chore: ⬆️ Update ikawrakow/ik_llama.cpp to 1a2a8604a6c6c6413c06bf9adfc2f64329af4366 (#11932)

⬆️ Update ikawrakow/ik_llama.cpp

Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: mudler <2420543+mudler@users.noreply.github.com>

.agents/ai-coding-assistants.md

@@ -49,6 +49,40 @@ AI agents MUST NOT add `Co-Authored-By` trailers for themselves either.
 A human reviewer owns the contribution; the AI's involvement is recorded
 via `Assisted-by` (see below).
 
+### Exception: automation operated by a maintainer
+
+The rule above addresses the common case, an AI assistant helping a human
+contributor who then signs off. It does not fit automation that a
+maintainer runs themselves, which opens pull requests with no human
+submitter to sign. Applied literally there, nothing ever signs and the
+DCO check blocks the pull request permanently.
+
+A maintainer-operated bot MUST therefore add a `Signed-off-by` trailer
+naming **the maintainer who operates it**, not the bot and not the model:
+
+```
+Assisted-by: Codex:gpt-5
+Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
+```
+
+This is not the AI certifying the DCO. The maintainer is, exactly as they
+do for a commit they typed by hand: they configured the automation, they
+own its output, and they take responsibility for it when they merge it.
+The `Assisted-by` trailer still records that a model produced the code, so
+the provenance trail is unchanged.
+
+The exception is narrow and does not widen the rule for anyone else:
+
+- It applies only to automation a LocalAI maintainer operates and whose
+  output that maintainer reviews before merge.
+- The sign-off names a real person who accepts DCO responsibility.
+- An AI assistant helping an outside contributor still MUST NOT sign off.
+  That contributor adds their own trailer.
+- A bot MUST NOT sign off on behalf of anyone other than its operator, and
+  MUST NOT add a trailer for a contributor whose branch it pushes to. If
+  automation contributes to someone else's branch, it leaves the sign-off
+  to that contributor.
+
 ## Attribution
 
 When AI tools contribute to LocalAI development, proper attribution helps

.agents/api-endpoints-and-auth.md

@@ -236,6 +236,58 @@ Use these HTTP status codes:
 
 If your endpoint should be tracked for usage (token counts, request counts), add the `usageMiddleware` to its middleware chain. See `core/http/middleware/usage.go` and how it's applied in `routes/openai.go`.
 
+## Control-plane database health metrics
+
+In distributed mode the frontend registers three OpenTelemetry gauges over the
+PostgreSQL control-plane database (`core/services/monitoring/control_plane_db.go`,
+wired in `core/application/distributed.go`). They reach `/metrics` through the
+same Prometheus exporter as the rest of the API metrics.
+
+| Metric | Meaning | Page when |
+|--------|---------|-----------|
+| `localai_control_plane_oldest_xmin_age` | Transactions elapsed since the oldest snapshot any backend still holds | above a few million, and rising |
+| `localai_control_plane_longest_transaction_seconds` | Age of the longest open transaction | above 3600 |
+| `localai_control_plane_dead_tuple_ratio` | Dead tuples per live tuple, labelled by `table`, on `backend_nodes`, `node_models` and `gallery_operations` | sustained above ~10 on a small table |
+
+A sustained high `localai_control_plane_oldest_xmin_age` is the one to page on.
+While it grows, autovacuum can reclaim nothing anywhere in the database no
+matter how often it runs, so the dead tuple ratio keeps climbing and a six-row
+registry table can reach hundreds of megabytes. Tuning autovacuum does not help.
+The fix is to find the transaction holding the horizon open and clear it:
+
+```sql
+SELECT pid, state, age(backend_xmin) AS xmin_age, now() - xact_start AS xact_age, query
+FROM pg_stat_activity
+WHERE backend_xmin IS NOT NULL
+ORDER BY age(backend_xmin) DESC;
+```
+
+Then `pg_terminate_backend(pid)` on the offenders, and `VACUUM (VERBOSE)` the
+bloated tables once the horizon has moved.
+
+**A healthy-looking xmin age does not on its own prove the horizon is free.**
+The gauge reads `pg_stat_activity`, which only sees live backends. Two other
+things pin the very same horizon and are invisible there, so either one can hold
+vacuum back while the gauge reads 0:
+
+```sql
+SELECT gid, prepared, database, transaction FROM pg_prepared_xacts;
+SELECT slot_name, active, xmin, catalog_xmin FROM pg_replication_slots;
+```
+
+An orphaned prepared transaction is cleared with `ROLLBACK PREPARED '<gid>'`,
+and a stale slot with `pg_drop_replication_slot('<slot_name>')`. Check both
+before concluding that a bloated table has some other cause.
+
+Sampling is scrape-driven behind a 30 second cache, so scrape frequency does not
+translate into database load. Failed and timed-out samples cost the same interval
+as successful ones, so a database that is already struggling is not retried on
+every scrape. A failed sample reports the last good values rather than failing the
+scrape, because these gauges matter most when the database is struggling. Before
+the first successful sample the gauges are absent rather than zero, since a zero
+xmin age would read as a healthy horizon: alert on `absent()` too if you need to
+distinguish "healthy" from "never sampled".
+
 ## Advertising surfaces — where to register a new capability
 
 Beyond routing and auth, LocalAI publishes its capability surface in **four independent places**. When you add an endpoint — especially one introducing a net-new capability like a new media type or a new auth-gated feature — you must update every relevant surface. These aren't optional: missing them means the endpoint works but is invisible to clients, admins, and the UI.

.agents/ds4-backend.md

@@ -77,6 +77,56 @@ spectrum. **Metal (Darwin) only** - it is a no-op on CUDA/CPU. Enable with
 budget). Gallery entries built on this: `deepseek-v4-flash-q4-ssd` (153 GB Flash
 on a 128 GB Mac) and `deepseek-v4-pro-q2-ssd` (433 GB Pro, experimental).
 
+## CUDA architecture (do not build without one)
+
+`backend/cpp/ds4/Makefile` drives upstream's **object targets** directly
+(`$(MAKE) -C ds4 ds4.o ds4_cuda.o ...`), which bypasses upstream's own guard:
+its `cuda` target refuses to build unless `CUDA_ARCH` is set, and offers
+`cuda-spark` (sm_121, DGX Spark / GB10) and `cuda-generic` (native) instead.
+Built with no `-arch`, nvcc targets its default architecture and the kernels run
+as JIT'd PTX. On GB10 that silently corrupted every prefill batch of >=128
+tokens - the model emitted text unrelated to the prompt and never closed its
+thinking block, so `content` came back empty - and cost close to two orders of
+magnitude of prefill throughput (4.21 t/s vs 325.70 t/s, same box, same model).
+Short prompts stayed correct, which is why it went unnoticed.
+
+The Makefile therefore picks a gencode list from `CUDA_MAJOR_VERSION` (a build
+arg the backend matrix already declares, forwarded by `Dockerfile.ds4`) and
+`uname -m`, and passes it as `NVCC_ARCH_FLAGS` to the sub-make. Upstream's
+`CUDA_ARCH` accepts a single value, so it cannot express the fat binary the
+shipped images need; a command-line assignment beats its `:=`. An empty
+`CUDA_MAJOR_VERSION` falls back to upstream's `native` for local developer
+builds, and an unrecognised one is a hard error - no CI runner has a GPU, so a
+silent `native` there is exactly the failure mode this guards against.
+
+`DS4_CUDA_HAVE_MXF4` is deliberately unset: upstream defines it only for
+single-arch sm_120/sm_121 builds and guards it with a plain `#ifdef` rather than
+`__CUDA_ARCH__`, so it cannot be combined with older archs. It gates an optional
+MXFP4 indexer fast path whose `#ifndef` branch returns 0, so omitting it costs
+speed, not correctness.
+
+### Verifying a build
+
+Check which flags a configuration resolves to, without compiling anything:
+
+```
+make -C backend/cpp/ds4 BUILD_TYPE=cublas CUDA_MAJOR_VERSION=13 NATIVE=false \
+  --eval='show: ; @echo [$(DS4_ARCH_MAKEVARS)]' show
+```
+
+Do not use `make -n` for this: the recipe is `+$(MAKE) ...`, and the `+` prefix
+makes it run even under `-n`.
+
+Then exercise the failure mode itself against a built backend. It only appears
+above one prefill batch, so the ordinary `predict` spec cannot catch it:
+
+```
+BACKEND_BINARY=$(pwd)/backend/cpp/ds4/package/run.sh \
+BACKEND_TEST_MODEL_FILE=/path/to/ds4flash.gguf \
+BACKEND_TEST_CAPS=health,load,predict,long_prefill \
+go test -count=1 -timeout=30m -v ./tests/e2e-backends/...
+```
+
 ## Build matrix
 
 | Build | Where | Notes |

.dockerignore

@@ -59,7 +59,9 @@ backend/rust/*/target
 backend-images
 local-backends
 local-ai
+.claude
 .crush
+.tools
 protoc
 tests
 

.github/PULL_REQUEST_TEMPLATE.md

@@ -5,7 +5,7 @@ This PR fixes #
 **Notes for Reviewers**
 
 
-**[Signed commits](../CONTRIBUTING.md#signing-off-on-commits-developer-certificate-of-origin)**
+**[Signed commits](../CONTRIBUTING.md#commit-messages)**
 - [ ] Yes, I signed my commits.
 - [ ] Documentation updated (docs/content/) for user-facing changes, or not applicable
  

.github/backend-matrix.yml

@@ -2584,7 +2584,7 @@ include:
     tag-latest: 'auto'
     tag-suffix: '-gpu-intel-vllm'
     runs-on: 'ubuntu-latest'
-    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    base-image: "intel/oneapi-basekit:2025.3.2-0-devel-ubuntu24.04"
     skip-drivers: 'false'
     backend: "vllm"
     dockerfile: "./backend/Dockerfile.python"
@@ -3754,6 +3754,19 @@ include:
     dockerfile: "./backend/Dockerfile.golang"
     context: "./"
     ubuntu-version: '2404'
+  - build-type: 'hipblas'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-rocm-hipblas-stablediffusion-ggml'
+    runs-on: 'ubuntu-latest'
+    base-image: "rocm/dev-ubuntu-24.04:7.2.1"
+    skip-drivers: 'false'
+    backend: "stablediffusion-ggml"
+    dockerfile: "./backend/Dockerfile.golang"
+    context: "./"
+    ubuntu-version: '2404'
   - build-type: 'sycl_f16'
     cuda-major-version: ""
     cuda-minor-version: ""

.github/bump_vllm_metal.sh

@@ -3,9 +3,9 @@
 # darwin (Apple Silicon) install path. The macOS/Metal build
 # (backend/python/vllm/install.sh, Darwin branch) installs vllm-metal, which is
 # version-locked to a specific vLLM source release. install.sh derives that vLLM
-# version at build time from vllm-metal's own installer (`vllm_v=`) at the pinned
-# tag, so there is only ONE value to bump here -- mirroring bump_vllm_wheel.sh,
-# which bumps the Linux cu130 wheel pin.
+# version, and the wheel asset name, at build time from the pinned tag, so there
+# is only ONE value to bump here -- mirroring bump_vllm_wheel.sh, which bumps the
+# Linux cu130 wheel pin.
 #
 # This deliberately tracks vllm-project/vllm-metal, NOT vllm-project/vllm: the
 # darwin build can only use the exact vLLM version vllm-metal supports, so it may
@@ -23,19 +23,24 @@ if [ -z "$FILE" ] || [ -z "$REPO" ] || [ -z "$VAR" ]; then
     exit 1
 fi
 
-# vllm-metal ships frequent dev releases, all flagged as non-prerelease, so
-# /releases/latest returns the newest one (with its cp312 wheel asset).
+# vllm-metal ships frequent .dev releases, flagged as prereleases, alongside the
+# stable ones. /releases/latest skips the prereleases and returns the newest
+# stable tag, which is what darwin should pin: upstream deletes and re-cuts .dev
+# tags, and a pin to a deleted tag 404s the whole build.
 LATEST_TAG=$(gh_curl -H "Accept: application/vnd.github+json" \
     "https://api.github.com/repos/$REPO/releases/latest" \
     | python3 -c "import json,sys; print(json.load(sys.stdin)['tag_name'])")
 
-# The coupled vLLM source version lives in vllm-metal's installer at that tag.
-NEW_VLLM_VERSION=$(gh_curl \
-    "https://raw.githubusercontent.com/$REPO/$LATEST_TAG/install.sh" \
-    | grep -oE 'vllm_v="[0-9]+\.[0-9]+\.[0-9]+"' | head -1 | cut -d'"' -f2)
+# The coupled vLLM release lives in .github/vllm-release-tag.commit at that tag
+# (since vllm-metal 0.28); releases predating that file pinned it inline in their
+# own install.sh. The extractor reads both forms.
+NEW_VLLM_VERSION=$( { gh_curl \
+    "https://raw.githubusercontent.com/$REPO/$LATEST_TAG/.github/vllm-release-tag.commit" \
+    || gh_curl "https://raw.githubusercontent.com/$REPO/$LATEST_TAG/install.sh"; } \
+    | "$(dirname "${BASH_SOURCE[0]}")/../scripts/lib/extract-vllm-metal-version.sh")
 
 if [ -z "$LATEST_TAG" ] || [ -z "$NEW_VLLM_VERSION" ]; then
-    echo "Could not resolve vllm-metal tag ($LATEST_TAG) or its vllm_v ($NEW_VLLM_VERSION)." >&2
+    echo "Could not resolve vllm-metal tag ($LATEST_TAG) or its vLLM version ($NEW_VLLM_VERSION)." >&2
     exit 1
 fi
 

.github/ci/test-website-review-143.sh

@@ -0,0 +1,44 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+python3 - <<'PY'
+from pathlib import Path
+
+home = Path("website/layouts/index.html").read_text()
+css = Path("website/static/css/site.css").read_text()
+install = Path("docs/content/getting-started/install.md").read_text()
+containers = Path("docs/content/getting-started/containers.md").read_text()
+
+def require(condition, message):
+    if not condition:
+        raise SystemExit(f"FAIL: {message}")
+
+require("Drop-in replacement for most upstream APIs." in home,
+        "homepage must use the requested drop-in API heading")
+require("Everything else plugs into LocalAI." not in home,
+        "old runtime heading must be removed")
+require("When the engine we need" not in home,
+        "hero must describe user outcomes instead of team implementation")
+require('href="mailto:contact@localai.io"' in home and "business" in home.lower(),
+        "homepage must provide a direct business contact action")
+require(home.index('id="localai"') < home.index('id="proof-quotes"') < home.index('id="mission"'),
+        "headline testimonials must directly follow the runtime section")
+require(home.count('id="proof-quotes"') == 1,
+        "headline testimonials must appear exactly once")
+require('id="engines"' not in home and "Engines we build" not in home,
+        "homepage engine showcase must be removed")
+require('href="/docs/installation/index.html"' in home,
+        "installation guide action must use the direct installation URL")
+require('<iframe' in install and "youtube.com/embed/cMVNnlqwfw4" in install,
+        "installation page must embed the walkthrough video")
+require("## Quick Start" not in install,
+        "installation landing page must not duplicate Quick Start")
+for text in ("CUDA 12", "CUDA 13", "ROCm", "Intel", "Jetson", "Vulkan", "fallback"):
+    require(text.lower() in containers.lower(), f"GPU chooser must explain {text}")
+require('class="sn__e"><a href="https://github.com/mudler/parakeet.cpp">parakeet.cpp</a>' in home,
+        "capability engine names must link to their repositories")
+require(".pane{min-height:" in css.replace(" ", ""),
+        "all installation panes must have a fixed minimum height")
+
+print("website review 143 source checks passed")
+PY

.github/dependabot.yml

@@ -29,10 +29,6 @@ updates:
     schedule:
       # Check for updates to GitHub Actions every weekday
       interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/backend/python/bark"
-    schedule:
-      interval: "weekly"
   - package-ecosystem: "pip"
     directory: "/backend/python/common/template"
     schedule:
@@ -55,30 +51,10 @@ updates:
     ignore:
     - dependency-name: "torch"
     - dependency-name: "transformers"
-  - package-ecosystem: "pip"
-    directory: "/backend/python/exllama"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/backend/python/exllama2"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/backend/python/mamba"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/backend/python/openvoice"
-    schedule:
-      interval: "weekly"
   - package-ecosystem: "pip"
     directory: "/backend/python/rerankers"
     schedule:
       interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/backend/python/sentencetransformers"
-    schedule:
-      interval: "weekly"
   - package-ecosystem: "pip"
     directory: "/backend/python/transformers"
     schedule:
@@ -86,44 +62,4 @@ updates:
   - package-ecosystem: "pip"
     directory: "/backend/python/vllm"
     schedule:
-      interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/examples/chainlit"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/examples/functions"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/examples/langchain/langchainpy-localai-example"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/examples/langchain-chroma"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/examples/streamlit-bot"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "docker"
-    directory: "/examples/k8sgpt"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "docker"
-    directory: "/examples/kubernetes"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "docker"
-    directory: "/examples/langchain"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "gomod"
-    directory: "/examples/semantic-todo"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "docker"
-    directory: "/examples/telegram-bot"
-    schedule:
-      interval: "weekly"
+      interval: "weekly"

.github/workflows/bump_deps.yaml

@@ -166,7 +166,7 @@ jobs:
           push-to-fork: ci-forks/LocalAI
           commit-message: ':arrow_up: Update ${{ matrix.repository }}'
           title: 'chore: :arrow_up: Update ${{ matrix.repository }} to `${{ steps.bump.outputs.commit }}`'
-          branch: "update/${{ matrix.variable }}"
+          branch: "bump/${{ matrix.variable }}"
           body: ${{ steps.bump.outputs.message }}
           signoff: true
 
@@ -203,7 +203,7 @@ jobs:
           push-to-fork: ci-forks/LocalAI
           commit-message: ':arrow_up: Update vllm-project/vllm cu130 wheel'
           title: 'chore: :arrow_up: Update vllm-project/vllm cu130 wheel to `${{ steps.bump.outputs.commit }}`'
-          branch: "update/VLLM_VERSION"
+          branch: "bump/VLLM_VERSION"
           body: ${{ steps.bump.outputs.message }}
           signoff: true
 
@@ -241,6 +241,6 @@ jobs:
           push-to-fork: ci-forks/LocalAI
           commit-message: ':arrow_up: Update vllm-project/vllm-metal (darwin)'
           title: 'chore: :arrow_up: Update vllm-metal (darwin) to `${{ steps.bump.outputs.commit }}`'
-          branch: "update/VLLM_METAL_VERSION"
+          branch: "bump/VLLM_METAL_VERSION"
           body: ${{ steps.bump.outputs.message }}
           signoff: true

.github/workflows/notify-releases.yaml

@@ -31,13 +31,14 @@ jobs:
             messages: [
                 {
                 role: "system",
-                content: "Write a discord message with a bullet point summary of the release notes."
+                content: "Write a Discord message with a bullet point summary of the release notes. Keep the complete message under 1800 characters."
                 },
                 {
                 role: "user",
                 content: $input
                 }
-            ]
+            ],
+            max_tokens: 450
             }')
 
             # Send the request to LocalAI API
@@ -46,7 +47,7 @@ jobs:
             -d "$json_payload")
 
             # Extract the summary from the response
-            summary=$(echo $response | jq -r '.choices[0].message.content')
+            summary=$(printf '%s' "$response" | jq -er '.choices[0].message.content | strings | .[0:1800]')
 
             # Print the summary
             #  -H "Authorization: Bearer $API_KEY" \

.github/workflows/refresh-site-counters.yml

@@ -14,6 +14,7 @@ on:
 
 permissions:
   contents: write
+  pull-requests: write
 
 concurrency:
   group: refresh-site-counters
@@ -30,15 +31,25 @@ jobs:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: ./.github/ci/refresh-site-counters.sh
 
-      - name: Commit only if something moved
+      - name: Show changes
         run: |
           if git diff --quiet -- website/data/stats.yaml; then
-            echo "counters unchanged, nothing to commit"
-            exit 0
+            echo "counters unchanged"
+          else
+            git diff --unified=0 -- website/data/stats.yaml
           fi
-          git diff --unified=0 -- website/data/stats.yaml
-          git config user.name  "github-actions[bot]"
-          git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
-          git add website/data/stats.yaml
-          git commit -m "chore(website): refresh the counters"
-          git push
+
+      - name: Create pull request when counters moved
+        uses: peter-evans/create-pull-request@v8
+        with:
+          token: ${{ secrets.UPDATE_BOT_TOKEN }}
+          push-to-fork: ci-forks/LocalAI
+          commit-message: "chore(website): refresh the counters"
+          title: "chore(website): refresh the counters"
+          body: |
+            Weekly refresh of the landing-page counters from the GitHub API.
+
+            This PR was created automatically by the `refresh-site-counters` workflow.
+          branch: update/site-counters
+          delete-branch: true
+          labels: automated

.github/workflows/test-extra.yml

@@ -526,6 +526,7 @@ jobs:
       - name: Build llama-cpp backend image and run gRPC e2e tests
         run: |
           make test-extra-backend-llama-cpp
+          make test-extra-backend-llama-cpp-embeddings
   tests-llama-cpp-grpc-transcription:
     needs: detect-changes
     if: needs.detect-changes.outputs.llama-cpp == 'true' || needs.detect-changes.outputs.run-all == 'true'

.github/workflows/test.yml

@@ -65,6 +65,12 @@ jobs:
       - name: Test (with coverage gate)
         run: |
           PATH="$PATH:/root/go/bin" make --jobs 5 --output-sync=target test-coverage-check
+      # tests/integration is outside the coverage roots because its store specs
+      # need a live backend. test-stores builds and installs local-store before
+      # running the complete suite, so new local-store specs are collected
+      # automatically without adding another workflow entry.
+      - name: Test local-store integration
+        run: PATH="$PATH:$HOME/go/bin" make test-stores
       - name: Upload coverage report
         if: ${{ always() }}
         uses: actions/upload-artifact@v4
@@ -74,8 +80,13 @@ jobs:
             coverage/coverage.out
             coverage/coverage.html
           if-no-files-found: ignore
+      # tmate keeps the runner busy until the 6 hour job limit, so a single
+      # failure costs a whole runner slot. Only open a session when someone
+      # asked for one by labelling the pull request `ci-debug`, and cap the
+      # session so a forgotten label cannot idle a runner either.
       - name: Setup tmate session if tests fail
-        if: ${{ failure() }}
+        if: ${{ failure() && contains(github.event.pull_request.labels.*.name, 'ci-debug') }}
+        timeout-minutes: 30
         uses: mxschmitt/action-tmate@v3.23
         with:
           detached: true
@@ -119,8 +130,13 @@ jobs:
           export PATH="/opt/homebrew/opt/make/libexec/gnubin:$PATH"
           PATH="$PATH:$HOME/go/bin" make protogen-go
           PATH="$PATH:$HOME/go/bin" BUILD_TYPE="GITHUB_CI_HAS_BROKEN_METAL" CMAKE_ARGS="-DGGML_F16C=OFF -DGGML_AVX512=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF" make --jobs 4 --output-sync=target test
+      # tmate keeps the runner busy until the 6 hour job limit, so a single
+      # failure costs a whole runner slot. Only open a session when someone
+      # asked for one by labelling the pull request `ci-debug`, and cap the
+      # session so a forgotten label cannot idle a runner either.
       - name: Setup tmate session if tests fail
-        if: ${{ failure() }}
+        if: ${{ failure() && contains(github.event.pull_request.labels.*.name, 'ci-debug') }}
+        timeout-minutes: 30
         uses: mxschmitt/action-tmate@v3.23
         with:
           detached: true

.github/workflows/tests-aio.yml

@@ -77,8 +77,13 @@ jobs:
       - name: Test
         run: |
             PATH="$PATH:$HOME/go/bin" make backends/local-store backends/silero-vad backends/llama-cpp backends/whisper backends/piper backends/stablediffusion-ggml docker-build-e2e e2e-aio
+      # tmate keeps the runner busy until the 6 hour job limit, so a single
+      # failure costs a whole runner slot. Only open a session when someone
+      # asked for one by labelling the pull request `ci-debug`, and cap the
+      # session so a forgotten label cannot idle a runner either.
       - name: Setup tmate session if tests fail
-        if: ${{ failure() }}
+        if: ${{ failure() && contains(github.event.pull_request.labels.*.name, 'ci-debug') }}
+        timeout-minutes: 30
         uses: mxschmitt/action-tmate@v3.23
         with:
           detached: true

.github/workflows/tests-e2e.yml

@@ -63,8 +63,13 @@ jobs:
       - name: Test Backend E2E
         run: |
           PATH="$PATH:$HOME/go/bin" make build-mock-backend test-e2e
+      # tmate keeps the runner busy until the 6 hour job limit, so a single
+      # failure costs a whole runner slot. Only open a session when someone
+      # asked for one by labelling the pull request `ci-debug`, and cap the
+      # session so a forgotten label cannot idle a runner either.
       - name: Setup tmate session if tests fail
-        if: ${{ failure() }}
+        if: ${{ failure() && contains(github.event.pull_request.labels.*.name, 'ci-debug') }}
+        timeout-minutes: 30
         uses: mxschmitt/action-tmate@v3.23
         with:
           detached: true

.github/workflows/tests-pii-ner-e2e.yml

@@ -88,8 +88,13 @@ jobs:
       # CPU and runs the token_classify capability spec (byte-offset contract).
       - name: Run live PII NER backend E2E
         run: PATH="$PATH:$HOME/go/bin" make test-extra-backend-privacy-filter
+      # tmate keeps the runner busy until the 6 hour job limit, so a single
+      # failure costs a whole runner slot. Only open a session when someone
+      # asked for one by labelling the pull request `ci-debug`, and cap the
+      # session so a forgotten label cannot idle a runner either.
       - name: Setup tmate session if tests fail
-        if: ${{ failure() }}
+        if: ${{ failure() && contains(github.event.pull_request.labels.*.name, 'ci-debug') }}
+        timeout-minutes: 30
         uses: mxschmitt/action-tmate@v3.23
         with:
           detached: true

.github/workflows/tests-ui-e2e.yml

@@ -52,6 +52,8 @@ jobs:
         run: |
           sudo apt-get update
           sudo apt-get install -y build-essential libopus-dev
+      - name: Run stale chunk recovery tests
+        run: PATH="$PATH:$HOME/go/bin" make test-ui-stale-chunk
       # Builds an instrumented UI bundle, runs the Playwright specs, and fails
       # if line coverage regressed beyond the jitter tolerance (the gate is
       # in `make test-ui-coverage-check`). PLAYWRIGHT_CHROMIUM_PATH is unset
@@ -73,8 +75,13 @@ jobs:
           path: core/http/react-ui/coverage/
           if-no-files-found: ignore
           retention-days: 7
+      # tmate keeps the runner busy until the 6 hour job limit, so a single
+      # failure costs a whole runner slot. Only open a session when someone
+      # asked for one by labelling the pull request `ci-debug`, and cap the
+      # session so a forgotten label cannot idle a runner either.
       - name: Setup tmate session if tests fail
-        if: ${{ failure() }}
+        if: ${{ failure() && contains(github.event.pull_request.labels.*.name, 'ci-debug') }}
+        timeout-minutes: 30
         uses: mxschmitt/action-tmate@v3.23
         with:
           detached: true

.impeccable.md

@@ -0,0 +1,21 @@
+## Design Context
+
+### Users
+
+LocalAI serves both single-host users who want to install and try models quickly and experienced developers, ML engineers, system administrators, and DevOps operators who manage production hosts or distributed clusters. The interface must support first-time discovery without hiding the runtime state, configuration, and control that returning operators need.
+
+### Brand Personality
+
+Capable, easy to use, and trustworthy. The interface should make sophisticated local-AI infrastructure feel understandable and under control. It should be direct and calm rather than playful, ornamental, or intimidating.
+
+### Aesthetic Direction
+
+Use LocalAI's established technical, editorial design language: Geist typography, compact information density, sharp geometry, deep blue-black surfaces, action blue, mint for healthy/local/live state, and amber only for decisions requiring attention. Support both dark and light themes. Avoid generic card dashboards, decorative gradients, glass effects, and visual noise.
+
+### Design Principles
+
+1. Use progressive disclosure to serve newcomers and operators in the same workflow: make the common path obvious, then reveal operational depth in context.
+2. Organize navigation around user intent and lifecycle state, not implementation concepts or nested containers.
+3. Give each resource one canonical home; expose discovery, installed state, and runtime state as clear views of that resource instead of duplicating management surfaces.
+4. Keep operational status visible and trustworthy through precise labels, explicit scope, and actionable state—not decoration.
+5. Preserve information density for expert use while flattening navigation and reducing repeated summaries, tabs, rails, and panels.

ADOPTERS.md

@@ -27,6 +27,7 @@ To be removed, open a pull request deleting your row, or email
 
 | Organisation | What they use it for | Status |
 |---|---|---|
+| [walcz.de](https://walcz.de) | Self-hosted appliance for a German B2B consultancy: local-only inference on AMD Strix Halo (gfx1151/ROCm), agents with MCP tools, RAG over an internal knowledge base, and a document/bookkeeping pipeline. | Production |
 | _Your organisation here_ | | |
 
 ## What this list is not

AGENTS.md

@@ -8,7 +8,7 @@ Human contributors: see [CONTRIBUTING.md](CONTRIBUTING.md) for the development w
 
 LocalAI follows the Linux kernel project's [guidelines for AI coding assistants](https://docs.kernel.org/process/coding-assistants.html). Before submitting AI-assisted code, read [.agents/ai-coding-assistants.md](.agents/ai-coding-assistants.md). Key rules:
 
-- **No `Signed-off-by` from AI.** Only the human submitter may sign off on the Developer Certificate of Origin.
+- **No `Signed-off-by` from AI.** Only the human submitter may sign off on the Developer Certificate of Origin. One exception: automation a maintainer operates signs off with *that maintainer's* identity, since no other human submitter exists to certify it. See [.agents/ai-coding-assistants.md](.agents/ai-coding-assistants.md).
 - **No `Co-Authored-By: <AI>` trailers.** The human contributor owns the change.
 - **Use an `Assisted-by:` trailer** to attribute AI involvement. Format: `Assisted-by: AGENT_NAME:MODEL_VERSION [TOOL1] [TOOL2]`.
 - **The human submitter is responsible** for reviewing, testing, and understanding every line of generated code.
@@ -33,6 +33,7 @@ LocalAI follows the Linux kernel project's [guidelines for AI coding assistants]
 | [.agents/localai-assistant-mcp.md](.agents/localai-assistant-mcp.md) | LocalAI Assistant chat modality — adding admin tools to the in-process MCP server, editing skill prompts, keeping REST + MCP + skills in sync |
 | [.agents/backend-signing.md](.agents/backend-signing.md) | Backend OCI image signing (keyless cosign + sigstore-go) — producer-side CI setup, consumer-side gallery `verification:` block, strict mode (`LOCALAI_REQUIRE_BACKEND_INTEGRITY`), revocation via `not_before` |
 | [.agents/preparing-a-release.md](.agents/preparing-a-release.md) | Cutting a release: PR labels, `RELEASE_NOTES_vX.Y.Z.md`, the blog post under `website/content/blog/`, and the demo clips under `website/static/media/` |
+| [.impeccable.md](.impeccable.md) | Design context for UI/UX work — users, brand personality, aesthetic direction, and design principles |
 
 ## Quick Reference
 

CONTRIBUTING.md

@@ -218,7 +218,7 @@ LocalAI follows the **same guidelines as the Linux kernel project** for AI-assis
 
 The full policy for this repository lives in [`.agents/ai-coding-assistants.md`](.agents/ai-coding-assistants.md). Summary:
 
-- **AI agents MUST NOT add `Signed-off-by` tags.** Only humans can certify the Developer Certificate of Origin.
+- **AI agents MUST NOT add `Signed-off-by` tags.** Only humans can certify the Developer Certificate of Origin. Automation operated by a maintainer is the one exception: it signs off with that maintainer's identity, because there is no other human submitter to certify it.
 - **AI agents MUST NOT add `Co-Authored-By` trailers** attributing themselves as co-authors.
 - **Attribute AI involvement with an `Assisted-by` trailer** in the commit message:
 

Makefile

@@ -34,6 +34,11 @@ TEST_FLAKES?=5
 RANDOM := $(shell bash -c 'echo $$RANDOM')
 
 VERSION?=$(shell git describe --always --tags || echo "dev" )
+# fyne package only accepts numeric x[.y[.z]] app versions, so reduce git
+# describe output (v4.9.0, v4.9.0-14-gabc1234, or a bare sha on untagged
+# checkouts) to its numeric core; anything non-numeric falls back to 0.0.0.
+# Without this the packaged launcher reports itself as version 0.0.0 (#11673).
+LAUNCHER_APP_VERSION?=$(shell v=$$(echo "$(VERSION)" | sed -E 's/^v//; s/[+-].*$$//'); echo "$$v" | grep -qE '^[0-9]+(\.[0-9]+){0,2}$$' && echo "$$v" || echo "0.0.0")
 # go tool nm ./local-ai | grep Commit
 LD_FLAGS?=-s -w
 override LD_FLAGS += -X "github.com/mudler/LocalAI/internal.Version=$(VERSION)"
@@ -103,7 +108,7 @@ COVERAGE_E2E_LABELS?=!real-models
 COVERAGE_EXCLUDE_RE?=grpc/proto/.*[.]pb[.]go
 
 
-.PHONY: all test test-coverage test-coverage-baseline test-coverage-check test-backend-cpp test-build-scripts test-ui test-ui-coverage-baseline test-ui-coverage-check build vendor lint lint-all
+.PHONY: all test test-coverage test-coverage-baseline test-coverage-check test-backend-cpp test-build-scripts test-ui test-ui-stale-chunk test-ui-coverage-baseline test-ui-coverage-check build vendor lint lint-all
 
 all: help
 
@@ -235,7 +240,7 @@ test-ci-scripts:
 ## pure stdlib on purpose so they run without any backend venv; the list is
 ## explicit because their siblings (model_identity_test) import grpc and the
 ## generated protobufs, which only exist inside a built backend.
-PYTHON_HELPER_TESTS?=python_utils_test vllm_utils_test model_utils_test mlx_utils_test parent_watch_test
+PYTHON_HELPER_TESTS?=python_utils_test vllm_utils_test model_utils_test mlx_utils_test parent_watch_test temp_utils_test
 test-python-helpers:
 	cd backend/python/common && python3 -m unittest $(PYTHON_HELPER_TESTS)
 
@@ -388,9 +393,17 @@ test-e2e: build-mock-backend build-cloud-proxy-backend prepare-e2e run-e2e-image
 	$(MAKE) teardown-e2e
 	docker rmi localai-tests
 
+# `docker stop` returns as soon as the container exits, but Docker reaps a
+# `--rm` container asynchronously after that. The `docker rmi localai-tests` in
+# test-e2e then loses the race against the reaper and fails on a still
+# referenced image, turning a green suite red. Removing the container ourselves
+# is synchronous, so the image reference is gone before we return. It also
+# covers the case where nothing is running, which `docker stop` could not
+# because it rejects an empty argument list.
 teardown-e2e:
 	rm -rf $(TEST_DIR) || true
-	docker stop $$(docker ps -q --filter ancestor=localai-tests)
+	@CONTAINERS=$$(docker ps -aq --filter ancestor=localai-tests 2>/dev/null); \
+	if [ -n "$$CONTAINERS" ]; then docker rm -f $$CONTAINERS || true; fi
 
 ########################################################
 ## Integration and unit tests
@@ -676,6 +689,7 @@ test-extra: prepare-test-extra
 ##   BACKEND_TEST_PROMPT      Override the prompt used in predict/stream specs.
 ##   BACKEND_TEST_OPTIONS     Comma-separated Options[] entries forwarded to LoadModel,
 ##                            e.g. "tool_parser:hermes,reasoning_parser:qwen3".
+##   BACKEND_TEST_EMBEDDING_LAYOUT Expected EmbeddingResult layout: "final" or "per_token".
 ##
 ## Direct usage (image already built, no docker-build-* dependency):
 ##
@@ -705,6 +719,7 @@ test-extra-backend: protogen-go
 	BACKEND_TEST_CAPS="$$BACKEND_TEST_CAPS" \
 	BACKEND_TEST_PROMPT="$$BACKEND_TEST_PROMPT" \
 	BACKEND_TEST_OPTIONS="$$BACKEND_TEST_OPTIONS" \
+	BACKEND_TEST_EMBEDDING_LAYOUT="$$BACKEND_TEST_EMBEDDING_LAYOUT" \
 	BACKEND_TEST_TOOL_PROMPT="$$BACKEND_TEST_TOOL_PROMPT" \
 	BACKEND_TEST_TOOL_NAME="$$BACKEND_TEST_TOOL_NAME" \
 	BACKEND_TEST_CACHE_TYPE_K="$$BACKEND_TEST_CACHE_TYPE_K" \
@@ -724,6 +739,15 @@ test-extra-backend-llama-cpp: docker-build-llama-cpp
 	BACKEND_TEST_CAPS=health,load,predict,stream,logprobs,logit_bias \
 	$(MAKE) test-extra-backend
 
+## Raw llama.cpp embeddings are required by Go-side pooling. This exercises the
+## real C++ backend and verifies that it marks the flattened matrix per-token.
+test-extra-backend-llama-cpp-embeddings: docker-build-llama-cpp
+	BACKEND_IMAGE=local-ai-backend:llama-cpp \
+	BACKEND_TEST_CAPS=health,load,embeddings \
+	BACKEND_TEST_OPTIONS=pooling:none \
+	BACKEND_TEST_EMBEDDING_LAYOUT=per_token \
+	$(MAKE) test-extra-backend
+
 test-extra-backend-ik-llama-cpp: docker-build-ik-llama-cpp
 	BACKEND_IMAGE=local-ai-backend:ik-llama-cpp $(MAKE) test-extra-backend
 
@@ -813,6 +837,7 @@ test-extra-backend-tinygrad-embeddings: docker-build-tinygrad
 	BACKEND_IMAGE=local-ai-backend:tinygrad \
 	BACKEND_TEST_MODEL_NAME=Qwen/Qwen3-0.6B \
 	BACKEND_TEST_CAPS=health,load,embeddings \
+	BACKEND_TEST_EMBEDDING_LAYOUT=final \
 	$(MAKE) test-extra-backend
 
 ## tinygrad — Stable Diffusion 1.5. The original CompVis/runwayml repos have
@@ -1505,6 +1530,13 @@ test-ui: build-mock-backend protogen-go
 	$(GOCMD) build -o tests/e2e-ui/ui-test-server ./tests/e2e-ui
 	cd core/http/react-ui && sh $(CURDIR)/scripts/ensure-playwright-browser.sh && bunx playwright test $(PLAYWRIGHT_WORKERS_FLAG)
 
+## The stale-chunk specs need the production code-split bundle. The V8 coverage
+## bundle below inlines dynamic imports to keep every page in its denominator.
+test-ui-stale-chunk: build-mock-backend protogen-go
+	cd core/http/react-ui && bun install && bun run build
+	$(GOCMD) build -o tests/e2e-ui/ui-test-server ./tests/e2e-ui
+	cd core/http/react-ui && sh $(CURDIR)/scripts/ensure-playwright-browser.sh && bunx playwright test --grep @production-chunks --workers=1
+
 ## React UI code coverage from the Playwright e2e suite. Builds a
 ## NON-instrumented bundle with source maps (COVERAGE_V8=true), re-embeds it
 ## into the ui-test-server (the dist is //go:embed'ed at compile time), runs the
@@ -1520,7 +1552,7 @@ test-ui-coverage: build-mock-backend protogen-go
 	$(GOCMD) build -o tests/e2e-ui/ui-test-server ./tests/e2e-ui && \
 	( cd core/http/react-ui && rm -rf .nyc_output coverage && \
 	    sh $(CURDIR)/scripts/ensure-playwright-browser.sh && \
-	    PW_V8_COVERAGE=1 bunx playwright test $(PLAYWRIGHT_WORKERS_FLAG) && bun run coverage:report )
+	    PW_V8_COVERAGE=1 bunx playwright test --grep-invert @production-chunks $(PLAYWRIGHT_WORKERS_FLAG) && bun run coverage:report )
 
 ## UI coverage baseline (committed) and the strict gate that compares against
 ## it — the React mirror of test-coverage-baseline / test-coverage-check.
@@ -1603,7 +1635,7 @@ site-serve: site
 build-launcher-darwin:
 	rm -rf dist/LocalAI.app cmd/launcher/LocalAI.app
 	mkdir -p dist
-	cd cmd/launcher && go run fyne.io/tools/cmd/fyne@latest package -os darwin -icon ../../core/http/static/logo.png --executable $(LAUNCHER_BINARY_NAME)
+	cd cmd/launcher && go run fyne.io/tools/cmd/fyne@latest package -os darwin -icon ../../core/http/static/logo.png --executable $(LAUNCHER_BINARY_NAME) --app-version $(LAUNCHER_APP_VERSION)
 	mv cmd/launcher/LocalAI.app dist/LocalAI.app
 	bash contrib/macos/sign-and-notarize.sh sign dist/LocalAI.app
 
@@ -1630,4 +1662,4 @@ release-launcher-darwin: notarize-launcher-darwin
 	@echo "dist/LocalAI.dmg is ready"
 
 build-launcher-linux:
-	cd cmd/launcher && go run fyne.io/tools/cmd/fyne@latest package -os linux -icon ../../core/http/static/logo.png --executable $(LAUNCHER_BINARY_NAME)-linux && mv LocalAI.tar.xz ../../$(LAUNCHER_BINARY_NAME)-linux.tar.xz
+	cd cmd/launcher && go run fyne.io/tools/cmd/fyne@latest package -os linux -icon ../../core/http/static/logo.png --executable $(LAUNCHER_BINARY_NAME)-linux --app-version $(LAUNCHER_APP_VERSION) && mv LocalAI.tar.xz ../../$(LAUNCHER_BINARY_NAME)-linux.tar.xz

README.md

@@ -5,9 +5,6 @@
 </h1>
 
 <p align="center">
-<a href="https://github.com/go-skynet/LocalAI/stargazers" target="blank">
-<img src="https://img.shields.io/github/stars/go-skynet/LocalAI?style=for-the-badge" alt="LocalAI stars"/>
-</a>
 <a href='https://github.com/go-skynet/LocalAI/releases'>
 <img src='https://img.shields.io/github/release/go-skynet/LocalAI?&label=Latest&style=for-the-badge'>
 </a>
@@ -231,7 +228,7 @@ Most backends wrap a best-in-class upstream engine. A handful of them are native
 
 | Backend | What it does |
 |---------|-------------|
-| [vllm.cpp](https://github.com/mudler/vllm.cpp) | From-scratch C++20 port of vLLM for text generation: paged KV cache, continuous batching, prefix caching, safetensors + GGUF loading, engine-enforced structured output, on CPU, CUDA, Metal and Vulkan |
+| [vllm.cpp](https://github.com/mudler/vllm.cpp) | From-scratch C++20 port of vLLM for text generation: paged KV cache, continuous batching, prefix caching, safetensors + GGUF loading, engine-enforced structured output, on CPU, CUDA, Metal and Vulkan. Also serves MiniMax-H3 joint video+audio generation |
 | [parakeet.cpp](https://github.com/mudler/parakeet.cpp) | C++/GGML port of NVIDIA NeMo Parakeet ASR (tdt/ctc/rnnt/hybrid), with cache-aware streaming transcription |
 | [moss-transcribe.cpp](https://github.com/localai-org/moss-transcribe.cpp) | C++/GGML port of OpenMOSS MOSS-Transcribe-Diarize: joint long-form transcription, speaker diarization and timestamping in a single pass |
 | [moss-tts.cpp](https://github.com/mudler/moss-tts.cpp) | C++/GGML port of the OpenMOSS MOSS-TTS family: text-to-speech (MOSS-TTS-Local v1.5, 48 kHz stereo) with reference-audio voice cloning, through the MOSS-Audio-Tokenizer neural codec |
@@ -318,10 +315,6 @@ Past sponsors
 
 A special thanks to individual sponsors, a full list is on [GitHub](https://github.com/sponsors/mudler) and [buymeacoffee](https://buymeacoffee.com/mudler). Special shout out to [drikster80](https://github.com/drikster80) for being generous. Thank you everyone!
 
