Add vLLM v0.25.0 W2 upgrade candidate

.github/workflows/bench-lint.yml

@@ -11,7 +11,9 @@ on:
       - "bench/**"
       - "docker/**"
       - "Dockerfile.recipes"
+      - "Dockerfile.sm120-v025"
       - "README.md"
+      - "docs/v025-port.md"
       - "docs/benchmarks/**"
       - "patch/**"
       - "tools/check_patch_files.py"
@@ -26,8 +28,10 @@ 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: Patch file lists match both frozen release overlays
+        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

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 ${VLLM_BASE}
+
+LABEL org.opencontainers.image.version="v0.25.0-w2candidate" \
+      ai.kostudios.vllm-moet.base="vllm/vllm-openai:v0.25.0" \
+      ai.kostudios.vllm-moet.patch-sha256="25ac6fea69d71c1a641b0d6343c01011bca3b481436e29f9f02f8e4c3ce639a4"
+
+# v0.25.0 already vendors the same SM120-capable DeepGEMM commit used by the
+# v0.24 recipe (a6b593d2826719dcf4892609af7b84ee23aaf32a), so no replacement
+# wheel is built here. git is needed only to apply the source overlay.
+RUN apt-get update && apt-get install -y --no-install-recommends git \
+    && rm -rf /var/lib/apt/lists/*
+
+# The W2 NVFP4 sparse-MLA patch was validated against FlashInfer 0.6.14's JIT
+# source/API layout. vLLM v0.25.0 pins 0.6.13, so preserve the proven 0.6.14
+# pair until the hardware canary explicitly qualifies the upstream pin.
+RUN pip uninstall -y --break-system-packages flashinfer-cubin || true
+RUN pip install --no-cache-dir --break-system-packages flashinfer-python==0.6.14 \
+    && pip install --no-cache-dir --break-system-packages \
+        --index-url https://flashinfer.ai/whl/cu130 \
+        "flashinfer-jit-cache==0.6.14+cu130"
+
+# Combined v0.25.0 overlay: W2 streaming/recovery and stores, DSpark confidence
+# scheduling, NVFP4 KV, and the SM120 fixes not absorbed by the release.
+COPY patch/vllm-moet-v0.25.0.patch /tmp/vllm-moet.patch
+RUN SP="$(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 60-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 +434,12 @@ 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 (60 files,
+  +13,042/-133 source lines) against exact official tag commit `702f4814`.
+- **`Dockerfile.sm120-v025`** — pinned official v0.25.0 image plus the candidate
+  overlay; built and qualified side-by-side with v0.24.
+- **`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 +

