* fix(vllm): don't stream raw tool-call markup as content when a tool parser is active
When a tool_parser is configured and the request carries tools, the streaming
loop emitted every text delta as delta.content — including the model's raw
tool-call markup (e.g. <tool_call>...) — because extract_tool_calls only runs
on the full output after the stream. Clients streaming a tool call therefore
saw the unparsed tool-call syntax as assistant content.
Buffer the text while a tool parser is active for the request; the existing
end-of-stream chat_delta already carries the parsed tool_calls (or the cleaned
content), which the Go side converts to SSE deltas. Non-tool-parser streaming
is unchanged.
Add a server-less regression test covering both the tool-call case (no raw
markup leaked as content) and the plain-text case (content delivered exactly
once — guards against double-emitting the buffered content).
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
* test(vllm): add expectedFailure test for progressive streaming with tool parser (Case 3, #582)
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
* test(vllm): add Cases 4+5 — marker split across chunks + false-positive prefix (TDD, Option B state machine, #582)
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
* feat(vllm): progressive streaming via parser.extract_tool_calls_streaming
When a tool parser is active for a tool-enabled streaming request,
#10346 buffers the entire generation and surfaces it on the final
chunk to prevent raw tool-call markup from leaking as delta.content.
This is correct but turns the request into effectively non-streaming
for plain-text responses — the client sees nothing until the model
stops.
Every concrete tool parser shipped with vLLM 0.23+ already implements
extract_tool_calls_streaming (Granite4, Qwen3Coder, DeepSeekV31, Jamba,
Ernie45, Hermes2Pro, llama3_json, mistral, …). Use it: instantiate
the parser before the streaming loop and call its streaming method per
delta, emitting DeltaMessage(content=…) or DeltaMessage(tool_calls=[…])
when the parser is ready.
Falls back to the existing #10346 buffer path when:
- the parser does not have extract_tool_calls_streaming, OR
- extract_tool_calls_streaming raises mid-stream (logged, the
rest of the request finishes via post-loop extract_tool_calls).
Tests (TestStreamingToolParser):
1. Buffer path: no markup leaked, no content duplication
2. Native streaming: plain-text response streams progressively
3. Native streaming: tool_call structured, no markup leaked
4. Native streaming exception → graceful fallback, no markup, no crash
5. No tool parser → unchanged per-delta content stream
E2E verified against qwen3_coder on vLLM 0.23.0 (NVIDIA GB10 / arm64 / CUDA 13).
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
* docs(vllm): add server-side TTFT benchmark for the streaming tool-parser path
Self-contained stdlib-only script that measures time-to-first-token (TTFT)
for the vLLM backend's two streaming scenarios:
- tool_call: request mentions a tool; model is expected to call it
- plain_text: request offers a tool but explicitly asks for prose
Use this to compare:
- the buffer-all path (#10346) → plain_text TTFT ≈ total response time
- the native-streaming path (this PR) → plain_text TTFT ≈ true first-token time
python examples/vllm-bench/ttft_streaming_tool_parser.py \\
--url http://localhost:8080 --model my-coder --runs 3
Lives under examples/ so it does not interfere with the test suite.
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
* examples/vllm-bench: add long-text scenario (8 paragraphs, 1500 tokens)
The long-text scenario shows the buffering vs streaming difference most
dramatically: with the buffer-all path, the client receives nothing for
20+ seconds and then the entire 1500-token response at once. With native
streaming, the first token arrives in tens of milliseconds and the
response flows progressively.
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
---------
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
Co-authored-by: Philipp Wacker <philipp.wacker@ibf-solutions.com>
* feat(nemo): enable word-level timestamps for ASR models
The nemo backend ignored timestamp_granularities and always returned a
single segment with start=0 end=0, making word-level timestamps
impossible to obtain even though the NeMo models (parakeet-tdt, etc.)
fully support them.
Changes:
- Add _get_stride_seconds() to compute frame duration from the model's
preprocessor window_stride and encoder subsampling_factor.
- Add _build_segments_with_words() that extracts word offsets from the
NeMo Hypothesis.timestamp dict and converts frame indices to
nanosecond timestamps.
- Support 'word' granularity (one segment per word) and 'segment'
granularity (merge at time-gap boundaries using a dynamic threshold).
- Populate TranscriptSegment.words with TranscriptWord entries so
callers get both segment-level and word-level timing.
- Only request timestamps from NeMo when the caller actually asks for
them (timestamp_granularities is non-empty), keeping the fast path
unchanged for callers that don't need timestamps.
Tested with nvidia/parakeet-tdt-0.6b-v3 on the JFK "ask not" clip:
curl -X POST /v1/audio/transcriptions \
-F file=@jfk.wav -F model=nemo-parakeet-tdt-0.6b \
-F 'timestamp_granularities[]=word' -F response_format=verbose_json
→ each word has correct start/end times in seconds.
Signed-off-by: fqscfqj <fqscfqj@outlook.com>
* fix(nemo): address Copilot review feedback
- Narrow exception handling in _get_stride_seconds to catch only
AttributeError, KeyError, TypeError instead of bare Exception, and
emit a warning when falling back to the hardcoded stride.
- Remove explicit return_hypotheses=False when timestamps are requested;
timestamps=True already forces NeMo to return Hypothesis objects.
- Add a warning when NeMo does not return Hypothesis objects despite
timestamps being requested.
Signed-off-by: fqscfqj <fqscfqj@outlook.com>
---------
Signed-off-by: fqscfqj <fqscfqj@outlook.com>
fix(vllm): structured outputs silently ignored on vLLM >= 0.23
vLLM >= 0.23 removed GuidedDecodingParams (now StructuredOutputsParams) and
renamed the SamplingParams field guided_decoding -> structured_outputs. The
import failed, HAS_GUIDED_DECODING became False, and the whole guided-decoding
block was skipped, so response_format / grammar constraints were silently
ignored. Adapt the existing request.Grammar path to the new class/field.
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
Squashed feat/pii-ner-tier-engine rebased onto master (was 45 commits; see
backup/pii-ner-tier-engine-prerebase). Net change:
- privacy-filter.cpp: standalone GGML engine for the openai-privacy-filter
PII/NER token classifier, wired as a LocalAI gRPC backend (CPU/CUDA/Vulkan).
TokenClassify moves off the patched llama.cpp path onto this backend.
- PII filter reworked to be NER-centric (encoder/NER detection tier scanning
whole conversations as one document), with a recreated bounded restricted-
regex secret-matching pattern detector tier alongside it (per-model
pii_detection.builtins / .patterns + core/services/routing/piipattern).
- Detection labelled by source (ner vs pattern); backend trace / confidence /
debug observability; analyze/redact exposed as a synchronous API.
- Instance-wide default detector policy + per-usecase default-on; request
filtering extended to completions, embeddings, edits & Ollama.
- React UI: NER-centric PII editor, detector-models table, pattern/builtins
editor, middleware default-policy UI.
- Gallery: privacy-filter-multilingual token-classify model + NER install
filter; token_classify known_usecase; batch sized to context for NER models.
privacy-filter backend registered in the backend gallery (cpu/vulkan/cuda-13
meta + image entries with a capabilities map) matching its CI matrix jobs,
and an /import-model auto-detect importer (PrivacyFilterImporter, narrow
privacy-filter GGUF detection) replacing the prior pref-only registration.
Reconciled against master's independent evolution:
- Dropped master's PIIPatternOverrides feature (global-pattern runtime
overrides + /api/pii/patterns API + runtime_settings.json persistence). The
per-model NER + pattern-detector design supersedes it; it was built on the
global redactor pattern set this branch replaced.
- Reverted the llama.cpp Score carry-patch (0006-server-task-type-score):
removed the patch and restored master's grpc-server.cpp Score RPC (direct
llama_decode, slot-loop bypass) and LLAMA_VERSION pin, plus master's
model_config validation forbidding score + chat/completion/embeddings on
llama-cpp. token_classify is unaffected (it runs on the privacy-filter
backend, not llama-cpp).
