Fix v0.25 W2 starvation and structured-output MTP failures

.gitattributes

@@ -0,0 +1,2 @@
+# Generated patch artifacts can end in a structural blank context line.
+patch/vllm-moet-v0.25.0.patch whitespace=-blank-at-eof

.github/workflows/bench-lint.yml

@@ -11,10 +11,14 @@ on:
       - "bench/**"
       - "docker/**"
       - "Dockerfile.recipes"
+      - "Dockerfile.sm120-v025"
       - "README.md"
+      - "docs/v025-port.md"
       - "docs/benchmarks/**"
       - "patch/**"
+      - "tests/test_check_patch_files.py"
       - "tools/check_patch_files.py"
+      - ".gitattributes"
       - ".github/workflows/bench-lint.yml"
 
 jobs:
@@ -26,8 +30,19 @@ jobs:
         with:
           python-version: "3.12"
       - run: pip install pyyaml
-      - name: Patch file list matches patch/FILES.txt (lost-line guard)
-        run: python3 tools/check_patch_files.py
+      - name: Clone the production v0.25 source lineage
+        run: |
+          git clone --filter=blob:none --single-branch --branch moet-v0.25.0 https://github.com/OmarB97/vllm.git .vllm-source-v025
+          git -C .vllm-source-v025 remote rename origin fork
+          git -C .vllm-source-v025 fetch --force fork refs/tags/v0.25.0:refs/tags/v0.25.0
+      - name: Patch guard regressions
+        run: python3 -m unittest discover -s tests -p "test_check_patch_files.py"
+      - name: Patch manifests and production source bindings match
+        env:
+          VLLM_MOET_FORK: ${{ github.workspace }}/.vllm-source-v025
+        run: |
+          python3 tools/check_patch_files.py
+          python3 tools/check_patch_files.py --version 0.25.0
       - name: Lint recipes, boxes, suites, matrix, results
         run: python3 bench/runner/lint.py
       - name: Verify README table + report match committed results

AGENTS.md

@@ -7,13 +7,16 @@ every rule below traces back to a real incident.
 
 ## The iron rule
 
-`patch/vllm-moet-v0.24.0.patch` is a **generated artifact**: byte-for-byte
-`git diff v0.24.0 moet-v0.24.0` from the vllm fork clone. It is **never
-edited by hand**, never patched incrementally, never regenerated from
-anything but the fork branch. The only sanctioned way to change it:
+`patch/vllm-moet-v0.24.0.patch` and
+`patch/vllm-moet-v0.25.0.patch` are **generated artifacts**. Each is the
+canonicalized `git diff` from its official release tag to the matching
+production-fork ship branch. They are **never edited by hand**, never patched
+incrementally, and never regenerated from anything but that branch. The only
+sanctioned commands are:
 
 ```bash
-python3 tools/check_patch_files.py --update
+python3 tools/check_patch_files.py --version 0.24.0 --update
+python3 tools/check_patch_files.py --version 0.25.0 --update
 ```
 
 If your change touches vLLM code, it goes to the **fork branch first**; the
@@ -25,21 +28,26 @@ WILL be erased by somebody's next regeneration.
 | repo | path | branch | role |
 |---|---|---|---|
 | **vLLM-Moet** (this one, public) | `/workspace/vllm-moet` | `main` | publication: generated patch, kernels + cubins, bench system, docs |
-| **vllm fork clone** | `/workspace/vllm-v0.24.0` | `moet-v0.24.0` | **source of truth for ALL vLLM code**; remotes: `fork` = `kacper-daftcode/vllm`, `origin` = `vllm-project/vllm` |
+| **vllm v0.24 fork clone** | `/workspace/vllm-v0.24.0` | `moet-v0.24.0` | legacy production source for the v0.24 overlay; remotes remain as documented in that clone |
+| **vllm v0.25 fork clone** | `/workspace/vllm-v0.25.0` | `moet-v0.25.0` | production source for the v0.25 overlay; `origin` = `vllm-project/vllm`, `fork` = `OmarB97/vllm` |
 
 `/root/workspace` is a symlink to `/workspace`. Upstream-PR branches and
 experiments live in worktrees off the same clone
 (`git -C /workspace/vllm-v0.24.0 worktree list`).
 
