chore(deps): bump llvmlite 0.43→0.49, numba 0.60→0.67, numpy>=2.5.2 in transformers

Merge dependabot PRs #11639, #11640, #11641 together: - #11639: llvmlite 0.43.0 → 0.49.0 (alone fails: numba 0.60 requires <0.44) - #11640: numba 0.60.0 → 0.67.0 (alone fails: requires llvmlite >=0.49) - #11641: numpy >=2.0.0 → >=2.5.2 (was blocked by scipy 1.15.1 <2.5, now resolved since scipy was bumped to 1.18.0 in #11399) Also bump outetts backend to match (same numba/llvmlite conflict). Signed-off-by: localai-org-maint-bot <bot@localai.io> Assisted-by: MAKI:GLM5.2 [gh]
fix(qwen-tts): limit CUDA compiler threads
2026-09-11 21:57:55 -04:00 · 2026-09-11 23:34:59 +00:00 · 2026-09-11 23:19:28 +00:00 · 2026-09-11 23:13:32 +00:00 · 2026-09-11 23:08:26 +00:00 · 2026-09-11 23:06:18 +00:00
398 changed files with 23767 additions and 1412 deletions
@@ -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
@@ -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.
@@ -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 |
@@ -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
 
@@ -166,6 +166,19 @@ include:
    dockerfile: "./backend/Dockerfile.python"
    context: "./"
    ubuntu-version: '2404'
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-whisper-medusa'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'true'
+    backend: "whisper-medusa"
+    dockerfile: "./backend/Dockerfile.python"
+    context: "./"
+    ubuntu-version: '2404'
  - build-type: ''
    cuda-major-version: ""
    cuda-minor-version: ""
@@ -354,6 +367,19 @@ include:
    dockerfile: "./backend/Dockerfile.python"
    context: "./"
    ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "12"
+    cuda-minor-version: "8"
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-nvidia-cuda-12-funasr'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "funasr"
+    dockerfile: "./backend/Dockerfile.python"
+    context: "./"
+    ubuntu-version: '2404'
  - build-type: 'cublas'
    cuda-major-version: "12"
    cuda-minor-version: "8"
@@ -626,6 +652,19 @@ include:
    dockerfile: "./backend/Dockerfile.python"
    context: "./"
    ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "12"
+    cuda-minor-version: "1"
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-nvidia-cuda-12-whisper-medusa'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "whisper-medusa"
+    dockerfile: "./backend/Dockerfile.python"
+    context: "./"
+    ubuntu-version: '2404'
  - build-type: 'cublas'
    cuda-major-version: "12"
    cuda-minor-version: "8"
@@ -1069,6 +1108,19 @@ include:
    dockerfile: "./backend/Dockerfile.python"
    context: "./"
    ubuntu-version: '2404'
+  - build-type: 'cublas'
+    cuda-major-version: "13"
+    cuda-minor-version: "0"
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-nvidia-cuda-13-funasr'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "funasr"
+    dockerfile: "./backend/Dockerfile.python"
+    context: "./"
+    ubuntu-version: '2404'
  - build-type: 'cublas'
    cuda-major-version: "13"
    cuda-minor-version: "0"
@@ -1404,6 +1456,19 @@ include:
    backend: "qwen-asr"
    dockerfile: "./backend/Dockerfile.python"
    context: "./"
+  - build-type: 'l4t'
+    cuda-major-version: "13"
+    cuda-minor-version: "0"
+    platforms: 'linux/arm64'
+    tag-latest: 'auto'
+    tag-suffix: '-nvidia-l4t-cuda-13-arm64-funasr'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    ubuntu-version: '2404'
+    backend: "funasr"
+    dockerfile: "./backend/Dockerfile.python"
+    context: "./"
  - build-type: 'l4t'
    cuda-major-version: "13"
    cuda-minor-version: "0"
@@ -2388,6 +2453,19 @@ include:
    dockerfile: "./backend/Dockerfile.python"
    context: "./"
    ubuntu-version: '2404'
+  - build-type: 'hipblas'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-rocm-hipblas-funasr'
+    runs-on: 'ubuntu-latest'
+    base-image: "rocm/dev-ubuntu-24.04:7.2.1"
+    skip-drivers: 'false'
+    backend: "funasr"
+    dockerfile: "./backend/Dockerfile.python"
+    context: "./"
+    ubuntu-version: '2404'
  - build-type: 'hipblas'
    cuda-major-version: ""
    cuda-minor-version: ""
@@ -2668,6 +2746,19 @@ include:
    dockerfile: "./backend/Dockerfile.python"
    context: "./"
    ubuntu-version: '2204'
+  - build-type: 'l4t'
+    cuda-major-version: "12"
+    cuda-minor-version: "0"
+    platforms: 'linux/arm64'
+    tag-latest: 'auto'
+    tag-suffix: '-nvidia-l4t-funasr'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
+    skip-drivers: 'true'
+    backend: "funasr"
+    dockerfile: "./backend/Dockerfile.python"
+    context: "./"
+    ubuntu-version: '2204'
  - build-type: 'l4t'
    cuda-major-version: "12"
    cuda-minor-version: "0"
@@ -2877,6 +2968,19 @@ include:
    dockerfile: "./backend/Dockerfile.python"
    context: "./"
    ubuntu-version: '2404'
+  - build-type: 'intel'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-intel-funasr'
+    runs-on: 'ubuntu-latest'
+    base-image: "intel/oneapi-basekit:2025.3.0-0-devel-ubuntu24.04"
+    skip-drivers: 'false'
+    backend: "funasr"
+    dockerfile: "./backend/Dockerfile.python"
+    context: "./"
+    ubuntu-version: '2404'
  - build-type: 'intel'
    cuda-major-version: ""
    cuda-minor-version: ""
@@ -3208,9 +3312,10 @@ include:
  # consumed. Same reason CUDA needs its toolkit in base-image rather than in a
  # builder image: this is the ds4 shape, not the llama-cpp one.
  #
-  # No ROCm entry: upstream has no HIP configuration. No CUDA arm64 or L4T
-  # entry: upstream documents and validates CUDA on x86 only. Darwin/Metal is in
-  # the includeDarwin matrix below, built by scripts/build/audio-cpp-darwin.sh.
+  # ROCm uses upstream's HIP backend and the project-wide ROCm 7.2.1 base. No
+  # CUDA arm64 or L4T entry: upstream documents and validates CUDA on x86 only.
+  # Darwin/Metal is in the includeDarwin matrix below, built by
+  # scripts/build/audio-cpp-darwin.sh.
  #
  # No vulkan entry either, though Dockerfile.audio-cpp and the backend Makefile
  # both handle BUILD_TYPE=vulkan for local builds. Every other vulkan backend
@@ -3278,6 +3383,19 @@ include:
    dockerfile: "./backend/Dockerfile.audio-cpp"
    context: "./"
    ubuntu-version: '2404'
+  - build-type: 'hipblas'
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-gpu-rocm-hipblas-audio-cpp'
+    runs-on: 'ubuntu-latest'
+    base-image: "rocm/dev-ubuntu-24.04:7.2.1"
+    skip-drivers: 'false'
+    backend: "audio-cpp"
+    dockerfile: "./backend/Dockerfile.audio-cpp"
+    context: "./"
+    ubuntu-version: '2404'
  - build-type: ''
    cuda-major-version: ""
    cuda-minor-version: ""
@@ -6071,6 +6189,20 @@ include:
    dockerfile: "./backend/Dockerfile.python"
    context: "./"
    ubuntu-version: '2404'
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/amd64'
+    platform-tag: 'amd64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-funasr'
+    runs-on: 'ubuntu-latest'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "funasr"
+    dockerfile: "./backend/Dockerfile.python"
+    context: "./"
+    ubuntu-version: '2404'
  - build-type: ''
    cuda-major-version: ""
    cuda-minor-version: ""
@@ -6085,6 +6217,20 @@ include:
    dockerfile: "./backend/Dockerfile.python"
    context: "./"
    ubuntu-version: '2404'
+  - build-type: ''
+    cuda-major-version: ""
+    cuda-minor-version: ""
+    platforms: 'linux/arm64'
+    platform-tag: 'arm64'
+    tag-latest: 'auto'
+    tag-suffix: '-cpu-funasr'
+    runs-on: 'ubuntu-24.04-arm'
+    base-image: "ubuntu:24.04"
+    skip-drivers: 'false'
+    backend: "funasr"
+    dockerfile: "./backend/Dockerfile.python"
+    context: "./"
+    ubuntu-version: '2404'
  - build-type: ''
    cuda-major-version: ""
    cuda-minor-version: ""
@@ -6324,6 +6470,9 @@ includeDarwin:
  - backend: "mlx"
    tag-suffix: "-metal-darwin-arm64-mlx"
    build-type: "mps"
+  - backend: "mlx-video"
+    tag-suffix: "-metal-darwin-arm64-mlx-video"
+    build-type: "mps"
  - backend: "chatterbox"
    tag-suffix: "-metal-darwin-arm64-chatterbox"
    build-type: "mps"
@@ -6460,6 +6609,9 @@ includeDarwin:
  - backend: "qwen-asr"
    tag-suffix: "-metal-darwin-arm64-qwen-asr"
    build-type: "mps"
+  - backend: "funasr"
+    tag-suffix: "-metal-darwin-arm64-funasr"
+    build-type: "mps"
  - backend: "nemo"
    tag-suffix: "-metal-darwin-arm64-nemo"
    build-type: "mps"
@@ -0,0 +1,49 @@
+#!/bin/bash
+# Bump the CTranslate2 ROCm Python wheel release pin used by faster-whisper.
+set -xe
+
+REPO=$1   # OpenNMT/CTranslate2
+FILE=$2   # backend/python/faster-whisper/install.sh
+VAR=$3    # CTRANSLATE2_VERSION (used for output file names so the workflow can read them)
+
+if [ -z "$FILE" ] || [ -z "$REPO" ] || [ -z "$VAR" ]; then
+    echo "usage: $0 <repo> <install-script> <var-name>" >&2
+    exit 1
+fi
+
+LATEST_RELEASE=$(curl -sS -H "Accept: application/vnd.github+json" \
+    "https://api.github.com/repos/$REPO/releases/latest")
+LATEST_TAG=$(python3 -c "import json,sys; print(json.load(sys.stdin)['tag_name'])" <<< "$LATEST_RELEASE")
+
+set +e
+CURRENT_VERSION=$(grep -m1 "^${VAR}=" "$FILE" | cut -d= -f2 | sed -E 's/^\$\{[^:]+:-([^}]+)\}$/\1/')
+CTRANSLATE2_ROCM_WHEEL_OS=$(grep -m1 '^CTRANSLATE2_ROCM_WHEEL_OS=' "$FILE" | cut -d= -f2 | sed -E 's/^\$\{[^:]+:-([^}]+)\}$/\1/')
+set -e
+
+if [ -z "$CURRENT_VERSION" ]; then
+    echo "Could not find $VAR in $FILE."
+    exit 0
+fi
+
+if [ -z "$CTRANSLATE2_ROCM_WHEEL_OS" ]; then
+    echo "Could not find CTRANSLATE2_ROCM_WHEEL_OS in $FILE."
+    exit 0
+fi
+
+ASSET_NAME="rocm-python-wheels-${CTRANSLATE2_ROCM_WHEEL_OS}.zip"
+LATEST_RELEASE="$LATEST_RELEASE" python3 - "$ASSET_NAME" <<'PY'
+import json
+import os
+import sys
+
+asset_name = sys.argv[1]
+release = json.loads(os.environ["LATEST_RELEASE"])
+assets = {asset.get("name") for asset in release.get("assets", [])}
+if asset_name not in assets:
+    raise SystemExit(f"Could not find release asset {asset_name!r}")
+PY
+
+sed -i "$FILE" -e "s/^${VAR}=.*/${VAR}=\${${VAR}:-${LATEST_TAG}}/"
+
+echo "Changes: https://github.com/$REPO/compare/${CURRENT_VERSION}...${LATEST_TAG}" >> "${VAR}_message.txt"
+echo "${LATEST_TAG}" >> "${VAR}_commit.txt"
@@ -3,9 +3,9 @@
 # darwin (Apple Silicon) install path. The macOS/Metal build
 # (backend/python/vllm/install.sh, Darwin branch) installs vllm-metal, which is
 # version-locked to a specific vLLM source release. install.sh derives that vLLM
-# version at build time from vllm-metal's own installer at the pinned
-# tag, so there is only ONE value to bump here -- mirroring bump_vllm_wheel.sh,
-# which bumps the Linux cu130 wheel pin.
+# version, and the wheel asset name, at build time from the pinned tag, so there
+# is only ONE value to bump here -- mirroring bump_vllm_wheel.sh, which bumps the
+# Linux cu130 wheel pin.
 #
 # This deliberately tracks vllm-project/vllm-metal, NOT vllm-project/vllm: the
 # darwin build can only use the exact vLLM version vllm-metal supports, so it may
@@ -23,15 +23,20 @@ if [ -z "$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/$LATEST_TAG/install.sh" \
+# The coupled vLLM release lives in .github/vllm-release-tag.commit at that tag
+# (since vllm-metal 0.28); releases predating that file pinned it inline in their
+# own install.sh. The extractor reads both forms.
+NEW_VLLM_VERSION=$( { 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
@@ -166,7 +166,40 @@ 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: ${{ steps.bump.outputs.message }}
+          signoff: true
+
+  bump-ctranslate2-rocm-wheel:
+    # CTranslate2's ROCm wheels are published as release assets, so the
+    # faster-whisper hipblas install path pins the release tag used in the URL.
+    if: github.repository == 'mudler/LocalAI'
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - name: Bump CTranslate2 ROCm wheel pin 🔧
+        id: bump
+        run: |
+          bash .github/bump_ctranslate2_rocm_wheel.sh OpenNMT/CTranslate2 backend/python/faster-whisper/install.sh CTRANSLATE2_VERSION
+          {
+            echo 'message<<EOF'
+            cat "CTRANSLATE2_VERSION_message.txt"
+            echo EOF
+          } >> "$GITHUB_OUTPUT"
+          {
+            echo 'commit<<EOF'
+            cat "CTRANSLATE2_VERSION_commit.txt"
+            echo EOF
+          } >> "$GITHUB_OUTPUT"
+          rm -rfv CTRANSLATE2_VERSION_message.txt CTRANSLATE2_VERSION_commit.txt
+      - name: Create Pull Request
+        uses: peter-evans/create-pull-request@v8
+        with:
+          token: ${{ secrets.UPDATE_BOT_TOKEN }}
+          push-to-fork: ci-forks/LocalAI
+          commit-message: ':arrow_up: Update OpenNMT/CTranslate2 ROCm wheel'
+          title: 'chore: :arrow_up: Update OpenNMT/CTranslate2 ROCm wheel to `${{ steps.bump.outputs.commit }}`'
+          branch: "update/CTRANSLATE2_VERSION"
          body: ${{ steps.bump.outputs.message }}
          signoff: true

