 3. Premier lancement
35+
36+Sur la machine hôte, à la racine du projet :
37+
38+```bash
39+sbx env create sbxenv.yaml
40+```
41+
42+Cela crée le sandbox `houla-olala`, télécharge le modèle (la première fois seulement, quelques minutes) et démarre le serveur sur le port 8080.
43+
44+Pour tester depuis l'hôte :
45+
46+```bash
47+# Le serveur répond ?
48+curl -s http://localhost:8080/health
49+
50+# Quel modèle est servi ?
51+curl -s http://localhost:8080/v1/models
52+
53+# Poser une question
54+curl -s http://localhost:8080/v1/chat/completions \
55+ -H 'Content-Type: application/json' \
56+ -d '{
57+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
58+ "messages": [{"role": "user", "content": "Explique le fonctionnement d'\''une pompe à chaleur en deux phrases."}],
59+ "reasoning_effort": "low"
60+ }' | jq -r '.choices[0].message.content'
61+
62+# Réponse en streaming
63+curl -sN http://localhost:8080/v1/chat/completions \
64+ -H 'Content-Type: application/json' \
65+ -d '{
66+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
67+ "stream": true,
68+ "reasoning_effort": "disabled",
69+ "messages": [{"role": "user", "content": "Dis bonjour"}]
70+ }'
71+```
72+
73+`reasoning_effort` peut valoir `disabled`, `low`, `medium` (par défaut) ou `high`.
74+
75+## 4. Relancer après avoir supprimé le sandbox
76+
77+```bash
78+sbx env rm sbxenv.yaml # supprime le sandbox (le modèle reste dans ./models)
79+sbx env create sbxenv.yaml # le recrée : pas de nouveau téléchargement, le serveur démarre directement
80+```
81+
82+Si le sandbox a seulement été **arrêté** (`sbx stop`, redémarrage de la machine), pas besoin de le recréer :
83+
84+```bash
85+sbx run -d --name houla-olala
86+sbx env exec --no-input sbxenv.yaml -- bash -lc "$PWD/scripts/olala.sh serve"
87+```
88+
89+## 5. Paramètres : port et modèle
90+
91+Passe-les avec `--env-arg` à la création du sandbox :
92+
93+```bash
94+# Un autre port sur l'hôte
95+sbx env create sbxenv.yaml --env-arg host_port=9090
96+
97+# Un autre fichier de modèle (du même dépôt Hugging Face)
98+sbx env create sbxenv.yaml --env-arg model_file=Olala-v1.0-mini-7A1B-q5_k_m.gguf
99+```
100+
101+| Argument | Défaut | Rôle |
102+| --- | --- | --- |
103+| `host_port` | `8080` | Port sur l'hôte |
104+| `model_repo` | `OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF` | Dépôt Hugging Face |
105+| `model_file` | `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` | Fichier GGUF à télécharger et servir |
106+
107+> Utilise les **mêmes** valeurs `--env-arg` pour toutes les commandes `sbx env` de ce sandbox (`rm`, `exec`…).
108+> Avec un autre port, remplace `8080` dans les URL curl. Avec un autre modèle, le nom `"model"` dans les requêtes est le nom du fichier sans `.gguf`.
109+
110+Autres paramètres (`models_dir`, `threads`, `ctx`) et détails : [docs/olala-sbx.fr.md](docs/olala-sbx.fr.md).
added README.md +110 -0
new file mode 100644
@@ -0,0 +1,110 @@
1+# Houla ... Houla
2+
3+🇫🇷 [Version française](README.fr.md)
4+
5+## 1. What is this project?
6+
7+Houla runs the **Olala-v1.0-mini-7A1B** language model (by OVHai) on your machine, inside a Docker sandbox (`sbx`), and exposes it as an **OpenAI-compatible API** on `http://localhost:8080`.
8+
9+- The model is served by `llama-server`, from OVHai's [llama.cpp fork](https://github.com/OVHai-LLM/llama.cpp) (CPU only, no GPU needed).
10+- The model file (GGUF, ~7 GB) is downloaded once into the `./models` folder and reused afterwards.
11+- Everything is described in [`sbxenv.yaml`](sbxenv.yaml): one command creates the sandbox, downloads the model and starts the server.
12+
13+## 2. Create the Hugging Face secret
14+
15+**Why?** The model is downloaded from [Hugging Face](https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF). With a token, downloads are authenticated (fewer rate limits, and access to gated or private models).
16+
17+The real token never enters the sandbox: the sandbox only sees a placeholder in `HF_TOKEN`, and the sbx proxy swaps in the real token when a request goes to `huggingface.co`.
18+
19+1. Create a **Read** token at <https://huggingface.co/settings/tokens>.
20+2. On the host, run once (replace `hf_xxx` with your token):
21+
22+```bash
23+printf '%s' "hf_xxx" | sbx secret set-custom \
24+ --host huggingface.co \
25+ --host '*.huggingface.co' \
26+ --env HF_TOKEN \
27+ --value-stdin
28+```
29+
30+The secret is global (it applies to every sandbox). Do this **before** creating the sandbox. More details: [docs/huggingface-token-sbx.en.md](docs/huggingface-token-sbx.en.md).
31+
32+> Never put the token in `sbxenv.yaml`: that file is committed.
33+
34+## 3. First launch
35+
36+On the host, at the project root:
37+
38+```bash
39+sbx env create sbxenv.yaml
40+```
41+
42+This creates the `houla-olala` sandbox, downloads the model (first time only, a few minutes) and starts the server on port 8080.
43+
44+Test it from the host:
45+
46+```bash
47+# Is the server up?
48+curl -s http://localhost:8080/health
49+
50+# Which model is served?
51+curl -s http://localhost:8080/v1/models
52+
53+# Ask a question
54+curl -s http://localhost:8080/v1/chat/completions \
55+ -H 'Content-Type: application/json' \
56+ -d '{
57+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
58+ "messages": [{"role": "user", "content": "Explain how a heat pump works in two sentences."}],
59+ "reasoning_effort": "low"
60+ }' | jq -r '.choices[0].message.content'
61+
62+# Streaming answer
63+curl -sN http://localhost:8080/v1/chat/completions \
64+ -H 'Content-Type: application/json' \
65+ -d '{
66+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
67+ "stream": true,
68+ "reasoning_effort": "disabled",
69+ "messages": [{"role": "user", "content": "Say hello"}]
70+ }'
71+```
72+
73+`reasoning_effort` can be `disabled`, `low`, `medium` (default) or `high`.
74+
75+## 4. Start again after deleting the sandbox
76+
77+```bash
78+sbx env rm sbxenv.yaml # delete the sandbox (the model stays in ./models)
79+sbx env create sbxenv.yaml # recreate it: no new download, the server starts directly
80+```
81+
82+If the sandbox was only **stopped** (`sbx stop`, reboot), no need to recreate it:
83+
84+```bash
85+sbx run -d --name houla-olala
86+sbx env exec --no-input sbxenv.yaml -- bash -lc "$PWD/scripts/olala.sh serve"
87+```
88+
89+## 5. Settings: port and model
90+
91+Pass them with `--env-arg` when creating the sandbox:
92+
93+```bash
94+# Another host port
95+sbx env create sbxenv.yaml --env-arg host_port=9090
96+
97+# Another model file (from the same Hugging Face repository)
98+sbx env create sbxenv.yaml --env-arg model_file=Olala-v1.0-mini-7A1B-q5_k_m.gguf
99+```
100+
101+| Argument | Default | Purpose |
102+| --- | --- | --- |
103+| `host_port` | `8080` | Port on the host |
104+| `model_repo` | `OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF` | Hugging Face repository |
105+| `model_file` | `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` | GGUF file to download and serve |
106+
107+> Use the **same** `--env-arg` values with every `sbx env` command of that sandbox (`rm`, `exec`…).
108+> With another port, replace `8080` in the curl URLs. With another model, the `"model"` name in the requests is the file name without `.gguf`.
109+
110+Other settings (`models_dir`, `threads`, `ctx`) and details: [docs/olala-sbx.en.md](docs/olala-sbx.en.md).
new file mode 100644
@@ -0,0 +1,110 @@
1+# Houla ... Houla
2+
3+🇫🇷 [Version française](README.fr.md)
4+
5+## 1. What is this project?
6+
7+Houla runs the **Olala-v1.0-mini-7A1B** language model (by OVHai) on your machine, inside a Docker sandbox (`sbx`), and exposes it as an **OpenAI-compatible API** on `http://localhost:8080`.
8+
9+- The model is served by `llama-server`, from OVHai's [llama.cpp fork](https://github.com/OVHai-LLM/llama.cpp) (CPU only, no GPU needed).
10+- The model file (GGUF, ~7 GB) is downloaded once into the `./models` folder and reused afterwards.
11+- Everything is described in [`sbxenv.yaml`](sbxenv.yaml): one command creates the sandbox, downloads the model and starts the server.
12+
13+## 2. Create the Hugging Face secret
14+
15+**Why?** The model is downloaded from [Hugging Face](https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF). With a token, downloads are authenticated (fewer rate limits, and access to gated or private models).
16+
17+The real token never enters the sandbox: the sandbox only sees a placeholder in `HF_TOKEN`, and the sbx proxy swaps in the real token when a request goes to `huggingface.co`.
18+
19+1. Create a **Read** token at <https://huggingface.co/settings/tokens>.
20+2. On the host, run once (replace `hf_xxx` with your token):
21+
22+```bash
23+printf '%s' "hf_xxx" | sbx secret set-custom \
24+ --host huggingface.co \
25+ --host '*.huggingface.co' \
26+ --env HF_TOKEN \
27+ --value-stdin
28+```
29+
30+The secret is global (it applies to every sandbox). Do this **before** creating the sandbox. More details: [docs/huggingface-token-sbx.en.md](docs/huggingface-token-sbx.en.md).
31+
32+> Never put the token in `sbxenv.yaml`: that file is committed.
33+
34+## 3. First launch
35+
36+On the host, at the project root:
37+
38+```bash
39+sbx env create sbxenv.yaml
40+```
41+
42+This creates the `houla-olala` sandbox, downloads the model (first time only, a few minutes) and starts the server on port 8080.
43+
44+Test it from the host:
45+
46+```bash
47+# Is the server up?
48+curl -s http://localhost:8080/health
49+
50+# Which model is served?
51+curl -s http://localhost:8080/v1/models
52+
53+# Ask a question
54+curl -s http://localhost:8080/v1/chat/completions \
55+ -H 'Content-Type: application/json' \
56+ -d '{
57+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
58+ "messages": [{"role": "user", "content": "Explain how a heat pump works in two sentences."}],
59+ "reasoning_effort": "low"
60+ }' | jq -r '.choices[0].message.content'
61+
62+# Streaming answer
63+curl -sN http://localhost:8080/v1/chat/completions \
64+ -H 'Content-Type: application/json' \
65+ -d '{
66+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
67+ "stream": true,
68+ "reasoning_effort": "disabled",
69+ "messages": [{"role": "user", "content": "Say hello"}]
70+ }'
71+```
72+
73+`reasoning_effort` can be `disabled`, `low`, `medium` (default) or `high`.
74+
75+## 4. Start again after deleting the sandbox
76+
77+```bash
78+sbx env rm sbxenv.yaml # delete the sandbox (the model stays in ./models)
79+sbx env create sbxenv.yaml # recreate it: no new download, the server starts directly
80+```
81+
82+If the sandbox was only **stopped** (`sbx stop`, reboot), no need to recreate it:
83+
84+```bash
85+sbx run -d --name houla-olala
86+sbx env exec --no-input sbxenv.yaml -- bash -lc "$PWD/scripts/olala.sh serve"
87+```
88+
89+## 5. Settings: port and model
90+
91+Pass them with `--env-arg` when creating the sandbox:
92+
93+```bash
94+# Another host port
95+sbx env create sbxenv.yaml --env-arg host_port=9090
96+
97+# Another model file (from the same Hugging Face repository)
98+sbx env create sbxenv.yaml --env-arg model_file=Olala-v1.0-mini-7A1B-q5_k_m.gguf
99+```
100+
101+| Argument | Default | Purpose |
102+| --- | --- | --- |
103+| `host_port` | `8080` | Port on the host |
104+| `model_repo` | `OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF` | Hugging Face repository |
105+| `model_file` | `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` | GGUF file to download and serve |
106+
107+> Use the **same** `--env-arg` values with every `sbx env` command of that sandbox (`rm`, `exec`…).
108+> With another port, replace `8080` in the curl URLs. With another model, the `"model"` name in the requests is the file name without `.gguf`.
109+
110+Other settings (`models_dir`, `threads`, `ctx`) and details: [docs/olala-sbx.en.md](docs/olala-sbx.en.md).
added docs/huggingface-token-sbx.en.md +74 -0
new file mode 100644
@@ -0,0 +1,74 @@
1+# Inject a Hugging Face token into an sbx sandbox
2+
3+This guide explains how to make a Hugging Face token usable in a Docker sandbox (`sbx`) without storing the real token in the container.
4+
5+> **Status:** `sbx secret set-custom` is **experimental**. Its behavior, options and placeholder format may change.
6+
7+## How it works
8+
9+Hugging Face is not one of the services built into `sbx` (like `github`). You therefore use a **custom secret**:
10+
11+- Inside the sandbox, the `HF_TOKEN` variable holds a **placeholder** (`sbx-cs-…`).
12+- The real token stays on the host machine, in the operating system's keychain.
13+- When a process in the sandbox sends an HTTP request to one of the declared domains, the proxy replaces the placeholder with the real token, **in HTTP headers only**, not in the request body.
