[fork] Bump base tag to v0.26.0 and refresh the patch series

.github/workflows/build-vllm-audio.yml

@@ -38,7 +38,7 @@ env:
   IMAGE_NAME: docker.io/openimage/vllm-openai-audio
   # Deterministic upstream release the patch series is generated against.
   # Bump this in lockstep with fork/scripts/refresh-patches.sh on each release.
-  DEFAULT_BASE_TAG: v0.25.1
+  DEFAULT_BASE_TAG: v0.26.0
 
 jobs:
   # ---------------------------------------------------------------------------

FORK.md

@@ -15,17 +15,20 @@ fork-owned file lives under `fork/` (plus one CI workflow). You can always
 **0001 — [#47953](https://github.com/vllm-project/vllm/pull/47953): Restrict
 embedding-width share guard to EAGLE drafts.**
 Fixes a Gemma-4 MTP boot crash (`mat1 and mat2 ... 6400/10752` at
-`pre_projection`). v0.25.1 carries the #43957 regression; without this, V1 + MTP
-won't boot.
+`pre_projection`). v0.26.0 still carries the #43957 regression; without this,
+V1 + MTP won't boot.
 
 **0002 — [#44993](https://github.com/vllm-project/vllm/pull/44993): Advance
 grammar across reasoning boundary.**
 Fixes structured-output `{{` / `{"{` corruption under reasoning + spec decode
 (#43388). The grammar must advance at the true reasoning boundary; the
 placeholder-derived delta window misses `</think>` when drafts are rejected.
 
-Both are **pure-Python** upstream backports. Patch 0002 carries only the PR's
-source changes (its test file is not present in the runtime image).
+Both are **pure-Python** upstream backports, byte-identical to the merged
+upstream commits. Patch 0002 carries only the PR's source changes (its test file
+is not present in the runtime image). Both PRs merged *after* `v0.26.0` was cut,
+so they are still not in any release — drop them once a release we rebase onto
+contains them.
 
 Each patch is filed with full context — impact, root cause, a **reproduce case**
 to re-check relevance, validation, and ruled-out theories — under
@@ -49,7 +52,7 @@ Two things are deliberately decoupled:
   patches. It is *not* what we build.
 - **The image** is built from a pinned release tag — `DEFAULT_BASE_TAG` in
   [`.github/workflows/build-vllm-audio.yml`](.github/workflows/build-vllm-audio.yml),
-  currently **`v0.25.1`**.
+  currently **`v0.26.0`**.
 
 The patch files in `fork/patches/` are generated against that exact tag, which
 is why they apply with no fuzz. If a patch ever fails to apply, the image build
@@ -58,22 +61,24 @@ ship an image whose patches silently did nothing.
 
 ## Lockstep with upstream releases
 
-When vLLM cuts a new release (e.g. `v0.26.0`):
+When vLLM cuts a new release (e.g. `v0.27.0`):
 
 ```bash
 # 1. Rebase the patch series onto the new tag (verifies + regenerates).
-fork/scripts/refresh-patches.sh v0.26.0
+fork/scripts/refresh-patches.sh v0.27.0
 
 # 2. Bump the base tag the image builds from.
-#    edit .github/workflows/build-vllm-audio.yml -> DEFAULT_BASE_TAG: v0.26.0
+#    edit .github/workflows/build-vllm-audio.yml -> DEFAULT_BASE_TAG: v0.27.0
 
 # 3. Review the regenerated patches, commit, push. Pushing to main (or running
-#    the workflow) builds and publishes openimage/vllm-openai-audio:v0.26.0.
+#    the workflow) builds and publishes openimage/vllm-openai-audio:v0.27.0.
 ```
 
-If `refresh-patches.sh` reports a patch no longer applies, rebase that patch by
-hand — or, if the upstream PR has since merged into the release, drop it from
-`fork/patches/series` entirely.
+Before step 1, check whether each patch's upstream PR already landed in the new
+release — `git merge-base --is-ancestor <merge-commit> <tag>`, with the merge
+commit recorded in the patch's note. If it did, drop the patch from
+`fork/patches/series` instead of rebasing it. Otherwise, if
+`refresh-patches.sh` reports a patch no longer applies, rebase it by hand.
 