-## Star history
-
-[![LocalAI Star history Chart](https://api.star-history.com/svg?repos=go-skynet/LocalAI&type=Date)](https://star-history.com/#go-skynet/LocalAI&Date)
-
 ## License
 
 LocalAI is a community-driven project created by [Ettore Di Giacinto](https://github.com/mudler/) and maintained by the [LocalAI team](#team).

backend/Dockerfile.ds4

@@ -10,6 +10,7 @@ FROM ${BASE_IMAGE} AS builder
 ARG BUILD_TYPE
 ARG TARGETARCH
 ARG TARGETVARIANT
+ARG CUDA_MAJOR_VERSION
 
 ENV BUILD_TYPE=${BUILD_TYPE} \
     DEBIAN_FRONTEND=noninteractive \
@@ -35,7 +36,8 @@ RUN apt-get update && \
 COPY . /LocalAI
 
 RUN --mount=type=cache,target=/root/.ccache,id=ds4-ccache-${TARGETARCH}-${BUILD_TYPE},sharing=locked \
-    make -C /LocalAI/backend/cpp/ds4 BUILD_TYPE=${BUILD_TYPE} NATIVE=false grpc-server package
+    make -C /LocalAI/backend/cpp/ds4 BUILD_TYPE=${BUILD_TYPE} \
+        CUDA_MAJOR_VERSION=${CUDA_MAJOR_VERSION} NATIVE=false grpc-server package
 
 FROM scratch
 COPY --from=builder /LocalAI/backend/cpp/ds4/package/. ./

backend/backend.proto

@@ -536,8 +536,28 @@ message Result {
   bool success = 2;
 }
 
+// EmbeddingLayout describes whether embeddings contains one final vector or
+// a matrix of per-token vectors. Go-side pooling must never infer this from
+// tokens/dim alone: a one-token raw matrix and a final vector have the same
+// shape.
+enum EmbeddingLayout {
+  EMBEDDING_LAYOUT_UNSPECIFIED = 0;
+  EMBEDDING_LAYOUT_FINAL = 1;
+  EMBEDDING_LAYOUT_PER_TOKEN = 2;
+}
+
 message EmbeddingResult {
   repeated float embeddings = 1;
+  // Shape of the payload above: dim is the embedding width, tokens is the
+  // number of vectors packed into `embeddings` (1 when the backend pooled
+  // server-side, N with pooling:none; total across prompts if a request
+  // carried several). tokens=0/dim=0 means the backend predates shape
+  // reporting. prompt_tokens is the number of prompt tokens evaluated, for
+  // usage accounting.
+  int32 tokens = 2;
+  int32 dim = 3;
+  int32 prompt_tokens = 4;
+  EmbeddingLayout layout = 5;
 }
 
 message TranscriptRequest {

backend/cpp/audio-cpp/Makefile

@@ -9,7 +9,7 @@
 # recipe is a make target (not a prepare.sh) so 'make purge && make' is a clean
 # rebuild and so the bump bot can see the pin.
 
-AUDIO_CPP_VERSION?=7efbb58def443722ea540d931dd3debee3e4d5e8
+AUDIO_CPP_VERSION?=05e508a70e3600b01454c647cdb122133ba8e64c
 AUDIO_CPP_REPO?=https://github.com/0xShug0/audio.cpp
 
 CURRENT_MAKEFILE_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))

backend/cpp/audio-cpp/capability_routing.cpp

@@ -29,6 +29,7 @@ const NamedTask kTaskNames[] = {
     {Task::VoiceDesign, "vdes"},
     {Task::SpeakerRecognition, "spk"},
     {Task::Svc, "svc"},
+    {Task::Midi, "midi"},
 };
 
 // Accepted on input but never emitted. "spkrec" was this backend's own earlier

backend/cpp/audio-cpp/capability_routing.h

@@ -25,6 +25,7 @@ enum class Task {
     VoiceDesign,
     SpeakerRecognition,
     Svc,
+    Midi,
 };
 
 // Mirrors engine::runtime::RunMode.

backend/cpp/audio-cpp/capability_routing_test.cpp

@@ -361,7 +361,7 @@ static void test_names_round_trip() {
                         Task::SourceSeparation, Task::AudioGeneration, Task::Tts,
                         Task::VoiceCloning, Task::VoiceConversion,
                         Task::SpeechToSpeech, Task::Alignment, Task::VoiceDesign,
-                        Task::SpeakerRecognition, Task::Svc};
+                        Task::SpeakerRecognition, Task::Svc, Task::Midi};
     for (const Task t : all) {
         Task parsed = Task::Vad;
         const bool ok = parse_task_name(task_name(t), parsed);

backend/cpp/audio-cpp/loaded_model.cpp

@@ -69,7 +69,8 @@ static_assert(kEngine(engine::runtime::VoiceTaskKind::VoiceDesign) == 10, "Voice
 static_assert(kEngine(engine::runtime::VoiceTaskKind::SpeakerRecognition) == 11, "VoiceTaskKind drifted");
 // The last member. Pinning it pins the member count too, as long as the
 // enumerators stay contiguous and unassigned, which upstream's declaration is.
-static_assert(kEngine(engine::runtime::VoiceTaskKind::Svc) == 12,
+static_assert(kEngine(engine::runtime::VoiceTaskKind::Svc) == 12, "VoiceTaskKind drifted");
+static_assert(kEngine(engine::runtime::VoiceTaskKind::Midi) == 13,
               "engine::runtime::VoiceTaskKind gained, lost or reordered a member. "
               "audiocpp_backend::Task mirrors it positionally: update capability_routing.h, "
               "to_engine_task and from_engine_task together, then move this pin.");
@@ -87,6 +88,7 @@ static_assert(kMirror(Task::Alignment) == 9, "Task drifted from VoiceTaskKind");
 static_assert(kMirror(Task::VoiceDesign) == 10, "Task drifted from VoiceTaskKind");
 static_assert(kMirror(Task::SpeakerRecognition) == 11, "Task drifted from VoiceTaskKind");
 static_assert(kMirror(Task::Svc) == 12, "Task drifted from VoiceTaskKind");
+static_assert(kMirror(Task::Midi) == 13, "Task drifted from VoiceTaskKind");
 
 static_assert(static_cast<int>(engine::runtime::RunMode::Offline) == 0, "RunMode drifted");
 static_assert(static_cast<int>(engine::runtime::RunMode::Streaming) == 1,
@@ -241,6 +243,7 @@ engine::runtime::VoiceTaskKind to_engine_task(Task task) {
     case Task::VoiceDesign:        return K::VoiceDesign;
     case Task::SpeakerRecognition: return K::SpeakerRecognition;
     case Task::Svc:                return K::Svc;
+    case Task::Midi:               return K::Midi;
     }
     // Unreachable for any valid enumerator. No `default:` label, so -Wswitch
     // still reports a member this switch stops covering.
@@ -263,6 +266,7 @@ Task from_engine_task(engine::runtime::VoiceTaskKind kind) {
     case K::VoiceDesign:        return Task::VoiceDesign;
     case K::SpeakerRecognition: return Task::SpeakerRecognition;
     case K::Svc:                return Task::Svc;
+    case K::Midi:               return Task::Midi;
     }
     return Task::Vad;
 }

backend/cpp/bonsai/Makefile

@@ -1,7 +1,7 @@
 
 # Pinned to the HEAD of the `prism` branch on https://github.com/PrismML-Eng/llama.cpp.
 # Auto-bumped nightly by .github/workflows/bump_deps.yaml.
-BONSAI_VERSION?=9ca265a57f85f2117942490f421f64a226dd9847
+BONSAI_VERSION?=312bb2a93ea2bf798333fa859614fbf913ecb9e2
 LLAMA_REPO?=https://github.com/PrismML-Eng/llama.cpp
 
 CMAKE_ARGS?=
@@ -41,7 +41,9 @@ define bonsai-build
 	# and are applied by apply-patches.sh below.
 	rm -rf $(CURRENT_MAKEFILE_DIR)/../bonsai-$(1)-build/patches
 	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../bonsai-$(1)-build purge
+	bash $(CURRENT_MAKEFILE_DIR)/patch-grpc-server.sh $(CURRENT_MAKEFILE_DIR)/../bonsai-$(1)-build/grpc-server.cpp
 	bash $(LLAMA_CPP_DIR)/disable-score-task.sh $(CURRENT_MAKEFILE_DIR)/../bonsai-$(1)-build/grpc-server.cpp
+	bash $(LLAMA_CPP_DIR)/disable-tts-task.sh $(CURRENT_MAKEFILE_DIR)/../bonsai-$(1)-build/grpc-server.cpp
 	$(info $(GREEN)I bonsai build info:$(1)$(RESET))
 	LLAMA_REPO=$(LLAMA_REPO) LLAMA_VERSION=$(BONSAI_VERSION) \
 	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../bonsai-$(1)-build llama.cpp
@@ -78,7 +80,9 @@ bonsai-cpu-all:
 	# and are applied by apply-patches.sh below.
 	rm -rf $(CURRENT_MAKEFILE_DIR)/../bonsai-cpu-all-build/patches
 	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../bonsai-cpu-all-build purge
+	bash $(CURRENT_MAKEFILE_DIR)/patch-grpc-server.sh $(CURRENT_MAKEFILE_DIR)/../bonsai-cpu-all-build/grpc-server.cpp
 	bash $(LLAMA_CPP_DIR)/disable-score-task.sh $(CURRENT_MAKEFILE_DIR)/../bonsai-cpu-all-build/grpc-server.cpp
+	bash $(LLAMA_CPP_DIR)/disable-tts-task.sh $(CURRENT_MAKEFILE_DIR)/../bonsai-cpu-all-build/grpc-server.cpp
 	$(info $(GREEN)I bonsai build info:cpu-all-variants$(RESET))
 	LLAMA_REPO=$(LLAMA_REPO) LLAMA_VERSION=$(BONSAI_VERSION) \
 	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../bonsai-cpu-all-build llama.cpp

backend/cpp/bonsai/patch-grpc-server.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+# Adapt the shared llama.cpp gRPC source to the older JSON API in Bonsai.
+
+set -euo pipefail
+
+if [[ $# -ne 1 ]]; then
+    echo "usage: $0 <grpc-server.cpp>" >&2
+    exit 2
+fi
+
+SRC=$1
+if [[ ! -f "$SRC" ]]; then
+    echo "grpc-server.cpp not found at $SRC" >&2
+    exit 2
+fi
+
+if grep -q 'common_json_error' "$SRC"; then
+    echo "==> patching $SRC to use the Bonsai JSON exception type"
+    awk '{ gsub(/common_json_error/, "json::parse_error"); print }' "$SRC" > "$SRC.tmp"
+    mv "$SRC.tmp" "$SRC"
+    echo "==> Bonsai JSON exception patch OK"
+else
+    echo "==> $SRC already uses a Bonsai-compatible JSON exception type, skipping"
+fi

backend/cpp/ds4/CMakeLists.txt

@@ -84,9 +84,10 @@ elseif(DS4_GPU STREQUAL "cpu")
     set(DS4_OBJS "${DS4_DIR}/ds4_cpu.o")
 endif()
 
-# Upstream splits distributed inference, tensor-parallel transport, the SSD
-# expert cache, and layer placement into GPU-agnostic translation units. Link
-# them regardless of DS4_GPU.
+# Upstream splits image preprocessing, distributed inference, tensor-parallel
+# transport, the SSD expert cache, and layer placement into GPU-agnostic
+# translation units. Link them regardless of DS4_GPU.
+list(APPEND DS4_OBJS "${DS4_DIR}/ds4_image.o")
 list(APPEND DS4_OBJS "${DS4_DIR}/ds4_distributed.o")
 list(APPEND DS4_OBJS "${DS4_DIR}/ds4_tp.o")
 list(APPEND DS4_OBJS "${DS4_DIR}/ds4_ssd.o")

backend/cpp/ds4/Makefile

@@ -1,10 +1,10 @@
 # ds4 backend Makefile.
 #
-# Upstream pin lives below as DS4_VERSION?=b0309611041655f4e45671cfd9c9886aff161406
+# Upstream pin lives below as DS4_VERSION?=6289c516273979173abbc062209a81dd3706b804
 # (.github/bump_deps.sh) can find and update it - matches the
 # llama-cpp / ik-llama-cpp / turboquant convention.
 
-DS4_VERSION?=b0309611041655f4e45671cfd9c9886aff161406
+DS4_VERSION?=6289c516273979173abbc062209a81dd3706b804
 DS4_REPO?=https://github.com/antirez/ds4
 
 CURRENT_MAKEFILE_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
@@ -18,21 +18,83 @@ UNAME_S := $(shell uname -s)
 
 CMAKE_ARGS ?= -DCMAKE_BUILD_TYPE=Release
 
-# Upstream splits distributed inference, tensor-parallel transport, the SSD
-# expert cache, and layer placement into GPU-agnostic translation units. They
-# are shared by every GPU mode, so append them unconditionally below.
+# nvcc must be told the target architecture explicitly for a cublas build, and
+# this is not a tuning knob. Upstream's Makefile leaves CUDA_ARCH empty and its
+# `cuda` target REFUSES to build without one, offering `cuda-spark`
+# (CUDA_ARCH=sm_121) and `cuda-generic` (CUDA_ARCH=native) instead. We drive its
+# object targets directly, which bypasses that guard: nvcc then compiles with no
+# -arch at all, and the kernels run as JIT'd PTX for its default architecture.
+# On GB10 (sm_121) that silently produced corrupt inference output above a
+# ~128-token prefill batch and ~77x slower prefill (4.21 t/s vs 325.70 t/s,
+# measured on the same box with the same model). No CI runner has a GPU, so
+# `native` has nothing to enumerate there.
+#
+# Upstream's CUDA_ARCH takes a SINGLE value (see its sm_120/sm_121 special cases
+# and the `-arch=$(CUDA_ARCH)` fallback), so it cannot express the fat binary
+# these images need. NVCC_ARCH_FLAGS is overridden instead: a command-line
+# assignment wins over the `:=` in upstream's Makefile, and its NVCCFLAGS
+# expands whatever we pass.
+#
+# The architecture lists are copied from backend/go/vllm-cpp/Makefile rather
+# than invented, so the two CUDA images cover the same GPUs: amd64 datacenter +
+# consumer, and l4t/arm64 covering Orin (87), Thor (110) and GB10 (121a).
+#
+# -DDS4_CUDA_HAVE_MXF4=1 is deliberately NOT set. Upstream only defines it for
+# single-arch sm_120/sm_121 builds and guards the code with a plain #ifdef
+# rather than __CUDA_ARCH__, so it cannot be combined with older archs in one
+# fat binary. It gates an optional MXFP4 indexer fast path whose #ifndef branch
+# returns 0 and falls back to the generic path, so omitting it costs some speed
+# on GB10, not correctness. Revisit if upstream adds __CUDA_ARCH__ guards.
+#
+# An EMPTY CUDA_MAJOR_VERSION means a local developer build, not CI: fall back
+# to upstream's own `native` handling, which needs a GPU present but is what a
+# developer building on their own machine wants. Both variables are `?=` so an
+# explicit value on the command line always wins.
+UNAME_M := $(shell uname -m)
+CUDA_MAJOR_VERSION ?=
+ifeq ($(BUILD_TYPE),cublas)
+ifeq ($(CUDA_MAJOR_VERSION),13)
+ifeq ($(UNAME_M),aarch64)
+    DS4_NVCC_ARCH_FLAGS ?= -gencode arch=compute_87,code=sm_87 \
+                           -gencode arch=compute_90a,code=sm_90a \
+                           -gencode arch=compute_100a,code=sm_100a \
+                           -gencode arch=compute_110,code=sm_110 \
+                           -gencode arch=compute_121a,code=sm_121a
+else
+    DS4_NVCC_ARCH_FLAGS ?= -gencode arch=compute_80,code=sm_80 \
+                           -gencode arch=compute_86,code=sm_86 \
+                           -gencode arch=compute_89,code=sm_89 \
+                           -gencode arch=compute_90a,code=sm_90a \
+                           -gencode arch=compute_100a,code=sm_100a \
+                           -gencode arch=compute_103a,code=sm_103a \
+                           -gencode arch=compute_120a,code=sm_120a \
+                           -gencode arch=compute_121a,code=sm_121a
+endif
+    DS4_ARCH_MAKEVARS := NVCC_ARCH_FLAGS="$(DS4_NVCC_ARCH_FLAGS)"
+else ifeq ($(CUDA_MAJOR_VERSION),)
+    # Local build: let upstream resolve the host GPU.
+    DS4_ARCH_MAKEVARS := CUDA_ARCH=native
+else
+    $(error CUDA_MAJOR_VERSION=$(CUDA_MAJOR_VERSION) has no architecture list here (13 does). Leave it empty for a native build, or pass DS4_NVCC_ARCH_FLAGS explicitly.)
+endif
+endif
+
+# Upstream splits image preprocessing, distributed inference, tensor-parallel
+# transport, the SSD expert cache, and layer placement into GPU-agnostic
+# translation units. They are shared by every GPU mode, so append them
+# unconditionally below.
 ifeq ($(BUILD_TYPE),cublas)
     CMAKE_ARGS += -DDS4_GPU=cuda
-    DS4_OBJ_TARGET := ds4.o ds4_cuda.o ds4_distributed.o ds4_tp.o ds4_ssd.o ds4_layer_pack.o \
+    DS4_OBJ_TARGET := ds4.o ds4_image.o ds4_cuda.o ds4_distributed.o ds4_tp.o ds4_ssd.o ds4_layer_pack.o \
         cuda/mmq/ds4_ggml_stubs.o cuda/mmq/ds4_mmq.o cuda/mmq/ds4_mmq_d2r.o \
         cuda/mmq/quantize.o cuda/mmq/mmid.o cuda/mmq/mmvq.o cuda/mmq/ds4_repack.o
 else ifeq ($(UNAME_S),Darwin)
     CMAKE_ARGS += -DDS4_GPU=metal
-    DS4_OBJ_TARGET := ds4.o ds4_metal.o ds4_distributed.o ds4_tp.o ds4_ssd.o ds4_layer_pack.o
+    DS4_OBJ_TARGET := ds4.o ds4_image.o ds4_metal.o ds4_distributed.o ds4_tp.o ds4_ssd.o ds4_layer_pack.o
 else
     # CPU reference path (Linux only - macOS CPU path is broken by VM bug per ds4 README).
     CMAKE_ARGS += -DDS4_GPU=cpu
-    DS4_OBJ_TARGET := ds4_cpu.o ds4_distributed.o ds4_tp.o ds4_ssd.o ds4_layer_pack.o
+    DS4_OBJ_TARGET := ds4_cpu.o ds4_image.o ds4_distributed.o ds4_tp.o ds4_ssd.o ds4_layer_pack.o
 endif
 
 ifneq ($(NATIVE),true)
@@ -57,11 +119,11 @@ ds4:
 # the right per-platform compile flags (Objective-C/Metal on Darwin, nvcc on Linux+CUDA).
 ds4/ds4.o: ds4
 ifeq ($(BUILD_TYPE),cublas)
-	+$(MAKE) -C ds4 $(DS4_OBJ_TARGET)
+	+$(MAKE) -C ds4 $(DS4_ARCH_MAKEVARS) $(DS4_OBJ_TARGET)
 else ifeq ($(UNAME_S),Darwin)
-	+$(MAKE) -C ds4 ds4.o ds4_metal.o ds4_distributed.o ds4_tp.o ds4_ssd.o ds4_layer_pack.o
+	+$(MAKE) -C ds4 ds4.o ds4_image.o ds4_metal.o ds4_distributed.o ds4_tp.o ds4_ssd.o ds4_layer_pack.o
 else
-	+$(MAKE) -C ds4 ds4_cpu.o ds4_distributed.o ds4_tp.o ds4_ssd.o ds4_layer_pack.o
+	+$(MAKE) -C ds4 ds4_cpu.o ds4_image.o ds4_distributed.o ds4_tp.o ds4_ssd.o ds4_layer_pack.o
 endif
 
 grpc-server: ds4/ds4.o

backend/cpp/ds4/dsml_parser.cpp

@@ -92,7 +92,8 @@ std::string json_escape(const std::string &in) {
 
 } // namespace
 
-DsmlParser::DsmlParser() = default;
+DsmlParser::DsmlParser(bool starts_in_thinking)
+    : state_(starts_in_thinking ? State::THINK : State::TEXT) {}
 
 bool DsmlParser::IsInDsmlStructural() const {
     switch (state_) {

backend/cpp/ds4/dsml_parser.h

@@ -17,7 +17,9 @@ struct ParserEvent {
 // Streaming parser. Stateless across instances; one per Predict call.
 class DsmlParser {
 public:
-    DsmlParser();
+    // The chat prompt may already contain the opening thinking marker, so the
+    // generated text can begin directly with reasoning bytes.
+    explicit DsmlParser(bool starts_in_thinking = false);
 
     // Feed a chunk of raw model-emitted text. Appends classified events to
     // `out`. May buffer the tail of `chunk` internally if it looks like a
@@ -43,7 +45,7 @@ public:
 
 private:
     enum class State { TEXT, THINK, TOOL_CALLS, INVOKE, PARAM_VALUE };
-    State state_ = State::TEXT;
+    State state_;
     std::string buf_;
     std::string current_tool_name_;
     int tool_index_ = -1;

backend/cpp/ds4/dsml_parser_test.cpp

@@ -0,0 +1,133 @@
+// SPDX-License-Identifier: MIT
+// Standalone regression tests for the DSML streaming parser.
+//
+// The repository's backend/cpp/run-unit-tests.sh harness compiles each
+// *_test.cpp as a single translation unit, so include the implementation here.
+
+#include "dsml_parser.cpp"
+
+#include <cstdio>
+#include <string>
+#include <type_traits>
+#include <vector>
+
+namespace {
+
+struct ParsedText {
+    std::string content;
+    std::string reasoning;
+};
+
+int failures = 0;
+
+void check_equal(const std::string &got, const std::string &want,
+                 const char *name) {
+    if (got == want) return;
+    std::fprintf(stderr, "FAIL %s: got \"%s\", want \"%s\"\n",
+                 name, got.c_str(), want.c_str());
+    failures++;
+}
+
+void collect_text(const std::vector<ds4cpp::ParserEvent> &events,
+                  ParsedText *parsed) {
+    for (const auto &event : events) {
+        if (event.type == ds4cpp::ParserEvent::CONTENT) {
+            parsed->content += event.text;
+        } else if (event.type == ds4cpp::ParserEvent::REASONING) {
+            parsed->reasoning += event.text;
+        }
+    }
+}
+
+ParsedText parse_chunks(ds4cpp::DsmlParser *parser,
+                        const std::vector<std::string> &chunks) {
+    ParsedText parsed;
+    for (const auto &chunk : chunks) {
+        std::vector<ds4cpp::ParserEvent> events;
+        parser->Feed(chunk, events);
+        collect_text(events, &parsed);
+    }
+    std::vector<ds4cpp::ParserEvent> events;
+    parser->Flush(events);
+    collect_text(events, &parsed);
+    return parsed;
+}
+
+template <typename Parser>
+void test_reasoning_opened_by_prompt() {
+    if constexpr (!std::is_constructible_v<Parser, bool>) {
+        std::fprintf(stderr,
+                     "FAIL reasoning_opened_by_prompt: parser cannot start in thinking state\n");
+        failures++;
+    } else {
+        Parser parser(true);
+        ParsedText parsed = parse_chunks(
+            &parser,
+            {"We need to calculate factorial recursively.</think>Here is the answer."});
+        check_equal(parsed.reasoning,
+                    "We need to calculate factorial recursively.",
+                    "reasoning_opened_by_prompt:reasoning");
+        check_equal(parsed.content, "Here is the answer.",
+                    "reasoning_opened_by_prompt:content");
+    }
+}
+
+template <typename Parser>
+Parser text_parser() {
+    if constexpr (std::is_constructible_v<Parser, bool>) {
+        return Parser(false);
+    } else {
+        return Parser();
+    }
+}
+
+void test_reasoning_disabled() {
+    auto parser = text_parser<ds4cpp::DsmlParser>();
+    ParsedText parsed = parse_chunks(&parser, {"Here is the answer."});
+    check_equal(parsed.reasoning, "", "reasoning_disabled:reasoning");
+    check_equal(parsed.content, "Here is the answer.",
+                "reasoning_disabled:content");
+}
+
+void test_explicit_think_tag() {
+    auto parser = text_parser<ds4cpp::DsmlParser>();
+    ParsedText parsed = parse_chunks(
+        &parser, {"<think>reasoning</think>answer"});
+    check_equal(parsed.reasoning, "reasoning", "explicit_think_tag:reasoning");
+    check_equal(parsed.content, "answer", "explicit_think_tag:content");
+}
+
+template <typename Parser>
+void test_split_think_close_marker() {
+    if constexpr (!std::is_constructible_v<Parser, bool>) {
+        std::fprintf(stderr,
+                     "FAIL split_think_close_marker: parser cannot start in thinking state\n");
+        failures++;
+    } else {
+        Parser parser(true);
+        ParsedText parsed = parse_chunks(
+            &parser,
+            {"We need ", "to calculate ", "factorial", "</thi", "nk>",
+             "Here is ", "the answer."});
+        check_equal(parsed.reasoning, "We need to calculate factorial",
+                    "split_think_close_marker:reasoning");
+        check_equal(parsed.content, "Here is the answer.",
+                    "split_think_close_marker:content");
+    }
+}
+
+} // namespace
+
+int main() {
+    test_reasoning_opened_by_prompt<ds4cpp::DsmlParser>();
+    test_reasoning_disabled();
+    test_explicit_think_tag();
+    test_split_think_close_marker<ds4cpp::DsmlParser>();
+
+    if (failures == 0) {
+        std::fprintf(stderr, "all dsml_parser checks passed\n");
+        return 0;
+    }
+    std::fprintf(stderr, "%d check(s) failed\n", failures);
+    return 1;
+}

backend/cpp/ds4/generation_limits.h

@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+#pragma once
+
+#include <algorithm>
+
+namespace ds4cpp {
+
+inline int EffectiveGenerationLimit(int requested, int context_size,
+                                    int session_position) {
+  const int limit = requested > 0 ? requested : 256;
+  const int room = context_size - session_position;
+  if (room <= 1) return 0;
+  return std::min(limit, room - 1);
+}
+
+inline int RemainingGenerationBudget(int effective_limit, int produced) {
+  if (effective_limit <= produced) return 0;
+  return effective_limit - produced;
+}
+
+inline int SpeculativeAcceptedCapacity(int remaining, int draft_allowance,
+                                       int buffer_capacity) {
+  if (remaining <= 0 || draft_allowance < 0 || buffer_capacity <= 0) return 0;
+  return std::min({remaining, draft_allowance + 1, buffer_capacity});
+}
+
+} // namespace ds4cpp

backend/cpp/ds4/generation_limits_test.cpp

@@ -0,0 +1,92 @@
+// SPDX-License-Identifier: MIT
+
+#include "generation_limits.h"
+
+#include <cstdio>
+
+namespace {
+
+int failures = 0;
+
+void check_equal(int got, int want, const char *name) {
+  if (got == want) return;
+  std::fprintf(stderr, "FAIL %s: got %d, want %d\n", name, got, want);
+  failures++;
+}
+
+// Mutation caught: treating omitted or negative max_tokens as unlimited instead
+// of preserving DS4's legacy 256-token default.
+void test_nonpositive_uses_legacy_default_when_space_permits() {
+  check_equal(ds4cpp::EffectiveGenerationLimit(0, 4096, 100), 256,
+              "zero max_tokens uses legacy default");
+  check_equal(ds4cpp::EffectiveGenerationLimit(-1, 4096, 100), 256,
+              "negative max_tokens uses legacy default");
+}
+
+// Mutation caught: applying the legacy default without clamping it to the
+// post-prefill context room and reserved slot.
+void test_legacy_default_is_clamped_by_context() {
+  check_equal(ds4cpp::EffectiveGenerationLimit(0, 300, 100), 199,
+              "legacy default is context-clamped");
+}
+
+// Mutation caught: allowing an explicitly large request to overrun the
+// post-prefill context boundary.
+void test_large_positive_limit_is_clamped_to_context() {
+  check_equal(ds4cpp::EffectiveGenerationLimit(32768, 32768, 100), 32667,
+              "large positive is context-clamped");
+}
+
+// Mutation caught: replacing every positive request with the legacy default
+// rather than preserving a smaller configured limit.
+void test_smaller_positive_limit_is_preserved() {
+  check_equal(ds4cpp::EffectiveGenerationLimit(64, 4096, 100), 64,
+              "smaller positive is preserved");
+}
+
+// Mutation caught: consuming the final context slot instead of reserving it as
+// required by DS4's generation loop.
+void test_no_usable_room_returns_zero() {
+  check_equal(ds4cpp::EffectiveGenerationLimit(32, 100, 99), 0,
+              "one remaining context slot is not usable");
+}
+
+// Mutation caught: sending the original generation limit to a later
+// speculative cycle instead of subtracting tokens already produced.
+void test_remaining_budget_accounts_for_produced_tokens() {
+  check_equal(ds4cpp::RemainingGenerationBudget(10, 4), 6,
+              "remaining budget subtracts produced tokens");
+  check_equal(ds4cpp::RemainingGenerationBudget(10, 12), 0,
+              "remaining budget never becomes negative");
+}
+
+// Mutation caught: giving speculative evaluation capacity beyond either the
+// output budget, the draft allowance plus its first target token, or the fixed
+// accepted-token buffer.
+void test_speculative_capacity_obeys_all_bounds() {
+  check_equal(ds4cpp::SpeculativeAcceptedCapacity(3, 8, 8), 3,
+              "capacity respects remaining output budget");
+  check_equal(ds4cpp::SpeculativeAcceptedCapacity(20, 4, 8), 5,
+              "capacity includes one target token beyond draft allowance");
+  check_equal(ds4cpp::SpeculativeAcceptedCapacity(20, 8, 6), 6,
+              "capacity respects fixed buffer");
+}
+
+} // namespace
+
+int main() {
+  test_nonpositive_uses_legacy_default_when_space_permits();
+  test_legacy_default_is_clamped_by_context();
+  test_large_positive_limit_is_clamped_to_context();
+  test_smaller_positive_limit_is_preserved();
+  test_no_usable_room_returns_zero();
+  test_remaining_budget_accounts_for_produced_tokens();
+  test_speculative_capacity_obeys_all_bounds();
+
+  if (failures == 0) {
+    std::fprintf(stderr, "all generation limit checks passed\n");
+    return 0;
+  }
+  std::fprintf(stderr, "%d check(s) failed\n", failures);
+  return 1;
+}

backend/cpp/ds4/grpc-server.cpp

@@ -10,7 +10,9 @@
 
 #include "dsml_parser.h"   // populated in Task 12
 #include "dsml_renderer.h" // populated in Task 16
+#include "generation_limits.h"
 #include "kv_cache.h"      // populated in Task 17
+#include "request_lifecycle.h"
 
 extern "C" {
 #include "ds4.h"
@@ -35,6 +37,7 @@ extern "C" {
 #include <mutex>
 #include <string>
 #include <thread>
+#include <utility>
 #include <vector>
 
 using grpc::Server;
@@ -69,6 +72,21 @@ int g_route_timeout_sec = 60;
 
 std::atomic<Server *> g_server{nullptr};
 
+static bool server_context_cancelled(void *ud) {
+    return static_cast<ServerContext *>(ud)->IsCancelled();
+}
+
+static void set_session_cancel(void *target, ds4cpp::CancelCallback callback,
+                               void *userdata) noexcept {
+    ds4_session_set_cancel(static_cast<ds4_session *>(target), callback, userdata);
+}
+
+static bool request_should_continue(ds4cpp::RequestLifecycle *request,
+                                    ServerContext *context) {
+    request->ObserveContextCancellation(context->IsCancelled());
+    return request->ShouldContinue();
+}
+
 // Parse a "key:value" option string. Returns empty when no colon.
 static std::pair<std::string, std::string> split_option(const std::string &opt) {
     auto colon = opt.find(':');
@@ -238,37 +256,58 @@ static bool apply_engine_option(ds4_engine_options *opt, const std::string &key,
 
 // When acting as a distributed coordinator, block until the worker route
 // covers all layers (ds4_session_distributed_route_ready == 1) or the timeout
-// elapses. Returns an empty string on success, or an error message to return
-// to the client. No-op when not distributed.
+// elapses. No-op when not distributed.
 //
 // Takes the g_engine_mu lock by reference and RELEASES it during each poll
 // sleep. The wait can span up to g_route_timeout_sec seconds while workers
 // connect; holding g_engine_mu the whole time would block the Status/Health
 // readiness probes (they also lock g_engine_mu), making LocalAI's loader treat
 // a still-starting worker as hung.
-static std::string wait_route_ready(std::unique_lock<std::mutex> &lock) {
-    if (!g_distributed) return "";
+struct RouteWaitResult {
+    ds4cpp::RouteWaitDecision decision;
+    std::string error;
+};
+
+static RouteWaitResult wait_route_ready(std::unique_lock<std::mutex> &lock,
+                                        ServerContext *context) {
+    if (!g_distributed) return {ds4cpp::RouteWaitDecision::Ready, ""};
     char err[256] = {0};
     const int deadline_polls = g_route_timeout_sec * 10; // 100ms per poll
     for (int i = 0; i <= deadline_polls; ++i) {
         int ready = ds4_session_distributed_route_ready(g_session, err, sizeof(err));
-        if (ready == 1) return "";
-        if (ready < 0) {
-            return std::string("ds4 distributed route error: ") +
-                   (err[0] ? err : "unknown");
+        switch (ds4cpp::DecideRouteWait(ready, context->IsCancelled())) {
+        case ds4cpp::RouteWaitDecision::Ready:
+            return {ds4cpp::RouteWaitDecision::Ready, ""};
+        case ds4cpp::RouteWaitDecision::Error:
+            return {ds4cpp::RouteWaitDecision::Error,
+                    std::string("ds4 distributed route error: ") +
+                        (err[0] ? err : "unknown")};
+        case ds4cpp::RouteWaitDecision::Cancelled:
+            return {ds4cpp::RouteWaitDecision::Cancelled, ""};
+        case ds4cpp::RouteWaitDecision::Pending:
+            break;
         }
+        if (i == deadline_polls) break;
         // Release the lock while sleeping so Status/Health and other RPCs can
         // interleave during worker startup.
         lock.unlock();
         struct timespec ts = {0, 100L * 1000L * 1000L}; // 100ms
         nanosleep(&ts, nullptr);
         lock.lock();
+        if (context->IsCancelled()) {
+            return {ds4cpp::RouteWaitDecision::Cancelled, ""};
+        }
         // A concurrent Free() may have torn down the engine while we slept.
         if (!g_engine || !g_session) {
-            return "ds4: model unloaded while waiting for distributed route";
+            return {ds4cpp::RouteWaitDecision::Error,
+                    "ds4: model unloaded while waiting for distributed route"};
         }
     }
-    return "ds4 distributed route incomplete: workers not connected (layers uncovered)";
+    if (context->IsCancelled()) {
+        return {ds4cpp::RouteWaitDecision::Cancelled, ""};
+    }
+    return {ds4cpp::RouteWaitDecision::Error,
+            "ds4 distributed route incomplete: workers not connected (layers uncovered)"};
 }
 
 static void append_token_text(ds4_engine *engine, int token, std::string &out) {
@@ -341,9 +380,9 @@ static void collect_done(void *) {}
 struct StreamCtx {
     ds4_engine *engine;
     ServerWriter<backend::Reply> *writer;
+    ds4cpp::RequestLifecycle *request;
     ds4cpp::DsmlParser parser;
     int tokens;
-    bool aborted;
     // Track which tool indices we've seen TOOL_START for, so subsequent
     // ARGS deltas can elide the redundant id/name fields.
     std::vector<bool> tool_started;
@@ -351,7 +390,7 @@ struct StreamCtx {
 
 static void stream_emit(void *ud, int token) {
     auto *s = static_cast<StreamCtx *>(ud);
-    if (s->aborted) return;
+    if (!s->request->ShouldContinue()) return;
     if (token == ds4_token_eos(s->engine)) return;
     size_t len = 0;
     const char *text = ds4_token_text(s->engine, token, &len);
@@ -401,7 +440,7 @@ static void stream_emit(void *ud, int token) {
     reply.set_message(chunk);
     reply.set_tokens(1);
     if (any_field) {
-        if (!s->writer->Write(reply)) s->aborted = true;
+        s->request->ObserveStreamWrite(s->writer->Write(reply));
     }
     s->tokens++;
 }
@@ -757,21 +796,30 @@ public:
         return GStatus::OK;
     }
 
-    GStatus Predict(ServerContext *, const backend::PredictOptions *request,
+    GStatus Predict(ServerContext *context, const backend::PredictOptions *request,
                    backend::Reply *reply) override {
         std::unique_lock<std::mutex> lock(g_engine_mu);
         if (!g_engine || !g_session) {
             return GStatus(StatusCode::FAILED_PRECONDITION, "ds4: model not loaded");
         }
         if (GStatus id = check_model_identity(request); !id.ok()) return id;
-        if (std::string route_err = wait_route_ready(lock); !route_err.empty()) {
-            return GStatus(StatusCode::UNAVAILABLE, route_err);
+        RouteWaitResult route = wait_route_ready(lock, context);
+        if (route.decision == ds4cpp::RouteWaitDecision::Cancelled) {
+            return GStatus(StatusCode::CANCELLED, "ds4 request cancelled");
+        }
+        if (route.decision == ds4cpp::RouteWaitDecision::Error) {
+            return GStatus(StatusCode::UNAVAILABLE, route.error);
         }
         ds4_tokens prompt = {};
         build_prompt(g_engine, request, &prompt);
-        int n_predict = request->tokens() > 0 ? request->tokens() : 256;
 
-        CollectCtx collect = {g_engine, "", {}, reply, 0, {}, "", ""};
+        const bool think_enabled = ds4_think_mode_enabled(parse_think_mode(request));
+        const bool starts_in_thinking = think_enabled &&
+            request->usetokenizertemplate() && request->messages_size() > 0;
+        CollectCtx collect = {
+            g_engine, "", ds4cpp::DsmlParser(starts_in_thinking),
+            reply, 0, {}, "", ""};
+        ds4cpp::RequestLifecycle lifecycle;
         std::string cache_key = render_prompt_text(request);
         size_t cache_hit = maybe_load_cache(cache_key);
         (void)cache_hit; // future: skip prompt prefix if hit covers full prompt
@@ -783,15 +831,27 @@ public:
         // Either way g_session advances so the disk KV cache picks up a
         // real checkpoint after the call (see maybe_save_cache below).
         char err[256] = {0};
-        int rc = ds4_session_sync(g_session, &prompt, err, sizeof(err));
+        int rc;
+        {
+            ds4cpp::CancelCallbackScope cancel_scope(
+                g_session, set_session_cancel, server_context_cancelled, context);
+            rc = ds4_session_sync(g_session, &prompt, err, sizeof(err));
+        }
         int prompt_len = prompt.len;
         ds4_tokens_free(&prompt);
-        if (rc == 0) {
+        if (rc == DS4_SESSION_SYNC_INTERRUPTED) {
+            lifecycle.ObserveContextCancellation(true);
+        }
+        const bool generation_started = rc == 0;
+        if (generation_started) {
+            const int n_predict = ds4cpp::EffectiveGenerationLimit(
+                request->tokens(), ds4_session_ctx(g_session),
+                ds4_session_pos(g_session));
             const int eos = ds4_token_eos(g_engine);
             const int draft_max = ds4_engine_mtp_draft_tokens(g_engine);
-            const bool think_enabled = ds4_think_mode_enabled(parse_think_mode(request));
             int produced = 0;
             while (produced < n_predict) {
+                if (!request_should_continue(&lifecycle, context)) break;
                 SampleParams sp = compute_sample_params(request, collect.parser, think_enabled);
                 int first;
                 if (sp.temperature <= 0.0f) {
@@ -806,13 +866,20 @@ public:
                 if (draft_max > 0 && sp.temperature <= 0.0f) {
                     constexpr int kAcceptedMax = 8;
                     int accepted[kAcceptedMax];
-                    int cap = std::min(kAcceptedMax, draft_max + 1);
+                    const int remaining = ds4cpp::RemainingGenerationBudget(
+                        n_predict, produced);
+                    const int cap = ds4cpp::SpeculativeAcceptedCapacity(
+                        remaining, draft_max, kAcceptedMax);
                     int n = ds4_session_eval_speculative_argmax(
-                        g_session, first, draft_max, eos,
+                        g_session, first, remaining, eos,
                         accepted, cap, err, sizeof(err));
                     if (n < 0) { rc = -1; break; }
                     bool stop = false;
                     for (int j = 0; j < n; ++j) {
+                        if (!request_should_continue(&lifecycle, context)) {
+                            stop = true;
+                            break;
+                        }
                         if (accepted[j] == eos) { stop = true; break; }
                         collect_emit(&collect, accepted[j]);
                         if (++produced >= n_predict) { stop = true; break; }
@@ -821,12 +888,26 @@ public:
                 } else {
                     collect_emit(&collect, first);
                     if (++produced >= n_predict) break;
+                    if (!request_should_continue(&lifecycle, context)) break;
                     rc = ds4_session_eval(g_session, first, err, sizeof(err));
                     if (rc != 0) break;
                 }
             }
-            collect_done(&collect);
         }
+
+        request_should_continue(&lifecycle, context);
+        ds4cpp::TerminalDecision terminal = ds4cpp::ResolveTerminalDecision(
+            rc == DS4_SESSION_SYNC_INTERRUPTED, rc != 0,
+            !lifecycle.ShouldFinalize());
+        if (!terminal.should_finalize) {
+            if (terminal.cause == ds4cpp::TerminalCause::EngineError) {
+                return GStatus(StatusCode::INTERNAL,
+                              std::string("ds4 generation failed: ") + err);
+            }
+            return GStatus(StatusCode::CANCELLED,
+                          "ds4 request cancelled");
+        }
+        if (generation_started) collect_done(&collect);
         maybe_save_cache(cache_key);
 
         // Flush any buffered parser state.
@@ -834,7 +915,7 @@ public:
         collect.parser.Flush(events);
         apply_events(&collect, events);
 
-        if (rc != 0) {
+        if (terminal.cause == ds4cpp::TerminalCause::EngineError) {
             return GStatus(StatusCode::INTERNAL,
                           std::string("ds4 generation failed: ") + err);
         }
@@ -857,21 +938,30 @@ public:
         return GStatus::OK;
     }
 
-    GStatus PredictStream(ServerContext *, const backend::PredictOptions *request,
+    GStatus PredictStream(ServerContext *context, const backend::PredictOptions *request,
                          ServerWriter<backend::Reply> *writer) override {
         std::unique_lock<std::mutex> lock(g_engine_mu);
         if (!g_engine || !g_session) {
             return GStatus(StatusCode::FAILED_PRECONDITION, "ds4: model not loaded");
         }
         if (GStatus id = check_model_identity(request); !id.ok()) return id;
-        if (std::string route_err = wait_route_ready(lock); !route_err.empty()) {
-            return GStatus(StatusCode::UNAVAILABLE, route_err);
+        RouteWaitResult route = wait_route_ready(lock, context);
+        if (route.decision == ds4cpp::RouteWaitDecision::Cancelled) {
+            return GStatus(StatusCode::CANCELLED, "ds4 request cancelled");
+        }
+        if (route.decision == ds4cpp::RouteWaitDecision::Error) {
+            return GStatus(StatusCode::UNAVAILABLE, route.error);
         }
         ds4_tokens prompt = {};
         build_prompt(g_engine, request, &prompt);
-        int n_predict = request->tokens() > 0 ? request->tokens() : 256;
 