-`moet-v0.24.0` is the ship lineage: everything committed there is meant to
-ship in the next patch regen. Park half-done work on a side branch or
-worktree, not on `moet-v0.24.0`.
+For v0.25, the production fork branch and its recorded SHA gate rollout; an
+optional PR to `kacper-daftcode/vllm` or `vllm-project/vllm` never does. This
+settlement does not migrate or redefine the legacy v0.24 remote contract.
+
+`moet-v0.24.0` and `moet-v0.25.0` are ship lineages: everything committed
+there is meant to ship in that release overlay. Park half-done work on a side
+branch or worktree.
 
 ## Where a change goes
 
 | change | commit where | must also update |
 |---|---|---|
-| vLLM runtime code (`moe_w2_*`, loaders, runner, attention, …) | fork branch `moet-v0.24.0` | regen the patch here — procedure below |
+| vLLM runtime code (`moe_w2_*`, loaders, runner, attention, …) | matching fork branch `moet-v0.24.0` or `moet-v0.25.0` | regen that release patch here — procedure below |
 | SASS kernels / cubins | `kernels/` | a `kernels/MANIFEST.md` row (generator + validation status) |
 | serve configs | `bench/recipes/` | `bench/models.yaml`, `bench/matrix.yaml`; run `bench/runner/lint.py` |
 | bench results | `bench/results/<release>/` | `bench/runner/render.py` — the README table and per-release report are **generated**; never hand-edit the marked README block |
@@ -51,21 +59,24 @@ smoke results (`bench/results/smoke/`).
 
 ## Shipping a vLLM code change — the procedure
 
-1. **Commit on the fork branch** (`/workspace/vllm-v0.24.0`,
-   `moet-v0.24.0`). If the remote may have moved, fetch and merge first —
-   the regen tool refuses to run when the local branch is missing pushed
-   commits.
+1. **Commit and push on the matching production fork branch**
+   (`moet-v0.24.0` or `moet-v0.25.0`). If the remote may have moved, fetch
+   and merge first. Strict releases refuse to regenerate from unpublished
+   source.
 2. **Regenerate** from this repo:
 
    ```bash
-   python3 tools/check_patch_files.py --update
+   VLLM_MOET_FORK=/workspace/vllm-v0.24.0 \
+     python3 tools/check_patch_files.py --version 0.24.0 --update
+   VLLM_MOET_FORK=/workspace/vllm-v0.25.0 \
+     python3 tools/check_patch_files.py --version 0.25.0 --update
    ```
 
-   This rewrites the patch from the branch tip and updates the two
-   committed fingerprints: `patch/FILES.txt` (file list) and
-   `patch/SOURCE.txt` (the fork SHA the patch was generated from). It
-   refuses to move `SOURCE.txt` backwards along the branch, so a
-   regeneration can never roll back work that already shipped.
+   This rewrites the patch from the published branch tip and updates that
+   release's committed file list and source fingerprint
+   (`FILES.txt`+`SOURCE.txt` or
+   `FILES-v025.txt`+`SOURCE-v025.txt`). It refuses to move a source
+   fingerprint backwards, so regeneration cannot roll back shipped work.
 3. **Review `git diff patch/`.** An entry *vanishing* from `FILES.txt`
    means the patch carried work that never reached the fork branch —
    someone skipped step 1. **Stop and merge that work into the branch**;
@@ -81,9 +92,10 @@ smoke results (`bench/results/smoke/`).
 
    > `Three-tier starvation fix ships: step-scoped seen windows (vllm 9736e4d34)`
 
-6. **Push both together** (`fork moet-v0.24.0` + `origin main`) once the
-   pre-push checklist passes — or leave both unpushed. Avoid a lasting
-   state where only one side is pushed.
+6. **Publish source first, then distribution.** The matching production fork
+   branch must contain the recorded source SHA before the distribution guard
+   can pass. Push the distribution branch only after every guard is green.
+   Optional contribution PRs are follow-up evidence, not rollout gates.
 