14+
15+For Hugging Face, that is enough: the token is sent in the `Authorization: Bearer …` header.
16+
17+## Set up the secret
18+
19+All `sbx` commands run **on the host machine**, not in the sandbox.
20+
21+The secret is **global**: without `--sandbox`, it applies to every sandbox, so both the dev sandbox (`houla-houla`, `sbxenv.dev.yaml`) and the Olala sandbox (`houla-olala`, `sbxenv.yaml`). There is only one command to run, once.
22+
23+```bash
24+printf '%s' "hf_xxx" | sbx secret set-custom \
25+ --host huggingface.co \
26+ --host '*.huggingface.co' \
27+ --env HF_TOKEN \
28+ --value-stdin
29+```
30+
31+| Option | Purpose |
32+| --- | --- |
33+| `--host` | Domain the token is sent to. Repeatable. `*` matches a single subdomain level (e.g. `router.huggingface.co`, `cdn-lfs.huggingface.co`), `**` matches several. |
34+| `--env` | Name of the environment variable set in the sandbox. |
35+| `--value-stdin` | Reads the token from standard input (recommended). Without this option, the command prompts for the value. |
36+| `--sandbox <name>` | Restricts the secret to a single sandbox. Not used here: we want the secret everywhere. |
37+| `--ref` | Reads the token from an external source when needed, e.g. `op://Vault/Item/credential` (1Password). |
38+| `--command` | Reads the token when needed by running a command on the host that prints it. |
39+| `--refresh` | How long a token read by `--ref` or `--command` is cached (`<duration>` or `on-demand`). |
40+
41+> **Warning:** avoid `--value hf_xxx`: the token would end up in your shell history and in the process list.
42+
43+To check that the secret is registered:
44+
45+```bash
46+sbx secret ls --host huggingface.co
47+```
48+
49+## Check inside the sandbox
50+
51+```bash
52+echo "$HF_TOKEN" # prints the placeholder, not the real token
53+hf auth whoami # should print your Hugging Face account
54+```
55+
56+## Possible problems
57+
58+- **`HF_TOKEN` is empty in an existing sandbox**: for built-in secrets (like `github`), a global secret only applies to sandboxes created afterwards. It is probably the same here (not verified): recreate the sandbox, e.g. `sbx env rm sbxenv.dev.yaml` then `sbx env create sbxenv.dev.yaml`. For a single existing sandbox, you can also re-run the command with `--sandbox <name>` (for `github`, a sandbox-scoped secret applies immediately).
59+- **A tool rejects the placeholder because it does not start with `hf_`**: a `--placeholder` option would let you choose its format (e.g. `--placeholder 'hf_{rand}'`). It does not appear on the detailed docs page: check it with `sbx secret set-custom --help`.
60+- **HTTP 403 error "no matching allow rule"**: the domain is blocked by the sandbox firewall. On the host:
61+
62+ ```bash
63+ sbx policy allow network huggingface.co,cdn-lfs.huggingface.co
64+ ```
65+
66+## Don't
67+
68+Don't put the token in the `env:` section of `sbxenv.yaml` or `sbxenv.dev.yaml`: these files are meant to be committed, and the token would end up in the repository.
69+
70+## References
71+
72+- [Manage credentials – Docker Sandboxes](https://docs.docker.com/ai/sandboxes/configuration/credentials/)
73+- [`sbx secret set-custom`](https://docs.docker.com/reference/cli/sbx/secret/set-custom/)
74+- [`sbx secret`](https://docs.docker.com/reference/cli/sbx/secret/)
new file mode 100644
@@ -0,0 +1,74 @@
1+# Inject a Hugging Face token into an sbx sandbox
2+
3+This guide explains how to make a Hugging Face token usable in a Docker sandbox (`sbx`) without storing the real token in the container.
4+
5+> **Status:** `sbx secret set-custom` is **experimental**. Its behavior, options and placeholder format may change.
6+
7+## How it works
8+
9+Hugging Face is not one of the services built into `sbx` (like `github`). You therefore use a **custom secret**:
10+
11+- Inside the sandbox, the `HF_TOKEN` variable holds a **placeholder** (`sbx-cs-…`).
12+- The real token stays on the host machine, in the operating system's keychain.
13+- When a process in the sandbox sends an HTTP request to one of the declared domains, the proxy replaces the placeholder with the real token, **in HTTP headers only**, not in the request body.
14+
15+For Hugging Face, that is enough: the token is sent in the `Authorization: Bearer …` header.
16+
17+## Set up the secret
18+
19+All `sbx` commands run **on the host machine**, not in the sandbox.
20+
21+The secret is **global**: without `--sandbox`, it applies to every sandbox, so both the dev sandbox (`houla-houla`, `sbxenv.dev.yaml`) and the Olala sandbox (`houla-olala`, `sbxenv.yaml`). There is only one command to run, once.
22+
23+```bash
24+printf '%s' "hf_xxx" | sbx secret set-custom \
25+ --host huggingface.co \
26+ --host '*.huggingface.co' \
27+ --env HF_TOKEN \
28+ --value-stdin
29+```
30+
31+| Option | Purpose |
32+| --- | --- |
33+| `--host` | Domain the token is sent to. Repeatable. `*` matches a single subdomain level (e.g. `router.huggingface.co`, `cdn-lfs.huggingface.co`), `**` matches several. |
34+| `--env` | Name of the environment variable set in the sandbox. |
35+| `--value-stdin` | Reads the token from standard input (recommended). Without this option, the command prompts for the value. |
36+| `--sandbox <name>` | Restricts the secret to a single sandbox. Not used here: we want the secret everywhere. |
37+| `--ref` | Reads the token from an external source when needed, e.g. `op://Vault/Item/credential` (1Password). |
38+| `--command` | Reads the token when needed by running a command on the host that prints it. |
39+| `--refresh` | How long a token read by `--ref` or `--command` is cached (`<duration>` or `on-demand`). |
40+
41+> **Warning:** avoid `--value hf_xxx`: the token would end up in your shell history and in the process list.
42+
43+To check that the secret is registered:
44+
45+```bash
46+sbx secret ls --host huggingface.co
47+```
48+
49+## Check inside the sandbox
50+
51+```bash
52+echo "$HF_TOKEN" # prints the placeholder, not the real token
53+hf auth whoami # should print your Hugging Face account
54+```
55+
56+## Possible problems
57+
58+- **`HF_TOKEN` is empty in an existing sandbox**: for built-in secrets (like `github`), a global secret only applies to sandboxes created afterwards. It is probably the same here (not verified): recreate the sandbox, e.g. `sbx env rm sbxenv.dev.yaml` then `sbx env create sbxenv.dev.yaml`. For a single existing sandbox, you can also re-run the command with `--sandbox <name>` (for `github`, a sandbox-scoped secret applies immediately).
59+- **A tool rejects the placeholder because it does not start with `hf_`**: a `--placeholder` option would let you choose its format (e.g. `--placeholder 'hf_{rand}'`). It does not appear on the detailed docs page: check it with `sbx secret set-custom --help`.
60+- **HTTP 403 error "no matching allow rule"**: the domain is blocked by the sandbox firewall. On the host:
61+
62+ ```bash
63+ sbx policy allow network huggingface.co,cdn-lfs.huggingface.co
64+ ```
65+
66+## Don't
67+
68+Don't put the token in the `env:` section of `sbxenv.yaml` or `sbxenv.dev.yaml`: these files are meant to be committed, and the token would end up in the repository.
69+
70+## References
71+
72+- [Manage credentials – Docker Sandboxes](https://docs.docker.com/ai/sandboxes/configuration/credentials/)
73+- [`sbx secret set-custom`](https://docs.docker.com/reference/cli/sbx/secret/set-custom/)
74+- [`sbx secret`](https://docs.docker.com/reference/cli/sbx/secret/)
added docs/huggingface-token-sbx.fr.md +74 -0
new file mode 100644
@@ -0,0 +1,74 @@
1+# Injecter un token Hugging Face dans un sandbox sbx
2+
3+Ce guide explique comment rendre un token Hugging Face utilisable dans un sandbox Docker (`sbx`) sans que le vrai token soit stocké dans le conteneur.
4+
5+> **Statut :** `sbx secret set-custom` est **expérimental**. Son comportement, ses options et le format de la valeur factice peuvent changer.
6+
7+## Principe
8+
9+Hugging Face ne fait pas partie des services intégrés à `sbx` (comme `github`). Il faut donc passer par un **secret personnalisé** :
10+
11+- Dans le sandbox, la variable `HF_TOKEN` contient une **valeur factice** (`sbx-cs-…`).
12+- Le vrai token reste sur la machine hôte, dans le trousseau du système d'exploitation.
13+- Quand un processus du sandbox envoie une requête HTTP vers un des domaines déclarés, le proxy remplace la valeur factice par le vrai token, **uniquement dans les en-têtes HTTP**, pas dans le corps de la requête.
14+
15+Pour Hugging Face, cela suffit : le token est envoyé dans l'en-tête `Authorization: Bearer …`.
16+
17+## Configurer le secret
18+
19+Toutes les commandes `sbx` se lancent **sur la machine hôte**, pas dans le sandbox.
20+
21+Le secret est **global** : sans `--sandbox`, il s'applique à tous les sandboxes, donc aussi bien au sandbox de dev (`houla-houla`, `sbxenv.dev.yaml`) qu'au sandbox Olala (`houla-olala`, `sbxenv.yaml`). Il n'y a qu'une commande à lancer, une seule fois.
22+
23+```bash
24+printf '%s' "hf_xxx" | sbx secret set-custom \
25+ --host huggingface.co \
26+ --host '*.huggingface.co' \
27+ --env HF_TOKEN \
28+ --value-stdin
29+```
30+
31+| Option | Rôle |
32+| --- | --- |
33+| `--host` | Domaine auquel le token est envoyé. Répétable. `*` correspond à un seul niveau de sous-domaine (ex. `router.huggingface.co`, `cdn-lfs.huggingface.co`), `**` à plusieurs. |
34+| `--env` | Nom de la variable d'environnement définie dans le sandbox. |
35+| `--value-stdin` | Lit le token sur l'entrée standard (méthode recommandée). Sans cette option, la commande demande la valeur. |
36+| `--sandbox <nom>` | Limite le secret à un seul sandbox. Non utilisé ici : on veut le secret partout. |
37+| `--ref` | Lit le token au moment où il en a besoin dans une source externe, par ex. `op://Vault/Item/credential` (1Password). |
38+| `--command` | Lit le token au moment où il en a besoin en lançant une commande sur l'hôte qui l'affiche. |
39+| `--refresh` | Durée de mise en cache d'un token lu par `--ref` ou `--command` (`<durée>` ou `on-demand`). |
40+
41+> **Attention :** évite `--value hf_xxx` : le token se retrouverait dans l'historique du shell et dans la liste des processus.
42+
43+Pour vérifier que le secret est enregistré :
44+
45+```bash
46+sbx secret ls --host huggingface.co
47+```
48+
49+## Vérifier dans le sandbox
50+
51+```bash
52+echo "$HF_TOKEN" # affiche la valeur factice, pas le vrai token
53+hf auth whoami # doit afficher ton compte Hugging Face
54+```
55+
56+## Problèmes possibles
57+
58+- **`HF_TOKEN` est vide dans un sandbox déjà créé** : pour les secrets intégrés (comme `github`), un secret global ne s'applique qu'aux sandboxes créés ensuite. C'est sans doute pareil ici (pas vérifié) : recrée le sandbox, par ex. `sbx env rm sbxenv.dev.yaml` puis `sbx env create sbxenv.dev.yaml`. Pour un seul sandbox existant, tu peux aussi relancer la commande avec `--sandbox <nom>` (pour `github`, un secret limité à un sandbox s'applique tout de suite).
59+- **Un outil refuse la valeur factice parce qu'elle ne commence pas par `hf_`** : une option `--placeholder` permettrait de choisir son format (par ex. `--placeholder 'hf_{rand}'`). Elle n'apparaît pas dans la page de doc détaillée : vérifie-la avec `sbx secret set-custom --help`.
60+- **Erreur HTTP 403 « no matching allow rule »** : le domaine est bloqué par le pare-feu du sandbox. Sur l'hôte :
61+
62+ ```bash
63+ sbx policy allow network huggingface.co,cdn-lfs.huggingface.co
64+ ```
65+
66+## À ne pas faire
67+
68+Ne mets pas le token dans la section `env:` de `sbxenv.yaml` ou `sbxenv.dev.yaml` : ces fichiers sont faits pour être commités, et le token se retrouverait dans le dépôt.
69+
70+## Références
71+
72+- [Manage credentials – Docker Sandboxes](https://docs.docker.com/ai/sandboxes/configuration/credentials/)
73+- [`sbx secret set-custom`](https://docs.docker.com/reference/cli/sbx/secret/set-custom/)
74+- [`sbx secret`](https://docs.docker.com/reference/cli/sbx/secret/)
new file mode 100644
@@ -0,0 +1,74 @@
1+# Injecter un token Hugging Face dans un sandbox sbx
2+
3+Ce guide explique comment rendre un token Hugging Face utilisable dans un sandbox Docker (`sbx`) sans que le vrai token soit stocké dans le conteneur.
4+
5+> **Statut :** `sbx secret set-custom` est **expérimental**. Son comportement, ses options et le format de la valeur factice peuvent changer.
6+
7+## Principe
8+
9+Hugging Face ne fait pas partie des services intégrés à `sbx` (comme `github`). Il faut donc passer par un **secret personnalisé** :
10+
11+- Dans le sandbox, la variable `HF_TOKEN` contient une **valeur factice** (`sbx-cs-…`).