-        StreamCtx s = {g_engine, writer, {}, 0, false, {}};
+        const bool think_enabled = ds4_think_mode_enabled(parse_think_mode(request));
+        const bool starts_in_thinking = think_enabled &&
+            request->usetokenizertemplate() && request->messages_size() > 0;
+        ds4cpp::RequestLifecycle lifecycle;
+        StreamCtx s = {
+            g_engine, writer, &lifecycle,
+            ds4cpp::DsmlParser(starts_in_thinking), 0, {}};
         std::string cache_key = render_prompt_text(request);
         size_t cache_hit = maybe_load_cache(cache_key);
         (void)cache_hit;
@@ -879,14 +969,26 @@ public:
         // Manual loop on g_session - see Predict() above for the rationale.
         // MTP speculative path used when ds4_engine_mtp_draft_tokens > 0.
         char err[256] = {0};
-        int rc = ds4_session_sync(g_session, &prompt, err, sizeof(err));
+        int rc;
+        {
+            ds4cpp::CancelCallbackScope cancel_scope(
+                g_session, set_session_cancel, server_context_cancelled, context);
+            rc = ds4_session_sync(g_session, &prompt, err, sizeof(err));
+        }
         ds4_tokens_free(&prompt);
-        if (rc == 0) {
+        if (rc == DS4_SESSION_SYNC_INTERRUPTED) {
+            lifecycle.ObserveContextCancellation(true);
+        }
+        const bool generation_started = rc == 0;
+        if (generation_started) {
+            const int n_predict = ds4cpp::EffectiveGenerationLimit(
+                request->tokens(), ds4_session_ctx(g_session),
+                ds4_session_pos(g_session));
             const int eos = ds4_token_eos(g_engine);
             const int draft_max = ds4_engine_mtp_draft_tokens(g_engine);
-            const bool think_enabled = ds4_think_mode_enabled(parse_think_mode(request));
             int produced = 0;
-            while (produced < n_predict && !s.aborted) {
+            while (produced < n_predict) {
+                if (!request_should_continue(&lifecycle, context)) break;
                 SampleParams sp = compute_sample_params(request, s.parser, think_enabled);
                 int first;
                 if (sp.temperature <= 0.0f) {
@@ -900,50 +1002,77 @@ public:
                 if (draft_max > 0 && sp.temperature <= 0.0f) {
                     constexpr int kAcceptedMax = 8;
                     int accepted[kAcceptedMax];
-                    int cap = std::min(kAcceptedMax, draft_max + 1);
+                    const int remaining = ds4cpp::RemainingGenerationBudget(
+                        n_predict, produced);
+                    const int cap = ds4cpp::SpeculativeAcceptedCapacity(
+                        remaining, draft_max, kAcceptedMax);
                     int n = ds4_session_eval_speculative_argmax(
-                        g_session, first, draft_max, eos,
+                        g_session, first, remaining, eos,
                         accepted, cap, err, sizeof(err));
                     if (n < 0) { rc = -1; break; }
                     bool stop = false;
                     for (int j = 0; j < n; ++j) {
+                        if (!request_should_continue(&lifecycle, context)) {
+                            stop = true;
+                            break;
+                        }
                         if (accepted[j] == eos) { stop = true; break; }
                         stream_emit(&s, accepted[j]);
-                        if (s.aborted) { stop = true; break; }
+                        if (!lifecycle.ShouldContinue()) { stop = true; break; }
                         if (++produced >= n_predict) { stop = true; break; }
                     }
                     if (stop) break;
                 } else {
                     stream_emit(&s, first);
-                    if (s.aborted || ++produced >= n_predict) break;
+                    if (!lifecycle.ShouldContinue() || ++produced >= n_predict) break;
+                    if (!request_should_continue(&lifecycle, context)) break;
                     rc = ds4_session_eval(g_session, first, err, sizeof(err));
                     if (rc != 0) break;
                 }
             }
-            stream_done(&s);
         }
-        maybe_save_cache(cache_key);
 
-        // Flush parser state.
-        std::vector<ds4cpp::ParserEvent> events;
-        s.parser.Flush(events);
-        if (!events.empty() && !s.aborted) {
-            backend::Reply reply;
-            auto *delta = reply.add_chat_deltas();
-            for (const auto &e : events) {
-                if (e.type == ds4cpp::ParserEvent::CONTENT) {
-                    delta->set_content(delta->content() + e.text);
-                } else if (e.type == ds4cpp::ParserEvent::REASONING) {
-                    delta->set_reasoning_content(delta->reasoning_content() + e.text);
+        request_should_continue(&lifecycle, context);
+        ds4cpp::TerminalDecision terminal = ds4cpp::ResolveTerminalDecision(
+            rc == DS4_SESSION_SYNC_INTERRUPTED, rc != 0,
+            !lifecycle.ShouldFinalize());
+        terminal = ds4cpp::RunPostlude(
+            terminal,
+            [&]() {
+                ds4cpp::DsmlParser staged_parser = s.parser;
+                std::vector<ds4cpp::ParserEvent> events;
+                staged_parser.Flush(events);
+                bool write_succeeded = true;
+                if (!events.empty()) {
+                    backend::Reply reply;
+                    auto *delta = reply.add_chat_deltas();
+                    for (const auto &e : events) {
+                        if (e.type == ds4cpp::ParserEvent::CONTENT) {
+                            delta->set_content(delta->content() + e.text);
+                        } else if (e.type == ds4cpp::ParserEvent::REASONING) {
+                            delta->set_reasoning_content(
+                                delta->reasoning_content() + e.text);
+                        }
+                    }
+                    write_succeeded = s.writer->Write(reply);
                 }
-            }
-            s.writer->Write(reply);
-        }
+                lifecycle.ObserveStreamWrite(write_succeeded);
+                request_should_continue(&lifecycle, context);
+                if (!lifecycle.ShouldFinalize()) return false;
+                s.parser = std::move(staged_parser);
+                if (generation_started) stream_done(&s);
+                return true;
+            },
+            [&]() { maybe_save_cache(cache_key); });
 
-        if (rc != 0 && !s.aborted) {
+        if (terminal.cause == ds4cpp::TerminalCause::EngineError) {
             return GStatus(StatusCode::INTERNAL,
                           std::string("ds4 generation failed: ") + err);
         }
+        if (terminal.cause == ds4cpp::TerminalCause::Cancelled) {
+            return GStatus(StatusCode::CANCELLED,
+                          "ds4 request cancelled");
+        }
         return GStatus::OK;
     }
 

backend/cpp/ds4/request_lifecycle.h

@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: MIT
+#pragma once
+
+namespace ds4cpp {
+
+using CancelCallback = bool (*)(void *);
+using CancelSetter = void (*)(void *, CancelCallback, void *) noexcept;
+
+class CancelCallbackScope {
+public:
+    CancelCallbackScope(void *target, CancelSetter setter,
+                        CancelCallback callback, void *userdata) noexcept
+        : target_(target), setter_(setter) {
+        setter_(target_, callback, userdata);
+    }
+
+    ~CancelCallbackScope() noexcept {
+        setter_(target_, nullptr, nullptr);
+    }
+
+    CancelCallbackScope(const CancelCallbackScope &) = delete;
+    CancelCallbackScope &operator=(const CancelCallbackScope &) = delete;
+
+private:
+    void *target_;
+    CancelSetter setter_;
+};
+
+enum class RouteWaitDecision {
+    Pending,
+    Ready,
+    Error,
+    Cancelled,
+};
+
+inline RouteWaitDecision DecideRouteWait(int route_status, bool cancelled) {
+    if (cancelled) return RouteWaitDecision::Cancelled;
+    if (route_status > 0) return RouteWaitDecision::Ready;
+    if (route_status < 0) return RouteWaitDecision::Error;
+    return RouteWaitDecision::Pending;
+}
+
+enum class TerminalCause {
+    Success,
+    Cancelled,
+    EngineError,
+};
+
+inline TerminalCause DecideTerminalCause(bool sync_interrupted,
+                                         bool engine_error,
+                                         bool abandoned) {
+    if (sync_interrupted) return TerminalCause::Cancelled;
+    if (engine_error) return TerminalCause::EngineError;
+    if (abandoned) return TerminalCause::Cancelled;
+    return TerminalCause::Success;
+}
+
+struct TerminalDecision {
+    TerminalCause cause;
+    bool should_finalize;
+};
+
+inline TerminalDecision ResolveTerminalDecision(bool sync_interrupted,
+                                                 bool engine_error,
+                                                 bool abandoned) {
+    return {
+        DecideTerminalCause(sync_interrupted, engine_error, abandoned),
+        !sync_interrupted && !abandoned,
+    };
+}
+
+template <typename Finalize, typename Persist>
+TerminalDecision RunPostlude(TerminalDecision terminal,
+                             Finalize transactional_finalize,
+                             Persist persist) {
+    if (!terminal.should_finalize) return terminal;
+    if (!transactional_finalize()) {
+        terminal.should_finalize = false;
+        if (terminal.cause != TerminalCause::EngineError) {
+            terminal.cause = TerminalCause::Cancelled;
+        }
+        return terminal;
+    }
+    persist();
+    return terminal;
+}
+
+class RequestLifecycle {
+public:
+    void ObserveContextCancellation(bool cancelled) {
+        context_cancelled_ = context_cancelled_ || cancelled;
+    }
+
+    void ObserveStreamWrite(bool succeeded) {
+        stream_write_aborted_ = stream_write_aborted_ || !succeeded;
+    }
+
+    bool ShouldContinue() const {
+        return !context_cancelled_ && !stream_write_aborted_;
+    }
+
+    bool ShouldFinalize() const {
+        return ShouldContinue();
+    }
+
+private:
+    bool context_cancelled_ = false;
+    bool stream_write_aborted_ = false;
+};
+
+} // namespace ds4cpp

backend/cpp/ds4/request_lifecycle_test.cpp

@@ -0,0 +1,414 @@
+// SPDX-License-Identifier: MIT
+// Standalone regression tests for DS4 request cancellation policy.
+
+#include "request_lifecycle.h"
+
+#include <cstdio>
+
+namespace {
+
+int failures = 0;
+
+struct FakeCancelTarget {
+    ds4cpp::CancelCallback callback = nullptr;
+    void *userdata = nullptr;
+    int installs = 0;
+    int clears = 0;
+};
+
+struct PostludeCounts {
+    int finalize_attempts = 0;
+    int finalize_commits = 0;
+    int cache_persists = 0;
+    bool cache_followed_commit = true;
+};
+
+ds4cpp::TerminalDecision run_fake_postlude(
+        ds4cpp::TerminalDecision terminal, bool finalize_succeeds,
+        PostludeCounts *counts) {
+    return ds4cpp::RunPostlude(
+        terminal,
+        [=]() {
+            counts->finalize_attempts++;
+            if (!finalize_succeeds) return false;
+            counts->finalize_commits++;
+            return true;
+        },
+        [=]() {
+            counts->cache_followed_commit = counts->finalize_commits == 1;
+            counts->cache_persists++;
+        });
+}
+
+bool fake_cancel(void *) {
+    return false;
+}
+
+void fake_set_cancel(void *target, ds4cpp::CancelCallback callback,
+                     void *userdata) noexcept {
+    auto *fake = static_cast<FakeCancelTarget *>(target);
+    fake->callback = callback;
+    fake->userdata = userdata;
+    if (callback) {
+        fake->installs++;
+    } else {
+        fake->clears++;
+    }
+}
+
+void check(bool condition, const char *name) {
+    if (condition) return;
+    std::fprintf(stderr, "FAIL %s\n", name);
+    failures++;
+}
+
+// Production mutation caught: treating an active request as abandoned would
+// skip its parser finalization and cache save.
+void test_active_request_continues_and_finalizes() {
+    ds4cpp::RequestLifecycle request;
+
+    check(request.ShouldContinue(), "active:continue");
+    check(request.ShouldFinalize(), "active:finalize");
+}
+
+// Production mutation caught: omitting the ServerContext cancellation branch
+// would continue decoding and finalize a partial response.
+void test_context_cancellation_stops_without_finalizing() {
+    ds4cpp::RequestLifecycle request;
+
+    request.ObserveContextCancellation(true);
+
+    check(!request.ShouldContinue(), "context_cancelled:stop");
+    check(!request.ShouldFinalize(), "context_cancelled:no_finalize");
+}
+
+// Production mutation caught: ignoring ServerWriter::Write failure would keep
+// streaming and finalize a response whose client has gone away.
+void test_stream_write_abort_stops_without_finalizing() {
+    ds4cpp::RequestLifecycle request;
+
+    request.ObserveStreamWrite(false);
+
+    check(!request.ShouldContinue(), "write_abort:stop");
+    check(!request.ShouldFinalize(), "write_abort:no_finalize");
+}
+
+// Production mutation caught: combining cancellation and write failure with
+// AND would fail to stop when either signal occurs on its own.
+void test_cancellation_and_write_abort_are_independent_or_conditions() {
+    ds4cpp::RequestLifecycle cancelled;
+    cancelled.ObserveContextCancellation(true);
+    cancelled.ObserveStreamWrite(true);
+
+    ds4cpp::RequestLifecycle write_aborted;
+    write_aborted.ObserveContextCancellation(false);
+    write_aborted.ObserveStreamWrite(false);
+
+    check(!cancelled.ShouldContinue(), "or:context_only");
+    check(!write_aborted.ShouldContinue(), "or:write_only");
+}
+
+// Production mutation caught: treating an incomplete distributed route as an
+// error would return before workers have time to connect.
+void test_route_wait_pending() {
+    check(ds4cpp::DecideRouteWait(0, false) ==
+              ds4cpp::RouteWaitDecision::Pending,
+          "route_wait:pending");
+}
+
+// Production mutation caught: failing to recognize a complete route would
+// keep a ready inference request in the polling loop.
+void test_route_wait_ready() {
+    check(ds4cpp::DecideRouteWait(1, false) ==
+              ds4cpp::RouteWaitDecision::Ready,
+          "route_wait:ready");
+}
+
+// Production mutation caught: ignoring a route probe error would poll until a
+// misleading timeout instead of returning UNAVAILABLE promptly.
+void test_route_wait_error() {
+    check(ds4cpp::DecideRouteWait(-1, false) ==
+              ds4cpp::RouteWaitDecision::Error,
+          "route_wait:error");
+}
+
+// Production mutation caught: omitting cancellation from route waiting would
+// leave an abandoned request blocked until the distributed timeout.
+void test_route_wait_cancellation() {
+    check(ds4cpp::DecideRouteWait(0, true) ==
+              ds4cpp::RouteWaitDecision::Cancelled,
+          "route_wait:cancelled");
+}
+
+// Production mutation caught: checking route errors before cancellation would
+// report UNAVAILABLE for a request the client already abandoned.
+void test_route_wait_cancellation_precedes_error() {
+    check(ds4cpp::DecideRouteWait(-1, true) ==
+              ds4cpp::RouteWaitDecision::Cancelled,
+          "route_wait:cancellation_precedence");
+}
+
+// Production mutation caught: classifying a successful active request as a
+// terminal failure would suppress its normal response finalization.
+void test_terminal_success() {
+    check(ds4cpp::DecideTerminalCause(false, false, false) ==
+              ds4cpp::TerminalCause::Success,
+          "terminal:success");
+}
+
+// Production mutation caught: treating DS4's cooperative sync interruption
+// as an ordinary engine error would return INTERNAL instead of CANCELLED.
+void test_terminal_sync_interruption_is_cancelled() {
+    check(ds4cpp::DecideTerminalCause(true, true, true) ==
+              ds4cpp::TerminalCause::Cancelled,
+          "terminal:sync_interrupted");
+}
+
+// Production mutation caught: treating every nonzero engine result as client
+// abandonment would hide genuine DS4 failures behind CANCELLED.
+void test_terminal_engine_error() {
+    check(ds4cpp::DecideTerminalCause(false, true, false) ==
+              ds4cpp::TerminalCause::EngineError,
+          "terminal:engine_error");
+}
+
+// Production mutation caught: ignoring an rc==0 context cancellation would
+// finalize and cache an abandoned request.
+void test_terminal_context_abandonment() {
+    ds4cpp::RequestLifecycle request;
+    request.ObserveContextCancellation(true);
+
+    check(ds4cpp::DecideTerminalCause(
+              false, false, !request.ShouldFinalize()) ==
+              ds4cpp::TerminalCause::Cancelled,
+          "terminal:context_abandonment");
+}
+
+// Production mutation caught: ignoring an rc==0 stream write failure would
+// finalize and cache an abandoned streaming request.
+void test_terminal_write_abandonment() {
+    ds4cpp::RequestLifecycle request;
+    request.ObserveStreamWrite(false);
+
+    check(ds4cpp::DecideTerminalCause(
+              false, false, !request.ShouldFinalize()) ==
+              ds4cpp::TerminalCause::Cancelled,
+          "terminal:write_abandonment");
+}
+
+// Production mutation caught: checking late cancellation or write failure
+// before a determined ordinary DS4 error would replace INTERNAL with CANCELLED.
+void test_terminal_engine_error_precedes_late_abandonment() {
+    ds4cpp::RequestLifecycle cancelled;
+    cancelled.ObserveContextCancellation(true);
+    ds4cpp::RequestLifecycle write_aborted;
+    write_aborted.ObserveStreamWrite(false);
+
+    check(ds4cpp::DecideTerminalCause(
+              false, true, !cancelled.ShouldFinalize()) ==
+              ds4cpp::TerminalCause::EngineError,
+          "terminal:engine_error_precedes_cancellation");
+    check(ds4cpp::DecideTerminalCause(
+              false, true, !write_aborted.ShouldFinalize()) ==
+              ds4cpp::TerminalCause::EngineError,
+          "terminal:engine_error_precedes_write_abort");
+}
+
+// Production mutation caught: using status precedence alone to gate side
+// effects would finalize and persist an engine-error request abandoned later.
+void test_abandoned_engine_error_keeps_internal_without_finalizing() {
+    ds4cpp::RequestLifecycle request;
+    request.ObserveContextCancellation(true);
+
+    ds4cpp::TerminalDecision terminal = ds4cpp::ResolveTerminalDecision(
+        false, true, !request.ShouldFinalize());
+
+    check(terminal.cause == ds4cpp::TerminalCause::EngineError,
+          "terminal_decision:abandoned_engine_error_status");
+    check(!terminal.should_finalize,
+          "terminal_decision:abandoned_engine_error_no_finalize");
+}
+
+// Production mutation caught: suppressing side effects for every engine error
+// would change the existing finalization and cache behavior of active failures.
+void test_active_engine_error_still_finalizes() {
+    ds4cpp::RequestLifecycle request;
+
+    ds4cpp::TerminalDecision terminal = ds4cpp::ResolveTerminalDecision(
+        false, true, !request.ShouldFinalize());
+
+    check(terminal.cause == ds4cpp::TerminalCause::EngineError,
+          "terminal_decision:active_engine_error_status");
+    check(terminal.should_finalize,
+          "terminal_decision:active_engine_error_finalize");
+}
+
+// Production mutation caught: persisting before committed finalization would
+// cache a state whose final buffered stream reply was never completed.
+void test_postlude_active_success_commits_then_persists() {
+    PostludeCounts counts;
+
+    ds4cpp::TerminalDecision terminal = run_fake_postlude(
+        {ds4cpp::TerminalCause::Success, true}, true, &counts);
+
+    check(terminal.cause == ds4cpp::TerminalCause::Success,
+          "postlude:success_outcome");
+    check(terminal.should_finalize, "postlude:success_committed");
+    check(counts.finalize_attempts == 1, "postlude:success_attempts");
+    check(counts.finalize_commits == 1, "postlude:success_commits");
+    check(counts.cache_persists == 1, "postlude:success_cache");
+    check(counts.cache_followed_commit, "postlude:success_cache_order");
+}
+
+// Production mutation caught: starting the postlude for an already-cancelled
+// request would flush buffered parser state or persist an abandoned session.
+void test_postlude_cancellation_skips_all_side_effects() {
+    PostludeCounts counts;
+
+    ds4cpp::TerminalDecision terminal = run_fake_postlude(
+        {ds4cpp::TerminalCause::Cancelled, false}, true, &counts);
+
+    check(terminal.cause == ds4cpp::TerminalCause::Cancelled,
+          "postlude:cancelled_outcome");
+    check(counts.finalize_attempts == 0, "postlude:cancelled_attempts");
+    check(counts.finalize_commits == 0, "postlude:cancelled_commits");
+    check(counts.cache_persists == 0, "postlude:cancelled_cache");
+}
+
+// Production mutation caught: committing the live parser or cache after a
+// failed final Write would publish an abandoned streaming postlude.
+void test_postlude_finalize_failure_cancels_without_commit_or_cache() {
+    PostludeCounts counts;
+
+    ds4cpp::TerminalDecision terminal = run_fake_postlude(
+        {ds4cpp::TerminalCause::Success, true}, false, &counts);
+
+    check(terminal.cause == ds4cpp::TerminalCause::Cancelled,
+          "postlude:write_failure_outcome");
+    check(!terminal.should_finalize, "postlude:write_failure_not_committed");
+    check(counts.finalize_attempts == 1, "postlude:write_failure_attempts");
+    check(counts.finalize_commits == 0, "postlude:write_failure_commits");
+    check(counts.cache_persists == 0, "postlude:write_failure_cache");
+}
+
+// Production mutation caught: skipping the postlude for every engine error
+// would change active internal-error finalization and cache behavior.
+void test_postlude_active_engine_error_finalizes_and_persists() {
+    PostludeCounts counts;
+
+    ds4cpp::TerminalDecision terminal = run_fake_postlude(
+        {ds4cpp::TerminalCause::EngineError, true}, true, &counts);
+
+    check(terminal.cause == ds4cpp::TerminalCause::EngineError,
+          "postlude:engine_error_outcome");
+    check(counts.finalize_attempts == 1, "postlude:engine_error_attempts");
+    check(counts.finalize_commits == 1, "postlude:engine_error_commits");
+    check(counts.cache_persists == 1, "postlude:engine_error_cache");
+    check(counts.cache_followed_commit, "postlude:engine_error_cache_order");
+}
+
+// Production mutation caught: replacing every failed transactional finalize
+// with cancellation would hide an already-determined engine error.
+void test_postlude_engine_error_finalize_failure_preserves_internal() {
+    PostludeCounts counts;
+
+    ds4cpp::TerminalDecision terminal = run_fake_postlude(
+        {ds4cpp::TerminalCause::EngineError, true}, false, &counts);
+
+    check(terminal.cause == ds4cpp::TerminalCause::EngineError,
+          "postlude:engine_error_write_failure_outcome");
+    check(!terminal.should_finalize,
+          "postlude:engine_error_write_failure_not_committed");
+    check(counts.finalize_attempts == 1,
+          "postlude:engine_error_write_failure_attempts");
+    check(counts.finalize_commits == 0,
+          "postlude:engine_error_write_failure_commits");
+    check(counts.cache_persists == 0,
+          "postlude:engine_error_write_failure_cache");
+}
+
+// Production mutation caught: status precedence must not grant side-effect
+// permission to an engine-error request that was also abandoned.
+void test_postlude_abandoned_engine_error_skips_all_side_effects() {
+    PostludeCounts counts;
+
+    ds4cpp::TerminalDecision terminal = run_fake_postlude(
+        {ds4cpp::TerminalCause::EngineError, false}, true, &counts);
+
+    check(terminal.cause == ds4cpp::TerminalCause::EngineError,
+          "postlude:abandoned_engine_error_outcome");
+    check(counts.finalize_attempts == 0,
+          "postlude:abandoned_engine_error_attempts");
+    check(counts.finalize_commits == 0,
+          "postlude:abandoned_engine_error_commits");
+    check(counts.cache_persists == 0,
+          "postlude:abandoned_engine_error_cache");
+}
+
+// Production mutation caught: failing to install the request callback would
+// make DS4 prompt synchronization unable to observe client cancellation.
+void test_cancel_callback_scope_installs_callback() {
+    FakeCancelTarget target;
+    int request_context = 42;
+
+    {
+        ds4cpp::CancelCallbackScope scope(
+            &target, fake_set_cancel, fake_cancel, &request_context);
+        check(target.callback == fake_cancel, "cancel_scope:callback_installed");
+        check(target.userdata == &request_context, "cancel_scope:userdata_installed");
+        check(target.installs == 1, "cancel_scope:installed_once");
+    }
+}
+
+// Production mutation caught: failing to clear the callback at every scope
+// exit would leave DS4 pointing at a destroyed stack-owned ServerContext.
+void test_cancel_callback_scope_clears_callback() {
+    FakeCancelTarget target;
+    int request_context = 42;
+
+    {
+        ds4cpp::CancelCallbackScope scope(
+            &target, fake_set_cancel, fake_cancel, &request_context);
+    }
+
+    check(target.callback == nullptr, "cancel_scope:callback_cleared");
+    check(target.userdata == nullptr, "cancel_scope:userdata_cleared");
+    check(target.clears == 1, "cancel_scope:cleared_once");
+}
+
+} // namespace
+
+int main() {
+    test_active_request_continues_and_finalizes();
+    test_context_cancellation_stops_without_finalizing();
+    test_stream_write_abort_stops_without_finalizing();
+    test_cancellation_and_write_abort_are_independent_or_conditions();
+    test_route_wait_pending();
+    test_route_wait_ready();
+    test_route_wait_error();
+    test_route_wait_cancellation();
+    test_route_wait_cancellation_precedes_error();
+    test_terminal_success();
+    test_terminal_sync_interruption_is_cancelled();
+    test_terminal_engine_error();
+    test_terminal_context_abandonment();
+    test_terminal_write_abandonment();
+    test_terminal_engine_error_precedes_late_abandonment();
+    test_abandoned_engine_error_keeps_internal_without_finalizing();
+    test_active_engine_error_still_finalizes();
+    test_postlude_active_success_commits_then_persists();
+    test_postlude_cancellation_skips_all_side_effects();
+    test_postlude_finalize_failure_cancels_without_commit_or_cache();
+    test_postlude_active_engine_error_finalizes_and_persists();
+    test_postlude_engine_error_finalize_failure_preserves_internal();
+    test_postlude_abandoned_engine_error_skips_all_side_effects();
+    test_cancel_callback_scope_installs_callback();
+    test_cancel_callback_scope_clears_callback();
+
+    if (failures == 0) {
+        std::fprintf(stderr, "all request_lifecycle checks passed\n");
+        return 0;
+    }
+    std::fprintf(stderr, "%d check(s) failed\n", failures);
+    return 1;
+}

backend/cpp/ik-llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-IK_LLAMA_VERSION?=cf1aa57e1a0fabfd015831718fc99d1aec01ada5
+IK_LLAMA_VERSION?=1a2a8604a6c6c6413c06bf9adfc2f64329af4366
 LLAMA_REPO?=https://github.com/ikawrakow/ik_llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/ik-llama-cpp/grpc-server.cpp

@@ -2565,6 +2565,7 @@ public:
     grpc::Status Embedding(ServerContext* context, const backend::PredictOptions* request, backend::EmbeddingResult* embeddingResult) {
         auto identity = checkModelIdentity(request);
         if (!identity.ok()) return identity;
+        embeddingResult->set_layout(backend::EMBEDDING_LAYOUT_FINAL);
         json data = parse_options(false, request, llama);
         const int task_id = llama.queue_tasks.get_new_id();
         llama.queue_results.add_waiting_task_id(task_id);

backend/cpp/llama-cpp/CMakeLists.txt

@@ -115,4 +115,14 @@ if(LLAMA_GRPC_BUILD_TESTS)
     target_include_directories(passthrough_options_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
     target_compile_features(passthrough_options_test PRIVATE cxx_std_17)
     add_test(NAME passthrough_options_test COMMAND passthrough_options_test)
+
+    add_executable(tts_request_options_test tts_request_options_test.cpp tts_request_options.h)
+    target_include_directories(tts_request_options_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+    target_compile_features(tts_request_options_test PRIVATE cxx_std_17)
+    add_test(NAME tts_request_options_test COMMAND tts_request_options_test)
+
+    add_executable(thread_params_test thread_params_test.cpp thread_params.h)
+    target_include_directories(thread_params_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+    target_compile_features(thread_params_test PRIVATE cxx_std_17)
+    add_test(NAME thread_params_test COMMAND thread_params_test)
 endif()

backend/cpp/llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=221f0f6356efe2260023208365705ec5d5a7c8f5
+LLAMA_VERSION?=67672dc5b76f8bc17785a19d3dc6d1463fc2902c
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/llama-cpp/disable-tts-task.sh

@@ -0,0 +1,43 @@
+#!/bin/bash
+# Mark a copied gRPC server as targeting a llama.cpp fork that does not carry
+# LocalAI's SERVER_TASK_TYPE_TTS patch. The RPCs remain present in the shared
+# protobuf service, but respond with UNIMPLEMENTED instead of referencing
+# server task types and mtmd gen-audio APIs absent from those forks.
+
+set -euo pipefail
+
+if [[ $# -ne 1 ]]; then
+    echo "usage: $0 <grpc-server.cpp>" >&2
+    exit 2
+fi
+
+SRC=$1
+
+if [[ ! -f "$SRC" ]]; then
+    echo "grpc-server.cpp not found at $SRC" >&2
+    exit 2
+fi
+
+if grep -q '^#define LOCALAI_LLAMA_CPP_NO_TTS_TASK' "$SRC"; then
+    echo "==> $SRC already disables the LocalAI TTS task, skipping"
+    exit 0
+fi
+
+awk '
+    !done && /^#include/ {
+        print "#define LOCALAI_LLAMA_CPP_NO_TTS_TASK 1"
+        print "// ^ injected by disable-tts-task.sh for an unpatched llama.cpp fork"
+        print ""
+        done = 1
+    }
+    { print }
+    END {
+        if (!done) {
+            print "disable-tts-task.sh: no #include anchor found" > "/dev/stderr"
+            exit 1
+        }
+    }
+' "$SRC" > "$SRC.tmp"
+mv "$SRC.tmp" "$SRC"
+
+echo "==> LocalAI TTS task disabled in $SRC"

backend/cpp/llama-cpp/grpc-server.cpp

@@ -53,8 +53,10 @@
 #include "arg.h"
 #include "chat-auto-parser.h"
 #include "llama_compat.h" // fork-skew switches, generated by prepare.sh
+#include "thread_params.h"
 #include "message_content.h"
 #include "passthrough_options.h"
+#include "tts_request_options.h"
 #include <getopt.h>
 #include <grpcpp/ext/proto_server_reflection_plugin.h>
 #include <grpcpp/grpcpp.h>
@@ -65,6 +67,7 @@
 #include <atomic>
 #include <cmath>
 #include <cstdlib>
+#include <cstring>
 #include <fstream>
 #include <iterator>
 #include <list>
@@ -85,6 +88,12 @@ using grpc::ServerBuilder;
 using grpc::ServerContext;
 using grpc::Status;
 
+#if LOCALAI_HAS_MTMD_INIT_OPT
+#define LOCALAI_MTMD_INIT_OPT_ARG(value) , value
+#else
+#define LOCALAI_MTMD_INIT_OPT_ARG(value)
+#endif
+
 // gRPC bearer token auth for distributed mode.
 // Reads LOCALAI_GRPC_AUTH_TOKEN from the environment. When set, rejects
 // requests without a matching "authorization: Bearer <token>" metadata header.
@@ -233,7 +242,15 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, const
     data["typical_p"] = predict->typicalp();
     data["temperature"] = predict->temperature();
     data["repeat_last_n"] = predict->repeat();
-    data["repeat_penalty"] = predict->penalty();
+    // PredictOptions.Penalty is a bare proto float, so a caller that names no
+    // repetition penalty sends 0 rather than omitting the field. Since
+    // llama.cpp 9de0fcf2b, common_sampler_init() rejects a non-positive
+    // penalty_repeat outright (it would divide logits by zero), which turned
+    // every such request into "Failed to initialize samplers". Treat 0 as
+    // "unset" and leave llama.cpp's own neutral default in place.
+    if (predict->penalty() > 0.0f) {
+        data["repeat_penalty"] = predict->penalty();
+    }
     data["frequency_penalty"] = predict->frequencypenalty();
     data["presence_penalty"] = predict->presencepenalty();
     data["mirostat"] = predict->mirostat();
@@ -283,7 +300,7 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, const
             } else {
                 SRV_WRN("[TOOLS DEBUG] parse_options: Parsed tools JSON is not an array: %s\n", tools_json.dump().c_str());
             }
-        } catch (const json::parse_error& e) {
+        } catch (const common_json_error& e) {
             SRV_WRN("Failed to parse tools JSON from proto: %s\n", e.what());
             SRV_WRN("[TOOLS DEBUG] parse_options: Tools string that failed to parse: %s\n", predict->tools().c_str());
         }
@@ -313,7 +330,7 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, const
                 SRV_DBG("[TOOLS DEBUG] Received tool_choice object from Go layer: %s\n", tool_choice_json.dump().c_str());
             }
             SRV_INF("Extracted tool_choice from proto: %s\n", predict->toolchoice().c_str());
-        } catch (const json::parse_error& e) {
+        } catch (const common_json_error& e) {
             // If parsing fails, treat as string
             data["tool_choice"] = predict->toolchoice();
             SRV_INF("Extracted tool_choice as string: %s\n", predict->toolchoice().c_str());
@@ -342,7 +359,7 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, const
             // Add to data - llama.cpp server expects it as an object (map)
             data["logit_bias"] = logit_bias_json;
             SRV_INF("Using logit_bias: %s\n", predict->logitbias().c_str());
-        } catch (const json::parse_error& e) {
+        } catch (const common_json_error& e) {
             SRV_ERR("Failed to parse logit_bias JSON from proto: %s\n", e.what());
         }
     }
@@ -387,7 +404,10 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, const
             });
     }
 
-    data["stop"] = predict->stopprompts();
+    data["stop"] = json::array();
+    for (const auto & stop : predict->stopprompts()) {
+        data["stop"].push_back(stop);
+    }
     // data["n_probs"] = predict->nprobs();
     //TODO: images,
 
@@ -1105,14 +1125,16 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
                 try {
                     int n = std::stoi(optval_str);
                     if (n < 0) n = 0;
-                    // Keep override-name storage alive for the lifetime of the params struct
-                    // (mirrors upstream arg.cpp behavior with a function-local static).
+#if LOCALAI_HAS_N_CPU_FFN_HELPER
+                    llm_add_n_cpu_ffn_overrides(n, LLM_FFN_EXPS_REGEX, params.speculative.draft.tensor_buft_overrides);
+#else
                     static std::list<std::string> buft_overrides_draft;
                     for (int i = 0; i < n; ++i) {
                         buft_overrides_draft.push_back(llm_ffn_exps_block_regex(i));
                         params.speculative.draft.tensor_buft_overrides.push_back(
                             {buft_overrides_draft.back().c_str(), ggml_backend_cpu_buffer_type()});
                     }
+#endif
                 } catch (...) {}
             }
 
@@ -1130,14 +1152,16 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
                 try {
                     int n = std::stoi(optval_str);
                     if (n < 0) n = 0;
-                    // Keep override-name storage alive for the lifetime of the
-                    // params struct (mirrors upstream arg.cpp's function-local static).
+#if LOCALAI_HAS_N_CPU_FFN_HELPER
+                    llm_add_n_cpu_ffn_overrides(n, LLM_FFN_EXPS_REGEX, params.tensor_buft_overrides);
+#else
                     static std::list<std::string> buft_overrides_main;
                     for (int i = 0; i < n; ++i) {
                         buft_overrides_main.push_back(llm_ffn_exps_block_regex(i));
                         params.tensor_buft_overrides.push_back(
                             {buft_overrides_main.back().c_str(), ggml_backend_cpu_buffer_type()});
                     }
+#endif
                 } catch (...) {}
             }
 
@@ -1402,6 +1426,12 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
             passthrough_draft_gpu_layers);
     }
 
+    // The library initializer now creates both threadpools before the server
+    // can apply llama_context's fallback for the -1 batch-thread sentinel.
+    params.cpuparams_batch.n_threads = llama_grpc::resolve_batch_threads(
+        params.cpuparams_batch.n_threads,
+        params.cpuparams.n_threads);
+
 #ifndef LOCALAI_LLAMA_CPP_NO_SCORE_TASK
     // Score-task suffix forking: reserve seq ids (and recurrent-state cells)
     // beyond the slots so one scoring call decodes all candidate tails in a
@@ -1445,6 +1475,26 @@ static void params_parse(server_context& /*ctx_server*/, const backend::ModelOpt
 }
 
 
+#ifndef LOCALAI_LLAMA_CPP_NO_TTS_TASK
+// MTMD_HELPER_GEN_AUDIO_OUTTYPE_PCM hands back raw float32 samples, but the
+// WAV header core/backend/tts.go builds around the streamed chunks announces
+// 16-bit samples, so the wire has to carry s16 or the client decodes floats as
+// integers and hears noise. The scaling matches write_wav16() in
+// tools/mtmd/mtmd-helper-gen.cpp, which is what the non-streaming path writes.
+static std::string tts_pcm_f32_to_s16(const std::string & samples) {
+    const size_t n = samples.size() / sizeof(float);
+    std::string out;
+    out.resize(n * sizeof(int16_t));
+    for (size_t i = 0; i < n; i++) {
+        float v = 0.0f;
+        std::memcpy(&v, samples.data() + i * sizeof(float), sizeof(float));
+        const int16_t s = (int16_t) (std::max(-1.0f, std::min(1.0f, v)) * 32767.0f);
+        std::memcpy(&out[i * sizeof(int16_t)], &s, sizeof(int16_t));
+    }
+    return out;
+}
+#endif
+
 // GRPC Server start
 class BackendServiceImpl final : public backend::Backend::Service {
 private:
@@ -1758,7 +1808,7 @@ public:
                         for (int j = 0; j < request->audios_size(); j++) rin.audios.push_back(request->audios(j));
                         for (int j = 0; j < request->videos_size(); j++) rin.videos.push_back(request->videos(j));
                     }
-                    messages_json.push_back(llama_grpc::build_reconstructed_message(rin));
+                    messages_json.push_back(json::parse(llama_grpc::build_reconstructed_message(rin).dump()));
                 }
 
                 // Final safety check: Ensure no message has null content (Jinja templates require strings)
@@ -1951,7 +2001,7 @@ public:
                             if (!body_json.contains("chat_template_kwargs")) {
                                 body_json["chat_template_kwargs"] = json::object();
                             }
-                            for (auto& el : ctk.items()) {
+                            for (auto el : ctk.items()) {
                                 body_json["chat_template_kwargs"][el.key()] = el.value();
                             }
                         }
@@ -2037,30 +2087,27 @@ public:
             // If not using chat templates, extract files from image_data/audio_data fields
             // (If using chat templates, files were already extracted by oaicompat_chat_params_parse)
             if (!request->usetokenizertemplate() || request->messages_size() == 0 || ctx_server.impl->chat_params.tmpls == nullptr) {
-                const auto &images_data = data.find("image_data");
-                if (images_data != data.end() && images_data->is_array())
+                if (data.contains("image_data") && data.at("image_data").is_array())
                 {
-                    for (const auto &img : *images_data)
+                    for (const auto &img : data.at("image_data"))
                     {
                         auto decoded_data = base64_decode(img["data"].get<std::string>());
                         files.push_back(decoded_data);
                     }
                 }
 
-                const auto &audio_data = data.find("audio_data");
-                if (audio_data != data.end() && audio_data->is_array())
+                if (data.contains("audio_data") && data.at("audio_data").is_array())
                 {
-                    for (const auto &audio : *audio_data)
+                    for (const auto &audio : data.at("audio_data"))
                     {
                         auto decoded_data = base64_decode(audio["data"].get<std::string>());
                         files.push_back(decoded_data);
                     }
                 }
 
-                const auto &video_data = data.find("video_data");
-                if (video_data != data.end() && video_data->is_array())
+                if (data.contains("video_data") && data.at("video_data").is_array())
                 {
-                    for (const auto &video : *video_data)
+                    for (const auto &video : data.at("video_data"))
                     {
                         auto decoded_data = base64_decode(video["data"].get<std::string>());
                         files.push_back(decoded_data);
@@ -2074,10 +2121,10 @@ public:
             std::vector<server_tokens> inputs;
             if (has_mtmd) {
                 // multimodal
-                inputs.push_back(process_mtmd_prompt(ctx_server.impl->mctx, prompt_str, files));
+                inputs.push_back(process_mtmd_prompt(ctx_server.impl->mctx, prompt_str, files LOCALAI_MTMD_INIT_OPT_ARG(ctx_server.impl->init_opt)));
             } else {
                  // Everything else, including multimodal completions.
-                inputs = tokenize_input_prompts(ctx_server.impl->vocab, ctx_server.impl->mctx, prompt_str, true, true);
+                inputs = tokenize_input_prompts(ctx_server.impl->vocab, ctx_server.impl->mctx, prompt_str, true, true LOCALAI_MTMD_INIT_OPT_ARG(ctx_server.impl->init_opt));
             }
 
             tasks.reserve(inputs.size());
@@ -2089,15 +2136,23 @@ public:
 
                 task.tokens    = std::move(inputs[i]);
 #ifdef LOCALAI_HAS_SERVER_SCHEMA
+                // The schema evaluator no longer takes the per-slot n_ctx: upstream
+                // dropped the parameter and server-schema stopped consulting n_ctx at
+                // all, leaving the context bound to the slot. Forks that predate the
+                // server-schema split still expect it, so only this branch loses it.
                 task.params           = server_schema::eval_llama_cmpl_schema(
+                        ctx_server.impl->vocab,
+                        params_base,
+                        ctx_server.get_meta().logit_bias_eog,
+                        data);
 #else
                 task.params           = server_task::params_from_json_cmpl(
-#endif
                         ctx_server.impl->vocab,
                         params_base,
                         ctx_server.get_meta().slot_n_ctx,
                         ctx_server.get_meta().logit_bias_eog,
                         data);
+#endif
                 task.id_slot = json_value(data, "id_slot", -1);
 
                 // OAI-compat: enable autoparser (PEG-based chat parsing) so that
@@ -2325,7 +2380,7 @@ public:
                         for (int j = 0; j < request->audios_size(); j++) rin.audios.push_back(request->audios(j));
                         for (int j = 0; j < request->videos_size(); j++) rin.videos.push_back(request->videos(j));
                     }
-                    messages_json.push_back(llama_grpc::build_reconstructed_message(rin));
+                    messages_json.push_back(json::parse(llama_grpc::build_reconstructed_message(rin).dump()));
                 }
 
                 // Final safety check: Ensure no message has null content (Jinja templates require strings)
@@ -2518,7 +2573,7 @@ public:
                             if (!body_json.contains("chat_template_kwargs")) {
                                 body_json["chat_template_kwargs"] = json::object();
                             }
-                            for (auto& el : ctk.items()) {
+                            for (auto el : ctk.items()) {
                                 body_json["chat_template_kwargs"][el.key()] = el.value();
                             }
                         }
@@ -2604,11 +2659,10 @@ public:
             // If not using chat templates, extract files from image_data/audio_data fields
             // (If using chat templates, files were already extracted by oaicompat_chat_params_parse)
             if (!request->usetokenizertemplate() || request->messages_size() == 0 || ctx_server.impl->chat_params.tmpls == nullptr) {
-                const auto &images_data = data.find("image_data");
-                if (images_data != data.end() && images_data->is_array())
+                if (data.contains("image_data") && data.at("image_data").is_array())
                 {
-                    std::cout << "[PREDICT] Processing " << images_data->size() << " images" << std::endl;
-                    for (const auto &img : *images_data)
+                    std::cout << "[PREDICT] Processing " << data.at("image_data").size() << " images" << std::endl;
+                    for (const auto &img : data.at("image_data"))
                     {
                         std::cout << "[PREDICT] Processing image" << std::endl;
                         auto decoded_data = base64_decode(img["data"].get<std::string>());
@@ -2616,20 +2670,18 @@ public:
                     }
                 }
 
