Prepare Web qualification for llama.cpp v0.9.0

.github/workflows/ci.yml

@@ -60,11 +60,11 @@ jobs:
   build-webgpu-bridge:
     needs: [changes, checks]
     if: needs.changes.outputs.native == 'true'
-    name: ${{ matrix.upstream == 'pinned' && 'Build WebGPU Bridge (WASM)' || 'Build WebGPU Bridge (WASM v0.7.0)' }}
+    name: ${{ matrix.upstream == 'pinned' && 'Build WebGPU Bridge (WASM)' || 'Build WebGPU Bridge (WASM v0.9.0)' }}
     strategy:
       fail-fast: false
       matrix:
-        upstream: [pinned, v0.7.0]
+        upstream: [pinned, v0.9.0]
     runs-on: ubuntu-latest
     env:
       LLAMA_WEBGPU_SMOKE_MODEL_URL: https://huggingface.co/aladar/llama-2-tiny-random-GGUF/resolve/main/llama-2-tiny-random.gguf
@@ -111,8 +111,8 @@ jobs:
             echo "error: llama_cpp.version is empty"
             exit 1
           fi
-          if [ "${{ matrix.upstream }}" = "v0.7.0" ]; then
-            LLAMA_CPP_TAG=v0.7.0
+          if [ "${{ matrix.upstream }}" = "v0.9.0" ]; then
+            LLAMA_CPP_TAG=v0.9.0
           fi
           echo "LLAMA_CPP_TAG=$LLAMA_CPP_TAG" >> "$GITHUB_ENV"
           echo "Using llama.cpp ${LLAMA_CPP_TAG}"
@@ -138,8 +138,8 @@ jobs:
       - name: Clone llama.cpp source
         run: |
           git clone --depth 1 --branch "$LLAMA_CPP_TAG" https://github.com/ggml-org/llama.cpp.git third_party/llama_cpp
-          if [ "${{ matrix.upstream }}" = "v0.7.0" ]; then
-            test "$(git -C third_party/llama_cpp rev-parse HEAD)" = 620d2225cc436ab0b001e3200e590e3b286ef9ab
+          if [ "${{ matrix.upstream }}" = "v0.9.0" ]; then
+            test "$(git -C third_party/llama_cpp rev-parse HEAD)" = cd959bae99ff4bed22bef2604f45241c09bbb8c8
           fi
 
       # Installed at a fixed path, not with setup-emsdk: its random temp
@@ -264,46 +264,46 @@ jobs:
         if: failure()
         uses: actions/upload-artifact@v4
         with:
-          name: ${{ matrix.upstream == 'pinned' && 'state-persistence-smoke-artifacts' || 'state-persistence-smoke-artifacts-v0.7.0' }}
+          name: ${{ matrix.upstream == 'pinned' && 'state-persistence-smoke-artifacts' || 'state-persistence-smoke-artifacts-v0.9.0' }}
           path: ${{ env.LLAMA_WEBGPU_SMOKE_ARTIFACTS_DIR }}
           if-no-files-found: ignore
 
       - name: Upload multimodal smoke diagnostics
         if: failure()
         uses: actions/upload-artifact@v4
         with:
-          name: ${{ matrix.upstream == 'pinned' && 'multimodal-smoke-artifacts' || 'multimodal-smoke-artifacts-v0.7.0' }}
+          name: ${{ matrix.upstream == 'pinned' && 'multimodal-smoke-artifacts' || 'multimodal-smoke-artifacts-v0.9.0' }}
           path: ${{ env.LLAMA_WEBGPU_MULTIMODAL_ARTIFACTS_DIR }}
           if-no-files-found: ignore
 
       - name: Upload grammar smoke diagnostics
         if: failure()
         uses: actions/upload-artifact@v4
         with:
-          name: ${{ matrix.upstream == 'pinned' && 'grammar-smoke-artifacts' || 'grammar-smoke-artifacts-v0.7.0' }}
+          name: ${{ matrix.upstream == 'pinned' && 'grammar-smoke-artifacts' || 'grammar-smoke-artifacts-v0.9.0' }}
           path: ${{ env.LLAMA_WEBGPU_GRAMMAR_ARTIFACTS_DIR }}
           if-no-files-found: ignore
 
       - name: Upload next-token scores smoke diagnostics
         if: failure()
         uses: actions/upload-artifact@v4
         with:
-          name: ${{ matrix.upstream == 'pinned' && 'next-token-scores-smoke-artifacts' || 'next-token-scores-smoke-artifacts-v0.7.0' }}
+          name: ${{ matrix.upstream == 'pinned' && 'next-token-scores-smoke-artifacts' || 'next-token-scores-smoke-artifacts-v0.9.0' }}
           path: ${{ env.LLAMA_WEBGPU_NEXT_TOKEN_SCORES_ARTIFACTS_DIR }}
           if-no-files-found: ignore
 
