bump(vllm-stable): v0.25.1 → v0.26.0 — DRAFT until live-boot validation

[noonghunna]

Mechanical half of the pin bump, per the checklist — do not merge until the GPU-session gates below pass.

Done (this PR)

• engines/vllm-stable.yml install.spec + display_name → v0.26.0
• 22 compose ${VLLM_IMAGE:-…} defaults bumped (launchers inject the profile image regardless — the compose default is for direct docker compose users)
• 7 fixture literals updated (test-launch-compat ×4, test-submit-bench ×3)
• Targeted guards green: launch-compat · submit-bench · diagnose-profile · patch-attribution · compose-image-drift
• Overlay-free engine (CONTRACT-5) → the re-validate-every-patch step is a no-op, same as Bump vllm-stable pin to v0.25.1 (arm waived; single-card live-validated) #682
• UPSTREAM.md prep row added

Gates before un-drafting (GPU session)

• Live boot via switch.sh (image ground-truthed via docker inspect), 3 warm-ups before judging output
• arch_patches.yml: vllm-stable@v0.26.0 loads:true rows — added ONLY after the boot proves it (v0.25.1 rows stay as history)
• verify-full (MODEL=<served-name>) + bench parity vs the fresh v0.25.1 baseline (dual-fast 68.5/92.1 decode, 2026-07-27 — same-rig, same-power comparison ready-made)
• Confirm the vendored pr48375 overlay still applies/is still needed (release notes show no #48375 merge; notable adjacent work: partial prefix-cache hits for hybrids, vllm#46384 — check interaction)
• While booted: check the All TurboQuant KV dtypes fail to boot on stock vLLM v0.25.1 — profiling path passes 'auto' to the TQ backend (no active slug affected; TQ composes are Genesis-era, archived) #788 TQ 'auto' regression status on this tag (academic — no active slug)
• dated learnings/qwen3.6-27b.md entry
• First planned workloads on this pin: ThinkingCap + Tess VLLM_USE_FLASHINFER_SAMPLER=0 paired bench (maintainer 2026-07-27)

🤖 Generated with Claude Code

[noonghunna]

Release-notes read (2026-07-27) — additions to the validation checklist:

• KV-cache layout refactor packing K/V into the content dim across backends (#44455) — exactly the class of change that can move our KV plumbing. While booted: re-check the #788 TQ 'auto' bug (may be fixed or reshaped), re-verify the backend-tier candidate lists (fp8→FLASHINFER, int8-PTH→TRITON-only) still hold on this tag, and treat the pr48375 overlay apply-check as mandatory (the caching work below touches the same region).
• Hybrid caching pair: partial prefix-cache hits (#46384) + selective hybrid cache retention (#47782) — changes hybrid prefix-cache economics (currently all-or-nothing per the 1600-token block rule). Interacts with the vendored pr48375 overlay AND may shift the mamba-align ~7% narrative-decode cost measured on ThinkingCap — the fleet mamba-align A/B should run on THIS pin, not v0.25.1.
• Qwen3.5-family perf work: RMSNorm+all-reduce fusion (#46998), MoE all-reduce→reduce-scatter (#47006), mamba-state-copy vectorization (#48110) — our exact arch class; bench parity vs 68.5/92.1 may come back better than parity. Don't misread a genuine engine win as noise.
• Sampler: logits no longer upcast to fp32 (#48641) — sampler cost structure changed; the v0.25.1-measured VLLM_USE_FLASHINFER_SAMPLER=0 deltas (fix(qwen3.6-27b): FlashInfer sampler OFF on the 2×3090 MTP composes #772) do NOT transplant. The planned ThinkingCap+Tess sampler A/B on this pin is now load-bearing for the qwen composes' =0 too — consider re-running the flagship arm.
• Hybrid (SWA+full) DFlash drafters (#47914) — scope-check needed: SWA+full ≈ Gemma-4-class, not our GDN hybrids (vllm#40898 likely still blocks Qwen/Tess DFlash). If it covers Gemma-4: a stock-vLLM DFlash path for gemma just as the beellama retirement removed the old one — worth a probe.
• Thinking-budget TPOT optimization with spec-decode (#46662) + separate drafter kv_cache_dtype (#48787) — free win for bounded-thinking users; new knob to note in DTYPE/compose docs later.
• Non-events confirmed: no #48375 (overlay rides), no custom-AR gate change, removals (TeleChat/Persimmon/Fuyu) don't touch us. Bundled FlashInfer moves to 0.6.14.

[noonghunna]

Closing without merging — superseded by timing, not by defect. Since this draft was opened we established that v0.26.0 (published 2026-07-27) does not contain the vllm#50021 fix our nine MTP-on hybrid slugs are caveated for (the PR was still open when that release was cut). Bumping to it would cost a full pin-bump validation cycle for a release that clears none of our standing exposure, and then a second full cycle for whichever release does. Maintainer policy set on #838/#847: one bump, to the first release containing #50021 — at which point the caveat-drop, the #848 status reverts, and the #758/#827 MTP-on re-validations all ride the same checklist. The branch stays as the checklist scaffold for that bump. Trigger recorded on the docs/UPSTREAM.md row.