 **CI hygiene.** This fork keeps only its own workflow
 (`.github/workflows/build-vllm-audio.yml`); upstream's governance/lint workflows
@@ -87,22 +92,22 @@ The canonical integrated tree is the `fork/<tag>` branch (the release tag with
 the patch series applied as discrete commits):
 
 ```bash
-git fetch origin fork/v0.25.1
-git log --oneline v0.25.1..origin/fork/v0.25.1   # exactly the two patches
+git fetch origin fork/v0.26.0
+git log --oneline v0.26.0..origin/fork/v0.26.0   # exactly the two patches
 ```
 
 Or apply a single patch against a fresh checkout to inspect it in isolation:
 
 ```bash
-git worktree add /tmp/v0.25.1 v0.25.1
-cd /tmp/v0.25.1
+git worktree add /tmp/v0.26.0 v0.26.0
+cd /tmp/v0.26.0
 git apply --check fork/patches/0001-restrict-embedding-width-guard-to-eagle-pr47953.patch
 ```
 
 ## The image
 
 - **Registry / name:** `docker.io/openimage/vllm-openai-audio`
-- **Tags:** the upstream base tag (e.g. `v0.25.1`) and `latest`.
+- **Tags:** the upstream base tag (e.g. `v0.26.0`) and `latest`.
 - **Drop-in:** entrypoint is inherited from `vllm/vllm-openai`, so it replaces
   the stock image directly.
 - **CI:** [`build-vllm-audio.yml`](.github/workflows/build-vllm-audio.yml) —

fork/README.md

@@ -43,8 +43,8 @@ against a new tag with `scripts/refresh-patches.sh <tag>`.
 ```bash
 # from the repo root
 docker build -f fork/docker/Dockerfile.audio \
-  --build-arg BASE_TAG=v0.25.1 \
-  -t openimage/vllm-openai-audio:v0.25.1 .
+  --build-arg BASE_TAG=v0.26.0 \
+  -t openimage/vllm-openai-audio:v0.26.0 .
 ```
 
 The build fails loudly if any patch does not apply to `BASE_TAG` — that is the

fork/docker/Dockerfile.audio

@@ -17,10 +17,10 @@
 # The entrypoint is inherited unchanged, so this is a drop-in replacement for
 # vllm/vllm-openai:${BASE_TAG}.
 #
-# BASE_TAG MUST be a real upstream release tag (e.g. v0.25.1). The patch series
+# BASE_TAG MUST be a real upstream release tag (e.g. v0.26.0). The patch series
 # is generated against that tag; bumping BASE_TAG without refreshing the series
 # (fork/scripts/refresh-patches.sh) will fail the build by design.
-ARG BASE_TAG=v0.25.1
+ARG BASE_TAG=v0.26.0
 FROM vllm/vllm-openai:${BASE_TAG}
 
 # `pip show` reads metadata WITHOUT importing vllm, so this stays fast and needs

fork/patches/0001-restrict-embedding-width-guard-to-eagle-pr47953.patch

@@ -1,13 +1,13 @@
 # Fork patch 0001 — Restrict embedding-width share guard to EAGLE drafts
 # Upstream PR: https://github.com/vllm-project/vllm/pull/47953
-# Generated against tag: v0.25.1
+# Generated against tag: v0.26.0
 # Apply from the vLLM package root (site-packages/ or repo root): patch -p1 < this-file
 #
 diff --git a/vllm/v1/spec_decode/llm_base_proposer.py b/vllm/v1/spec_decode/llm_base_proposer.py
-index 756c5f3b3..45b463737 100644
+index f8b52d079..6ee442af5 100644
 --- a/vllm/v1/spec_decode/llm_base_proposer.py
 +++ b/vllm/v1/spec_decode/llm_base_proposer.py
-@@ -1468,9 +1468,13 @@ class SpecDecodeBaseProposer:
+@@ -1477,9 +1477,13 @@ class SpecDecodeBaseProposer:
                      "Sharing target model embedding weights with the draft model."
                  )
 

fork/patches/0002-advance-grammar-across-reasoning-boundary-pr44993.patch

@@ -1,13 +1,13 @@
 # Fork patch 0002 — Advance grammar across reasoning boundary (source only)
 # Upstream PR: https://github.com/vllm-project/vllm/pull/44993
-# Generated against tag: v0.25.1
+# Generated against tag: v0.26.0
 # Apply from the vLLM package root (site-packages/ or repo root): patch -p1 < this-file
 #
 diff --git a/vllm/v1/core/sched/scheduler.py b/vllm/v1/core/sched/scheduler.py
-index 8e01035ae..d079a100e 100644
+index b77d01410..d979bd7d1 100644
 --- a/vllm/v1/core/sched/scheduler.py
 +++ b/vllm/v1/core/sched/scheduler.py
-@@ -1635,7 +1635,9 @@ class Scheduler(SchedulerInterface):
+@@ -1729,7 +1729,9 @@ class Scheduler(SchedulerInterface):
                  request.status = RequestStatus.FINISHED_STOPPED
                  stopped = True
 
@@ -19,7 +19,7 @@ index 8e01035ae..d079a100e 100644
                  assert struct_output_request is not None
                  grammar = struct_output_request.grammar
 diff --git a/vllm/v1/structured_output/__init__.py b/vllm/v1/structured_output/__init__.py
-index 34f775257..27a4b26f8 100644
+index 939a387ee..7619a7fc2 100644
 --- a/vllm/v1/structured_output/__init__.py
 +++ b/vllm/v1/structured_output/__init__.py
 @@ -7,7 +7,6 @@ from concurrent.futures import Future, ThreadPoolExecutor
@@ -48,7 +48,7 @@ index 34f775257..27a4b26f8 100644
  class StructuredOutputManager:
      """Engine-level manager for structured output requests."""
 