 ## Concurrency — several agents, one checkout
 
@@ -92,10 +104,11 @@ smoke results (`bench/results/smoke/`).
   `git commit -a`, no `git stash`, no `git checkout --` / `git reset` over
   someone else's files, ever.
 - Stage **explicit paths only**: `git add <file> <file> …`.
-- `main` and `moet-v0.24.0` are shared trunks: no amending commits you did
-  not just create, no rebase, no force-push, no history rewrite.
-- The `patch/` trio (patch, `FILES.txt`, `SOURCE.txt`) changes **only** via
-  `--update`. If your commit would touch any of them for another reason,
+- `main`, `moet-v0.24.0`, and `moet-v0.25.0` are shared trunks: no
+  amending commits you did not just create, no rebase, no force-push, no
+  history rewrite.
+- Each release's patch, FILES, and SOURCE artifacts change **only** via the
+  matching versioned `--update`. If your commit would touch them another way,
   you are doing something wrong.
 - A pre-commit hook in this checkout runs the patch guard whenever `patch/`
   is staged. Do not bypass it with `--no-verify`.
@@ -113,6 +126,8 @@ smoke results (`bench/results/smoke/`).
 
 ```bash
 python3 tools/check_patch_files.py       # patch <-> FILES.txt <-> SOURCE.txt
+python3 tools/check_patch_files.py --version 0.25.0
+python3 -m unittest discover -s tests -p "test_check_patch_files.py"
 python3 bench/runner/lint.py             # recipes/boxes/suites/results schemas
 python3 bench/runner/render.py --check   # README table == committed results
 ```

Dockerfile.sm120-v025

@@ -0,0 +1,63 @@
+# vLLM-Moet on official vLLM v0.25.0 — SM120 (RTX PRO 6000 / RTX 5090)
+#
+# This is a side-by-side upgrade candidate. Dockerfile.sm120-v024 remains the
+# proven rollback until the v0.25 image passes the baked SM120 and live canary
+# gates documented in docs/v025-port.md.
+#
+# Build (linux/amd64 NVIDIA host, from the repo root):
+#   DOCKER_BUILDKIT=1 docker build -f Dockerfile.sm120-v025 \
+#     -t vllm-moet-sm120:v025-w2candidate .
+
+ARG VLLM_BASE=vllm/vllm-openai:v0.25.0@sha256:e1c1ff1af9a15921bfa11d1d95047258c1797392cdbfa296e7639da446b23f97
+FROM ${VLLM_BASE}
+
+LABEL org.opencontainers.image.version="v0.25.0-w2candidate" \
+      ai.kostudios.vllm-moet.base="vllm/vllm-openai:v0.25.0" \
+      ai.kostudios.vllm-moet.patch-sha256="ea1e8462008e8d3530e8938483a4f8974258196acc6a0bbcc4124bc4a719ed5d"
+
+# v0.25.0 already vendors the same SM120-capable DeepGEMM commit used by the
+# v0.24 recipe (a6b593d2826719dcf4892609af7b84ee23aaf32a), so no replacement
+# wheel is built here. git is needed only to apply the source overlay.
+RUN apt-get update && apt-get install -y --no-install-recommends git \
+    && rm -rf /var/lib/apt/lists/*
+
+# The W2 NVFP4 sparse-MLA patch was validated against FlashInfer 0.6.14's JIT
+# source/API layout. vLLM v0.25.0 pins 0.6.13, so preserve the proven 0.6.14
+# pair until the hardware canary explicitly qualifies the upstream pin.
+RUN pip uninstall -y --break-system-packages flashinfer-cubin || true
+RUN pip install --no-cache-dir --break-system-packages flashinfer-python==0.6.14 \
+    && pip install --no-cache-dir --break-system-packages \
+        --index-url https://flashinfer.ai/whl/cu130 \
+        "flashinfer-jit-cache==0.6.14+cu130"
+
+# Combined v0.25.0 overlay: W2 streaming/recovery and stores, DSpark confidence
+# scheduling, NVFP4 KV, and the SM120 fixes not absorbed by the release.
+COPY patch/vllm-moet-v0.25.0.patch /tmp/vllm-moet.patch
+RUN SP="$(python3 -c 'import vllm, os; print(os.path.dirname(os.path.dirname(vllm.__file__)))')" \
+    && cd "$SP" \
+    && git apply --check /tmp/vllm-moet.patch \
+    && git apply --verbose /tmp/vllm-moet.patch \
+    && python3 -m py_compile \
+       vllm/v1/worker/gpu/spec_decode/dspark/scheduler.py \
+       vllm/model_executor/layers/quantization/utils/moe_w2_cubit.py \
+       vllm/model_executor/layers/quantization/utils/moe_w2_store.py \
+    && python3 -c "from vllm.model_executor.layers.quantization.utils import moe_w2_cubit; print('moe_w2 hook OK')" \
+    && python3 -c "import vllm.v1.worker.gpu.spec_decode.dspark.speculator; print('dspark OK')" \
+    && rm /tmp/vllm-moet.patch
+
+# Prebuilt SM120 W2/W4 GEMM cubins, including K=6144 for GLM-5.x.
+COPY kernels/cubins-sm120/ /cubit-share/
+ENV VLLM_MOE_W2_CUBIT_DIR=/cubit-share
+
+# Bake the NVFP4 sparse-MLA read and packed-write kernels so the first serve
+# does not pay JIT compilation. See Dockerfile.sm120-v024 for layout details.
+RUN SP="$(python3 -c 'import vllm, os; print(os.path.dirname(os.path.dirname(vllm.__file__)))')" \
+    && cd "$SP" \
+    && python3 tools/nvfp4_flashinfer_sm120/patch_flashinfer.py \
+    && rm -f "$SP/flashinfer_jit_cache/jit_cache/sparse_mla_sm120/sparse_mla_sm120.so" \
+    && FLASHINFER_CUDA_ARCH_LIST=12.0f python3 -c \
+       "from flashinfer.jit.mla import gen_sparse_mla_sm120_module; gen_sparse_mla_sm120_module().build(verbose=False)" \
+    && mkdir -p /opt/nvfp4-ds-mla \
+    && TORCH_CUDA_ARCH_LIST=12.0a python3 -c \
+       "from torch.utils.cpp_extension import load; load(name='nvfp4_ds_mla_cache_ext', sources=['$SP/csrc/nvfp4_ds_mla/concat_and_cache_nvfp4_ds_mla.cu'], extra_cuda_cflags=['-O3', '--generate-code=arch=compute_120a,code=sm_120a'], build_directory='/opt/nvfp4-ds-mla')"
+ENV VLLM_NVFP4_DS_MLA_EXT_DIR=/opt/nvfp4-ds-mla