12+- Le vrai token reste sur la machine hôte, dans le trousseau du système d'exploitation.
13+- Quand un processus du sandbox envoie une requête HTTP vers un des domaines déclarés, le proxy remplace la valeur factice par le vrai token, **uniquement dans les en-têtes HTTP**, pas dans le corps de la requête.
14+
15+Pour Hugging Face, cela suffit : le token est envoyé dans l'en-tête `Authorization: Bearer …`.
16+
17+## Configurer le secret
18+
19+Toutes les commandes `sbx` se lancent **sur la machine hôte**, pas dans le sandbox.
20+
21+Le secret est **global** : sans `--sandbox`, il s'applique à tous les sandboxes, donc aussi bien au sandbox de dev (`houla-houla`, `sbxenv.dev.yaml`) qu'au sandbox Olala (`houla-olala`, `sbxenv.yaml`). Il n'y a qu'une commande à lancer, une seule fois.
22+
23+```bash
24+printf '%s' "hf_xxx" | sbx secret set-custom \
25+ --host huggingface.co \
26+ --host '*.huggingface.co' \
27+ --env HF_TOKEN \
28+ --value-stdin
29+```
30+
31+| Option | Rôle |
32+| --- | --- |
33+| `--host` | Domaine auquel le token est envoyé. Répétable. `*` correspond à un seul niveau de sous-domaine (ex. `router.huggingface.co`, `cdn-lfs.huggingface.co`), `**` à plusieurs. |
34+| `--env` | Nom de la variable d'environnement définie dans le sandbox. |
35+| `--value-stdin` | Lit le token sur l'entrée standard (méthode recommandée). Sans cette option, la commande demande la valeur. |
36+| `--sandbox <nom>` | Limite le secret à un seul sandbox. Non utilisé ici : on veut le secret partout. |
37+| `--ref` | Lit le token au moment où il en a besoin dans une source externe, par ex. `op://Vault/Item/credential` (1Password). |
38+| `--command` | Lit le token au moment où il en a besoin en lançant une commande sur l'hôte qui l'affiche. |
39+| `--refresh` | Durée de mise en cache d'un token lu par `--ref` ou `--command` (`<durée>` ou `on-demand`). |
40+
41+> **Attention :** évite `--value hf_xxx` : le token se retrouverait dans l'historique du shell et dans la liste des processus.
42+
43+Pour vérifier que le secret est enregistré :
44+
45+```bash
46+sbx secret ls --host huggingface.co
47+```
48+
49+## Vérifier dans le sandbox
50+
51+```bash
52+echo "$HF_TOKEN" # affiche la valeur factice, pas le vrai token
53+hf auth whoami # doit afficher ton compte Hugging Face
54+```
55+
56+## Problèmes possibles
57+
58+- **`HF_TOKEN` est vide dans un sandbox déjà créé** : pour les secrets intégrés (comme `github`), un secret global ne s'applique qu'aux sandboxes créés ensuite. C'est sans doute pareil ici (pas vérifié) : recrée le sandbox, par ex. `sbx env rm sbxenv.dev.yaml` puis `sbx env create sbxenv.dev.yaml`. Pour un seul sandbox existant, tu peux aussi relancer la commande avec `--sandbox <nom>` (pour `github`, un secret limité à un sandbox s'applique tout de suite).
59+- **Un outil refuse la valeur factice parce qu'elle ne commence pas par `hf_`** : une option `--placeholder` permettrait de choisir son format (par ex. `--placeholder 'hf_{rand}'`). Elle n'apparaît pas dans la page de doc détaillée : vérifie-la avec `sbx secret set-custom --help`.
60+- **Erreur HTTP 403 « no matching allow rule »** : le domaine est bloqué par le pare-feu du sandbox. Sur l'hôte :
61+
62+ ```bash
63+ sbx policy allow network huggingface.co,cdn-lfs.huggingface.co
64+ ```
65+
66+## À ne pas faire
67+
68+Ne mets pas le token dans la section `env:` de `sbxenv.yaml` ou `sbxenv.dev.yaml` : ces fichiers sont faits pour être commités, et le token se retrouverait dans le dépôt.
69+
70+## Références
71+
72+- [Manage credentials – Docker Sandboxes](https://docs.docker.com/ai/sandboxes/configuration/credentials/)
73+- [`sbx secret set-custom`](https://docs.docker.com/reference/cli/sbx/secret/set-custom/)
74+- [`sbx secret`](https://docs.docker.com/reference/cli/sbx/secret/)
added docs/olala-sbx.en.md +123 -0
new file mode 100644
@@ -0,0 +1,123 @@
1+# Serve Olala-v1.0-mini-7A1B in an sbx sandbox
2+
3+This guide creates a sandbox that serves an Olala GGUF model with `llama-server` (OpenAI-compatible API), reachable from the host. `llama-server` comes from OVHai's llama.cpp fork (CPU only), shipped by the [`kits/llama-cpp-olala`](../kits/llama-cpp-olala/) kit.
4+
5+> Upstream llama.cpp, Ollama and LM Studio do not support the `olala` architecture yet: the [OVHai-LLM/llama.cpp](https://github.com/OVHai-LLM/llama.cpp) fork is required (see the [GGUF model card](https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF)).
6+
7+## Create the sandbox
8+
9+On the host, at the project root:
10+
11+```bash
12+sbx env create sbxenv.yaml
13+```
14+
15+> The development sandbox (Claude Code + Go/Python/mini-me toolchains, `houla-houla`) is described separately in `sbxenv.dev.yaml`: `sbx env run sbxenv.dev.yaml`.
16+
17+The `postCreate` hook runs [`scripts/olala.sh up`](../scripts/olala.sh) in the sandbox:
18+
19+1. uses the kit's `llama-server` (without the kit, it builds the fork: a few minutes);
20+2. downloads `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` (6.9 GB) into `./models`, in the repository;
21+3. starts `llama-server` on `0.0.0.0:8080`, published on host port 8080.
22+
23+Each step is skipped when already done (`llama-server` found, model present, server running). Models stay in `./models` (ignored by git) when the sandbox is removed: a new sandbox does not download them again.
24+
25+> The repository folder is on kDrive, so the models are synced too. Exclude `models/` from kDrive sync if you do not want that.
26+
27+Options (pass them again to every `sbx env` command of this sandbox):
28+
29+| Argument | Default | Purpose |
30+| --- | --- | --- |
31+| `model_repo` | `OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF` | Hugging Face repository |
32+| `model_file` | `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` | GGUF file |
33+| `models_dir` | `./models` | Models folder (relative to the repository root, or absolute) |
34+| `host_port` | `8080` | Host port |
35+| `threads` | `16` | CPU threads (keep 2 CPUs free: using all of them is ~30% slower) |
36+| `ctx` | `32768` | Context size (shared by the 4 slots) |
37+
38+```bash
39+sbx env create sbxenv.yaml --env-arg model_file=Olala-v1.0-mini-7A1B-q5_k_m.gguf --env-arg host_port=9090
40+```
41+
42+## In an existing sandbox
43+
44+Inside the sandbox (or through `sbx env exec sbxenv.yaml -- bash -lc '$SBX_WS/scripts/olala.sh <command>'`):
45+
46+```bash
47+scripts/olala.sh up # build + download + serve
48+scripts/olala.sh status # state and /health
49+scripts/olala.sh restart # after a model change
50+scripts/olala.sh stop
51+scripts/olala.sh logs
52+```
53+
54+Switch model:
55+
56+```bash
57+OLALA_MODEL_FILE=Olala-v1.0-mini-7A1B-q6_k.gguf scripts/olala.sh download
58+scripts/olala.sh stop
59+OLALA_MODEL_FILE=Olala-v1.0-mini-7A1B-q6_k.gguf scripts/olala.sh serve
60+```
61+
62+The same settings exist as environment variables: `OLALA_MODELS_DIR`, `OLALA_THREADS`, `OLALA_CTX`, `OLALA_PORT`, `OLALA_LOAD_MODE`, `OLALA_EXTRA_ARGS`… (listed at the top of [`scripts/olala.sh`](../scripts/olala.sh)).
63+
64+> **Model loading:** the script runs `llama-server --load-mode none` (the model is read into RAM, ~7 GB). With `mmap` (llama.cpp's default) the file stays on the repository mount and generation drops to ~2 tokens/s.
65+
66+A local sandbox stops as soon as no session keeps it running, and `llama-server` with it. The `postCreate` hook therefore switches the sandbox to detached mode (`sbx run -d --name houla-olala`, permanent for this sandbox): it keeps running until `sbx stop`. After `sbx stop` or a reboot, start it again, then the server:
67+
68+```bash
69+sbx run -d --name houla-olala
70+sbx env exec --no-input sbxenv.yaml -- bash -lc "$PWD/scripts/olala.sh serve"
71+```
72+
73+For a sandbox created without `sbxenv.yaml` (e.g. `houla-houla`, the dev sandbox of `sbxenv.dev.yaml`), publish the port from the host:
74+
75+```bash
76+sbx ports houla-houla --publish 8080:8080
77+```
78+
79+## Test with curl
80+
81+From the host (or the sandbox):
82+
83+```bash
84+# Health
85+curl -s http://localhost:8080/health
86+
87+# Served models (name = file name without .gguf)
88+curl -s http://localhost:8080/v1/models
89+
90+# Chat
91+curl -s http://localhost:8080/v1/chat/completions \
92+ -H 'Content-Type: application/json' \
93+ -d '{
94+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
95+ "messages": [{"role": "user", "content": "Explain how a heat pump works in two sentences."}],
96+ "temperature": 0.8, "top_p": 0.9, "top_k": 40,
97+ "reasoning_effort": "low"
98+ }' | jq -r '.choices[0].message.content'
99+
100+# Streaming, no reasoning
101+curl -sN http://localhost:8080/v1/chat/completions \
102+ -H 'Content-Type: application/json' \
103+ -d '{
104+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
105+ "stream": true,
106+ "reasoning_effort": "disabled",
107+ "messages": [{"role": "user", "content": "Say hello"}]
108+ }'
109+```
110+
111+`reasoning_effort` is `disabled`, `low`, `medium` (default) or `high`. The reasoning comes back in `choices[0].message.reasoning_content`.
112+
113+Measured in an aarch64 sandbox (18 CPUs, 31 GB) with q8_0-safe2, the kit's `llama-server`, 16 threads: 35 to 44 tokens/s generation.
114+
115+## The `llama-cpp-olala` kit
116+
117+An sbx mixin (kits v3) that builds the fork once, when the kit is built, and puts `llama-server` and `llama-cli` in `/opt/llama-cpp-olala` (links in `/usr/local/bin`). Portable build (`GGML_NATIVE=OFF`, every CPU variant loaded at runtime), arm64 (tested) and amd64 (not tested yet). The `ref` argument pins the fork commit.
118+
119+To use it in another sandbox (syntax not yet tried with `sbx`):
120+
121+```bash
122+sbx run claude --kit ./kits/llama-cpp-olala
123+```
new file mode 100644
@@ -0,0 +1,123 @@
1+# Serve Olala-v1.0-mini-7A1B in an sbx sandbox
2+
3+This guide creates a sandbox that serves an Olala GGUF model with `llama-server` (OpenAI-compatible API), reachable from the host. `llama-server` comes from OVHai's llama.cpp fork (CPU only), shipped by the [`kits/llama-cpp-olala`](../kits/llama-cpp-olala/) kit.
4+
5+> Upstream llama.cpp, Ollama and LM Studio do not support the `olala` architecture yet: the [OVHai-LLM/llama.cpp](https://github.com/OVHai-LLM/llama.cpp) fork is required (see the [GGUF model card](https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF)).
6+
7+## Create the sandbox
8+
9+On the host, at the project root:
10+
11+```bash
12+sbx env create sbxenv.yaml
13+```
14+
15+> The development sandbox (Claude Code + Go/Python/mini-me toolchains, `houla-houla`) is described separately in `sbxenv.dev.yaml`: `sbx env run sbxenv.dev.yaml`.
16+
17+The `postCreate` hook runs [`scripts/olala.sh up`](../scripts/olala.sh) in the sandbox:
18+
19+1. uses the kit's `llama-server` (without the kit, it builds the fork: a few minutes);
20+2. downloads `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` (6.9 GB) into `./models`, in the repository;
21+3. starts `llama-server` on `0.0.0.0:8080`, published on host port 8080.
22+
23+Each step is skipped when already done (`llama-server` found, model present, server running). Models stay in `./models` (ignored by git) when the sandbox is removed: a new sandbox does not download them again.
24+
25+> The repository folder is on kDrive, so the models are synced too. Exclude `models/` from kDrive sync if you do not want that.
26+
27+Options (pass them again to every `sbx env` command of this sandbox):
28+
29+| Argument | Default | Purpose |
30+| --- | --- | --- |
31+| `model_repo` | `OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF` | Hugging Face repository |
32+| `model_file` | `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` | GGUF file |
33+| `models_dir` | `./models` | Models folder (relative to the repository root, or absolute) |
34+| `host_port` | `8080` | Host port |
35+| `threads` | `16` | CPU threads (keep 2 CPUs free: using all of them is ~30% slower) |
36+| `ctx` | `32768` | Context size (shared by the 4 slots) |
37+
38+```bash
39+sbx env create sbxenv.yaml --env-arg model_file=Olala-v1.0-mini-7A1B-q5_k_m.gguf --env-arg host_port=9090
40+```
41+
42+## In an existing sandbox
43+
44+Inside the sandbox (or through `sbx env exec sbxenv.yaml -- bash -lc '$SBX_WS/scripts/olala.sh <command>'`):
45+
46+```bash
47+scripts/olala.sh up # build + download + serve
48+scripts/olala.sh status # state and /health
49+scripts/olala.sh restart # after a model change
50+scripts/olala.sh stop
51+scripts/olala.sh logs
52+```
53+
54+Switch model:
55+
56+```bash
57+OLALA_MODEL_FILE=Olala-v1.0-mini-7A1B-q6_k.gguf scripts/olala.sh download
58+scripts/olala.sh stop
59+OLALA_MODEL_FILE=Olala-v1.0-mini-7A1B-q6_k.gguf scripts/olala.sh serve
60+```
61+
62+The same settings exist as environment variables: `OLALA_MODELS_DIR`, `OLALA_THREADS`, `OLALA_CTX`, `OLALA_PORT`, `OLALA_LOAD_MODE`, `OLALA_EXTRA_ARGS`… (listed at the top of [`scripts/olala.sh`](../scripts/olala.sh)).