@@ -203,7 +236,7 @@ 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 +274,6 @@ 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
@@ -31,7 +31,7 @@ jobs:
        if: ${{ !github.repository.fork && github.actor != 'dependabot[bot]' }}
      - name: Run Gosec Security Scanner
        if: ${{ !github.repository.fork && github.actor != 'dependabot[bot]' }}
-        uses: securego/gosec@v2.27.1
+        uses: securego/gosec@v2.29.0
        with:
          # we let the report trigger content trigger a failure using the GitHub Security features.
          # backend/go/supertonic is excluded: it vendors upstream supertone-inc/supertonic
@@ -80,8 +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: ${{ failure() && contains(github.event.pull_request.labels.*.name, 'ci-debug') }}
+        timeout-minutes: 30
        uses: mxschmitt/action-tmate@v3.23
        with:
          detached: true
@@ -125,8 +130,13 @@ jobs:
          export PATH="/opt/homebrew/opt/make/libexec/gnubin:$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: ${{ failure() && contains(github.event.pull_request.labels.*.name, 'ci-debug') }}
+        timeout-minutes: 30
        uses: mxschmitt/action-tmate@v3.23
        with:
          detached: true
@@ -77,8 +77,13 @@ jobs:
      - name: Test
        run: |
            PATH="$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: ${{ failure() && contains(github.event.pull_request.labels.*.name, 'ci-debug') }}
+        timeout-minutes: 30
        uses: mxschmitt/action-tmate@v3.23
        with:
          detached: true
@@ -63,8 +63,13 @@ jobs:
      - name: Test Backend E2E
        run: |
          PATH="$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: ${{ failure() && contains(github.event.pull_request.labels.*.name, 'ci-debug') }}
+        timeout-minutes: 30
        uses: mxschmitt/action-tmate@v3.23
        with:
          detached: true
@@ -88,8 +88,13 @@ jobs:
      # CPU and runs the token_classify capability spec (byte-offset contract).
      - name: Run live PII NER backend E2E
        run: PATH="$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
@@ -75,8 +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
@@ -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.
@@ -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:

@@ -1,5 +1,7 @@
 # Disable parallel execution for backend builds
-.NOTPARALLEL: backends/diffusers backends/llama-cpp backends/turboquant backends/bonsai backends/outetts backends/piper backends/stablediffusion-ggml backends/trellis2cpp backends/trellis2cpp-darwin backends/whisper backends/crispasr backends/parakeet-cpp backends/moss-transcribe-cpp backends/nemo-speech-cpp backends/faster-whisper backends/silero-vad backends/local-store backends/valkey-store backends/cloud-proxy backends/huggingface backends/rfdetr backends/rfdetr-cpp backends/insightface backends/speaker-recognition backends/kitten-tts backends/kokoro backends/chatterbox backends/llama-cpp-darwin backends/neutts build-darwin-python-backend build-darwin-go-backend backends/mlx backends/diffuser-darwin backends/mlx-vlm backends/mlx-audio backends/mlx-distributed backends/stablediffusion-ggml-darwin backends/vllm backends/vllm-omni backends/longcat-video backends/sglang backends/moonshine backends/pocket-tts backends/qwen-tts backends/faster-qwen3-tts backends/qwen-asr backends/nemo backends/voxcpm backends/whisperx backends/ace-step backends/acestep-cpp backends/fish-speech backends/voxtral backends/opus backends/trl backends/llama-cpp-quantization backends/kokoros backends/sam3-cpp backends/qwen3-tts-cpp backends/moss-tts-cpp backends/magpie-tts-cpp backends/vllm-cpp backends/omnivoice-cpp backends/vibevoice-cpp backends/localvqe backends/tinygrad backends/sherpa-onnx backends/ds4 backends/ds4-darwin backends/liquid-audio backends/supertonic backends/depth-anything-cpp backends/privacy-filter backends/privacy-filter-darwin backends/audio-cpp backends/audio-cpp-darwin
+.NOTPARALLEL: backends/diffusers backends/llama-cpp backends/turboquant backends/bonsai backends/outetts backends/piper backends/stablediffusion-ggml backends/trellis2cpp backends/trellis2cpp-darwin backends/whisper backends/crispasr backends/parakeet-cpp backends/moss-transcribe-cpp backends/nemo-speech-cpp backends/faster-whisper backends/silero-vad backends/local-store backends/valkey-store backends/cloud-proxy backends/huggingface backends/rfdetr backends/rfdetr-cpp backends/insightface backends/speaker-recognition backends/kitten-tts backends/kokoro backends/chatterbox backends/llama-cpp-darwin backends/neutts build-darwin-python-backend build-darwin-go-backend backends/mlx backends/mlx-video backends/diffuser-darwin backends/mlx-vlm backends/mlx-audio backends/mlx-distributed backends/stablediffusion-ggml-darwin backends/vllm backends/vllm-omni backends/longcat-video backends/sglang backends/moonshine backends/pocket-tts backends/qwen-tts backends/faster-qwen3-tts backends/qwen-asr backends/nemo backends/voxcpm backends/whisperx backends/ace-step backends/acestep-cpp backends/fish-speech backends/voxtral backends/opus backends/trl backends/llama-cpp-quantization backends/kokoros backends/sam3-cpp backends/qwen3-tts-cpp backends/moss-tts-cpp backends/magpie-tts-cpp backends/vllm-cpp backends/omnivoice-cpp backends/vibevoice-cpp backends/localvqe backends/tinygrad backends/sherpa-onnx backends/ds4 backends/ds4-darwin backends/liquid-audio backends/supertonic backends/depth-anything-cpp backends/privacy-filter backends/privacy-filter-darwin backends/audio-cpp backends/audio-cpp-darwin
+.NOTPARALLEL: backends/whisper-medusa
+.NOTPARALLEL: backends/funasr

 GOCMD=go
 GOTEST=$(GOCMD) test
@@ -240,7 +242,7 @@ 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)

@@ -393,9 +395,17 @@ 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
@@ -604,6 +614,7 @@ prepare-test-extra: protogen-python
 	$(MAKE) -C backend/python/vllm
 	$(MAKE) -C backend/python/vllm-omni
 	$(MAKE) -C backend/python/longcat-video
+	$(MAKE) -C backend/python/mlx-video
 	$(MAKE) -C backend/python/sglang
 	$(MAKE) -C backend/python/vibevoice
 	$(MAKE) -C backend/python/liquid-audio
@@ -613,9 +624,11 @@ prepare-test-extra: protogen-python
 	$(MAKE) -C backend/python/fish-speech
 	$(MAKE) -C backend/python/faster-qwen3-tts
 	$(MAKE) -C backend/python/qwen-asr
+	$(MAKE) -C backend/python/funasr
 	$(MAKE) -C backend/python/nemo
 	$(MAKE) -C backend/python/voxcpm
 	$(MAKE) -C backend/python/faster-whisper
+	$(MAKE) -C backend/python/whisper-medusa
 	$(MAKE) -C backend/python/whisperx
 	$(MAKE) -C backend/python/ace-step
 	$(MAKE) -C backend/python/trl
@@ -636,6 +649,7 @@ test-extra: prepare-test-extra
 	$(MAKE) -C backend/python/vllm test
 	$(MAKE) -C backend/python/vllm-omni test
 	$(MAKE) -C backend/python/longcat-video test
+	$(MAKE) -C backend/python/mlx-video test
 	$(MAKE) -C backend/python/vibevoice test
 	$(MAKE) -C backend/python/liquid-audio test
 	$(MAKE) -C backend/python/moonshine test
@@ -644,9 +658,11 @@ test-extra: prepare-test-extra
 	$(MAKE) -C backend/python/fish-speech test
 	$(MAKE) -C backend/python/faster-qwen3-tts test
 	$(MAKE) -C backend/python/qwen-asr test
+	$(MAKE) -C backend/python/funasr test
 	$(MAKE) -C backend/python/nemo test
 	$(MAKE) -C backend/python/voxcpm test
 	$(MAKE) -C backend/python/faster-whisper test
+	$(MAKE) -C backend/python/whisper-medusa test
 	$(MAKE) -C backend/python/whisperx test
 	$(MAKE) -C backend/python/ace-step test
 	$(MAKE) -C backend/python/trl test
@@ -1250,6 +1266,10 @@ backends/mlx:
 	BACKEND=mlx $(MAKE) build-darwin-python-backend
 	./local-ai backends install "ocifile://$(abspath ./backend-images/mlx.tar)"

