chore(lib): bump vllm/vllm-openai to v0.26.0

[cswaney]

Summary

Two-year jump from `v0.20.0` (Aug 2024) to `v0.26.0` (Jul 2026). The launch template's flags (`--model`, `--port`, `--revision`, `--trust-remote-code`, `--tensor-parallel-size`) are unchanged between the two versions, and the Disable Thinking toggle's `--default-chat-template-kwargs` shape was not touched.

Motivation: v0.20 pre-dates many models we now cache. Concretely, Gemma 4 (v0.21+), gpt-oss (v0.24), and TranslateGemma (v0.26) either fail or are hit-and-miss on the current pin.

Testing status

Runtime override for pre-merge validation:

```shell
BLACKFISH_TEXT_GENERATION_IMAGE=vllm/vllm-openai:v0.26.0 blackfish start
```

Full checklist tracked on the issue — #512. Highlights already green:

• `google/gemma-4-12b-it`
• `google/gemma-3-4b-it`
• `deepseek-ai/DeepSeek-OCR-2`

Still to smoke test before this leaves draft:

• `TinyLlama/TinyLlama-1.1B-Chat-v1.0` (sanity)
• `meta-llama/Llama-3.3-70B-Instruct` (flagship dense regression + multi-GPU)
• `Qwen/Qwen2.5-72B-Instruct` (Qwen 2.5 regression + Disable Thinking flag)
• `openai/gpt-oss-20b` (new arch, v0.24)
• `google/translategemma-4b-it` (new arch, v0.26)
• Parameter checklist (temperature/max_tokens/stop/stream/seed/penalties in one request; deterministic seed check)

Notes

• Follows the earlier tigerflow-ml pattern for tests that used to hard-code the default tag: assert against `DEFAULT_IMAGES` instead of the literal so future bumps don't touch `test_cli_run.py`.
• Opaque test fixtures elsewhere (`test_image_probe.py`, `ImageVersionSelect.test.jsx`, etc.) intentionally not touched — the strings are arbitrary inputs to the units under test, not claims about the default.
• Sub-task of Update container versions #451. chore(web): sync text generation launcher params with the bumped vLLM version #515 (launcher param sync after this bump) is the natural follow-up.

Refs #512

[claude]

Review

Small, well-scoped chore PR — nice work keeping the diff minimal by relying on DEFAULT_IMAGES as the single source of truth.

Code quality

• Clean change: bumps the pin in images.py and updates the two docstring examples (cli/__main__.py, client.py) that referenced the old tag literally, so nothing is left inconsistent.
• Nice touch in test_cli_run.py asserting DEFAULT_IMAGES["text_generation"].sif instead of hardcoding "vllm-openai_v0.20.0" — matches the tigerflow-ml precedent and means the next version bump won't need to touch this test again.
• Leaving the opaque fixtures alone (test_image_probe.py, ImageVersionSelect.test.jsx, test_containers.py, test_utils.py) is the right call — those strings are arbitrary inputs to the units under test, not assertions about the current default, so churning them would add noise without value.

Minor nit

• image_probe.py:45 docstring still cites v0.20.0 as the example of vLLM's tag style ("vLLM publishes v0.20.0, the DDSS images publish a bare 0.1.2"). Purely illustrative and not load-bearing, but since you're already touching version references in this PR it might be worth a quick pass to avoid a stale-looking example (e.g. swap in v0.26.0). Not a blocker either way.

Potential issues / risk

• This is a 2-major-version jump for vLLM (v0.20.0 → v0.26.0), so the main risk isn't in this diff itself but in behavioral drift between versions — flags could have changed defaults, error formats, or health-check timing even where CLI flag names stayed stable. The PR body already tracks this well with the chore(lib): bump vllm/vllm-openai to v0.26 or v0.27 #512 checklist, and the PR is correctly still in draft with several validations outstanding (Llama-3.3-70B multi-GPU, Qwen2.5-72B + Disable Thinking, gpt-oss-20b, translategemma, and the parameter checklist). Worth explicitly confirming the --tensor-parallel-size / multi-GPU path and the --default-chat-template-kwargs (Disable Thinking) behavior before flipping out of draft, since those are the two flags most likely to have semantic (not just presence) changes across a jump this large.
• No CI signal in this diff that actually pulls/runs the new vLLM image (understandably, since that likely requires GPU runners) — just flagging that the checklist in the PR description is effectively the only verification gate here, so it's worth holding the line on completing it before merge.

Test coverage

• Good: the CLI dry-run test now guards the image reference literal via DEFAULT_IMAGES, preventing regressions where the pin flag is echoed correctly but the rendered script body still points at the old default.
• No functional/behavioral tests are needed here since this is purely a config value change — the real verification is the manual model-launch checklist tracked on chore(lib): bump vllm/vllm-openai to v0.26 or v0.27 #512.

Security

• No concerns. The image remains pinned to an explicit tag (not latest), which is good practice for reproducibility and supply-chain hygiene.

Overall: LGTM pending the outstanding smoke tests noted in the PR description. Happy to take another look once the checklist is complete and this comes out of draft.