63+
64+> **Model loading:** the script runs `llama-server --load-mode none` (the model is read into RAM, ~7 GB). With `mmap` (llama.cpp's default) the file stays on the repository mount and generation drops to ~2 tokens/s.
65+
66+A local sandbox stops as soon as no session keeps it running, and `llama-server` with it. The `postCreate` hook therefore switches the sandbox to detached mode (`sbx run -d --name houla-olala`, permanent for this sandbox): it keeps running until `sbx stop`. After `sbx stop` or a reboot, start it again, then the server:
67+
68+```bash
69+sbx run -d --name houla-olala
70+sbx env exec --no-input sbxenv.yaml -- bash -lc "$PWD/scripts/olala.sh serve"
71+```
72+
73+For a sandbox created without `sbxenv.yaml` (e.g. `houla-houla`, the dev sandbox of `sbxenv.dev.yaml`), publish the port from the host:
74+
75+```bash
76+sbx ports houla-houla --publish 8080:8080
77+```
78+
79+## Test with curl
80+
81+From the host (or the sandbox):
82+
83+```bash
84+# Health
85+curl -s http://localhost:8080/health
86+
87+# Served models (name = file name without .gguf)
88+curl -s http://localhost:8080/v1/models
89+
90+# Chat
91+curl -s http://localhost:8080/v1/chat/completions \
92+ -H 'Content-Type: application/json' \
93+ -d '{
94+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
95+ "messages": [{"role": "user", "content": "Explain how a heat pump works in two sentences."}],
96+ "temperature": 0.8, "top_p": 0.9, "top_k": 40,
97+ "reasoning_effort": "low"
98+ }' | jq -r '.choices[0].message.content'
99+
100+# Streaming, no reasoning
101+curl -sN http://localhost:8080/v1/chat/completions \
102+ -H 'Content-Type: application/json' \
103+ -d '{
104+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
105+ "stream": true,
106+ "reasoning_effort": "disabled",
107+ "messages": [{"role": "user", "content": "Say hello"}]
108+ }'
109+```
110+
111+`reasoning_effort` is `disabled`, `low`, `medium` (default) or `high`. The reasoning comes back in `choices[0].message.reasoning_content`.
112+
113+Measured in an aarch64 sandbox (18 CPUs, 31 GB) with q8_0-safe2, the kit's `llama-server`, 16 threads: 35 to 44 tokens/s generation.
114+
115+## The `llama-cpp-olala` kit
116+
117+An sbx mixin (kits v3) that builds the fork once, when the kit is built, and puts `llama-server` and `llama-cli` in `/opt/llama-cpp-olala` (links in `/usr/local/bin`). Portable build (`GGML_NATIVE=OFF`, every CPU variant loaded at runtime), arm64 (tested) and amd64 (not tested yet). The `ref` argument pins the fork commit.
118+
119+To use it in another sandbox (syntax not yet tried with `sbx`):
120+
121+```bash
122+sbx run claude --kit ./kits/llama-cpp-olala
123+```
added docs/olala-sbx.fr.md +123 -0
new file mode 100644
@@ -0,0 +1,123 @@
1+# Servir Olala-v1.0-mini-7A1B dans un sandbox sbx
2+
3+Ce guide crée un sandbox qui sert un modèle GGUF Olala avec `llama-server` (API compatible OpenAI), accessible depuis la machine hôte. `llama-server` vient du fork llama.cpp d'OVHai (CPU uniquement), fourni par le kit [`kits/llama-cpp-olala`](../kits/llama-cpp-olala/).
4+
5+> Upstream llama.cpp, Ollama et LM Studio ne connaissent pas encore l'architecture `olala` : il faut le fork [OVHai-LLM/llama.cpp](https://github.com/OVHai-LLM/llama.cpp) (voir la [fiche du modèle GGUF](https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF)).
6+
7+## Créer le sandbox
8+
9+Sur la machine hôte, à la racine du projet :
10+
11+```bash
12+sbx env create sbxenv.yaml
13+```
14+
15+> Le sandbox de développement (Claude Code + toolchains Go/Python/mini-me, `houla-houla`) est décrit à part dans `sbxenv.dev.yaml` : `sbx env run sbxenv.dev.yaml`.
16+
17+Le hook `postCreate` lance [`scripts/olala.sh up`](../scripts/olala.sh) dans le sandbox :
18+
19+1. utilise le `llama-server` du kit (sans le kit, il compile le fork : quelques minutes) ;
20+2. télécharge `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` (6,9 Go) dans `./models`, dans le dépôt ;
21+3. démarre `llama-server` sur `0.0.0.0:8080`, publié sur le port 8080 de l'hôte.
22+
23+Chaque étape est sautée si elle est déjà faite (`llama-server` trouvé, modèle présent, serveur lancé). Les modèles restent dans `./models` (ignoré par git) quand le sandbox est supprimé : un nouveau sandbox ne les retélécharge pas.
24+
25+> Le dossier du dépôt est sur kDrive : les modèles y sont donc aussi synchronisés. Exclus `models/` de la synchronisation kDrive si tu ne le veux pas.
26+
27+Options (à repasser à chaque commande `sbx env` de ce sandbox) :
28+
29+| Argument | Défaut | Rôle |
30+| --- | --- | --- |
31+| `model_repo` | `OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF` | Dépôt Hugging Face |
32+| `model_file` | `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` | Fichier GGUF |
33+| `models_dir` | `./models` | Dossier des modèles (relatif à la racine du dépôt, ou absolu) |
34+| `host_port` | `8080` | Port de l'hôte |
35+| `threads` | `16` | Threads CPU (garde 2 CPU libres : tous les utiliser est ~30 % plus lent) |
36+| `ctx` | `32768` | Taille du contexte (partagée par les 4 slots) |
37+
38+```bash
39+sbx env create sbxenv.yaml --env-arg model_file=Olala-v1.0-mini-7A1B-q5_k_m.gguf --env-arg host_port=9090
40+```
41+
42+## Dans un sandbox existant
43+
44+Dans le sandbox (ou via `sbx env exec sbxenv.yaml -- bash -lc '$SBX_WS/scripts/olala.sh <commande>'`) :
45+
46+```bash
47+scripts/olala.sh up # compile + télécharge + sert
48+scripts/olala.sh status # état et /health
49+scripts/olala.sh restart # après un changement de modèle
50+scripts/olala.sh stop
51+scripts/olala.sh logs
52+```
53+
54+Changer de modèle :
55+
56+```bash
57+OLALA_MODEL_FILE=Olala-v1.0-mini-7A1B-q6_k.gguf scripts/olala.sh download
58+scripts/olala.sh stop
59+OLALA_MODEL_FILE=Olala-v1.0-mini-7A1B-q6_k.gguf scripts/olala.sh serve
60+```
61+
62+Les mêmes réglages existent en variables d'environnement : `OLALA_MODELS_DIR`, `OLALA_THREADS`, `OLALA_CTX`, `OLALA_PORT`, `OLALA_LOAD_MODE`, `OLALA_EXTRA_ARGS`… (liste en tête de [`scripts/olala.sh`](../scripts/olala.sh)).
63+
64+> **Chargement du modèle :** le script lance `llama-server --load-mode none` (le modèle est lu en RAM, ~7 Go). En `mmap` (défaut de llama.cpp), le fichier reste sur le montage du dépôt et la génération tombe à ~2 tokens/s.
65+
66+Un sandbox local s'arrête dès qu'aucune session ne le garde en vie, et `llama-server` avec lui. Le hook `postCreate` passe donc le sandbox en mode détaché (`sbx run -d --name houla-olala`, permanent pour ce sandbox) : il tourne jusqu'à `sbx stop`. Après un `sbx stop` ou un redémarrage de la machine, relance-le puis relance le serveur :
67+
68+```bash
69+sbx run -d --name houla-olala
70+sbx env exec --no-input sbxenv.yaml -- bash -lc "$PWD/scripts/olala.sh serve"
71+```
72+
73+Pour un sandbox créé sans `sbxenv.yaml` (par ex. `houla-houla`, le sandbox de dev de `sbxenv.dev.yaml`), publie le port depuis l'hôte :
74+
75+```bash
76+sbx ports houla-houla --publish 8080:8080
77+```
78+
79+## Tester avec curl
80+
81+Depuis l'hôte (ou le sandbox) :
82+
83+```bash
84+# Santé
85+curl -s http://localhost:8080/health
86+
87+# Modèles servis (le nom = fichier sans .gguf)
88+curl -s http://localhost:8080/v1/models
89+
90+# Chat
91+curl -s http://localhost:8080/v1/chat/completions \
92+ -H 'Content-Type: application/json' \
93+ -d '{
94+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
95+ "messages": [{"role": "user", "content": "Explique en deux phrases ce qu est une pompe à chaleur."}],
96+ "temperature": 0.8, "top_p": 0.9, "top_k": 40,
97+ "reasoning_effort": "low"
98+ }' | jq -r '.choices[0].message.content'
99+
100+# Streaming, sans raisonnement
101+curl -sN http://localhost:8080/v1/chat/completions \
102+ -H 'Content-Type: application/json' \
103+ -d '{
104+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
105+ "stream": true,
106+ "reasoning_effort": "disabled",
107+ "messages": [{"role": "user", "content": "Dis bonjour"}]
108+ }'
109+```
110+
111+`reasoning_effort` vaut `disabled`, `low`, `medium` (défaut) ou `high`. Le raisonnement est renvoyé dans `choices[0].message.reasoning_content`.
112+
113+Mesuré dans un sandbox aarch64 (18 CPU, 31 Go) avec q8_0-safe2, `llama-server` du kit, 16 threads : 35 à 44 tokens/s en génération.
114+
115+## Le kit `llama-cpp-olala`
116+
117+Mixin sbx (kits v3) qui compile le fork une fois, au build du kit, et pose `llama-server` et `llama-cli` dans `/opt/llama-cpp-olala` (liens dans `/usr/local/bin`). Build portable (`GGML_NATIVE=OFF`, toutes les variantes CPU chargées à l'exécution), arm64 (testé) et amd64 (pas encore testé). L'argument `ref` fixe le commit du fork.
118+
119+Pour l'utiliser dans un autre sandbox (syntaxe pas encore testée avec `sbx`) :
120+
121+```bash
122+sbx run claude --kit ./kits/llama-cpp-olala
123+```
new file mode 100644
@@ -0,0 +1,123 @@
1+# Servir Olala-v1.0-mini-7A1B dans un sandbox sbx
2+
3+Ce guide crée un sandbox qui sert un modèle GGUF Olala avec `llama-server` (API compatible OpenAI), accessible depuis la machine hôte. `llama-server` vient du fork llama.cpp d'OVHai (CPU uniquement), fourni par le kit [`kits/llama-cpp-olala`](../kits/llama-cpp-olala/).
4+
5+> Upstream llama.cpp, Ollama et LM Studio ne connaissent pas encore l'architecture `olala` : il faut le fork [OVHai-LLM/llama.cpp](https://github.com/OVHai-LLM/llama.cpp) (voir la [fiche du modèle GGUF](https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF)).
6+
7+## Créer le sandbox
8+
9+Sur la machine hôte, à la racine du projet :
10+
11+```bash
12+sbx env create sbxenv.yaml
13+```
14+
15+> Le sandbox de développement (Claude Code + toolchains Go/Python/mini-me, `houla-houla`) est décrit à part dans `sbxenv.dev.yaml` : `sbx env run sbxenv.dev.yaml`.
16+
17+Le hook `postCreate` lance [`scripts/olala.sh up`](../scripts/olala.sh) dans le sandbox :
18+
19+1. utilise le `llama-server` du kit (sans le kit, il compile le fork : quelques minutes) ;
20+2. télécharge `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` (6,9 Go) dans `./models`, dans le dépôt ;
21+3. démarre `llama-server` sur `0.0.0.0:8080`, publié sur le port 8080 de l'hôte.
22+
23+Chaque étape est sautée si elle est déjà faite (`llama-server` trouvé, modèle présent, serveur lancé). Les modèles restent dans `./models` (ignoré par git) quand le sandbox est supprimé : un nouveau sandbox ne les retélécharge pas.
24+
25+> Le dossier du dépôt est sur kDrive : les modèles y sont donc aussi synchronisés. Exclus `models/` de la synchronisation kDrive si tu ne le veux pas.