+backends/mlx-video:
+	BACKEND=mlx-video $(MAKE) build-darwin-python-backend
+	./local-ai backends install "ocifile://$(abspath ./backend-images/mlx-video.tar)"
+
 backends/diffuser-darwin:
 	BACKEND=diffusers $(MAKE) build-darwin-python-backend
 	./local-ai backends install "ocifile://$(abspath ./backend-images/diffusers.tar)"
@@ -1336,6 +1356,7 @@ BACKEND_RERANKERS = rerankers|python|.|false|true
 BACKEND_TRANSFORMERS = transformers|python|.|false|true
 BACKEND_OUTETTS = outetts|python|.|false|true
 BACKEND_FASTER_WHISPER = faster-whisper|python|.|false|true
+BACKEND_WHISPER_MEDUSA = whisper-medusa|python|.|false|true
 BACKEND_COQUI = coqui|python|.|false|true
 BACKEND_RFDETR = rfdetr|python|.|false|true
 BACKEND_INSIGHTFACE = insightface|python|.|false|true
@@ -1357,6 +1378,7 @@ BACKEND_QWEN_TTS = qwen-tts|python|.|false|true
 BACKEND_FISH_SPEECH = fish-speech|python|.|false|true
 BACKEND_FASTER_QWEN3_TTS = faster-qwen3-tts|python|.|false|true
 BACKEND_QWEN_ASR = qwen-asr|python|.|false|true
+BACKEND_FUNASR = funasr|python|.|false|true
 BACKEND_NEMO = nemo|python|.|false|true
 BACKEND_VOXCPM = voxcpm|python|.|false|true
 BACKEND_WHISPERX = whisperx|python|.|false|true
@@ -1427,6 +1449,7 @@ $(eval $(call generate-docker-build-target,$(BACKEND_RERANKERS)))
 $(eval $(call generate-docker-build-target,$(BACKEND_TRANSFORMERS)))
 $(eval $(call generate-docker-build-target,$(BACKEND_OUTETTS)))
 $(eval $(call generate-docker-build-target,$(BACKEND_FASTER_WHISPER)))
+$(eval $(call generate-docker-build-target,$(BACKEND_WHISPER_MEDUSA)))
 $(eval $(call generate-docker-build-target,$(BACKEND_COQUI)))
 $(eval $(call generate-docker-build-target,$(BACKEND_RFDETR)))
 $(eval $(call generate-docker-build-target,$(BACKEND_INSIGHTFACE)))
@@ -1448,6 +1471,7 @@ $(eval $(call generate-docker-build-target,$(BACKEND_QWEN_TTS)))
 $(eval $(call generate-docker-build-target,$(BACKEND_FISH_SPEECH)))
 $(eval $(call generate-docker-build-target,$(BACKEND_FASTER_QWEN3_TTS)))
 $(eval $(call generate-docker-build-target,$(BACKEND_QWEN_ASR)))
+$(eval $(call generate-docker-build-target,$(BACKEND_FUNASR)))
 $(eval $(call generate-docker-build-target,$(BACKEND_NEMO)))
 $(eval $(call generate-docker-build-target,$(BACKEND_VOXCPM)))
 $(eval $(call generate-docker-build-target,$(BACKEND_WHISPERX)))
@@ -1477,6 +1501,8 @@ docker-save-%: backend-images
 	docker save local-ai-backend:$* -o backend-images/$*.tar

 docker-build-backends: docker-build-llama-cpp docker-build-ik-llama-cpp docker-build-turboquant docker-build-bonsai docker-build-ds4 docker-build-rerankers docker-build-vllm docker-build-vllm-omni docker-build-longcat-video docker-build-sglang docker-build-transformers docker-build-outetts docker-build-diffusers docker-build-kokoro docker-build-faster-whisper docker-build-crispasr docker-build-coqui docker-build-chatterbox docker-build-vibevoice docker-build-liquid-audio docker-build-moonshine docker-build-pocket-tts docker-build-qwen-tts docker-build-fish-speech docker-build-faster-qwen3-tts docker-build-qwen-asr docker-build-nemo docker-build-voxcpm docker-build-whisperx docker-build-ace-step docker-build-acestep-cpp docker-build-voxtral docker-build-mlx-distributed docker-build-trl docker-build-llama-cpp-quantization docker-build-tinygrad docker-build-kokoros docker-build-sam3-cpp docker-build-rfdetr-cpp docker-build-qwen3-tts-cpp docker-build-moss-tts-cpp docker-build-magpie-tts-cpp docker-build-vllm-cpp docker-build-omnivoice-cpp docker-build-vibevoice-cpp docker-build-localvqe docker-build-insightface docker-build-speaker-recognition docker-build-sherpa-onnx docker-build-cloud-proxy docker-build-supertonic docker-build-depth-anything-cpp docker-build-moss-transcribe-cpp docker-build-nemo-speech-cpp docker-build-privacy-filter docker-build-trellis2cpp docker-build-valkey-store docker-build-audio-cpp
+docker-build-backends: docker-build-whisper-medusa
+docker-build-backends: docker-build-funasr

 ########################################################
 ### Mock Backend for E2E Tests
@@ -6,13 +6,14 @@ ARG APT_PORTS_MIRROR=""
 # ASR, VAD, diarization, source separation and music generation, wrapped as a
 # LocalAI gRPC backend.
 #
-# BASE_IMAGE is ubuntu:24.04 for cpu and vulkan builds, or
-# nvidia/cuda:<ver>-devel-ubuntu24.04 for cublas builds; both ship apt and
-# Ubuntu Noble packages, and the CUDA base additionally provides
-# /usr/local/cuda. BUILD_TYPE selects the engine backend in the Makefile:
-# "" = portable CPU with all ggml CPU variants, "cublas" ->
-# -DENGINE_ENABLE_CUDA=ON, "vulkan" -> -DENGINE_ENABLE_VULKAN=ON. Darwin
-# (Metal) builds bypass this Dockerfile entirely.
+# BASE_IMAGE is ubuntu:24.04 for cpu and vulkan builds,
+# nvidia/cuda:<ver>-devel-ubuntu24.04 for cublas builds, or
+# rocm/dev-ubuntu-24.04:<ver> for hipblas builds. All ship apt and Ubuntu Noble
+# packages; the GPU bases also provide their toolkits. BUILD_TYPE selects the
+# engine backend in the Makefile: "" = portable CPU with all ggml CPU variants,
+# "cublas" -> -DENGINE_ENABLE_CUDA=ON, "hipblas" -> -DENGINE_ENABLE_HIP=ON,
+# and "vulkan" -> -DENGINE_ENABLE_VULKAN=ON. Darwin (Metal) builds bypass this
+# Dockerfile entirely.
 #
 # Upstream needs GCC 13 or newer, which ubuntu:24.04 and the CUDA 12/13
 # devel-ubuntu24.04 images all provide.
@@ -62,7 +63,7 @@ ENV BUILD_TYPE=${BUILD_TYPE} \
    APT_MIRROR=${APT_MIRROR} \
    APT_PORTS_MIRROR=${APT_PORTS_MIRROR} \
    DEBIAN_FRONTEND=noninteractive \
-    PATH=/usr/local/cuda/bin:${PATH}
+    PATH=/opt/rocm/bin:/usr/local/cuda/bin:${PATH}

 WORKDIR /build

@@ -73,7 +74,7 @@ WORKDIR /build
 # fallback of its own.
 #
 # BUILD_TYPE=vulkan additionally needs the loader headers and glslc; both are in
-# Noble. The CUDA toolkit for BUILD_TYPE=cublas comes from BASE_IMAGE.
+# Noble. The CUDA and ROCm toolkits come from their matching BASE_IMAGE.
 RUN --mount=type=bind,source=.docker/apt-mirror.sh,target=/usr/local/sbin/apt-mirror \
    sh /usr/local/sbin/apt-mirror && \
    apt-get update && \
@@ -83,6 +84,9 @@ RUN --mount=type=bind,source=.docker/apt-mirror.sh,target=/usr/local/sbin/apt-mi
    if [ "${BUILD_TYPE}" = "vulkan" ]; then \
        apt-get install -y --no-install-recommends libvulkan-dev glslc; \
    fi && \
+    if [ "${BUILD_TYPE}" = "hipblas" ]; then \
+        apt-get install -y --no-install-recommends hipblas-dev rocblas-dev; \
+    fi && \
    if [ "${TARGETARCH}" = "arm64" ]; then \
        apt-get install -y --no-install-recommends gcc-14 g++-14; \
    fi && \
@@ -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/. ./
@@ -47,7 +47,7 @@ The backend system provides language-specific Dockerfiles that handle the build
 - **mlx**: Apple Silicon optimization
 - **diffusers**: Stable Diffusion models
 - **longcat-video**: CUDA text/image-to-video and speech-driven avatar generation
- **Audio**: coqui, faster-whisper, kitten-tts
+- **Audio**: coqui, faster-whisper, funasr, kitten-tts
 - **Vision**: mlx-vlm, rfdetr
 - **Specialized**: rerankers, chatterbox, kokoro

@@ -499,6 +499,10 @@ message ModelOptions {
  // applied verbatim to the backend's engine constructor (e.g. vLLM AsyncEngineArgs).
  // Unknown keys produce an error at LoadModel time.
  string EngineArgs = 73;
+  string OriginalConfigFile = 77;
+
+  // EnvVars carries environment variables to be passed to the backend process.
+  map<string, string> EnvVars = 76;

  // Proxy carries the cloud-proxy backend's per-model configuration.
  // Empty for non-proxy backends.
@@ -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?=89a0e9803380880305e9e1b83c93614f9df2c893
+AUDIO_CPP_VERSION?=efb04233dab73aeee4b2912042a90e7b36329061
 AUDIO_CPP_REPO?=https://github.com/0xShug0/audio.cpp

 CURRENT_MAKEFILE_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
@@ -77,6 +77,16 @@ endif

 ifeq ($(BUILD_TYPE),cublas)
    CMAKE_ARGS += -DENGINE_ENABLE_CUDA=ON "-DCMAKE_CUDA_ARCHITECTURES=$(CUDA_ARCHITECTURES)"
+else ifeq ($(BUILD_TYPE),hipblas)
+    ROCM_HOME ?= /opt/rocm
+    ROCM_PATH ?= /opt/rocm
+    export CXX=$(ROCM_HOME)/llvm/bin/clang++
+    export CC=$(ROCM_HOME)/llvm/bin/clang
+    AMDGPU_TARGETS ?= gfx908,gfx90a,gfx942,gfx950,gfx1030,gfx1100,gfx1101,gfx1102,gfx1151,gfx1200,gfx1201
+    # audio.cpp forwards GPU_TARGETS to CMake's semicolon-delimited HIP list.
+    comma := ,
+    HIP_TARGETS := $(subst $(comma),;,$(AMDGPU_TARGETS))
+    CMAKE_ARGS += -DENGINE_ENABLE_HIP=ON "-DGPU_TARGETS=$(HIP_TARGETS)"
 else ifeq ($(BUILD_TYPE),vulkan)
    CMAKE_ARGS += -DENGINE_ENABLE_VULKAN=ON
 else ifeq ($(UNAME_S),Darwin)
@@ -22,7 +22,8 @@ bool voice_is_reference_file(const std::string &voice) {

 RequestShape build_tts_shape(const backend::TTSRequest &request) {
    RequestShape shape;
-    shape.has_voice_reference = voice_is_reference_file(request.voice());
+    shape.has_voice_reference = voice_is_reference_file(request.voice()) ||
+                                request.params().find("multi_reference_cond") != request.params().end();
    // !empty() as well as has_instructions(), and it must match the guard in
    // build_tts_request: a request whose instructions are an empty string
    // carries no style condition, so telling routing to prefer VoiceDesign for
@@ -128,6 +128,14 @@ static void test_tts_shape() {
        check(!shape.has_voice_reference,
              "shape: a preset name is not a voice reference");
    }
+    {
+        backend::TTSRequest request;
+        (*request.mutable_params())["multi_reference_cond"] =
+            R"([{"audio":"one.wav","text":"one"}])";
+        const auto shape = build_tts_shape(request);
+        check(shape.has_voice_reference,
+              "shape: multi-reference conditioning is a voice reference");
+    }
    {
        backend::TTSRequest request;
        request.set_voice(dir.string());
@@ -103,6 +103,9 @@ engine::core::BackendType parse_backend_type(const std::string &value) {
    if (value == "cuda") {
        return engine::core::BackendType::Cuda;
    }
+    if (value == "hip" || value == "rocm") {
+        return engine::core::BackendType::Hip;
+    }
    if (value == "vulkan") {
        return engine::core::BackendType::Vulkan;
    }
@@ -116,7 +119,7 @@ engine::core::BackendType parse_backend_type(const std::string &value) {
        return engine::core::BackendType::Cpu;
    }
    throw ConfigError("audio-cpp: unknown backend option '" + value +
-                      "'. Known backends: cpu, cuda, vulkan, metal, best");
+                      "'. Known backends: cpu, cuda, hip, rocm, vulkan, metal, best");
 }

 std::filesystem::path executable_directory() {
@@ -59,6 +59,12 @@ bool starts_with(const std::string &value, const std::string &prefix) {
           value.compare(0, prefix.size(), prefix) == 0;
 }

+bool is_known_backend(const std::string &value) {
+    return value == "cpu" || value == "cuda" || value == "hip" ||
+           value == "rocm" || value == "vulkan" || value == "metal" ||
+           value == "best";
+}
+
 } // namespace