-                const auto &audio_data = data.find("audio_data");
-                if (audio_data != data.end() && audio_data->is_array())
+                if (data.contains("audio_data") && data.at("audio_data").is_array())
                 {
-                    for (const auto &audio : *audio_data)
+                    for (const auto &audio : data.at("audio_data"))
                     {
                         auto decoded_data = base64_decode(audio["data"].get<std::string>());
                         files.push_back(decoded_data);
                     }
                 }
 
-                const auto &video_data = data.find("video_data");
-                if (video_data != data.end() && video_data->is_array())
+                if (data.contains("video_data") && data.at("video_data").is_array())
                 {
-                    for (const auto &video : *video_data)
+                    for (const auto &video : data.at("video_data"))
                     {
                         auto decoded_data = base64_decode(video["data"].get<std::string>());
                         files.push_back(decoded_data);
@@ -2644,10 +2696,10 @@ public:
             std::vector<server_tokens> inputs;
             if (has_mtmd) {
                 // multimodal
-                inputs.push_back(process_mtmd_prompt(ctx_server.impl->mctx, prompt_str, files));
+                inputs.push_back(process_mtmd_prompt(ctx_server.impl->mctx, prompt_str, files LOCALAI_MTMD_INIT_OPT_ARG(ctx_server.impl->init_opt)));
             } else {
                  // Everything else, including multimodal completions.
-                inputs = tokenize_input_prompts(ctx_server.impl->vocab, ctx_server.impl->mctx, prompt_str, true, true);
+                inputs = tokenize_input_prompts(ctx_server.impl->vocab, ctx_server.impl->mctx, prompt_str, true, true LOCALAI_MTMD_INIT_OPT_ARG(ctx_server.impl->init_opt));
             }
 
             tasks.reserve(inputs.size());
@@ -2659,15 +2711,23 @@ public:
 
                 task.tokens    = std::move(inputs[i]);
 #ifdef LOCALAI_HAS_SERVER_SCHEMA
+                // The schema evaluator no longer takes the per-slot n_ctx: upstream
+                // dropped the parameter and server-schema stopped consulting n_ctx at
+                // all, leaving the context bound to the slot. Forks that predate the
+                // server-schema split still expect it, so only this branch loses it.
                 task.params           = server_schema::eval_llama_cmpl_schema(
+                        ctx_server.impl->vocab,
+                        params_base,
+                        ctx_server.get_meta().logit_bias_eog,
+                        data);
 #else
                 task.params           = server_task::params_from_json_cmpl(
-#endif
                         ctx_server.impl->vocab,
                         params_base,
                         ctx_server.get_meta().slot_n_ctx,
                         ctx_server.get_meta().logit_bias_eog,
                         data);
+#endif
                 task.id_slot = json_value(data, "id_slot", -1);
 
                 // OAI-compat: enable autoparser (PEG-based chat parsing) so that
@@ -2826,7 +2886,7 @@ public:
         json prompt = body.at("embeddings");
 
 
-        auto tokenized_prompts = tokenize_input_prompts(ctx_server.impl->vocab, ctx_server.impl->mctx, prompt, true, true);
+        auto tokenized_prompts = tokenize_input_prompts(ctx_server.impl->vocab, ctx_server.impl->mctx, prompt, true, true LOCALAI_MTMD_INIT_OPT_ARG(ctx_server.impl->init_opt));
         for (const auto & tokens : tokenized_prompts) {
             // this check is necessary for models that do not add BOS token to the input
             if (tokens.empty()) {
@@ -2865,42 +2925,40 @@ public:
             return grpc::Status(grpc::StatusCode::INTERNAL, all_results.error->to_json().value("message", "Error in receiving results"));
         }
 
-        // Collect responses
-        json responses = json::array();
+        // Extract the embeddings typed, straight from the task results (no
+        // JSON round-trip), and report the payload shape alongside the same
+        // flat float array as before: dim is the embedding width, tokens the
+        // number of vectors packed into `embeddings` (1 per prompt when the
+        // server pooled, one per token with pooling:none; summed across
+        // prompts if the request carried several), prompt_tokens the prompt
+        // tokens evaluated, for usage accounting. Consumers seeing 0/0 know
+        // the backend predates shape reporting.
+        int32_t n_vectors = 0;
+        int32_t dim = 0;
+        int32_t prompt_tokens = 0;
         for (auto & res : all_results.results) {
-            GGML_ASSERT(dynamic_cast<server_task_result_embd*>(res.get()) != nullptr);
-            responses.push_back(res->to_json());
-        }
-
-        std::cout << "[DEBUG] Responses size: " << responses.size() << std::endl;
-
-        // Process the responses and extract embeddings
-        for (const auto & response_elem : responses) {
-            // Check if the response has an "embedding" field
-            if (response_elem.contains("embedding")) {
-                json embedding_data = json_value(response_elem, "embedding", json::array());
-
-                if (embedding_data.is_array() && !embedding_data.empty()) {
-                    for (const auto & embedding_vector : embedding_data) {
-                        if (embedding_vector.is_array()) {
-                            for (const auto & embedding_value : embedding_vector) {
-                                embeddingResult->add_embeddings(embedding_value.get<float>());
-                            }
-                        }
-                    }
+            auto * embd_res = dynamic_cast<server_task_result_embd*>(res.get());
+            GGML_ASSERT(embd_res != nullptr);
+            prompt_tokens += embd_res->n_tokens;
+            for (const auto & vec : embd_res->embedding) {
+                for (const float value : vec) {
+                    embeddingResult->add_embeddings(value);
                 }
-            } else {
-                // Check if the response itself contains the embedding data directly
-                if (response_elem.is_array()) {
-                    for (const auto & embedding_value : response_elem) {
-                        embeddingResult->add_embeddings(embedding_value.get<float>());
-                    }
+                if (!vec.empty()) {
+                    n_vectors++;
+                    dim = (int32_t) vec.size();
                 }
             }
         }
+        embeddingResult->set_tokens(n_vectors);
+        embeddingResult->set_dim(dim);
+        embeddingResult->set_prompt_tokens(prompt_tokens);
+        embeddingResult->set_layout(
+            llama_pooling_type(ctx_server.get_llama_context()) == LLAMA_POOLING_TYPE_NONE
+                ? backend::EMBEDDING_LAYOUT_PER_TOKEN
+                : backend::EMBEDDING_LAYOUT_FINAL);
 
-
-
+        std::cout << "[DEBUG] Embedding vectors: " << n_vectors << " x " << dim << std::endl;
 
         return grpc::Status::OK;
     }
@@ -2933,7 +2991,7 @@ public:
 
             tasks.reserve(documents.size());
             for (size_t i = 0; i < documents.size(); i++) {
-                auto tmp = format_prompt_rerank(ctx_server.impl->model_tgt, ctx_server.impl->vocab, ctx_server.impl->mctx, request->query(), documents[i]);
+                auto tmp = format_prompt_rerank(ctx_server.impl->model_tgt, ctx_server.impl->vocab, ctx_server.impl->mctx, request->query(), documents[i] LOCALAI_MTMD_INIT_OPT_ARG(ctx_server.impl->init_opt));
                 server_task task = server_task(SERVER_TASK_TYPE_RERANK);
                 task.id = rd.queue_tasks.get_new_id();
                 task.index = i;
@@ -2954,7 +3012,7 @@ public:
         }
 
         // Collect responses
-        json responses = json::array();
+        std::vector<json> responses;
         for (auto & res : all_results.results) {
             GGML_ASSERT(dynamic_cast<server_task_result_rerank*>(res.get()) != nullptr);
             responses.push_back(res->to_json());
@@ -2967,7 +3025,7 @@ public:
         // Crop results by request.top_n if specified
         int top_n = request->top_n();
         if (top_n > 0 && top_n < static_cast<int>(responses.size())) {
-            responses = json(responses.begin(), responses.begin() + top_n);
+            responses.resize(top_n);
         }
         // Set usage information
         backend::Usage* usage = rerankResult->mutable_usage();
@@ -2994,6 +3052,229 @@ public:
         return grpc::Status::OK;
     }
 
+#ifndef LOCALAI_LLAMA_CPP_NO_TTS_TASK
+    // Builds the shared TTS task from a request. Returns a non-OK status and
+    // leaves `task` untouched when the request is malformed or the loaded model
+    // cannot synthesise audio.
+    grpc::Status prepareTTSTask(const backend::TTSRequest* request, bool stream, server_task & task) {
+        if (!ctx_server.get_meta().has_cap_tts) {
+            return grpc::Status(grpc::StatusCode::UNIMPLEMENTED,
+                "the loaded model does not support audio generation (no gen-audio mmproj)");
+        }
+
+        std::map<std::string, std::string> params(request->params().begin(), request->params().end());
+        const auto opts = llama_grpc::parse_tts_request_options(
+            request->text(),
+            request->voice(),
+            request->has_language() ? request->language() : std::string(),
+            params);
+        if (!opts.ok) {
+            return grpc::Status(grpc::StatusCode::INVALID_ARGUMENT, opts.error);
+        }
+
+        auto wrapper = mtmd_helper_bitmap_init_from_file(ctx_server.impl->mctx, opts.voice_path.c_str(), false LOCALAI_MTMD_INIT_OPT_ARG(ctx_server.impl->init_opt));
+        if (!wrapper.bitmap) {
+            return grpc::Status(grpc::StatusCode::INVALID_ARGUMENT,
+                "failed to read speaker reference audio: " + opts.voice_path);
+        }
+
+        task.tts_inp.set_prompt(opts.text);
+        // core/backend/tts.go always sets TTSRequest.language, so has_language()
+        // is true even when the caller named no language and the string is empty.
+        // gen_audio::inp::get() already maps a stored blank to nullptr, so this
+        // guard is behavior-preserving rather than behavior-fixing. It is kept
+        // so the "unset" intent is visible at the call site instead of resting
+        // on a detail of the helper.
+        if (!opts.language.empty()) {
+            task.tts_inp.set_lang(opts.language);
+        }
+        task.tts_inp.set_speaker_ref(mtmd::bitmap_ptr(wrapper.bitmap));
+        task.tts_inp.data.top_k    = opts.top_k;
+        task.tts_inp.data.top_p    = opts.top_p;
+        task.tts_inp.data.stream   = stream;
+        task.tts_inp.data.out_type = stream
+            ? MTMD_HELPER_GEN_AUDIO_OUTTYPE_PCM   // Go prepends its own WAV header, see core/backend/tts.go
+            : MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV;
+
+        task.params.stream    = stream;
+        // -1 keeps upstream's 512-frame default. The model does not always emit
+        // its codec EOS, so a short input can otherwise generate the full cap.
+        task.params.n_predict = opts.max_frames > 0 ? opts.max_frames : -1;
+        task.params.sampling  = params_base.sampling;
+        // Both values mirror upstream's draft POST /tts handler. Note that the
+        // pair is INERT at this pin: llama_sampler_init_penalties() clamps
+        // penalty_last_n with std::max(penalty_last_n, 0), so -1 means "off",
+        // not "the whole generation", and the penalty sampler is then built
+        // disabled. No repetition penalty is actually applied.
+        //
+        // That is deliberate. Dropping the second line lets the sampling
+        // default of 64 apply and genuinely engages the 1.05 penalty, which was
+        // measured here against the model's habit of never emitting its codec
+        // EOS and running to the frame cap: 0 of 15 short requests ran away
+        // with the penalty inert, 1 of 15 with it active over the last 64
+        // tokens. It does not fix the runaway, so the line stays for parity
+        // with the draft. Use max_frames to bound the output instead.
+        task.params.sampling.penalty_repeat = 1.05f;
+        task.params.sampling.penalty_last_n = -1;
+        if (opts.top_k > 0) {
+            task.params.sampling.top_k = opts.top_k;
+        }
+        if (opts.top_p > 0) {
+            task.params.sampling.top_p = opts.top_p;
+        }
+
+        return grpc::Status::OK;
+    }
+
+    grpc::Status TTS(ServerContext* context, const backend::TTSRequest* request, backend::Result* result) override {
+        auto auth = checkAuth(context);
+        if (!auth.ok()) return auth;
+        auto identity = checkModelIdentity(request);
+        if (!identity.ok()) return identity;
+        if (params_base.model.path.empty()) {
+            return grpc::Status(grpc::StatusCode::FAILED_PRECONDITION, "Model not loaded");
+        }
+        if (request->dst().empty()) {
+            return grpc::Status(grpc::StatusCode::INVALID_ARGUMENT, "dst must name an output file path");
+        }
+
+        server_task task(SERVER_TASK_TYPE_TTS);
+        auto prepared = prepareTTSTask(request, /* stream= */ false, task);
+        if (!prepared.ok()) return prepared;
+
+        auto rd = ctx_server.get_response_reader();
+        task.id = rd.get_new_id();
+        rd.post_task(std::move(task));
+
+        auto should_stop = [context]() { return context->IsCancelled(); };
+
+        std::string audio;
+        while (true) {
+            auto res = rd.next(should_stop);
+            if (!res) {
+                return grpc::Status(grpc::StatusCode::CANCELLED, "TTS request cancelled");
+            }
+            if (res->is_error()) {
+                return grpc::Status(grpc::StatusCode::INTERNAL, res->to_json().dump());
+            }
+            auto * tts_res = dynamic_cast<server_task_result_tts *>(res.get());
+            if (tts_res == nullptr) {
+                return grpc::Status(grpc::StatusCode::INTERNAL, "unexpected result type for a TTS task");
+            }
+            audio.append(tts_res->audio);
+            if (tts_res->final) {
+                break;
+            }
+        }
+
+        std::ofstream out(request->dst(), std::ios::binary | std::ios::trunc);
+        if (!out) {
+            return grpc::Status(grpc::StatusCode::INTERNAL, "failed to open output file: " + request->dst());
+        }
+        out.write(audio.data(), (std::streamsize) audio.size());
+        if (!out) {
+            return grpc::Status(grpc::StatusCode::INTERNAL, "failed to write output file: " + request->dst());
+        }
+        // Buffered data is flushed here, so a full disk or a failing device can
+        // surface for the first time on close. Reporting success then would
+        // leave a truncated file behind under the name the caller will read.
+        out.close();
+        if (!out) {
+            return grpc::Status(grpc::StatusCode::INTERNAL, "failed to close output file: " + request->dst());
+        }
+
+        result->set_success(true);
+        result->set_message("TTS audio generated");
+        return grpc::Status::OK;
+    }
+
+    grpc::Status TTSStream(ServerContext* context, const backend::TTSRequest* request, grpc::ServerWriter<backend::Reply>* writer) override {
+        auto auth = checkAuth(context);
+        if (!auth.ok()) return auth;
+        auto identity = checkModelIdentity(request);
+        if (!identity.ok()) return identity;
+        if (params_base.model.path.empty()) {
+            return grpc::Status(grpc::StatusCode::FAILED_PRECONDITION, "Model not loaded");
+        }
+
+        server_task task(SERVER_TASK_TYPE_TTS);
+        auto prepared = prepareTTSTask(request, /* stream= */ true, task);
+        if (!prepared.ok()) return prepared;
+
+        auto rd = ctx_server.get_response_reader();
+        task.id = rd.get_new_id();
+        rd.post_task(std::move(task));
+
+        auto should_stop = [context]() { return context->IsCancelled(); };
+
+        // core/backend/tts.go:ModelTTSStream builds the WAV header itself from
+        // the sample rate in the first reply's Message, then concatenates every
+        // Reply.Audio verbatim. So the rate goes out once, up front, and the
+        // chunks stay raw PCM.
+        //
+        // Send it before draining rather than off the first audio result: a
+        // chunk needs a whole 72-frame window, about 5.8 s of audio and far
+        // longer in wall time on CPU, and the Go side cannot emit the WAV
+        // header until this reply lands. Waiting would hold the client at zero
+        // bytes for that entire stretch. The rate is a property of the loaded
+        // model, available synchronously, so there is nothing to wait for.
+        {
+            backend::Reply header;
+            const json info = { {"sample_rate", mtmd_gen_audio_get_info(ctx_server.impl->mctx).sample_rate} };
+            header.set_message(info.dump());
+            if (!writer->Write(header)) {
+                return grpc::Status(grpc::StatusCode::CANCELLED, "client closed the TTS stream");
+            }
+        }
+
+        while (true) {
+            auto res = rd.next(should_stop);
+            if (!res) {
+                return grpc::Status(grpc::StatusCode::CANCELLED, "TTS request cancelled");
+            }
+            if (res->is_error()) {
+                return grpc::Status(grpc::StatusCode::INTERNAL, res->to_json().dump());
+            }
+            auto * tts_res = dynamic_cast<server_task_result_tts *>(res.get());
+            if (tts_res == nullptr) {
+                return grpc::Status(grpc::StatusCode::INTERNAL, "unexpected result type for a TTS task");
+            }
+
+            if (!tts_res->audio.empty()) {
+                backend::Reply chunk;
+                chunk.set_audio(tts_pcm_f32_to_s16(tts_res->audio));
+                if (!writer->Write(chunk)) {
+                    return grpc::Status(grpc::StatusCode::CANCELLED, "client closed the TTS stream");
+                }
+            }
+
+            if (tts_res->final) {
+                break;
+            }
+        }
+
+        return grpc::Status::OK;
+    }
+#else
+    grpc::Status TTS(ServerContext* context, const backend::TTSRequest* request, backend::Result* result) override {
+        auto auth = checkAuth(context);
+        if (!auth.ok()) return auth;
+        (void) request;
+        (void) result;
+        return grpc::Status(grpc::StatusCode::UNIMPLEMENTED,
+            "TTS is unavailable in this llama.cpp fork backend");
+    }
+
+    grpc::Status TTSStream(ServerContext* context, const backend::TTSRequest* request, grpc::ServerWriter<backend::Reply>* writer) override {
+        auto auth = checkAuth(context);
+        if (!auth.ok()) return auth;
+        (void) request;
+        (void) writer;
+        return grpc::Status(grpc::StatusCode::UNIMPLEMENTED,
+            "TTSStream is unavailable in this llama.cpp fork backend");
+    }
+#endif
+
     // Score returns the model's joint log-probability of each candidate
     // continuation given a shared prompt.
     //
@@ -3328,9 +3609,15 @@ public:
         // Populate the response with metrics
         response->set_slot_id(0);
         response->set_prompt_json_for_slot("");
+#if LOCALAI_HAS_SERVER_METRICS
+        response->set_tokens_per_second(res_metrics->metrics.prompt_bucket.n_per_second());
+        response->set_tokens_generated(res_metrics->metrics.predict.count);
+        response->set_prompt_tokens_processed(res_metrics->metrics.prompt.count);
+#else
         response->set_tokens_per_second(res_metrics->n_prompt_tokens_processed ? 1.e3 / res_metrics->t_prompt_processing * res_metrics->n_prompt_tokens_processed : 0.);
         response->set_tokens_generated(res_metrics->n_tokens_predicted_total);
         response->set_prompt_tokens_processed(res_metrics->n_prompt_tokens_processed_total);
+#endif
 
 
         return grpc::Status::OK;

backend/cpp/llama-cpp/message_content.h

@@ -52,14 +52,15 @@ inline nlohmann::ordered_json normalize_message_content(const std::string& role,
 // (#7528). A multimodal user message legitimately carries a typed-part array
 // ({type:text}, {type:image_url}, ...), which must be left intact. Shared by the
 // streaming and non-streaming paths so this invariant cannot drift between them.
-inline void normalize_template_message(nlohmann::ordered_json& msg) {
+template <typename Json>
+inline void normalize_template_message(Json& msg) {
     if (!msg.contains("content")) {
         msg["content"] = ""; // templates expect the field to exist
         return;
     }
-    nlohmann::ordered_json& content = msg["content"];
+    auto& content = msg["content"];
     const std::string role = (msg.contains("role") && msg["role"].is_string())
-                                 ? msg["role"].get<std::string>()
+                                 ? msg["role"].template get<std::string>()
                                  : std::string();
     if (content.is_null()) {
         content = ""; // #7324: null would crash content[:N] slicing

backend/cpp/llama-cpp/patches/0001-add-server-task-type-score.patch

@@ -1,8 +1,20 @@
+From 75220a0d74892e3315f4042274b1efa6195868d8 Mon Sep 17 00:00:00 2001
+From: Codex <codex@local>
+Date: Mon, 10 Aug 2026 23:05:52 +0000
+Subject: [PATCH 1/2] score-patch
+
+---
+ common/common.cpp               |   6 +-
+ common/common.h                 |   3 +
+ tools/server/server-context.cpp | 358 +++++++++++++++++++++++++++++++-
+ tools/server/server-task.h      |  47 +++++
+ 4 files changed, 405 insertions(+), 9 deletions(-)
+
 diff --git a/common/common.cpp b/common/common.cpp
-index 8f13217..fc584e1 100644
+index 2e3f14c..0cec0dc 100644
 --- a/common/common.cpp
 +++ b/common/common.cpp
-@@ -1591,8 +1591,10 @@ struct llama_context_params common_context_params_to_llama(const common_params &
+@@ -1636,8 +1636,10 @@ struct llama_context_params common_context_params_to_llama(const common_params &
      auto cparams = llama_context_default_params();
 
      cparams.n_ctx             = params.n_ctx;
@@ -13,13 +25,13 @@ index 8f13217..fc584e1 100644
 +    cparams.n_seq_max         = params.n_parallel + params.n_seq_score_forks;
 +    cparams.n_rs_seq          = std::max(params.speculative.need_n_rs_seq(), (uint32_t) std::max(0, params.n_rs_seq));
      cparams.n_outputs_max     = std::max(params.n_outputs_max, 0);
+     cparams.n_outputs_max_per_seq = std::max(params.n_outputs_max_per_seq, 0);
      cparams.n_batch           = params.n_batch;
-     cparams.n_ubatch          = params.n_ubatch;
 diff --git a/common/common.h b/common/common.h
-index bffc176..e313bd6 100644
+index 878534d..4001df2 100644
 --- a/common/common.h
 +++ b/common/common.h
-@@ -455,6 +455,9 @@ struct common_params {
+@@ -445,6 +445,9 @@ struct common_params {
      int32_t n_keep                =     0; // number of tokens to keep from initial prompt
      int32_t n_chunks              =    -1; // max number of chunks to process (-1 = unlimited)
      int32_t n_parallel            =     1; // number of parallel sequences to decode
@@ -28,39 +40,26 @@ index bffc176..e313bd6 100644
 +    bool    score_enabled          = false; // reserve server resources for the Score task type
      int32_t n_sequences           =     1; // number of sequences to decode
      int32_t n_outputs_max         =     0; // max outputs in a batch (0 = n_batch)
-     int32_t grp_attn_n            =     1; // group-attention factor
-diff --git a/tools/CMakeLists.txt b/tools/CMakeLists.txt
-index 780df32..1d2fe8f 100644
---- a/tools/CMakeLists.txt
-+++ b/tools/CMakeLists.txt
-@@ -41,3 +41,4 @@ else()
-     add_subdirectory(fit-params)
-     add_subdirectory(results)
- endif()
-+add_subdirectory(grpc-server)
+     int32_t n_outputs_max_per_seq =     1; // max outputs per sequence
 diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
-index 715477e..de5bed8 100644
+index 3b5f6a1..d0e18e6 100644
 --- a/tools/server/server-context.cpp
 +++ b/tools/server/server-context.cpp
-@@ -49,7 +49,16 @@ static uint32_t server_n_outputs_max(const common_params & params) {
+@@ -48,6 +48,13 @@ static common_speculative_output_limits server_output_limits(const common_params
+     auto result = common_speculative_get_output_limits(
+             params.n_batch, params.n_parallel, common_speculative_n_max(&params.speculative));
 
-     const uint32_t n_outputs_per_seq = 1 + common_speculative_n_max(&params.speculative);
-
--    const uint64_t n_outputs = (uint64_t) params.n_parallel * n_outputs_per_seq;
-+    // score tasks (SERVER_TASK_TYPE_SCORE) output logits for every candidate
-+    // token, so reserve room for a bounded candidate tail per parallel slot
-+    if (!params.score_enabled) {
-+        return std::max<uint32_t>(1, std::min<uint64_t>(n_batch,
-+            (uint64_t) params.n_parallel * n_outputs_per_seq));
++    // Score tasks output logits for every candidate token, so reserve room
++    // for a bounded candidate tail per parallel slot.
++    if (params.score_enabled) {
++        result.per_seq = std::max<int32_t>(result.per_seq, 1 + SERVER_SCORE_MAX_CAND_TOKENS);
++        result.total   = std::min<int32_t>(params.n_batch, params.n_parallel * result.per_seq);
 +    }
 +
-+    const uint32_t n_outputs_score_seq = 1 + SERVER_SCORE_MAX_CAND_TOKENS;
-+
-+    const uint64_t n_outputs = (uint64_t) params.n_parallel * std::max(n_outputs_per_seq, n_outputs_score_seq);
-
-     return std::max<uint32_t>(1, std::min<uint64_t>(n_batch, n_outputs));
- }
-@@ -202,6 +211,26 @@ struct server_slot {
+     result.total   = std::max<int32_t>(1, result.total);
+     result.per_seq = std::max<int32_t>(1, result.per_seq);
+     return result;
+@@ -239,6 +246,26 @@ struct server_slot {
 
      std::vector<completion_token_output> generated_token_probs;
 
@@ -87,7 +86,7 @@ index 715477e..de5bed8 100644
      bool has_next_token = true;
      bool has_new_line   = false;
      bool truncated      = false;
-@@ -311,6 +340,10 @@ struct server_slot {
+@@ -341,6 +368,10 @@ struct server_slot {
          }
          generated_tokens.clear();
          generated_token_probs.clear();
@@ -97,8 +96,8 @@ index 715477e..de5bed8 100644
 +        score_divergence = -1;
          json_schema = json();
 
-         // clear speculative decoding stats
-@@ -2205,6 +2238,229 @@ private:
+         task_prev = std::move(task);
+@@ -2271,6 +2302,229 @@ private:
          queue_results.send(std::move(res));
      }
 
@@ -328,7 +327,7 @@ index 715477e..de5bed8 100644
      //
      // Functions to process the task
      //
-@@ -2341,6 +2597,7 @@ private:
+@@ -2407,6 +2661,7 @@ private:
              case SERVER_TASK_TYPE_INFILL:
              case SERVER_TASK_TYPE_EMBEDDING:
              case SERVER_TASK_TYPE_RERANK:
@@ -336,7 +335,7 @@ index 715477e..de5bed8 100644
                  {
                      // special case: if input is provided via CLI, tokenize it first
                      // otherwise, no need to tokenize as it's already done inside the HTTP thread
-@@ -2832,6 +3089,13 @@ private:
+@@ -2903,6 +3158,13 @@ private:
                  break; // stop any further processing
              }
          }
@@ -350,7 +349,7 @@ index 715477e..de5bed8 100644
      }
 
      void pre_decode() {
-@@ -3154,6 +3418,16 @@ private:
+@@ -3222,6 +3484,16 @@ private:
                                      n_past = std::min(n_past, slot.alora_invocation_start - 1);
                                  }
 
@@ -367,7 +366,7 @@ index 715477e..de5bed8 100644
                                  const auto n_cache_reuse = slot.task->params.n_cache_reuse;
 
                                  const bool can_cache_reuse =
-@@ -3395,8 +3669,12 @@ private:
+@@ -3455,8 +3727,12 @@ private:
 
                      bool do_checkpoint = params_base.n_ctx_checkpoints > 0;
 
@@ -382,7 +381,7 @@ index 715477e..de5bed8 100644
 
                      // make a checkpoint of the parts of the memory that cannot be rolled back.
                      // checkpoints are created only if:
-@@ -3463,10 +3741,17 @@ private:
+@@ -3444,9 +3720,16 @@ private:
                          // embedding requires all tokens in the batch to be output;
                          // MTP also wants logits at every prompt position so the
                          // streaming hook can mirror t_h_nextn into ctx_dft.
@@ -395,16 +394,12 @@ index 715477e..de5bed8 100644
 +                            slot.prompt.n_tokens() + 1 <  slot.task->n_tokens();
                          add_ok &= batch.add(slot.id,
                              cur_tok,
-                             slot.prompt.tokens.pos_next(),
--                            slot.need_embd());
-+                            slot.need_embd() || need_score_logit);
+                             /* pos       = */ slot.prompt.tokens.pos_next(),
+-                            /* output    = */ slot.need_embd(),
++                            /* output    = */ slot.need_embd() || need_score_logit,
+                             /* is_prompt = */ true);
                          slot.prompt.tokens.push_back(cur_tok);
-
-                         slot.n_prompt_tokens_processed++;
-@@ -3481,6 +3766,32 @@ private:
-                             }
-                         }
-
+@@ -3454,2 +3737,28 @@ private:
 +                        // score tasks: break at the shared-prompt boundary so the checkpoint
 +                        // below lands exactly there — the other candidates of the same
 +                        // scoring call re-process only their own tokens. Also break at the
@@ -431,10 +426,9 @@ index 715477e..de5bed8 100644
 +                            }
 +                        }
 +
-                         // process the last few tokens of the prompt separately in order to allow for a checkpoint to be created.
-                         // create checkpoints that many tokens before the end of the prompt:
-                         //  - 4 + n_ubatch
-@@ -3513,6 +3824,15 @@ private:
+                         // break at the last user message, or at user messages at least min step past the last checkpoint
+                         if (do_checkpoint && spans.is_user_start(slot.prompt.n_tokens())) {
+@@ -3573,6 +3882,15 @@ private:
                      const bool is_user_start = spans.is_user_start(n_tokens_start);
                      const bool is_last_user_message = n_tokens_start == last_user_pos;
 
@@ -450,7 +444,7 @@ index 715477e..de5bed8 100644
                      // entire prompt has been processed
                      if (slot.prompt.n_tokens() == slot.task->n_tokens()) {
                          slot.state = SLOT_STATE_DONE_PROMPT;
-@@ -3528,8 +3848,8 @@ private:
+@@ -3588,8 +3906,8 @@ private:
                          slot.init_sampler();
                      } else {
                          // skip ordinary mid-prompt checkpoints, unless the batch starts a user
@@ -461,7 +455,7 @@ index 715477e..de5bed8 100644
                              do_checkpoint = false;
                          }
                      }
-@@ -3546,10 +3866,10 @@ private:
+@@ -3606,10 +3924,10 @@ private:
                      // do not checkpoint after mtmd chunks
                      do_checkpoint = do_checkpoint && !has_mtmd;
 
@@ -474,7 +468,7 @@ index 715477e..de5bed8 100644
                              n_tokens_start > slot.prompt.checkpoints.back().n_tokens + params_base.checkpoint_min_step);
                      SLT_DBG(slot, "main/do_checkpoint = %s, pos_min = %d, pos_max = %d\n", do_checkpoint ? "yes" : "no", pos_min, pos_max);
 
-@@ -3703,6 +4023,13 @@ private:
+@@ -3772,6 +4090,13 @@ private:
                  }
              }
 
@@ -488,7 +482,7 @@ index 715477e..de5bed8 100644
              if (!is_inside_view(slot.i_batch)) {
                  // the required token not in this sub-batch, skip
                  return;
-@@ -3724,6 +4051,25 @@ private:
+@@ -3793,6 +4118,25 @@ private:
                      return;
                  }
 
@@ -515,7 +509,7 @@ index 715477e..de5bed8 100644
 
                  // prompt evaluated for next-token prediction
 diff --git a/tools/server/server-task.h b/tools/server/server-task.h
-index c3eea2e..fb3c178 100644
+index 6275ec7..5bedf19 100644
 --- a/tools/server/server-task.h
 +++ b/tools/server/server-task.h
 @@ -13,10 +13,25 @@
@@ -597,3 +591,5 @@ index c3eea2e..fb3c178 100644
  struct server_task_result_error : server_task_result {
      error_type err_type = ERROR_TYPE_SERVER;
      std::string err_msg;
+--
+2.39.5