README.md

@@ -18,6 +18,12 @@ cannot even fit on. Three ideas carry it:
    bit‑deterministic), an **NVFP4 KV cache** (352 B/token), and agent‑ready tool/reasoning
    parsing.
 
+> **v0.25.0 upgrade candidate:** [`Dockerfile.sm120-v025`](Dockerfile.sm120-v025)
+> and [`docs/v025-port.md`](docs/v025-port.md) carry the rebased 61-file overlay
+> on official vLLM v0.25.0. The results below remain v0.24 measurements until
+> the candidate passes its own SM120 model-load, 128K, quality, and performance
+> gates; the v0.24 image stays the rollback.
+
 ---
 
 ## GLM‑5.2 (753B) — the headline model
@@ -250,6 +256,10 @@ git clone https://github.com/kacper-daftcode/vLLM-Moet && cd vLLM-Moet
 
 # official vllm-openai:v0.24.0 image + patch + pins + SM120 cubins
 DOCKER_BUILDKIT=1 docker build -f Dockerfile.sm120-v024 -t vllm-moet-sm120:v024 .
+
+# side-by-side v0.25.0 candidate (do not replace the v0.24 live image yet)
+DOCKER_BUILDKIT=1 docker build -f Dockerfile.sm120-v025 \
+  -t vllm-moet-sm120:v025-w2candidate .
 ```
 
 **Easiest path — run a benchmarked recipe.** The recipes image downloads the
@@ -424,6 +434,13 @@ Release **`baseline-2026-07-10`** — one row per supported recipe (`bench/recip
 <!-- bench:table:end -->
 
 ## Repository layout
+- **`patch/vllm-moet-v0.25.0.patch`** — the v0.25 candidate delta (61 files,
+  +13,342/-134 source lines) against exact official tag commit `702f4814`.
+- **`Dockerfile.sm120-v025`** — pinned official v0.25.0 image plus the candidate
+  overlay; the exact image is built side-by-side with v0.24 and still requires
+  its documented disposable serve gate before production promotion.
+- **`docs/v025-port.md`** — exact identities, absorbed-upstream inventory,
+  compatibility decisions, completed source gates, and remaining promotion gates.
 - **`patch/vllm-moet-v0.24.0.patch`** — the delta vs official vLLM `v0.24.0` (37 files,
   +7.4k lines; applies clean on the tag). Goes with the pins above.
 - **`Dockerfile.sm120-v024`** — the image: official `vllm/vllm-openai:v0.24.0` + patch + pins +