 ParsedOptions parse_model_options(const std::vector<std::string> &entries) {
@@ -107,6 +113,12 @@ ParsedOptions parse_model_options(const std::vector<std::string> &entries) {
        } else if (key == "task") {
            parsed.options.task = value;
        } else if (key == "backend") {
+            if (!is_known_backend(value)) {
+                parsed.error = "audio-cpp: unknown backend option '" + value +
+                               "'. Known backends: cpu, cuda, hip, rocm, "
+                               "vulkan, metal, best";
+                return parsed;
+            }
            parsed.options.backend = value;
        } else if (key == "model_spec_override") {
            parsed.options.model_spec_override = value;
@@ -17,7 +17,7 @@ struct ModelOptions {
    std::string family;
    // Pins the audio.cpp task, overriding RPC-based routing. Empty means route.
    std::string task;
-    // ggml backend: cpu, cuda, vulkan, metal, best.
+    // ggml backend: cpu, cuda, hip (or rocm), vulkan, metal, best.
    std::string backend = "cpu";
    int device = 0;
    // True once a `device:` entry has been seen. 0 is both the default and a
@@ -75,6 +75,11 @@ static void test_scalar_options() {
    check(parse_model_options({"live_idle_timeout_ms:0"}).options.live_idle_timeout_ms == 0,
          "an explicit 0 turns the live idle limit off rather than reverting to "
          "the default");
+
+    check(parse_model_options({"backend:hip"}).error.empty(),
+          "HIP backend option is accepted");
+    check(parse_model_options({"backend:rocm"}).error.empty(),
+          "ROCm backend alias is accepted");
 }

 // Values containing colons must survive: split on the FIRST colon only.
@@ -124,6 +129,8 @@ static void test_errors() {
          "negative device is rejected");
    check(!parse_model_options({"threads:x"}).error.empty(),
          "non-numeric threads is rejected");
+    check(!parse_model_options({"backend:unknown"}).error.empty(),
+          "unknown compute backend is rejected before model loading");

    // Values too large for int must be rejected, not silently wrapped into a
    // negative device index that then reaches the ggml backend selector.
@@ -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")
@@ -1,10 +1,10 @@
 # ds4 backend Makefile.
 #
-# Upstream pin lives below as DS4_VERSION?=8db89fe083ae4d17c9a2428ccd29803d3ae8f577
+# 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?=8db89fe083ae4d17c9a2428ccd29803d3ae8f577
+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
@@ -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_) {
@@ -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;
@@ -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;
+}
@@ -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
@@ -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;
+}
@@ -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;
    }

@@ -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
@@ -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;
+}
@@ -1,5 +1,5 @@

-IK_LLAMA_VERSION?=15dddc60b3fc937a9e2a210359ecce392ccdf446
+IK_LLAMA_VERSION?=3bb386eb68ffee0a5dc7db21da0735d594929eeb
 LLAMA_REPO?=https://github.com/ikawrakow/ik_llama.cpp

 CMAKE_ARGS?=
@@ -125,4 +125,15 @@ if(LLAMA_GRPC_BUILD_TESTS)
    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)
+
+    add_executable(model_load_error_test model_load_error_test.cpp model_load_error.h)
+    target_include_directories(model_load_error_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+    target_compile_features(model_load_error_test PRIVATE cxx_std_17)
+    add_test(NAME model_load_error_test COMMAND model_load_error_test)
+
+    # Dead-stream tracker (standard library only).
+    add_executable(stream_peer_test stream_peer_test.cpp stream_peer.h)
+    target_include_directories(stream_peer_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
+    target_compile_features(stream_peer_test PRIVATE cxx_std_17)
+    add_test(NAME stream_peer_test COMMAND stream_peer_test)
 endif()
@@ -1,5 +1,5 @@

-LLAMA_VERSION?=d7bd3bfcad3e29c7e49fd26f38c79ee3e9a3fd6b
+LLAMA_VERSION?=df03399b885831b2a1603b3abb0d8c156808e363
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp

 CMAKE_ARGS?=
@@ -53,9 +53,11 @@
 #include "arg.h"
 #include "chat-auto-parser.h"
 #include "llama_compat.h" // fork-skew switches, generated by prepare.sh
+#include "model_load_error.h"
 #include "thread_params.h"
 #include "message_content.h"
 #include "passthrough_options.h"
+#include "stream_peer.h"
 #include "tts_request_options.h"
 #include <getopt.h>
 #include <grpcpp/ext/proto_server_reflection_plugin.h>
@@ -1628,7 +1630,8 @@ public:
            {
                std::lock_guard<std::mutex> lock(error_capture_data.error_mutex);
                if (!error_capture_data.captured_error.empty()) {
-                    error_msg += ". Error: " + error_capture_data.captured_error;
+                    error_msg += ". Error: " +
+                                 localai::model_load_error_with_hint(error_capture_data.captured_error);
                } else {
                    error_msg += ". Model file may not exist or be invalid.";
                }
@@ -2266,6 +2269,11 @@ public:
        // such concept, so there is nothing to emit — the real tokens arrive in
        // the loop below. Feeding this null into build_reply_from_json would
        // throw (uncaught) and surface as a generic RPC error.
+        // A write that returns false means the peer is gone for good. Track it
+        // so the loop below stops decoding instead of feeding a dead stream —
+        // see stream_peer.h for why that matters to everyone else's requests.
+        llama_grpc::StreamPeer peer;
+
        if (first_res_json.is_null()) {
            // skip the begin-of-stream marker
        } else if (first_res_json.is_array()) {
@@ -2278,17 +2286,21 @@ public:
                if (!is_role_init) {
                    attach_chat_deltas(reply, first_result.get());
                }
-                writer->Write(reply);
+                peer.observe_write(writer->Write(reply));
+                if (peer.gone()) {
+                    break;
+                }
            }
        } else {
            auto reply = build_reply_from_json(first_res_json, first_result.get());
            attach_chat_deltas(reply, first_result.get());
-            writer->Write(reply);
+            peer.observe_write(writer->Write(reply));
        }

        // Process subsequent results
        while (rd.has_next()) {
-            if (context->IsCancelled()) {
+            peer.observe_cancelled(context->IsCancelled());
+            if (peer.gone()) {
                break;
            }

@@ -2309,17 +2321,22 @@ public:
                    if (!is_role_init) {
                        attach_chat_deltas(reply, result.get());
                    }
-                    writer->Write(reply);
+                    peer.observe_write(writer->Write(reply));
+                    if (peer.gone()) {
+                        break;
+                    }
                }
            } else {
                auto reply = build_reply_from_json(res_json, result.get());
                attach_chat_deltas(reply, result.get());
-                writer->Write(reply);
+                peer.observe_write(writer->Write(reply));
            }
        }

-        // Check if context was cancelled during processing
-        if (context->IsCancelled()) {
+        // Returning here is what releases the slot: ~server_response_reader()
+        // posts SERVER_TASK_TYPE_CANCEL for whatever is still decoding.
+        peer.observe_cancelled(context->IsCancelled());
+        if (peer.gone()) {
            return grpc::Status(grpc::StatusCode::CANCELLED, "Request cancelled by client");
        }

@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include <string>
+
+namespace localai {
+
+inline std::string model_load_error_with_hint(const std::string& error) {
+  const std::string mismatch = "wrong number of tensors; expected ";
+  const std::string got = ", got ";
+  const std::string::size_type mismatch_pos = error.find(mismatch);
+  if (mismatch_pos == std::string::npos ||
+      error.find(got, mismatch_pos + mismatch.size()) == std::string::npos) {
+    return error;
+  }
+
+  return error +
+         " Hint: the model may be incompatible with this llama.cpp backend "
+         "or the GGUF file may be corrupt. Try a newer compatible backend "
+         "and verify or re-download the model file.";
+}
+
+}  // namespace localai
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+
+#include "model_load_error.h"
+
+#include <cassert>
+#include <string>
+
+int main() {
+  const std::string issue_error =
+      "llama_model_load: error loading model: done_getting_tensors: wrong number of tensors; expected 2131, got 720; "
+      "llama_model_load_from_file_impl: failed to load model";
+  const std::string issue_result = localai::model_load_error_with_hint(issue_error);
+  assert(issue_result.compare(0, issue_error.size(), issue_error) == 0);
+  assert(issue_result.find("incompatible") != std::string::npos);
+  assert(issue_result.find("corrupt") != std::string::npos);
+
+  const std::string generic_error =
+      "wrong number of tensors; expected 42, got 17";
+  const std::string generic_result = localai::model_load_error_with_hint(generic_error);
+  assert(generic_result.compare(0, generic_error.size(), generic_error) == 0);
+  assert(generic_result.size() > generic_error.size());
+
+  const std::string unrelated_error = "failed to open GGUF file";
+  assert(localai::model_load_error_with_hint(unrelated_error) == unrelated_error);
+}
@@ -15,12 +15,40 @@ 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

 cp -r CMakeLists.txt llama.cpp/tools/grpc-server/
 cp -r grpc-server.cpp llama.cpp/tools/grpc-server/
+# Model-load diagnostics (included by grpc-server.cpp) and their standalone
+# regression test.
+cp -r model_load_error.h llama.cpp/tools/grpc-server/
+cp -r model_load_error_test.cpp llama.cpp/tools/grpc-server/
 # Shared message-reconstruction helpers (included by grpc-server.cpp) and their
 # unit test (compiled only when -DLLAMA_GRPC_BUILD_TESTS=ON).
 cp -r message_content.h llama.cpp/tools/grpc-server/
@@ -39,6 +67,10 @@ cp -r thread_params_test.cpp llama.cpp/tools/grpc-server/
 # test (run via backend/cpp/run-unit-tests.sh; also buildable under ctest).
 cp -r parent_watch.h llama.cpp/tools/grpc-server/
 cp -r parent_watch_test.cpp llama.cpp/tools/grpc-server/
+# Dead-stream tracker (included by grpc-server.cpp) and its standalone unit
+# test (run via backend/cpp/run-unit-tests.sh; also buildable under ctest).
+cp -r stream_peer.h llama.cpp/tools/grpc-server/
+cp -r stream_peer_test.cpp llama.cpp/tools/grpc-server/
 cp -rfv llama.cpp/vendor/nlohmann/json.hpp llama.cpp/tools/grpc-server/
 cp -rfv llama.cpp/vendor/cpp-httplib/httplib.h llama.cpp/tools/grpc-server/

@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: MIT
+#pragma once
+
+namespace llama_grpc {
+
+// Tracks whether a server-streaming RPC still has somewhere to send tokens.
+//
+// grpc::ServerWriter::Write() returns false once the peer is gone, and a
+// stream never recovers afterwards. Ignoring that result is not harmless: the
+// handler goes on draining decoded tokens into a dead stream, so the llama.cpp
+// slot stays busy for the rest of the request's token budget. A model config
+// with no max_tokens and a large context turns that into tens of minutes per
+// abandoned request, and the slots are exactly what every other request queues
+// behind.
+//
+// Returning as soon as the peer is gone is what frees the slot: the handler's
+// server_response_reader then goes out of scope and its destructor posts
+// SERVER_TASK_TYPE_CANCEL for whatever is still decoding.
+class StreamPeer {
+public:
+    // Records the outcome of a Write(). Once a write has failed the peer stays
+    // gone -- a later write cannot succeed on a broken stream.
+    void observe_write(bool ok) noexcept {
+        if (!ok) {
+            gone_ = true;
+        }
+    }
+
+    // Folds in the RPC's own cancellation flag, so callers have a single
+    // predicate to test rather than two that can disagree.
+    void observe_cancelled(bool cancelled) noexcept {
+        if (cancelled) {
+            gone_ = true;
+        }
+    }
+
+    bool gone() const noexcept { return gone_; }
+    bool alive() const noexcept { return !gone_; }
+
+private:
+    bool gone_ = false;
+};
+
+} // namespace llama_grpc
@@ -0,0 +1,67 @@
+#include "stream_peer.h"
+
+#include <cstdio>
+
+namespace {
+
+int failures = 0;
+
+void check(bool condition, const char *what) {
+    if (!condition) {
+        std::fprintf(stderr, "FAIL: %s\n", what);
+        ++failures;
+    }
+}
+
+} // namespace
+
+int main() {
+    {
+        llama_grpc::StreamPeer peer;
+        check(peer.alive(), "a fresh peer is alive");
+        check(!peer.gone(), "a fresh peer is not gone");
+    }
+
+    {
+        llama_grpc::StreamPeer peer;
+        peer.observe_write(true);
+        peer.observe_write(true);
+        check(peer.alive(), "successful writes keep the peer alive");
+    }
+
+    {
+        llama_grpc::StreamPeer peer;
+        peer.observe_write(false);
+        check(peer.gone(), "a failed write marks the peer gone");
+    }
+
+    {
+        // The whole point of the guard: a stream never comes back, so a later
+        // success must not resurrect a peer an earlier failure retired.
+        llama_grpc::StreamPeer peer;
+        peer.observe_write(false);
+        peer.observe_write(true);
+        check(peer.gone(), "a failed write is sticky across later writes");
+    }
+
+    {
+        llama_grpc::StreamPeer peer;
+        peer.observe_cancelled(false);
+        check(peer.alive(), "an uncancelled RPC keeps the peer alive");
+        peer.observe_cancelled(true);
+        check(peer.gone(), "cancellation marks the peer gone");
+    }
+
+    {
+        llama_grpc::StreamPeer peer;
+        peer.observe_cancelled(true);
+        peer.observe_cancelled(false);
+        check(peer.gone(), "cancellation is sticky across later checks");
+    }
+
+    if (failures != 0) {
+        std::fprintf(stderr, "%d check(s) failed\n", failures);
+        return 1;
+    }
+    return 0;
+}
@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)