       - name: Upload speculative decoding smoke diagnostics
         if: failure()
         uses: actions/upload-artifact@v4
         with:
-          name: ${{ matrix.upstream == 'pinned' && 'speculative-smoke-artifacts' || 'speculative-smoke-artifacts-v0.7.0' }}
+          name: ${{ matrix.upstream == 'pinned' && 'speculative-smoke-artifacts' || 'speculative-smoke-artifacts-v0.9.0' }}
           path: ${{ env.LLAMA_WEBGPU_SPECULATIVE_ARTIFACTS_DIR }}
           if-no-files-found: ignore
 
       - name: Upload bridge artifacts
         uses: actions/upload-artifact@v4
         with:
-          name: ${{ matrix.upstream == 'pinned' && 'webgpu-bridge-dist' || 'webgpu-bridge-dist-v0.7.0' }}
+          name: ${{ matrix.upstream == 'pinned' && 'webgpu-bridge-dist' || 'webgpu-bridge-dist-v0.9.0' }}
           path: |
             ${{ runner.temp }}/webgpu_bridge_dist/llama_webgpu_bridge.js
             ${{ runner.temp }}/webgpu_bridge_dist/llama_webgpu_bridge_worker.js

AGENTS.md

@@ -82,11 +82,11 @@ GGUFs or smoke artifacts.
 
 - CI build gate: `.github/workflows/ci.yml`
   - Builds wasm32/memory64 against the `llama_cpp.version` pin and exact
-    v0.7.0 candidate in separate lanes. Its exact SHA lives in the workflow.
+    v0.9.0 candidate in separate lanes. Its exact SHA lives in the workflow.
     Preserve its check/artifact identities (`Build WebGPU Bridge (WASM)`,
     `webgpu-bridge-dist`). Neither lane publishes or changes the source pin; candidate check
-    and artifact identities use the `v0.7.0` suffix. See
-    `docs/v070_qualification.md` for the preparation evidence and limits.
+    and artifact identities use the `v0.9.0` suffix. See
+    `docs/v090_qualification.md` for the preparation evidence and limits.
     Builds against llama.cpp v0.4.0 are no longer tested in CI; the
     `src/llama_webgpu_mtmd_compat.h` shim and its static contract remain.
   - Verifies the active `emcc` identity against `emsdk.version` and

README.md

@@ -215,16 +215,16 @@ This repo includes a wasm build gate in:
 - `.github/workflows/ci.yml`
 
 After the JS/compatibility contracts pass, it builds wasm32 and memory64 against
-the pinned `llama.cpp` tag in `llama_cpp.version` and the exact v0.7.0
+the pinned `llama.cpp` tag in `llama_cpp.version` and the exact v0.9.0
 candidate in separate lanes. Both run real
 state-persistence, multimodal, grammar and next-token-score browser smokes. A
 successful run uploads its seven built files as `webgpu-bridge-dist`; a failed
 run uploads no dist and instead the smoke diagnostics that exist
 (`state-persistence-smoke-artifacts`, `multimodal-smoke-artifacts`,
-`grammar-smoke-artifacts`, `next-token-scores-smoke-artifacts`). Candidate checks and artifacts use the `v0.7.0` suffix;
+`grammar-smoke-artifacts`, `next-token-scores-smoke-artifacts`). Candidate checks and artifacts use the `v0.9.0` suffix;
 the pinned identities remain unchanged. CI never changes a pin or publishes
 assets. Builds against llama.cpp v0.4.0 are no longer tested in CI. See
-[the v0.7.0 preparation evidence](docs/v070_qualification.md) for validation
+[the v0.9.0 preparation evidence](docs/v090_qualification.md) for validation
 and qualification limits.
 To run the media-helper and static CI contracts locally:
 