backend/cpp/llama-cpp/patches/0002-add-server-task-type-tts.patch

@@ -0,0 +1,845 @@
+diff --git a/tools/mtmd/mtmd-helper-gen.cpp b/tools/mtmd/mtmd-helper-gen.cpp
+index 1c58d3ae1..196cbd433 100644
+--- a/tools/mtmd/mtmd-helper-gen.cpp
++++ b/tools/mtmd/mtmd-helper-gen.cpp
+@@ -50,29 +50,38 @@ static llama_token find_special_token(const llama_vocab * vocab, const std::stri
+     return LLAMA_TOKEN_NULL;
+ }
+
++static void put_bytes(std::vector<char> & buf, const void * p, size_t n) {
++    const char * c = (const char *) p;
++    buf.insert(buf.end(), c, c + n);
++}
++
++// data_sz == UINT32_MAX writes the "unknown length" sentinel (streaming), same as ffmpeg does on a pipe
++static void write_wav16_header(std::vector<char> & buf, uint32_t data_sz, int32_t rate) {
++    const uint32_t riff_sz   = data_sz == UINT32_MAX ? UINT32_MAX : 36 + data_sz;
++    const uint32_t fmt_sz    = 16, byte_rate = (uint32_t) rate * 2;
++    const uint16_t fmt = 1, ch = 1, align = 2, bits = 16;
++    const uint32_t rate32    = (uint32_t) rate;
++    put_bytes(buf, "RIFF", 4); put_bytes(buf, &riff_sz, 4); put_bytes(buf, "WAVE", 4);
++    put_bytes(buf, "fmt ", 4); put_bytes(buf, &fmt_sz, 4);
++    put_bytes(buf, &fmt, 2); put_bytes(buf, &ch, 2); put_bytes(buf, &rate32, 4);
++    put_bytes(buf, &byte_rate, 4); put_bytes(buf, &align, 2); put_bytes(buf, &bits, 2);
++    put_bytes(buf, "data", 4); put_bytes(buf, &data_sz, 4);
++}
++
++static void append_wav16_pcm(std::vector<char> & buf, const float * pcm, size_t n) {
++    for (size_t i = 0; i < n; i++) {
++        int16_t s = (int16_t) (std::max(-1.0f, std::min(1.0f, pcm[i])) * 32767.0f);
++        put_bytes(buf, &s, 2);
++    }
++}
++
+ static bool write_wav16(std::vector<char> & buf, const std::vector<float> & pcm, int32_t rate) {
+     // RIFF chunk sizes are 32-bit; refuse to emit a file with a truncated header
+     if (pcm.size() > ((size_t) UINT32_MAX - 36) / 2) {
+         return false;
+     }
+-    const uint32_t data_sz   = (uint32_t) (pcm.size() * 2);
+-    const uint32_t riff_sz   = 36 + data_sz;
+-    const uint32_t fmt_sz    = 16, byte_rate = (uint32_t) rate * 2;
+-    const uint16_t fmt = 1, ch = 1, align = 2, bits = 16;
+-    const uint32_t rate32    = (uint32_t) rate;
+-    auto put = [&](const void * p, size_t n) {
+-        const char * c = (const char *) p;
+-        buf.insert(buf.end(), c, c + n);
+-    };
+-    put("RIFF", 4); put(&riff_sz, 4); put("WAVE", 4);
+-    put("fmt ", 4); put(&fmt_sz, 4);
+-    put(&fmt, 2); put(&ch, 2); put(&rate32, 4);
+-    put(&byte_rate, 4); put(&align, 2); put(&bits, 2);
+-    put("data", 4); put(&data_sz, 4);
+-    for (float v : pcm) {
+-        int16_t s = (int16_t) (std::max(-1.0f, std::min(1.0f, v)) * 32767.0f);
+-        put(&s, 2);
+-    }
++    write_wav16_header(buf, (uint32_t) (pcm.size() * 2), rate);
++    append_wav16_pcm(buf, pcm.data(), pcm.size());
+     return true;
+ }
+
+@@ -92,6 +101,8 @@ public:
+     // set out_stop on end-of-speech, h_state_out must be null if no frame is generated
+     virtual int32_t step_gen(llama_token sampled, const float * h_state_in, const float ** h_state_out, bool * out_stop) = 0;
+     virtual int32_t get_output(int32_t * out_sample_rate, const char ** out_data, size_t * out_data_len, int64_t * out_n_samples) = 0;
++    // forces any buffered codes through code2wav now, regardless of window_frames
++    virtual int32_t flush() { return 0; }
+
+ protected:
+     llama_context * lctx;
+@@ -121,6 +132,9 @@ public:
+         prompt_batch.reset();
+         n_prompt = 0;
+         prompt_pos = 0;
++        stream = false;
++        pcm_sent = 0;
++        wav_header_sent = false;
+     }
+
+     int32_t set_input(const mtmd_helper_gen_audio_inp * inp) override {
+@@ -208,6 +222,7 @@ public:
+         top_p = inp->top_p > 0 ? inp->top_p : def.top_p;
+         seed  = inp->seed;
+         out_type = inp->out_type;
++        stream = inp->stream;
+
+         // the prompt above holds the whole text stream up to tts_eos, so every generated
+         // frame adds tts_pad on top of the codes embedding
+@@ -302,31 +317,60 @@ public:
+     }
+
+     int32_t get_output(int32_t * out_sample_rate, const char ** out_data, size_t * out_data_len, int64_t * out_n_samples) override {
+-        if (!flush_gen_wav()) {
+-            return 1;
++        *out_sample_rate = info.sample_rate;
++
++        if (!stream) {
++            // one-shot call: force out whatever's left, regardless of window_frames
++            if (!flush_gen_wav()) {
++                return 1;
++            }
++            if (out_n_samples) {
++                *out_n_samples = (int64_t) audio_pcm.size();
++            }
++            if (out_type == MTMD_HELPER_GEN_AUDIO_OUTTYPE_PCM) {
++                *out_data     = (const char *) audio_pcm.data();
++                *out_data_len = audio_pcm.size() * sizeof(float);
++                return 0;
++            }
++            out_buf.clear();
++            if (!write_wav16(out_buf, audio_pcm, info.sample_rate)) {
++                LOG_ERR("mtmd_helper_gen_audio: output too large for WAV\n");
++                return 1;
++            }
++            *out_data     = out_buf.data();
++            *out_data_len = out_buf.size();
++            return 0;
+         }
+
+-        *out_sample_rate = info.sample_rate;
++        // streaming: only return audio produced since the previous call
++        const size_t n_new = audio_pcm.size() - pcm_sent;
+         if (out_n_samples) {
+-            *out_n_samples = (int64_t) audio_pcm.size();
++            *out_n_samples = (int64_t) n_new;
+         }
+
+         if (out_type == MTMD_HELPER_GEN_AUDIO_OUTTYPE_PCM) {
+-            *out_data     = (const char *) audio_pcm.data();
+-            *out_data_len = audio_pcm.size() * sizeof(float);
++            *out_data     = (const char *) (audio_pcm.data() + pcm_sent);
++            *out_data_len = n_new * sizeof(float);
++            pcm_sent = audio_pcm.size();
+             return 0;
+         }
+
+         out_buf.clear();
+-        if (!write_wav16(out_buf, audio_pcm, info.sample_rate)) {
+-            LOG_ERR("mtmd_helper_gen_audio: output too large for WAV\n");
+-            return 1;
++        if (!wav_header_sent) {
++            write_wav16_header(out_buf, UINT32_MAX, info.sample_rate);
++            wav_header_sent = true;
+         }
++        append_wav16_pcm(out_buf, audio_pcm.data() + pcm_sent, n_new);
++        pcm_sent = audio_pcm.size();
+         *out_data     = out_buf.data();
+         *out_data_len = out_buf.size();
+         return 0;
+     }
+
++    int32_t flush() override {
++        return flush_gen_wav() ? 0 : 1;
++    }
++
+ private:
+     bool ensure_cache() {
+         if (specials_ok) {
+@@ -370,7 +414,7 @@ private:
+             LOG_ERR("mtmd_helper_gen_audio: mmproj has no speaker/audio encoder\n");
+             return false;
+         }
+-        const std::string  marker = mtmd_default_marker();
++        const std::string  marker = mtmd_get_marker(mctx);
+         mtmd_input_text     text{ marker.c_str(), marker.size(), false, true };
+         mtmd_input_chunks * chunks = mtmd_input_chunks_init();
+         const mtmd_bitmap * bptr = bitmap;
+@@ -456,6 +500,9 @@ private:
+     std::vector<float> h_state_buf;
+     mtmd_helper_gen_audio_outtype out_type = MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV;
+     std::vector<char> out_buf;
++    bool   stream          = false;
++    size_t pcm_sent        = 0; // samples already returned by get_output()
++    bool   wav_header_sent = false;
+ };
+
+ // settings that only live in the reference's per-pack yaml, not in the checkpoint
+@@ -1024,6 +1071,14 @@ void mtmd_helper_gen_audio_reset(mtmd_helper_gen_audio * ctx) {
+     }
+ }
+
++struct mtmd_helper_gen_audio_inp mtmd_helper_gen_audio_inp_default(void) {
++    mtmd_helper_gen_audio_inp inp{};
++    inp.top_k    = 50;
++    inp.top_p    = 1.0f;
++    inp.out_type = MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV;
++    return inp;
++}
++
+ int32_t mtmd_helper_gen_audio_set_input(mtmd_helper_gen_audio * ctx, const mtmd_helper_gen_audio_inp * inp) {
+     if (!ctx->pipeline) {
+         LOG_ERR("mtmd_helper_gen_audio: unsupported or missing gen-audio pipeline\n");
+@@ -1060,3 +1115,10 @@ int32_t mtmd_helper_gen_audio_get_output(mtmd_helper_gen_audio * ctx, int32_t *
+     }
+     return ctx->pipeline->get_output(out_sample_rate, out_data, out_data_len, out_n_samples);
+ }
++
++int32_t mtmd_helper_gen_audio_flush(mtmd_helper_gen_audio * ctx) {
++    if (!ctx->pipeline) {
++        return 1;
++    }
++    return ctx->pipeline->flush();
++}
+diff --git a/tools/mtmd/mtmd-helper.h b/tools/mtmd/mtmd-helper.h
+index 832f7171a..3eaa01aab 100644
+--- a/tools/mtmd/mtmd-helper.h
++++ b/tools/mtmd/mtmd-helper.h
+@@ -175,6 +175,7 @@ enum mtmd_helper_gen_audio_outtype {
+     MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV, // WAV PCM 16-bit LE, mono
+ };
+ struct mtmd_helper_gen_audio_inp {
++    bool stream; // if true, output() must be called after each step_gen()
+     llama_seq_id seq_id;
+
+     const char * prompt;
+@@ -190,6 +191,8 @@ struct mtmd_helper_gen_audio_inp {
+     enum mtmd_helper_gen_audio_outtype out_type;
+ };
+
++MTMD_API struct mtmd_helper_gen_audio_inp mtmd_helper_gen_audio_inp_default(void);
++
+ MTMD_API mtmd_helper_gen_audio * mtmd_helper_gen_audio_init(
+                                     struct llama_context * lctx,
+                                     struct mtmd_context * mctx);
+@@ -221,6 +224,8 @@ MTMD_API int32_t mtmd_helper_gen_audio_step_gen(
+
+ // out_data valid until next get_output() or reset() call
+ // out_n_samples (optional, can be NULL) receives the number of generated PCM samples
++// if inp->stream is true: returns only audio produced since the previous call, and
++// *out_data_len == 0 whenever a full window_frames batch hasn't accumulated yet
+ MTMD_API int32_t mtmd_helper_gen_audio_get_output(
+                         mtmd_helper_gen_audio * ctx,
+                         int32_t * out_sample_rate,
+@@ -228,6 +233,10 @@ MTMD_API int32_t mtmd_helper_gen_audio_get_output(
+                         size_t * out_data_len,
+                         int64_t * out_n_samples);
+
++// forces any buffered codes through code2wav now, regardless of window_frames;
++// call once when generation has ended, before the last get_output() in stream mode
++MTMD_API int32_t mtmd_helper_gen_audio_flush(mtmd_helper_gen_audio * ctx);
++
+ #ifdef __cplusplus
+ } // extern "C"
+ #endif
+@@ -254,8 +263,41 @@ struct mtmd_helper_gen_audio_deleter {
+ };
+ using gen_audio_ptr = std::unique_ptr<mtmd_helper_gen_audio, mtmd_helper_gen_audio_deleter>;
+ struct gen_audio {
++
++    // sub-struct, RAII wrapper for mtmd_helper_gen_audio_inp
++    struct inp {
++        mtmd_helper_gen_audio_inp data = mtmd_helper_gen_audio_inp_default();
++        std::string prompt_str;
++        std::string lang_str;
++        mtmd::bitmap_ptr speaker_ref_ptr;
++
++        inp()                             = default;
++        inp(inp &&)                       = default;
++        inp & operator=(inp &&)           = default;
++        inp(const inp &)                  = delete;
++        inp & operator=(const inp &)      = delete;
++
++        void set_prompt     (std::string p)        { prompt_str = std::move(p); }
++        void set_lang       (std::string l)        { lang_str   = std::move(l); }
++        void set_speaker_ref(mtmd::bitmap_ptr bmp) { speaker_ref_ptr = std::move(bmp); }
++
++        // pointers are only valid as long as *this is alive
++        const mtmd_helper_gen_audio_inp * get() {
++            data.prompt      = prompt_str.c_str();
++            data.prompt_len  = prompt_str.size();
++            data.lang        = lang_str.empty() ? nullptr : lang_str.c_str();
++            data.speaker_ref = speaker_ref_ptr.get();
++            return &data;
++        }
++    };
++
+     gen_audio_ptr ctx;
+-    gen_audio(struct llama_context * lctx, struct mtmd_context * mctx) : ctx(mtmd_helper_gen_audio_init(lctx, mctx)) {}
++    void init(struct llama_context * lctx, struct mtmd_context * mctx) {
++        ctx.reset(mtmd_helper_gen_audio_init(lctx, mctx));
++    }
++    bool valid() const {
++        return ctx.get() != nullptr;
++    }
+     void reset() {
+         mtmd_helper_gen_audio_reset(ctx.get());
+     }
+@@ -271,6 +313,9 @@ struct gen_audio {
+     int32_t get_output(int32_t * out_sample_rate, const char ** out_data, size_t * out_data_len, int64_t * out_n_samples = nullptr) {
+         return mtmd_helper_gen_audio_get_output(ctx.get(), out_sample_rate, out_data, out_data_len, out_n_samples);
+     }
++    int32_t flush() {
++        return mtmd_helper_gen_audio_flush(ctx.get());
++    }
+ };
+
+ } // namespace mtmd_helper
+diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
+index 9069463fe..b7fa1e534 100644
+--- a/tools/server/server-context.cpp
++++ b/tools/server/server-context.cpp
+@@ -16,6 +16,7 @@
+ #include "speculative.h"
+ #include "mtmd.h"
+ #include "mtmd-helper.h"
++#include "base64.hpp"
+
+ #include <algorithm>
+ #include <cstddef>
+@@ -41,8 +42,9 @@ constexpr int HTTP_POLLING_SECONDS = 1;
+
+ static common_speculative_output_limits server_output_limits(const common_params & params) {
+     if (params.embedding ||
+-            (params.pooling_type != LLAMA_POOLING_TYPE_UNSPECIFIED && params.pooling_type != LLAMA_POOLING_TYPE_NONE)) {
+-        return { params.n_batch, 1 };
++            (params.pooling_type != LLAMA_POOLING_TYPE_UNSPECIFIED && params.pooling_type != LLAMA_POOLING_TYPE_NONE) ||
++            !params.mmproj.path.empty()) { // gen-audio (TTS) capability isn't known until the mmproj loads, size generously
++        return { params.n_batch, params.n_batch };
+     }
+
+     auto result = common_speculative_get_output_limits(
+@@ -212,6 +214,30 @@ struct server_slot {
+     mtmd_context * mctx = nullptr;
+     mtmd::batch_ptr mbatch = nullptr;
+
++    struct tts_ctx {
++        mtmd_helper::gen_audio ctx;
++        const float * h_state;
++        llama_token   sampled;
++        int32_t       n_decoded;
++        bool is_supported() const {
++            return ctx.valid();
++        }
++        void reset() {
++            // mtmd_helper_gen_audio_reset() dereferences its argument before it
++            // null-checks the pipeline, and the pipeline is only allocated for
++            // models that actually carry a gen-audio mmproj. server_slot::reset()
++            // runs for every slot of every model, so without this guard any
++            // non-TTS model segfaults during slot initialization.
++            if (is_supported()) {
++                ctx.reset();
++            }
++            h_state   = nullptr;
++            sampled   = LLAMA_TOKEN_NULL;
++            n_decoded = 0;
++        }
++    };
++    tts_ctx tts;
++
+     // speculative decoding
+     common_speculative * spec;
+
+@@ -391,6 +417,8 @@ struct server_slot {
+
+         // clear multimodal state
+         mbatch.reset();
++
++        tts.reset();
+     }
+
+     void init_sampler() const {
+@@ -829,6 +857,14 @@ public:
+     mtmd_context * mctx = nullptr;
+     const llama_vocab * vocab = nullptr;
+
++    bool has_cap_tts() const {
++        return mctx != nullptr && mtmd_gen_audio_get_info(mctx).type != MTMD_GEN_AUDIO_TYPE_NONE;
++    }
++
++    bool has_cap_chat() const {
++        return mctx == nullptr || mtmd_helper_model_can_chat(ctx_tgt, mctx);
++    }
++
+     server_queue    queue_tasks;
+     server_response queue_results;
+
+@@ -1288,6 +1324,10 @@ private:
+             slot.mctx                   = mctx;
+             slot.prompt.tokens.has_mtmd = mctx != nullptr;
+
++            if (has_cap_tts()) {
++                slot.tts.ctx.init(ctx_tgt, mctx);
++            }
++
+             SLT_TRC(slot, "new slot, n_ctx = %d\n", slot.n_ctx);
+
+             slot.callback_on_release = [this](int id_slot) {
+@@ -1748,6 +1788,28 @@ private:
+
+         SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str());
+
++        if (task.type == SERVER_TASK_TYPE_TTS) {
++            GGML_ASSERT(has_cap_tts()); // should already checked in route handler
++            if (!slot.tts.is_supported()) {
++                slot.tts.ctx.init(ctx_tgt, slot.mctx);
++            }
++
++            // TTS slots never enter the shared batch: pre_decode() returns early for
++            // them and process_tts_slots() drives them instead, so they skip the
++            // prompt-cache bookkeeping that clears this sequence between requests.
++            // The gen-audio pipeline always decodes from position 0, and its own
++            // reset() only clears host-side buffers, so without this the second and
++            // later tasks on a slot decode over the previous request's tokens and
++            // step_prompt() fails immediately.
++            slot.prompt_clear();
++
++            task.tts_inp.data.seq_id = slot.id;
++            if (slot.tts.ctx.set_input(task.tts_inp.get()) != 0) {
++                send_error(task, "failed to process TTS prompt", ERROR_TYPE_SERVER);
++                return false;
++            }
++        }
++
+         // initialize samplers
+         if (task.need_sampling()) {
+             try {
+@@ -1765,6 +1827,9 @@ private:
+             // TODO: getting pre sampling logits is not yet supported with backend sampling
+             use_backend_sampling &= !need_pre_sample_logits;
+
++            // TODO: check verify if this actually works with TTS
++            use_backend_sampling &= task.type != SERVER_TASK_TYPE_TTS;
++
+             // TODO: tmp until backend sampling is fully implemented
+             if (use_backend_sampling) {
+                 llama_set_sampler(ctx_tgt, slot.id, common_sampler_get(slot.smpl.get()));
+@@ -1783,9 +1848,13 @@ private:
+
+         slot.task = std::make_unique<const server_task>(std::move(task));
+
+-        slot.state = slot.task->is_child()
+-            ? SLOT_STATE_WAIT_OTHER // wait for the parent to process prompt
+-            : SLOT_STATE_STARTED;
++        if (slot.task->type == SERVER_TASK_TYPE_TTS) {
++            slot.state = SLOT_STATE_PROCESSING_PROMPT;
++        } else {
++            slot.state = slot.task->is_child()
++                ? SLOT_STATE_WAIT_OTHER // wait for the parent to process prompt
++                : SLOT_STATE_STARTED;
++        }
+
+         // reset server kill-switch counter
+         n_empty_consecutive = 0;
+@@ -2050,6 +2119,18 @@ private:
+         queue_results.send(std::move(res));
+     }
+
++    void send_tts_result(server_slot & slot, int32_t sample_rate, const char * data, size_t data_len, bool final) {
++        auto res = std::make_unique<server_task_result_tts>();
++
++        res->id          = slot.task->id;
++        res->index       = slot.task->index;
++        res->sample_rate = sample_rate;
++        res->audio.assign(data, data_len);
++        res->final       = final;
++
++        queue_results.send(std::move(res));
++    }
++
+     void send_final_response(server_slot & slot) {
+         auto res = std::make_unique<server_task_result_cmpl_final>();
+
+@@ -2556,6 +2637,7 @@ private:
+             case SERVER_TASK_TYPE_EMBEDDING:
+             case SERVER_TASK_TYPE_RERANK:
+             case SERVER_TASK_TYPE_SCORE:
++            case SERVER_TASK_TYPE_TTS:
+                 {
+                     // special case: if input is provided via CLI, tokenize it first
+                     // otherwise, no need to tokenize as it's already done inside the HTTP thread
+@@ -3007,1 +3089,9 @@ private:
++        // note: TTS slots bypass the shared batch entirely
++        try {
++            process_tts_slots();
++        } catch (const std::exception & e) {
++            SRV_ERR("process_tts_slots() failed: %s\n", e.what());
++            abort_all_slots("process_tts_slots() failed: " + std::string(e.what()));
++        }
++
+         GGML_ASSERT(batch.slot_batched || batch.size() == 0);
+@@ -3074,10 +3164,77 @@ private:
+         }
+     }
+
++    void process_tts_slots() {
++        iterate(slots, [&](server_slot & slot) {
++            if (!slot.is_processing() || slot.task->type != SERVER_TASK_TYPE_TTS) {
++                return;
++            }
++
++            llama_set_embeddings(ctx_tgt, true);
++
++            if (slot.state == SLOT_STATE_PROCESSING_PROMPT) {
++                const int32_t ret = slot.tts.ctx.step_prompt(llama_n_batch(ctx_tgt));
++                if (ret < 0) {
++                    send_error(slot, "TTS prompt processing failed", ERROR_TYPE_SERVER);
++                    slot.release();
++                } else if (ret == 0) {
++                    slot.tts.sampled = common_sampler_sample(slot.smpl.get(), ctx_tgt, -1);
++                    common_sampler_accept(slot.smpl.get(), slot.tts.sampled, true);
++                    slot.tts.h_state = llama_get_embeddings_ith(ctx_tgt, -1);
++                    slot.state = SLOT_STATE_GENERATING;
++                }
++                return;
++            }
++
++            const int32_t n_predict = slot.task->params.n_predict > 0 ? slot.task->params.n_predict : 512;
++            if (slot.tts.n_decoded >= n_predict || llama_vocab_is_eog(vocab, slot.tts.sampled)) {
++                int32_t      sample_rate = 0;
++                const char * data        = nullptr;
++                size_t       data_len    = 0;
++                // generation truly ends here: force out any sub-window remainder still buffered
++                if (slot.tts.ctx.flush() != 0 || slot.tts.ctx.get_output(&sample_rate, &data, &data_len) != 0) {
++                    send_error(slot, "failed to finalize TTS output", ERROR_TYPE_SERVER);
++                } else {
++                    send_tts_result(slot, sample_rate, data, data_len, true);
++                }
++                slot.release();
++                return;
++            }
++
++            const float * h_state_next = nullptr;
++            if (slot.tts.ctx.step_gen(slot.tts.sampled, slot.tts.h_state, &h_state_next) != 0) {
++                send_error(slot, "TTS generation failed", ERROR_TYPE_SERVER);
++                slot.release();
++                return;
++            }
++            slot.tts.h_state = h_state_next;
++            slot.tts.n_decoded++;
++
++            slot.tts.sampled = common_sampler_sample(slot.smpl.get(), ctx_tgt, -1);
++            common_sampler_accept(slot.smpl.get(), slot.tts.sampled, true);
++
++            if (slot.task->params.stream) {
++                int32_t      sample_rate = 0;
++                const char * data        = nullptr;
++                size_t       data_len    = 0;
++                if (slot.tts.ctx.get_output(&sample_rate, &data, &data_len) != 0) {
++                    send_error(slot, "TTS streaming output failed", ERROR_TYPE_SERVER);
++                    slot.release();
++                } else if (data_len > 0) {
++                    send_tts_result(slot, sample_rate, data, data_len, false);
++                }
++            }
++        });
++    }
++
+     void pre_decode() {
+         // apply context-shift if needed
+         // TODO: simplify and improve
+         iterate(slots, [&](server_slot & slot) {
++            if (slot.task && slot.task->type == SERVER_TASK_TYPE_TTS) {
++                // TTS slots drive their own decode loop in process_tts_slots(), never enter the shared batch
++                return;
++            }
+             if (slot.state == SLOT_STATE_GENERATING && slot.prompt.n_tokens() + 1 >= slot.n_ctx) {
+                 if (!params_base.ctx_shift) {
+                     // this check is redundant (for good)
+@@ -3150,7 +3307,7 @@ private:
+
+         // determine which slots are generating and drafting
+         iterate(slots, [&](server_slot & slot) {
+-            if (slot.state != SLOT_STATE_GENERATING) {
++            if (slot.state != SLOT_STATE_GENERATING || slot.task->type == SERVER_TASK_TYPE_TTS) {
+                 return;
+             }
+
+@@ -3284,7 +3441,7 @@ private:
+                     return; // batch is full, skip remaining slots
+                 }
+
+-                if (!slot.is_processing()) {
++                if (!slot.is_processing() || slot.task->type == SERVER_TASK_TYPE_TTS) {
+                     return;
+                 }
+
+@@ -4433,6 +4590,8 @@ server_context_meta server_context::get_meta() const {
+         /* has_inp_image          */ impl->chat_params.allow_image,
+         /* has_inp_audio          */ impl->chat_params.allow_audio,
+         /* has_inp_video          */ impl->chat_params.allow_video,
++        /* has_cap_chat           */ impl->has_cap_chat(),
++        /* has_cap_tts            */ impl->has_cap_tts(),
+         /* json_ui_settings       */ impl->json_ui_settings,
+         /* slot_n_ctx             */ impl->get_slot_n_ctx(),
+         /* pooling_type           */ llama_pooling_type(impl->ctx_tgt),
+@@ -4512,6 +4671,11 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
+
+     res->set_req(&req); // will also set spipe if needed
+
++    if (!ctx_server.has_cap_chat()) {
++        res->error(format_error_response("this server does not support chat/completions", ERROR_TYPE_NOT_SUPPORTED));
++        return res;
++    }
++
+     int32_t sse_ping_interval = params.sse_ping_interval;
+
+     try {
+@@ -5399,6 +5563,150 @@ void server_routes::init_routes() {
+         return res;
+     };
+
++    this->post_tts = [this](const server_http_req & req) {
++        auto res = create_response();
++        res->set_req(&req); // will also set spipe if needed
++
++        if (!ctx_server.has_cap_tts()) {
++            res->error(format_error_response("this server does not support audio generation", ERROR_TYPE_NOT_SUPPORTED));
++            return res;
++        }
++
++        const json body = json::parse(req.body);
++
++        std::string prompt = json_value(body, "input", json_value(body, "prompt", std::string()));
++        if (prompt.empty()) {
++            res->error(format_error_response("\"input\" must be a non-empty string", ERROR_TYPE_INVALID_REQUEST));
++            return res;
++        }
++
++        const std::string response_format = json_value(body, "response_format", std::string("wav"));
++        const bool stream = json_value(body, "stream", false);
++
++        server_task task(SERVER_TASK_TYPE_TTS);
++        task.tts_inp.set_prompt(prompt);
++        task.tts_inp.set_lang(json_value(body, "lang", std::string()));
++        task.tts_inp.data.top_k     = json_value(body, "top_k", 0);
++        task.tts_inp.data.top_p     = json_value(body, "top_p", 0.0f);
++        task.tts_inp.data.stream    = stream;
++        task.tts_inp.data.out_type  = response_format == "pcm"
++            ? MTMD_HELPER_GEN_AUDIO_OUTTYPE_PCM
++            : MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV;
++        task.params.stream    = stream;
++        task.params.n_predict = json_value(body, "n_predict", -1);
++        task.params.sampling  = params.sampling; // baseline defaults, then apply overrides below
++        task.params.sampling.penalty_repeat  = json_value(body, "repeat_penalty", 1.05f);
++        task.params.sampling.penalty_last_n  = -1;
++        if (task.tts_inp.data.top_k > 0) {
++            task.params.sampling.top_k = task.tts_inp.data.top_k;
++        }
++        if (task.tts_inp.data.top_p > 0) {
++            task.params.sampling.top_p = task.tts_inp.data.top_p;
++        }
++
++        // speaker reference: either an uploaded form file ("speaker_ref") or a base64 JSON field ("speaker_ref_b64")
++        const unsigned char * speaker_ref_data = nullptr;
++        size_t speaker_ref_len  = 0;
++        std::string speaker_ref_b64_decoded;
++
++        auto speaker_ref_file = req.files.find("speaker_ref");
++        if (speaker_ref_file != req.files.end()) {
++            speaker_ref_data = speaker_ref_file->second.data.data();
++            speaker_ref_len  = speaker_ref_file->second.data.size();
++        } else {
++            std::string speaker_ref_b64 = json_value(body, "speaker_ref_b64", std::string());
++            if (!speaker_ref_b64.empty()) {
++                speaker_ref_b64_decoded = base64::decode(speaker_ref_b64);
++                speaker_ref_data = (const unsigned char *) speaker_ref_b64_decoded.data();
++                speaker_ref_len  = speaker_ref_b64_decoded.size();
++            }
++        }
++
++        if (speaker_ref_len > 0) {
++            auto wrapper = mtmd_helper_bitmap_init_from_buf(ctx_server.mctx, speaker_ref_data, speaker_ref_len, false, ctx_server.init_opt);
++            if (!wrapper.bitmap) {
++                res->error(format_error_response("failed to decode \"speaker_ref\"", ERROR_TYPE_INVALID_REQUEST));
++                return res;
++            }
++            task.tts_inp.set_speaker_ref(mtmd::bitmap_ptr(wrapper.bitmap));
++        } else {
++            // SRV_WRN expands __VA_ARGS__ without the GNU comma-elision extension,
++            // so a bare format string leaves a trailing comma and will not compile
++            SRV_WRN("%s", "no speaker reference provided, the model may behave randomly\n");
++        }
++
++        auto & rd = res->rd;
++        task.id = rd.get_new_id();
++        rd.post_task(std::move(task));
++
++        const std::string content_type = response_format == "pcm" ? "audio/L16" : "audio/wav";
++
++        if (!stream) {
++            auto result = rd.next(req.should_stop);
++            if (!result) {
++                GGML_ASSERT(req.should_stop());
++                return res; // connection is closed
++            }
++            if (result->is_error()) {
++                res->error(result->to_json());
++                return res;
++            }
++            auto * tts_res = dynamic_cast<server_task_result_tts *>(result.get());
++            GGML_ASSERT(tts_res != nullptr);
++            res->status       = 200;
++            res->content_type = content_type;
++            res->data         = std::move(tts_res->audio);
++            return res;
++        } else {
++            auto first_result = rd.next(req.should_stop);
++            if (!first_result) {
++                GGML_ASSERT(req.should_stop());
++                return res; // connection is closed
++            }
++            if (first_result->is_error()) {
++                res->error(first_result->to_json());
++                return res;
++            }
++            auto * first_tts_res = dynamic_cast<server_task_result_tts *>(first_result.get());
++            GGML_ASSERT(first_tts_res != nullptr);
++
++            res->status       = 200;
++            res->content_type = content_type;
++            res->data          = std::move(first_tts_res->audio);
++            bool is_done       = first_tts_res->final;
++
++            res->set_next([res_this = res.get(), is_done](std::string & output) mutable -> bool {
++                if (is_done) {
++                    return false;
++                }
++                if (res_this->should_stop()) {
++                    return false;
++                }
++                if (!res_this->data.empty()) {
++                    output = std::move(res_this->data);
++                    res_this->data.clear();
++                    return true;
++                }
++
++                server_response_reader & rd = res_this->rd;
++                if (!rd.has_next()) {
++                    return false;
++                }
++                auto result = rd.next([&res_this]() { return res_this->should_stop(); });
++                if (!result || result->is_error()) {
++                    return false;
++                }
++                auto * tts_res = dynamic_cast<server_task_result_tts *>(result.get());
++                GGML_ASSERT(tts_res != nullptr);
++                output  = std::move(tts_res->audio);
++                is_done = tts_res->final;
++                return true;
++            });
++        }
++
++        return res;
++    };
++
+     this->get_lora_adapters = [this](const server_http_req & req) {
+         auto res = create_response();
+
+diff --git a/tools/server/server-context.h b/tools/server/server-context.h
+index f9ab1132b..610512678 100644
+--- a/tools/server/server-context.h
++++ b/tools/server/server-context.h
+@@ -22,6 +22,8 @@ struct server_context_meta {
+     bool has_inp_image;
+     bool has_inp_audio;
+     bool has_inp_video;
++    bool has_cap_chat;
++    bool has_cap_tts;
+     json json_ui_settings;
+     int slot_n_ctx;
+     enum llama_pooling_type pooling_type;
+@@ -151,6 +153,7 @@ struct server_routes {
+     server_http_context::handler_t post_embeddings;
+     server_http_context::handler_t post_embeddings_oai;
+     server_http_context::handler_t post_rerank;
++    server_http_context::handler_t post_tts;
+     server_http_context::handler_t get_lora_adapters;
+     server_http_context::handler_t post_lora_adapters;
+
+diff --git a/tools/server/server-task.cpp b/tools/server/server-task.cpp
+index 1ee677553..939630b8b 100644
+--- a/tools/server/server-task.cpp
++++ b/tools/server/server-task.cpp
+@@ -1497,6 +1497,17 @@ json server_task_result_rerank::to_json() {
+     };
+ }
+
++//
++// server_task_result_tts
++//
++json server_task_result_tts::to_json() {
++    return json {
++        {"sample_rate", sample_rate},
++        {"n_bytes",     audio.size()},
++        {"final",       final},
++    };
++}
++
+ //
+ // server_task_result_error
+ //
+diff --git a/tools/server/server-task.h b/tools/server/server-task.h
+index 5bedf1987..e6ca67a65 100644
+--- a/tools/server/server-task.h
++++ b/tools/server/server-task.h
+@@ -10,6 +10,7 @@
+
+ // TODO: prevent including the whole server-common.h as we only use server_tokens
+ #include "server-common.h"
++#include "mtmd-helper.h"
+
+ using json = nlohmann::ordered_json;
+
+@@ -42,6 +43,7 @@ enum server_task_type {
+     SERVER_TASK_TYPE_SLOT_ERASE,
+     SERVER_TASK_TYPE_GET_LORA,
+     SERVER_TASK_TYPE_SET_LORA,
++    SERVER_TASK_TYPE_TTS,
+ };
+
+ // TODO: change this to more generic "response_format" to replace the "format_response_*" in server-common
+@@ -202,6 +204,9 @@ struct server_task {
+     // used by SERVER_TASK_TYPE_SET_LORA
+     std::map<int, float> set_lora; // mapping adapter ID -> scale
+
++    // used by SERVER_TASK_TYPE_TTS
++    mtmd_helper::gen_audio::inp tts_inp;
++
+     server_task() = default;
+
+     server_task(server_task_type type) : type(type) {}
+@@ -235,6 +240,7 @@ struct server_task {
+         switch (type) {
+             case SERVER_TASK_TYPE_COMPLETION:
+             case SERVER_TASK_TYPE_INFILL:
++            case SERVER_TASK_TYPE_TTS:
+                 return true;
+             default:
+                 return false;
+@@ -494,5 +500,15 @@ struct server_task_result_embd : server_task_result {
+     json to_json_oaicompat();
+ };
+
++struct server_task_result_tts : server_task_result {
++    std::string audio; // raw bytes for this chunk (WAV or PCM, per request's out_type)
++    int32_t     sample_rate = 0;
++    bool        final = false; // true for the last chunk of a request
++
++    virtual bool is_stop() override { return final; }
++
++    virtual json to_json() override;
++};
++
+ struct server_task_result_rerank : server_task_result {
+     float score = -1e6;

backend/cpp/llama-cpp/prepare.sh

@@ -15,6 +15,30 @@ if [ -d "patches" ]; then
     done
 fi
 
+## Apple RDMA link fixup.
+
+## ggml-rpc hands Apple's librdma to the linker with
+## target_link_options(ggml-rpc PRIVATE "LINKER:-weak_library,..."). Link options are not
+## a usage requirement of a static library, so in our BUILD_SHARED_LIBS=OFF build the flag
+## dies with libggml-rpc.a and every ibv_* symbol transport-apple.cpp reaches for comes out
+## undefined when grpc-server and ggml-rpc-server link. Re-declare the same weak link as
+## INTERFACE so it travels to whoever links the static library.
+##
+## Guarded on the marker so a second prepare.sh over the same checkout is a no-op, and on
+## GGML_RPC_RDMA_APPLE so forks that branched before the Apple RDMA transport (turboquant,
+## bonsai) are left alone.
+RPC_CMAKE=llama.cpp/ggml/src/ggml-rpc/CMakeLists.txt
+if [ -f "$RPC_CMAKE" ] && grep -q "GGML_RPC_RDMA_APPLE" "$RPC_CMAKE" && ! grep -q "LOCALAI_RDMA_IFACE" "$RPC_CMAKE"; then
+    echo "==> ggml-rpc carries the Apple RDMA transport, re-declaring its weak librdma link as INTERFACE"
+    cat >> "$RPC_CMAKE" <<'EOF'
+
+# LOCALAI_RDMA_IFACE: added by backend/cpp/llama-cpp/prepare.sh
+if (GGML_RPC_RDMA AND APPLE AND NOT BUILD_SHARED_LIBS)
+    target_link_options(ggml-rpc INTERFACE "LINKER:-weak_library,${RDMA_LIB}")
+endif()
+EOF
+fi
+
 for file in $(ls llama.cpp/tools/server/); do
     cp -rfv llama.cpp/tools/server/$file llama.cpp/tools/grpc-server/
 done
@@ -28,6 +52,13 @@ cp -r message_content_test.cpp llama.cpp/tools/grpc-server/
 # Generic passthrough parser staging and its standalone regression test.
 cp -r passthrough_options.h llama.cpp/tools/grpc-server/
 cp -r passthrough_options_test.cpp llama.cpp/tools/grpc-server/
+# TTS request validation (included by grpc-server.cpp) and its standalone
+# regression test.
+cp -r tts_request_options.h llama.cpp/tools/grpc-server/
+cp -r tts_request_options_test.cpp llama.cpp/tools/grpc-server/
+# Thread-count default normalization and its standalone regression test.
+cp -r thread_params.h llama.cpp/tools/grpc-server/
+cp -r thread_params_test.cpp llama.cpp/tools/grpc-server/
 # Parent-death watcher (included by grpc-server.cpp) and its standalone unit
 # test (run via backend/cpp/run-unit-tests.sh; also buildable under ctest).
 cp -r parent_watch.h llama.cpp/tools/grpc-server/
@@ -49,10 +80,28 @@ else
     echo "==> llama.cpp predates the load-mode enum, using the legacy mmap/mlock/direct-io booleans"
     LEGACY_LOAD_MODE=1
 fi
+if grep -q "server_metrics metrics;" llama.cpp/tools/server/server-task.h; then
+    HAS_SERVER_METRICS=1
+else
+    HAS_SERVER_METRICS=0
+fi
+if grep -q "mtmd_helper_init_opt" llama.cpp/tools/mtmd/mtmd-helper.h; then
+    HAS_MTMD_INIT_OPT=1
+else
+    HAS_MTMD_INIT_OPT=0
+fi
+if grep -q "llm_add_n_cpu_ffn_overrides" llama.cpp/common/common.h; then
+    HAS_N_CPU_FFN_HELPER=1
+else
+    HAS_N_CPU_FFN_HELPER=0
+fi
 cat > llama.cpp/tools/grpc-server/llama_compat.h <<EOF
 // Generated by backend/cpp/llama-cpp/prepare.sh. Do not edit.
 #pragma once
 #define LOCALAI_LEGACY_LOAD_MODE ${LEGACY_LOAD_MODE}
+#define LOCALAI_HAS_SERVER_METRICS ${HAS_SERVER_METRICS}
+#define LOCALAI_HAS_MTMD_INIT_OPT ${HAS_MTMD_INIT_OPT}
+#define LOCALAI_HAS_N_CPU_FFN_HELPER ${HAS_N_CPU_FFN_HELPER}
 EOF
 
 set +e

backend/cpp/llama-cpp/thread_params.h

@@ -0,0 +1,11 @@
+#pragma once
+
+#include <cstdint>
+
+namespace llama_grpc {
+
+inline int32_t resolve_batch_threads(int32_t batch_threads, int32_t inference_threads) {
+    return batch_threads < 0 ? inference_threads : batch_threads;
+}
+
+} // namespace llama_grpc

backend/cpp/llama-cpp/thread_params_test.cpp

@@ -0,0 +1,15 @@
+#include "thread_params.h"
+
+#include <cstdio>
+
+int main() {
+    if (llama_grpc::resolve_batch_threads(-1, 4) != 4) {
+        std::fprintf(stderr, "default batch threads did not inherit inference threads\n");
+        return 1;
+    }
+    if (llama_grpc::resolve_batch_threads(2, 4) != 2) {
+        std::fprintf(stderr, "explicit batch threads were overwritten\n");
+        return 1;
+    }
+    return 0;
+}

backend/cpp/llama-cpp/tts_request_options.h

@@ -0,0 +1,149 @@
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include <cstdint>
+#include <exception>
+#include <map>
+#include <string>
+
+namespace llama_grpc {
+
+// Validated, parsed form of a backend::TTSRequest, kept free of llama.cpp,
+// mtmd and gRPC headers so backend/cpp/run-unit-tests.sh can compile it as a
+// standalone translation unit. grpc-server.cpp turns this into a
+// mtmd_helper::gen_audio::inp.
+struct tts_request_options {
+    bool        ok = false;
+    std::string error;
+
+    std::string text;
+    std::string voice_path;
+    std::string language;
+
+    // 0 / 0.0f mean "unset": upstream only overrides the sampler defaults when
+    // the value is strictly positive.
+    int32_t top_k = 0;
+    float   top_p = 0.0f;
+
+    // Upper bound on generated audio frames, exposed because the model does not
+    // always emit its codec EOS and will otherwise run to the 512-frame default,
+    // which is roughly 41 s at the 12.5 Hz frame rate. 0 means unset, leaving
+    // that default in place.
+    int32_t max_frames = 0;
+};
+
+namespace detail {
+
+// Strict whole-string numeric parsing. std::stoi/stof accept trailing garbage
+// ("40abc" -> 40), which would silently honour a typo'd request.
+inline bool parse_whole_int32(const std::string & value, int32_t & out) {
+    if (value.empty()) {
+        return false;
+    }
+    try {
+        size_t     consumed = 0;
+        const long parsed   = std::stol(value, &consumed);
+        if (consumed != value.size()) {
+            return false;
+        }
+        if (parsed < INT32_MIN || parsed > INT32_MAX) {
+            return false;
+        }
+        out = static_cast<int32_t>(parsed);
+        return true;
+    } catch (const std::exception &) {
+        return false;
+    }
+}
+
+inline bool parse_whole_float(const std::string & value, float & out) {
+    if (value.empty()) {
+        return false;
+    }
+    try {
+        size_t      consumed = 0;
+        const float parsed   = std::stof(value, &consumed);
+        if (consumed != value.size()) {
+            return false;
+        }
+        out = parsed;
+        return true;
+    } catch (const std::exception &) {
+        return false;
+    }
+}
+
+inline tts_request_options reject(const std::string & message) {
+    tts_request_options opts;
+    opts.ok    = false;
+    opts.error = message;
+    return opts;
+}
+
+} // namespace detail
+
+inline tts_request_options parse_tts_request_options(
+        const std::string & text,
+        const std::string & voice,
+        const std::string & language,
+        const std::map<std::string, std::string> & params) {
+    if (text.empty()) {
+        return detail::reject("text must be a non-empty string");
+    }
+
+    // The Qwen3-TTS Base checkpoints have no built-in speaker. Without a
+    // reference clip the model picks an arbitrary voice, so an unset voice is
+    // a request error rather than a defaulted one.
+    if (voice.empty()) {
+        return detail::reject("voice must name a speaker reference audio file");
+    }
+
+    tts_request_options opts;
+    opts.text       = text;
+    opts.voice_path = voice;
+    opts.language   = language;
+
+    // Both values are range-checked here rather than left to the caller: the
+    // consumer copies them straight into mtmd_helper::gen_audio::inp, and only
+    // its separate sampler assignment is guarded by "> 0". An out-of-range or
+    // non-finite value would slip past that guard and reach llama.cpp.
+    const auto top_k_it = params.find("top_k");
+    if (top_k_it != params.end()) {
+        if (!detail::parse_whole_int32(top_k_it->second, opts.top_k)) {
+            return detail::reject("top_k must be an integer, got \"" + top_k_it->second + "\"");
+        }
+        if (opts.top_k < 0) {
+            return detail::reject("top_k must be >= 0, got \"" + top_k_it->second + "\"");
+        }
+    }
+
+    const auto top_p_it = params.find("top_p");
+    if (top_p_it != params.end()) {
+        if (!detail::parse_whole_float(top_p_it->second, opts.top_p)) {
+            return detail::reject("top_p must be a number, got \"" + top_p_it->second + "\"");
+        }
+        // Phrased as a negated in-range test, not "p < 0.0f || p > 1.0f",
+        // because every comparison against NaN is false: the obvious form
+        // would accept NaN, and NaN then defeats the consumer's "> 0" guard
+        // too, since that comparison is false as well.
+        if (!(opts.top_p >= 0.0f && opts.top_p <= 1.0f)) {
+            return detail::reject("top_p must be between 0.0 and 1.0, got \"" + top_p_it->second + "\"");
+        }
+    }
+
+    const auto max_frames_it = params.find("max_frames");
+    if (max_frames_it != params.end()) {
+        if (!detail::parse_whole_int32(max_frames_it->second, opts.max_frames)) {
+            return detail::reject("max_frames must be an integer, got \"" + max_frames_it->second + "\"");
+        }
+        if (opts.max_frames < 0) {
+            return detail::reject("max_frames must be >= 0, got \"" + max_frames_it->second + "\"");
+        }
+    }
+
+    opts.ok = true;
+    return opts;
+}
+
+} // namespace llama_grpc

backend/cpp/llama-cpp/tts_request_options_test.cpp

@@ -0,0 +1,209 @@
+// SPDX-License-Identifier: MIT
+
+#include <cstdio>
+#include <map>
+#include <string>
+
+#include "tts_request_options.h"
+
+static int failures = 0;
+
+static void check(bool ok, const char * name) {
+    if (!ok) {
+        ++failures;
+        std::fprintf(stderr, "FAIL: %s\n", name);
+    }
+}
+
+static void test_accepts_a_minimal_valid_request() {
+    const auto opts = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "en", {});
+
+    check(opts.ok, "minimal request is accepted");
+    check(opts.error.empty(), "minimal request has no error");
+    check(opts.text == "Hello world", "text passes through");
+    check(opts.voice_path == "/models/voices/ref.wav", "voice path passes through");
+    check(opts.language == "en", "language passes through");
+    check(opts.top_k == 0, "top_k defaults to the unset sentinel");
+    check(opts.top_p == 0.0f, "top_p defaults to the unset sentinel");
+    check(opts.max_frames == 0, "max_frames defaults to the unset sentinel");
+}
+
+static void test_rejects_empty_text() {
+    const auto opts = llama_grpc::parse_tts_request_options(
+        "", "/models/voices/ref.wav", "en", {});
+
+    check(!opts.ok, "empty text is rejected");
+    check(opts.error.find("text") != std::string::npos, "empty-text error names the field");
+}
+
+static void test_rejects_missing_speaker_reference() {
+    // Qwen3-TTS Base has no built-in speaker; without a reference it produces
+    // an arbitrary voice, so this must be a hard error rather than a surprise.
+    const auto opts = llama_grpc::parse_tts_request_options(
+        "Hello world", "", "en", {});
+
+    check(!opts.ok, "missing voice is rejected");
+    check(opts.error.find("voice") != std::string::npos, "missing-voice error names the field");
+}
+
+static void test_parses_sampling_params() {
+    const std::map<std::string, std::string> params{
+        {"top_k", "40"},
+        {"top_p", "0.85"},
+    };
+    const auto opts = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", params);
+
+    check(opts.ok, "sampling params are accepted");
+    check(opts.top_k == 40, "top_k is parsed");
+    check(opts.top_p > 0.849f && opts.top_p < 0.851f, "top_p is parsed");
+    check(opts.language.empty(), "absent language stays empty");
+}
+
+static void test_rejects_malformed_sampling_params() {
+    const auto bad_top_k = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_k", "forty"}});
+    check(!bad_top_k.ok, "non-numeric top_k is rejected");
+    check(bad_top_k.error.find("top_k") != std::string::npos, "top_k error names the field");
+
+    const auto bad_top_p = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_p", ""}});
+    check(!bad_top_p.ok, "empty top_p is rejected");
+
+    const auto trailing = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_k", "40abc"}});
+    check(!trailing.ok, "top_k with trailing garbage is rejected");
+
+    const auto trailing_float = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_p", "0.8abc"}});
+    check(!trailing_float.ok, "top_p with trailing garbage is rejected");
+
+    // std::stol returns a long, which is wider than int32_t on 64-bit hosts, so
+    // an in-range-for-long value still has to be caught before the narrowing.
+    const auto overflow_top_k = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_k", "99999999999"}});
+    check(!overflow_top_k.ok, "top_k beyond int32 range is rejected");
+    check(overflow_top_k.error.find("top_k") != std::string::npos,
+          "top_k overflow error names the field");
+}
+
+static void test_rejects_out_of_range_sampling_params() {
+    // These reach mtmd_helper::gen_audio::inp unconditionally downstream, where
+    // the "> 0" sampler guard does not screen them, so they must die here.
+    const auto negative_top_k = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_k", "-5"}});
+    check(!negative_top_k.ok, "negative top_k is rejected");
+    check(negative_top_k.error.find("top_k") != std::string::npos,
+          "negative top_k error names the field");
+
+    const auto negative_top_p = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_p", "-0.1"}});
+    check(!negative_top_p.ok, "negative top_p is rejected");
+    check(negative_top_p.error.find("top_p") != std::string::npos,
+          "negative top_p error names the field");
+
+    const auto large_top_p = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_p", "1.5"}});
+    check(!large_top_p.ok, "top_p above 1.0 is rejected");
+
+    // NaN survives a naive "p < 0.0f || p > 1.0f" range test because every
+    // comparison against NaN is false. This case pins the correct form.
+    const auto nan_top_p = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_p", "nan"}});
+    check(!nan_top_p.ok, "NaN top_p is rejected");
+
+    const auto inf_top_p = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_p", "inf"}});
+    check(!inf_top_p.ok, "infinite top_p is rejected");
+}
+
+static void test_accepts_sampling_param_boundaries() {
+    const auto zero_top_p = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_p", "0.0"}});
+    check(zero_top_p.ok, "top_p of 0.0 is accepted");
+    check(zero_top_p.top_p == 0.0f, "top_p of 0.0 round-trips");
+
+    const auto one_top_p = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_p", "1.0"}});
+    check(one_top_p.ok, "top_p of 1.0 is accepted");
+    check(one_top_p.top_p == 1.0f, "top_p of 1.0 round-trips");
+
+    const auto zero_top_k = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_k", "0"}});
+    check(zero_top_k.ok, "top_k of 0 is accepted");
+}
+
+static void test_parses_max_frames() {
+    // The consumer maps a positive value onto n_predict and leaves upstream's
+    // 512-frame default in place when it is unset, so the sentinel matters as
+    // much as the parsed value.
+    const auto opts = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"max_frames", "120"}});
+
+    check(opts.ok, "max_frames is accepted");
+    check(opts.max_frames == 120, "max_frames is parsed");
+
+    const auto absent = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"top_k", "40"}});
+    check(absent.ok, "a request without max_frames is accepted");
+    check(absent.max_frames == 0, "absent max_frames leaves the unset sentinel");
+
+    const auto zero = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"max_frames", "0"}});
+    check(zero.ok, "max_frames of 0 is accepted");
+    check(zero.max_frames == 0, "max_frames of 0 means unset");
+}
+
+static void test_rejects_malformed_max_frames() {
+    const auto negative = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"max_frames", "-1"}});
+    check(!negative.ok, "negative max_frames is rejected");
+    check(negative.error.find("max_frames") != std::string::npos,
+          "negative max_frames error names the field");
+
+    const auto non_numeric = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"max_frames", "many"}});
+    check(!non_numeric.ok, "non-numeric max_frames is rejected");
+    check(non_numeric.error.find("max_frames") != std::string::npos,
+          "non-numeric max_frames error names the field");
+
+    const auto trailing = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"max_frames", "120abc"}});
+    check(!trailing.ok, "max_frames with trailing garbage is rejected");
+
+    const auto empty = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"max_frames", ""}});
+    check(!empty.ok, "empty max_frames is rejected");
+
+    const auto overflow = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"max_frames", "99999999999"}});
+    check(!overflow.ok, "max_frames beyond int32 range is rejected");
+}
+
+static void test_ignores_unknown_params() {
+    // Unknown keys are backend-specific knobs meant for other TTS engines. A
+    // request routed here must not fail just because it carries them.
+    const auto opts = llama_grpc::parse_tts_request_options(
+        "Hello world", "/models/voices/ref.wav", "", {{"exaggeration", "0.7"}});
+
+    check(opts.ok, "unknown params are ignored, not rejected");
+}
+
+int main() {
+    test_accepts_a_minimal_valid_request();
+    test_rejects_empty_text();
+    test_rejects_missing_speaker_reference();
+    test_parses_sampling_params();
+    test_rejects_malformed_sampling_params();
+    test_rejects_out_of_range_sampling_params();
+    test_accepts_sampling_param_boundaries();
+    test_parses_max_frames();
+    test_rejects_malformed_max_frames();
+    test_ignores_unknown_params();
+
+    if (failures == 0) {
+        std::printf("tts_request_options_test: all checks passed\n");
+    }
+    return failures;
+}

backend/cpp/turboquant/Makefile

@@ -48,6 +48,7 @@ define turboquant-build
 	# stays compiling against vanilla upstream.
 	bash $(CURRENT_MAKEFILE_DIR)/patch-grpc-server.sh $(CURRENT_MAKEFILE_DIR)/../turboquant-$(1)-build/grpc-server.cpp
 	bash $(LLAMA_CPP_DIR)/disable-score-task.sh $(CURRENT_MAKEFILE_DIR)/../turboquant-$(1)-build/grpc-server.cpp
+	bash $(LLAMA_CPP_DIR)/disable-tts-task.sh $(CURRENT_MAKEFILE_DIR)/../turboquant-$(1)-build/grpc-server.cpp
 	$(info $(GREEN)I turboquant build info:$(1)$(RESET))
 	LLAMA_REPO=$(LLAMA_REPO) LLAMA_VERSION=$(TURBOQUANT_VERSION) \
 	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../turboquant-$(1)-build llama.cpp
@@ -86,6 +87,7 @@ turboquant-cpu-all:
 	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build purge
 	bash $(CURRENT_MAKEFILE_DIR)/patch-grpc-server.sh $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build/grpc-server.cpp
 	bash $(LLAMA_CPP_DIR)/disable-score-task.sh $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build/grpc-server.cpp
+	bash $(LLAMA_CPP_DIR)/disable-tts-task.sh $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build/grpc-server.cpp
 	$(info $(GREEN)I turboquant build info:cpu-all-variants$(RESET))
 	LLAMA_REPO=$(LLAMA_REPO) LLAMA_VERSION=$(TURBOQUANT_VERSION) \
 	$(MAKE) -C $(CURRENT_MAKEFILE_DIR)/../turboquant-cpu-all-build llama.cpp

backend/cpp/turboquant/patch-grpc-server.sh

@@ -8,6 +8,8 @@
 #      so the grpc-server option parser skips the two references to
 #      common_params::checkpoint_min_step (the default and the option handler).
 #      That field does not exist in the fork yet; drop this once it does.
+#   3. Use nlohmann's parse_error type in JSON catch clauses because the fork
+#      predates upstream's common_json_error wrapper.
 #
 # The fork used to lag upstream on the whole common_params_speculative refactor
 # (ggml-org/llama.cpp#22397/#22838/#22964), the model_tgt rename (#22838) and
@@ -100,4 +102,16 @@ else
     echo "==> LOCALAI_TURBOQUANT_NO_CHECKPOINT_MIN_STEP define OK"
 fi
 