docs/v025-port.md

@@ -0,0 +1,145 @@
+# The v0.25.0 upgrade candidate
+
+This repository now carries a side-by-side W2 overlay for official vLLM
+`v0.25.0`. It is an upgrade candidate, not yet the production default. The
+proven v0.24 image, patch, recipes, and benchmark receipts remain intact as the
+rollback boundary until the v0.25 candidate passes the SM120 hardware canary.
+
+## Exact source identity
+
+- Official tag: `v0.25.0`
+- Official tag commit: `702f4814fe54fabff350d43cb753ae3e47c0c276`
+- Linux/amd64 base image manifest: `sha256:e1c1ff1af9a15921bfa11d1d95047258c1797392cdbfa296e7639da446b23f97`
+- W2 overlay: `patch/vllm-moet-v0.25.0.patch`
+- Overlay SHA-256: `25ac6fea69d71c1a641b0d6343c01011bca3b481436e29f9f02f8e4c3ce639a4`
+- Overlay scope: 60 files, 13,042 insertions, 133 deletions
+
+Apply it directly to an official checkout with:
+
+```bash
+git clone --branch v0.25.0 https://github.com/vllm-project/vllm && cd vllm
+git apply --check /path/to/vLLM-Moet/patch/vllm-moet-v0.25.0.patch
+git apply /path/to/vLLM-Moet/patch/vllm-moet-v0.25.0.patch
+```
+
+Or build the pinned serving image:
+
+```bash
+DOCKER_BUILDKIT=1 docker build -f Dockerfile.sm120-v025 \
+  -t vllm-moet-sm120:v025-w2candidate .
+```
+
+## What v0.25 absorbed
+
+The port was produced by applying the frozen v0.24 overlay to `v0.24.0`, then
+rebasing that exact tree onto `v0.25.0` and resolving conflicts against the new
+Model Runner V2 paths. Ten old overlay files disappeared because v0.25 now owns
+their behavior, including the core DSpark/DFlash model registrations, DeepSeek
+V4 DSpark implementation, Gumbel sampling, and SM120 cooperative-top-k guard.
+The v0.25 overlay therefore drops those redundant hunks instead of shadowing
+upstream.
+
+Exact paths removed from the overlay:
+
+```text
+vllm/model_executor/models/qwen3_dflash.py
+vllm/model_executor/models/registry.py
+vllm/models/deepseek_v4/__init__.py
+vllm/models/deepseek_v4/nvidia/dspark.py
+vllm/models/deepseek_v4/nvidia/model.py
+vllm/transformers_utils/configs/speculators/algos.py
+vllm/v1/worker/gpu/sample/gumbel.py
+vllm/v1/worker/gpu/spec_decode/__init__.py
+vllm/v1/worker/gpu/spec_decode/dspark/__init__.py
+vllm/v1/worker/gpu/spec_decode/eagle/eagle3_utils.py
+```
+
+The retained delta is the project-specific W2 stack: 2-bit planes, FP4
+recovery and confidence gate, tiered/NVMe expert stores, persistent pack-cache
+safety, SM120 cubins, NVFP4 KV, pipeline-aware replay, and the optional
+hardware-aware DSpark confidence scheduler.
+
+## Compatibility decisions
+
+- **Model Runner V2 stays enabled.** The port preserves v0.25's new default and
+  composes W2 padded-slot, prefill, replay, Mamba-preprocess, and graph metadata
+  with it. It does not restore a V1/PagedAttention escape hatch.
+- **DeepGEMM uses the release copy.** v0.25 already vendors exact commit
+  `a6b593d2826719dcf4892609af7b84ee23aaf32a`, the same SM120-capable commit the
+  v0.24 recipe built separately. The v0.25 Dockerfile removes that duplicate
+  wheel build.
+- **FlashInfer remains 0.6.14 temporarily.** Official v0.25 pins 0.6.13, while
+  the W2 NVFP4 sparse-MLA source patch and JIT kwargs were hardware-validated on
+  0.6.14. The candidate preserves the proven pair and makes that deviation
+  explicit. Qualifying 0.6.13 is a separate canary, not an assumption.
+- **SM120 raw FP8 scales remain.** The v0.25 release still uses the SM100 packed
+  scale recipe in the DeepSeek V4 output projection. Consumer Blackwell needs
+  the raw row-major scale layout carried by this overlay.
+- **The DSpark extensions remain optional.** v0.25 supplies the core DSpark
+  engine; the overlay adds per-request confidence widths, profiled cost tables,
+  online calibration, hysteresis, and live dynamic-SD re-derivation.
+- **Tier managers stop before interpreter teardown.** The v0.25 stable-libtorch
+  extension can abort if a daemon manager still owns Torch tensors during
+  Python shutdown. Each tier now has an explicit stop/join boundary and the
+  module registers a deduplicated `atexit` shutdown for serving workers.
+
+## Verification completed before image build
+
+The source port passed:
+
+- `git diff --check` against the exact v0.25 tag;
+- Python compilation across every changed Python file;
+- 22 passed / 1 skipped focused W2 memory, padded-route, step-pin, and manager
+  shutdown tests;
+- 6 passed CPU DSpark scheduling and live-re-derivation regressions;
+- clean patch application and a committed 60-file lost-line manifest.
+
+These are source gates only. They do **not** establish CUDA kernel, model-load,
+quality, context, or throughput parity.
+
+## Bounded SM120 image receipt (2026-07-12)
+
+The digest-pinned recipe built on taro as
+`vllm-moet-sm120:v025-w2candidate-25ac6fea`, local image ID
+`sha256:1b3dc4a340a6`. On its RTX 5090 (SM120), the exact image passed:
+
+- stable-libtorch native extension import with zero allocated GPU bytes;
+- the baked 22 passed / 1 skipped W2 suite and 6 passed DSpark suite;
+- bounded W2/W4 decode (`max_rel` 0.01358 / cosine 0.999911), full-FP4 delta
+  (`0.01611` / `0.999906`), and split-FP4 delta (`0.01333` / `0.999922`);
+- split three-tier mixed dispatch, base-miss zeroing, coupled eviction, and
+  clean interpreter shutdown;
+- byte-identical pinned, pack, reboot, tiered arena, eviction, overflow,
+  scan-resistance, and preheat store paths;
+- baked NVFP4 packed-cache writes and FlashInfer sparse-MLA JIT-cache load.
+
+The first v0.25 candidate exposed the manager teardown abort after its
+three-tier assertions passed. The same test exited clean on frozen v0.24; the
+explicit stop/join fix then exited clean on corrected v0.25 and is covered by
+two CPU regressions. The superseded image tag was removed. Throughout these
+bounded checks, taro's live llama-swap Qwen process stayed at 23,114 MiB and
+was not restarted or rerouted.
+
+This receipt still does **not** establish a DS4 checkpoint load, 128K context,
+quality, or performance result on v0.25.
+
+## Promotion gates
+
+The v0.25 candidate must stay side-by-side with the live v0.24 image. Promotion
+requires, in order:
+
+1. build the pinned image on an SM120 host and run the baked import/compile
+   checks;
+2. run CUDA op tests for W2/W4 cubins, raw-scale output projection, NVFP4 cache
+   write/read, and CUDA-graph capture;
+3. cold-start a disposable DS4 canary without replacing the live router lane;
+4. prove the 128K serve configuration, exact retrieval, frozen-rule quality,
+   memory/cgroup safety, and no pack corruption on the real endpoint;
+5. compare decode, prefill, MTP acceptance, replay rate, and memory receipts to
+   the frozen v0.24 baseline;
+6. only then move the router lane, retaining the v0.24 image and packs for an
+   immediate rollback.
+
+No existing seed or benchmark receipt is relabeled as v0.25 evidence. The
+upgrade reuses the test definitions, but the candidate must earn its own
+runtime receipts because the execution engine and dependency base changed.