 # CrispASR version (release tag)
 CRISPASR_REPO?=https://github.com/CrispStrobe/CrispASR
-CRISPASR_VERSION?=ae4474dd8306384a0e697183d863dfc52e69a2fb
+CRISPASR_VERSION?=301acd87b036764973b8bfba71e0a21818036d33
 SO_TARGET?=libgocrispasr.so

 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
@@ -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
@@ -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?=739992d10bf9472c46dcd4622b14d2b20766c58d
+DEPTHANYTHING_VERSION?=14f7461d1f704761a038ac9f50dbde8fdb7275e2

 ifeq ($(NATIVE),false)
 	CMAKE_ARGS+=-DGGML_NATIVE=OFF
@@ -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)
 	}
@@ -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"))
+	})
+})
@@ -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?=4f9676226f667d14608487df744f375db87127f8
+NEMO_SPEECH_VERSION?=a5b6953c4a579a2bbd1c0913ad8a85c2a4d99953
 NEMO_SPEECH_REPO?=https://github.com/NVIDIA/NeMo-Speech.cpp

 GOCMD?=go
@@ -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
@@ -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)
 	}
@@ -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"))
+	})
+})
@@ -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?=be0e34480dada95f8ce9a021bbb95c5de85d67c7
+STABLEDIFFUSION_GGML_VERSION?=b68d58624d227682eb4b95ef8bcf569cd1311eb5

 CMAKE_ARGS+=-DGGML_MAX_NAME=128

@@ -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;
 }
-
@@ -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?=150b37852c123f7855fb219b37347572ca9427e7
+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
@@ -128,9 +128,40 @@ 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) {
@@ -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())
+	})
+})
@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)

 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggml-org/whisper.cpp
-WHISPER_CPP_VERSION?=978113305b2ead22249b881deafa131dc8884911
+WHISPER_CPP_VERSION?=c44b60b8053bbf2a5c1e014f11323fb3f2485177
 SO_TARGET?=libgowhisper.so

 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
@@ -0,0 +1,35 @@
+package main
+
+import (
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+// Specs for resolveAddr run under the suite bootstrap in gowhisper_test.go
+// (TestWhisper); they need no native library, so they never skip.
+var _ = Describe("resolveAddr", func() {
+	It("prefers an explicitly set -addr over a positional argument", func() {
+		Expect(resolveAddr("127.0.0.1:12345", true, []string{"127.0.0.1:59999"})).To(Equal("127.0.0.1:12345"))
+	})
+
+	It("keeps an explicit -addr equal to the default over a positional argument", func() {
+		Expect(resolveAddr(defaultAddr, true, []string{"127.0.0.1:59999"})).To(Equal(defaultAddr))
+	})
+
+	It("falls back to the positional argument when -addr is unset", func() {
+		Expect(resolveAddr(defaultAddr, false, []string{"127.0.0.1:59999"})).To(Equal("127.0.0.1:59999"))
+	})
+
+	It("keeps the default when -addr is unset and no positional argument is given", func() {
+		Expect(resolveAddr(defaultAddr, false, nil)).To(Equal(defaultAddr))
+	})
+
+	It("uses the first positional argument among several", func() {
+		Expect(resolveAddr(defaultAddr, false, []string{"127.0.0.1:59999", "extra"})).To(Equal("127.0.0.1:59999"))
+	})
+
+	It("treats an explicitly empty -addr as unset", func() {
+		Expect(resolveAddr("", true, []string{"127.0.0.1:59999"})).To(Equal("127.0.0.1:59999"))
+		Expect(resolveAddr("", true, nil)).To(Equal(defaultAddr))
+	})
+})
@@ -10,10 +10,31 @@ import (
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
 )

-var (
-	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+const (
+	defaultAddr = "localhost:50051"
 )

+var (
+	addr = flag.String("addr", defaultAddr, "the address to listen on")
+)
+
+// resolveAddr picks the address the gRPC server binds to. An explicitly set
+// -addr always wins. Launchers may hand us the listen address as a bare
+// positional argument, which Go's flag package silently drops — honour it
+// next so the server binds the port its caller actually allocated instead of
+// the default one (#11623). An explicitly empty -addr counts as unset:
+// binding the empty address would listen on an OS-chosen port on every
+// interface instead of the one the caller allocated.
+func resolveAddr(flagAddr string, addrSet bool, args []string) string {
+	if addrSet && flagAddr != "" {
+		return flagAddr
+	}
+	if len(args) > 0 {
+		return args[0]
+	}
+	return defaultAddr
+}
+
 type LibFuncs struct {
 	FuncPtr any
 	Name    string
@@ -61,7 +82,16 @@ func main() {

 	flag.Parse()

