#6103 [Other] Publish the Docker images for v0.2.14 and v0.2.15 (ghcr still stops at 0.2.13-cuda13)

open · @simlu · 3 comentários · No GitHub

Setup & installServer & APINVIDIA / CUDAModels & quantsSecurityLinux

Descrição

**Area:** [Other] (setup and installation — Docker images)

Could you run the manual **Docker images** workflow for v0.2.14 and v0.2.15? As of today (2026-11-16) the container registry still stops at `0.2.13-cuda13`, so container users cannot move off the 0.2.13 engine at all, while a forward pin to a not-yet-built tag breaks deployments outright (see the failure under "Workaround today", including the exact error).

Why these two publishes matter for containers specifically:

- The main engine is baked into the image (Dockerfile header: "built from the engine COMPONENT PACKAGE"), so unlike app/Terminal users, container users take engine behavior changes **only** when an image tag exists and is pulled. Numeric image tags are also the only lever container users have to control **when** the v0.2.14 prompt-kernel rewrite lands ("Answers can differ from 0.2.13 in some words") and to hold back v0.2.14's DeepSeek retirement. Right now the newest image that can be pulled is 0.2.13 (engine 0.1.59), and even that has to be pinned explicitly because `latest-cuda13` is only correct by accident.
- v0.2.15's control-side fixes that do affect the container path (Terminal/serve run inside the image) can only reach us via a new image: "a drafter you choose stays chosen" (the persisted `drafting` value now survives starts without a drafting word — this rig pins MTP through exactly that persisted value) and serve #6055 (requests arriving during shutdown get a 503 `server_draining` instead of being accepted and cut with 0 tokens).

Both images look cheap to build:

- The **v0.2.14** release already carries every engine package the workflow needs (including `strata-engine-0.1.60-linux-x86_64-cuda-cuda13.tar.gz`, 641 MiB), so a dispatch with `release_tag: v0.2.14` should work today.
- The **v0.2.15** release carries no engine assets of its own ("The engines stay at 0.2.14 ... the same files, which the update feed downloads from the 0.2.14 release"), so the v0.2.15 image would presumably run with the v0.2.14 packages plus `source_ref: v0.2.15` for the Terminal/control sources — whatever combination fits `docker/fetch-engines.py`; your call. The two images share the identical 0.1.60 engine regardless.

Related existing thread (so this does not look like an unsearched duplicate): #5164 (open PR) addresses the root cause by auto-triggering the workflow on `release: published`. This issue is only the immediate ask to run the **manual** dispatch for v0.2.14/v0.2.15 in the meantime. Happy to move this evidence into #5164 and close this if you prefer one thread.

Verified evidence (all of the below run today, 2026-11-16; the verification commands are under "How to verify on your side" below):

- `ghcr.io/v2/niko1221/strata/tags/list` newest cuda13 tag: `0.2.13-cuda13`; manifests for `0.2.14-cuda13` and `0.2.15-cuda13` return HTTP 404.
- `docker-content-digest` of `latest-cuda13` is byte-identical to `0.2.13-cuda13` (`sha256:b6e1876515d5663047a7cd8dfee5223c8b2d117710eb2a9c285a433fb6e4a703`).
- Actions history for "Docker images": last run 2026-11-09T12:57:14Z (workflow_dispatch, success, main@37d5c9e8) — nothing since, consistent with the workflow header "Run by hand only ... Nothing starts it on its own."

Reproduction status: reproduced locally (real failed deploy on 2026-11-16, error under "Workaround today") plus the registry/Actions state checked today.

Minor optional idea, no strong feelings: a one-line note on the release ("Docker images for this release will follow when the manual workflow runs") would let deployers distinguish "not yet published" from "published with a bad tag" without probing the registry.


### Workaround today, and the deploy failure

We pin `ghcr.io/niko1221/strata:0.2.13-cuda13` — the newest published tag — which keeps the deployment on engine 0.1.59 and defers the v0.2.14 prompt-path changes indefinitely. The alternative we tried first is what fails:

Environment: Ubuntu Linux x86_64, 4x RTX 3090 (24 GB each), 256 GB RAM, NVIDIA CUDA 13 driver; Strata container running unsloth Qwen `Q8_0` at 256K context, layer split across the four cards, MTP drafter pinned.

```text
$ docker compose up llm-gateway llama-deepseek strata-qwen-flash
Error response from daemon: failed to resolve reference
"ghcr.io/niko1221/strata:0.2.15-cuda13":
ghcr.io/niko1221/strata:0.2.15-cuda13: not found
```

Worth noting for other container deployers: `docker compose up` resolves **every** service's image first, so one forward-pinned tag like this aborts unrelated services in the same project (our gateway and the llama.cpp service refused to start too). Building locally with `docker/build.sh` from the v0.2.14 package would work but means self-maintaining an image that used to be one `pull` away, so we would rather wait for the published tag.

How to verify on your side:

```text
# token: curl -s "https://ghcr.io/token?scope=repository:niko1221/strata:pull"
curl -s -H "Authorization: Bearer $TOK" https://ghcr.io/v2/niko1221/strata/tags/list
curl -s -o /dev/null -w "%{http_code}\n" -H "Authorization: Bearer $TOK" \
  -H "Accept: application/vnd.oci.image.index.v1+json" \
  https://ghcr.io/v2/niko1221/strata/manifests/0.2.15-cuda13        # 404 today
```

No site

Links install, modelos, releases.