26+
27+Options (à repasser à chaque commande `sbx env` de ce sandbox) :
28+
29+| Argument | Défaut | Rôle |
30+| --- | --- | --- |
31+| `model_repo` | `OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF` | Dépôt Hugging Face |
32+| `model_file` | `Olala-v1.0-mini-7A1B-q8_0-safe2.gguf` | Fichier GGUF |
33+| `models_dir` | `./models` | Dossier des modèles (relatif à la racine du dépôt, ou absolu) |
34+| `host_port` | `8080` | Port de l'hôte |
35+| `threads` | `16` | Threads CPU (garde 2 CPU libres : tous les utiliser est ~30 % plus lent) |
36+| `ctx` | `32768` | Taille du contexte (partagée par les 4 slots) |
37+
38+```bash
39+sbx env create sbxenv.yaml --env-arg model_file=Olala-v1.0-mini-7A1B-q5_k_m.gguf --env-arg host_port=9090
40+```
41+
42+## Dans un sandbox existant
43+
44+Dans le sandbox (ou via `sbx env exec sbxenv.yaml -- bash -lc '$SBX_WS/scripts/olala.sh <commande>'`) :
45+
46+```bash
47+scripts/olala.sh up # compile + télécharge + sert
48+scripts/olala.sh status # état et /health
49+scripts/olala.sh restart # après un changement de modèle
50+scripts/olala.sh stop
51+scripts/olala.sh logs
52+```
53+
54+Changer de modèle :
55+
56+```bash
57+OLALA_MODEL_FILE=Olala-v1.0-mini-7A1B-q6_k.gguf scripts/olala.sh download
58+scripts/olala.sh stop
59+OLALA_MODEL_FILE=Olala-v1.0-mini-7A1B-q6_k.gguf scripts/olala.sh serve
60+```
61+
62+Les mêmes réglages existent en variables d'environnement : `OLALA_MODELS_DIR`, `OLALA_THREADS`, `OLALA_CTX`, `OLALA_PORT`, `OLALA_LOAD_MODE`, `OLALA_EXTRA_ARGS`… (liste en tête de [`scripts/olala.sh`](../scripts/olala.sh)).
63+
64+> **Chargement du modèle :** le script lance `llama-server --load-mode none` (le modèle est lu en RAM, ~7 Go). En `mmap` (défaut de llama.cpp), le fichier reste sur le montage du dépôt et la génération tombe à ~2 tokens/s.
65+
66+Un sandbox local s'arrête dès qu'aucune session ne le garde en vie, et `llama-server` avec lui. Le hook `postCreate` passe donc le sandbox en mode détaché (`sbx run -d --name houla-olala`, permanent pour ce sandbox) : il tourne jusqu'à `sbx stop`. Après un `sbx stop` ou un redémarrage de la machine, relance-le puis relance le serveur :
67+
68+```bash
69+sbx run -d --name houla-olala
70+sbx env exec --no-input sbxenv.yaml -- bash -lc "$PWD/scripts/olala.sh serve"
71+```
72+
73+Pour un sandbox créé sans `sbxenv.yaml` (par ex. `houla-houla`, le sandbox de dev de `sbxenv.dev.yaml`), publie le port depuis l'hôte :
74+
75+```bash
76+sbx ports houla-houla --publish 8080:8080
77+```
78+
79+## Tester avec curl
80+
81+Depuis l'hôte (ou le sandbox) :
82+
83+```bash
84+# Santé
85+curl -s http://localhost:8080/health
86+
87+# Modèles servis (le nom = fichier sans .gguf)
88+curl -s http://localhost:8080/v1/models
89+
90+# Chat
91+curl -s http://localhost:8080/v1/chat/completions \
92+ -H 'Content-Type: application/json' \
93+ -d '{
94+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
95+ "messages": [{"role": "user", "content": "Explique en deux phrases ce qu est une pompe à chaleur."}],
96+ "temperature": 0.8, "top_p": 0.9, "top_k": 40,
97+ "reasoning_effort": "low"
98+ }' | jq -r '.choices[0].message.content'
99+
100+# Streaming, sans raisonnement
101+curl -sN http://localhost:8080/v1/chat/completions \
102+ -H 'Content-Type: application/json' \
103+ -d '{
104+ "model": "Olala-v1.0-mini-7A1B-q8_0-safe2",
105+ "stream": true,
106+ "reasoning_effort": "disabled",
107+ "messages": [{"role": "user", "content": "Dis bonjour"}]
108+ }'
109+```
110+
111+`reasoning_effort` vaut `disabled`, `low`, `medium` (défaut) ou `high`. Le raisonnement est renvoyé dans `choices[0].message.reasoning_content`.
112+
113+Mesuré dans un sandbox aarch64 (18 CPU, 31 Go) avec q8_0-safe2, `llama-server` du kit, 16 threads : 35 à 44 tokens/s en génération.
114+
115+## Le kit `llama-cpp-olala`
116+
117+Mixin sbx (kits v3) qui compile le fork une fois, au build du kit, et pose `llama-server` et `llama-cli` dans `/opt/llama-cpp-olala` (liens dans `/usr/local/bin`). Build portable (`GGML_NATIVE=OFF`, toutes les variantes CPU chargées à l'exécution), arm64 (testé) et amd64 (pas encore testé). L'argument `ref` fixe le commit du fork.
118+
119+Pour l'utiliser dans un autre sandbox (syntaxe pas encore testée avec `sbx`) :
120+
121+```bash
122+sbx run claude --kit ./kits/llama-cpp-olala
123+```
added git.sh +136 -0
new file mode 100755
@@ -0,0 +1,136 @@
1+#!/bin/bash
2+message=""
3+case $1 in
4+
5+ # 🎨: art
6+ art)
7+ message="Improve structure / format of the code"
8+ emoji="🎨"
9+ ;;
10+
11+ # 🐛: bug
12+ bug|fix)
13+ message="Fix a bug"
14+ emoji="🐛"
15+ ;;
16+
17+ # 🤓: geek
18+ human|human-fixed)
19+ message="Human Fixed"
20+ emoji="🤓"
21+ ;;
22+
23+ # 🤖: robot
24+ ai|ai-generated)
25+ message="AI generated"
26+ emoji="🤖"
27+ ;;
28+
29+ # ✨: sparkles
30+ sparkles|feature)
31+ message="Introduce new feature(s)"
32+ emoji="✨"
33+ ;;
34+
35+ # 🧩: jigsaw
36+ jigsaw|example|examples|demo|demos)
37+ message="Introduce new example(s)"
38+ emoji="🧩"
39+ ;;
40+
41+
42+ # 📝: memo
43+ memo|doc|documentation)
44+ message="Add or update documentation"
45+ emoji="📝"
46+ ;;
47+
48+ # 🌸: cherry_blossom
49+ gardening|garden|clean|cleaning)
50+ message="Gardening"
51+ emoji="🌸"
52+ ;;
53+
54+ # 🚀: rocket
55+ rocket|deploy)
56+ message="Deploy stuff"
57+ emoji="🚀"
58+ ;;
59+
60+ # 🎉: tada
61+ tada|first)
62+ message="Begin a project"
63+ emoji="🎉"
64+ ;;
65+
66+ # 🚧: construction
67+ construction|wip)
68+ message="Work in progress"
69+ emoji="🚧"
70+ ;;
71+
72+ # 📦️: package
73+ package|build)
74+ message="Add or update compiled files or packages"
75+ emoji="📦️"
76+ ;;
77+
78+ # 📦️: package
79+ release)
80+ message="Create a release"
81+ emoji="📦️"
82+ ;;
83+
84+ # 👽️: alien
85+ alien|api)
86+ message="Update code due to external API changes"
87+ emoji="👽️"
88+ ;;
89+
90+ # 🐳: whale
91+ docker|container)
92+ message="Docker"
93+ emoji="🐳"
94+ ;;
95+
96+ # 🍊: tangerine
97+ gitpod|gitpodify)
98+ message="Gitpodify"
99+ emoji="🍊"
100+ ;;
101+
102+ # 🧪: test tube
103+ alembic|experiments|experiment|xp)
104+ message="Perform experiments"
105+ emoji="🧪"
106+ ;;
107+
108+ # ✅: check mark
109+ test|tests|testing)
110+ message="Add or update tests"
111+ emoji="✅"
112+ ;;
113+
114+ # 💾: floppy-disk
115+ save)
116+ message="Saved"
117+ emoji="💾"
118+ ;;
119+
120+ *)
121+ message="Updated"
122+ emoji="🛟"
123+ ;;
124+
125+esac
126+
127+find . -name '.DS_Store' -type f -delete
128+
129+if [ -z "$2" ]
130+then
131+ # empty
132+ git add .; git commit -m "$emoji $message."; git push
133+else
134+ # not empty
135+ git add .; git commit -m "$emoji $message: $2"; git push
136+fi
new file mode 100755
@@ -0,0 +1,136 @@
1+#!/bin/bash
2+message=""
3+case $1 in
4+
5+ # 🎨: art
6+ art)
7+ message="Improve structure / format of the code"
8+ emoji="🎨"
9+ ;;
10+
11+ # 🐛: bug
12+ bug|fix)
13+ message="Fix a bug"
14+ emoji="🐛"
15+ ;;
16+
17+ # 🤓: geek
18+ human|human-fixed)
19+ message="Human Fixed"
20+ emoji="🤓"
21+ ;;
22+
23+ # 🤖: robot
24+ ai|ai-generated)
25+ message="AI generated"
26+ emoji="🤖"
27+ ;;
28+
29+ # ✨: sparkles
30+ sparkles|feature)
31+ message="Introduce new feature(s)"
32+ emoji="✨"
33+ ;;
34+
35+ # 🧩: jigsaw
36+ jigsaw|example|examples|demo|demos)
37+ message="Introduce new example(s)"
38+ emoji="🧩"
39+ ;;
40+
41+
42+ # 📝: memo
43+ memo|doc|documentation)
44+ message="Add or update documentation"
45+ emoji="📝"
46+ ;;
47+
48+ # 🌸: cherry_blossom
49+ gardening|garden|clean|cleaning)
50+ message="Gardening"
51+ emoji="🌸"
52+ ;;
53+
54+ # 🚀: rocket
55+ rocket|deploy)
56+ message="Deploy stuff"
57+ emoji="🚀"
58+ ;;
59+
60+ # 🎉: tada
61+ tada|first)
62+ message="Begin a project"
63+ emoji="🎉"
64+ ;;
65+
66+ # 🚧: construction
67+ construction|wip)
68+ message="Work in progress"
69+ emoji="🚧"
70+ ;;
71+
72+ # 📦️: package
73+ package|build)
74+ message="Add or update compiled files or packages"
75+ emoji="📦️"
76+ ;;
77+
78+ # 📦️: package
79+ release)
80+ message="Create a release"
81+ emoji="📦️"
82+ ;;
83+
84+ # 👽️: alien
85+ alien|api)
86+ message="Update code due to external API changes"
87+ emoji="👽️"
88+ ;;
89+
90+ # 🐳: whale
91+ docker|container)
92+ message="Docker"
93+ emoji="🐳"
94+ ;;
95+
96+ # 🍊: tangerine
97+ gitpod|gitpodify)
98+ message="Gitpodify"
99+ emoji="🍊"
100+ ;;
101+
102+ # 🧪: test tube
103+ alembic|experiments|experiment|xp)
104+ message="Perform experiments"
105+ emoji="🧪"
106+ ;;
107+
108+ # ✅: check mark
109+ test|tests|testing)
110+ message="Add or update tests"
111+ emoji="✅"
112+ ;;
113+
114+ # 💾: floppy-disk
115+ save)
116+ message="Saved"
117+ emoji="💾"
118+ ;;
119+
120+ *)
121+ message="Updated"
122+ emoji="🛟"
123+ ;;
124+
125+esac
126+
127+find . -name '.DS_Store' -type f -delete
128+
129+if [ -z "$2" ]
130+then
131+ # empty
132+ git add .; git commit -m "$emoji $message."; git push
133+else
134+ # not empty
135+ git add .; git commit -m "$emoji $message: $2"; git push
136+fi
added kits/llama-cpp-olala/llama-cpp-olala-context.md +9 -0
new file mode 100644
@@ -0,0 +1,9 @@
1+# llama.cpp (OVHai Olala fork)
2+
3+`llama-server` and `llama-cli` on the PATH come from the OVHai fork of llama.cpp
4+(`/opt/llama-cpp-olala`, commit in `/opt/llama-cpp-olala/REVISION`), CPU only.
5+They load `olala` GGUF models, which upstream llama.cpp cannot.
6+
7+- Serve: `llama-server -m <file.gguf> -t <threads> -fa 1 -nr -ctk q8_0 -ctv q8_0 -c 32768 -np 4 --host 0.0.0.0 --port 8080`
8+- In this project, prefer `scripts/olala.sh up|serve|stop|status` (downloads the model if needed).
9+- Recommended sampling: `temperature=0.8`, `top_p=0.9`, `top_k=40`; `reasoning_effort`: disabled, low, medium, high.
new file mode 100644
@@ -0,0 +1,9 @@
1+# llama.cpp (OVHai Olala fork)
2+
3+`llama-server` and `llama-cli` on the PATH come from the OVHai fork of llama.cpp
4+(`/opt/llama-cpp-olala`, commit in `/opt/llama-cpp-olala/REVISION`), CPU only.
5+They load `olala` GGUF models, which upstream llama.cpp cannot.
6+
7+- Serve: `llama-server -m <file.gguf> -t <threads> -fa 1 -nr -ctk q8_0 -ctv q8_0 -c 32768 -np 4 --host 0.0.0.0 --port 8080`
8+- In this project, prefer `scripts/olala.sh up|serve|stop|status` (downloads the model if needed).