docs/v090_qualification.md

@@ -0,0 +1,46 @@
+# llama.cpp v0.9.0 compatibility preparation
+
+The candidate lane builds upstream `cd959bae99ff4bed22bef2604f45241c09bbb8c8`
+(v0.9.0) alongside the unchanged v0.5.0 development pin, matching published
+native `v0.9.0`. The bridge source needs no change for v0.9.0. No Web assets or
+Dart packages are published by this change.
+
+## Upstream and native patches
+
+- The v0.7.0 embedding-mode reserve repair
+  ([llamadart-native#114](https://github.com/leehack/llamadart-native/issues/114),
+  [v0.7.0 notes](v070_qualification.md)) is upstream since v0.8.0:
+  `set_embeddings` re-reserves on every toggle. `cmake/embedding_mode_reserve.cmake`
+  finds no affected guard and stays inactive.
+- The other native patches do not apply to the Web build: 0001/0002 are
+  Vulkan, 0004 Metal, and 0005/0006 (F16 dot and GEMM F32 accumulation,
+  [llamadart-native#174](https://github.com/leehack/llamadart-native/issues/174))
+  change code only under `__ARM_NEON`/`__ARM_FEATURE_FP16_VECTOR_ARITHMETIC`,
+  which WASM never defines.
+- The mtmd-audio single-thread rewrite still finds both `n_threads` markers.
+- v0.9.0 drops `mtmd_helper_gen_audio_inp::voice` and adds
+  `mtmd_gen_audio_info::n_voices` and Chatterbox voices. The bridge TTS path
+  sets neither field, so Qwen3-TTS, with or without a speaker reference, is
+  unaffected.
+
+## Validation
+
+Local evidence uses Emscripten 6.0.8, Node 24 and the locked Playwright
+Chromium on macOS arm64, against a clean upstream checkout at the commit above.
+Model/projector pins and commands live in
+[CONTRIBUTING.md](../CONTRIBUTING.md#validate-outputs).
+
+| Check | Result |
+| --- | --- |
+| `npm run check:js` | 602 passed, 4 existing skips, zero failures |
+| CI reliability contract | PASS |
+| v0.9.0 production build | PASS, wasm32 and memory64 |
+| Extraction/state/generation smoke | PASS, direct and worker on both memory modes |
+| Qwen3.5 image inference | PASS, direct and worker on wasm32 |
+| Tiny and Qwen3.5 grammar, next-token scores, tiny speculative | PASS |
+| Qwen3-ASR transcript, silence and cancellation | PASS, four runtime rows |
+| Qwen3-TTS synthesis, cancellation and unload, WebGPU offload | PASS, four runtime rows with a speaker reference; direct and worker on memory64 without one |
+
+Four runtime rows mean direct and worker execution on wasm32 and memory64.
+Hosted exact-head CI, candidate qualification, publication approval and Dart
+adoption remain separate gates.

tests/ci/ci_scope_test.mjs

@@ -156,8 +156,8 @@ const rootDir = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '../.
   const { jobs } = workflow;
   const build = jobs['build-webgpu-bridge'];
   const { checks } = jobs;
-  assert.equal(build.name, "${{ matrix.upstream == 'pinned' && 'Build WebGPU Bridge (WASM)' || 'Build WebGPU Bridge (WASM v0.7.0)' }}");
-  assert.deepEqual(build.strategy.matrix.upstream, ['pinned', 'v0.7.0']);
+  assert.equal(build.name, "${{ matrix.upstream == 'pinned' && 'Build WebGPU Bridge (WASM)' || 'Build WebGPU Bridge (WASM v0.9.0)' }}");
+  assert.deepEqual(build.strategy.matrix.upstream, ['pinned', 'v0.9.0']);
   assert.equal(build.strategy['fail-fast'], false);
   assert.deepEqual(build.needs, ['changes', 'checks']);
   assert.equal(build.if, "needs.changes.outputs.native == 'true'");
@@ -170,15 +170,16 @@ const rootDir = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '../.
     1,
   );
   const clone = build.steps.find((step) => step.name === 'Clone llama.cpp source').run;
-  assert.match(clone, /620d2225cc436ab0b001e3200e590e3b286ef9ab/);
+  assert.match(clone, /cd959bae99ff4bed22bef2604f45241c09bbb8c8/);
+  assert.match(clone, /"\$\{\{ matrix.upstream \}\}" = "v0.9.0"/);
   assert.match(clone, /git -C third_party\/llama_cpp rev-parse HEAD/);
   const resolver = build.steps.find((step) => step.name === 'Resolve llama.cpp pin').run;
   assert.match(resolver, /llama_cpp.version/);
   assert.match(resolver, /matrix.upstream/);
-  assert.match(resolver, /LLAMA_CPP_TAG=v0.7.0/);
+  assert.match(resolver, /LLAMA_CPP_TAG=v0.9.0/);
   for (const step of build.steps.filter((step) => step.uses?.startsWith('actions/upload-artifact@'))) {
     assert.match(step.with.name, /matrix.upstream == 'pinned'/);
-    assert.match(step.with.name, /-v0.7.0'/);
+    assert.match(step.with.name, /-v0.9.0'/);
   }
   const buildSteps = Object.fromEntries(build.steps.map((step) => [step.name, step]));
   for (const name of ['Run state persistence browser smoke', 'Run multimodal browser smoke', 'Build bridge artifacts', 'Verify outputs']) {