-	if err := grpc.StartServer(*addr, &Whisper{}); err != nil {
+	// flag.Visit reports only flags that were explicitly set, so an -addr
+	// equal to the default is still distinguished from an untouched one.
+	addrSet := false
+	flag.Visit(func(f *flag.Flag) {
+		if f.Name == "addr" {
+			addrSet = true
+		}
+	})
+
+	if err := grpc.StartServer(resolveAddr(*addr, addrSet, flag.Args()), &Whisper{}); err != nil {
 		panic(err)
 	}
 }
@@ -150,6 +150,7 @@
    - audio-transcription
    - CPU
    - CUDA
+    - HIP
    - Metal
  # No vulkan key: the vulkan image would carry a Vulkan loader with no Mesa ICD
  # (see the audio-cpp block in .github/backend-matrix.yml). Pointing a
@@ -161,6 +162,7 @@
    nvidia: "cuda12-audio-cpp"
    nvidia-cuda-12: "cuda12-audio-cpp"
    nvidia-cuda-13: "cuda13-audio-cpp"
+    amd: "rocm-audio-cpp"
    metal: "metal-audio-cpp"
    metal-darwin-arm64: "metal-audio-cpp"
 - &whispercpp
@@ -1060,6 +1062,21 @@
    nvidia-l4t: "nvidia-l4t-mlx"
    nvidia-l4t-cuda-12: "nvidia-l4t-mlx"
    nvidia-l4t-cuda-13: "cuda13-nvidia-l4t-arm64-mlx"
+- &mlx-video
+  name: "mlx-video"
+  icon: https://avatars.githubusercontent.com/u/102832242?s=200&v=4
+  urls:
+    - https://github.com/Blaizzy/mlx-video
+  license: MIT
+  description: |
+      Generate videos with LTX-2 and converted Wan2.1/Wan2.2 checkpoints using
+      MLX on Apple Silicon.
+  tags:
+    - text-to-video
+    - image-to-video
+    - MLX
+  capabilities:
+    metal: "metal-mlx-video"
 - &mlx-vlm
  name: "mlx-vlm"
  icon: https://avatars.githubusercontent.com/u/102832242?s=200&v=4
@@ -1194,6 +1211,7 @@
  tags:
    - image-generation
    - video-generation
+    - sound-generation
    - diffusion-models
  license: apache-2.0
  alias: "diffusers"
@@ -1518,6 +1536,25 @@
    nvidia-cuda-12: "cuda12-faster-whisper"
    nvidia-l4t: "nvidia-l4t-arm64-faster-whisper"
    nvidia-l4t-cuda-12: "nvidia-l4t-arm64-faster-whisper"
+- &whisper-medusa
+  description: |
+    Whisper-Medusa accelerates Whisper transcription by predicting multiple tokens per decoding step.
+    The upstream checkpoints use custom Transformers generation code and accept audio clips up to 30 seconds.
+  urls:
+    - https://github.com/aiola-lab/whisper-medusa
+    - https://huggingface.co/collections/aiola/whisper-medusa
+  tags:
+    - speech-to-text
+    - transcription
+    - Whisper
+    - Medusa
+  license: MIT
+  name: "whisper-medusa"
+  alias: "whisper-medusa"
+  capabilities:
+    default: "cpu-whisper-medusa"
+    nvidia: "cuda12-whisper-medusa"
+    nvidia-cuda-12: "cuda12-whisper-medusa"
 - &moonshine
  description: |
    Moonshine is a fast, accurate, and efficient speech-to-text transcription model using ONNX Runtime.
@@ -1785,6 +1822,32 @@
    nvidia-l4t-cuda-12: "nvidia-l4t-faster-qwen3-tts"
    nvidia-l4t-cuda-13: "cuda13-nvidia-l4t-arm64-faster-qwen3-tts"
  icon: https://cdn-avatars.huggingface.co/v1/production/uploads/620760a26e3b7210c2ff1943/-s1gyJfvbE1RgO5iBeNOi.png
+- &funasr
+  urls:
+    - https://github.com/modelscope/FunASR
+  description: |
+    FunASR is an industrial-grade speech recognition toolkit supporting 50+ languages.
+    Includes SenseVoice (170x realtime, emotion detection), Paraformer (highest Chinese
+    accuracy), and built-in VAD, punctuation restoration, and speaker diarization.
+  tags:
+    - speech-recognition
+    - ASR
+    - multilingual
+  license: mit
+  name: "funasr"
+  alias: "funasr"
+  capabilities:
+    nvidia: "cuda12-funasr"
+    intel: "intel-funasr"
+    amd: "rocm-funasr"
+    metal: "metal-funasr"
+    default: "cpu-funasr"
+    nvidia-cuda-13: "cuda13-funasr"
+    nvidia-cuda-12: "cuda12-funasr"
+    nvidia-l4t: "nvidia-l4t-funasr"
+    nvidia-l4t-cuda-12: "nvidia-l4t-funasr"
+    nvidia-l4t-cuda-13: "cuda13-nvidia-l4t-arm64-funasr"
+  icon: https://avatars.githubusercontent.com/u/109454077
 - &qwen-asr
  urls:
    - https://github.com/QwenLM/Qwen3-ASR
@@ -2101,6 +2164,7 @@
    nvidia: "cuda12-audio-cpp-development"
    nvidia-cuda-12: "cuda12-audio-cpp-development"
    nvidia-cuda-13: "cuda13-audio-cpp-development"
+    amd: "rocm-audio-cpp-development"
    metal: "metal-audio-cpp-development"
    metal-darwin-arm64: "metal-audio-cpp-development"
 - !!merge <<: *stablediffusionggml
@@ -2164,6 +2228,20 @@
  uri: "quay.io/go-skynet/local-ai-backends:latest-metal-darwin-arm64-mlx"
  mirrors:
    - localai/localai-backends:latest-metal-darwin-arm64-mlx
+- !!merge <<: *mlx-video
+  name: "metal-mlx-video"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-metal-darwin-arm64-mlx-video"
+  mirrors:
+    - localai/localai-backends:latest-metal-darwin-arm64-mlx-video
+- !!merge <<: *mlx-video
+  name: "metal-mlx-video-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-metal-darwin-arm64-mlx-video"
+  mirrors:
+    - localai/localai-backends:master-metal-darwin-arm64-mlx-video
+- !!merge <<: *mlx-video
+  name: "mlx-video-development"
+  capabilities:
+    metal: "metal-mlx-video-development"
 - !!merge <<: *mlx
  name: "metal-mlx-development"
  uri: "quay.io/go-skynet/local-ai-backends:master-metal-darwin-arm64-mlx"
@@ -6368,6 +6446,34 @@
  uri: "quay.io/go-skynet/local-ai-backends:master-nvidia-l4t-cuda-13-arm64-faster-qwen3-tts"
  mirrors:
    - localai/localai-backends:master-nvidia-l4t-cuda-13-arm64-faster-qwen3-tts
+## whisper-medusa
+- !!merge <<: *whisper-medusa
+  name: "whisper-medusa-development"
+  capabilities:
+    default: "cpu-whisper-medusa-development"
+    nvidia: "cuda12-whisper-medusa-development"
+    nvidia-cuda-12: "cuda12-whisper-medusa-development"
+- !!merge <<: *whisper-medusa
+  name: "cpu-whisper-medusa"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-cpu-whisper-medusa"
+  mirrors:
+    - localai/localai-backends:latest-cpu-whisper-medusa
+- !!merge <<: *whisper-medusa
+  name: "cpu-whisper-medusa-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-cpu-whisper-medusa"
+  mirrors:
+    - localai/localai-backends:master-cpu-whisper-medusa
+- !!merge <<: *whisper-medusa
+  name: "cuda12-whisper-medusa"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-gpu-nvidia-cuda-12-whisper-medusa"
+  mirrors:
+    - localai/localai-backends:latest-gpu-nvidia-cuda-12-whisper-medusa
+- !!merge <<: *whisper-medusa
+  name: "cuda12-whisper-medusa-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-gpu-nvidia-cuda-12-whisper-medusa"
+  mirrors:
+    - localai/localai-backends:master-gpu-nvidia-cuda-12-whisper-medusa
+
 ## qwen-asr
 - !!merge <<: *qwen-asr
  name: "qwen-asr-development"
@@ -7054,6 +7160,16 @@
  uri: "quay.io/go-skynet/local-ai-backends:master-gpu-nvidia-cuda-13-audio-cpp"
  mirrors:
    - localai/localai-backends:master-gpu-nvidia-cuda-13-audio-cpp
+- !!merge <<: *audiocpp
+  name: "rocm-audio-cpp"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-gpu-rocm-hipblas-audio-cpp"
+  mirrors:
+    - localai/localai-backends:latest-gpu-rocm-hipblas-audio-cpp
+- !!merge <<: *audiocpp
+  name: "rocm-audio-cpp-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-gpu-rocm-hipblas-audio-cpp"
+  mirrors:
+    - localai/localai-backends:master-gpu-rocm-hipblas-audio-cpp
 - !!merge <<: *audiocpp
  name: "metal-audio-cpp"
  uri: "quay.io/go-skynet/local-ai-backends:latest-metal-darwin-arm64-audio-cpp"
@@ -7064,3 +7180,97 @@
  uri: "quay.io/go-skynet/local-ai-backends:master-metal-darwin-arm64-audio-cpp"
  mirrors:
    - localai/localai-backends:master-metal-darwin-arm64-audio-cpp
+## funasr
+- !!merge <<: *funasr
+  name: "funasr-development"
+  capabilities:
+    nvidia: "cuda12-funasr-development"
+    intel: "intel-funasr-development"
+    amd: "rocm-funasr-development"
+    nvidia-l4t: "nvidia-l4t-funasr-development"
+    metal: "metal-funasr-development"
+    default: "cpu-funasr-development"
+    nvidia-cuda-13: "cuda13-funasr-development"
+    nvidia-cuda-12: "cuda12-funasr-development"
+    nvidia-l4t-cuda-12: "nvidia-l4t-funasr-development"
+    nvidia-l4t-cuda-13: "cuda13-nvidia-l4t-arm64-funasr-development"
+- !!merge <<: *funasr
+  name: "cpu-funasr"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-cpu-funasr"
+  mirrors:
+    - localai/localai-backends:latest-cpu-funasr
+- !!merge <<: *funasr
+  name: "cpu-funasr-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-cpu-funasr"
+  mirrors:
+    - localai/localai-backends:master-cpu-funasr
+- !!merge <<: *funasr
+  name: "cuda12-funasr"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-gpu-nvidia-cuda-12-funasr"
+  mirrors:
+    - localai/localai-backends:latest-gpu-nvidia-cuda-12-funasr
+- !!merge <<: *funasr
+  name: "cuda12-funasr-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-gpu-nvidia-cuda-12-funasr"
+  mirrors:
+    - localai/localai-backends:master-gpu-nvidia-cuda-12-funasr
+- !!merge <<: *funasr
+  name: "cuda13-funasr"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-gpu-nvidia-cuda-13-funasr"
+  mirrors:
+    - localai/localai-backends:latest-gpu-nvidia-cuda-13-funasr
+- !!merge <<: *funasr
+  name: "cuda13-funasr-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-gpu-nvidia-cuda-13-funasr"
+  mirrors:
+    - localai/localai-backends:master-gpu-nvidia-cuda-13-funasr
+- !!merge <<: *funasr
+  name: "intel-funasr"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-gpu-intel-funasr"
+  mirrors:
+    - localai/localai-backends:latest-gpu-intel-funasr
+- !!merge <<: *funasr
+  name: "intel-funasr-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-gpu-intel-funasr"
+  mirrors:
+    - localai/localai-backends:master-gpu-intel-funasr
+- !!merge <<: *funasr
+  name: "rocm-funasr"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-gpu-rocm-hipblas-funasr"
+  mirrors:
+    - localai/localai-backends:latest-gpu-rocm-hipblas-funasr
+- !!merge <<: *funasr
+  name: "rocm-funasr-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-gpu-rocm-hipblas-funasr"
+  mirrors:
+    - localai/localai-backends:master-gpu-rocm-hipblas-funasr
+- !!merge <<: *funasr
+  name: "nvidia-l4t-funasr"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-nvidia-l4t-funasr"
+  mirrors:
+    - localai/localai-backends:latest-nvidia-l4t-funasr
+- !!merge <<: *funasr
+  name: "nvidia-l4t-funasr-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-nvidia-l4t-funasr"
+  mirrors:
+    - localai/localai-backends:master-nvidia-l4t-funasr
+- !!merge <<: *funasr
+  name: "cuda13-nvidia-l4t-arm64-funasr"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-nvidia-l4t-cuda-13-arm64-funasr"
+  mirrors:
+    - localai/localai-backends:latest-nvidia-l4t-cuda-13-arm64-funasr
+- !!merge <<: *funasr
+  name: "cuda13-nvidia-l4t-arm64-funasr-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-nvidia-l4t-cuda-13-arm64-funasr"
+  mirrors:
+    - localai/localai-backends:master-nvidia-l4t-cuda-13-arm64-funasr
+- !!merge <<: *funasr
+  name: "metal-funasr"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-metal-darwin-arm64-funasr"
+  mirrors:
+    - localai/localai-backends:latest-metal-darwin-arm64-funasr
+- !!merge <<: *funasr
+  name: "metal-funasr-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-metal-darwin-arm64-funasr"
+  mirrors:
+    - localai/localai-backends:master-metal-darwin-arm64-funasr
@@ -20,6 +20,7 @@ The Python backends use a unified build system based on `libbackend.sh` that pro
 ### Audio & Speech
 - **coqui** - Coqui TTS models
 - **faster-whisper** - Fast Whisper speech recognition
+- **funasr** - Local multilingual transcription with FunASR and SenseVoice
 - **kitten-tts** - Lightweight TTS
 - **mlx-audio** - Apple Silicon audio processing
 - **chatterbox** - TTS model
@@ -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)
@@ -563,6 +563,18 @@ function startBackend() {
        echo "Added ${EDIR}/lib to LD_LIBRARY_PATH for GPU libraries"
    fi

+    # XPU wheels carry a matching SYCL/Unified Runtime set. Prefer it over
+    # packaged oneAPI or host libraries, which may expose an older loader ABI.
+    if [ "$(uname -s)" = "Linux" ]; then
+        local sycl_runtime
+        for sycl_runtime in "${EDIR}/venv/lib"/libsycl.so*; do
+            if [ -f "${sycl_runtime}" ]; then
+                export LD_LIBRARY_PATH="${EDIR}/venv/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
+                break
+            fi
+        done
+    fi
+
    if [ ! -z "${BACKEND_FILE:-}" ]; then
        exec "${EDIR}/venv/bin/python" "${BACKEND_FILE}" "$@"
    elif [ -e "${MY_DIR}/server.py" ]; then
@@ -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``.

@@ -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()
@@ -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
@@ -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()
@@ -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]
@@ -1,3 +1,3 @@
-grpcio==1.82.1
+grpcio==1.83.1
 protobuf
 grpcio-tools
@@ -1,4 +1,4 @@
-grpcio==1.83.0
+grpcio==1.83.1
 protobuf
 certifi
 packaging==26.3
@@ -0,0 +1,24 @@
+import array
+import sys
+import wave
+
+
+def write_pcm_wav(destination, samples, sampling_rate):
+    """Write normalized floating-point audio samples as mono 16-bit PCM."""
+    pcm = array.array(
+        "h",
+        (
+            max(-32768, min(32767, round(float(sample) * 32768)))
+            for sample in samples
+        ),
+    )
+    if pcm.itemsize != 2:
+        raise RuntimeError("16-bit PCM requires two-byte signed integers")
+    if sys.byteorder != "little":
+        pcm.byteswap()
+
+    with wave.open(destination, "wb") as output:
+        output.setnchannels(1)
+        output.setsampwidth(2)
+        output.setframerate(sampling_rate)
+        output.writeframes(pcm.tobytes())
@@ -26,6 +26,7 @@ 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 model_utils import resolve_model_reference
+from audio_utils import write_pcm_wav


 # Import dynamic loader for pipeline discovery
@@ -35,6 +36,7 @@ from diffusers_dynamic_loader import (
    get_available_pipelines,
    load_diffusers_pipeline,
 )
+from load_options import single_file_load_kwargs

 # Import specific items still needed for special cases and safety checker
 from diffusers import DiffusionPipeline, ControlNetModel
@@ -122,6 +124,21 @@ 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:
@@ -464,6 +481,9 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):

        # Build kwargs for dynamic loading
        load_kwargs = {"torch_dtype": torchType}
+        load_kwargs.update(
+            single_file_load_kwargs(request.OriginalConfigFile, from_single_file)
+        )

        # Add variant if not loading from single file
        if not from_single_file and variant:
@@ -627,12 +647,13 @@ 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 +821,12 @@ 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 +889,7 @@ 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]
@@ -883,6 +904,49 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):

        return backend_pb2.Result(message="Media generated", success=True)