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
fix(neutts): pin torchaudio to match torch to avoid ABI mismatch (#9798)
neucodec pulls torchaudio transitively but it was unpinned, so an
incompatible torchaudio could be resolved against the pinned torch==2.8.0,
producing the 'undefined symbol: torch_library_impl' load failure. Pin
torchaudio==2.8.0 alongside torch in the cpu and cublas12 requirements.
Assisted-by: claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
Vision-language checkpoints such as mlx-community/gemma-4-E4B-it-qat-4bit
declare the "image-text-to-text" pipeline tag on HuggingFace. The mlx
importer hardcoded backend "mlx" for every mlx-community model, so these
VLMs were served by the text-only mlx-lm backend whose tokenizer does not
carry the processor chat template. The template was never applied and the
model produced degenerate, looping output that echoed the prompt.
Detect the "image-text-to-text" pipeline tag in the importer and route those
models to mlx-vlm, which applies the processor-aware chat template. An
explicit backend preference still wins.
As a defensive backstop, the mlx backend now warns loudly when the loaded
model has no chat template, so a misrouted VLM surfaces the problem instead
of silently looping.
Fixes#10269
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
Neither the sherpa-onnx nor the speaker-recognition backend had a
darwin/arm64 image, so `local-ai backends install` failed with "no child
with platform darwin/arm64" on macOS. This left /v1/audio/diarization (the
sherpa-onnx path) and /v1/voice/embed without any usable backend on Apple
Silicon.
Both backends build on darwin/arm64:
- sherpa-onnx (Go) already fetches the onnxruntime osx-arm64 runtime in its
Makefile; it only needed a darwin matrix entry (build-type metal, lang go,
like whisper and silero-vad).
- speaker-recognition (Python) needed a requirements-mps.txt so the mps build
installs plain onnxruntime (which ships a macOS arm64 wheel) instead of the
onnxruntime-gpu pulled by its base requirements (which does not).
Add both to the includeDarwin build matrix, wire the metal capability and
metal image aliases into the gallery, and add the speaker-recognition
requirements-mps.txt.
Fixes#10268
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
fix(vllm): restore compatibility with vLLM >= 0.22 (get_tokenizer moved)
vLLM 0.22 moved get_tokenizer from vllm.transformers_utils.tokenizer
to vllm.tokenizers. Since the backend requirements install vllm
unpinned, freshly built/installed vllm backends currently fail to
start with ModuleNotFoundError: No module named
'vllm.transformers_utils.tokenizer' (surfacing as 'grpc service not
ready' when loading a model).
Use the same try/except version-compat import pattern already used
elsewhere in this file: try the new vllm.tokenizers location first and
fall back to the pre-0.22 path.
Tested on a DGX Spark (GB10, ARM64) with the
cuda13-nvidia-l4t-arm64-vllm backend and vllm 0.22.0: model load, chat
completions and tool calls all work with this patch applied.
Signed-off-by: pos-ei-don <1822533+pos-ei-don@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The OpenAI-compatible TTS endpoint accepts an `instructions` field, but it
was silently dropped at the HTTP->gRPC boundary: neither schema.TTSRequest
nor the gRPC TTSRequest proto carried it, so backends could only read such a
value from static YAML options (identical for every request). This blocked
per-line emotion/style and, for Qwen3-TTS VoiceDesign, limited a model config
to a single designed voice.
Plumb a generic per-request instruction string end to end, plus an optional
backend-specific params map:
- proto: add `optional string instructions` and `map<string,string> params`
to TTSRequest.
- schema: add Instructions (maps OpenAI `instructions`) and Params (LocalAI
extension) to schema.TTSRequest.
- core: thread both through ModelTTS/ModelTTSStream via a newTTSRequest helper
that attaches instructions only when non-empty (so backends can fall back to
YAML when unset); forward them from the /v1/audio/speech handler.
- qwen-tts: prefer the per-request instruction over the YAML `instruct` option
(used by both mode detection and generation) and merge per-request params.
- chatterbox: merge per-request params (coerced to float/int/bool) over YAML
options into generate() kwargs.
Fully backward compatible: empty instructions fall back to the YAML option and
backends that don't support style/voice instructions ignore the field.
Closes#10164
Assisted-by: Claude:claude-opus-4-8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
Pin texterrors==1.1.6 before nemo_toolkit[asr] in requirements-cublas13.txt.
The texterrors package (a NeMo transitive dependency) contains a compiled
C++ extension (texterrors_align.so) that may be built from source during
OCI image creation. When built on systems with GCC 14+ (e.g. Ubuntu 24.04),
the resulting binary requires GLIBCXX_3.4.32, which is not available in
the default LocalAI container (Ubuntu 22.04, GLIBCXX up to 3.4.30).
Pinning to 1.1.6 (the latest release) ensures:
- Reproducible builds across environments
- pip resolves the pre-built manylinux2014 wheel (needs only GLIBCXX_3.4.11)
instead of potentially building from source with a newer toolchain
Fixes#10056
Signed-off-by: 番茄摔成番茄酱 <fqscfqj@outlook.com>
* fix(nemo): extract Hypothesis.text for TDT/RNNT ASR models
CTC models (e.g. Whisper) return List[str] from transcribe(), but
TDT/RNNT models (e.g. parakeet-tdt-0.6b-v3) return List[Hypothesis]
where the decoded text lives in the Hypothesis.text attribute.
Previously, results[0] was assigned directly to the protobuf string
field, causing silent empty output for non-CTC models.
Now checks the return type and extracts .text from Hypothesis objects,
with a safe fallback via getattr().
* refactor: simplify Hypothesis text extraction per Copilot review
Use single getattr() call instead of hasattr() + double access,
and return empty string for unknown types instead of str(result)
to avoid leaking internal repr to clients.
* fix(qwen-asr): enable timestamp output when forced_aligner is configured
Two bugs prevented timestamps from working in the qwen-asr backend:
1. transcribe() was called without return_time_stamps=True, so the
forced aligner was loaded but never invoked. Now we pass
return_time_stamps=True when a forced_aligner is present.
2. The timestamp parsing code expected (list, tuple) items, but the
qwen_asr library returns ForcedAlignItem dataclass instances with
.text, .start_time, .end_time attributes. Added hasattr() check
to handle this correctly, falling back to tuple parsing for
backward compatibility.
* refactor: address Copilot review for qwen-asr timestamps
- Wrap return_time_stamps kwarg in try/except TypeError for safety
- Add defensive float() normalization for timestamp times
- Use str() for text extraction to ensure string type
* fix(qwen-asr): convert seconds to nanoseconds for Go time.Duration
The Go server reads TranscriptSegment.start/end via time.Duration,
which is in nanoseconds. Previously the backend sent milliseconds
(* 1000), causing timestamps to be 1000x too small (e.g. 8e-8
instead of 0.08). Convert seconds → nanoseconds (* 1e9) instead.
Also applies to the legacy tuple path for consistency.
* feat(qwen-asr): respect timestamp_granularities (segment vs word)
Read request.timestamp_granularities from the gRPC request.
- 'word': return one segment per aligned item (character / word)
- 'segment' (default): merge consecutive items at sentence boundaries
Sentence boundaries detected via CJK punctuation (。！？；…)
and Latin endings (. ! ? ;). This matches the OpenAI Whisper API
contract where omitting the parameter defaults to segment-level.
* fix(qwen-asr): escape smart quotes in punctuation set
Unicode curly quotes (U+2018/2019) were being interpreted as Python
string delimiters, causing SyntaxError. Use explicit unicode escapes.
* fix(qwen-asr): use time-gap threshold for segment boundaries
The forced aligner strips punctuation from its output, so text-based
sentence detection doesn't work. Instead, detect segment boundaries
by measuring time gaps between consecutive aligned items.
Threshold = max(median_gap * 4, 0.3s). This cleanly separates
intra-sentence gaps (< 0.24s) from inter-sentence gaps (> 0.3s)
across Chinese, English, and other languages.
* fix(qwen-asr): smart join with spaces for non-CJK tokens
The forced aligner strips whitespace from tokenized text, so English
words like ['hello', 'world'] were joined as 'helloworld'. Add
_smart_join() that inserts spaces between non-CJK tokens while
keeping CJK characters and punctuation unspaced. Works for Chinese,
English, Korean, Japanese, and mixed-language text.