patch/FILES-v025.txt

@@ -0,0 +1,64 @@
+# Files touched by patch/vllm-moet-v0.25.0.patch (sorted; generated by
+# tools/check_patch_files.py --version 0.25.0 --update, verified by CI
+# bench-lint). A file DISAPPEARING from this list means work was dropped by
+# regeneration; restore it in the source branch instead of hiding the loss.
+csrc/nvfp4_ds_mla/concat_and_cache_nvfp4_ds_mla.cu
+tests/model_executor/layers/quantization/test_moe_w2_cgroup_memory.py
+tests/model_executor/layers/quantization/test_moe_w2_padded_routes.py
+tests/model_executor/layers/quantization/test_moe_w2_step_pins.py
+tests/v1/spec_decode/test_dspark_scheduler.py
+tools/nvfp4_flashinfer_sm120/README.md
+tools/nvfp4_flashinfer_sm120/nvfp4_expand.cuh
+tools/nvfp4_flashinfer_sm120/patch_flashinfer.py
+vllm/compilation/breakable_cudagraph.py
+vllm/compilation/cuda_graph.py
+vllm/config/speculative.py
+vllm/config/vllm.py
+vllm/envs.py
+vllm/forward_context.py
+vllm/model_executor/layers/attention/mla_attention.py
+vllm/model_executor/layers/quantization/fp8.py
+vllm/model_executor/layers/quantization/modelopt.py
+vllm/model_executor/layers/quantization/mxfp4.py
+vllm/model_executor/layers/quantization/utils/fp8_utils.py
+vllm/model_executor/layers/quantization/utils/moe_w2_cubit.py
+vllm/model_executor/layers/quantization/utils/moe_w2_delta.py
+vllm/model_executor/layers/quantization/utils/moe_w2_gate.py
+vllm/model_executor/layers/quantization/utils/moe_w2_looka.py
+vllm/model_executor/layers/quantization/utils/moe_w2_planes.py
+vllm/model_executor/layers/quantization/utils/moe_w2_planes_cache.py
+vllm/model_executor/layers/quantization/utils/moe_w2_store.py
+vllm/model_executor/layers/quantization/utils/prefill_timers.py
+vllm/model_executor/layers/quantization/utils/skinny_fp8_cubit.py
+vllm/model_executor/model_loader/__init__.py
+vllm/model_executor/model_loader/default_loader.py
+vllm/model_executor/model_loader/weight_utils.py
+vllm/model_executor/models/deepseek_mtp.py
+vllm/model_executor/models/qwen3_dspark.py
+vllm/models/deepseek_v4/nvidia/mtp.py
+vllm/models/deepseek_v4/nvidia/ops/o_proj.py
+vllm/v1/attention/backends/flashinfer.py
+vllm/v1/attention/backends/mla/cubit_sparse_mla.py
+vllm/v1/attention/backends/mla/flashinfer_mla_sparse.py
+vllm/v1/attention/backends/mla/flashinfer_mla_sparse_sm120.py
+vllm/v1/attention/backends/mla/nvfp4_ds_mla_cache.py
+vllm/v1/attention/backends/mla/sparse_mla_env.py
+vllm/v1/attention/backends/mla/sparse_swa.py
+vllm/v1/attention/ops/merge_attn_states.py
+vllm/v1/attention/ops/triton_decode_attention.py
+vllm/v1/core/sched/scheduler.py
+vllm/v1/engine/core.py
+vllm/v1/kv_cache_interface.py
+vllm/v1/spec_decode/dynamic/utils.py
+vllm/v1/worker/gpu/cudagraph_utils.py
+vllm/v1/worker/gpu/input_batch.py
+vllm/v1/worker/gpu/model_runner.py
+vllm/v1/worker/gpu/spec_decode/dflash/speculator.py
+vllm/v1/worker/gpu/spec_decode/dspark/scheduler.py
+vllm/v1/worker/gpu/spec_decode/dspark/speculator.py
+vllm/v1/worker/gpu/spec_decode/dspark/utils.py
+vllm/v1/worker/gpu/spec_decode/rejection_sampler.py
+vllm/v1/worker/gpu/spec_decode/utils.py
+vllm/v1/worker/gpu_model_runner.py
+vllm/v1/worker/gpu_ubatch_wrapper.py
+vllm/v1/worker/gpu_worker.py