+# 3. The shared source follows current upstream and catches common_json_error.
+#    TurboQuant still exposes nlohmann::json directly, so its equivalent parse
+#    failures use json::parse_error instead.
+if grep -q 'common_json_error' "$SRC"; then
+    echo "==> patching $SRC to use the TurboQuant JSON exception type"
+    awk '{ gsub(/common_json_error/, "json::parse_error"); print }' "$SRC" > "$SRC.tmp"
+    mv "$SRC.tmp" "$SRC"
+    echo "==> TurboQuant JSON exception patch OK"
+else
+    echo "==> $SRC already uses a TurboQuant-compatible JSON exception type, skipping"
+fi
+
 echo "==> all patches applied"

backend/go/crispasr/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # CrispASR version (release tag)
 CRISPASR_REPO?=https://github.com/CrispStrobe/CrispASR
-CRISPASR_VERSION?=21901d3f7c23554f072964828363e49ddbc2dc68
+CRISPASR_VERSION?=301acd87b036764973b8bfba71e0a21818036d33
 SO_TARGET?=libgocrispasr.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/go/crispasr/gocrispasr.go

@@ -615,10 +615,10 @@ func (w *CrispASR) TTSStream(req *pb.TTSRequest, results chan []byte) error {
 		return fmt.Errorf("crispasr: tempfile: %w", err)
 	}
 	dst := tmp.Name()
+	defer func() { _ = os.Remove(dst) }()
 	if err := tmp.Close(); err != nil {
 		return fmt.Errorf("crispasr: close tempfile: %w", err)
 	}
-	defer func() { _ = os.Remove(dst) }()
 
 	if err := writeWAV(dst, pcm, w.sampleRate); err != nil {
 		return err

backend/go/depth-anything-cpp/Makefile

@@ -14,7 +14,7 @@ JOBS?=$(shell nproc --ignore=1)
 # It is kept alive by the upstream tag da2-support (survives a squash-merge);
 # repoint to the master merge commit once mudler/depth-anything.cpp PR #1 lands.
 DEPTHANYTHING_REPO?=https://github.com/mudler/depth-anything.cpp.git
-DEPTHANYTHING_VERSION?=2028b47ac75a8659c6a9aa617baf09be193eb55f
+DEPTHANYTHING_VERSION?=14f7461d1f704761a038ac9f50dbde8fdb7275e2
 
 ifeq ($(NATIVE),false)
 	CMAKE_ARGS+=-DGGML_NATIVE=OFF

backend/go/local-store/store.go

@@ -38,8 +38,9 @@ type Store struct {
 	// keysAreNormalized stays true until any non-unit-magnitude key
 	// is added; once false, the magnitude-aware fallback path is
 	// used by Find. Re-evaluated only at Set time, never again on
-	// its own — a deletion of the offending key does NOT flip it
-	// back to true (the bookkeeping cost would dominate the gain).
+	// its own — a partial deletion of the offending key does NOT flip
+	// it back to true (the bookkeeping cost would dominate the gain).
+	// An empty store returns to its initial state.
 	keysAreNormalized bool
 
 	// keyLen is the dimension of every stored key. -1 means "no
@@ -142,6 +143,10 @@ func (s *Store) StoresDelete(opts *pb.StoresDeleteOptions) error {
 	mergedV = append(mergedV, tailV...)
 	s.keys = mergedK
 	s.values = mergedV
+	if len(s.keys) == 0 {
+		s.keyLen = -1
+		s.keysAreNormalized = true
+	}
 	assert(slices.IsSortedFunc(s.keys, slices.Compare[[]float32]), "Delete: s.keys not sorted post-merge")
 	assert(len(s.keys) == len(s.values), "Delete: keys/values length skew")
 	return nil

backend/go/local-store/store_test.go

@@ -105,6 +105,46 @@ var _ = Describe("StoresDelete", func() {
 		})).To(Succeed(), "delete of missing key should succeed")
 		Expect(s.keys).To(HaveLen(1))
 	})
+
+	It("reopens the dimension after deleting every key", func() {
+		s := NewStore()
+		oldKey := []float32{2, 0, 0}
+		mustSet(s, [][]float32{oldKey}, [][]byte{[]byte("3d")})
+		Expect(s.keysAreNormalized).To(BeFalse())
+
+		Expect(s.StoresDelete(&pb.StoresDeleteOptions{
+			Keys: wrapKeys([][]float32{oldKey}),
+		})).To(Succeed())
+		Expect(s.keys).To(BeEmpty())
+		Expect(s.keyLen).To(Equal(-1))
+		Expect(s.keysAreNormalized).To(BeTrue())
+
+		newKey := normalizeVec([]float32{1, 1})
+		mustSet(s, [][]float32{newKey}, [][]byte{[]byte("2d")})
+		res, err := s.StoresFind(&pb.StoresFindOptions{
+			Key:  &pb.StoresKey{Floats: newKey},
+			TopK: 1,
+		})
+		Expect(err).NotTo(HaveOccurred())
+		Expect(res.Values).To(HaveLen(1))
+		Expect(string(res.Values[0].Bytes)).To(Equal("2d"))
+	})
+
+	It("retains the dimension after a partial delete", func() {
+		s := NewStore()
+		mustSet(s,
+			[][]float32{{1, 0, 0}, {0, 1, 0}},
+			[][]byte{[]byte("x"), []byte("y")},
+		)
+		Expect(s.StoresDelete(&pb.StoresDeleteOptions{
+			Keys: wrapKeys([][]float32{{1, 0, 0}}),
+		})).To(Succeed())
+		Expect(s.keyLen).To(Equal(3))
+		Expect(s.StoresSet(&pb.StoresSetOptions{
+			Keys:   wrapKeys([][]float32{{1, 0}}),
+			Values: wrapValues([][]byte{[]byte("2d")}),
+		})).NotTo(Succeed())
+	})
 })
 
 var _ = Describe("StoresFind", func() {

backend/go/locate-anything-cpp/golocateanythingcpp.go

@@ -11,6 +11,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"runtime"
 	"unsafe"
 
 	"github.com/mudler/LocalAI/pkg/grpc/base"
@@ -109,30 +110,25 @@ func (r *LocateAnythingCpp) Detect(opts *pb.DetectOptions) (pb.DetectResponse, e
 		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: a text prompt is required (open-vocabulary detection)")
 	}
 
-	// Decode base64 image and write to temp file.
 	imgData, err := base64.StdEncoding.DecodeString(opts.Src)
 	if err != nil {
 		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: failed to decode base64 image: %w", err)
 	}
-
-	tmpFile, err := os.CreateTemp("", "locate-anything-*.img")
-	if err != nil {
-		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: failed to create temp file: %w", err)
-	}
-	defer func() { _ = os.Remove(tmpFile.Name()) }()
-
-	if _, err := tmpFile.Write(imgData); err != nil {
-		_ = tmpFile.Close()
-		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: failed to write temp file: %w", err)
-	}
-	if err := tmpFile.Close(); err != nil {
-		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: failed to close temp file: %w", err)
+	if len(imgData) == 0 {
+		return pb.DetectResponse{}, fmt.Errorf("locate-anything-cpp: decoded image is empty")
 	}
 
 	// mode 0 = hybrid (Parallel Box Decoding). The JSON return value is unused:
 	// structured detections are read via the accessor functions. Still must
 	// free the returned string.
-	jsonPtr := CapiLocatePath(r.handle, tmpFile.Name(), prompt, 0)
+	jsonPtr := CapiLocateBuffer(
+		r.handle,
+		uintptr(unsafe.Pointer(unsafe.SliceData(imgData))),
+		uintptr(len(imgData)),
+		prompt,
+		0,
+	)
+	runtime.KeepAlive(imgData)
 	if jsonPtr != 0 {
 		CapiFreeString(jsonPtr)
 	}

backend/go/locate-anything-cpp/golocateanythingcpp_unit_test.go

@@ -0,0 +1,54 @@
+package main
+
+import (
+	"encoding/base64"
+	"path/filepath"
+
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("LocateAnythingCpp detection input", func() {
+	It("detects from memory when the temporary directory is unavailable", func() {
+		originalLocateBuffer := CapiLocateBuffer
+		originalLocatePath := CapiLocatePath
+		originalGetNDetections := CapiGetNDetections
+		defer func() {
+			CapiLocateBuffer = originalLocateBuffer
+			CapiLocatePath = originalLocatePath
+			CapiGetNDetections = originalGetNDetections
+		}()
+
+		image := []byte("encoded-image")
+		var receivedData uintptr
+		var receivedLength uintptr
+		CapiLocateBuffer = func(_ uintptr, data uintptr, length uintptr, _ string, _ int32) uintptr {
+			receivedData = data
+			receivedLength = length
+			return 0
+		}
+		CapiLocatePath = func(_ uintptr, _ string, _ string, _ int32) uintptr {
+			Fail("path-based detection must not be called")
+			return 0
+		}
+		CapiGetNDetections = func(uintptr) int32 { return 0 }
+		GinkgoT().Setenv("TMPDIR", filepath.Join(GinkgoT().TempDir(), "missing"))
+
+		result, err := (&LocateAnythingCpp{handle: 1}).Detect(&pb.DetectOptions{
+			Src:    base64.StdEncoding.EncodeToString(image),
+			Prompt: "the object",
+		})
+
+		Expect(err).NotTo(HaveOccurred())
+		Expect(result.Detections).To(BeEmpty())
+		Expect(receivedData).NotTo(BeZero())
+		Expect(receivedLength).To(Equal(uintptr(len(image))))
+	})
+
+	It("rejects an empty decoded image", func() {
+		_, err := (&LocateAnythingCpp{handle: 1}).Detect(&pb.DetectOptions{Prompt: "the object"})
+
+		Expect(err).To(MatchError("locate-anything-cpp: decoded image is empty"))
+	})
+})

backend/go/nemo-speech-cpp/Makefile

@@ -12,7 +12,7 @@
 # runs 'make -C backend/go/$(BACKEND) build' and then copies package/), so it
 # has to produce the binary and the package, not just the shared libraries.
 
-NEMO_SPEECH_VERSION?=2e12e2def8a98ed06666f7ee3ca94e7193e04be4
+NEMO_SPEECH_VERSION?=a5b6953c4a579a2bbd1c0913ad8a85c2a4d99953
 NEMO_SPEECH_REPO?=https://github.com/NVIDIA/NeMo-Speech.cpp
 
 GOCMD?=go
@@ -88,6 +88,18 @@ ITN_LIB_DIR=$(ITN_PREFIX)/lib
 ITN_MARKER=$(ITN_LIB_DIR)/libsparrowhawk.so
 ITN_FST_HEADER=$(ITN_PREFIX)/include/fst/fst.h
 
+# SentencePiece became a core ASR dependency in 5be7bfb: RNNT context biasing
+# uses it even when Flashlight and text normalization are disabled. Build the
+# pinned static archive provided by upstream so every platform gets the same
+# dependency instead of relying on an undeclared system package.
+SENTENCEPIECE_PREFIX=sources/NeMo-Speech.cpp/.deps/sentencepiece
+SENTENCEPIECE_MARKER=$(SENTENCEPIECE_PREFIX)/lib/libsentencepiece.a
+
+# Linux's ASR CMake block looks in NEMO_SPEECH_DEPENDENCY_PREFIX directly, but
+# the Apple branch uses generic find_library()/find_path(). Put the same private
+# prefix on CMake's search path so Darwin consumes the archive built above too.
+CMAKE_ARGS+=-DCMAKE_PREFIX_PATH=$(abspath $(SENTENCEPIECE_PREFIX))
+
 ITN_CC?=gcc-12
 ITN_CXX?=g++-12
 
@@ -152,7 +164,7 @@ else
 endif
 CMAKE_ARGS+=-DNEMO_SPEECH_GGML_PATCHED=$(GGML_PATCHED)
 
-.PHONY: nemo-speech-cpp-grpc package build clean purge test all stage-libs patch-ggml engine itn patch-itn-headers
+.PHONY: nemo-speech-cpp-grpc package build clean purge test all stage-libs patch-ggml engine itn sentencepiece patch-itn-headers
 
 all: nemo-speech-cpp-grpc package
 
@@ -266,11 +278,28 @@ patch-itn-headers:
 
 itn: $(ITN_MARKER)
 
+$(SENTENCEPIECE_MARKER): | sources/NeMo-Speech.cpp
+	# Upstream's license copies use GNU install's -D flag, which BSD install
+	# does not support. Homebrew CMake 4 also rejects SentencePiece's old policy
+	# floor. Patch both incompatibilities before running the helper on Darwin.
+	@if [ "$(shell uname -s)" = Darwin ]; then \
+		cd sources/NeMo-Speech.cpp && \
+		mkdir -p .deps/sentencepiece/share/licenses/nemo-speech/third_party/sentencepiece && \
+		perl -pi \
+			-e 's/install -Dm0644/install -m 0644/g;' \
+			-e 's/-DCMAKE_BUILD_TYPE=Release /-DCMAKE_BUILD_TYPE=Release -DCMAKE_POLICY_VERSION_MINIMUM=3.5 /;' \
+			scripts/build_sentencepiece_static.sh; \
+	fi
+	cd sources/NeMo-Speech.cpp && JOBS=$(JOBS) scripts/build_sentencepiece_static.sh
+
+sentencepiece: $(SENTENCEPIECE_MARKER)
+
 # Only a WITH_NORM=ON build needs the ITN stack, and it must exist before cmake
 # configures, since the WITH_NORM cmake block find_library()s into the prefix
 # with REQUIRED.
+NEMO_RUNTIME_PREREQS=$(SENTENCEPIECE_MARKER)
 ifeq ($(WITH_NORM),ON)
-NEMO_RUNTIME_PREREQS=$(ITN_MARKER)
+NEMO_RUNTIME_PREREQS+=$(ITN_MARKER)
 endif
 
 # Upstream sets CMAKE_LIBRARY_OUTPUT_DIRECTORY to ${CMAKE_BINARY_DIR}/bin, so the

backend/go/omnivoice-cpp/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # omnivoice.cpp version
 OMNIVOICE_REPO?=https://github.com/ServeurpersoCom/omnivoice.cpp
-OMNIVOICE_VERSION?=4f33af825d66e6ef1cb185e87b4589cacf747291
+OMNIVOICE_VERSION?=040c8b344d8c670ce1475194751d119b5ef82c78
 SO_TARGET?=libgomnivoicecpp.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/go/parakeet-cpp/Makefile

@@ -1,6 +1,6 @@
 # parakeet-cpp backend Makefile.
 #
-# Upstream pin lives below as PARAKEET_VERSION?=1bfbebfaaf493866f49597cd3b7901959d395c60
+# Upstream pin lives below as PARAKEET_VERSION?=e75de9b6b9b688fd293aa22f7e27aa724ea286f8
 # (.github/bump_deps.sh) can find and update it - matches the
 # whisper.cpp / ds4 / vibevoice-cpp convention.
 #
@@ -15,7 +15,7 @@
 # That's what the L0 smoke test uses. The default target below does the
 # proper clone-at-pin + cmake build so CI doesn't need a side-checkout.
 
-PARAKEET_VERSION?=1bfbebfaaf493866f49597cd3b7901959d395c60
+PARAKEET_VERSION?=e75de9b6b9b688fd293aa22f7e27aa724ea286f8
 PARAKEET_REPO?=https://github.com/mudler/parakeet.cpp
 
 GOCMD?=go
@@ -49,6 +49,8 @@ else ifeq ($(BUILD_TYPE),hipblas)
 	CMAKE_ARGS+=-DPARAKEET_GGML_HIP=ON
 else ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DPARAKEET_GGML_VULKAN=ON
+else ifeq ($(BUILD_TYPE),metal)
+	CMAKE_ARGS+=-DPARAKEET_GGML_METAL=ON
 endif
 
 .PHONY: parakeet-cpp-grpc package build clean purge test all

backend/go/rfdetr-cpp/gorfdetrcpp.go

@@ -10,6 +10,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"runtime"
 	"strconv"
 	"unsafe"
 
@@ -102,24 +103,12 @@ func (r *RFDetrCpp) Detect(opts *pb.DetectOptions) (pb.DetectResponse, error) {
 		return pb.DetectResponse{}, fmt.Errorf("rfdetr-cpp: model not loaded")
 	}
 
-	// Decode base64 image and write to temp file.
 	imgData, err := base64.StdEncoding.DecodeString(opts.Src)
 	if err != nil {
 		return pb.DetectResponse{}, fmt.Errorf("rfdetr-cpp: failed to decode base64 image: %w", err)
 	}
-
-	tmpFile, err := os.CreateTemp("", "rfdetr-*.img")
-	if err != nil {
-		return pb.DetectResponse{}, fmt.Errorf("rfdetr-cpp: failed to create temp file: %w", err)
-	}
-	defer func() { _ = os.Remove(tmpFile.Name()) }()
-
-	if _, err := tmpFile.Write(imgData); err != nil {
-		_ = tmpFile.Close()
-		return pb.DetectResponse{}, fmt.Errorf("rfdetr-cpp: failed to write temp file: %w", err)
-	}
-	if err := tmpFile.Close(); err != nil {
-		return pb.DetectResponse{}, fmt.Errorf("rfdetr-cpp: failed to close temp file: %w", err)
+	if len(imgData) == 0 {
+		return pb.DetectResponse{}, fmt.Errorf("rfdetr-cpp: decoded image is empty")
 	}
 
 	threshold := opts.Threshold
@@ -127,10 +116,18 @@ func (r *RFDetrCpp) Detect(opts *pb.DetectOptions) (pb.DetectResponse, error) {
 		threshold = 0.5
 	}
 
-	// JSON output from detect_path is unused: we read structured detections via
+	// JSON output from the detection ABI is unused: we read structured detections via
 	// the accessor functions. Still must free the returned string.
 	var jsonPtr uintptr
-	rc := CapiDetectPath(r.handle, tmpFile.Name(), threshold, uint32(defaultTopK), &jsonPtr)
+	rc := CapiDetectBuffer(
+		r.handle,
+		uintptr(unsafe.Pointer(unsafe.SliceData(imgData))),
+		uintptr(len(imgData)),
+		threshold,
+		uint32(defaultTopK),
+		&jsonPtr,
+	)
+	runtime.KeepAlive(imgData)
 	if jsonPtr != 0 {
 		CapiFreeString(jsonPtr)
 	}

backend/go/rfdetr-cpp/gorfdetrcpp_unit_test.go

@@ -0,0 +1,56 @@
+package main
+
+import (
+	"encoding/base64"
+	"path/filepath"
+
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("RFDetrCpp detection input", func() {
+	It("detects from memory when the temporary directory is unavailable", func() {
+		originalDetectBuffer := CapiDetectBuffer
+		originalDetectPath := CapiDetectPath
+		originalFreeString := CapiFreeString
+		originalGetNDetections := CapiGetNDetections
+		defer func() {
+			CapiDetectBuffer = originalDetectBuffer
+			CapiDetectPath = originalDetectPath
+			CapiFreeString = originalFreeString
+			CapiGetNDetections = originalGetNDetections
+		}()
+
+		image := []byte("encoded-image")
+		var receivedData uintptr
+		var receivedLength uintptr
+		CapiDetectBuffer = func(_ uintptr, data uintptr, length uintptr, _ float32, _ uint32, _ *uintptr) int32 {
+			receivedData = data
+			receivedLength = length
+			return 0
+		}
+		CapiDetectPath = func(_ uintptr, _ string, _ float32, _ uint32, _ *uintptr) int32 {
+			Fail("path-based detection must not be called")
+			return -1
+		}
+		CapiFreeString = func(uintptr) {}
+		CapiGetNDetections = func(uintptr) int32 { return 0 }
+		GinkgoT().Setenv("TMPDIR", filepath.Join(GinkgoT().TempDir(), "missing"))
+
+		result, err := (&RFDetrCpp{handle: 1}).Detect(&pb.DetectOptions{
+			Src: base64.StdEncoding.EncodeToString(image),
+		})
+
+		Expect(err).NotTo(HaveOccurred())
+		Expect(result.Detections).To(BeEmpty())
+		Expect(receivedData).NotTo(BeZero())
+		Expect(receivedLength).To(Equal(uintptr(len(image))))
+	})
+
+	It("rejects an empty decoded image", func() {
+		_, err := (&RFDetrCpp{handle: 1}).Detect(&pb.DetectOptions{})
+
+		Expect(err).To(MatchError("rfdetr-cpp: decoded image is empty"))
+	})
+})

backend/go/stablediffusion-ggml/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # stablediffusion.cpp (ggml)
 STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
-STABLEDIFFUSION_GGML_VERSION?=c6beeef35526c6dc94b74a7fb69f9d2e6a2a7a12
+STABLEDIFFUSION_GGML_VERSION?=d04e8950c1ec8d30248cbe996682b3182fb1adf6
 
 CMAKE_ARGS+=-DGGML_MAX_NAME=128
 
@@ -38,17 +38,16 @@ else ifeq ($(BUILD_TYPE),hipblas)
 	ROCM_PATH ?= /opt/rocm
 	export CXX=$(ROCM_HOME)/llvm/bin/clang++
 	export CC=$(ROCM_HOME)/llvm/bin/clang
-	AMDGPU_TARGETS?=gfx908,gfx90a,gfx942,gfx950,gfx1030,gfx1100,gfx1101,gfx1102,gfx1200,gfx1201
-	CMAKE_ARGS+=-DSD_HIPBLAS=ON -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=$(AMDGPU_TARGETS)
+	AMDGPU_TARGETS?=gfx908,gfx90a,gfx942,gfx950,gfx1030,gfx1100,gfx1101,gfx1102,gfx1151,gfx1200,gfx1201
+	# SD_HIPBLAS turns on ggml's HIP backend itself; GGML_HIPBLAS is the name ggml
+	# used before it was renamed to GGML_HIP, so passing it here only produced an
+	# unused-variable warning.
+	CMAKE_ARGS+=-DSD_HIPBLAS=ON -DAMDGPU_TARGETS=$(AMDGPU_TARGETS)
 else ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DSD_VULKAN=ON -DGGML_VULKAN=ON
-else ifeq ($(OS),Darwin)
-	ifneq ($(BUILD_TYPE),metal)
-		CMAKE_ARGS+=-DSD_METAL=OFF -DGGML_METAL=OFF
-	else
-		CMAKE_ARGS+=-DSD_METAL=ON -DGGML_METAL=ON
-		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
-	endif
+else ifeq ($(BUILD_TYPE),metal)
+	CMAKE_ARGS+=-DSD_METAL=ON -DGGML_METAL=ON
+	CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
 endif
 
 ifeq ($(BUILD_TYPE),sycl_f16)
@@ -72,7 +71,6 @@ sources/stablediffusion-ggml.cpp:
 	git checkout $(STABLEDIFFUSION_GGML_VERSION) && \
 	git submodule update --init --recursive --depth 1 --single-branch
 
-# Detect OS
 UNAME_S := $(shell uname -s)
 
 # Only build CPU variants on Linux
@@ -134,4 +132,4 @@ libgosd-custom: CMakeLists.txt cpp/gosd.cpp cpp/gosd.h
 	(mv build-$(SO_TARGET)/libgosd.so ./$(SO_TARGET) 2>/dev/null || \
 	 mv build-$(SO_TARGET)/libgosd.dylib ./$(SO_TARGET) 2>/dev/null)
 
-all: stablediffusion-ggml package
+all: stablediffusion-ggml package

backend/go/stablediffusion-ggml/cpp/gosd.cpp

@@ -401,7 +401,6 @@ int load_model(const char *model, char *model_path, char* options[], int threads
     const char *params_backend_arg = "";
     const char *rpc_servers_arg = "";
     const char *max_vram_arg = "";
-    bool stream_layers = false;
 
     int n_threads = threads;
     enum sd_type_t wtype = SD_TYPE_COUNT;
@@ -510,7 +509,10 @@ int load_model(const char *model, char *model_path, char* options[], int threads
         if (!strcmp(optname, "params_backend")) params_backend_arg = strdup(optval);
         if (!strcmp(optname, "rpc_servers")) rpc_servers_arg = strdup(optval);
         if (!strcmp(optname, "max_vram")) max_vram_arg = strdup(optval);
-        if (!strcmp(optname, "stream_layers")) stream_layers = (strcmp(optval, "true") == 0 || strcmp(optval, "1") == 0);
+        if (!strcmp(optname, "stream_layers")) {
+            // Retained as a no-op for existing configurations. Upstream now
+            // selects segmented weight streaming automatically.
+        }
 
         // vae_decode_only is still accepted for backwards compatibility with
         // existing gallery configs, but upstream dropped the option (the model
@@ -650,11 +652,9 @@ int load_model(const char *model, char *model_path, char* options[], int threads
             ctx_params.rpc_servers = env_rpc_servers;
         }
     }
-    // max_vram: GiB budget or per-backend spec for graph-cut segmented param
-    // offload ("0" = disabled, "-1" = auto). stream_layers only has effect when
-    // max_vram is set.
+    // max_vram is an optional GiB budget or per-backend spec for automatic
+    // graph-cut execution. A zero value uses the live free-VRAM budget.
     if (strlen(max_vram_arg) > 0) ctx_params.max_vram = max_vram_arg;
-    ctx_params.stream_layers = stream_layers;
     ctx_params.diffusion_flash_attn = diffusion_flash_attn;
     ctx_params.tae_preview_only = tae_preview_only;
     ctx_params.diffusion_conv_direct = diffusion_conv_direct;
@@ -1144,17 +1144,25 @@ static uint8_t* load_and_resize_image(const char* path, int target_width, int ta
 // Write sd.cpp's audio buffer to a temp WAV file (IEEE float, interleaved).
 // sd_audio_t.data is planar (all channel 0 samples, then channel 1, etc.) — we
 // interleave on the fly so ffmpeg's standard wav demuxer can read it directly.
-// Returns 0 on success and fills wav_path (must be at least 64 bytes).
+// Returns 0 on success and fills wav_path.
 static int write_planar_float_wav(const sd_audio_t* a, char* wav_path, size_t wav_path_sz) {
     if (!a || !a->data || a->sample_count == 0 || a->channels == 0 || a->sample_rate == 0) {
         return -1;
     }
 
-    snprintf(wav_path, wav_path_sz, "/tmp/gosd-audio-XXXXXX.wav");
+    const char* temp_dir = getenv("TMPDIR");
+    if (!temp_dir || temp_dir[0] == '\0') {
+        temp_dir = "/tmp";
+    }
+    int path_len = snprintf(wav_path, wav_path_sz, "%s/gosd-audio-XXXXXX.wav", temp_dir);
+    if (path_len < 0 || (size_t)path_len >= wav_path_sz) {
+        fprintf(stderr, "temporary directory path is too long\n");
+        return -1;
+    }
     int fd = mkstemps(wav_path, 4);
     if (fd < 0) { perror("mkstemps wav"); return -1; }
     FILE* f = fdopen(fd, "wb");
-    if (!f) { perror("fdopen wav"); close(fd); return -1; }
+    if (!f) { perror("fdopen wav"); close(fd); unlink(wav_path); return -1; }
 
     uint64_t frames = a->sample_count;
     uint32_t channels = a->channels;
@@ -1221,7 +1229,7 @@ static int ffmpeg_mux_raw_to_mp4(sd_image_t* frames, int num_frames, int fps,
     snprintf(fps_str, sizeof(fps_str), "%d", fps);
 
     // Optional audio: write a temp WAV file if the model produced audio.
-    char wav_path[64] = {0};
+    char wav_path[4096] = {0};
     bool have_audio = false;
     if (audio && audio->data && audio->sample_count > 0 && audio->channels > 0 && audio->sample_rate > 0) {
         if (write_planar_float_wav(audio, wav_path, sizeof(wav_path)) == 0) {
@@ -1438,4 +1446,3 @@ int unload() {
     free_sd_ctx(sd_c);
     return 0;
 }
-

backend/go/vllm-cpp/Makefile

@@ -11,7 +11,7 @@ JOBS?=$(shell nproc --ignore=1 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || e
 
 # vllm.cpp version
 VLLM_CPP_REPO?=https://github.com/mudler/vllm.cpp
-VLLM_CPP_VERSION?=0757cac231ecd571a83c4fd2f50805c9251fc225
+VLLM_CPP_VERSION?=6bf3abb580982f4fd2e4525ef37802ee0ce28981
 
 # MLX GEMM provider (darwin/metal only; see the metal branch below for why).
 # Consumed as the prebuilt pip wheel: building MLX from source needs `xcrun
@@ -47,26 +47,35 @@ CMAKE_ARGS+=-DCMAKE_BUILD_TYPE=Release
 UNAME_M := $(shell uname -m)
 
 ifeq ($(BUILD_TYPE),cublas)
-	# Blackwell-family targets only: other CUDA arches are build-supported
-	# upstream but have no runtime-proven fast path. amd64 gets the consumer
-	# (120a) + GB10 (121a) fat binary; arm64 CUDA (l4t-style images, DGX
-	# Spark) is GB10 only. Triton-AOT GDN cubins are vendored per-arch, no
-	# Python needed to consume them.
+	# Every CUDA architecture upstream builds that the platform can actually
+	# host, split by where the silicon exists: Jetson (87 Orin, 110 Thor) is
+	# arm64-only, desktop 120a is amd64-only, and 90a/100a appear on both
+	# because of the SBSA parts (GH200, GB200).
+	#
+	# This deliberately matches vllm.cpp's own release archive rather than
+	# narrowing to the boxes we benchmark on. A narrower list does not degrade
+	# on an unlisted card, it dies at the first request with "no kernel image
+	# is available for execution on the device", long after `backends install`
+	# reported success -- so an arch we merely lack numbers for still belongs
+	# in the binary.
+	#
+	# Triton-AOT stays ON for both. A fat build is supported on the BUILDER
+	# path: it embeds every vendored cubin tree (sm_80/86/89/90a/100a/121a) and
+	# selects by exact SM at runtime, so the arches with no tree (87, 103a,
+	# 110, 120a) take the portable CUDA kernels and can never load a
+	# neighbouring cubin. Only maintainer REGEN needs a single pinned arch.
+	# See vllm.cpp cmake/TritonAOT.cmake `_triton_aot_arch_names`.
+	#
+	# CUDA builds REQUIRE the CUDA 13 toolchain: 12.x nvcc lacks compute_121a
+	# (GB10) and its ptxas rejects the sm_120a NVFP4 MMA kernels ("Vector type
+	# too large"), so no cuda-12 variant is shipped.
+	ifeq ($(CUDA_MAJOR_VERSION),12)
+    $(error vllm.cpp needs the CUDA 13 toolchain: CUDA 12.x cannot compile the Blackwell fp4 kernels)
+	endif
 	ifeq ($(UNAME_M),x86_64)
-		# NO -DVLLM_CPP_TRITON on fat builds: the vendored Triton-AOT cubin
-		# trees are per-arch and the engine refuses a multi-arch build unless
-		# pinned to one tree (unsound for the other arch). The non-AOT GDN
-		# path serves the fat binary; single-arch builds keep the cubins.
-		#
-		# CUDA builds REQUIRE the CUDA 13 toolchain: 12.x nvcc lacks
-		# compute_121a (GB10) and its ptxas rejects the sm_120a NVFP4 MMA
-		# kernels ("Vector type too large"), so no cuda-12 variant is shipped.
-		ifeq ($(CUDA_MAJOR_VERSION),12)
-      $(error vllm.cpp needs the CUDA 13 toolchain: CUDA 12.x cannot compile the Blackwell fp4 kernels)
-		endif
-		CMAKE_ARGS+=-DVLLM_CPP_CUDA=ON "-DVLLM_CPP_CUDA_ARCHITECTURES=120a;121a"
+		CMAKE_ARGS+=-DVLLM_CPP_CUDA=ON "-DVLLM_CPP_CUDA_ARCHITECTURES=80;86;89;90a;100a;103a;120a;121a" -DVLLM_CPP_TRITON=ON
 	else
-		CMAKE_ARGS+=-DVLLM_CPP_CUDA=ON -DVLLM_CPP_CUDA_ARCHITECTURES=121a -DVLLM_CPP_TRITON=ON
+		CMAKE_ARGS+=-DVLLM_CPP_CUDA=ON "-DVLLM_CPP_CUDA_ARCHITECTURES=87;90a;100a;110;121a" -DVLLM_CPP_TRITON=ON
 	endif
 else ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DVLLM_CPP_VULKAN=ON -DVLLM_CPP_CUDA=OFF
@@ -106,13 +115,24 @@ else
 	LIB=libvllm.so
 endif
 
-sources/vllm.cpp:
+# patches/ carries fixes the pinned engine SHA does not have yet. `git apply`
+# is deliberately unguarded: a patch that no longer applies must FAIL the clone
+# loudly, because the alternative is a pin that silently ships without a fix it
+# is documented to carry. Each patch header says which pin retires it.
+VLLM_CPP_PATCHES=$(wildcard patches/*.patch)
+
+sources/vllm.cpp: $(VLLM_CPP_PATCHES)
+	rm -rf sources/vllm.cpp
 	mkdir -p sources/vllm.cpp
 	cd sources/vllm.cpp && \
 	git init && \
 	git remote add origin $(VLLM_CPP_REPO) && \
 	git fetch --depth 1 origin $(VLLM_CPP_VERSION) && \
-	git checkout FETCH_HEAD
+	git checkout FETCH_HEAD && \
+	for p in $(VLLM_CPP_PATCHES); do \
+		echo "==> applying $$p"; \
+		git apply ../../$$p || exit 1; \
+	done
 
 ifeq ($(MLX_ENABLED),1)
 # A stamp FILE, not a phony target: a phony prerequisite is always "newer" than
@@ -165,7 +185,7 @@ $(LIB): sources/vllm.cpp $(MLX_STAMP)
 	cmake --build . --config Release -j$(JOBS) --target vllm_shared
 	cp -fL build/$(LIB) ./$(LIB)
 
-vllm-cpp: main.go govllmcpp.go backend.go options.go $(LIB)
+vllm-cpp: main.go govllmcpp.go backend.go chat.go options.go video.go $(LIB)
 	CGO_ENABLED=0 $(GOCMD) build -tags "$(GO_TAGS)" -o vllm-cpp ./
 
 package: vllm-cpp

backend/go/vllm-cpp/README.md

@@ -1,12 +1,15 @@
 # vllm-cpp backend
 
-LocalAI text-generation backend for [vllm.cpp](https://github.com/mudler/vllm.cpp),
+LocalAI backend for [vllm.cpp](https://github.com/mudler/vllm.cpp),
 the LocalAI-team C++20 port of vLLM (paged KV cache, continuous batching,
 safetensors + GGUF loading, CUDA / CPU / Metal / Vulkan) with no Python at
 inference time.
 
+It serves two things: text generation, and MiniMax-H3 joint video+audio
+generation.
+
 The backend dlopens the engine's stable C ABI (`libvllm`, `include/vllm.h`,
-ABI v10) through purego:
+ABI v20) through purego:
 
 - `Load` -> `vllm_engine_load`: accepts a `.gguf` file or a HF-style model
   directory (`config.json` + safetensors). `context_size` maps to
@@ -29,6 +32,12 @@ ABI v10) through purego:
   LocalAI's Go-side grammar-constrained tool calling; JSON-schema / regex /
   choice constraints are also exposed by the ABI.
 