+    def SoundGeneration(self, request, context):
+        if not request.dst:
+            return backend_pb2.Result(success=False, message="request.dst is required")
+
+        prompt = request.text or request.caption
+        if not prompt:
+            return backend_pb2.Result(success=False, message="request.text is required")
+
+        try:
+            generation_options = dict(self.options)
+            if "num_inference_steps" in generation_options:
+                generation_options["num_inference_steps"] = int(
+                    generation_options["num_inference_steps"]
+                )
+            generation_options["prompt"] = prompt
+            if request.HasField("duration"):
+                generation_options["audio_length_in_s"] = request.duration
+            if request.HasField("temperature"):
+                generation_options["guidance_scale"] = request.temperature
+
+            generated = self.pipe(**generation_options)
+            if not hasattr(generated, "audios") or len(generated.audios) == 0:
+                return backend_pb2.Result(
+                    success=False,
+                    message="The diffusers pipeline returned no audio",
+                )
+
+            samples = generated.audios[0]
+            if hasattr(samples, "reshape"):
+                samples = samples.reshape(-1)
+            if hasattr(samples, "tolist"):
+                samples = samples.tolist()
+
+            sampling_rate = getattr(
+                getattr(getattr(self.pipe, "vae", None), "config", None),
+                "sampling_rate",
+                16000,
+            )
+            write_pcm_wav(request.dst, samples, sampling_rate)
+            return backend_pb2.Result(success=True, message="Sound generated successfully")
+        except Exception as err:
+            return backend_pb2.Result(success=False, message=f"SoundGeneration error: {err}")
+
    def UpscaleImage(self, request, context):
        try:
            if not request.src:
@@ -0,0 +1,7 @@
+# SPDX-License-Identifier: MIT
+
+
+def single_file_load_kwargs(original_config_file: str, from_single_file: bool) -> dict:
+    if from_single_file and original_config_file:
+        return {"original_config_file": original_config_file}
+    return {}
@@ -4,9 +4,13 @@ A test script to test the gRPC service and dynamic loader
 import unittest
 import subprocess
 import time
+import os
+import tempfile
+import wave
 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 +377,135 @@ 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")
+
+
+class TestWritePcmWav(unittest.TestCase):
+    def test_writes_clipped_float_samples_as_mono_pcm(self):
+        from audio_utils import write_pcm_wav
+
+        destination = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+        destination.close()
+
+        try:
+            write_pcm_wav(destination.name, [0.0, 0.5, -0.5, 2.0], 16000)
+
+            with wave.open(destination.name, "rb") as generated:
+                self.assertEqual(generated.getframerate(), 16000)
+                self.assertEqual(generated.getnchannels(), 1)
+                self.assertEqual(generated.getsampwidth(), 2)
+                self.assertEqual(generated.getnframes(), 4)
+                self.assertEqual(
+                    generated.readframes(4),
+                    b"\x00\x00\x00@\x00\xc0\xff\x7f",
+                )
+        finally:
+            os.unlink(destination.name)
+
+
+@unittest.skipUnless(GRPC_AVAILABLE, "gRPC modules not available")
+class TestSoundGeneration(unittest.TestCase):
+    def test_maps_request_options_and_writes_pipeline_audio(self):
+        from backend import BackendServicer
+
+        service = BackendServicer.__new__(BackendServicer)
+        service.options = {"num_inference_steps": 200.0}
+        service.pipe = MagicMock()
+        service.pipe.return_value.audios = [[0.0, 0.5, -0.5]]
+        service.pipe.vae.config.sampling_rate = 16000
+
+        destination = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+        destination.close()
+
+        try:
+            request = backend_pb2.SoundGenerationRequest(
+                text="ocean waves",
+                dst=destination.name,
+                duration=2.5,
+                temperature=0,
+            )
+
+            result = service.SoundGeneration(request, context=None)
+
+            self.assertTrue(result.success, result.message)
+            service.pipe.assert_called_once_with(
+                num_inference_steps=200,
+                prompt="ocean waves",
+                audio_length_in_s=2.5,
+                guidance_scale=0,
+            )
+            with wave.open(destination.name, "rb") as generated:
+                self.assertEqual(generated.getframerate(), 16000)
+                self.assertEqual(generated.getnframes(), 3)
+        finally:
+            os.unlink(destination.name)
@@ -0,0 +1,26 @@
+# SPDX-License-Identifier: MIT
+
+import unittest
+
+from backend.python.diffusers.load_options import single_file_load_kwargs
+
+
+class SingleFileLoadKwargsTest(unittest.TestCase):
+    def test_includes_original_config_for_single_file(self):
+        self.assertEqual(
+            single_file_load_kwargs("configs/v1-inference.yaml", True),
+            {"original_config_file": "configs/v1-inference.yaml"},
+        )
+
+    def test_omits_original_config_for_pretrained_model(self):
+        self.assertEqual(
+            single_file_load_kwargs("configs/v1-inference.yaml", False),
+            {},
+        )
+
+    def test_omits_empty_original_config(self):
+        self.assertEqual(single_file_load_kwargs("", True), {})
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -26,4 +26,17 @@ if [ "x${BUILD_PROFILE}" == "xl4t12" ]; then
    USE_PIP=true
 fi

+CTRANSLATE2_VERSION=${CTRANSLATE2_VERSION:-v4.7.1}
+CTRANSLATE2_ROCM_WHEEL_OS=${CTRANSLATE2_ROCM_WHEEL_OS:-Linux}
+CTRANSLATE2_ROCM_WHEEL_ARCHIVE="rocm-python-wheels-${CTRANSLATE2_ROCM_WHEEL_OS}.zip"
+
+if [ "x${BUILD_PROFILE}" == "xhipblas" ]; then
+  ensureVenv
+  mkdir /tmp/ctranslate2-rocm
+  wget -O "/tmp/ctranslate2-rocm/${CTRANSLATE2_ROCM_WHEEL_ARCHIVE}" "https://github.com/OpenNMT/CTranslate2/releases/download/${CTRANSLATE2_VERSION}/${CTRANSLATE2_ROCM_WHEEL_ARCHIVE}"
+  unzip "/tmp/ctranslate2-rocm/${CTRANSLATE2_ROCM_WHEEL_ARCHIVE}" -d /tmp/ctranslate2-rocm/
+  python3 -m ensurepip
+  python3 -m pip install --no-dependencies --no-index --find-links=/tmp/ctranslate2-rocm/temp-linux/ ctranslate2
+fi
+
 installRequirements
@@ -1,18 +1,26 @@
-.PHONY: fish-speech test-source-preparation
-fish-speech: test-source-preparation
+.PHONY: fish-speech test-source-preparation test-runtime-import
+fish-speech: test-source-preparation test-runtime-import
 	bash install.sh

 test-source-preparation:
 	bash prepare-source_test.sh

+test-runtime-import:
+	bash run_test.sh
+
 .PHONY: run
 run: fish-speech
 	@echo "Running fish-speech..."
 	bash run.sh
 	@echo "fish-speech run."

+.PHONY: test-unit
+test-unit:
+	python3 -m unittest -v prepare_upstream_test.py
+	bash run_test.sh
+
 .PHONY: test
-test: fish-speech
+test: fish-speech test-unit
 	@echo "Testing fish-speech..."
 	bash test.sh
 	@echo "fish-speech tested."
@@ -326,11 +326,29 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
            max_new_tokens = self.options.get("max_new_tokens", 1024)
            chunk_length = self.options.get("chunk_length", 200)

-            # Build references list for voice cloning
+            # Build references list for voice cloning. Saved LocalAI
+            # personalities use the same ordered JSON shape as audio.cpp.
            references = []
            voice_name = request.voice if request.voice else None

-            if voice_name and os.path.isfile(voice_name):
+            multi_reference_cond = request.params.get("multi_reference_cond", "") if hasattr(request, "params") else ""
+            if multi_reference_cond:
+                reference_entries = json.loads(multi_reference_cond)
+                if not isinstance(reference_entries, list) or not reference_entries:
+                    raise ValueError("multi_reference_cond must be a non-empty JSON array")
+                for entry in reference_entries:
+                    if not isinstance(entry, dict) or not entry.get("audio") or not entry.get("text"):
+                        raise ValueError("multi_reference_cond entries require audio and text")
+                    ref_audio_path = self._get_ref_audio_path(entry["audio"])
+                    with open(ref_audio_path, "rb") as f:
+                        audio_bytes = f.read()
+                    references.append(ServeReferenceAudio(audio=audio_bytes, text=entry["text"]))
+                print(
+                    f"[INFO] Using {len(references)} per-request reference audios",
+                    file=sys.stderr,
+                )
+
+            elif voice_name and os.path.isfile(voice_name):
                ref_audio_path = self._get_ref_audio_path(voice_name)
                with open(ref_audio_path, "rb") as f:
                    audio_bytes = f.read()
@@ -44,6 +44,13 @@ fi
 # editable install. pyaudio is only used by the upstream playback client.
 bash "${backend_dir}/prepare-source.sh" "${BUILD_TYPE:-}" "${FISH_SPEECH_DIR}/pyproject.toml"

+# CUDA 13 has no torch 2.8 wheels, so fish-speech's exact upstream pin would
+# make pip select the CPU-only aarch64 wheel from PyPI. Prepare the cloned tree
+# before resolving it, and use soundfile for reference audio because torchcodec
+# does not publish Linux aarch64 wheels.
+python3 "${backend_dir}/prepare_upstream.py" "${FISH_SPEECH_DIR}" \
+    --cuda-major "${CUDA_MAJOR_VERSION:-}"
+
 # Install fish-speech deps from source (without the package itself since we use PYTHONPATH)
 ensureVenv
 if [ "x${USE_PIP}" == "xtrue" ]; then
@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: MIT
+
+import argparse
+from pathlib import Path
+
+
+TORCH_28 = '"torch==2.8.0"'
+TORCH_29 = '"torch==2.9.1"'
+TORCHAUDIO_28 = '"torchaudio==2.8.0"'
+TORCHAUDIO_29 = '"torchaudio==2.9.1"'
+TORCHAUDIO_LOAD = (
+    "        waveform, original_sr = "
+    "torchaudio.load(reference_audio, backend=self.backend)"
+)
+SOUNDFILE_LOAD = "\n".join(
+    (
+        "        import soundfile as _sf",
+        "        import torch as _torch",
+        "",
+        "        data, original_sr = _sf.read(",
+        '            reference_audio, dtype="float32", always_2d=True',
+        "        )",
+        "        waveform = _torch.from_numpy(data.T.copy())",
+    )
+)
+
+
+def patch_cuda13_dependencies(pyproject: Path) -> None:
+    content = pyproject.read_text()
+    if (
+        TORCH_28 not in content
+        and TORCHAUDIO_28 not in content
+        and TORCH_29 in content
+        and TORCHAUDIO_29 in content
+    ):
+        return
+    if TORCH_28 not in content or TORCHAUDIO_28 not in content:
+        raise RuntimeError("fish-speech's torch 2.8 dependency pins have changed")
+
+    content = content.replace(TORCH_28, TORCH_29)
+    content = content.replace(TORCHAUDIO_28, TORCHAUDIO_29)
+    pyproject.write_text(content)
+
+
+def patch_reference_loader(loader: Path) -> None:
+    content = loader.read_text()
+    if TORCHAUDIO_LOAD not in content and content.count(SOUNDFILE_LOAD) == 1:
+        return
+    if content.count(TORCHAUDIO_LOAD) != 1:
+        raise RuntimeError("fish-speech's torchaudio.load call has changed")
+
+    loader.write_text(content.replace(TORCHAUDIO_LOAD, SOUNDFILE_LOAD))
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("source", type=Path)
+    parser.add_argument("--cuda-major")
+    args = parser.parse_args()
+
+    if args.cuda_major == "13":
+        patch_cuda13_dependencies(args.source / "pyproject.toml")
+    patch_reference_loader(
+        args.source / "fish_speech/inference_engine/reference_loader.py"
+    )
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,136 @@
+# SPDX-License-Identifier: MIT
+
+import importlib.util
+import sys
+import tempfile
+import types
+import unittest
+from pathlib import Path
+
+
+MODULE_PATH = Path(__file__).with_name("prepare_upstream.py")
+
+
+def load_prepare_upstream():
+    if not MODULE_PATH.exists():
+        raise AssertionError("prepare_upstream.py is missing")
+    spec = importlib.util.spec_from_file_location("prepare_upstream", MODULE_PATH)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+class FakeAudioData:
+    @property
+    def T(self):
+        return self
+
+    def copy(self):
+        return "channels-first"
+
+
+class PrepareUpstreamTests(unittest.TestCase):
+    def test_cuda13_dependencies_follow_available_pytorch_wheels(self):
+        prepare_upstream = load_prepare_upstream()
+
+        with tempfile.TemporaryDirectory() as tmp:
+            pyproject = Path(tmp) / "pyproject.toml"
+            pyproject.write_text(
+                'dependencies = [\n    "torch==2.8.0",\n    "torchaudio==2.8.0",\n]\n'
+                'stable = [\n    "torch==2.8.0",\n    "torchaudio",\n]\n'
+            )
+
+            prepare_upstream.patch_cuda13_dependencies(pyproject)
+
+            self.assertEqual(
+                pyproject.read_text(),
+                'dependencies = [\n    "torch==2.9.1",\n    "torchaudio==2.9.1",\n]\n'
+                'stable = [\n    "torch==2.9.1",\n    "torchaudio",\n]\n',
+            )
+
+    def test_reference_audio_uses_soundfile_without_torchcodec(self):
+        prepare_upstream = load_prepare_upstream()
+
+        with tempfile.TemporaryDirectory() as tmp:
+            loader = Path(tmp) / "reference_loader.py"
+            loader.write_text(
+                "class ReferenceLoader:\n"
+                "    def load_audio(self, reference_audio):\n"
+                "        waveform, original_sr = torchaudio.load(reference_audio, backend=self.backend)\n"
+                "        return waveform, original_sr\n"
+            )
+            prepare_upstream.patch_reference_loader(loader)
+
+            calls = []
+            fake_soundfile = types.SimpleNamespace(
+                read=lambda source, **kwargs: (
+                    calls.append((source, kwargs)) or FakeAudioData(),
+                    24000,
+                )
+            )
+            fake_torch = types.SimpleNamespace(
+                from_numpy=lambda data: ("tensor", data),
+            )
+            previous_soundfile = sys.modules.get("soundfile")
+            previous_torch = sys.modules.get("torch")
+            sys.modules["soundfile"] = fake_soundfile
+            sys.modules["torch"] = fake_torch
+            try:
+                namespace = {"torchaudio": None}
+                exec(compile(loader.read_text(), str(loader), "exec"), namespace)
+                instance = namespace["ReferenceLoader"]()
+                instance.backend = "soundfile"
+
+                waveform, sample_rate = instance.load_audio("voice.wav")
+            finally:
+                if previous_soundfile is None:
+                    del sys.modules["soundfile"]
+                else:
+                    sys.modules["soundfile"] = previous_soundfile
+                if previous_torch is None:
+                    del sys.modules["torch"]
+                else:
+                    sys.modules["torch"] = previous_torch
+
+            self.assertEqual(waveform, ("tensor", "channels-first"))
+            self.assertEqual(sample_rate, 24000)
+            self.assertEqual(
+                calls,
+                [("voice.wav", {"dtype": "float32", "always_2d": True})],
+            )
+
+    def test_reference_loader_drift_fails_the_build(self):
+        prepare_upstream = load_prepare_upstream()
+
+        with tempfile.TemporaryDirectory() as tmp:
+            loader = Path(tmp) / "reference_loader.py"
+            loader.write_text("def load_audio():\n    pass\n")
+
+            with self.assertRaisesRegex(RuntimeError, "torchaudio.load call"):
+                prepare_upstream.patch_reference_loader(loader)
+
+    def test_preparation_can_be_repeated(self):
+        prepare_upstream = load_prepare_upstream()
+
+        with tempfile.TemporaryDirectory() as tmp:
+            pyproject = Path(tmp) / "pyproject.toml"
+            pyproject.write_text(
+                'dependencies = ["torch==2.8.0", "torchaudio==2.8.0"]\n'
+            )
+            loader = Path(tmp) / "reference_loader.py"
+            loader.write_text(
+                "def load_audio(reference_audio):\n"
+                "        waveform, original_sr = torchaudio.load(reference_audio, backend=self.backend)\n"
+            )
+
+            prepare_upstream.patch_cuda13_dependencies(pyproject)
+            prepare_upstream.patch_reference_loader(loader)
+            try:
+                prepare_upstream.patch_cuda13_dependencies(pyproject)
+                prepare_upstream.patch_reference_loader(loader)
+            except RuntimeError as err:
+                self.fail(f"preparation is not idempotent: {err}")
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -1,3 +1,3 @@
 --extra-index-url https://download.pytorch.org/whl/cu130
-torch
-torchaudio
+torch==2.9.1+cu130
+torchaudio==2.9.1
@@ -1,3 +1,3 @@
 --extra-index-url https://download.pytorch.org/whl/cu130
-torch
-torchaudio
+torch==2.9.1+cu130
+torchaudio==2.9.1
@@ -1,9 +1,18 @@
 #!/bin/bash
-backend_dir=$(dirname $0)
+backend_dir=$(dirname "$(realpath "$0")")
 if [ -d $backend_dir/common ]; then
    source $backend_dir/common/libbackend.sh
 else
    source $backend_dir/../common/libbackend.sh
 fi