---------
Co-authored-by: fqscfqj <fqsfqj@outlook.com>
Add a routing middleware stack and a cloud-proxy backend.
* cloud-proxy: a Go gRPC backend that forwards OpenAI- and
Anthropic-shaped chat requests to upstream providers, with an
optional translate mode (OpenAI request -> Anthropic /v1/messages
-> OpenAI response) and full tool-calling support.
* routing: admission control, content-aware model routing
(embedding cache + classifier + rerank + Arch-Router score),
PII detection/redaction (regex + NER) with streaming filter and
OpenAI/Anthropic adapters, and a per-user/per-key billing recorder
backed by GORM or in-memory storage.
* middleware: UsageMiddleware records usage via the billing recorder,
plus admission, route-model, usage-stamp and trace middlewares.
* observability: BackendTrace ring buffer stores full request bodies
(capped), MITM proxy emits structured trace events, and router
classifier decisions surface at /api/router/decide.
* gallery: Arch-Router-1.5B (Q4_K_M and Q8_0).
* UI: cloud-proxy model-editor fields, classifier system-prompt and
score-normalization config, and a Traces page rendering request
bodies.
Assisted-by: claude-code:claude-opus-4-7 [Read] [Edit] [Bash]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* fix(vllm): switch L4T13 backend to PyPI aarch64+cu130 wheels
The L4T13 vllm backend pulled torch / torchvision / torchaudio / vllm from
pypi.jetson-ai-lab.io's sbsa/cu130 mirror via [tool.uv.sources] with no
version pins. That mirror started shipping torch 2.11.0 next to a
vllm-0.20.0+cu130 wheel that was still compiled against torch 2.10's c10
ABI, so uv landed on the mismatched pair and vllm crashed at import:
ImportError: vllm/_C.abi3.so: undefined symbol:
_ZN3c1013MessageLoggerC1EPKciib
(c10::MessageLogger's constructor signature changed between torch 2.10 and
2.11; the vllm wheel referenced the 2.10 form, the installed libc10.so
exported only the 2.11 form.)
Since torch 2.11 (April 2026) PyPI publishes its own aarch64 + cu130
manylinux wheels, and vllm 0.20.0 ships an aarch64 wheel whose Requires-
Dist locks torch==2.11.0 / torchvision==0.26.0 / torchaudio==2.11.0. That
makes uv's resolver produce an ABI-consistent set automatically, so the
mirror and the [tool.uv.sources] pinning are no longer needed.
flash-attn is dropped from the dep list: PyPI has no aarch64 wheel, but
vLLM 0.20+ already bundles its own vllm_flash_attn (fa2 + fa3) inside the
main wheel, so the Dao-AILab package isn't required at runtime.
Reference: https://pytorch.org/blog/vllm-and-pytorch-work-together-to-improve-the-developer-experience-on-aarch64/
Assisted-by: Claude:claude-opus-4-7 [Read] [Edit] [Write] [Bash] [WebFetch]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* refactor(vllm): retire l4t13 pyproject.toml in favor of requirements-*.txt
pyproject.toml only existed because uv pip install -r requirements.txt
doesn't honor [tool.uv.sources]. The previous commit dropped [tool.uv.
sources] (PyPI now serves the aarch64 + cu130 wheels directly), so the
file no longer carries any logic the requirements-*.txt path can't.
Replace with the same two-file pattern every other build profile uses:
- requirements-l4t13.txt (accelerate / torch / transformers /
bitsandbytes - matches cublas13's split)
- requirements-l4t13-after.txt (vllm; runs after the base resolve so
the cu130 torch wheel lands first)
install.sh's whole l4t13 elif branch goes away; libbackend.sh's
installRequirements already handles the requirements-install.txt build-
deps pass, the C_INCLUDE_PATH export for PORTABLE_PYTHON, and the
runProtogen call, so falling through to the standard else: branch
produces identical install behavior with less surface area.
No functional change at install time - same wheels, same order.
Assisted-by: Claude:claude-opus-4-7 [Read] [Edit] [Write] [Bash]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix(sglang,vllm-omni): switch L4T13 backends to PyPI aarch64+cu130 wheels
Same root cause and same fix as the vllm backend in the previous commits:
the L4T13 sglang and vllm-omni backends both pulled their accelerator
stack from pypi.jetson-ai-lab.io's sbsa/cu130 mirror with no version
pins, so they would silently land on the same torch 2.11 vs cu130-built
wheel ABI mismatch the moment the mirror published an out-of-sync pair.
sglang
------
- Drop pyproject.toml + [tool.uv.sources]. The historical comment said
the [all] extra was unsafe on aarch64 because of decord, but sglang
0.5.x now uses `decord2` on aarch64/arm/armv7l (which ships cp312
aarch64 wheels), so we can match cublas13's sglang[all]>=0.5.11 pin
and stop being capped at the 0.5.1.post2 the L4T mirror shipped.
That unblocks Gemma 4 / MTP recipes on Jetson Thor.
- New requirements-l4t13.txt mirrors the cublas13 split (accelerate /
torch / torchvision / torchaudio / transformers), requirements-l4t13-
after.txt carries sglang[all]>=0.5.11.
- install.sh's l4t13 elif branch goes away; falls through to the
standard installRequirements path.
vllm-omni
---------
- requirements-l4t13.txt drops --extra-index-url to jetson-ai-lab and
drops flash-attn (PyPI has no aarch64 wheel, vLLM 0.20+ bundles its
own vllm_flash_attn fa2 + fa3 internally).
- install.sh's l4t13 vllm-install branch collapses into the cublas13
branch since both now just run `pip install vllm --torch-backend=auto`
against PyPI.
- --index-strategy=unsafe-best-match is dropped from the top-level
l4t13 guard; without the L4T mirror in the picture it had no purpose.
The from-source vllm-omni install on top still keeps its existing
`sed -i '/^fa3-fwd[[:space:]]*==/d' requirements/cuda.txt` workaround -
fa3-fwd has no aarch64 wheel and no sdist, unrelated to flash-attn.
Reference: https://pytorch.org/blog/vllm-and-pytorch-work-together-to-improve-the-developer-experience-on-aarch64/
Assisted-by: Claude:claude-opus-4-7 [Read] [Edit] [Write] [Bash] [WebFetch]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix(sglang): drop [all] extra on l4t13 - xatlas has no aarch64 wheel
CI revealed that sglang[all]==0.5.12 transitively pulls xatlas via the
[diffusion] sub-extra, and xatlas ships no aarch64 wheel. Its sdist
depends on scikit_build_core without declaring it in build-system.
requires, so under --no-build-isolation uv can't build it from source:
× Failed to build `xatlas==0.0.11`
├─▶ The build backend returned an error
╰─▶ Call to `scikit_build_core.build.build_wheel` failed (exit status: 1)
ModuleNotFoundError: No module named 'scikit_build_core'
help: `xatlas` (v0.0.11) was included because `sglang[all]` (v0.5.12)
depends on `xatlas`
Upstream sglang explicitly gates st_attn and vsa on
`platform_machine != aarch64` inside the same [diffusion] extra but
forgot xatlas - same class of bug that bit the old decord pin.
Use plain `sglang>=0.5.11` on l4t13. backend.py imports only base
sglang.srt symbols (Engine, ServerArgs, FunctionCallParser,
ReasoningParser); the [all] extras are optional accelerators not
required at import time. cublas13 (x86_64) keeps [all] because xatlas
has x86_64 wheels there.
Assisted-by: Claude:claude-opus-4-7 [Read] [Edit] [Write] [Bash]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* feat(liquid-audio): add LFM2.5-Audio any-to-any backend + realtime_audio usecase
Wires LiquidAI's LFM2.5-Audio-1.5B as a self-contained Realtime API model:
single engine handles VAD, transcription, LLM, and TTS in one bidirectional
stream — drop-in alternative to a VAD+STT+LLM+TTS pipeline.
Backend
- backend/python/liquid-audio/ — new Python gRPC backend wrapping the
`liquid-audio` package. Modes: chat / asr / tts / s2s, voice presets,
Load/Predict/PredictStream/AudioTranscription/TTS/VAD/AudioToAudioStream/
Free and StartFineTune/FineTuneProgress/StopFineTune. Runtime monkey-patch
on `liquid_audio.utils.snapshot_download` so absolute local paths from
LocalAI's gallery resolve without a HF round-trip. soundfile in place of
torchaudio.load/save (torchcodec drags NVIDIA NPP we don't bundle).
- backend/backend.proto + pkg/grpc/{backend,client,server,base,embed,
interface}.go — new AudioToAudioStream RPC mirroring AudioTransformStream
(config/frame/control oneof in; typed event+pcm+meta out).
- core/services/nodes/{health_mock,inflight}_test.go — add stubs for the
new RPC to the test fakes.
Config + capabilities
- core/config/backend_capabilities.go — UsecaseRealtimeAudio, MethodAudio
ToAudioStream, UsecaseInfoMap entry, liquid-audio BackendCapability row.
- core/config/model_config.go — FLAG_REALTIME_AUDIO bitmask, ModalityGroups
membership in both speech-input and audio-output groups so a lone flag
still reads as multimodal, GetAllModelConfigUsecases entry, GuessUsecases
branch.
Realtime endpoint
- core/http/endpoints/openai/realtime.go — extract prepareRealtimeConfig()
so the gate is unit-testable; accept realtime_audio models and self-fill
empty pipeline slots with the model's own name (user-pinned slots win).
- core/http/endpoints/openai/realtime_gate_test.go — six specs covering nil
cfg, empty pipeline, legacy pipeline, self-contained realtime_audio,
user-pinned VAD slot, and partial legacy pipeline.
UI + endpoints
- core/http/routes/ui.go — /api/pipeline-models accepts either a legacy
VAD+STT+LLM+TTS pipeline or a realtime_audio model; surfaces a
self_contained flag so the Talk page can collapse the four cards.
- core/http/routes/ui_api.go — realtime_audio in usecaseFilters.
- core/http/routes/ui_pipeline_models_test.go — covers both code paths.
- core/http/react-ui/src/pages/Talk.jsx — self-contained badge instead of
the four-slot grid; rename Edit Pipeline → Edit Model Config; less
pipeline-specific wording.
- core/http/react-ui/src/pages/Models.jsx + locales/en/models.json — new
realtime_audio filter button + i18n.
- core/http/react-ui/src/utils/capabilities.js — CAP_REALTIME_AUDIO.
- core/http/react-ui/src/pages/FineTune.jsx — voice + validation-dataset
fields, surfaced when backend === liquid-audio, plumbed via
extra_options on submit/export/import.
Gallery + importer
- gallery/liquid-audio.yaml — config template with known_usecases:
[realtime_audio, chat, tts, transcript, vad].
- gallery/index.yaml — four model entries (realtime/chat/asr/tts) keyed by
mode option. Fixed pre-existing `transcribe` typo on the asr entry
(loader silently dropped the unknown string → entry never surfaced as a
transcript model).
- gallery/lfm.yaml — function block for the LFM2 Pythonic tool-call format
`<|tool_call_start|>[name(k="v")]<|tool_call_end|>` matching
common_chat_params_init_lfm2 in vendored llama.cpp.
- core/gallery/importers/{liquid-audio,liquid-audio_test}.go — detector
matches LFM2-Audio HF repos (excludes -gguf mirrors); mode/voice
preferences plumbed through to options.
- core/gallery/importers/importers.go — register LiquidAudioImporter
before LlamaCPPImporter.
- pkg/functions/parse_lfm2_test.go — seven specs for the response/argument
regex pair on the LFM2 pythonic format.
Build matrix
- .github/backend-matrix.yml — seven liquid-audio targets (cuda12, cuda13,
l4t-cuda-13, hipblas, intel, cpu amd64, cpu arm64). Jetpack r36 cuda-12
is skipped (Ubuntu 22.04 / Python 3.10 incompatible with liquid-audio's
3.12 floor).
- backend/index.yaml — anchor + 13 image entries.
- Makefile — .NOTPARALLEL, prepare-test-extra, test-extra,
docker-build-liquid-audio.
Docs
- .agents/plans/liquid-audio-integration.md — phased plan; PR-D (real
any-to-any wiring via AudioToAudioStream), PR-E (mid-audio tool-call
detector), PR-G (GGUF entries once upstream llama.cpp PR #18641 lands)
remain.
- .agents/api-endpoints-and-auth.md — expand the capability-surface
checklist with every place a new FLAG_* needs to be registered.
Assisted-by: claude-code:claude-opus-4-7-1m [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* feat(realtime): function calling + history cap for any-to-any models
Three pieces, all on the realtime_audio path that just landed:
1. liquid-audio backend (backend/python/liquid-audio/backend.py):
- _build_chat_state grows a `tools_prelude` arg.
- new _render_tools_prelude parses request.Tools (the OpenAI Chat
Completions function array realtime.go already serialises) and
emits an LFM2 `<|tool_list_start|>…<|tool_list_end|>` system turn
ahead of the user history. Mirrors gallery/lfm.yaml's `function:`
template so the model sees the same prompt shape whether served
via llama-cpp or here. Without this the backend silently dropped
tools — function calling was wired end-to-end on the Go side but
the model never saw a tool list.
2. Realtime history cap (core/http/endpoints/openai/realtime.go):
- Session grows MaxHistoryItems int; default picked by new
defaultMaxHistoryItems(cfg) — 6 for realtime_audio models (LFM2.5
1.5B degrades quickly past a handful of turns), 0/unlimited for
legacy pipelines composing larger LLMs.
- triggerResponse runs conv.Items through trimRealtimeItems before
building conversationHistory. Helper walks the cut left if it
would orphan a function_call_output, so tool result + call pairs
stay intact.
- realtime_gate_test.go: specs for defaultMaxHistoryItems and
trimRealtimeItems (zero cap, under cap, over cap, tool-call pair
preservation).
3. Talk page (core/http/react-ui/src/pages/Talk.jsx):
- Reuses the chat page's MCP plumbing — useMCPClient hook,
ClientMCPDropdown component, same auto-connect/disconnect effect
pattern. No bespoke tool registry, no new REST endpoints; tools
come from whichever MCP servers the user toggles on, exactly as
on the chat page.
- sendSessionUpdate now passes session.tools=getToolsForLLM(); the
update re-fires when the active server set changes mid-session.
- New response.function_call_arguments.done handler executes via
the hook's executeTool (which round-trips through the MCP client
SDK), then replies with conversation.item.create
{type:function_call_output} + response.create so the model
completes its turn with the tool output. Mirrors chat's
client-side agentic loop, translated to the realtime wire shape.
UI changes require a LocalAI image rebuild (Dockerfile:308-313 bakes
react-ui/dist into the runtime image). Backend.py changes can be
swapped live in /backends/<id>/backend.py + /backend/shutdown.
Assisted-by: claude-code:claude-opus-4-7-1m [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* feat(realtime): LocalAI Assistant ("Manage Mode") for the Talk page
Mirrors the chat-page metadata.localai_assistant flow so users can ask the
realtime model what's loaded / installed / configured. Tools are run
server-side via the same in-process MCP holder that powers the chat
modality — no transport switch, no proxy, no new wire protocol.
Wire:
- core/http/endpoints/openai/realtime.go:
- RealtimeSessionOptions{LocalAIAssistant,IsAdmin}; isCurrentUserAdmin
helper mirrors chat.go's requireAssistantAccess (no-op when auth
disabled, else requires auth.RoleAdmin).
- Session grows AssistantExecutor mcpTools.ToolExecutor.
- runRealtimeSession, when opts.LocalAIAssistant is set: gate on admin,
fail closed if DisableLocalAIAssistant or the holder has no tools,
DiscoverTools and inject into session.Tools, prepend
holder.SystemPrompt() to instructions.
- Tool-call dispatch loop: when AssistantExecutor.IsTool(name), run
ExecuteTool inproc, append a FunctionCallOutput to conv.Items, skip
the function_call_arguments client emit (the client can't execute
these — it doesn't know about them). After the loop, if any
assistant tool ran, trigger another response so the model speaks the
result. Mirrors chat's agentic loop, driven server-side rather than
via client round-trip.
- core/http/endpoints/openai/realtime_webrtc.go: RealtimeCallRequest
gains `localai_assistant` (JSON omitempty). Handshake calls
isCurrentUserAdmin and builds RealtimeSessionOptions.
- core/http/react-ui/src/pages/Talk.jsx: admin-only "Manage Mode"
checkbox under the Tools dropdown; passes localai_assistant: true to
realtimeApi.call's body, captured in the connect callback's deps.
Mirroring chat's pattern means the in-process MCP tools surface "just
works" for the Talk page without exposing a Streamable-HTTP MCP endpoint
(which was the alternative). Clients with their own MCP servers can
still use the existing ClientMCPDropdown path in parallel; the realtime
handler distinguishes them by AssistantExecutor.IsTool() at dispatch
time.
Assisted-by: claude-code:claude-opus-4-7-1m [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* feat(realtime): render Manage Mode tool calls in the Talk transcript
Previously the realtime endpoint only emitted response.output_item.added
for the FunctionCall item, and Talk.jsx's switch ignored the event — so
server-side tool runs were invisible in the UI. The model would speak
the result but the user had no way to see what tool was actually
called.
realtime.go: after executing an assistant tool inproc, emit a second
output_item.added/.done pair for the FunctionCallOutput item. Mirrors
the way the chat page displays tool_call + tool_result blocks.
Talk.jsx: handle both response.output_item.added and .done. Render
FunctionCall (with arguments) and FunctionCallOutput (pretty-printed
JSON when possible) as two transcript entries — `tool_call` with the
wrench icon, `tool_result` with the clipboard icon, both in mono-space
secondary-colour. Resets streamingRef after the result so the next
assistant text delta starts a fresh transcript entry instead of
appending to the previous turn.
Assisted-by: claude-code:claude-opus-4-7-1m [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* refactor(realtime): bound the Manage Mode tool-loop + preserve assistant tools
Fallout from a review pass on the Manage Mode patches:
- Bound the server-side agentic loop. triggerResponse used to recurse on
executedAssistantTool with no cap — a model that kept calling tools
would blow the goroutine stack. New maxAssistantToolTurns = 10 (mirrors
useChat.js's maxToolTurns). Public triggerResponse is now a thin shim
over triggerResponseAtTurn(toolTurn int); recursion increments the
counter and stops at the cap with an xlog.Warn.
- Preserve Manage Mode tools across client session.update. The handler
used to blindly overwrite session.Tools, so toggling a client MCP
server mid-session silently wiped the in-process admin tools. Session
now caches the original AssistantTools slice at session creation and
the session.update handler merges them back in (client names win on
collision — the client is explicit).
- strconv.ParseBool for the localai_assistant query param instead of
hand-rolled "1" || "true". Mirrors LocalAIAssistantFromMetadata.
- Talk.jsx: render both tool_call and tool_result on
response.output_item.done instead of splitting them across .added and
.done. The server's event pairing (added → done) stays correct; the
UI just doesn't need to inspect both phases of the same item. One
switch case instead of two, no behavioural change.
Out of scope (noted for follow-ups): extract a shared assistant-tools
helper between chat.go and realtime.go (duplication is small enough
that two parallel implementations stay readable for now), and an i18n
key for the Manage Mode helper text (Talk.jsx doesn't use i18n
anywhere else yet).
Assisted-by: claude-code:claude-opus-4-7-1m [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* ci(test-extra): wire liquid-audio backend smoke test
The backend ships test.py + a `make test` target and is listed in
backend-matrix.yml, so scripts/changed-backends.js already writes a
`liquid-audio=true|false` output when files under backend/python/liquid-audio/
change. The workflow just wasn't reading it.
- Expose the `liquid-audio` output on the detect-changes job
- Add a tests-liquid-audio job that runs `make` + `make test` in
backend/python/liquid-audio, gated on the per-backend detect flag
The smoke covers Health() and LoadModel(mode:finetune); fine-tune mode
short-circuits before any HuggingFace download (backend.py:192), so the
job needs neither weights nor a GPU. The full-inference path remains
gated on LIQUID_AUDIO_MODEL_ID, which CI doesn't set.
The four new Go test files (core/gallery/importers/liquid-audio_test.go,
core/http/endpoints/openai/realtime_gate_test.go,
core/http/routes/ui_pipeline_models_test.go, pkg/functions/parse_lfm2_test.go)
are already picked up by the existing test.yml workflow via `make test` →
`ginkgo -r ./pkg/... ./core/...`; their packages all carry RunSpecs entries.
Assisted-by: Claude:claude-opus-4-7
Signed-off-by: Richard Palethorpe <io@richiejp.com>
---------
Signed-off-by: Richard Palethorpe <io@richiejp.com>
The PR that introduced the per-arch + manifest-merge pilot (#9727)
only touched CI infrastructure files, so the path filter correctly
skipped backend builds on its merge commit. To observe the new
backend-merge-jobs flow assemble a real manifest list, this commit
touches faster-whisper's Makefile so its two new per-arch entries
schedule and the merge job runs.
The trailing comment is the smallest possible diff and is harmless
to the build.
Assisted-by: Claude:claude-opus-4-7
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Bring the sglang Python backend up to feature parity with vllm by adding
the same engine_args:-map plumbing the vLLM backend already has. Any
ServerArgs field (~380 in sglang 0.5.11) becomes settable from a model
YAML, including the speculative-decoding flags needed for Multi-Token
Prediction. Validation matches the vllm backend's: keys are checked
against dataclasses.fields(ServerArgs), unknown keys raise ValueError
with a difflib close-match suggestion at LoadModel time, and the typed
ModelOptions fields keep their existing meaning with engine_args
overriding them.
Backend code:
* backend/python/sglang/backend.py: add _apply_engine_args, import
dataclasses/difflib/ServerArgs, call from LoadModel; rename Seed ->
sampling_seed (sglang 0.5.11 renamed the SamplingParams field).
* backend/python/sglang/test.py + test.sh + Makefile: six unit tests
exercising the helper directly (no engine load required).
Build / CI / backend gallery (cuda13 + l4t13 paths are now first-class):
* backend/python/sglang/install.sh: add --prerelease=allow because
sglang 0.5.11 hard-pins flash-attn-4 which only ships beta wheels;
add --index-strategy=unsafe-best-match for cublas12 so the cu128
torch index wins over default-PyPI's cu130; new pyproject.toml-driven
l4t13 install path so [tool.uv.sources] can pin torch/torchvision/
torchaudio/sglang to the jetson-ai-lab index without forcing every
transitive PyPI dep through the L4T mirror's flaky proxy (mirrors the
equivalent fix in backend/python/vllm/install.sh).
* backend/python/sglang/pyproject.toml (new): L4T project spec with
explicit-source jetson-ai-lab index. Replaces requirements-l4t13.txt
for the l4t13 BUILD_PROFILE; other profiles still go through the
requirements-*.txt pipeline via libbackend.sh's installRequirements.
* backend/python/sglang/requirements-l4t13.txt: removed; superseded
by pyproject.toml.
* backend/python/sglang/requirements-cublas{12,13}{,-after}.txt: pin
sglang>=0.5.11 (Gemma 4 floor); add cu130 torch index for cublas13
(new files) and cu128 torch index for cublas12 (default PyPI now
ships cu130 torch wheels by default and breaks cu12 hosts).
* backend/index.yaml: add cuda13-sglang and cuda13-sglang-development
capability mappings + image entries pointing at
quay.io/.../-gpu-nvidia-cuda-13-sglang.
* .github/workflows/backend.yml: new cublas13 sglang matrix entry,
mirroring vllm's cuda13 build.
Model gallery + docs:
* gallery/sglang.yaml: base sglang config template, mirrors vllm.yaml.
* gallery/sglang-gemma-4-{e2b,e4b}-mtp.yaml: Gemma 4 MTP demos
transcribed verbatim from the SGLang Gemma 4 cookbook MTP commands.
* gallery/sglang-mimo-7b-mtp.yaml: MiMo-7B-RL with built-in MTP heads
+ online fp8 weight quantization, verified end-to-end on a 16 GB
RTX 5070 Ti at ~88 tok/s. Uses mem_fraction_static: 0.7 because the
MTP draft worker's vocab embedding is loaded unquantised and OOMs
the static reservation at sglang's 0.85 default.
* gallery/index.yaml: three new entries (gemma-4-e2b-it:sglang-mtp,
gemma-4-e4b-it:sglang-mtp, mimo-7b-mtp:sglang).
* docs/content/features/text-generation.md: new SGLang section with
setup, engine_args reference, MTP demos, version requirements.
* .agents/sglang-backend.md (new): agent one-pager covering the flat
ServerArgs structure, the typed-vs-engine_args precedence, the
speculative-decoding cheatsheet, and the mem_fraction_static gotcha
documented above.
* AGENTS.md: index entry for the new agent doc.
Known limitation: the two Gemma 4 MTP gallery entries ship a recipe
that doesn't yet run on stock libraries. The drafter checkpoints
(google/gemma-4-{E2B,E4B}-it-assistant) declare
model_type: gemma4_assistant / Gemma4AssistantForCausalLM, which
neither transformers (<=5.6.0, including the SGLang cookbook's pinned
commit 91b1ab1f... and main HEAD) nor sglang's own model registry
(<=0.5.11) registers as of 2026-05-06. They will start working when
HF or sglang upstream registers the architecture -- no LocalAI
changes needed. The MiMo MTP demo and the non-MTP Gemma 4 paths work
today on this build (verified on RTX 5070 Ti, 16 GB).
Assisted-by: Claude:claude-opus-4-7 [Read] [Edit] [Bash] [WebFetch] [WebSearch]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
The previous omegaconf pin only addressed one symptom of a deeper problem:
chatterbox-tts upstream depends on `russian-text-stresser` (unpinned git URL),
which transitively pins `spacy==3.6.*` and other ancient packages. That cascade
forces pip to backtrack through Jinja2/MarkupSafe/omegaconf into Python-2-era
sdists that no longer build (e.g. ruamel.yaml<0.15, Jinja2 2.6 importing the
long-removed `setuptools.Feature`).
Install chatterbox-tts itself with --no-deps in install.sh and list its real
runtime deps explicitly in each requirements-*.txt, dropping the optional
russian-text-stresser. This unblocks the darwin (and other) builds without
playing whack-a-mole on each newly-discovered transitive pin.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
The previous pin in requirements.txt was ineffective: installRequirements
runs a separate `pip install --requirement` per file, so resolution does
not carry over to the per-profile file where chatterbox-tts is declared.
With chatterbox-tts's unpinned `omegaconf` dep, pip backtracked through
1.x sdists into ruamel.yaml<0.15, whose Python-2-era setup.py fails on
Python 3.10+.
Pin omegaconf==2.3.0 next to chatterbox-tts in every profile file
(matches what upstream chatterbox uses). Drop the dead pin from
requirements.txt.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Without an upper-floor pin, pip's resolver backtracks through omegaconf 1.x
sdists when installing chatterbox-tts. Old 1.x setups depend on
ruamel.yaml<0.15, whose setup.py uses Python-2-era names (Str, Bytes) and
fails to build on Python 3.10+, breaking the darwin python backend build.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Two unrelated CI breakages bundled together since both are one-liners:
- rerankers: bump torch 2.4.1 -> 2.7.1 on cpu/cublas12. The unpinned
transformers resolves to 5.x, whose moe.py registers a custom_op with
string-typed `'torch.Tensor'` annotations that torch 2.4.1's
infer_schema rejects, blocking the gRPC server from starting and
failing all 5 backend tests with "Connection refused" on :50051.
Matches the version used by the transformers backend.
- vllm-omni: strip fa3-fwd from the upstream requirements/cuda.txt
before resolving on aarch64. fa3-fwd 0.0.3 ships only an
x86_64 wheel and has no sdist, making the cuda profile unsatisfiable
on Jetson/SBSA. fa3-fwd is a soft runtime dep — vllm-omni's
attention backends fall back to FA2 then SDPA when it's missing.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
Two related runtime fixes for Python backends that JIT-compile CUDA
kernels at first model load (FlashInfer, PyTorch inductor, triton):
1. libbackend.sh: replace `source ${EDIR}/venv/bin/activate` with a
minimal manual setup (_activateVenv: export VIRTUAL_ENV, prepend
PATH, unset PYTHONHOME) computed from $EDIR at runtime. `uv venv`
and `python -m venv` both bake the create-time absolute path into
bin/activate (e.g. VIRTUAL_ENV='/vllm/venv' from the Docker build
stage), so sourcing activate on a relocated venv — copied out of
the build container and unpacked at an arbitrary backend dir —
prepends a stale, non-existent path to $PATH. Pip-installed CLI
tools (e.g. ninja, used by FlashInfer's NVFP4 GEMM JIT) are then
never found and the load aborts with FileNotFoundError. Doing the
env setup ourselves matches what `uv run` does internally and
sidesteps the relocation problem entirely. Generic — every Python
backend benefits.
2. vllm/run.sh: replace ninja's default -j$(nproc)+2 with an adaptive
MAX_JOBS = min(nproc, (MemAvailable-4)/4). Each concurrent
nvcc/cudafe++ peaks at multiple GiB; the default OOM-kills on
memory-tight hosts (e.g. a 16 GiB desktop loading a 27B NVFP4
model) but underutilises 100-core / 1 TB boxes. User-set MAX_JOBS
still wins. Also pin NVCC_THREADS=2 unless overridden.
Refs: https://github.com/vllm-project/vllm/issues/20079
Assisted-by: Claude:claude-opus-4-7 [Edit] [Bash]
* feat(vllm): build vllm from source for Intel XPU
Upstream publishes no XPU wheels for vllm. The Intel profile was
silently picking up a non-XPU wheel that imported but errored at
engine init, and several runtime deps (pillow, charset-normalizer,
chardet) were missing on Intel -- backend.py crashed at import time
before the gRPC server came up.
Switch the Intel profile to upstream's documented from-source
procedure (docs/getting_started/installation/gpu.xpu.inc.md in
vllm-project/vllm):
- Bump portable Python to 3.12 -- vllm-xpu-kernels ships only a
cp312 wheel.
- Source /opt/intel/oneapi/setvars.sh so vllm's CMake build sees
the dpcpp/sycl compiler from the oneapi-basekit base image.
- Hide requirements-intel-after.txt during installRequirements
(it used to 'pip install vllm'); install vllm's deps from a
fresh git clone of vllm via 'uv pip install -r
requirements/xpu.txt', swap stock triton for
triton-xpu==3.7.0, then 'VLLM_TARGET_DEVICE=xpu uv pip install
--no-deps .'.
- requirements-intel.txt trimmed to LocalAI's direct deps
(accelerate / transformers / bitsandbytes); torch-xpu, vllm,
vllm_xpu_kernels and the rest come from upstream's xpu.txt
during the source build.
- requirements.txt: add pillow + charset-normalizer + chardet --
used by backend.py and missing on the Intel install profile.
- run.sh: 'set -x' so backend startup is visible in container
logs (the gRPC startup error path was previously opaque).
Also adds a one-line docs example for engine_args.attention_backend
under the vLLM section, since older XE-HPG GPUs (e.g. Arc A770)
need TRITON_ATTN to bypass the cutlass path in vllm_xpu_kernels.
Tested end-to-end on an Intel Arc A770 with Qwen2.5-0.5B-Instruct
via LocalAI's /v1/chat/completions.
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* feat(vllm): add multi-node data-parallel follower worker
vLLM v1's multi-node story is one process per node sharing a DP
coordinator over ZMQ -- the head runs the API server with
data_parallel_size > 1 and followers run `vllm serve --headless ...`
with matching topology. Today LocalAI can already configure DP on the
head via the engine_args YAML map, but there's no way to bring up the
follower nodes -- so the head sits waiting for ranks that never
handshake.
Add `local-ai p2p-worker vllm`, mirroring MLXDistributed's structural
precedent (operator-launched, static config, no NATS placement). The
worker:
- Optionally self-registers with the frontend as an agent-type node
tagged `node.role=vllm-follower` so it's visible in the admin UI
and operators can scope ordinary models away via inverse
selectors.
- Resolves the platform-specific vllm backend via the gallery's
"vllm" meta-entry (cuda*, intel-vllm, rocm-vllm, ...).
- Runs vLLM as a child process so the heartbeat goroutine survives
until vLLM exits; forwards SIGINT/SIGTERM so vLLM can clean up its
ZMQ sockets before we tear down.
- Validates --headless + --start-rank 0 is rejected (rank 0 is the
head and must serve the API).
Backend run.sh dispatches `serve` as the first arg to vllm's own CLI
instead of LocalAI's backend.py gRPC server -- the follower speaks
ZMQ directly to the head, there is no LocalAI gRPC on the follower
side. Single-node usage is unchanged.
Generalises the gallery resolution helper into findBackendPath()
shared by MLX and vLLM workers; extracts ParseNodeLabels for the
comma-separated label parsing both use.
Ships with two compose recipes (`docker-compose.vllm-multinode.yaml`
for NVIDIA, `docker-compose.vllm-multinode.intel.yaml` for Intel
XPU/xccl) plus `tests/e2e/vllm-multinode/smoke.sh`. Both vendors are
supported (NCCL for CUDA/ROCm, xccl for XPU) but mixed-vendor DP is
not -- PyTorch's process group requires every rank to use the same
collective backend, and NCCL/xccl/gloo don't interoperate.
Out of scope (deferred): SmartRouter-driven placement of follower
ranks via NATS backend.install events, follower log streaming through
/api/backend-logs, tensor-parallel across nodes, disaggregated
prefill via KVTransferConfig.
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* test(vllm): CPU-only end-to-end test for multi-node DP
Adds tests/e2e/vllm-multinode/, a Ginkgo + testcontainers-go suite
that brings up a head + headless follower from the locally-built
local-ai:tests image, bind-mounts the cpu-vllm backend extracted by
make extract-backend-vllm so it's seen as a system backend (no gallery
fetch, no registry server), and asserts a chat completion across both
DP ranks. New `make test-e2e-vllm-multinode` target wires the docker
build, backend extract, and ginkgo run together; BuildKit caches both
images so re-runs only rebuild what changed. Tagged Label("VLLMMultinode")
so the existing distributed suite isn't pulled along.
Two pre-existing bugs surfaced by the test:
1. extract-backend-% (Makefile) failed for every backend, because all
backend images end with `FROM scratch` and `docker create` rejects
an image with no CMD/ENTRYPOINT. Fixed by passing
--entrypoint=/run.sh -- the container is never started, only
docker-cp'd, so the path doesn't have to exist; we just need
anything that satisfies the daemon's create-time validation.
2. backend/python/vllm/run.sh's `serve` shortcut for the multi-node DP
follower exec'd ${EDIR}/venv/bin/vllm directly, but uv bakes an
absolute build-time shebang (`#!/vllm/venv/bin/python3`) that no
longer resolves once the backend is relocated to BackendsPath.
_makeVenvPortable's shebang rewriter only matches paths that
already point at ${EDIR}, so the original shebang slips through
unchanged. Fixed by exec-ing ${EDIR}/venv/bin/python with the script
as an argument -- Python ignores the script's shebang in that case.
The test fixture caps memory aggressively (max_model_len=512,
VLLM_CPU_KVCACHE_SPACE=1, TORCH_COMPILE_DISABLE=1) so two CPU engines
fit on a 32 GB box. TORCH_COMPILE_DISABLE is currently mandatory for
cpu-vllm: torch._inductor's CPU-ISA probe runs even with
enforce_eager=True and needs g++ on PATH, which the LocalAI runtime
image doesn't ship -- to be addressed in a follow-up that bundles a
toolchain in the cpu-vllm backend.
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* feat(vllm): bundle a g++ toolchain in the cpu-vllm backend image
torch._inductor's CPU-ISA probe (`cpu_model_runner.py:65 "Warming up
model for the compilation"`) shells out to `g++` at vllm engine
startup, regardless of `enforce_eager=True` -- the eager flag only
disables CUDA graphs, not inductor's first-batch warmup. The LocalAI
CPU runtime image (Dockerfile, unconditional apt list) does not ship
build-essential, and the cpu-vllm backend image is `FROM scratch`,
so any non-trivial inference on cpu-vllm crashes with:
torch._inductor.exc.InductorError:
InvalidCxxCompiler: No working C++ compiler found in
torch._inductor.config.cpp.cxx: (None, 'g++')
Bundling the toolchain in the CPU runtime image would bloat every
non-vllm-CPU deployment and force a single GCC version on backends
that may want clang or a different version. So this lives in the
backend, gated to BUILD_TYPE=='' (the CPU profile).
`package.sh` snapshots g++ + binutils + cc1plus + libstdc++ + libc6
(runtime + dev) + the math libs cc1plus links (libisl/libmpc/libmpfr/
libjansson) into ${BACKEND}/toolchain/, mirroring /usr/... layout. The
unversioned binaries on Debian/Ubuntu are symlink chains pointing into
multiarch packages (`g++` -> `g++-13` -> `x86_64-linux-gnu-g++-13`,
the latter in `g++-13-x86-64-linux-gnu`), so the package list resolves
both the version and the arch-triplet variant. Symlinks /lib ->
usr/lib and /lib64 -> usr/lib64 are recreated under the toolchain
root because Ubuntu's UsrMerge keeps them at /, and ld scripts
(`libc.so`, `libm.so`) hardcode `/lib/...` paths that --sysroot
re-roots into the toolchain.
The unversioned `g++`/`gcc`/`cpp` symlinks are replaced with wrapper
shell scripts that resolve their own location at runtime and pass
`--sysroot=<toolchain>` and `-B <toolchain>/usr/lib/gcc/<triplet>/<ver>/`
to the underlying versioned binary. That's how torch's bare `g++ foo.cpp
-o foo` invocation finds cc1plus (-B), system headers (--sysroot), and
the bundled libstdc++ (--sysroot, --sysroot is recursive into linker).
`run.sh` adds the toolchain bin dir to PATH and the toolchain's
shared-lib dir to LD_LIBRARY_PATH -- everything else (header search,
linker search, executable search) is encapsulated in the wrappers.
No-op for non-CPU builds, the dir doesn't exist there.
The cpu-vllm image grows by ~217 MB. Tradeoff is acceptable -- cpu-vllm
is already a niche profile (few users compared to GPU vllm) and the
alternative is a backend that crashes at first inference unless the
operator manually sets TORCH_COMPILE_DISABLE=1, which silently disables
all torch.compile optimizations.
Drops `TORCH_COMPILE_DISABLE=1` from tests/e2e/vllm-multinode -- the
smoke now exercises the real compile path through the bundled toolchain.
Test runtime is +20s for the warmup compile, still <90s end to end.
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* fix(vllm): scope jetson-ai-lab index to L4T-specific wheels via pyproject.toml
The L4T arm64 build resolves dependencies through pypi.jetson-ai-lab.io,
which hosts the L4T-specific torch / vllm / flash-attn wheels but also
transparently proxies the rest of PyPI through `/+f/<sha>/<filename>`
URLs. With `--extra-index-url` + `--index-strategy=unsafe-best-match`
uv would pick those proxy URLs for ordinary PyPI packages —
anthropic/openai/propcache/annotated-types — and fail when the proxy
503s. Master is hitting the same bug on its own l4t-vllm matrix entry.
Switch the l4t13 install path to a pyproject.toml that marks the
jetson-ai-lab index `explicit = true` and pins only torch, torchvision,
torchaudio, flash-attn, and vllm to it via [tool.uv.sources]. uv won't
consult the L4T mirror for anything else, so transitive deps fall back
to PyPI as the default index — no exposure to the proxy 503s.
`uv pip install -r requirements.txt` ignores [tool.uv.sources], so the
l4t13 branch in install.sh now invokes `uv pip install --requirement
pyproject.toml` directly, replacing the old requirements-l4t13*.txt
files. Other BUILD_PROFILEs continue using libbackend.sh's
installRequirements and never read pyproject.toml.
Local resolution test (x86_64, dry-run) confirms uv hits the L4T
index for torch and falls through to PyPI for everything else.
Assisted-by: claude-code:claude-opus-4-7-1m [Read] [Edit] [Bash] [Write]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
---------
Signed-off-by: Richard Palethorpe <io@richiejp.com>
`int(x) * 1e9` returns a float because `1e9` is a float literal, but
TranscriptSegment.start/end are integer protobuf fields. This caused
every transcription request to fail with:
TypeError: 'float' object cannot be interpreted as an integer
Multiply first, then cast — `int(x * 1e9)` — to get an int as required.
compel 2.3.1 (latest, Nov 2025) declares transformers~=4.25 in its
metadata, i.e. >=4.25,<5.0. After transformers 5.0 (2026-01-26) and
huggingface-hub 1.0 (2025-10-27) shipped, the weekly DEPS_REFRESH
cache rotation in CI started seeing the new majors and pip's resolver
went into multi-hour backtracking storms walking every transformers
4.x candidate against every accelerate/hf-hub/tokenizers combination
to find a set compel would accept. The 2026-04-29 backend-build for
the diffusers backend (darwin-mps + l4t + cublas13-turboquant matrix
cells) hit the GitHub Actions 6h job timeout still inside pip
install — the build itself never started.
compel is the only hard upper bound on transformers in this stack
(diffusers, accelerate, peft, optimum-quanto are all flexible), and
upstream support for transformers 5 is still in flight: damian0815/
compel#129 ("Modernize Compel for Transformers 5") and #128 ("Bump
transformers version to >5.0") are both open as of today.
backend.py only constructs Compel() when COMPEL=1 is set in the env
(default off), so make compel a true optional extra:
- Wrap the top-level `from compel import ...` in try/except
ImportError, mirroring the existing sd_embed pattern.
- Auto-disable COMPEL with a warning when the module isn't
installed, instead of crashing on module load.
- Drop compel from all eight requirements-*.txt variants so the
resolver no longer has to satisfy its transformers cap.
- Leave a TODO in backend.py and in each requirements file
pointing at the upstream PR/issue, so the dependency can be
reinstated once compel supports transformers >= 5.
Users who rely on weighted-prompt embeddings can opt in with a
manual `pip install compel` alongside COMPEL=1; the warning emitted
on startup tells them how.
Assisted-by: Claude:claude-opus-4-7 [Bash Read Edit WebFetch]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Closes#9601
Makes the temporary scratch paths in vllm, vllm-omni, tinygrad, and pocket-tts
backends configurable via the standard TMPDIR env var, instead of always writing
to /tmp. This is a one-line change per call site that calls tempfile.gettempdir()
for the directory and keeps the same filename suffix.
Users who run on systems with a small root partition (or want to relocate scratch
files to a larger volume) can now redirect these by setting TMPDIR
(e.g. TMPDIR=/data/tmp), without affecting the existing LOCALAI_GENERATED_CONTENT_PATH
or LOCALAI_UPLOAD_PATH options that already cover other temp paths.
Files touched:
- backend/python/vllm/backend.py (1 site: video base64 scratch)
- backend/python/tinygrad/backend.py (1 site: image fallback dst)
- backend/python/pocket-tts/backend.py (1 site: tts wav fallback dst)
- backend/python/vllm-omni/backend.py (2 sites: video + audio scratch)
* feat(vllm): expose AsyncEngineArgs via generic engine_args YAML map
LocalAI's vLLM backend wraps a small typed subset of vLLM's
AsyncEngineArgs (quantization, tensor_parallel_size, dtype, etc.).
Anything outside that subset -- pipeline/data/expert parallelism,
speculative_config, kv_transfer_config, all2all_backend, prefix
caching, chunked prefill, etc. -- requires a new protobuf field, a
Go struct field, an options.go line, and a backend.py mapping per
feature. That cadence is the bottleneck on shipping vLLM's
production feature set.
Add a generic `engine_args:` map on the model YAML that is
JSON-serialised into a new ModelOptions.EngineArgs proto field and
applied verbatim to AsyncEngineArgs at LoadModel time. Validation
is done by the Python backend via dataclasses.fields(); unknown
keys fail with the closest valid name as a hint.
dataclasses.replace() is used so vLLM's __post_init__ re-runs and
auto-converts dict values into nested config dataclasses
(CompilationConfig, AttentionConfig, ...). speculative_config and
kv_transfer_config flow through as dicts; vLLM converts them at
engine init.
Operators can now write:
engine_args:
data_parallel_size: 8
enable_expert_parallel: true
all2all_backend: deepep_low_latency
speculative_config:
method: deepseek_mtp
num_speculative_tokens: 3
kv_cache_dtype: fp8
without further proto/Go/Python plumbing per field.
Production defaults seeded by hooks_vllm.go: enable_prefix_caching
and enable_chunked_prefill default to true unless explicitly set.
Existing typed YAML fields (gpu_memory_utilization,
tensor_parallel_size, etc.) remain for back-compat; engine_args
overrides them when both are set.
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* chore(vllm): pin cublas13 to vLLM 0.20.0 cu130 wheel
vLLM's PyPI wheel is built against CUDA 12 (libcudart.so.12) and won't
load on a cu130 host. Switch the cublas13 build to vLLM's per-tag cu130
simple-index (https://wheels.vllm.ai/0.20.0/cu130/) and pin
vllm==0.20.0. The cu130-flavoured wheel ships libcudart.so.13 and
includes the DFlash speculative-decoding method that landed in 0.20.0.
cublas13 install gets --index-strategy=unsafe-best-match so uv consults
both the cu130 index and PyPI when resolving — PyPI also publishes
vllm==0.20.0, but with cu12 binaries that error at import time.
Verified: Qwen3.5-4B + z-lab/Qwen3.5-4B-DFlash loads and serves chat
completions on RTX 5070 Ti (sm_120, cu130).
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* ci(vllm): bot job to bump cublas13 vLLM wheel pin
vLLM's cu130 wheel index URL is itself version-locked
(wheels.vllm.ai/<TAG>/cu130/, no /latest/ alias upstream), so a vLLM
bump means rewriting two values atomically — the URL segment and the
version constraint. bump_deps.sh handles git-sha-in-Makefile only;
add a sibling bump_vllm_wheel.sh and a matching workflow job that
mirrors the existing matrix's PR-creation pattern.
The bumper queries /releases/latest (which excludes prereleases),
strips the leading 'v', and seds both lines unconditionally. When the
file is already on the latest tag the rewrite is a no-op and
peter-evans/create-pull-request opens no PR.
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* docs(vllm): document engine_args and speculative decoding
The new engine_args: map plumbs arbitrary AsyncEngineArgs through to
vLLM, but the public docs only covered the basic typed fields. Add a
short subsection in the vLLM section explaining the typed/generic
split and showing a worked DFlash speculative-decoding config, with
pointers to vLLM's SpeculativeConfig reference and z-lab's drafter
collection.
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Richard Palethorpe <io@richiejp.com>
---------
Signed-off-by: Richard Palethorpe <io@richiejp.com>
Co-authored-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
fastsafetensors==0.3 (transitive dep of vllm) imports pybind11 in
setup.py without declaring it in build-system.requires. With
--no-build-isolation it has to already exist in the venv, otherwise the
wheel build fails with ModuleNotFoundError on arm64 L4T CUDA 13 (and
any other profile that picks up vllm 0.20.0).
Assisted-by: Claude:claude-opus-4-7 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Blaizzy/mlx-vlm git HEAD bumped its constraint to mlx>=0.31.2, but
mlx-cuda-12 and mlx-cuda-13 are only published up to 0.31.1 on PyPI.
Since mlx[cudaXX]==0.31.2 forces a sibling wheel that doesn't exist,
pip backtracks through every older mlx[cudaXX], none of which satisfy
mlx>=0.31.2, producing ResolutionImpossible.
Pin all variants to the v0.4.4 tag (mlx>=0.30.0), which resolves
cleanly against mlx[cuda13]==0.31.1. cpu/mps weren't broken yet but
are pinned for consistency.
Assisted-by: Claude:claude-opus-4-7
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