+`patches/` carries fixes the pinned engine SHA does not have yet, applied to
+the clone the same way `longcat-video` patches its upstream. `git apply` is
+unguarded on purpose: a patch that stops applying must fail the clone loudly
+rather than leave a pin silently missing a fix it is documented to carry. Each
+patch header says what retires it.
+
 The struct mirrors in `govllmcpp.go` are hand-written against one ABI version,
 and the engine refuses to load against any other. Moving `VLLM_CPP_VERSION` in
 the Makefile therefore means updating `abiVersion` plus the mirrors (and their
@@ -47,6 +56,68 @@ options:
 - max_num_seqs:16
 ```
 
+## MiniMax-H3 video+audio generation
+
+`GenerateVideo` -> `vllm_video_generate` (ABI v12). H3 renders picture and sound
+together, so the output MP4 carries a real AAC track.
+
+The video engine is a SECOND handle (`vllm_video_engine`), not a mode of the
+text one, because H3 is a checkpoint SET rather than a model directory: the DiT,
+the text encoder and two VAEs are separate artifacts, and vllm.cpp has the two
+loaders refuse each other's checkpoints. `Load` takes the video branch when the
+model config carries any of the video options below; `parameters.model` is the
+DiT and everything else is named in `options:`.
+
+```yaml
+name: minimax-h3-fl2va-q4
+backend: vllm-cpp
+cuda: true
+known_usecases: [video]
+parameters:
+  model: minimax-h3/MiniMax-H3-FL2VA-Q4_K_M.gguf
+options:
+- video_encoder:minimax-h3/qwen3vl-32B-MiniMax-H3-Q4_K_M.gguf
+- video_tokenizer:minimax-h3/tokenizer.json
+- video_vae:minimax-h3/video_vae.safetensors
+- video_vae_config:minimax-h3/video_vae_config.json
+- audio_vae:minimax-h3/audio_vae.safetensors
+- audio_vae_config:minimax-h3/audio_vae_config.json
+- video_partition:fl2va
+- video_device:cuda
+- video_dequant_bf16:true
+- video_width:1344
+- video_height:768
+- video_num_frames:124
+```
+
+Three things are worth knowing before touching this path.
+
+**The partition is declared, not detected, and a mismatch does not fail
+cleanly.** The FL2VA DiT serves `t2va` and `fl2va`; `ref2va` is a different
+checkpoint. The community GGUF/NVFP4 quantisations strip the release metadata
+and the two DiTs are byte-structurally identical, so the engine refuses every
+generate until `video_partition` says which one it has. Handing reference
+conditioning to an FL2VA DiT renders for hours and returns a coloured lattice
+over the frame, so `checkPartitionConditioning` refuses that combination here,
+before the engine is called.
+
+**ffmpeg comes from the host.** libvllm writes the frames and the WAV and
+COMPOSES the mux argv, then spawns nothing — that process boundary is upstream's
+decision. `muxVideo` takes the composed argv, substitutes `argv[0]` with the
+resolved binary and execs it; the backend image is `FROM scratch` and carries no
+ffmpeg, the same arrangement `vibevoice-cpp` uses for transcoding. ffmpeg also
+converts a `start_image`/`end_image` upload into the binary PPM at the exact
+output canvas the engine requires, since libvllm vendors neither an image codec
+nor a resampler.
+
+**It is slow.** Roughly 176 s per denoise step at 1344x768 on a 20-SM device, so
+the 50-step default is hours. Nothing here imposes a deadline.
+
+Geometry mirrors the engine so the two agree: the canvas is truncated onto a
+32-pixel grid, the frame count sits on the 17n+5 grid, and an unspecified canvas
+with a keyframe is derived from that image's aspect on a 768-pixel short edge
+(`MiniMaxH3ResolveShape`, `minimax_h3_planner.cpp`).
+
 ## Apple Silicon: the MLX GEMM provider (ON by default, gated to prefill)
 
 `BUILD_TYPE=metal` builds vllm.cpp's MLX provider for the dense GEMM

backend/go/vllm-cpp/backend.go

@@ -28,7 +28,12 @@ type VllmCpp struct {
 	base.Base
 
 	engine uintptr
-	opts   loadOptions
+	// videoEngine is the MiniMax-H3 handle (ABI v12). It is deliberately a
+	// SECOND handle, not a mode of the first: H3 is a checkpoint set rather
+	// than a model directory, and vllm.cpp has the two loaders refuse each
+	// other's checkpoints. Exactly one of the two is ever non-zero.
+	videoEngine uintptr
+	opts        loadOptions
 }
 
 // Stream registry: the per-request bridge between the C token callback and
@@ -109,6 +114,14 @@ func (v *VllmCpp) Load(opts *pb.ModelOptions) error {
 
 	v.opts = parseOptions(opts)
 
+	// MiniMax-H3 is a checkpoint SET behind its own engine handle, so the
+	// branch is taken before any text-engine knob is resolved. The two loaders
+	// refuse each other's checkpoints, which is why this is decided from the
+	// config rather than probed.
+	if v.opts.video.engaged() {
+		return v.loadVideo(opts, model)
+	}
+
 	// A DFlash draft is a second checkpoint the engine opens by path, and the
 	// engine never downloads one. Resolve it against LocalAI's models directory
 	// now so a repo-id spelling works, and so a missing draft fails here with an
@@ -194,6 +207,10 @@ func (v *VllmCpp) Free() error {
 		vllmEngineFree(v.engine)
 		v.engine = 0
 	}
+	if v.videoEngine != 0 {
+		vllmVideoEngineFree(v.videoEngine)
+		v.videoEngine = 0
+	}
 	return nil
 }
 

backend/go/vllm-cpp/govllmcpp.go

@@ -1,6 +1,6 @@
 package main
 
-// purego bindings for the vllm.cpp stable C ABI (include/vllm.h, ABI v10).
+// purego bindings for the vllm.cpp stable C ABI (include/vllm.h, ABI v23).
 //
 // The structs below are hand-mirrored PODs of the C declarations, with
 // explicit padding so the Go layout matches the C layout on linux/darwin
@@ -21,7 +21,7 @@ import (
 // the header of the VLLM_CPP_VERSION pinned in the Makefile: the build checks
 // the two against each other, because a mismatch is only caught at runtime by
 // registerLib, where it takes the backend down on every load (issue #11379).
-const abiVersion = 10
+const abiVersion = 23
 
 // The ABI's tri-state toggles (enable_prefix_caching ABI v7,
 // enable_jump_forward ABI v10) share one encoding: 0 is NOT "off", it is
@@ -69,8 +69,21 @@ type cModelParams struct {
 	MaxNumBatchedTokens int32   // <= 0 = per-arch default (ABI v9)
 	SchedulingPolicy    uintptr // const char*; NULL = "fcfs" (ABI v9)
 	KVTransferConfig    uintptr // const char* JSON; NULL = no connector (ABI v9)
+	OffloadConfig       uintptr // const char* JSON; NULL = no weight offload
 	EnableJumpForward   int32   // tri-state 0/1/2 (ABI v10)
-	_                   [4]byte // trailing pad to the struct's 8-byte alignment
+	// v14/v16 tail. LocalAI sets none of these (0 is "auto" for the device and
+	// "unset" for both sizing knobs, i.e. the pre-v14 engine byte for byte), but
+	// the fields MUST be mirrored: the C side reads sizeof(vllm_model_params)
+	// bytes off the pointer we hand it, so a Go struct that stopped at
+	// EnableJumpForward would have vllm_engine_load read 24 bytes past our
+	// allocation and size the KV pool from whatever sat there.
+	Device             int32   // 0 auto, 1 cpu, 2 cuda (ABI v14)
+	GPUMemoryUtil      float64 // 0 => 0.92 (ABI v16)
+	KVCacheMemoryBytes int64   // 0 => unset (ABI v16)
+	LanguageModelOnly  int32   // 0 = multimodal inputs enabled (ABI v19)
+	_                  [4]byte
+	LimitMMPerPrompt   uintptr // const char* JSON; NULL = default limits (ABI v19)
+	MMProjPath         uintptr // const char*; NULL = no GGUF projector (ABI v22)
 }
 
 // cSamplingParams mirrors vllm_sampling_params (structured fields included).
@@ -117,6 +130,88 @@ type cCompletion struct {
 	CompletionTokens int32
 }
 
+// ── Video+audio generation (ABI v12, MiniMax-H3) ────────────────────────────
+//
+// A video engine is a SEPARATE handle from vllm_engine: H3 is a checkpoint SET
+// (DiT + text encoder + two VAEs), not one model directory, and the two loaders
+// refuse each other's checkpoints on purpose. Offsets are asserted in
+// video_test.go the same way the text PODs are in vllmcpp_test.go.
+
+// cVideoModelParams mirrors vllm_video_model_params. Nine pointers then three
+// int32s, so only the trailing pad is implicit.
+type cVideoModelParams struct {
+	DitPath            uintptr // const char*
+	EncoderPath        uintptr // const char*
+	TokenizerPath      uintptr // const char*
+	VideoVaePath       uintptr // const char*
+	VideoVaeConfigPath uintptr // const char*
+	AudioVaePath       uintptr // const char*
+	AudioVaeConfigPath uintptr // const char*
+	PromptEmbedsPath   uintptr // const char*
+	Partition          uintptr // const char*; "fl2va" | "ref2va", REQUIRED
+	Device             int32   // 0 cpu, 1 cuda
+	DequantBf16        int32   // 0 keep-quant, 1 dequant/stream bf16
+	Fp4Resident        int32   // NVFP4+cuda: keep FP4 packed, Marlin W4A16
+	_                  [4]byte
+	Family             uintptr // const char*; NULL = detect (ABI v18)
+	ExtraKeys          uintptr // const char* const* (ABI v18)
+	ExtraValues        uintptr // const char* const* (ABI v18)
+	NExtras            int32   // 0 = none (ABI v18)
+	_                  [4]byte // trailing pad to the struct's 8-byte alignment
+}
+
+// cVideoParams mirrors vllm_video_params. `width`/`height` and `num_frames`/
+// `steps` pair up into 8-byte slots; the uint64 seed forces the alignment after
+// them, and the float noise_aug leaves a pad before output_dir.
+type cVideoParams struct {
+	Prompt      uintptr // const char*
+	Width       int32
+	Height      int32
+	NumFrames   int32 // <= 1 => per-task default (124 for t2va/fl2va)
+	Steps       int32 // <= 0 => the H3 default (50)
+	Seed        uint64
+	HasSeed     int32
+	_           [4]byte
+	FirstFrame  uintptr // const char*; fl2va keyframe, binary PPM (P6)
+	LastFrame   uintptr // const char*
+	RefImage    uintptr // const char*; ref2va only
+	RefVideo    uintptr // const char*; ref2va only, a frame_%06d.ppm DIRECTORY
+	RefAudio    uintptr // const char*; ref2va only, 16-bit PCM WAV
+	NoiseAug    float32 // <= 0 => 1.0
+	_           [4]byte
+	OutputDir   uintptr // const char*; REQUIRED
+	ExtraKeys   uintptr // const char* const* (ABI v18)
+	ExtraValues uintptr // const char* const* (ABI v18)
+	NExtras     int32   // 0 = none (ABI v18)
+	_           [4]byte
+}
+
+// cVideoResult mirrors vllm_video_result. Every member is library-allocated and
+// released together by vllm_video_result_free.
+type cVideoResult struct {
+	FrameDir   uintptr // char*, holds frame_%06d.ppm
+	AudioPath  uintptr // char*, 16-bit PCM WAV
+	FrameCount int32
+	Width      int32
+	Height     int32
+	Fps        int32
+	SampleRate int32
+	_          [4]byte
+	MuxArgv    uintptr // char**, NULL-terminated at MuxArgc
+	MuxArgc    int32
+	_          [4]byte
+}
+
+// cVideoMuxParams mirrors vllm_video_mux_params. The library composes the argv;
+// spawning it is the CALLER's job, which is why no ffmpeg lives in libvllm.
+type cVideoMuxParams struct {
+	Frames     uintptr // const char*; printf pattern, dir/frame_%06d.ppm
+	AudioPath  uintptr // const char*; NULL/empty => a silent clip
+	OutputPath uintptr // const char*; the .mp4 to write
+	Fps        int32   // <= 0 => the H3 default (24)
+	Crf        int32   // <= 0 => the library default (18)
+}
+
 // defaultSamplingParams mirrors vllm_sampling_params_default().
 func defaultSamplingParams() cSamplingParams {
 	return cSamplingParams{
@@ -148,6 +243,14 @@ var (
 	vllmLastError      func() string
 	vllmVersion        func() string
 	vllmABIVersion     func() int32
+
+	// Video+audio generation (ABI v12).
+	vllmVideoEngineLoad func(params, out unsafe.Pointer) int32
+	vllmVideoEngineFree func(engine uintptr)
+	vllmVideoGenerate   func(engine uintptr, params, out unsafe.Pointer) int32
+	vllmVideoResultFree func(out unsafe.Pointer)
+	vllmVideoMuxArgv    func(params, outArgv, outArgc unsafe.Pointer) int32
+	vllmVideoMuxArgvFre func(argv uintptr, argc int32)
 )
 
 type libFunc struct {
@@ -175,6 +278,12 @@ func registerLib(libName string) error {
 		{&vllmLastError, "vllm_last_error"},
 		{&vllmVersion, "vllm_version"},
 		{&vllmABIVersion, "vllm_abi_version"},
+		{&vllmVideoEngineLoad, "vllm_video_engine_load"},
+		{&vllmVideoEngineFree, "vllm_video_engine_free"},
+		{&vllmVideoGenerate, "vllm_video_generate"},
+		{&vllmVideoResultFree, "vllm_video_result_free"},
+		{&vllmVideoMuxArgv, "vllm_video_mux_argv"},
+		{&vllmVideoMuxArgvFre, "vllm_video_mux_argv_free"},
 	} {
 		purego.RegisterLibFunc(lf.ptr, lib, lf.name)
 	}
@@ -222,3 +331,19 @@ func goString(p uintptr) string {
 	}
 	return string(unsafe.Slice((*byte)(base), n))
 }
+
+// goStringSlice copies a C `char*` array of n entries. Used for the ffmpeg argv
+// the library composes: it is copied out immediately so the caller can free the
+// C allocation before ever spawning the process.
+func goStringSlice(p uintptr, n int32) []string {
+	if p == 0 || n <= 0 {
+		return nil
+	}
+	//nolint:govet // C-owned pointer handed over by purego, valid for this call
+	entries := unsafe.Slice((**byte)(unsafe.Pointer(p)), int(n)) // #nosec G103 -- C-owned, copied out immediately
+	out := make([]string, 0, n)
+	for _, e := range entries {
+		out = append(out, goString(uintptr(unsafe.Pointer(e)))) // #nosec G103 -- ditto
+	}
+	return out
+}

backend/go/vllm-cpp/options.go

@@ -62,15 +62,106 @@ type loadOptions struct {
 	// Override for the tokenizer_config.json the chat template is read from
 	// (ABI v9). Empty = <model_dir>/tokenizer_config.json.
 	tokenizerConfigPath string
+	// MiniMax-H3 video+audio generation (ABI v12). Present only when the config
+	// carries at least one of its keys; see videoOptions.engaged.
+	video videoOptions
+}
+
+// videoOptions is the MiniMax-H3 checkpoint SET plus its generation defaults.
+//
+// H3 is not one model directory: the DiT, the text encoder and the two VAEs are
+// separate artifacts, which is why vllm.cpp gives video its own engine handle
+// (vllm_video_engine, ABI v12) rather than another vllm_engine. The DiT is the
+// model config's `parameters.model`; everything else arrives through these
+// options, so one gallery entry can name five files.
+//
+// The geometry/frame defaults exist because H3's trained canvas is nothing like
+// the generic /video defaults: 1344x768 at 124 frames is a ~5.2 s clip, and the
+// frame count must sit on the 17n+5 grid. A request that leaves a field unset
+// gets the model's own default from here instead of a canvas the checkpoint was
+// never trained at.
+type videoOptions struct {
+	encoderPath      string // H3-Encoder GGUF or bf16 shard dir
+	tokenizerPath    string // tokenizer.json, needed with an encoder
+	videoVaePath     string
+	videoVaeConfig   string
+	audioVaePath     string
+	audioVaeConfig   string
+	promptEmbedsPath string // fallback conditioning when there is no encoder
+	// The served checkpoint PARTITION. Community GGUF/NVFP4 files strip the
+	// release metadata and the FL2VA/Ref2VA DiTs are byte-structurally
+	// identical, so the engine refuses every generate until it is DECLARED.
+	// "fl2va" serves t2va + fl2va; "ref2va" serves reference conditioning.
+	partition   string
+	device      int32 // 0 cpu, 1 cuda (the ABI's own encoding, no auto slot)
+	deviceSet   bool
+	dequantBf16 int32
+	fp4Resident int32
+	// Per-model generation defaults, applied when the request leaves the field
+	// at 0.
+	width     int32
+	height    int32
+	numFrames int32
+	steps     int32
+	// Where frames + WAV are written. Empty = a temporary directory beside the
+	// requested output, removed once the mux succeeds. Set it to keep the
+	// frame_%06d.ppm runs around (they are what ref2va's ref_video consumes).
+	workdir string
+	// The ffmpeg binary the composed mux argv is exec'd with. Empty = "ffmpeg"
+	// from PATH. libvllm composes the argv and spawns nothing, by design.
+	ffmpeg string
+	crf    int32
+}
+
+// engaged reports whether this config describes an H3 video engine. Load uses
+// it to choose which of the two mutually exclusive engine handles to open: the
+// checkpoints refuse each other, so guessing is not an option, and every key
+// below is meaningless to the text engine.
+func (v videoOptions) engaged() bool {
+	return v.encoderPath != "" || v.tokenizerPath != "" ||
+		v.videoVaePath != "" || v.videoVaeConfig != "" ||
+		v.audioVaePath != "" || v.audioVaeConfig != "" ||
+		v.promptEmbedsPath != "" || v.partition != ""
 }
 
 func parseOptions(opts *pb.ModelOptions) loadOptions {
 	lo := loadOptions{}
 	applyOptionsList(&lo, opts.GetOptions())
 	applyEngineArgs(&lo, opts.GetEngineArgs())
+	applyDraftModelOption(&lo, opts.GetOptions())
 	return lo
 }
 
+// applyDraftModelOption binds a managed companion snapshot after engine_args
+// has supplied the speculative document. Companion paths do not exist until
+// LocalAI materializes the artifact, so they must replace the gallery's static
+// repository reference without disturbing the method or token budget.
+func applyDraftModelOption(lo *loadOptions, options []string) {
+	if strings.TrimSpace(lo.speculativeConfig) == "" {
+		return
+	}
+	var draftModel string
+	for _, option := range options {
+		key, value, found := strings.Cut(option, ":")
+		if found && strings.TrimSpace(key) == "draft_model" {
+			draftModel = strings.TrimSpace(value)
+		}
+	}
+	if draftModel == "" {
+		return
+	}
+
+	var spec map[string]any
+	if err := json.Unmarshal([]byte(lo.speculativeConfig), &spec); err != nil {
+		return
+	}
+	spec["model"] = draftModel
+	encoded, err := json.Marshal(spec)
+	if err == nil {
+		lo.speculativeConfig = string(encoded)
+	}
+}
+
 // applyOptionsList reads the legacy free-form "key:value" list. strings.Cut
 // splits on the FIRST colon only, so a JSON object value survives intact.
 func applyOptionsList(lo *loadOptions, options []string) {
@@ -110,10 +201,94 @@ func applyOptionsList(lo *loadOptions, options []string) {
 			if b, err := strconv.ParseBool(strings.TrimSpace(v)); err == nil {
 				lo.enableJumpForward = boolTriState(b)
 			}
+		default:
+			applyVideoOption(&lo.video, strings.TrimSpace(k), v)
 		}
 	}
 }
 
+// applyVideoOption reads one MiniMax-H3 key. Split out of applyOptionsList so
+// the video surface stays legible next to the videoOptions it fills, and so
+// video_test.go can exercise it directly.
+func applyVideoOption(vo *videoOptions, key, value string) bool {
+	v := strings.TrimSpace(value)
+	switch key {
+	case "video_encoder":
+		vo.encoderPath = v
+	case "video_tokenizer":
+		vo.tokenizerPath = v
+	case "video_vae":
+		vo.videoVaePath = v
+	case "video_vae_config":
+		vo.videoVaeConfig = v
+	case "audio_vae":
+		vo.audioVaePath = v
+	case "audio_vae_config":
+		vo.audioVaeConfig = v
+	case "video_prompt_embeds":
+		vo.promptEmbedsPath = v
+	case "video_partition":
+		vo.partition = strings.ToLower(v)
+	case "video_device":
+		switch strings.ToLower(v) {
+		case "cpu":
+			vo.device, vo.deviceSet = videoDeviceCPU, true
+		case "cuda", "gpu":
+			vo.device, vo.deviceSet = videoDeviceCUDA, true
+		default:
+			xlog.Warn("[vllm-cpp] ignoring unknown video_device", "value", v)
+		}
+	case "video_dequant_bf16":
+		if b, err := strconv.ParseBool(v); err == nil {
+			vo.dequantBf16 = boolInt32(b)
+		}
+	case "video_fp4_resident":
+		if b, err := strconv.ParseBool(v); err == nil {
+			vo.fp4Resident = boolInt32(b)
+		}
+	case "video_width":
+		vo.width = parseInt32(v, vo.width)
+	case "video_height":
+		vo.height = parseInt32(v, vo.height)
+	case "video_num_frames":
+		vo.numFrames = parseInt32(v, vo.numFrames)
+	case "video_steps":
+		vo.steps = parseInt32(v, vo.steps)
+	case "video_workdir":
+		vo.workdir = v
+	case "video_crf":
+		vo.crf = parseInt32(v, vo.crf)
+	case "ffmpeg", "ffmpeg_path":
+		vo.ffmpeg = v
+	default:
+		return false
+	}
+	return true
+}
+
+// videoScalarString renders an engine_args scalar so the video keys can share
+// one parser with the "key:value" list. Objects and arrays have no video
+// meaning and are left to the caller's unknown-key path.
+func videoScalarString(v any) (string, bool) {
+	switch t := v.(type) {
+	case string:
+		return t, true
+	case bool:
+		return strconv.FormatBool(t), true
+	case float64:
+		return strconv.FormatFloat(t, 'f', -1, 64), true
+	default:
+		return "", false
+	}
+}
+
+func boolInt32(b bool) int32 {
+	if b {
+		return 1
+	}
+	return 0
+}
+
 // applyEngineArgs overlays the `engine_args:` JSON object. A document that does
 // not parse is logged and skipped: engine_args is shared with the other engines
 // (the vLLM and SGLang backends read the same field), so a stray key must not
@@ -160,6 +335,9 @@ func applyEngineArgs(lo *loadOptions, engineArgs string) {
 				lo.enableJumpForward = boolTriState(b)
 			}
 		default:
+			if s, ok := videoScalarString(v); ok && applyVideoOption(&lo.video, k, s) {
+				continue
+			}
 			xlog.Debug("[vllm-cpp] ignoring unknown engine_args key", "key", k)
 		}
 	}

backend/go/vllm-cpp/options_test.go

@@ -0,0 +1,38 @@
+package main
+
+import (
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+)
+
+var _ = Describe("managed DFlash companion options", func() {
+	It("replaces only the draft model in an existing speculative configuration", func() {
+		managedPath := ".artifacts/huggingface/0123456789abcdef/snapshot"
+		lo := parseOptions(&pb.ModelOptions{
+			Options: []string{"draft_model:" + managedPath},
+			EngineArgs: `{
+				"speculative_config": {
+					"method": "dflash",
+					"model": "Mia-AiLab/Qwen3.8-27B-DFlash2-EXL3-5.0bpw",
+					"num_speculative_tokens": 7
+				}
+			}`,
+		})
+
+		Expect(lo.speculativeConfig).To(MatchJSON(`{
+			"method": "dflash",
+			"model": ".artifacts/huggingface/0123456789abcdef/snapshot",
+			"num_speculative_tokens": 7
+		}`))
+	})
+
+	It("ignores a draft companion when speculative decoding is not configured", func() {
+		lo := parseOptions(&pb.ModelOptions{
+			Options: []string{"draft_model:.artifacts/huggingface/0123456789abcdef/snapshot"},
+		})
+
+		Expect(lo.speculativeConfig).To(BeEmpty())
+	})
+})

backend/go/vllm-cpp/video.go

@@ -0,0 +1,634 @@
+package main
+
+// MiniMax-H3 video+audio generation over the vllm.cpp C ABI (v12).
+//
+// Two things make this different from the text path, and both come from the
+// engine's own shape rather than from LocalAI:
+//
+//  1. A video engine is loaded from a checkpoint SET - the DiT, the text
+//     encoder and two VAEs are separate artifacts - so it is its own handle
+//     (vllm_video_engine) and its own Load branch. The two loaders refuse each
+//     other's checkpoints on purpose.
+//  2. libvllm writes frames + a WAV and COMPOSES the ffmpeg argv, but spawns
+//     nothing. That process boundary is deliberate upstream, so the mux lives
+//     here: we take the composed argv, substitute argv[0], and exec it. ffmpeg
+//     comes from PATH the same way the vibevoice-cpp backend takes it.
+//
+// Generation is SLOW - roughly 176 s per denoise step at 1344x768 on a 20-SM
+// device, so a default 50-step render is hours, not seconds. Nothing here
+// imposes a deadline: GenerateVideo blocks for as long as the engine needs and
+// the gRPC call carries LocalAI's application context.
+
+import (
+	"fmt"
+	"image"
+	"math"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"strings"
+	"unsafe"
+
+	// Registered for image.DecodeConfig only: a staged keyframe arrives as
+	// whatever the caller uploaded, and we need its geometry to size the canvas.
+	_ "image/gif"
+	_ "image/jpeg"
+	_ "image/png"
+
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	"github.com/mudler/xlog"
+)
+
+// vllm_video_model_params.device (vllm.h): no auto slot, unlike the text
+// engine's v14 device field.
+const (
+	videoDeviceCPU  int32 = 0
+	videoDeviceCUDA int32 = 1
+)
+
+// H3's shipped geometry. The canvas is truncated onto a 32-pixel grid and the
+// frame count onto the 17n+5 grid by the engine itself
+// (MiniMaxH3ResolveShape / MiniMaxH3AlignFrameCount in
+// src/vllm/model_executor/models/minimax_h3_planner.cpp); mirrored here only so
+// a keyframe can be resampled to the exact canvas the engine will render at.
+const (
+	h3CanvasMultiple int32 = 32
+	h3FrameGrid      int32 = 17
+	h3FrameOffset    int32 = 5
+	h3ShortEdge      int32 = 768
+)
+
+// videoPartitions are the two DECLARED partitions of the H3 release. The FL2VA
+// checkpoint serves t2va and fl2va; ref2va is a different checkpoint. Passing
+// reference conditioning against an fl2va DiT is a partition mismatch that
+// renders a coloured lattice over the frame rather than failing cleanly, which
+// is why it is refused here before the engine is ever called.
+const (
+	partitionFL2VA  = "fl2va"
+	partitionRef2VA = "ref2va"
+)
+
+// videoRequestParams are the per-request `params` keys this backend accepts.
+// Unknown keys are an error rather than a silent drop: a misspelled reference
+// path would otherwise produce a perfectly successful render of the wrong
+// thing, hours later.
+var videoRequestParams = []string{"noise_aug", "ref_image", "ref_video", "crf"}
+
+// loadVideo opens the H3 checkpoint set. `dit` is the model config's
+// parameters.model; every other artifact comes from the options.
+func (v *VllmCpp) loadVideo(opts *pb.ModelOptions, dit string) error {
+	vo := &v.opts.video
+
+	// Relative option paths resolve against LocalAI's models directory, which
+	// is where the gallery lands the five H3 files.
+	resolve := func(p string) string {
+		if p == "" || filepath.IsAbs(p) || opts.ModelPath == "" {
+			return p
+		}
+		return filepath.Join(opts.ModelPath, p)
+	}
+	vo.encoderPath = resolve(vo.encoderPath)
+	vo.tokenizerPath = resolve(vo.tokenizerPath)
+	vo.videoVaePath = resolve(vo.videoVaePath)
+	vo.videoVaeConfig = resolve(vo.videoVaeConfig)
+	vo.audioVaePath = resolve(vo.audioVaePath)
+	vo.audioVaeConfig = resolve(vo.audioVaeConfig)
+	vo.promptEmbedsPath = resolve(vo.promptEmbedsPath)
+	vo.workdir = resolve(vo.workdir)
+
+	// A VAE config carries the per-channel latents_mean/latents_std and the
+	// temporal clip_length/token_drop; decode is wrong without it. The release
+	// ships it beside the weights, so default to that rather than making every
+	// config repeat it.
+	if vo.videoVaeConfig == "" && vo.videoVaePath != "" {
+		vo.videoVaeConfig = siblingConfigJSON(vo.videoVaePath)
+	}
+	if vo.audioVaeConfig == "" && vo.audioVaePath != "" {
+		vo.audioVaeConfig = siblingConfigJSON(vo.audioVaePath)
+	}
+
+	if vo.partition == "" {
+		// The community GGUF/NVFP4 quantisations strip the release metadata and
+		// the two DiTs are byte-structurally identical, so the engine cannot
+		// infer this and refuses every generate until it is declared. The
+		// shipped FL2VA checkpoint is the one the gallery entry installs.
+		vo.partition = partitionFL2VA
+		xlog.Warn("[vllm-cpp] video partition not declared, assuming the FL2VA checkpoint",
+			"hint", "set options: [video_partition:fl2va] or [video_partition:ref2va] to match the DiT you installed")
+	}
+	if vo.partition != partitionFL2VA && vo.partition != partitionRef2VA {
+		return fmt.Errorf("vllm-cpp: video_partition must be %q or %q, got %q",
+			partitionFL2VA, partitionRef2VA, vo.partition)
+	}
+	if vo.videoVaePath == "" || vo.audioVaePath == "" {
+		return fmt.Errorf("vllm-cpp: MiniMax-H3 needs both VAEs: set options: " +
+			"[video_vae:<video vae .safetensors>, audio_vae:<audio vae .safetensors>]")
+	}
+	if vo.encoderPath == "" && vo.promptEmbedsPath == "" {
+		return fmt.Errorf("vllm-cpp: MiniMax-H3 needs text conditioning: set options: " +
+			"[video_encoder:<encoder .gguf>, video_tokenizer:<tokenizer.json>] " +
+			"or [video_prompt_embeds:<f32 embeddings>]")
+	}
+	if !vo.deviceSet && opts.GetCUDA() {
+		vo.device = videoDeviceCUDA
+	}
+
+	mp := cVideoModelParams{
+		Device:      vo.device,
+		DequantBf16: vo.dequantBf16,
+		Fp4Resident: vo.fp4Resident,
+	}
+	var keep [][]byte
+	setStr := func(dst *uintptr, s string) {
+		if s == "" {
+			return
+		}
+		b := cString(s)
+		keep = append(keep, b)
+		*dst = uintptr(unsafe.Pointer(&b[0])) // #nosec G103 -- borrowed by C for the load call only
+	}
+	setStr(&mp.DitPath, dit)
+	setStr(&mp.EncoderPath, vo.encoderPath)
+	setStr(&mp.TokenizerPath, vo.tokenizerPath)
+	setStr(&mp.VideoVaePath, vo.videoVaePath)
+	setStr(&mp.VideoVaeConfigPath, vo.videoVaeConfig)
+	setStr(&mp.AudioVaePath, vo.audioVaePath)
+	setStr(&mp.AudioVaeConfigPath, vo.audioVaeConfig)
+	setStr(&mp.PromptEmbedsPath, vo.promptEmbedsPath)
+	setStr(&mp.Partition, vo.partition)
+
+	xlog.Info("[vllm-cpp] Load (MiniMax-H3 video)", "dit", dit, "engine", vllmVersion(),
+		"encoder", vo.encoderPath, "tokenizer", vo.tokenizerPath,
+		"videoVae", vo.videoVaePath, "audioVae", vo.audioVaePath,
+		"partition", vo.partition, "device", videoDeviceName(vo.device),
+		"dequantBf16", vo.dequantBf16 == 1, "fp4Resident", vo.fp4Resident == 1)
+
+	var engine uintptr
+	rc := vllmVideoEngineLoad(unsafe.Pointer(&mp), unsafe.Pointer(&engine)) // #nosec G103 -- POD out-params
+	runtime.KeepAlive(keep)
+	if rc != vllmOK {
+		return fmt.Errorf("vllm-cpp: video engine load failed: %s", vllmLastError())
+	}
+	v.videoEngine = engine
+	return nil
+}
+
+// GenerateVideo renders one clip and muxes it to opts.Dst as an MP4 carrying
+// H3's jointly generated AAC audio track. It blocks for the whole render.
+func (v *VllmCpp) GenerateVideo(opts *pb.GenerateVideoRequest) error {
+	if v.videoEngine == 0 {
+		return fmt.Errorf("vllm-cpp: this model is not a MiniMax-H3 video engine " +
+			"(load it with the video_vae / audio_vae / video_encoder options)")
+	}
+	if strings.TrimSpace(opts.GetPrompt()) == "" {
+		return fmt.Errorf("vllm-cpp: video generation needs a prompt")
+	}
+	dst := opts.GetDst()
+	if dst == "" {
+		return fmt.Errorf("vllm-cpp: video generation needs an output path")
+	}
+	vo := v.opts.video
+
+	extra, err := parseVideoRequestParams(opts.GetParams())
+	if err != nil {
+		return err
+	}
+	if err := checkPartitionConditioning(vo.partition, opts, extra); err != nil {
+		return err
+	}
+	if opts.GetNegativePrompt() != "" {
+		xlog.Warn("[vllm-cpp] MiniMax-H3 has no negative prompt; ignoring it")
+	}
+	if opts.GetCfgScale() != 0 {
+		xlog.Warn("[vllm-cpp] MiniMax-H3 has no classifier-free guidance scale; ignoring cfg_scale")
+	}
+
+	workdir, cleanup, err := v.videoWorkdir(dst)
+	if err != nil {
+		return err
+	}
+	defer cleanup()
+
+	width, height := firstPositive(opts.GetWidth(), vo.width), firstPositive(opts.GetHeight(), vo.height)
+	frames := firstPositive(opts.GetNumFrames(), vo.numFrames)
+	steps := firstPositive(opts.GetStep(), vo.steps)
+
+	vp := cVideoParams{
+		NumFrames: frames,
+		Steps:     steps,
+		NoiseAug:  extra.noiseAug,
+	}
+	if opts.GetSeed() > 0 {
+		vp.Seed = uint64(opts.GetSeed())
+		vp.HasSeed = 1
+	}
+	if aligned := alignFrameCount(frames); aligned != frames {
+		xlog.Warn("[vllm-cpp] frame count is not on H3's 17n+5 grid; the engine rounds up",
+			"requested", frames, "rendered", aligned)
+	}
+
+	// Keyframes must be binary PPM (P6) at the exact output canvas: no image
+	// codec and no resampler is vendored in libvllm. Resolve the canvas first,
+	// then stage the frames through ffmpeg into it.
+	//
+	// The REQUEST's geometry is what is honoured here, not the model-level
+	// default: that default is a t2va canvas, and applying it to a keyframe
+	// would stretch a portrait photo into a 1344x768 letterbox. With no
+	// requested geometry the canvas comes from the keyframe's own aspect, which
+	// is the rule the engine itself applies (MiniMaxH3ResolveShape).
+	first, last := opts.GetStartImage(), opts.GetEndImage()
+	if first != "" || last != "" {
+		width, height, err = resolveCanvas(opts.GetWidth(), opts.GetHeight(), first, last)
+		if err != nil {
+			return err
+		}
+		if first, err = stageKeyframe(vo.ffmpeg, first, width, height, workdir, "first"); err != nil {
+			return err
+		}
+		if last, err = stageKeyframe(vo.ffmpeg, last, width, height, workdir, "last"); err != nil {
+			return err
+		}
+	}
+	vp.Width, vp.Height = truncateToGrid(width), truncateToGrid(height)
+
+	var keep [][]byte
+	setStr := func(dst *uintptr, s string) {
+		if s == "" {
+			return
+		}
+		b := cString(s)
+		keep = append(keep, b)
+		*dst = uintptr(unsafe.Pointer(&b[0])) // #nosec G103 -- borrowed by C for the call only
+	}
+	setStr(&vp.Prompt, opts.GetPrompt())
+	setStr(&vp.OutputDir, workdir)
+	setStr(&vp.FirstFrame, first)
+	setStr(&vp.LastFrame, last)
+	setStr(&vp.RefImage, extra.refImage)
+	setStr(&vp.RefVideo, extra.refVideo)
+	setStr(&vp.RefAudio, opts.GetAudio())
+
+	xlog.Info("[vllm-cpp] GenerateVideo", "dst", dst, "workdir", workdir,
+		"width", vp.Width, "height", vp.Height, "frames", vp.NumFrames,
+		"steps", vp.Steps, "seeded", vp.HasSeed == 1, "partition", vo.partition)
+
+	var out cVideoResult
+	rc := vllmVideoGenerate(v.videoEngine, unsafe.Pointer(&vp), unsafe.Pointer(&out)) // #nosec G103 -- POD in/out params
+	runtime.KeepAlive(keep)
+	if rc != vllmOK {
+		return fmt.Errorf("vllm-cpp: video generation failed: %s", vllmLastError())
+	}
+	defer vllmVideoResultFree(unsafe.Pointer(&out)) // #nosec G103 -- frees the library-owned members
+
+	frameDir, audioPath := goString(out.FrameDir), goString(out.AudioPath)
+	xlog.Info("[vllm-cpp] rendered", "frames", out.FrameCount,
+		"width", out.Width, "height", out.Height, "fps", out.Fps,
+		"audio", audioPath, "sampleRate", out.SampleRate)
+	if opts.GetFps() > 0 && opts.GetFps() != out.Fps {
+		// Muxing at any other rate desynchronises the jointly generated audio.
+		xlog.Warn("[vllm-cpp] MiniMax-H3 renders at a fixed frame rate; ignoring the requested fps",
+			"requested", opts.GetFps(), "rendered", out.Fps)
+	}
+
+	return v.muxVideo(frameDir, audioPath, dst, out.Fps, extra.crf)
+}
+
+// muxVideo execs the argv libvllm composed. The encoding contract (h264 /
+// yuv420p + AAC, -shortest, +faststart) belongs to the library; only the spawn
+// is ours.
+func (v *VllmCpp) muxVideo(frameDir, audioPath, dst string, fps, crf int32) error {
+	mx := cVideoMuxParams{Fps: fps, Crf: crf}
+	var keep [][]byte
+	setStr := func(dst *uintptr, s string) {
+		if s == "" {
+			return
+		}
+		b := cString(s)
+		keep = append(keep, b)
+		*dst = uintptr(unsafe.Pointer(&b[0])) // #nosec G103 -- borrowed by C for the call only
+	}
+	setStr(&mx.Frames, filepath.Join(frameDir, "frame_%06d.ppm"))
+	setStr(&mx.AudioPath, audioPath)
+	setStr(&mx.OutputPath, dst)
+
+	var argvPtr uintptr
+	var argc int32
+	rc := vllmVideoMuxArgv(unsafe.Pointer(&mx), unsafe.Pointer(&argvPtr), unsafe.Pointer(&argc)) // #nosec G103 -- POD out-params
+	runtime.KeepAlive(keep)
+	if rc != vllmOK {
+		return fmt.Errorf("vllm-cpp: composing the mux command failed: %s", vllmLastError())
+	}
+	argv := goStringSlice(argvPtr, argc)
+	vllmVideoMuxArgvFre(argvPtr, argc)
+	if len(argv) == 0 {
+		return fmt.Errorf("vllm-cpp: the library composed an empty mux command")
+	}
+
+	ffmpegBin, err := resolveFfmpeg(v.opts.video.ffmpeg)
+	if err != nil {
+		return err
+	}
+	argv[0] = ffmpegBin
+
+	xlog.Debug("[vllm-cpp] muxing", "argv", argv)
+	output, err := exec.Command(argv[0], argv[1:]...).CombinedOutput() // #nosec G204 -- argv is composed by libvllm, argv[0] is a resolved binary
+	if err != nil {
+		return fmt.Errorf("vllm-cpp: ffmpeg mux failed: %w (output: %s)", err, strings.TrimSpace(string(output)))
+	}
+	return nil
+}
+
+// resolveFfmpeg locates the mux binary. The backend image is FROM scratch and
+// carries no ffmpeg, exactly like vibevoice-cpp's transcode path: the host must
+// provide one, and saying so plainly beats a bare "exec: not found" after an
+// hours-long render.
+func resolveFfmpeg(configured string) (string, error) {
+	name := configured
+	if name == "" {
+		name = "ffmpeg"
+	}
+	path, err := exec.LookPath(name)
+	if err != nil {
+		return "", fmt.Errorf("vllm-cpp: %q not found: MiniMax-H3 output is muxed with ffmpeg, "+
+			"install it on the host or point options: [ffmpeg:<path>] at a binary: %w", name, err)
+	}
+	return path, nil
+}
+
+// videoWorkdir returns the directory the engine writes frame_%06d.ppm and
+// audio.wav into, plus its cleanup.
+//
+// It is ALWAYS a fresh directory. Reusing one would leave a longer previous
+// run's trailing frames in place for the mux to pick up, silently splicing two
+// renders together. With video_workdir set the run is kept (its frames are what
+// ref2va's ref_video consumes); otherwise it is removed once the mux succeeds.
+func (v *VllmCpp) videoWorkdir(dst string) (string, func(), error) {
+	parent := v.opts.video.workdir
+	keep := parent != ""
+	if parent == "" {
+		parent = filepath.Dir(dst)
+	}
+	if err := os.MkdirAll(parent, 0o750); err != nil {
+		return "", nil, fmt.Errorf("vllm-cpp: creating the video work directory: %w", err)
+	}
+	dir, err := os.MkdirTemp(parent, "vllm-cpp-h3-")
+	if err != nil {
+		return "", nil, fmt.Errorf("vllm-cpp: creating the video work directory: %w", err)
+	}
+	if keep {
+		return dir, func() {}, nil
+	}
+	return dir, func() {
+		if err := os.RemoveAll(dir); err != nil {
+			xlog.Warn("[vllm-cpp] could not remove the video work directory", "dir", dir, "error", err)
+		}
+	}, nil
+}
+
+// videoExtraParams holds the per-request knobs that have no proto field.
+type videoExtraParams struct {
+	noiseAug float32
+	refImage string
+	refVideo string
+	crf      int32
+}
+
+func parseVideoRequestParams(params map[string]string) (videoExtraParams, error) {
+	var extra videoExtraParams
+	for k, raw := range params {
+		v := strings.TrimSpace(raw)
+		switch k {
+		case "noise_aug":
+			f, err := strconv.ParseFloat(v, 32)
+			if err != nil {
+				return extra, fmt.Errorf("vllm-cpp: params.noise_aug must be a number, got %q", raw)
+			}
+			extra.noiseAug = float32(f)
+		case "ref_image":
+			extra.refImage = v
+		case "ref_video":
+			extra.refVideo = v
+		case "crf":
+			n, err := strconv.ParseInt(v, 10, 32)
+			if err != nil {
+				return extra, fmt.Errorf("vllm-cpp: params.crf must be an integer, got %q", raw)
+			}
+			extra.crf = int32(n)
+		default:
+			return extra, fmt.Errorf("vllm-cpp: unknown params key %q (accepted: %s)",
+				k, strings.Join(videoRequestParams, ", "))
+		}
+	}
+	return extra, nil
+}
+
+// checkPartitionConditioning refuses conditioning the loaded checkpoint cannot
+// serve.
+//
+// This is the failure this backend most needs to catch early. The FL2VA
+// partition serves t2va and fl2va; handing it a reference image or audio is a
+// partition mismatch, and H3 does not fail cleanly on one - it renders, for
+// hours, and returns a coloured lattice over the frame. The engine's own #77
+// guard covers a missing declaration; this covers a declaration that does not
+// match the request.
+func checkPartitionConditioning(partition string, opts *pb.GenerateVideoRequest, extra videoExtraParams) error {
+	hasKeyframe := opts.GetStartImage() != "" || opts.GetEndImage() != ""
+	hasReference := extra.refImage != "" || extra.refVideo != "" || opts.GetAudio() != ""
+
+	if hasKeyframe && hasReference {
+		return fmt.Errorf("vllm-cpp: fl2va keyframes (start_image/end_image) and ref2va reference " +
+			"conditioning (params.ref_image/params.ref_video/audio) are exclusive in the H3 pipeline")
+	}
+	switch partition {
+	case partitionFL2VA:
+		if hasReference {
+			return fmt.Errorf("vllm-cpp: the FL2VA checkpoint serves t2va and fl2va only - " +
+				"reference conditioning (params.ref_image/params.ref_video/audio) needs a ref2va DiT. " +
+				"Use start_image for first-frame conditioning instead")
+		}
+	case partitionRef2VA:
+		if hasKeyframe {
+			return fmt.Errorf("vllm-cpp: the Ref2VA checkpoint does not serve fl2va keyframes - " +
+				"pass the image as params.ref_image, or install the FL2VA checkpoint")
+		}
+	}
+	return nil
+}
+
+// resolveCanvas settles the output geometry BEFORE a keyframe is resampled,
+// because the two have to agree exactly: the engine refuses a keyframe that is
+// not already at the output resolution, and when no geometry is requested it
+// derives one from the keyframe's own aspect. Mirrors _resolve_shape
+// (src/vllm/model_executor/models/minimax_h3_planner.cpp:264-308).
+func resolveCanvas(width, height int32, keyframes ...string) (int32, int32, error) {
+	if width > 0 && height > 0 {
+		return width, height, nil
+	}
+	for _, k := range keyframes {
+		if k == "" {
+			continue
+		}
+		w, h, err := imageDimensions(k)
+		if err != nil {
+			return 0, 0, err
+		}
+		if w <= 0 || h <= 0 {
+			continue
+		}
+		// A 768 short edge, the long edge snapped onto the 32 grid.
+		if w >= h {
+			return alignMultiple(float64(h3ShortEdge)*float64(w)/float64(h), h3CanvasMultiple), h3ShortEdge, nil
+		}
+		return h3ShortEdge, alignMultiple(float64(h3ShortEdge)*float64(h)/float64(w), h3CanvasMultiple), nil
+	}
+	// The shipped canvas.
+	return 1344, h3ShortEdge, nil
+}
+
+// stageKeyframe converts a staged upload into the binary PPM (P6) at exactly
+// width x height that the engine requires. libvllm vendors no image codec and
+// no resampler, so ffmpeg does both; a P6 already at the canvas passes through
+// untouched.
+func stageKeyframe(ffmpegPath, src string, width, height int32, workdir, name string) (string, error) {
+	if src == "" {
+		return "", nil
+	}
+	if w, h, err := ppmDimensions(src); err == nil && w == width && h == height {
+		return src, nil
+	}
+	ffmpegBin, err := resolveFfmpeg(ffmpegPath)
+	if err != nil {
+		return "", fmt.Errorf("converting the %s keyframe to PPM: %w", name, err)
+	}
+	out := filepath.Join(workdir, name+"_frame.ppm")
+	// -frames:v 1 because an animated upload (GIF) would otherwise write a
+	// sequence; -pix_fmt rgb24 is what the image2/ppm muxer needs for P6.
+	cmd := exec.Command(ffmpegBin, "-y", "-loglevel", "error", "-i", src, // #nosec G204 -- the binary is resolved, the rest are literals and staged paths
+		"-frames:v", "1",
+		"-vf", fmt.Sprintf("scale=%d:%d", width, height),
+		"-pix_fmt", "rgb24", "-f", "image2", out)