-@@ -368,7 +363,11 @@ class StructuredOutputManager:
+@@ -379,7 +374,11 @@ class StructuredOutputManager:
              return request.structured_output_request.reasoning_ended
          return True
 
@@ -61,7 +61,7 @@ index 34f775257..27a4b26f8 100644
          if not request.use_structured_output:
              return False
 
-@@ -391,37 +390,40 @@ class StructuredOutputManager:
+@@ -402,37 +401,36 @@ class StructuredOutputManager:
          if structured_req.reasoning_ended:
              return True
 
@@ -88,9 +88,7 @@ index 34f775257..27a4b26f8 100644
 +            start = len(all_token_ids) - len(new_token_ids)
 +            delta_ids: Iterable[int] = new_token_ids
 +        else:
-+            delta_from = (
-+                request.num_computed_tokens - request.num_output_placeholders
-+            )
++            delta_from = request.num_computed_tokens - request.num_output_placeholders
 +            start = (
 +                delta_from
 +                if delta_from >= 0
@@ -121,9 +119,7 @@ index 34f775257..27a4b26f8 100644
 -                )
 -                return True
 +            # Record the boundary so the scheduler can exclude reasoning tokens.
-+            end_index = self._find_reasoning_end_index(
-+                reasoner, all_token_ids, start
-+            )
++            end_index = self._find_reasoning_end_index(reasoner, all_token_ids, start)
 +
 +            structured_req.reasoning_end_token_index = end_index
 +            return True

fork/patches/notes/0001-gemma4-mtp-boot-crash.md

@@ -5,9 +5,9 @@
 | **Patch file** | [`../0001-restrict-embedding-width-guard-to-eagle-pr47953.patch`](../0001-restrict-embedding-width-guard-to-eagle-pr47953.patch) |
 | **Upstream PR** | <https://github.com/vllm-project/vllm/pull/47953> |
 | **File touched** | `vllm/v1/spec_decode/llm_base_proposer.py` |
-| **Applied on** | `v0.25.1` |
-| **Upstream status** | Open, mergeable. Related: issue #47794, sibling PR #47833 ("Always share target embeddings for MTP"). |
-| **Drop this patch when** | #47953 (or an equivalent such as #47833) lands in a release we rebase onto — verify with the reproduce below, then remove it from `../series`. |
+| **Applied on** | `v0.26.0` |
+| **Upstream status** | **Merged** 2026-07-21 as `b2b8f679d` — *after* `v0.26.0` was cut (2026-07-20), so it is in no release yet. Related: issue #47794, sibling PR #47833 ("Always share target embeddings for MTP"). |
+| **Drop this patch when** | `b2b8f679d` is an ancestor of the tag we rebase onto (`git merge-base --is-ancestor b2b8f679d <tag>`) — expected in the release after `v0.26.0`. Then remove it from `../series`. |
 
 ## Why it hurts us (impact)
 
@@ -42,11 +42,13 @@ MTP drafts, which is wrong for Gemma-4:
 4. `pre_projection` is `Linear(2 * backbone_hidden_size = 10752, ...)`, so the
    **6400**-wide input cannot be multiplied by the **10752**-wide weight. Crash.
 
-This is a **`0.25.x`-only regression**: before #43957 (e.g. the pre-regression
-nightly `34b560b72`, `v0.23.1rc1.dev786`), vLLM shared the target embedding
-unconditionally for MTP, so `combined = 10752` and it worked. The `pre_projection`
-weight is genuinely **10752**-wide in the checkpoint, so the correct fix is to
-**share** the embedding, not to resize anything.
+This regression is present in every release from `0.25.0` through **`v0.26.0`**
+(fixed on `main` only by #47953, merged after the `v0.26.0` cut): before #43957
+(e.g. the pre-regression nightly `34b560b72`, `v0.23.1rc1.dev786`), vLLM shared
+the target embedding unconditionally for MTP, so `combined = 10752` and it
+worked. The `pre_projection` weight is genuinely **10752**-wide in the
+checkpoint, so the correct fix is to **share** the embedding, not to resize
+anything.
 