9+- Recommended sampling: `temperature=0.8`, `top_p=0.9`, `top_k=40`; `reasoning_effort`: disabled, low, medium, high.
added kits/llama-cpp-olala/llama-cpp-olala.dockerfile +61 -0
new file mode 100644
@@ -0,0 +1,61 @@
1+# syntax=docker/dockerfile:1
2+
3+# --- build stage: the OVHai llama.cpp fork, CPU only, portable ------------------
4+# Same flags as the fork's .devops/cpu.Dockerfile: no -march=native, and every CPU
5+# variant built as a backend library loaded at runtime for the CPU it runs on.
6+FROM debian:trixie-slim AS build
7+ARG LLAMA_REF
8+ARG TARGETARCH
9+
10+# A heredoc RUN does not stop at the first failing line by itself: every
11+# script starts with `set -eu`.
12+RUN <<EOF
13+set -eu
14+apt-get update
15+apt-get install -y --no-install-recommends build-essential cmake git ca-certificates libgomp1
16+rm -rf /var/lib/apt/lists/*
17+EOF
18+
19+RUN <<EOF
20+set -eu
21+git init -q /src
22+cd /src
23+git remote add origin https://github.com/OVHai-LLM/llama.cpp.git
24+git fetch -q --depth 1 origin "${LLAMA_REF}"
25+git checkout -q FETCH_HEAD
26+EOF
27+
28+# -DLLAMA_USE_PREBUILT_UI=OFF: the prebuilt web UI does not match the fork (model card).
29+# -DLLAMA_OPENSSL=OFF: llama-server serves plain HTTP; downloads are done with `hf`.
30+# RPATH=$ORIGIN: binaries find their libraries next to them, wherever they are installed.
31+RUN <<EOF
32+set -eu
33+case "${TARGETARCH}" in
34+ amd64|arm64) ;;
35+ *) echo "unsupported TARGETARCH: ${TARGETARCH}" >&2; exit 1 ;;
36+esac
37+cd /src
38+cmake -S . -B build -DCMAKE_BUILD_TYPE=Release \
39+ -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
40+ -DLLAMA_USE_PREBUILT_UI=OFF -DLLAMA_OPENSSL=OFF \
41+ -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_EXAMPLES=OFF \
42+ -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DCMAKE_INSTALL_RPATH='$ORIGIN'
43+cmake --build build -j "$(nproc)" --target llama-server llama-cli
44+EOF
45+
46+RUN <<EOF
47+set -eu
48+dest=/out/opt/llama-cpp-olala
49+mkdir -p "$dest" /out/usr/local/bin
50+cp -P /src/build/bin/llama-server /src/build/bin/llama-cli "$dest"/
51+find /src/build/bin -name '*.so*' -exec cp -P {} "$dest"/ \;
52+# OpenMP runtime: not guaranteed on the sandbox image.
53+cp -L "$(gcc -print-file-name=libgomp.so.1)" "$dest"/libgomp.so.1
54+git -C /src rev-parse HEAD > "$dest"/REVISION
55+ln -s /opt/llama-cpp-olala/llama-server /out/usr/local/bin/llama-server
56+ln -s /opt/llama-cpp-olala/llama-cli /out/usr/local/bin/llama-cli
57+EOF
58+
59+# --- overlay: only the binaries and their libraries land on the workload ---------
60+FROM scratch
61+COPY --from=build /out/ /
new file mode 100644
@@ -0,0 +1,61 @@
1+# syntax=docker/dockerfile:1
2+
3+# --- build stage: the OVHai llama.cpp fork, CPU only, portable ------------------
4+# Same flags as the fork's .devops/cpu.Dockerfile: no -march=native, and every CPU
5+# variant built as a backend library loaded at runtime for the CPU it runs on.
6+FROM debian:trixie-slim AS build
7+ARG LLAMA_REF
8+ARG TARGETARCH
9+
10+# A heredoc RUN does not stop at the first failing line by itself: every
11+# script starts with `set -eu`.
12+RUN <<EOF
13+set -eu
14+apt-get update
15+apt-get install -y --no-install-recommends build-essential cmake git ca-certificates libgomp1
16+rm -rf /var/lib/apt/lists/*
17+EOF
18+
19+RUN <<EOF
20+set -eu
21+git init -q /src
22+cd /src
23+git remote add origin https://github.com/OVHai-LLM/llama.cpp.git
24+git fetch -q --depth 1 origin "${LLAMA_REF}"
25+git checkout -q FETCH_HEAD
26+EOF
27+
28+# -DLLAMA_USE_PREBUILT_UI=OFF: the prebuilt web UI does not match the fork (model card).
29+# -DLLAMA_OPENSSL=OFF: llama-server serves plain HTTP; downloads are done with `hf`.
30+# RPATH=$ORIGIN: binaries find their libraries next to them, wherever they are installed.
31+RUN <<EOF
32+set -eu
33+case "${TARGETARCH}" in
34+ amd64|arm64) ;;
35+ *) echo "unsupported TARGETARCH: ${TARGETARCH}" >&2; exit 1 ;;
36+esac
37+cd /src
38+cmake -S . -B build -DCMAKE_BUILD_TYPE=Release \
39+ -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
40+ -DLLAMA_USE_PREBUILT_UI=OFF -DLLAMA_OPENSSL=OFF \
41+ -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_EXAMPLES=OFF \
42+ -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DCMAKE_INSTALL_RPATH='$ORIGIN'
43+cmake --build build -j "$(nproc)" --target llama-server llama-cli
44+EOF
45+
46+RUN <<EOF
47+set -eu
48+dest=/out/opt/llama-cpp-olala
49+mkdir -p "$dest" /out/usr/local/bin
50+cp -P /src/build/bin/llama-server /src/build/bin/llama-cli "$dest"/
51+find /src/build/bin -name '*.so*' -exec cp -P {} "$dest"/ \;
52+# OpenMP runtime: not guaranteed on the sandbox image.
53+cp -L "$(gcc -print-file-name=libgomp.so.1)" "$dest"/libgomp.so.1
54+git -C /src rev-parse HEAD > "$dest"/REVISION
55+ln -s /opt/llama-cpp-olala/llama-server /out/usr/local/bin/llama-server
56+ln -s /opt/llama-cpp-olala/llama-cli /out/usr/local/bin/llama-cli
57+EOF
58+
59+# --- overlay: only the binaries and their libraries land on the workload ---------
60+FROM scratch
61+COPY --from=build /out/ /
added kits/llama-cpp-olala/llama-cpp-olala.yaml +29 -0
new file mode 100644
@@ -0,0 +1,29 @@
1+# syntax=docker/sandbox-kit:3
2+schemaVersion: "3"
3+kind: mixin
4+displayName: llama.cpp (OVHai Olala fork, CPU)
5+description: llama-server and llama-cli from the OVHai llama.cpp fork, which supports the `olala` architecture (CPU only)
6+
7+args:
8+ ref:
9+ # Commit of https://github.com/OVHai-LLM/llama.cpp to build. Pin a full SHA for reproducible builds.
10+ default: "55c22dca0c987f247c21b427c3b3a09e1a28187f"
11+ pattern: '^[0-9a-f]{40}$'
12+ buildArg: LLAMA_REF
13+
14+# provides needs an X.Y.Z version (a commit SHA is rejected by the kit frontend):
15+# 0.5.0 = what the fork reports (`llama-server --version`: 0.5.0-dev). The commit is pinned by `ref`.
16+provides: ["llama-cpp-olala@0.5.0"]
17+
18+capabilities:
19+ # Runtime access to download GGUF models from Hugging Face (hf download, llama-server -hf).
20+ - type: com.docker.sandbox/network-policy@1
21+ config:
22+ runtime:
23+ allow:
24+ - huggingface.co
25+ - "*.huggingface.co"
26+ - "*.hf.co"
27+ - type: com.docker.sandbox/agent-context@1
28+ config:
29+ contentFile: ./llama-cpp-olala-context.md
new file mode 100644
@@ -0,0 +1,29 @@
1+# syntax=docker/sandbox-kit:3
2+schemaVersion: "3"
3+kind: mixin
4+displayName: llama.cpp (OVHai Olala fork, CPU)
5+description: llama-server and llama-cli from the OVHai llama.cpp fork, which supports the `olala` architecture (CPU only)
6+
7+args:
8+ ref:
9+ # Commit of https://github.com/OVHai-LLM/llama.cpp to build. Pin a full SHA for reproducible builds.
10+ default: "55c22dca0c987f247c21b427c3b3a09e1a28187f"
11+ pattern: '^[0-9a-f]{40}$'
12+ buildArg: LLAMA_REF
13+
14+# provides needs an X.Y.Z version (a commit SHA is rejected by the kit frontend):
15+# 0.5.0 = what the fork reports (`llama-server --version`: 0.5.0-dev). The commit is pinned by `ref`.
16+provides: ["llama-cpp-olala@0.5.0"]
17+
18+capabilities:
19+ # Runtime access to download GGUF models from Hugging Face (hf download, llama-server -hf).
20+ - type: com.docker.sandbox/network-policy@1
21+ config:
22+ runtime:
23+ allow:
24+ - huggingface.co
25+ - "*.huggingface.co"
26+ - "*.hf.co"
27+ - type: com.docker.sandbox/agent-context@1
28+ config:
29+ contentFile: ./llama-cpp-olala-context.md
added sbxenv.dev.yaml +39 -0
new file mode 100644
@@ -0,0 +1,39 @@
1+schemaVersion: "1"
2+
3+# Development sandbox (Claude Code + Go/mini-me toolchains).
4+# Python and uv come from the claude agent kit itself: do not add python-toolchain-kit
5+# (two kits cannot provide the same capability).
6+# The Olala model server is described by sbxenv.yaml (sandbox houla-olala).
7+
8+name: houla-houla
9+
10+agent: claude
11+
12+# Primary workspace: this folder (relative paths resolve against this file).
13+workspace: .
14+
15+additionalWorkspaces:
16+ - path: /Users/k33g/Documents
17+ - path: /Users/k33g/kDrive/Rickub/bots-garden/mini-me
18+
19+kits:
20+ - docker.io/k33g/go-toolchain-kit:0.2.0
21+ - docker.io/k33g/mini-me-agent-kit:0.2.0
22+
23+env:
24+ # Claude Code default model
25+ ANTHROPIC_MODEL: claude-opus-5-5
26+
27+# Lifecycle commands run on the HOST, so we reach into the sandbox with `sbx exec`.
28+# postCreate runs once, right after a new sandbox is created (not on re-attach).
29+lifecycle:
30+ postCreate:
31+ - name: Install the Hugging Face CLI (hf)
32+ command: sbx exec "$SBX_SANDBOX_NAME" bash -lc 'uv tool install "huggingface_hub[cli]" && hf version'
33+ timeout: 5m
34+
35+
36+
37+# sbx env create sbxenv.dev.yaml # shows the plan and asks for confirmation (--force to skip)
38+# sbx env run sbxenv.dev.yaml # re-attach (creates the sandbox first if it does not exist)
39+# sbx env rm sbxenv.dev.yaml # delete the sandbox
new file mode 100644
@@ -0,0 +1,39 @@
1+schemaVersion: "1"
2+
3+# Development sandbox (Claude Code + Go/mini-me toolchains).
4+# Python and uv come from the claude agent kit itself: do not add python-toolchain-kit
5+# (two kits cannot provide the same capability).
6+# The Olala model server is described by sbxenv.yaml (sandbox houla-olala).
7+
8+name: houla-houla
9+
10+agent: claude
11+
12+# Primary workspace: this folder (relative paths resolve against this file).
13+workspace: .
14+
15+additionalWorkspaces:
16+ - path: /Users/k33g/Documents
17+ - path: /Users/k33g/kDrive/Rickub/bots-garden/mini-me
18+
19+kits:
20+ - docker.io/k33g/go-toolchain-kit:0.2.0
21+ - docker.io/k33g/mini-me-agent-kit:0.2.0
22+
23+env:
24+ # Claude Code default model
25+ ANTHROPIC_MODEL: claude-opus-5-5
26+
27+# Lifecycle commands run on the HOST, so we reach into the sandbox with `sbx exec`.
28+# postCreate runs once, right after a new sandbox is created (not on re-attach).
29+lifecycle:
30+ postCreate:
31+ - name: Install the Hugging Face CLI (hf)
32+ command: sbx exec "$SBX_SANDBOX_NAME" bash -lc 'uv tool install "huggingface_hub[cli]" && hf version'
33+ timeout: 5m
34+
35+
36+
37+# sbx env create sbxenv.dev.yaml # shows the plan and asks for confirmation (--force to skip)
38+# sbx env run sbxenv.dev.yaml # re-attach (creates the sandbox first if it does not exist)
39+# sbx env rm sbxenv.dev.yaml # delete the sandbox
added sbxenv.yaml +76 -0
new file mode 100644
@@ -0,0 +1,76 @@
1+schemaVersion: "1"
2+
3+# Sandbox that serves an Olala GGUF model with llama-server (OpenAI-compatible API)
4+# on the host, using the OVHai llama.cpp fork (CPU only) from kits/llama-cpp-olala.
5+#
6+# sbx env create sbxenv.yaml # create + download + serve
7+# sbx env create sbxenv.yaml --env-arg host_port=9090 # another host port
8+# sbx env create sbxenv.yaml --env-arg models_dir=./my-models # another models folder
9+# sbx env run sbxenv.yaml # attach Claude Code
10+# sbx env exec --no-input sbxenv.yaml -- bash -lc '$SBX_WS/scripts/olala.sh serve' # restart after `sbx stop`
11+# sbx env rm sbxenv.yaml # models stay in models_dir
12+#
13+# Every step of scripts/olala.sh is skipped when already done (llama-server found,
14+# model present, server running), so re-running it is cheap.
15+# Use the same --env-arg values with every sbx env command of a given sandbox.
16+
17+name: houla-olala
18+
19+agent: claude
20+
21+workspace: .
22+
23+kits:
24+ # llama-server / llama-cli from the OVHai fork (built by sbx from the local kit).
25+ - ./kits/llama-cpp-olala
26+
27+args:
28+ model_repo:
29+ default: OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF
30+ description: Hugging Face repository of the GGUF file
31+ model_file:
32+ default: Olala-v1.0-mini-7A1B-q8_0-safe2.gguf
33+ description: GGUF file to download and serve (q8_0-safe2 = recommended for CPUs)
34+ models_dir:
35+ default: ./models
36+ description: Where models are stored; relative to the repository root, kept when the sandbox is removed
37+ host_port:
38+ default: "8080"
39+ pattern: "[1-9][0-9]*"
40+ description: Host port that publishes llama-server
41+ threads:
42+ default: "16"
43+ pattern: "[1-9][0-9]*"
44+ description: CPU threads used by llama-server (leave a couple of CPUs free)
45+ ctx:
46+ default: "32768"
47+ pattern: "[1-9][0-9]*"
48+ description: Context size (shared by the parallel slots)
49+
50+env:
51+ ANTHROPIC_MODEL: claude-opus-5-5
52+ OLALA_MODEL_REPO: "${{ env.args.model_repo }}"
53+ OLALA_MODEL_FILE: "${{ env.args.model_file }}"
54+ OLALA_MODELS_DIR: "${{ env.args.models_dir }}"
55+ OLALA_THREADS: "${{ env.args.threads }}"
56+ OLALA_CTX: "${{ env.args.ctx }}"
57+ OLALA_PORT: "8080"
58+ SBX_WS: "${{ env.fileDir }}"
59+
60+ports:
61+ - sandbox: 8080
62+ host: ${{ env.args.host_port }}
63+
64+# Lifecycle commands run on the HOST; they reach into the sandbox with `sbx env exec`.
65+# postCreate runs once, after creation (not on re-attach).
66+lifecycle:
67+ postCreate:
68+ # A local sandbox stops as soon as no session keeps it running (and llama-server with it).
69+ # Detached mode keeps it running until `sbx stop`; it is permanent for this sandbox.
70+ - name: Keep the sandbox running in the background (detached)
71+ command: sbx run -d --name "$SBX_SANDBOX_NAME"
72+ timeout: 2m
73+ - name: Download the model (if missing) and start llama-server
74+ # Flags go BEFORE the env file path: after it, sbx reads every word up to `--` as another path.
75+ command: sbx env exec --name "$SBX_SANDBOX_NAME" --no-input "$SBX_ENV_FILE" -- bash -lc "$SBX_WORKSPACE/scripts/olala.sh up"
76+ timeout: 30m
new file mode 100644
@@ -0,0 +1,76 @@
1+schemaVersion: "1"
2+
3+# Sandbox that serves an Olala GGUF model with llama-server (OpenAI-compatible API)
4+# on the host, using the OVHai llama.cpp fork (CPU only) from kits/llama-cpp-olala.
5+#
6+# sbx env create sbxenv.yaml # create + download + serve
7+# sbx env create sbxenv.yaml --env-arg host_port=9090 # another host port
8+# sbx env create sbxenv.yaml --env-arg models_dir=./my-models # another models folder
9+# sbx env run sbxenv.yaml # attach Claude Code
10+# sbx env exec --no-input sbxenv.yaml -- bash -lc '$SBX_WS/scripts/olala.sh serve' # restart after `sbx stop`
11+# sbx env rm sbxenv.yaml # models stay in models_dir
12+#
13+# Every step of scripts/olala.sh is skipped when already done (llama-server found,
14+# model present, server running), so re-running it is cheap.
15+# Use the same --env-arg values with every sbx env command of a given sandbox.
16+
17+name: houla-olala
18+
19+agent: claude
20+
21+workspace: .
22+
23+kits:
24+ # llama-server / llama-cli from the OVHai fork (built by sbx from the local kit).
25+ - ./kits/llama-cpp-olala
26+
27+args:
28+ model_repo:
29+ default: OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF
30+ description: Hugging Face repository of the GGUF file
31+ model_file:
32+ default: Olala-v1.0-mini-7A1B-q8_0-safe2.gguf
33+ description: GGUF file to download and serve (q8_0-safe2 = recommended for CPUs)
34+ models_dir:
35+ default: ./models
36+ description: Where models are stored; relative to the repository root, kept when the sandbox is removed
37+ host_port:
38+ default: "8080"
39+ pattern: "[1-9][0-9]*"
40+ description: Host port that publishes llama-server
41+ threads:
42+ default: "16"
43+ pattern: "[1-9][0-9]*"
44+ description: CPU threads used by llama-server (leave a couple of CPUs free)
45+ ctx:
46+ default: "32768"
47+ pattern: "[1-9][0-9]*"
48+ description: Context size (shared by the parallel slots)
49+
50+env:
51+ ANTHROPIC_MODEL: claude-opus-5-5
52+ OLALA_MODEL_REPO: "${{ env.args.model_repo }}"
53+ OLALA_MODEL_FILE: "${{ env.args.model_file }}"
54+ OLALA_MODELS_DIR: "${{ env.args.models_dir }}"
55+ OLALA_THREADS: "${{ env.args.threads }}"
56+ OLALA_CTX: "${{ env.args.ctx }}"
57+ OLALA_PORT: "8080"
58+ SBX_WS: "${{ env.fileDir }}"
59+
60+ports:
61+ - sandbox: 8080
62+ host: ${{ env.args.host_port }}
63+
64+# Lifecycle commands run on the HOST; they reach into the sandbox with `sbx env exec`.
65+# postCreate runs once, after creation (not on re-attach).
66+lifecycle:
67+ postCreate:
68+ # A local sandbox stops as soon as no session keeps it running (and llama-server with it).
69+ # Detached mode keeps it running until `sbx stop`; it is permanent for this sandbox.
70+ - name: Keep the sandbox running in the background (detached)
71+ command: sbx run -d --name "$SBX_SANDBOX_NAME"
72+ timeout: 2m
73+ - name: Download the model (if missing) and start llama-server
74+ # Flags go BEFORE the env file path: after it, sbx reads every word up to `--` as another path.
75+ command: sbx env exec --name "$SBX_SANDBOX_NAME" --no-input "$SBX_ENV_FILE" -- bash -lc "$SBX_WORKSPACE/scripts/olala.sh up"
76+ timeout: 30m
added scripts/olala.sh +190 -0
new file mode 100755
@@ -0,0 +1,190 @@
1+#!/usr/bin/env bash
2+# Download a GGUF model and serve it with llama-server (OpenAI-compatible API),
3+# using the OVHai llama.cpp fork (CPU only). Runs INSIDE the sandbox.
4+#
5+# Every step is idempotent: it is skipped when its result already exists.
6+#
7+# scripts/olala.sh setup # find llama-server (kit) or build the fork (skipped if found/built)
8+# scripts/olala.sh download # download the model (skipped if the file exists)
9+# scripts/olala.sh serve # start llama-server in the background (skipped if running)
10+# scripts/olala.sh up # setup + download + serve
11+# scripts/olala.sh stop | restart | status | logs
12+#
13+# Configuration (environment variables, defaults below):
14+# OLALA_MODEL_REPO Hugging Face repo of the GGUF file
15+# OLALA_MODEL_FILE GGUF file name in that repo
16+# OLALA_MODELS_DIR where models are stored; relative paths are relative to the repo root
17+# (default: ./models, kept in the workspace so dropping the sandbox keeps them)
18+# OLALA_PORT port llama-server listens on (0.0.0.0)
19+# OLALA_THREADS CPU threads (default: nproc - 2; using every CPU is ~30% slower)
20+# OLALA_CTX context size
21+# OLALA_PARALLEL parallel slots
22+# OLALA_ALIAS model name exposed by the API (default: file name without .gguf)
23+# OLALA_LOAD_MODE llama-server --load-mode (default: none = read the model into RAM;
24+# mmap over the workspace mount is ~20x slower at generation)
25+# OLALA_EXTRA_ARGS extra llama-server arguments
26+# OLALA_SERVER_BIN llama-server to use (default: the one on PATH, e.g. from the
27+# kits/llama-cpp-olala kit, else the one built in OLALA_SRC_DIR)
28+# OLALA_SRC_DIR where the fork is cloned and built when no llama-server is found
29+set -euo pipefail
30+
31+REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
32+
33+OLALA_MODEL_REPO="${OLALA_MODEL_REPO:-OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF}"
34+OLALA_MODEL_FILE="${OLALA_MODEL_FILE:-Olala-v1.0-mini-7A1B-q8_0-safe2.gguf}"
35+OLALA_MODELS_DIR="${OLALA_MODELS_DIR:-./models}"
36+OLALA_PORT="${OLALA_PORT:-8080}"
37+OLALA_THREADS="${OLALA_THREADS:-$(( $(nproc) > 2 ? $(nproc) - 2 : 1 ))}"
38+OLALA_CTX="${OLALA_CTX:-32768}"
39+OLALA_PARALLEL="${OLALA_PARALLEL:-4}"
40+OLALA_ALIAS="${OLALA_ALIAS:-${OLALA_MODEL_FILE%.gguf}}"
41+OLALA_LOAD_MODE="${OLALA_LOAD_MODE:-none}"
42+OLALA_EXTRA_ARGS="${OLALA_EXTRA_ARGS:-}"
43+OLALA_SRC_DIR="${OLALA_SRC_DIR:-$HOME/olala/llama.cpp-olala}"
44+OLALA_REPO_URL="${OLALA_REPO_URL:-https://github.com/OVHai-LLM/llama.cpp.git}"
45+
46+case "$OLALA_MODELS_DIR" in
47+ /*) ;;
48+ *) OLALA_MODELS_DIR="$REPO_ROOT/${OLALA_MODELS_DIR#./}" ;;
49+esac
50+
51+RUN_DIR="$HOME/.olala"
52+PID_FILE="$RUN_DIR/llama-server.pid"
53+LOG_FILE="$RUN_DIR/llama-server.log"
54+BUILT_BIN="$OLALA_SRC_DIR/build/bin/llama-server"
55+MODEL_PATH="$OLALA_MODELS_DIR/$OLALA_MODEL_FILE"
56+
57+log() { printf '[olala] %s\n' "$*" >&2; }
58+
59+# Prints the llama-server to use, or nothing.
60+server_bin() {
61+ if [ -n "${OLALA_SERVER_BIN:-}" ]; then
62+ echo "$OLALA_SERVER_BIN"
63+ elif command -v llama-server >/dev/null; then
64+ command -v llama-server
65+ elif [ -x "$BUILT_BIN" ]; then
66+ echo "$BUILT_BIN"
67+ fi
68+}
69+
70+setup() {
71+ local bin
72+ bin="$(server_bin)"
73+ if [ -n "$bin" ]; then
74+ log "llama-server found: $bin"
75+ return
76+ fi
77+ log "no llama-server on PATH (kits/llama-cpp-olala not installed): building the fork"
78+ if ! command -v cmake >/dev/null || ! command -v g++ >/dev/null; then
79+ log "installing build tools (build-essential, cmake, libssl-dev)"
80+ sudo apt-get update -qq
81+ sudo apt-get install -y -qq build-essential cmake libssl-dev pkg-config git >/dev/null
82+ fi
83+ if [ ! -d "$OLALA_SRC_DIR/.git" ]; then
84+ log "cloning $OLALA_REPO_URL"
85+ mkdir -p "$(dirname "$OLALA_SRC_DIR")"
86+ git clone --depth 1 "$OLALA_REPO_URL" "$OLALA_SRC_DIR"
87+ fi
88+ log "building (CPU only)"
89+ # -DLLAMA_USE_PREBUILT_UI=OFF: the prebuilt web UI does not match the fork (per the model card).
90+ cmake -S "$OLALA_SRC_DIR" -B "$OLALA_SRC_DIR/build" -DCMAKE_BUILD_TYPE=Release -DLLAMA_USE_PREBUILT_UI=OFF >/dev/null
91+ cmake --build "$OLALA_SRC_DIR/build" -j "$(nproc)" --target llama-server llama-cli >/dev/null
92+ log "built: $BUILT_BIN"
93+}
94+
95+download() {
96+ if [ -s "$MODEL_PATH" ]; then
97+ log "model already present: $MODEL_PATH"
98+ return
99+ fi
100+ if ! command -v hf >/dev/null; then
101+ log "installing the Hugging Face CLI"
102+ uv tool install "huggingface_hub[cli]" >/dev/null
103+ export PATH="$HOME/.local/bin:$PATH"
104+ fi
105+ mkdir -p "$OLALA_MODELS_DIR"
106+ log "downloading $OLALA_MODEL_REPO / $OLALA_MODEL_FILE -> $OLALA_MODELS_DIR"
107+ hf download "$OLALA_MODEL_REPO" "$OLALA_MODEL_FILE" --local-dir "$OLALA_MODELS_DIR"
108+}
109+
110+is_running() {
111+ [ -f "$PID_FILE" ] && kill -0 "$(cat "$PID_FILE")" 2>/dev/null
112+}
113+
114+serve() {
115+ if is_running; then
116+ log "llama-server already running (pid $(cat "$PID_FILE")) on port $OLALA_PORT"
117+ return
118+ fi
119+ local bin
120+ bin="$(server_bin)"
121+ [ -n "$bin" ] || { log "llama-server not found: run '$0 setup'"; exit 1; }
122+ [ -s "$MODEL_PATH" ] || { log "model missing: run '$0 download'"; exit 1; }
123+ mkdir -p "$RUN_DIR"
124+ log "starting $bin: $MODEL_PATH on 0.0.0.0:$OLALA_PORT (threads=$OLALA_THREADS ctx=$OLALA_CTX load-mode=$OLALA_LOAD_MODE)"
125+ # Flags from the model card's "CPU" server example; --host 0.0.0.0 so the port can be published.