-startBackend $@
+# Editable installs record their build-time absolute source path, which becomes
+# stale when the backend is relocated under /backends at install time.
+export PYTHONPATH="${backend_dir}/fish-speech-src${PYTHONPATH:+:${PYTHONPATH}}"
+
+cuda_home=${CUDA_HOME:-/usr/local/cuda}
+if [ -z "${TRITON_PTXAS_PATH:-}" ] && [ -x "$cuda_home/bin/ptxas" ]; then
+    export TRITON_PTXAS_PATH="$cuda_home/bin/ptxas"
+fi
+
+startBackend "$@"
@@ -0,0 +1,23 @@
+#!/bin/bash
+set -euo pipefail
+
+SCRIPT_DIR=$(dirname "$(realpath "$0")")
+WORK_DIR=$(mktemp -d)
+trap 'rm -rf "$WORK_DIR"' EXIT
+
+mkdir -p "$WORK_DIR/runtime/common" "$WORK_DIR/runtime/fish-speech-src/fish_speech"
+cp "$SCRIPT_DIR/run.sh" "$WORK_DIR/runtime/run.sh"
+
+cat > "$WORK_DIR/runtime/fish-speech-src/fish_speech/inference_engine.py" <<'PY'
+IMPORT_OK = True
+PY
+
+cat > "$WORK_DIR/runtime/common/libbackend.sh" <<'SH'
+startBackend() {
+    python3 -c 'from fish_speech.inference_engine import IMPORT_OK; assert IMPORT_OK'
+}
+SH
+
+bash "$WORK_DIR/runtime/run.sh"
+
+echo "PASS: runtime launcher imports relocated fish-speech source"
@@ -0,0 +1,23 @@
+.PHONY: funasr
+funasr:
+	bash install.sh
+
+.PHONY: run
+run: funasr
+	@echo "Running funasr..."
+	bash run.sh
+	@echo "funasr run."
+
+.PHONY: test
+test: funasr
+	@echo "Testing funasr..."
+	bash test.sh
+	@echo "funasr tested."
+
+.PHONY: protogen-clean
+protogen-clean:
+	$(RM) backend_pb2_grpc.py backend_pb2.py
+
+.PHONY: clean
+clean: protogen-clean
+	rm -rf venv __pycache__
@@ -0,0 +1,152 @@
+#!/usr/bin/env python3
+"""
+gRPC backend for LocalAI wrapping FunASR (SenseVoice / Paraformer).
+"""
+from concurrent import futures
+import time
+import argparse
+import signal
+import sys
+import os
+import backend_pb2
+import backend_pb2_grpc
+import torch
+
+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
+
+
+_ONE_DAY_IN_SECONDS = 60 * 60 * 24
+MAX_WORKERS = int(os.environ.get('PYTHON_GRPC_MAX_WORKERS', '1'))
+
+
+class BackendServicer(backend_pb2_grpc.BackendServicer):
+    def Health(self, request, context):
+        return backend_pb2.Reply(message=bytes("OK", 'utf-8'))
+
+    def LoadModel(self, request, context):
+        from funasr import AutoModel
+
+        device = "cpu"
+        if request.CUDA and torch.cuda.is_available():
+            device = "cuda"
+        elif hasattr(torch, "xpu") and torch.xpu.is_available():
+            device = "xpu"
+        elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+            device = "mps"
+
+        model_id = request.Model or "iic/SenseVoiceSmall"
+        candidate_devices = [device]
+        if device in ("xpu", "mps"):
+            candidate_devices.append("cpu")
+
+        for candidate_device in candidate_devices:
+            try:
+                print(
+                    f"Loading FunASR model: {model_id} on {candidate_device}",
+                    file=sys.stderr,
+                )
+                self.model = AutoModel(
+                    model=model_id,
+                    vad_model="fsmn-vad",
+                    device=candidate_device,
+                    disable_update=True,
+                )
+                print("FunASR model loaded successfully", file=sys.stderr)
+                break
+            except Exception as err:
+                if candidate_device != candidate_devices[-1]:
+                    print(
+                        f"[WARN] FunASR {candidate_device} initialization failed: "
+                        f"{err}; retrying on cpu",
+                        file=sys.stderr,
+                    )
+                    continue
+                print(f"[ERROR] LoadModel failed: {err}", file=sys.stderr)
+                import traceback
+                traceback.print_exc(file=sys.stderr)
+                return backend_pb2.Result(success=False, message=str(err))
+
+        return backend_pb2.Result(message="Model loaded successfully", success=True)
+
+    def AudioTranscription(self, request, context):
+        from funasr.utils.postprocess_utils import rich_transcription_postprocess
+
+        result_segments = []
+        text = ""
+        try:
+            audio_path = request.dst
+            if not audio_path or not os.path.exists(audio_path):
+                print(f"Error: Audio file not found: {audio_path}", file=sys.stderr)
+                return backend_pb2.TranscriptResult(segments=[], text="")
+
+            language = None
+            if request.language and request.language.strip():
+                language = request.language.strip()
+
+            kwargs = {}
+            if language:
+                kwargs["language"] = language
+
+            results = self.model.generate(input=audio_path, **kwargs)
+
+            if not results:
+                return backend_pb2.TranscriptResult(segments=[], text="")
+
+            for idx, r in enumerate(results):
+                seg_text = r.get("text", "") if isinstance(r, dict) else str(r)
+                seg_text = rich_transcription_postprocess(seg_text)
+                text += seg_text
+                result_segments.append(backend_pb2.TranscriptSegment(
+                    id=idx,
+                    start=0,
+                    end=0,
+                    text=seg_text,
+                ))
+
+        except Exception as err:
+            print(f"Error in AudioTranscription: {err}", file=sys.stderr)
+            import traceback
+            traceback.print_exc(file=sys.stderr)
+            return backend_pb2.TranscriptResult(segments=[], text="")
+
+        return backend_pb2.TranscriptResult(segments=result_segments, text=text)
+
+
+def serve(address):
+    server = grpc.server(
+        futures.ThreadPoolExecutor(max_workers=MAX_WORKERS),
+        options=[
+            ('grpc.max_message_length', 50 * 1024 * 1024),
+            ('grpc.max_send_message_length', 50 * 1024 * 1024),
+            ('grpc.max_receive_message_length', 50 * 1024 * 1024),
+        ],
+        interceptors=get_auth_interceptors(),
+    )
+    backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
+    server.add_insecure_port(address)
+    server.start()
+    print("Server started. Listening on: " + address, file=sys.stderr)
+
+    def signal_handler(sig, frame):
+        print("Received termination signal. Shutting down...")
+        server.stop(0)
+        sys.exit(0)
+
+    signal.signal(signal.SIGINT, signal_handler)
+    signal.signal(signal.SIGTERM, signal_handler)
+
+    try:
+        while True:
+            time.sleep(_ONE_DAY_IN_SECONDS)
+    except KeyboardInterrupt:
+        server.stop(0)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run the gRPC server.")
+    parser.add_argument("--addr", default="localhost:50051", help="The address to bind the server to.")
+    args = parser.parse_args()
+    serve(args.addr)
@@ -0,0 +1,21 @@
+#!/bin/bash
+set -e
+
+EXTRA_PIP_INSTALL_FLAGS="--no-build-isolation"
+
+backend_dir=$(dirname $0)
+if [ -d $backend_dir/common ]; then
+    source $backend_dir/common/libbackend.sh
+else
+    source $backend_dir/../common/libbackend.sh
+fi
+
+if [ "x${BUILD_PROFILE}" == "xintel" ]; then
+    EXTRA_PIP_INSTALL_FLAGS+=" --upgrade --index-strategy=unsafe-first-match"
+fi
+
+PYTHON_VERSION="3.12"
+PYTHON_PATCH="12"
+PY_STANDALONE_TAG="20251120"
+
+installRequirements
@@ -0,0 +1,4 @@
+--extra-index-url https://download.pytorch.org/whl/cpu
+torch
+torchaudio
+funasr
@@ -0,0 +1,4 @@
+--extra-index-url https://download.pytorch.org/whl/cu121
+torch
+torchaudio
+funasr
@@ -0,0 +1,4 @@
+--extra-index-url https://download.pytorch.org/whl/cu130
+torch
+torchaudio
+funasr
@@ -0,0 +1,4 @@
+--extra-index-url https://download.pytorch.org/whl/rocm7.0
+torch
+torchaudio
+funasr
@@ -0,0 +1,4 @@
+--extra-index-url https://download.pytorch.org/whl/xpu
+torch
+torchaudio
+funasr
@@ -0,0 +1,4 @@
+--extra-index-url https://pypi.jetson-ai-lab.io/jp6/cu129/
+torch
+torchaudio
+funasr
@@ -0,0 +1,4 @@
+--extra-index-url https://download.pytorch.org/whl/cu130
+torch
+torchaudio
+funasr