 **The fix** restricts the width guard to EAGLE drafts. EAGLE draft modules define
 `has_own_embed_tokens`; MTP drafts do not (the code already branches on
@@ -71,8 +73,9 @@ VLLM_USE_V2_MODEL_RUNNER=0 vllm serve <gemma-4-31B-it-FP8-block> \
   --max-model-len 8192
 ```
 
-- **Stock `v0.25.1`:** crashes at boot; the log contains `Keeping separate
-  embedding weights` and the `6400`/`10752` shape error.
+- **Stock `v0.26.0`:** crashes at boot; the log contains `Keeping separate
+  embedding weights` and the `6400`/`10752` shape error. (The guard is byte-for-byte
+  unchanged since `v0.25.1`, where this was observed directly.)
 - **With patch 0001:** boots; the log contains only `Sharing target model
   embedding weights with the draft model`, and the server serves a completion.
 
@@ -87,6 +90,11 @@ Applied on `vllm/vllm-openai:v0.25.1` with the 31B main and its MTP draft: befor
 patch is effective. Confirmed on H100 (SM90) and L40S (SM89). The built fork image
 was re-validated the same way on **2026-07-18**.
 
+**2026-07-25 — `v0.26.0` rebase, static verification only.** `_maybe_share_embeddings`
+is byte-identical in `v0.25.1` and `v0.26.0`, so the crash carries over unchanged;
+the regenerated patch matches upstream's merged commit `b2b8f679d` exactly and
+applies to `v0.26.0` with no fuzz. Not re-run on hardware.
+
 ## Ruled out (do not re-explore)
 
 An early hypothesis that the MTP draft was inheriting the fp8-block main's

fork/patches/notes/0002-structured-output-reasoning-corruption.md

@@ -5,9 +5,9 @@
 | **Patch file** | [`../0002-advance-grammar-across-reasoning-boundary-pr44993.patch`](../0002-advance-grammar-across-reasoning-boundary-pr44993.patch) |
 | **Upstream PR** | <https://github.com/vllm-project/vllm/pull/44993> |
 | **Files touched** | `vllm/v1/core/sched/scheduler.py`, `vllm/v1/structured_output/__init__.py` |
-| **Applied on** | `v0.25.1` (source changes only; the PR's test file is not needed in the runtime image) |
-| **Upstream status** | Open, mergeable. Closes issues #43388, #34650, #48228. Related: #41967 (Gemma-4 + MTP dropping the first tool-call arguments — same root cause in the tool-call path). |
-| **Drop this patch when** | #44993 lands in a release we rebase onto — verify with the reproduce below, then remove it from `../series`. |
+| **Applied on** | `v0.26.0` (source changes only; the PR's test file is not needed in the runtime image) |
+| **Upstream status** | **Merged** 2026-07-24 as `0416dab27` — *after* `v0.26.0` was cut (2026-07-20), so it is in no release yet. Closes issues #43388, #34650, #48228. Related: #41967 (Gemma-4 + MTP dropping the first tool-call arguments — same root cause in the tool-call path). |
+| **Drop this patch when** | `0416dab27` is an ancestor of the tag we rebase onto (`git merge-base --is-ancestor 0416dab27 <tag>`) — expected in the release after `v0.26.0`. Then remove it from `../series`. |
 
 ## Why it hurts us (impact)
 
@@ -77,7 +77,8 @@ VLLM_USE_V2_MODEL_RUNNER=0 vllm serve google/gemma-4-E2B-it \
 Drive it with structured (native) JSON requests that also enable thinking, at
 concurrency around 16.
 
-- **Stock `v0.25.1`:** roughly 75–100% of responses corrupt with `{{`.
+- **Stock `v0.26.0`:** roughly 75–100% of responses corrupt with `{{`. (Measured
+  on `v0.25.1`; both touched paths are unchanged in `v0.26.0`.)
 - **With patch 0002:** 0% corrupt.
 
 **Relevance check:** if stock (unpatched) shows 0% corruption here, the upstream
@@ -90,6 +91,12 @@ On Gemma-4-E2B, applying both coupled changes moved corruption from
 with no errors. Because the fault is in the shared grammar-bitmask path and not in
 any model-specific code, the fix carries to the 31B production model.
 
+**2026-07-25 — `v0.26.0` rebase, static verification only.** `should_advance` and
+the scheduler call site are byte-identical in `v0.25.1` and `v0.26.0`, so the bug
+carries over unchanged; the regenerated patch matches upstream's merged commit
+`0416dab27` exactly, and the patched files are byte-identical to `upstream/main`.
+Not re-run on hardware.
+
 ## Ruled out (do not re-explore)
 
 - **Parallel-vs-serial bitmask path** — the parallel threshold is 128 and the bug