+	if output, err := cmd.CombinedOutput(); err != nil {
+		return "", fmt.Errorf("vllm-cpp: converting the %s keyframe to PPM failed: %w (output: %s)",
+			name, err, strings.TrimSpace(string(output)))
+	}
+	return out, nil
+}
+
+// imageDimensions reads geometry from a staged upload, PPM included (the Go
+// standard library has no netpbm decoder).
+func imageDimensions(path string) (int32, int32, error) {
+	if w, h, err := ppmDimensions(path); err == nil {
+		return w, h, nil
+	}
+	f, err := os.Open(path) // #nosec G304 -- a path staged by LocalAI for this request
+	if err != nil {
+		return 0, 0, fmt.Errorf("vllm-cpp: reading the keyframe %q: %w", path, err)
+	}
+	defer func() { _ = f.Close() }()
+	cfg, _, err := image.DecodeConfig(f)
+	if err != nil {
+		return 0, 0, fmt.Errorf("vllm-cpp: the keyframe %q is not a PNG, JPEG, GIF or binary PPM: %w", path, err)
+	}
+	return int32(cfg.Width), int32(cfg.Height), nil
+}
+
+// ppmDimensions parses a binary PPM (P6) header: magic, then width, height and
+// maxval as ASCII decimals separated by whitespace, with # comments allowed.
+func ppmDimensions(path string) (int32, int32, error) {
+	f, err := os.Open(path) // #nosec G304 -- a path staged by LocalAI for this request
+	if err != nil {
+		return 0, 0, err
+	}
+	defer func() { _ = f.Close() }()
+
+	// A P6 header is a handful of bytes; 512 covers any sane comment run.
+	buf := make([]byte, 512)
+	n, err := f.Read(buf)
+	if n < 2 || (err != nil && n == 0) {
+		return 0, 0, fmt.Errorf("not a PPM")
+	}
+	if buf[0] != 'P' || buf[1] != '6' {
+		return 0, 0, fmt.Errorf("not a binary PPM (P6)")
+	}
+	fields := make([]int32, 0, 2)
+	for i := 2; i < n && len(fields) < 2; {
+		switch {
+		case buf[i] == '#':
+			for i < n && buf[i] != '\n' {
+				i++
+			}
+		case buf[i] >= '0' && buf[i] <= '9':
+			value := int32(0)
+			for i < n && buf[i] >= '0' && buf[i] <= '9' {
+				value = value*10 + int32(buf[i]-'0')
+				i++
+			}
+			fields = append(fields, value)
+		default:
+			i++
+		}
+	}
+	if len(fields) < 2 {
+		return 0, 0, fmt.Errorf("truncated PPM header")
+	}
+	return fields[0], fields[1], nil
+}
+
+// alignMultiple mirrors MiniMaxH3AlignMultiple: round-half-to-even onto the
+// multiple, floored at one multiple. Half-to-even, not half-away-from-zero,
+// because the reference pipeline uses Python's round().
+func alignMultiple(value float64, multiple int32) int32 {
+	snapped := int32(math.RoundToEven(value/float64(multiple))) * multiple
+	if snapped < multiple {
+		return multiple
+	}
+	return snapped
+}
+
+// truncateToGrid mirrors the engine's canvas snap: truncation, not rounding.
+func truncateToGrid(v int32) int32 {
+	if v <= 0 {
+		return 0
+	}
+	return v / h3CanvasMultiple * h3CanvasMultiple
+}
+
+// alignFrameCount mirrors MiniMaxH3AlignFrameCount: the next value on the
+// 17n+5 grid. Used only to warn - the engine does the real alignment.
+func alignFrameCount(frames int32) int32 {
+	if frames <= 0 {
+		return frames
+	}
+	for frames%h3FrameGrid != h3FrameOffset {
+		frames++
+	}
+	return frames
+}
+
+func firstPositive(values ...int32) int32 {
+	for _, v := range values {
+		if v > 0 {
+			return v
+		}
+	}
+	return 0
+}
+
+func videoDeviceName(device int32) string {
+	if device == videoDeviceCUDA {
+		return "cuda"
+	}
+	return "cpu"
+}
+
+// siblingConfigJSON is the release layout: each VAE ships its config.json in
+// the directory holding its weights.
+func siblingConfigJSON(weights string) string {
+	candidate := filepath.Join(filepath.Dir(weights), "config.json")
+	if _, err := os.Stat(candidate); err != nil {
+		return ""
+	}
+	return candidate
+}

backend/go/vllm-cpp/video_test.go

@@ -0,0 +1,305 @@
+package main
+
+import (
+	"os"
+	"path/filepath"
+	"unsafe"
+
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+// The video PODs carry the same contract as the text ones in vllmcpp_test.go:
+// these are the C offsets of vllm.h on LP64, and a drift here is silent memory
+// corruption rather than a compile error.
+var _ = Describe("C ABI video struct mirrors", func() {
+	It("cVideoModelParams matches vllm_video_model_params", func() {
+		var p cVideoModelParams
+		Expect(unsafe.Offsetof(p.DitPath)).To(Equal(uintptr(0)))
+		Expect(unsafe.Offsetof(p.EncoderPath)).To(Equal(uintptr(8)))
+		Expect(unsafe.Offsetof(p.TokenizerPath)).To(Equal(uintptr(16)))
+		Expect(unsafe.Offsetof(p.VideoVaePath)).To(Equal(uintptr(24)))
+		Expect(unsafe.Offsetof(p.VideoVaeConfigPath)).To(Equal(uintptr(32)))
+		Expect(unsafe.Offsetof(p.AudioVaePath)).To(Equal(uintptr(40)))
+		Expect(unsafe.Offsetof(p.AudioVaeConfigPath)).To(Equal(uintptr(48)))
+		Expect(unsafe.Offsetof(p.PromptEmbedsPath)).To(Equal(uintptr(56)))
+		Expect(unsafe.Offsetof(p.Partition)).To(Equal(uintptr(64)))
+		Expect(unsafe.Offsetof(p.Device)).To(Equal(uintptr(72)))
+		Expect(unsafe.Offsetof(p.DequantBf16)).To(Equal(uintptr(76)))
+		Expect(unsafe.Offsetof(p.Fp4Resident)).To(Equal(uintptr(80)))
+		Expect(unsafe.Offsetof(p.Family)).To(Equal(uintptr(88)))
+		Expect(unsafe.Offsetof(p.ExtraKeys)).To(Equal(uintptr(96)))
+		Expect(unsafe.Offsetof(p.ExtraValues)).To(Equal(uintptr(104)))
+		Expect(unsafe.Offsetof(p.NExtras)).To(Equal(uintptr(112)))
+		Expect(unsafe.Sizeof(p)).To(Equal(uintptr(120)))
+	})
+
+	It("cVideoParams matches vllm_video_params", func() {
+		var p cVideoParams
+		Expect(unsafe.Offsetof(p.Prompt)).To(Equal(uintptr(0)))
+		Expect(unsafe.Offsetof(p.Width)).To(Equal(uintptr(8)))
+		Expect(unsafe.Offsetof(p.Height)).To(Equal(uintptr(12)))
+		Expect(unsafe.Offsetof(p.NumFrames)).To(Equal(uintptr(16)))
+		Expect(unsafe.Offsetof(p.Steps)).To(Equal(uintptr(20)))
+		Expect(unsafe.Offsetof(p.Seed)).To(Equal(uintptr(24)))
+		Expect(unsafe.Offsetof(p.HasSeed)).To(Equal(uintptr(32)))
+		Expect(unsafe.Offsetof(p.FirstFrame)).To(Equal(uintptr(40)))
+		Expect(unsafe.Offsetof(p.LastFrame)).To(Equal(uintptr(48)))
+		Expect(unsafe.Offsetof(p.RefImage)).To(Equal(uintptr(56)))
+		Expect(unsafe.Offsetof(p.RefVideo)).To(Equal(uintptr(64)))
+		Expect(unsafe.Offsetof(p.RefAudio)).To(Equal(uintptr(72)))
+		Expect(unsafe.Offsetof(p.NoiseAug)).To(Equal(uintptr(80)))
+		Expect(unsafe.Offsetof(p.OutputDir)).To(Equal(uintptr(88)))
+		Expect(unsafe.Offsetof(p.ExtraKeys)).To(Equal(uintptr(96)))
+		Expect(unsafe.Offsetof(p.ExtraValues)).To(Equal(uintptr(104)))
+		Expect(unsafe.Offsetof(p.NExtras)).To(Equal(uintptr(112)))
+		Expect(unsafe.Sizeof(p)).To(Equal(uintptr(120)))
+	})
+
+	It("cVideoResult matches vllm_video_result", func() {
+		var r cVideoResult
+		Expect(unsafe.Offsetof(r.FrameDir)).To(Equal(uintptr(0)))
+		Expect(unsafe.Offsetof(r.AudioPath)).To(Equal(uintptr(8)))
+		Expect(unsafe.Offsetof(r.FrameCount)).To(Equal(uintptr(16)))
+		Expect(unsafe.Offsetof(r.Width)).To(Equal(uintptr(20)))
+		Expect(unsafe.Offsetof(r.Height)).To(Equal(uintptr(24)))
+		Expect(unsafe.Offsetof(r.Fps)).To(Equal(uintptr(28)))
+		Expect(unsafe.Offsetof(r.SampleRate)).To(Equal(uintptr(32)))
+		Expect(unsafe.Offsetof(r.MuxArgv)).To(Equal(uintptr(40)))
+		Expect(unsafe.Offsetof(r.MuxArgc)).To(Equal(uintptr(48)))
+		Expect(unsafe.Sizeof(r)).To(Equal(uintptr(56)))
+	})
+
+	It("cVideoMuxParams matches vllm_video_mux_params", func() {
+		var p cVideoMuxParams
+		Expect(unsafe.Offsetof(p.Frames)).To(Equal(uintptr(0)))
+		Expect(unsafe.Offsetof(p.AudioPath)).To(Equal(uintptr(8)))
+		Expect(unsafe.Offsetof(p.OutputPath)).To(Equal(uintptr(16)))
+		Expect(unsafe.Offsetof(p.Fps)).To(Equal(uintptr(24)))
+		Expect(unsafe.Offsetof(p.Crf)).To(Equal(uintptr(28)))
+		Expect(unsafe.Sizeof(p)).To(Equal(uintptr(32)))
+	})
+})
+
+var _ = Describe("video load options", func() {
+	It("stays disengaged for a plain text config", func() {
+		lo := parseOptions(&pb.ModelOptions{Options: []string{"max_num_seqs:16"}})
+		Expect(lo.video.engaged()).To(BeFalse())
+	})
+
+	It("reads the H3 checkpoint set from the options list", func() {
+		lo := parseOptions(&pb.ModelOptions{Options: []string{
+			"video_encoder:qwen3vl-32B-MiniMax-H3-Q4_K_M.gguf",
+			"video_tokenizer:tokenizer.json",
+			"video_vae:vae/diffusion_pytorch_model.safetensors",
+			"audio_vae:audio_vae/model.safetensors",
+			"video_partition:fl2va",
+			"video_device:cuda",
+			"video_dequant_bf16:true",
+			"video_width:1344",
+			"video_height:768",
+			"video_num_frames:124",
+			"video_steps:50",
+		}})
+		Expect(lo.video.engaged()).To(BeTrue())
+		Expect(lo.video.encoderPath).To(Equal("qwen3vl-32B-MiniMax-H3-Q4_K_M.gguf"))
+		Expect(lo.video.tokenizerPath).To(Equal("tokenizer.json"))
+		Expect(lo.video.videoVaePath).To(Equal("vae/diffusion_pytorch_model.safetensors"))
+		Expect(lo.video.audioVaePath).To(Equal("audio_vae/model.safetensors"))
+		Expect(lo.video.partition).To(Equal(partitionFL2VA))
+		Expect(lo.video.device).To(Equal(videoDeviceCUDA))
+		Expect(lo.video.deviceSet).To(BeTrue())
+		Expect(lo.video.dequantBf16).To(Equal(int32(1)))
+		Expect(lo.video.width).To(Equal(int32(1344)))
+		Expect(lo.video.height).To(Equal(int32(768)))
+		Expect(lo.video.numFrames).To(Equal(int32(124)))
+		Expect(lo.video.steps).To(Equal(int32(50)))
+	})
+
+	It("reads the same keys from engine_args", func() {
+		lo := parseOptions(&pb.ModelOptions{
+			EngineArgs: `{"video_vae":"vae/v.safetensors","audio_vae":"a.safetensors","video_num_frames":124,"video_dequant_bf16":true}`,
+		})
+		Expect(lo.video.engaged()).To(BeTrue())
+		Expect(lo.video.videoVaePath).To(Equal("vae/v.safetensors"))
+		Expect(lo.video.audioVaePath).To(Equal("a.safetensors"))
+		Expect(lo.video.numFrames).To(Equal(int32(124)))
+		Expect(lo.video.dequantBf16).To(Equal(int32(1)))
+	})
+
+	It("ignores an unknown video_device rather than guessing", func() {
+		lo := parseOptions(&pb.ModelOptions{Options: []string{"video_vae:v", "video_device:tpu"}})
+		Expect(lo.video.deviceSet).To(BeFalse())
+		Expect(lo.video.device).To(Equal(videoDeviceCPU))
+	})
+})
+
+var _ = Describe("per-request params", func() {
+	It("maps the accepted keys", func() {
+		extra, err := parseVideoRequestParams(map[string]string{
+			"noise_aug": "0.5", "ref_image": "/tmp/ref.ppm", "crf": "20",
+		})
+		Expect(err).ToNot(HaveOccurred())
+		Expect(extra.noiseAug).To(BeNumerically("~", 0.5, 1e-6))
+		Expect(extra.refImage).To(Equal("/tmp/ref.ppm"))
+		Expect(extra.crf).To(Equal(int32(20)))
+	})
+
+	It("refuses an unknown key instead of dropping it", func() {
+		_, err := parseVideoRequestParams(map[string]string{"resolution": "480p"})
+		Expect(err).To(MatchError(ContainSubstring("unknown params key")))
+	})
+
+	It("refuses a non-numeric noise_aug", func() {
+		_, err := parseVideoRequestParams(map[string]string{"noise_aug": "high"})
+		Expect(err).To(HaveOccurred())
+	})
+})
+
+// The partition guard is the correctness rule this backend exists to enforce:
+// the FL2VA DiT serves t2va and fl2va, and handing it reference conditioning
+// renders a broken lattice over the frame after a multi-hour generation rather
+// than failing.
+var _ = Describe("partition conditioning guard", func() {
+	It("accepts a plain t2va request on fl2va", func() {
+		Expect(checkPartitionConditioning(partitionFL2VA,
+			&pb.GenerateVideoRequest{Prompt: "a llama"}, videoExtraParams{})).To(Succeed())
+	})
+
+	It("accepts fl2va keyframes on fl2va", func() {
+		Expect(checkPartitionConditioning(partitionFL2VA,
+			&pb.GenerateVideoRequest{StartImage: "/tmp/a.png"}, videoExtraParams{})).To(Succeed())
+	})
+
+	It("refuses a reference image on fl2va", func() {
+		err := checkPartitionConditioning(partitionFL2VA,
+			&pb.GenerateVideoRequest{}, videoExtraParams{refImage: "/tmp/ref.ppm"})
+		Expect(err).To(MatchError(ContainSubstring("ref2va")))
+	})
+
+	It("refuses reference audio on fl2va", func() {
+		err := checkPartitionConditioning(partitionFL2VA,
+			&pb.GenerateVideoRequest{Audio: "/tmp/voice.wav"}, videoExtraParams{})
+		Expect(err).To(HaveOccurred())
+	})
+
+	It("refuses fl2va keyframes on ref2va", func() {
+		err := checkPartitionConditioning(partitionRef2VA,
+			&pb.GenerateVideoRequest{StartImage: "/tmp/a.png"}, videoExtraParams{})
+		Expect(err).To(HaveOccurred())
+	})
+
+	It("refuses keyframes and references together on either partition", func() {
+		err := checkPartitionConditioning(partitionRef2VA,
+			&pb.GenerateVideoRequest{StartImage: "/tmp/a.png"}, videoExtraParams{refVideo: "/tmp/clip"})
+		Expect(err).To(MatchError(ContainSubstring("exclusive")))
+	})
+})
+
+var _ = Describe("H3 geometry", func() {
+	It("keeps an explicitly requested canvas", func() {
+		w, h, err := resolveCanvas(1280, 720)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(w).To(Equal(int32(1280)))
+		Expect(h).To(Equal(int32(720)))
+	})
+
+	It("falls back to the shipped 1344x768 canvas", func() {
+		w, h, err := resolveCanvas(0, 0)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(w).To(Equal(int32(1344)))
+		Expect(h).To(Equal(int32(768)))
+	})
+
+	It("derives a landscape canvas from a keyframe's aspect", func() {
+		path := writePPM(1920, 1080)
+		w, h, err := resolveCanvas(0, 0, path)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(h).To(Equal(int32(768)))
+		// 768 * 16/9 = 1365.33; /32 = 42.67, round-half-to-even to 43, x32.
+		Expect(w).To(Equal(int32(1376)))
+	})
+
+	It("derives a portrait canvas from a keyframe's aspect", func() {
+		path := writePPM(1080, 1920)
+		w, h, err := resolveCanvas(0, 0, path)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(w).To(Equal(int32(768)))
+		Expect(h).To(Equal(int32(1376)))
+	})
+
+	It("truncates onto the 32 grid the way the engine does", func() {
+		Expect(truncateToGrid(1000)).To(Equal(int32(992)))
+		Expect(truncateToGrid(768)).To(Equal(int32(768)))
+	})
+
+	It("reports the 17n+5 frame grid", func() {
+		Expect(alignFrameCount(124)).To(Equal(int32(124)))
+		Expect(alignFrameCount(120)).To(Equal(int32(124)))
+		Expect(alignFrameCount(100)).To(Equal(int32(107)))
+	})
+})
+
+var _ = Describe("keyframe staging", func() {
+	It("parses a binary PPM header, comments included", func() {
+		dir := GinkgoT().TempDir()
+		path := filepath.Join(dir, "commented.ppm")
+		Expect(os.WriteFile(path, []byte("P6\n# made by a test\n64 32\n255\n"), 0o600)).To(Succeed())
+		w, h, err := ppmDimensions(path)
+		Expect(err).ToNot(HaveOccurred())
+		Expect(w).To(Equal(int32(64)))
+		Expect(h).To(Equal(int32(32)))
+	})
+
+	It("refuses an ASCII PPM (P3): the engine reads P6 only", func() {
+		dir := GinkgoT().TempDir()
+		path := filepath.Join(dir, "ascii.ppm")
+		Expect(os.WriteFile(path, []byte("P3\n64 32\n255\n"), 0o600)).To(Succeed())
+		_, _, err := ppmDimensions(path)
+		Expect(err).To(HaveOccurred())
+	})
+
+	It("passes a P6 already at the canvas straight through, without ffmpeg", func() {
+		path := writePPM(64, 32)
+		out, err := stageKeyframe("", path, 64, 32, GinkgoT().TempDir(), "first")
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(Equal(path))
+	})
+
+	It("is a no-op for an absent keyframe", func() {
+		out, err := stageKeyframe("", "", 64, 32, GinkgoT().TempDir(), "first")
+		Expect(err).ToNot(HaveOccurred())
+		Expect(out).To(BeEmpty())
+	})
+})
+
+var _ = Describe("GenerateVideo preconditions", func() {
+	It("refuses when the model is not a video engine", func() {
+		v := &VllmCpp{}
+		Expect(v.GenerateVideo(&pb.GenerateVideoRequest{Prompt: "x", Dst: "/tmp/o.mp4"})).
+			To(MatchError(ContainSubstring("not a MiniMax-H3 video engine")))
+	})
+})
+
+// writePPM writes a valid P6 header of the given geometry. Only the header is
+// read by anything under test, so the pixel payload is left off.
+func writePPM(width, height int) string {
+	dir := GinkgoT().TempDir()
+	path := filepath.Join(dir, "frame.ppm")
+	header := []byte("P6\n" + itoa(width) + " " + itoa(height) + "\n255\n")
+	Expect(os.WriteFile(path, header, 0o600)).To(Succeed())
+	return path
+}
+
+func itoa(v int) string {
+	if v == 0 {
+		return "0"
+	}
+	digits := ""
+	for v > 0 {
+		digits = string(rune('0'+v%10)) + digits
+		v /= 10
+	}
+	return digits
+}

backend/go/vllm-cpp/vllmcpp_test.go

@@ -16,7 +16,7 @@ func TestVllmCpp(t *testing.T) {
 	RunSpecs(t, "vllm-cpp suite")
 }
 
-// The Go POD mirrors must match the C struct layout of vllm.h (ABI v10)
+// The Go POD mirrors must match the C struct layout of vllm.h (ABI v23)
 // byte-for-byte: these offsets are the C offsets on LP64 (linux/darwin
 // amd64+arm64). A failure here means govllmcpp.go drifted from vllm.h.
 var _ = Describe("C ABI struct mirrors", func() {
@@ -24,7 +24,7 @@ var _ = Describe("C ABI struct mirrors", func() {
 		// VLLM_ABI_VERSION in the vllm.h of VLLM_CPP_VERSION (Makefile).
 		// Moving the pin past this without growing the mirrors below ships a
 		// backend that refuses every load at startup (issue #11379).
-		Expect(abiVersion).To(Equal(10))
+		Expect(abiVersion).To(Equal(23))
 	})
 
 	It("cModelParams matches vllm_model_params", func() {
@@ -42,10 +42,17 @@ var _ = Describe("C ABI struct mirrors", func() {
 		Expect(unsafe.Offsetof(p.MaxNumBatchedTokens)).To(Equal(uintptr(60)))
 		Expect(unsafe.Offsetof(p.SchedulingPolicy)).To(Equal(uintptr(64)))
 		Expect(unsafe.Offsetof(p.KVTransferConfig)).To(Equal(uintptr(72)))
-		Expect(unsafe.Offsetof(p.EnableJumpForward)).To(Equal(uintptr(80)))
-		// 88, not 84: the struct is 8-aligned (it holds pointers), so the
-		// trailing int32 is padded out. Go pads identically.
-		Expect(unsafe.Sizeof(p)).To(Equal(uintptr(88)))
+		Expect(unsafe.Offsetof(p.OffloadConfig)).To(Equal(uintptr(80)))
+		Expect(unsafe.Offsetof(p.EnableJumpForward)).To(Equal(uintptr(88)))
+		Expect(unsafe.Offsetof(p.Device)).To(Equal(uintptr(92)))
+		// 96: gpu_memory_utilization is a double, so it takes the next
+		// 8-aligned slot after the int32 pair. Go pads identically.
+		Expect(unsafe.Offsetof(p.GPUMemoryUtil)).To(Equal(uintptr(96)))
+		Expect(unsafe.Offsetof(p.KVCacheMemoryBytes)).To(Equal(uintptr(104)))
+		Expect(unsafe.Offsetof(p.LanguageModelOnly)).To(Equal(uintptr(112)))
+		Expect(unsafe.Offsetof(p.LimitMMPerPrompt)).To(Equal(uintptr(120)))
+		Expect(unsafe.Offsetof(p.MMProjPath)).To(Equal(uintptr(128)))
+		Expect(unsafe.Sizeof(p)).To(Equal(uintptr(136)))
 	})
 
 	It("cSamplingParams matches vllm_sampling_params (ABI v8)", func() {

backend/go/whisper/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggml-org/whisper.cpp
-WHISPER_CPP_VERSION?=306c88f4d1286aec1bf96e544632897886af5501
+WHISPER_CPP_VERSION?=c44b60b8053bbf2a5c1e014f11323fb3f2485177
 SO_TARGET?=libgowhisper.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/index.yaml

@@ -11,6 +11,8 @@
     - https://github.com/ggerganov/llama.cpp
   tags:
     - text-to-text
+    - text-to-speech
+    - TTS
     - LLM
     - CPU
     - GPU
@@ -508,7 +510,7 @@
     default: "cpu-stablediffusion-ggml"
     nvidia: "cuda12-stablediffusion-ggml"
     intel: "intel-sycl-f16-stablediffusion-ggml"
-    # amd: "rocm-stablediffusion-ggml"
+    amd: "rocm-stablediffusion-ggml"
     vulkan: "vulkan-stablediffusion-ggml"
     nvidia-l4t: "nvidia-l4t-arm64-stablediffusion-ggml"
     metal: "metal-stablediffusion-ggml"
@@ -2107,7 +2109,7 @@
     default: "cpu-stablediffusion-ggml-development"
     nvidia: "cuda12-stablediffusion-ggml-development"
     intel: "intel-sycl-f16-stablediffusion-ggml-development"
-    # amd: "rocm-stablediffusion-ggml-development"
+    amd: "rocm-stablediffusion-ggml-development"
     vulkan: "vulkan-stablediffusion-ggml-development"
     nvidia-l4t: "nvidia-l4t-arm64-stablediffusion-ggml-development"
     metal: "metal-stablediffusion-ggml-development"
@@ -3902,6 +3904,11 @@
   uri: "quay.io/go-skynet/local-ai-backends:latest-gpu-nvidia-cuda-12-stablediffusion-ggml"
   mirrors:
     - localai/localai-backends:latest-gpu-nvidia-cuda-12-stablediffusion-ggml
+- !!merge <<: *stablediffusionggml
+  name: "rocm-stablediffusion-ggml"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-gpu-rocm-hipblas-stablediffusion-ggml"
+  mirrors:
+    - localai/localai-backends:latest-gpu-rocm-hipblas-stablediffusion-ggml
 - !!merge <<: *stablediffusionggml
   name: "intel-sycl-f32-stablediffusion-ggml"
   uri: "quay.io/go-skynet/local-ai-backends:latest-gpu-intel-sycl-f32-stablediffusion-ggml"
@@ -3915,6 +3922,11 @@
   uri: "quay.io/go-skynet/local-ai-backends:master-gpu-nvidia-cuda-12-stablediffusion-ggml"
   mirrors:
     - localai/localai-backends:master-gpu-nvidia-cuda-12-stablediffusion-ggml
+- !!merge <<: *stablediffusionggml
+  name: "rocm-stablediffusion-ggml-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-gpu-rocm-hipblas-stablediffusion-ggml"
+  mirrors:
+    - localai/localai-backends:master-gpu-rocm-hipblas-stablediffusion-ggml
 - !!merge <<: *stablediffusionggml
   name: "intel-sycl-f32-stablediffusion-ggml-development"
   uri: "quay.io/go-skynet/local-ai-backends:master-gpu-intel-sycl-f32-stablediffusion-ggml"

backend/python/chatterbox/backend.py

@@ -19,6 +19,7 @@ import grpc
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'common'))
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), 'common'))
 from grpc_auth import get_auth_interceptors
+from temp_utils import cleanup_paths
 
 import tempfile
 
@@ -115,11 +116,6 @@ def merge_audio_files(audio_files, output_path, sample_rate):
     # Save the merged audio
     ta.save(output_path, merged_waveform, sample_rate)
     
-    # Clean up temporary files
-    for audio_file in audio_files:
-        if os.path.exists(audio_file):
-            os.remove(audio_file)
-
 _ONE_DAY_IN_SECONDS = 60 * 60 * 24
 
 # If MAX_WORKERS are specified in the environment use it, otherwise default to 1
@@ -226,19 +222,20 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                 text_chunks = split_text_at_word_boundary(request.text, max_length=250)
                 print(f"Splitting text into chunks of 250 characters: {len(text_chunks)}", file=sys.stderr)
                 # Generate audio for each chunk
-                temp_audio_files = []
-                for i, chunk in enumerate(text_chunks):
-                    # Generate audio for this chunk
-                    wav = self.model.generate(chunk, **kwargs)
-                    
-                    # Create temporary file for this chunk
-                    temp_file = tempfile.NamedTemporaryFile(delete=False, suffix='.wav')
-                    temp_file.close()
-                    ta.save(temp_file.name, wav, self.model.sr)
-                    temp_audio_files.append(temp_file.name)
-                
-                # Merge all audio files
-                merge_audio_files(temp_audio_files, request.dst, self.model.sr)
+                with cleanup_paths() as temp_audio_files:
+                    for i, chunk in enumerate(text_chunks):
+                        # Generate audio for this chunk
+                        wav = self.model.generate(chunk, **kwargs)
+
+                        # Register ownership before saving so a partial write is
+                        # removed too when generation or encoding fails.
+                        temp_file = tempfile.NamedTemporaryFile(delete=False, suffix='.wav')
+                        temp_file.close()
+                        temp_audio_files.append(temp_file.name)
+                        ta.save(temp_file.name, wav, self.model.sr)
+
+                    # Merge all audio files
+                    merge_audio_files(temp_audio_files, request.dst, self.model.sr)
             else:
                 # Generate audio using ChatterboxTTS for short text
                 wav = self.model.generate(request.text, **kwargs)

backend/python/common/python_utils.py

@@ -37,6 +37,46 @@ def parse_options(options_list):
     return opts
 
 
+def attach_media_parts(messages_dicts, n_images=0, n_videos=0):
+    """Rebuild the last user message as content *parts* carrying media markers.
+
+    Backends that let the tokenizer do the templating hand plain string content
+    to ``apply_chat_template``, but a chat template only emits the model's own
+    media tokens (``<|vision_start|><|image_pad|><|vision_end|>`` for the
+    Qwen-VL family, and the equivalents elsewhere) when the content is a list
+    of parts. Without those markers the engine's multimodal processor finds
+    nothing to substitute and silently discards the pixels, even though they
+    were forwarded correctly out of band.
+
+    Returns a new list whose last user message has
+    ``[{"type": "image"} * n_images, {"type": "video"} * n_videos, text]`` as
+    its content, or ``None`` when there is nothing to attach - no media, no
+    user turn, or content that is already a list of parts - so the caller can
+    keep using the original string-content list.
+    """
+    if not n_images and not n_videos:
+        return None
+    idx = next(
+        (
+            i
+            for i in reversed(range(len(messages_dicts)))
+            if messages_dicts[i].get("role") == "user"
+        ),
+        None,
+    )
+    if idx is None:
+        return None
+    text = messages_dicts[idx].get("content") or ""
+    if not isinstance(text, str):
+        return None
+    parts = [{"type": "image"}] * n_images + [{"type": "video"}] * n_videos
+    if text:
+        parts.append({"type": "text", "text": text})
+    patched = list(messages_dicts)
+    patched[idx] = dict(patched[idx], content=parts)
+    return patched
+
+
 def messages_to_dicts(proto_messages):
     """Convert proto ``Message`` objects to dicts suitable for ``apply_chat_template``.
 

backend/python/common/python_utils_test.py

@@ -14,7 +14,7 @@ import json
 import types
 import unittest
 
-from python_utils import messages_to_dicts, parse_options
+from python_utils import attach_media_parts, messages_to_dicts, parse_options
 
 
 def _msg(**fields):
@@ -118,5 +118,63 @@ class TestMessagesToDicts(unittest.TestCase):
         self.assertNotIn("tool_calls", out[0])
 
 
+class TestAttachMediaParts(unittest.TestCase):
+    def test_image_marker_added_to_last_user_turn(self):
+        messages = [
+            {"role": "system", "content": "be brief"},
+            {"role": "user", "content": "first"},
+            {"role": "assistant", "content": "ok"},
+            {"role": "user", "content": "how high is the water?"},
+        ]
+        out = attach_media_parts(messages, n_images=1)
+        self.assertEqual(
+            out[3]["content"],
+            [{"type": "image"}, {"type": "text", "text": "how high is the water?"}],
+        )
+        # Earlier turns and the input list itself are untouched.
+        self.assertEqual(out[:3], messages[:3])
+        self.assertEqual(messages[3]["content"], "how high is the water?")
+
+    def test_counts_and_order_images_then_videos(self):
+        out = attach_media_parts(
+            [{"role": "user", "content": "describe"}], n_images=2, n_videos=1
+        )
+        self.assertEqual(
+            out[0]["content"],
+            [
+                {"type": "image"},
+                {"type": "image"},
+                {"type": "video"},
+                {"type": "text", "text": "describe"},
+            ],
+        )
+
+    def test_empty_text_yields_media_only_parts(self):
+        out = attach_media_parts([{"role": "user", "content": ""}], n_images=1)
+        self.assertEqual(out[0]["content"], [{"type": "image"}])
+
+    def test_other_message_keys_are_preserved(self):
+        out = attach_media_parts(
+            [{"role": "user", "content": "hi", "name": "bob"}], n_images=1
+        )
+        self.assertEqual(out[0]["name"], "bob")
+
+    def test_no_media_is_a_no_op(self):
+        self.assertIsNone(attach_media_parts([{"role": "user", "content": "hi"}]))
+
+    def test_no_user_turn_is_a_no_op(self):
+        self.assertIsNone(
+            attach_media_parts([{"role": "system", "content": "hi"}], n_images=1)
+        )
+
+    def test_content_already_parts_is_a_no_op(self):
+        self.assertIsNone(
+            attach_media_parts(
+                [{"role": "user", "content": [{"type": "text", "text": "hi"}]}],
+                n_images=1,
+            )
+        )
+
+
 if __name__ == "__main__":
     unittest.main()

backend/python/common/temp_utils.py

@@ -0,0 +1,36 @@
+import base64
+import contextlib
+import os
+import tempfile
+
+
+@contextlib.contextmanager
+def materialize_base64(data, suffix=""):
+    """Materialize base64 data for a path-only library and always remove it."""
+    descriptor, path = tempfile.mkstemp(prefix="localai-media-", suffix=suffix)
+    try:
+        with os.fdopen(descriptor, "wb") as output:
+            descriptor = None
+            output.write(base64.b64decode(data))
+        yield path
+    finally:
+        if descriptor is not None:
+            os.close(descriptor)
+        try:
+            os.remove(path)
+        except OSError:
+            pass
+
+
+@contextlib.contextmanager
+def cleanup_paths():
+    """Collect temporary paths and remove them on success or failure."""
+    paths = []
+    try:
+        yield paths
+    finally:
+        for path in paths:
+            try:
+                os.remove(path)
+            except OSError:
+                pass

backend/python/common/temp_utils_test.py

@@ -0,0 +1,41 @@
+import os
+import tempfile
+import unittest
+from unittest import mock
+
+from temp_utils import cleanup_paths, materialize_base64
+
+
+class MaterializeBase64Test(unittest.TestCase):
+    def test_removes_materialized_file_after_success(self):
+        with tempfile.TemporaryDirectory() as directory:
+            with mock.patch.object(tempfile, "tempdir", directory):
+                with materialize_base64("aGVsbG8=", suffix=".data") as path:
+                    with open(path, "rb") as materialized:
+                        self.assertEqual(materialized.read(), b"hello")
+                self.assertFalse(os.path.exists(path))
+
+    def test_removes_materialized_file_when_consumer_fails(self):
+        with tempfile.TemporaryDirectory() as directory:
+            with mock.patch.object(tempfile, "tempdir", directory):
+                with self.assertRaisesRegex(RuntimeError, "decode failed"):
+                    with materialize_base64("aGVsbG8="):
+                        raise RuntimeError("decode failed")
+                self.assertEqual(os.listdir(directory), [])
+
+
+class CleanupPathsTest(unittest.TestCase):
+    def test_removes_every_registered_path_after_failure(self):
+        with tempfile.TemporaryDirectory() as directory:
+            paths = [os.path.join(directory, name) for name in ("one.wav", "two.wav")]
+            with self.assertRaisesRegex(RuntimeError, "merge failed"):
+                with cleanup_paths() as registered:
+                    for path in paths:
+                        open(path, "wb").close()
+                        registered.append(path)
+                    raise RuntimeError("merge failed")
+            self.assertEqual(os.listdir(directory), [])
+
+
+if __name__ == "__main__":
+    unittest.main()

backend/python/common/template/requirements-intel.txt

@@ -1,4 +1,4 @@
 --extra-index-url https://download.pytorch.org/whl/xpu
-torch==2.13.0+xpu
+torch==2.14.0+xpu
 oneccl_bind_pt==2.8.0+xpu
 optimum[openvino]

backend/python/common/template/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.82.1
+grpcio==1.83.1
 protobuf
 grpcio-tools

backend/python/coqui/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.83.0
+grpcio==1.83.1
 protobuf
 certifi
-packaging==26.2
+packaging==26.3

backend/python/diffusers/backend.py

@@ -122,6 +122,21 @@ from diffusers.schedulers import (
     UniPCMultistepScheduler,
 )
 
+def select_device(request_cuda, device_option, cuda_available, xpu, mps_available):
+    """Pick the pipeline device. An explicit `device:` model option wins;
+    otherwise CUDA is used whenever torch reports it available (ROCm
+    builds included) or the model config forces it with `cuda: true`,
+    keeping the pre-existing XPU/MPS overrides. CPU is the fallback, not
+    the default."""
+    if device_option:
+        return device_option
+    device = "cuda" if (request_cuda or cuda_available) else "cpu"
+    if xpu:
+        device = "xpu"
+    if mps_available:
+        device = "mps"
+    return device
+
 def is_float(s):
     """Check if a string can be converted to float."""
     try:
@@ -627,12 +642,13 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                 # modify LoraAdapter to be relative to modelFileBase
                 request.LoraAdapter = os.path.join(request.ModelPath, request.LoraAdapter)
 
-            device = "cpu" if not request.CUDA else "cuda"
-            if XPU:
-                device = "xpu"
-            mps_available = hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
-            if mps_available:
-                device = "mps"
+            device = select_device(
+                request.CUDA,
+                self.options.pop("device", None),
+                torch.cuda.is_available(),
+                XPU,
+                hasattr(torch.backends, "mps") and torch.backends.mps.is_available(),
+            )
             self.device = device
             if request.LoraAdapter:
                 # Check if its a local file and not a directory ( we load lora differently for a safetensor file )
@@ -800,12 +816,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             image = image.resize((1024, 576))
 
             generator = torch.manual_seed(request.seed)
-            frames = self.pipe(image, guidance_scale=self.cfg_scale, decode_chunk_size=CHUNK_SIZE, generator=generator).frames[0]
+            frames = self.pipe(image=image, guidance_scale=self.cfg_scale, decode_chunk_size=CHUNK_SIZE, generator=generator).frames[0]
             export_to_video(frames, request.dst, fps=FPS)
             return backend_pb2.Result(message="Media generated successfully", success=True)
 
         if self.txt2vid:
-            video_frames = self.pipe(prompt, guidance_scale=self.cfg_scale, num_inference_steps=steps, num_frames=int(FRAMES)).frames
+            video_frames = self.pipe(prompt=prompt, guidance_scale=self.cfg_scale, num_inference_steps=steps, num_frames=int(FRAMES)).frames
             export_to_video(video_frames, request.dst)
             return backend_pb2.Result(message="Media generated successfully", success=True)
 
@@ -868,7 +884,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         else:
             # pass the kwargs dictionary to the self.pipe method
             image = self.pipe(
-                prompt,
+                prompt=prompt,
                 guidance_scale=self.cfg_scale,
                 **kwargs
             ).images[0]

backend/python/diffusers/test.py

@@ -7,6 +7,7 @@ import time
 from unittest.mock import patch, MagicMock
 
 # Import dynamic loader for testing (these don't need gRPC)
+import backend
 import diffusers_dynamic_loader as loader
 from diffusers import DiffusionPipeline, StableDiffusionPipeline
 
@@ -373,3 +374,74 @@ class TestGenerateImageOptionsKwargsMerge(unittest.TestCase):
         finally:
             os.unlink(src_file.name)
             os.unlink(dst_file.name)
+
+    def test_text_to_image_prompt_is_passed_by_keyword(self):
+        """Test compatibility with pipelines that take image before prompt."""
+        import os
+        import tempfile
+
+        from PIL import Image
+
+        from backend import BackendServicer
+
+        class Flux2CompatiblePipeline:
+            """Model the FLUX.2 call signature: image is before prompt."""
+
+            def __call__(self, image=None, prompt=None, **kwargs):
+                if prompt is None:
+                    raise ValueError("prompt was not passed by keyword")
+                self.prompt = prompt
+                self.kwargs = kwargs
+                return MagicMock(images=[Image.new("RGB", (4, 4))])
+
+        pipeline = Flux2CompatiblePipeline()
+        svc = BackendServicer.__new__(BackendServicer)
+        svc.pipe = pipeline
+        svc.cfg_scale = 7.5
+        svc.controlnet = None
+        svc.img2vid = False
+        svc.txt2vid = False
+        svc.clip_skip = 0
+        svc.PipelineType = "Flux2KleinPipeline"
+        svc.options = {}
+
+        with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as dst_file:
+            dst_path = dst_file.name
+
+        try:
+            request = MagicMock()
+            request.positive_prompt = "a red apple on a wooden table"
+            request.negative_prompt = ""
+            request.step = 4
+            request.seed = 0
+            request.width = 0
+            request.height = 0
+            request.src = ""
+            request.ref_images = []
+            request.dst = dst_path
+
+            svc.GenerateImage(request, context=None)
+
+            self.assertEqual(pipeline.prompt, request.positive_prompt)
+            self.assertEqual(pipeline.kwargs["num_inference_steps"], 4)
+        finally:
+            os.unlink(dst_path)
+
+
+class TestDeviceSelection(unittest.TestCase):
+    """Unit tests for backend.select_device (no GPU required)."""
+
+    def test_autodetect_cuda(self):
+        self.assertEqual(backend.select_device(False, None, True, False, False), "cuda")
+
+    def test_cpu_fallback(self):
+        self.assertEqual(backend.select_device(False, None, False, False, False), "cpu")
+
+    def test_forced_cuda(self):
+        self.assertEqual(backend.select_device(True, None, False, False, False), "cuda")
+
+    def test_device_option_wins(self):
+        self.assertEqual(backend.select_device(True, "cpu", True, True, True), "cpu")
+
+    def test_mps_overrides(self):
+        self.assertEqual(backend.select_device(False, None, True, False, True), "mps")

backend/python/fish-speech/Makefile

@@ -1,7 +1,10 @@
-.PHONY: fish-speech
-fish-speech:
+.PHONY: fish-speech test-source-preparation
+fish-speech: test-source-preparation
 	bash install.sh
 
+test-source-preparation:
+	bash prepare-source_test.sh
+
 .PHONY: run
 run: fish-speech
 	@echo "Running fish-speech..."

backend/python/fish-speech/install.sh

@@ -39,10 +39,10 @@ else
     cd "${FISH_SPEECH_DIR}" && git pull && cd -
 fi
 
-# Remove pyaudio from fish-speech deps — it's only used by the upstream client tool
-# (tools/api_client.py) for speaker playback, not by our gRPC backend server.
-# It requires native portaudio libs which aren't available on all build environments.
-sed -i.bak '/"pyaudio"/d' "${FISH_SPEECH_DIR}/pyproject.toml"
+# Keep the platform-specific PyTorch installed above. Upstream pins the generic
+# PyPI torch wheel, which replaces ROCm builds with a CUDA wheel during the
+# editable install. pyaudio is only used by the upstream playback client.
+bash "${backend_dir}/prepare-source.sh" "${BUILD_TYPE:-}" "${FISH_SPEECH_DIR}/pyproject.toml"
 
 # Install fish-speech deps from source (without the package itself since we use PYTHONPATH)
 ensureVenv

backend/python/fish-speech/prepare-source.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+set -euo pipefail
+
+build_type=${1:-}
+pyproject=${2:?usage: prepare-source.sh BUILD_TYPE PYPROJECT}
+prepared=$(mktemp "${pyproject}.XXXXXX")
+trap 'rm -f "$prepared"' EXIT
+
+awk -v build_type="$build_type" '
+    /^dependencies = \[$/ { in_project_dependencies = 1 }
+    build_type == "hipblas" && in_project_dependencies && /^[[:space:]]*"(torch|torchaudio)[^"]*",?[[:space:]]*$/ { next }
+    in_project_dependencies && /^[[:space:]]*"pyaudio",?[[:space:]]*$/ { next }
+    { print }
+    in_project_dependencies && /^\]$/ { in_project_dependencies = 0 }
+' "$pyproject" > "$prepared"
+
+mv "$prepared" "$pyproject"
+trap - EXIT

backend/python/fish-speech/prepare-source_test.sh

@@ -0,0 +1,66 @@
+#!/bin/bash
+set -euo pipefail
+
+SCRIPT_DIR=$(dirname "$(realpath "$0")")
+WORK_DIR=$(mktemp -d)
+trap 'rm -rf "$WORK_DIR"' EXIT
+
+write_fixture() {
+    cat > "$1" <<'EOF'
+[project]
+dependencies = [
+    "numpy",
+    "torch==2.8.0",
+    "torchaudio==2.8.0",
+    "pyaudio",
+]
+
+[project.optional-dependencies]
+stable = [
+    "torch==2.8.0",
+    "torchaudio",
+]
+EOF
+}
+
+write_fixture "$WORK_DIR/rocm.toml"
+write_fixture "$WORK_DIR/cuda.toml"
+write_fixture "$WORK_DIR/cpu.toml"
+
+bash "$SCRIPT_DIR/prepare-source.sh" hipblas "$WORK_DIR/rocm.toml"
+bash "$SCRIPT_DIR/prepare-source.sh" cublas "$WORK_DIR/cuda.toml"
+bash "$SCRIPT_DIR/prepare-source.sh" "" "$WORK_DIR/cpu.toml"
+
+cat > "$WORK_DIR/expected-rocm.toml" <<'EOF'
+[project]
+dependencies = [
+    "numpy",
+]
+
+[project.optional-dependencies]
+stable = [
+    "torch==2.8.0",
+    "torchaudio",
+]
+EOF
+
+cat > "$WORK_DIR/expected-default.toml" <<'EOF'
+[project]
+dependencies = [
+    "numpy",
+    "torch==2.8.0",
+    "torchaudio==2.8.0",
+]
+
+[project.optional-dependencies]
+stable = [
+    "torch==2.8.0",
+    "torchaudio",
+]
+EOF
+
+diff -u "$WORK_DIR/expected-rocm.toml" "$WORK_DIR/rocm.toml"
+diff -u "$WORK_DIR/expected-default.toml" "$WORK_DIR/cuda.toml"
+diff -u "$WORK_DIR/expected-default.toml" "$WORK_DIR/cpu.toml"
+
+echo "PASS: source preparation preserves each platform's PyTorch dependencies"

backend/python/fish-speech/test.sh

@@ -8,4 +8,5 @@ else
     source $backend_dir/../common/libbackend.sh
 fi
 
+bash "${backend_dir}/prepare-source_test.sh"
 runUnittests

backend/python/insightface/backend.py

@@ -127,7 +127,10 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             context.set_code(grpc.StatusCode.NOT_FOUND)
             context.set_details("no face detected")
             return backend_pb2.EmbeddingResult()
-        return backend_pb2.EmbeddingResult(embeddings=[float(x) for x in vec])
+        return backend_pb2.EmbeddingResult(
+            embeddings=[float(x) for x in vec],
+            layout=backend_pb2.EMBEDDING_LAYOUT_FINAL,
+        )
 
     def Detect(self, request, context):
         if self.engine is None: