* feat(backend): add tinygrad multimodal backend
Wire tinygrad as a new Python backend covering LLM text generation with
native tool-call extraction, embeddings, Stable Diffusion 1.x image
generation, and Whisper speech-to-text from a single self-contained
container.
Backend (`backend/python/tinygrad/`):
- `backend.py` gRPC servicer with LLM Predict/PredictStream (auto-detects
Llama / Qwen2 / Mistral architecture from `config.json`, supports
safetensors and GGUF), Embedding via mean-pooled last hidden state,
GenerateImage via the vendored SD1.x pipeline, AudioTranscription +
AudioTranscriptionStream via the vendored Whisper inference loop, plus
Tokenize / ModelMetadata / Status / Free.
- Vendored upstream model code under `vendor/` (MIT, headers preserved):
llama.py with an added `qkv_bias` flag for Qwen2-family bias support
and an `embed()` method that returns the last hidden state, plus
clip.py, unet.py, stable_diffusion.py (trimmed to drop the MLPerf
training branch that pulls `mlperf.initializers`), audio_helpers.py
and whisper.py (trimmed to drop the pyaudio listener).
- Pluggable tool-call parsers under `tool_parsers/`: hermes (Qwen2.5 /
Hermes), llama3_json (Llama 3.1+), qwen3_xml (Qwen 3), mistral
(Mistral / Mixtral). Auto-selected from model architecture or `Options`.
- `install.sh` pins Python 3.11.14 (tinygrad >=0.12 needs >=3.11; the
default portable python is 3.10).
- `package.sh` bundles libLLVM.so.1 + libedit/libtinfo/libgomp/libsndfile
into the scratch image. `run.sh` sets `CPU_LLVM=1` and `LLVM_PATH` so
tinygrad's CPU device uses the in-process libLLVM JIT instead of
shelling out to the missing `clang` binary.
- Local unit tests for Health and the four parsers in `test.py`.
Build wiring:
- Root `Makefile`: `.NOTPARALLEL`, `prepare-test-extra`, `test-extra`,
`BACKEND_TINYGRAD = tinygrad|python|.|false|true`,
docker-build-target eval, and `docker-build-backends` aggregator.
- `.github/workflows/backend.yml`: cpu / cuda12 / cuda13 build matrix
entries (mirrors the transformers backend placement).
- `backend/index.yaml`: `&tinygrad` meta + cpu/cuda12/cuda13 image
entries (latest + development).
E2E test wiring:
- `tests/e2e-backends/backend_test.go` gains an `image` capability that
exercises GenerateImage and asserts a non-empty PNG is written to
`dst`. New `BACKEND_TEST_IMAGE_PROMPT` / `BACKEND_TEST_IMAGE_STEPS`
knobs.
- Five new make targets next to `test-extra-backend-vllm`:
- `test-extra-backend-tinygrad` — Qwen2.5-0.5B-Instruct + hermes,
mirrors the vllm target 1:1 (5/9 specs in ~57s).
- `test-extra-backend-tinygrad-embeddings` — same model, embeddings
via LLM hidden state (3/9 in ~10s).
- `test-extra-backend-tinygrad-sd` — stable-diffusion-v1-5 mirror,
health/load/image (3/9 in ~10min, 4 diffusion steps on CPU).
- `test-extra-backend-tinygrad-whisper` — openai/whisper-tiny.en
against jfk.wav from whisper.cpp samples (4/9 in ~49s).
- `test-extra-backend-tinygrad-all` aggregate.
All four targets land green on the first MVP pass: 15 specs total, 0
failures across LLM+tools, embeddings, image generation, and speech
transcription.
* refactor(tinygrad): collapse to a single backend image
tinygrad generates its own GPU kernels (PTX renderer for CUDA, the
autogen ctypes wrappers for HIP / Metal / WebGPU) and never links
against cuDNN, cuBLAS, or any toolkit-version-tied library. The only
runtime dependency that varies across hosts is the driver's libcuda.so.1
/ libamdhip64.so, which are injected into the container at run time by
the nvidia-container / rocm runtimes. So unlike torch- or vLLM-based
backends, there is no reason to ship per-CUDA-version images.
- Drop the cuda12-tinygrad and cuda13-tinygrad build-matrix entries
from .github/workflows/backend.yml. The sole remaining entry is
renamed to -tinygrad (from -cpu-tinygrad) since it is no longer
CPU-only.
- Collapse backend/index.yaml to a single meta + development pair.
The meta anchor carries the latest uri directly; the development
entry points at the master tag.
- run.sh picks the tinygrad device at launch time by probing
/usr/lib/... for libcuda.so.1 / libamdhip64.so. When libcuda is
visible we set CUDA=1 + CUDA_PTX=1 so tinygrad uses its own PTX
renderer (avoids any nvrtc/toolkit dependency); otherwise we fall
back to HIP or CLANG. CPU_LLVM=1 + LLVM_PATH keep the in-process
libLLVM JIT for the CLANG path.
- backend.py's _select_tinygrad_device() is trimmed to a CLANG-only
fallback since production device selection happens in run.sh.
Re-ran test-extra-backend-tinygrad after the change:
Ran 5 of 9 Specs in 56.541 seconds — 5 Passed, 0 Failed
* feat(backend): add turboquant llama.cpp-fork backend
turboquant is a llama.cpp fork (TheTom/llama-cpp-turboquant, branch
feature/turboquant-kv-cache) that adds a TurboQuant KV-cache scheme.
It ships as a first-class backend reusing backend/cpp/llama-cpp sources
via a thin wrapper Makefile: each variant target copies ../llama-cpp
into a sibling build dir and invokes llama-cpp's build-llama-cpp-grpc-server
with LLAMA_REPO/LLAMA_VERSION overridden to point at the fork. No
duplication of grpc-server.cpp — upstream fixes flow through automatically.
Wires up the full matrix (CPU, CUDA 12/13, L4T, L4T-CUDA13, ROCm, SYCL
f32/f16, Vulkan) in backend.yml and the gallery entries in index.yaml,
adds a tests-turboquant-grpc e2e job driven by BACKEND_TEST_CACHE_TYPE_K/V=q8_0
to exercise the KV-cache config path (backend_test.go gains dedicated env
vars wired into ModelOptions.CacheTypeKey/Value — a generic improvement
usable by any llama.cpp-family backend), and registers a nightly auto-bump
PR in bump_deps.yaml tracking feature/turboquant-kv-cache.
scripts/changed-backends.js gets a special-case so edits to
backend/cpp/llama-cpp/ also retrigger the turboquant CI pipeline, since
the wrapper reuses those sources.
* feat(turboquant): carry upstream patches against fork API drift
turboquant branched from llama.cpp before upstream commit 66060008
("server: respect the ignore eos flag", #21203) which added the
`logit_bias_eog` field to `server_context_meta` and a matching
parameter to `server_task::params_from_json_cmpl`. The shared
backend/cpp/llama-cpp/grpc-server.cpp depends on that field, so
building it against the fork unmodified fails.
Cherry-pick that commit as a patch file under
backend/cpp/turboquant/patches/ and apply it to the cloned fork
sources via a new apply-patches.sh hook called from the wrapper
Makefile. Simplifies the build flow too: instead of hopping through
llama-cpp's build-llama-cpp-grpc-server indirection, the wrapper now
drives the copied Makefile directly (clone -> patch -> build).
Drop the corresponding patch whenever the fork catches up with
upstream — the build fails fast if a patch stops applying, which
is the signal to retire it.
* docs: add turboquant backend section + clarify cache_type_k/v
Document the new turboquant (llama.cpp fork with TurboQuant KV-cache)
backend alongside the existing llama-cpp / ik-llama-cpp sections in
features/text-generation.md: when to pick it, how to install it from
the gallery, and a YAML example showing backend: turboquant together
with cache_type_k / cache_type_v.
Also expand the cache_type_k / cache_type_v table rows in
advanced/model-configuration.md to spell out the accepted llama.cpp
quantization values and note that these fields apply to all
llama.cpp-family backends, not just vLLM.
* feat(turboquant): patch ggml-rpc GGML_OP_COUNT assertion
The fork adds new GGML ops bringing GGML_OP_COUNT to 97, but
ggml/include/ggml-rpc.h static-asserts it equals 96, breaking
the GGML_RPC=ON build paths (turboquant-grpc / turboquant-rpc-server).
Carry a one-line patch that updates the expected count so the
assertion holds. Drop this patch whenever the fork fixes it upstream.
* feat(turboquant): allow turbo* KV-cache types and exercise them in e2e
The shared backend/cpp/llama-cpp/grpc-server.cpp carries its own
allow-list of accepted KV-cache types (kv_cache_types[]) and rejects
anything outside it before the value reaches llama.cpp's parser. That
list only contains the standard llama.cpp types — turbo2/turbo3/turbo4
would throw "Unsupported cache type" at LoadModel time, meaning
nothing the LocalAI gRPC layer accepted was actually fork-specific.
Add a build-time augmentation step (patch-grpc-server.sh, called from
the turboquant wrapper Makefile) that inserts GGML_TYPE_TURBO2_0/3_0/4_0
into the allow-list of the *copied* grpc-server.cpp under
turboquant-<flavor>-build/. The original file under backend/cpp/llama-cpp/
is never touched, so the stock llama-cpp build keeps compiling against
vanilla upstream which has no notion of those enum values.
Switch test-extra-backend-turboquant to set
BACKEND_TEST_CACHE_TYPE_K=turbo3 / _V=turbo3 so the e2e gRPC suite
actually runs the fork's TurboQuant KV-cache code paths (turbo3 also
auto-enables flash_attention in the fork). Picking q8_0 here would
only re-test the standard llama.cpp path that the upstream llama-cpp
backend already covers.
Refresh the docs (text-generation.md + model-configuration.md) to
list turbo2/turbo3/turbo4 explicitly and call out that you only get
the TurboQuant code path with this backend + a turbo* cache type.
* fix(turboquant): rewrite patch-grpc-server.sh in awk, not python3
The builder image (ubuntu:24.04 stage-2 in Dockerfile.turboquant)
does not install python3, so the python-based augmentation step
errored with `python3: command not found` at make time. Switch to
awk, which ships in coreutils and is already available everywhere
the rest of the wrapper Makefile runs.
* Apply suggestion from @mudler
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
---------
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
* refactor(backends): extract python_utils + add mlx_utils shared helpers
Move parse_options() and messages_to_dicts() out of vllm_utils.py into a
new framework-agnostic python_utils.py, and re-export them from vllm_utils
so existing vllm / vllm-omni imports keep working.
Add mlx_utils.py with split_reasoning() and parse_tool_calls() — ported
from mlx_vlm/server.py's process_tool_calls. These work with any
mlx-lm / mlx-vlm tool module (anything exposing tool_call_start,
tool_call_end, parse_tool_call). Used by the mlx and mlx-vlm backends in
later commits to emit structured ChatDelta.tool_calls without
reimplementing per-model parsing.
Shared smoke tests confirm:
- parse_options round-trips bool/int/float/string
- vllm_utils re-exports are identity-equal to python_utils originals
- mlx_utils parse_tool_calls handles <tool_call>...</tool_call> with a
shim module and produces a correctly-indexed list with JSON arguments
- mlx_utils split_reasoning extracts <think> blocks and leaves clean
content
* feat(mlx): wire native tool parsers + ChatDelta + token usage + logprobs
Bring the MLX backend up to the same structured-output contract as vLLM
and llama.cpp: emit Reply.chat_deltas so the OpenAI HTTP layer sees
tool_calls and reasoning_content, not just raw text.
Key insight: mlx_lm.load() returns a TokenizerWrapper that already auto-
detects the right tool parser from the model's chat template
(_infer_tool_parser in mlx_lm/tokenizer_utils.py). The wrapper exposes
has_tool_calling, has_thinking, tool_parser, tool_call_start,
tool_call_end, think_start, think_end — no user configuration needed,
unlike vLLM.
Changes in backend/python/mlx/backend.py:
- Imports: replace inline parse_options / messages_to_dicts with the
shared helpers from python_utils. Pull split_reasoning / parse_tool_calls
from the new mlx_utils shared module.
- LoadModel: log the auto-detected has_tool_calling / has_thinking /
tool_parser_type for observability. Drop the local is_float / is_int
duplicates.
- _prepare_prompt: run request.Messages through messages_to_dicts so
tool_call_id / tool_calls / reasoning_content survive the conversion,
and pass tools=json.loads(request.Tools) + enable_thinking=True (when
request.Metadata says so) to apply_chat_template. Falls back on
TypeError for tokenizers whose template doesn't accept those kwargs.
- _build_generation_params: return an additional (logits_params,
stop_words) pair. Maps RepetitionPenalty / PresencePenalty /
FrequencyPenalty to mlx_lm.sample_utils.make_logits_processors and
threads StopPrompts through to post-decode truncation.
- New _tool_module_from_tokenizer / _finalize_output / _truncate_at_stop
helpers. _finalize_output runs split_reasoning when has_thinking is
true and parse_tool_calls (using a SimpleNamespace shim around the
wrapper's tool_parser callable) when has_tool_calling is true, then
extracts prompt_tokens, generation_tokens and (best-effort) logprobs
from the last GenerationResponse chunk.
- Predict: use make_logits_processors, accumulate text + last_response,
finalize into a structured Reply carrying chat_deltas,
prompt_tokens, tokens, logprobs. Early-stops on user stop sequences.
- PredictStream: per-chunk Reply still carries raw message bytes for
back-compat but now also emits chat_deltas=[ChatDelta(content=delta)].
On loop exit, emit a terminal Reply with structured
reasoning_content / tool_calls / token counts / logprobs — so the Go
side sees tool calls without needing the regex fallback.
- TokenizeString RPC: uses the TokenizerWrapper's encode(); returns
length + tokens or FAILED_PRECONDITION if the model isn't loaded.
- Free RPC: drops model / tokenizer / lru_cache, runs gc.collect(),
calls mx.metal.clear_cache() when available, and best-effort clears
torch.cuda as a belt-and-suspenders.
* feat(mlx-vlm): mirror MLX parity (tool parsers + ChatDelta + samplers)
Same treatment as the MLX backend: emit structured Reply.chat_deltas,
tool_calls, reasoning_content, token counts and logprobs, and extend
sampling parameter coverage beyond the temp/top_p pair the backend
used to handle.
- Imports: drop the inline is_float/is_int helpers, pull parse_options /
messages_to_dicts from python_utils and split_reasoning /
parse_tool_calls from mlx_utils. Also import make_sampler and
make_logits_processors from mlx_lm.sample_utils — mlx-vlm re-uses them.
- LoadModel: use parse_options; call mlx_vlm.tool_parsers._infer_tool_parser
/ load_tool_module to auto-detect a tool module from the processor's
chat_template. Stash think_start / think_end / has_thinking so later
finalisation can split reasoning blocks without duck-typing on each
call. Logs the detected parser type.
- _prepare_prompt: convert proto Messages via messages_to_dicts (so
tool_call_id / tool_calls survive), pass tools=json.loads(request.Tools)
and enable_thinking=True to apply_chat_template when present, fall
back on TypeError for older mlx-vlm versions. Also handle the
prompt-only + media and empty-prompt + media paths consistently.
- _build_generation_params: return (max_tokens, sampler_params,
logits_params, stop_words). Maps repetition_penalty / presence_penalty /
frequency_penalty and passes them through make_logits_processors.
- _finalize_output / _truncate_at_stop: common helper used by Predict
and PredictStream to split reasoning, run parse_tool_calls against the
auto-detected tool module, build ToolCallDelta list, and extract token
counts + logprobs from the last GenerationResult.
- Predict / PredictStream: switch from mlx_vlm.generate to mlx_vlm.stream_generate
in both paths, accumulate text + last_response, pass sampler and
logits_processors through, emit content-only ChatDelta per streaming
chunk followed by a terminal Reply carrying reasoning_content,
tool_calls, prompt_tokens, tokens and logprobs. Non-streaming Predict
returns the same structured Reply shape.
- New helper _collect_media extracted from the duplicated base64 image /
audio decode loop.
- New TokenizeString RPC using the processor's tokenizer.encode and
Free RPC that drops model/processor/config, runs gc + Metal cache
clear + best-effort torch.cuda cache clear.
* feat(importer/mlx): auto-set tool_parser/reasoning_parser on import
Mirror what core/gallery/importers/vllm.go does: after applying the
shared inference defaults, look up the model URI in parser_defaults.json
and append matching tool_parser:/reasoning_parser: entries to Options.
The MLX backends auto-detect tool parsers from the chat template at
runtime so they don't actually consume these options — but surfacing
them in the generated YAML:
- keeps the import experience consistent with vllm
- gives users a single visible place to override
- documents the intended parser for a given model family
* test(mlx): add helper unit tests + TokenizeString/Free + e2e make targets
- backend/python/mlx/test.py: add TestSharedHelpers with server-less
unit tests for parse_options, messages_to_dicts, split_reasoning and
parse_tool_calls (using a SimpleNamespace shim to fake a tool module
without requiring a model). Plus test_tokenize_string and test_free
RPC tests that load a tiny MLX-quantized Llama and exercise the new
RPCs end-to-end.
- backend/python/mlx-vlm/test.py: same helper unit tests + cleanup of
the duplicated import block at the top of the file.
- Makefile: register BACKEND_MLX and BACKEND_MLX_VLM (they were missing
from the docker-build-target eval list — only mlx-distributed had a
generated target before). Add test-extra-backend-mlx and
test-extra-backend-mlx-vlm convenience targets that build the
respective image and run tests/e2e-backends with the tools capability
against mlx-community/Qwen2.5-0.5B-Instruct-4bit. The MLX backend
auto-detects the tool parser from the chat template so no
BACKEND_TEST_OPTIONS is needed (unlike vllm).
* fix(libbackend): don't pass --copies to venv unless PORTABLE_PYTHON=true
backend/python/common/libbackend.sh:ensureVenv() always invoked
'python -m venv --copies', but macOS system python (and some other
builds) refuses with:
Error: This build of python cannot create venvs without using symlinks
--copies only matters when _makeVenvPortable later relocates the venv,
which only happens when PORTABLE_PYTHON=true. Make --copies conditional
on that flag and fall back to default (symlinked) venv otherwise.
Caught while bringing up the mlx backend on Apple Silicon — the same
build path is used by every Python backend with USE_PIP=true.
* fix(mlx): support mlx-lm 0.29.x tool calling + drop deprecated clear_cache
The released mlx-lm 0.29.x ships a much simpler tool-calling API than
HEAD: TokenizerWrapper detects the <tool_call>...</tool_call> markers
from the tokenizer vocab and exposes has_tool_calling /
tool_call_start / tool_call_end, but does NOT expose a tool_parser
callable on the wrapper and does NOT ship a mlx_lm.tool_parsers
subpackage at all (those only exist on main).
Caught while running the smoke test on Apple Silicon with the
released mlx-lm 0.29.1: tokenizer.tool_parser raised AttributeError
(falling through to the underlying HF tokenizer), so
_tool_module_from_tokenizer always returned None and tool calls slipped
through as raw <tool_call>...</tool_call> text in Reply.message instead
of being parsed into ChatDelta.tool_calls.
Fix: when has_tool_calling is True but tokenizer.tool_parser is missing,
default the parse_tool_call callable to json.loads(body.strip()) — that's
exactly what mlx_lm.tool_parsers.json_tools.parse_tool_call does on HEAD
and covers the only format 0.29 detects (<tool_call>JSON</tool_call>).
Future mlx-lm releases that ship more parsers will be picked up
automatically via the tokenizer.tool_parser attribute when present.
Also tighten the LoadModel logging — the old log line read
init_kwargs.get('tool_parser_type') which doesn't exist on 0.29 and
showed None even when has_tool_calling was True. Log the actual
tool_call_start / tool_call_end markers instead.
While here, switch Free()'s Metal cache clear from the deprecated
mx.metal.clear_cache to mx.clear_cache (mlx >= 0.30), with a
fallback for older releases. Mirrored to the mlx-vlm backend.
* feat(mlx-distributed): mirror MLX parity (tool calls + ChatDelta + sampler)
Same treatment as the mlx and mlx-vlm backends: emit Reply.chat_deltas
with structured tool_calls / reasoning_content / token counts /
logprobs, expand sampling parameter coverage beyond temp+top_p, and
add the missing TokenizeString and Free RPCs.
Notes specific to mlx-distributed:
- Rank 0 is the only rank that owns a sampler — workers participate in
the pipeline-parallel forward pass via mx.distributed and don't
re-implement sampling. So the new logits_params (repetition_penalty,
presence_penalty, frequency_penalty) and stop_words apply on rank 0
only; we don't need to extend coordinator.broadcast_generation_params,
which still ships only max_tokens / temperature / top_p to workers
(everything else is a rank-0 concern).
- Free() now broadcasts CMD_SHUTDOWN to workers when a coordinator is
active, so they release the model on their end too. The constant is
already defined and handled by the existing worker loop in
backend.py:633 (CMD_SHUTDOWN = -1).
- Drop the locally-defined is_float / is_int / parse_options trio in
favor of python_utils.parse_options, re-exported under the module
name for back-compat with anything that imported it directly.
- _prepare_prompt: route through messages_to_dicts so tool_call_id /
tool_calls / reasoning_content survive, pass tools=json.loads(
request.Tools) and enable_thinking=True to apply_chat_template, fall
back on TypeError for templates that don't accept those kwargs.
- New _tool_module_from_tokenizer (with the json.loads fallback for
mlx-lm 0.29.x), _finalize_output, _truncate_at_stop helpers — same
contract as the mlx backend.
- LoadModel logs the auto-detected has_tool_calling / has_thinking /
tool_call_start / tool_call_end so users can see what the wrapper
picked up for the loaded model.
- backend/python/mlx-distributed/test.py: add the same TestSharedHelpers
unit tests (parse_options, messages_to_dicts, split_reasoning,
parse_tool_calls) that exist for mlx and mlx-vlm.
* fix(schema): serialize ToolCallID and Reasoning in Messages.ToProto
The ToProto conversion was dropping tool_call_id and reasoning_content
even though both proto and Go fields existed, breaking multi-turn tool
calling and reasoning passthrough to backends.
* refactor(config): introduce backend hook system and migrate llama-cpp defaults
Adds RegisterBackendHook/runBackendHooks so each backend can register
default-filling functions that run during ModelConfig.SetDefaults().
Migrates the existing GGUF guessing logic into hooks_llamacpp.go,
registered for both 'llama-cpp' and the empty backend (auto-detect).
Removes the old guesser.go shim.
* feat(config): add vLLM parser defaults hook and importer auto-detection
Introduces parser_defaults.json mapping model families to vLLM
tool_parser/reasoning_parser names, with longest-pattern-first matching.
The vllmDefaults hook auto-fills tool_parser and reasoning_parser
options at load time for known families, while the VLLMImporter writes
the same values into generated YAML so users can review and edit them.
Adds tests covering MatchParserDefaults, hook registration via
SetDefaults, and the user-override behavior.
* feat(vllm): wire native tool/reasoning parsers + chat deltas + logprobs
- Use vLLM's ToolParserManager/ReasoningParserManager to extract structured
output (tool calls, reasoning content) instead of reimplementing parsing
- Convert proto Messages to dicts and pass tools to apply_chat_template
- Emit ChatDelta with content/reasoning_content/tool_calls in Reply
- Extract prompt_tokens, completion_tokens, and logprobs from output
- Replace boolean GuidedDecoding with proper GuidedDecodingParams from Grammar
- Add TokenizeString and Free RPC methods
- Fix missing `time` import used by load_video()
* feat(vllm): CPU support + shared utils + vllm-omni feature parity
- Split vllm install per acceleration: move generic `vllm` out of
requirements-after.txt into per-profile after files (cublas12, hipblas,
intel) and add CPU wheel URL for cpu-after.txt
- requirements-cpu.txt now pulls torch==2.7.0+cpu from PyTorch CPU index
- backend/index.yaml: register cpu-vllm / cpu-vllm-development variants
- New backend/python/common/vllm_utils.py: shared parse_options,
messages_to_dicts, setup_parsers helpers (used by both vllm backends)
- vllm-omni: replace hardcoded chat template with tokenizer.apply_chat_template,
wire native parsers via shared utils, emit ChatDelta with token counts,
add TokenizeString and Free RPCs, detect CPU and set VLLM_TARGET_DEVICE
- Add test_cpu_inference.py: standalone script to validate CPU build with
a small model (Qwen2.5-0.5B-Instruct)
* fix(vllm): CPU build compatibility with vllm 0.14.1
Validated end-to-end on CPU with Qwen2.5-0.5B-Instruct (LoadModel, Predict,
TokenizeString, Free all working).
- requirements-cpu-after.txt: pin vllm to 0.14.1+cpu (pre-built wheel from
GitHub releases) for x86_64 and aarch64. vllm 0.14.1 is the newest CPU
wheel whose torch dependency resolves against published PyTorch builds
(torch==2.9.1+cpu). Later vllm CPU wheels currently require
torch==2.10.0+cpu which is only available on the PyTorch test channel
with incompatible torchvision.
- requirements-cpu.txt: bump torch to 2.9.1+cpu, add torchvision/torchaudio
so uv resolves them consistently from the PyTorch CPU index.
- install.sh: add --index-strategy=unsafe-best-match for CPU builds so uv
can mix the PyTorch index and PyPI for transitive deps (matches the
existing intel profile behaviour).
- backend.py LoadModel: vllm >= 0.14 removed AsyncLLMEngine.get_model_config
so the old code path errored out with AttributeError on model load.
Switch to the new get_tokenizer()/tokenizer accessor with a fallback
to building the tokenizer directly from request.Model.
* fix(vllm): tool parser constructor compat + e2e tool calling test
Concrete vLLM tool parsers override the abstract base's __init__ and
drop the tools kwarg (e.g. Hermes2ProToolParser only takes tokenizer).
Instantiating with tools= raised TypeError which was silently caught,
leaving chat_deltas.tool_calls empty.
Retry the constructor without the tools kwarg on TypeError — tools
aren't required by these parsers since extract_tool_calls finds tool
syntax in the raw model output directly.
Validated with Qwen/Qwen2.5-0.5B-Instruct + hermes parser on CPU:
the backend correctly returns ToolCallDelta{name='get_weather',
arguments='{"location": "Paris, France"}'} in ChatDelta.
test_tool_calls.py is a standalone smoke test that spawns the gRPC
backend, sends a chat completion with tools, and asserts the response
contains a structured tool call.
* ci(backend): build cpu-vllm container image
Add the cpu-vllm variant to the backend container build matrix so the
image registered in backend/index.yaml (cpu-vllm / cpu-vllm-development)
is actually produced by CI.
Follows the same pattern as the other CPU python backends
(cpu-diffusers, cpu-chatterbox, etc.) with build-type='' and no CUDA.
backend_pr.yml auto-picks this up via its matrix filter from backend.yml.
* test(e2e-backends): add tools capability + HF model name support
Extends tests/e2e-backends to cover backends that:
- Resolve HuggingFace model ids natively (vllm, vllm-omni) instead of
loading a local file: BACKEND_TEST_MODEL_NAME is passed verbatim as
ModelOptions.Model with no download/ModelFile.
- Parse tool calls into ChatDelta.tool_calls: new "tools" capability
sends a Predict with a get_weather function definition and asserts
the Reply contains a matching ToolCallDelta. Uses UseTokenizerTemplate
with OpenAI-style Messages so the backend can wire tools into the
model's chat template.
- Need backend-specific Options[]: BACKEND_TEST_OPTIONS lets a test set
e.g. "tool_parser:hermes,reasoning_parser:qwen3" at LoadModel time.
Adds make target test-extra-backend-vllm that:
- docker-build-vllm
- loads Qwen/Qwen2.5-0.5B-Instruct
- runs health,load,predict,stream,tools with tool_parser:hermes
Drops backend/python/vllm/test_{cpu_inference,tool_calls}.py — those
standalone scripts were scaffolding used while bringing up the Python
backend; the e2e-backends harness now covers the same ground uniformly
alongside llama-cpp and ik-llama-cpp.
* ci(test-extra): run vllm e2e tests on CPU
Adds tests-vllm-grpc to the test-extra workflow, mirroring the
llama-cpp and ik-llama-cpp gRPC jobs. Triggers when files under
backend/python/vllm/ change (or on run-all), builds the local-ai
vllm container image, and runs the tests/e2e-backends harness with
BACKEND_TEST_MODEL_NAME=Qwen/Qwen2.5-0.5B-Instruct, tool_parser:hermes,
and the tools capability enabled.
Uses ubuntu-latest (no GPU) — vllm runs on CPU via the cpu-vllm
wheel we pinned in requirements-cpu-after.txt. Frees disk space
before the build since the docker image + torch + vllm wheel is
sizeable.
* fix(vllm): build from source on CI to avoid SIGILL on prebuilt wheel
The prebuilt vllm 0.14.1+cpu wheel from GitHub releases is compiled with
SIMD instructions (AVX-512 VNNI/BF16 or AMX-BF16) that not every CPU
supports. GitHub Actions ubuntu-latest runners SIGILL when vllm spawns
the model_executor.models.registry subprocess for introspection, so
LoadModel never reaches the actual inference path.
- install.sh: when FROM_SOURCE=true on a CPU build, temporarily hide
requirements-cpu-after.txt so installRequirements installs the base
deps + torch CPU without pulling the prebuilt wheel, then clone vllm
and compile it with VLLM_TARGET_DEVICE=cpu. The resulting binaries
target the host's actual CPU.
- backend/Dockerfile.python: accept a FROM_SOURCE build-arg and expose
it as an ENV so install.sh sees it during `make`.
- Makefile docker-build-backend: forward FROM_SOURCE as --build-arg
when set, so backends that need source builds can opt in.
- Makefile test-extra-backend-vllm: call docker-build-vllm via a
recursive $(MAKE) invocation so FROM_SOURCE flows through.
- .github/workflows/test-extra.yml: set FROM_SOURCE=true on the
tests-vllm-grpc job. Slower but reliable — the prebuilt wheel only
works on hosts that share the build-time SIMD baseline.
Answers 'did you test locally?': yes, end-to-end on my local machine
with the prebuilt wheel (CPU supports AVX-512 VNNI). The CI runner CPU
gap was not covered locally — this commit plugs that gap.
* ci(vllm): use bigger-runner instead of source build
The prebuilt vllm 0.14.1+cpu wheel requires SIMD instructions (AVX-512
VNNI/BF16) that stock ubuntu-latest GitHub runners don't support —
vllm.model_executor.models.registry SIGILLs on import during LoadModel.
Source compilation works but takes 30-40 minutes per CI run, which is
too slow for an e2e smoke test. Instead, switch tests-vllm-grpc to the
bigger-runner self-hosted label (already used by backend.yml for the
llama-cpp CUDA build) — that hardware has the required SIMD baseline
and the prebuilt wheel runs cleanly.
FROM_SOURCE=true is kept as an opt-in escape hatch:
- install.sh still has the CPU source-build path for hosts that need it
- backend/Dockerfile.python still declares the ARG + ENV
- Makefile docker-build-backend still forwards the build-arg when set
Default CI path uses the fast prebuilt wheel; source build can be
re-enabled by exporting FROM_SOURCE=true in the environment.
* ci(vllm): install make + build deps on bigger-runner
bigger-runner is a bare self-hosted runner used by backend.yml for
docker image builds — it has docker but not the usual ubuntu-latest
toolchain. The make-based test target needs make, build-essential
(cgo in 'go test'), and curl/unzip (the Makefile protoc target
downloads protoc from github releases).
protoc-gen-go and protoc-gen-go-grpc come via 'go install' in the
install-go-tools target, which setup-go makes possible.
* ci(vllm): install libnuma1 + libgomp1 on bigger-runner
The vllm 0.14.1+cpu wheel ships a _C C++ extension that dlopens
libnuma.so.1 at import time. When the runner host doesn't have it,
the extension silently fails to register its torch ops, so
EngineCore crashes on init_device with:
AttributeError: '_OpNamespace' '_C_utils' object has no attribute
'init_cpu_threads_env'
Also add libgomp1 (OpenMP runtime, used by torch CPU kernels) to be
safe on stripped-down runners.
* feat(vllm): bundle libnuma/libgomp via package.sh
The vllm CPU wheel ships a _C extension that dlopens libnuma.so.1 at
import time; torch's CPU kernels in turn use libgomp.so.1 (OpenMP).
Without these on the host, vllm._C silently fails to register its
torch ops and EngineCore crashes with:
AttributeError: '_OpNamespace' '_C_utils' object has no attribute
'init_cpu_threads_env'
Rather than asking every user to install libnuma1/libgomp1 on their
host (or every LocalAI base image to ship them), bundle them into
the backend image itself — same pattern fish-speech and the GPU libs
already use. libbackend.sh adds ${EDIR}/lib to LD_LIBRARY_PATH at
run time so the bundled copies are picked up automatically.
- backend/python/vllm/package.sh (new): copies libnuma.so.1 and
libgomp.so.1 from the builder's multilib paths into ${BACKEND}/lib,
preserving soname symlinks. Runs during Dockerfile.python's
'Run backend-specific packaging' step (which already invokes
package.sh if present).
- backend/Dockerfile.python: install libnuma1 + libgomp1 in the
builder stage so package.sh has something to copy (the Ubuntu
base image otherwise only has libgomp in the gcc dep chain).
- test-extra.yml: drop the workaround that installed these libs on
the runner host — with the backend image self-contained, the
runner no longer needs them, and the test now exercises the
packaging path end-to-end the way a production host would.
* ci(vllm): disable tests-vllm-grpc job (heterogeneous runners)
Both ubuntu-latest and bigger-runner have inconsistent CPU baselines:
some instances support the AVX-512 VNNI/BF16 instructions the prebuilt
vllm 0.14.1+cpu wheel was compiled with, others SIGILL on import of
vllm.model_executor.models.registry. The libnuma packaging fix doesn't
help when the wheel itself can't be loaded.
FROM_SOURCE=true compiles vllm against the actual host CPU and works
everywhere, but takes 30-50 minutes per run — too slow for a smoke
test on every PR.
Comment out the job for now. The test itself is intact and passes
locally; run it via 'make test-extra-backend-vllm' on a host with the
required SIMD baseline. Re-enable when:
- we have a self-hosted runner label with guaranteed AVX-512 VNNI/BF16, or
- vllm publishes a CPU wheel with a wider baseline, or
- we set up a docker layer cache that makes FROM_SOURCE acceptable
The detect-changes vllm output, the test harness changes (tests/
e2e-backends + tools cap), the make target (test-extra-backend-vllm),
the package.sh and the Dockerfile/install.sh plumbing all stay in
place.
* feat: wire min_p
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat: inferencing defaults
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore(refactor): re-use iterative parser
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore: generate automatically inference defaults from unsloth
Instead of trying to re-invent the wheel and maintain here the inference
defaults, prefer to consume unsloth ones, and contribute there as
necessary.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore: apply defaults also to models installed via gallery
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore: be consistent and apply fallback to all endpoint
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat: add fine-tuning endpoint
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(experimental): add fine-tuning endpoint and TRL support
This changeset defines new GRPC signatues for Fine tuning backends, and
add TRL backend as initial fine-tuning engine. This implementation also
supports exporting to GGUF and automatically importing it to LocalAI
after fine-tuning.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* commit TRL backend, stop by killing process
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* move fine-tune to generic features
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* add evals, reorder menu
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Fix tests
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
AIO images are behind, and takes effort to maintain these. Wizard and
installation of models have been semplified massively, so AIO images
lost their purpose.
This allows us to be more laser focused on main images and reliefes
stress from CI.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(realtime): WebRTC support
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* fix(tracing): Show full LLM opts and deltas
Signed-off-by: Richard Palethorpe <io@richiejp.com>
---------
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* feat(mlx-distributed): add new MLX-distributed backend
Add new MLX distributed backend with support for both TCP and RDMA for
model sharding.
This implementation ties in the discovery implementation already in
place, and re-uses the same P2P mechanism for the TCP MLX-distributed
inferencing.
The Auto-parallel implementation is inspired by Exo's
ones (who have been added to acknowledgement for the great work!)
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* expose a CLI to facilitate backend starting
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat: make manual rank0 configurable via model configs
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Add missing features from mlx backend
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Apply suggestion from @mudler
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
* feat(backends): add faster-qwen3-tts
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix: this backend is CUDA only
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix: add requirements-install.txt with setuptools for build isolation
The faster-qwen3-tts backend requires setuptools to build packages
like sox that have setuptools as a build dependency. This ensures
the build completes successfully in CI.
Signed-off-by: LocalAI Bot <localai-bot@users.noreply.github.com>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Signed-off-by: LocalAI Bot <localai-bot@users.noreply.github.com>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>
* feat(musicgen): add ace-step and UI interface
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Correctly handle model dir
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Drop auto-download
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Fixups
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Add to models, fixup UIs icons
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fixups
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Update docs
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* l4t13 is incompatbile
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* avoid pinning version for cuda12
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Drop l4t12
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(metal): try to extend support to remaining backends
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* neutts doesn't work
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* split outetts out of transformers
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Remove torch pin to whisperx
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(proto): add speaker field to TranscriptSegment for diarization
Add speaker field to the gRPC TranscriptSegment message and map it
through the Go schema, enabling backends to return speaker labels.
Signed-off-by: eureka928 <meobius123@gmail.com>
* feat(whisperx): add whisperx backend for transcription with diarization
Add Python gRPC backend using WhisperX for speech-to-text with
word-level timestamps, forced alignment, and speaker diarization
via pyannote-audio when HF_TOKEN is provided.
Signed-off-by: eureka928 <meobius123@gmail.com>
* feat(whisperx): register whisperx backend in Makefile
Signed-off-by: eureka928 <meobius123@gmail.com>
* feat(whisperx): add whisperx meta and image entries to index.yaml
Signed-off-by: eureka928 <meobius123@gmail.com>
* ci(whisperx): add build matrix entries for CPU, CUDA 12/13, and ROCm
Signed-off-by: eureka928 <meobius123@gmail.com>
* fix(whisperx): unpin torch versions and use CPU index for cpu requirements
Address review feedback:
- Use --extra-index-url for CPU torch wheels to reduce size
- Remove torch version pins, let uv resolve compatible versions
Signed-off-by: eureka928 <meobius123@gmail.com>
* fix(whisperx): pin torch ROCm variant to fix CI build failure
Signed-off-by: eureka928 <meobius123@gmail.com>
* fix(whisperx): pin torch CPU variant to fix uv resolution failure
Pin torch==2.8.0+cpu so uv resolves the CPU wheel from the extra
index instead of picking torch==2.8.0+cu128 from PyPI, which pulls
unresolvable CUDA dependencies.
Signed-off-by: eureka928 <meobius123@gmail.com>
* fix(whisperx): use unsafe-best-match index strategy to fix uv resolution failure
uv's default first-match strategy finds torch on PyPI before checking
the extra index, causing it to pick torch==2.8.0+cu128 instead of the
CPU variant. This makes whisperx's transitive torch dependency
unresolvable. Using unsafe-best-match lets uv consider all indexes.
Signed-off-by: eureka928 <meobius123@gmail.com>
* fix(whisperx): drop +cpu local version suffix to fix uv resolution failure
PEP 440 ==2.8.0 matches 2.8.0+cpu from the extra index, avoiding the
issue where uv cannot locate an explicit +cpu local version specifier.
This aligns with the pattern used by all other CPU backends.
Signed-off-by: eureka928 <meobius123@gmail.com>
* fix(backends): drop +rocm local version suffixes from hipblas requirements to fix uv resolution
uv cannot resolve PEP 440 local version specifiers (e.g. +rocm6.4,
+rocm6.3) in pinned requirements. The --extra-index-url already points
to the correct ROCm wheel index and --index-strategy unsafe-best-match
(set in libbackend.sh) ensures the ROCm variant is preferred.
Applies the same fix as 7f5d72e8 (which resolved this for +cpu) across
all 14 hipblas requirements files.
Signed-off-by: eureka928 <meobius123@gmail.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Signed-off-by: eureka928 <meobius123@gmail.com>
* revert: scope hipblas suffix fix to whisperx only
Reverts changes to non-whisperx hipblas requirements files per
maintainer review — other backends are building fine with the +rocm
local version suffix.
Signed-off-by: eureka928 <meobius123@gmail.com>
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Signed-off-by: eureka928 <meobius123@gmail.com>
---------
Signed-off-by: eureka928 <meobius123@gmail.com>
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Some datacenter setups might be stuck with the 5.x kernel which doesn't
play well with CUDA >=12.9. To incrase compatibility with the CUDA 12.x
branch, downgrade to 12.8. For newer systems, it is still suggested to
use CUDA 13.x wherever compatible.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(vllm-omni: add new backend
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* default to py3.12
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
exllama2 development has stalled and only old architectures are
supported. exllamav3 is still in development, meanwhile cleaning up
exllama2 from the gallery.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(backends): add moonshine backend for faster transcription
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Add backend to CI, update AGENTS.md from this exercise
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore(makefile): Add buildargs for sd and cuda when building backend
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* feat(whisper): Add prompt to condition transcription output
Signed-off-by: Richard Palethorpe <io@richiejp.com>
---------
Signed-off-by: Richard Palethorpe <io@richiejp.com>
* chore(deps): bump stable-diffusion.cpp to '8823dc48bcc1598eb9671da7b69e45338d0cc5a5'
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix(Dockerfile.golang): Make curl noisy to see when download fails
Signed-off-by: Richard Palethorpe <io@richiejp.com>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Signed-off-by: Richard Palethorpe <io@richiejp.com>
Co-authored-by: Richard Palethorpe <io@richiejp.com>
* feat(vibevoice): add backend
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore: add workflow and backend index
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore(gallery): add vibevoice
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Use self-hosted for intel builds
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Pin python version for l4t
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore(ci): add cuda13 jobs
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Add to pipelines and to capabilities. Start to work on the gallery
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* gallery
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* capabilities: try to detect by looking at /usr/local
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* neutts
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* backends.yaml
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* add cuda13 l4t requirements.txt
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* add cuda13 requirements.txt
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Fixups
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Fixups
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Pin vllm
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Not all backends are compatible
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* add vllm to requirements
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* vllm is not pre-compiled for cuda 13
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore(tests): check embeddings differs in llama.cpp
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix(llama.cpp): use the correct field for embedding
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* fix(llama.cpp): use embedding type none
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* chore(tests): add test-cases in aio-e2e suite
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