126+ # The inner shell writes its own pid then execs llama-server, so the pid file is exact
127+ # (setsid may fork, which made $! point at a process that had already exited).
128+ # shellcheck disable=SC2016,SC2086
129+ nohup setsid bash -c 'echo $$ >"$0"; exec "$@"' "$PID_FILE" \
130+ "$bin" \
131+ -m "$MODEL_PATH" \
132+ --load-mode "$OLALA_LOAD_MODE" \
133+ -t "$OLALA_THREADS" -fa 1 -nr -ctk q8_0 -ctv q8_0 \
134+ -c "$OLALA_CTX" -np "$OLALA_PARALLEL" \
135+ --temp 0.8 --top-p 0.9 --top-k 40 \
136+ --alias "$OLALA_ALIAS" \
137+ --host 0.0.0.0 --port "$OLALA_PORT" \
138+ $OLALA_EXTRA_ARGS \
139+ >"$LOG_FILE" 2>&1 </dev/null &
140+ log "waiting for /health (log: $LOG_FILE)"
141+ for _ in $(seq 1 300); do
142+ if curl -fsS "http://127.0.0.1:$OLALA_PORT/health" >/dev/null 2>&1; then
143+ log "ready: http://127.0.0.1:$OLALA_PORT/v1"
144+ return
145+ fi
146+ if [ -f "$PID_FILE" ] && ! is_running; then
147+ log "llama-server exited, last lines:"
148+ tail -20 "$LOG_FILE" >&2
149+ exit 1
150+ fi
151+ sleep 1
152+ done
153+ log "not ready after 300 s, see $LOG_FILE"
154+ exit 1
155+}
156+
157+stop() {
158+ if is_running; then
159+ local pid
160+ pid="$(cat "$PID_FILE")"
161+ kill "$pid"
162+ for _ in $(seq 1 30); do kill -0 "$pid" 2>/dev/null || break; sleep 0.5; done
163+ kill -0 "$pid" 2>/dev/null && kill -9 "$pid"
164+ rm -f "$PID_FILE"
165+ log "stopped"
166+ else
167+ log "not running"
168+ fi
169+}
170+
171+status() {
172+ if is_running; then
173+ log "running (pid $(cat "$PID_FILE")), model $MODEL_PATH, port $OLALA_PORT"
174+ curl -fsS "http://127.0.0.1:$OLALA_PORT/health" && echo
175+ else
176+ log "not running"
177+ fi
178+}
179+
180+case "${1:-up}" in
181+ setup) setup ;;
182+ download) download ;;
183+ serve) serve ;;
184+ up) setup; download; serve ;;
185+ stop) stop ;;
186+ restart) stop; serve ;;
187+ status) status ;;
188+ logs) tail -f "$LOG_FILE" ;;
189+ *) echo "usage: $0 {up|setup|download|serve|stop|restart|status|logs}" >&2; exit 2 ;;
190+esac
new file mode 100755
@@ -0,0 +1,190 @@
1+#!/usr/bin/env bash
2+# Download a GGUF model and serve it with llama-server (OpenAI-compatible API),
3+# using the OVHai llama.cpp fork (CPU only). Runs INSIDE the sandbox.
4+#
5+# Every step is idempotent: it is skipped when its result already exists.
6+#
7+# scripts/olala.sh setup # find llama-server (kit) or build the fork (skipped if found/built)
8+# scripts/olala.sh download # download the model (skipped if the file exists)
9+# scripts/olala.sh serve # start llama-server in the background (skipped if running)
10+# scripts/olala.sh up # setup + download + serve
11+# scripts/olala.sh stop | restart | status | logs
12+#
13+# Configuration (environment variables, defaults below):
14+# OLALA_MODEL_REPO Hugging Face repo of the GGUF file
15+# OLALA_MODEL_FILE GGUF file name in that repo
16+# OLALA_MODELS_DIR where models are stored; relative paths are relative to the repo root
17+# (default: ./models, kept in the workspace so dropping the sandbox keeps them)
18+# OLALA_PORT port llama-server listens on (0.0.0.0)
19+# OLALA_THREADS CPU threads (default: nproc - 2; using every CPU is ~30% slower)
20+# OLALA_CTX context size
21+# OLALA_PARALLEL parallel slots
22+# OLALA_ALIAS model name exposed by the API (default: file name without .gguf)
23+# OLALA_LOAD_MODE llama-server --load-mode (default: none = read the model into RAM;
24+# mmap over the workspace mount is ~20x slower at generation)
25+# OLALA_EXTRA_ARGS extra llama-server arguments
26+# OLALA_SERVER_BIN llama-server to use (default: the one on PATH, e.g. from the
27+# kits/llama-cpp-olala kit, else the one built in OLALA_SRC_DIR)
28+# OLALA_SRC_DIR where the fork is cloned and built when no llama-server is found
29+set -euo pipefail
30+
31+REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
32+
33+OLALA_MODEL_REPO="${OLALA_MODEL_REPO:-OVHaiLLM/Olala-v1.0-mini-7A1B-GGUF}"
34+OLALA_MODEL_FILE="${OLALA_MODEL_FILE:-Olala-v1.0-mini-7A1B-q8_0-safe2.gguf}"
35+OLALA_MODELS_DIR="${OLALA_MODELS_DIR:-./models}"
36+OLALA_PORT="${OLALA_PORT:-8080}"
37+OLALA_THREADS="${OLALA_THREADS:-$(( $(nproc) > 2 ? $(nproc) - 2 : 1 ))}"
38+OLALA_CTX="${OLALA_CTX:-32768}"
39+OLALA_PARALLEL="${OLALA_PARALLEL:-4}"
40+OLALA_ALIAS="${OLALA_ALIAS:-${OLALA_MODEL_FILE%.gguf}}"
41+OLALA_LOAD_MODE="${OLALA_LOAD_MODE:-none}"
42+OLALA_EXTRA_ARGS="${OLALA_EXTRA_ARGS:-}"
43+OLALA_SRC_DIR="${OLALA_SRC_DIR:-$HOME/olala/llama.cpp-olala}"
44+OLALA_REPO_URL="${OLALA_REPO_URL:-https://github.com/OVHai-LLM/llama.cpp.git}"
45+
46+case "$OLALA_MODELS_DIR" in
47+ /*) ;;
48+ *) OLALA_MODELS_DIR="$REPO_ROOT/${OLALA_MODELS_DIR#./}" ;;
49+esac
50+
51+RUN_DIR="$HOME/.olala"
52+PID_FILE="$RUN_DIR/llama-server.pid"
53+LOG_FILE="$RUN_DIR/llama-server.log"
54+BUILT_BIN="$OLALA_SRC_DIR/build/bin/llama-server"
55+MODEL_PATH="$OLALA_MODELS_DIR/$OLALA_MODEL_FILE"
56+
57+log() { printf '[olala] %s\n' "$*" >&2; }
58+
59+# Prints the llama-server to use, or nothing.
60+server_bin() {
61+ if [ -n "${OLALA_SERVER_BIN:-}" ]; then
62+ echo "$OLALA_SERVER_BIN"
63+ elif command -v llama-server >/dev/null; then
64+ command -v llama-server
65+ elif [ -x "$BUILT_BIN" ]; then
66+ echo "$BUILT_BIN"
67+ fi
68+}
69+
70+setup() {
71+ local bin
72+ bin="$(server_bin)"
73+ if [ -n "$bin" ]; then
74+ log "llama-server found: $bin"
75+ return
76+ fi
77+ log "no llama-server on PATH (kits/llama-cpp-olala not installed): building the fork"
78+ if ! command -v cmake >/dev/null || ! command -v g++ >/dev/null; then
79+ log "installing build tools (build-essential, cmake, libssl-dev)"
80+ sudo apt-get update -qq
81+ sudo apt-get install -y -qq build-essential cmake libssl-dev pkg-config git >/dev/null
82+ fi
83+ if [ ! -d "$OLALA_SRC_DIR/.git" ]; then
84+ log "cloning $OLALA_REPO_URL"
85+ mkdir -p "$(dirname "$OLALA_SRC_DIR")"
86+ git clone --depth 1 "$OLALA_REPO_URL" "$OLALA_SRC_DIR"
87+ fi
88+ log "building (CPU only)"
89+ # -DLLAMA_USE_PREBUILT_UI=OFF: the prebuilt web UI does not match the fork (per the model card).
90+ cmake -S "$OLALA_SRC_DIR" -B "$OLALA_SRC_DIR/build" -DCMAKE_BUILD_TYPE=Release -DLLAMA_USE_PREBUILT_UI=OFF >/dev/null
91+ cmake --build "$OLALA_SRC_DIR/build" -j "$(nproc)" --target llama-server llama-cli >/dev/null
92+ log "built: $BUILT_BIN"
93+}
94+
95+download() {
96+ if [ -s "$MODEL_PATH" ]; then
97+ log "model already present: $MODEL_PATH"
98+ return
99+ fi
100+ if ! command -v hf >/dev/null; then
101+ log "installing the Hugging Face CLI"
102+ uv tool install "huggingface_hub[cli]" >/dev/null
103+ export PATH="$HOME/.local/bin:$PATH"
104+ fi
105+ mkdir -p "$OLALA_MODELS_DIR"
106+ log "downloading $OLALA_MODEL_REPO / $OLALA_MODEL_FILE -> $OLALA_MODELS_DIR"
107+ hf download "$OLALA_MODEL_REPO" "$OLALA_MODEL_FILE" --local-dir "$OLALA_MODELS_DIR"
108+}
109+
110+is_running() {
111+ [ -f "$PID_FILE" ] && kill -0 "$(cat "$PID_FILE")" 2>/dev/null
112+}
113+
114+serve() {
115+ if is_running; then
116+ log "llama-server already running (pid $(cat "$PID_FILE")) on port $OLALA_PORT"
117+ return
118+ fi
119+ local bin
120+ bin="$(server_bin)"
121+ [ -n "$bin" ] || { log "llama-server not found: run '$0 setup'"; exit 1; }
122+ [ -s "$MODEL_PATH" ] || { log "model missing: run '$0 download'"; exit 1; }
123+ mkdir -p "$RUN_DIR"
124+ log "starting $bin: $MODEL_PATH on 0.0.0.0:$OLALA_PORT (threads=$OLALA_THREADS ctx=$OLALA_CTX load-mode=$OLALA_LOAD_MODE)"
125+ # Flags from the model card's "CPU" server example; --host 0.0.0.0 so the port can be published.
126+ # The inner shell writes its own pid then execs llama-server, so the pid file is exact
127+ # (setsid may fork, which made $! point at a process that had already exited).
128+ # shellcheck disable=SC2016,SC2086
129+ nohup setsid bash -c 'echo $$ >"$0"; exec "$@"' "$PID_FILE" \
130+ "$bin" \
131+ -m "$MODEL_PATH" \
132+ --load-mode "$OLALA_LOAD_MODE" \
133+ -t "$OLALA_THREADS" -fa 1 -nr -ctk q8_0 -ctv q8_0 \
134+ -c "$OLALA_CTX" -np "$OLALA_PARALLEL" \
135+ --temp 0.8 --top-p 0.9 --top-k 40 \
136+ --alias "$OLALA_ALIAS" \
137+ --host 0.0.0.0 --port "$OLALA_PORT" \
138+ $OLALA_EXTRA_ARGS \
139+ >"$LOG_FILE" 2>&1 </dev/null &
140+ log "waiting for /health (log: $LOG_FILE)"
141+ for _ in $(seq 1 300); do
142+ if curl -fsS "http://127.0.0.1:$OLALA_PORT/health" >/dev/null 2>&1; then
143+ log "ready: http://127.0.0.1:$OLALA_PORT/v1"
144+ return
145+ fi
146+ if [ -f "$PID_FILE" ] && ! is_running; then
147+ log "llama-server exited, last lines:"
148+ tail -20 "$LOG_FILE" >&2
149+ exit 1
150+ fi
151+ sleep 1
152+ done
153+ log "not ready after 300 s, see $LOG_FILE"
154+ exit 1
155+}
156+
157+stop() {
158+ if is_running; then
159+ local pid
160+ pid="$(cat "$PID_FILE")"
161+ kill "$pid"
162+ for _ in $(seq 1 30); do kill -0 "$pid" 2>/dev/null || break; sleep 0.5; done
163+ kill -0 "$pid" 2>/dev/null && kill -9 "$pid"
164+ rm -f "$PID_FILE"
165+ log "stopped"
166+ else
167+ log "not running"
168+ fi
169+}
170+
171+status() {
172+ if is_running; then
173+ log "running (pid $(cat "$PID_FILE")), model $MODEL_PATH, port $OLALA_PORT"
174+ curl -fsS "http://127.0.0.1:$OLALA_PORT/health" && echo
175+ else
176+ log "not running"
177+ fi
178+}
179+
180+case "${1:-up}" in
181+ setup) setup ;;
182+ download) download ;;
183+ serve) serve ;;
184+ up) setup; download; serve ;;
185+ stop) stop ;;
186+ restart) stop; serve ;;
187+ status) status ;;
188+ logs) tail -f "$LOG_FILE" ;;
189+ *) echo "usage: $0 {up|setup|download|serve|stop|restart|status|logs}" >&2; exit 2 ;;
190+esac