mirror of
https://github.com/mudler/LocalAI.git
synced 2026-04-16 12:59:33 -04:00
* 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
87 lines
2.9 KiB
Python
87 lines
2.9 KiB
Python
"""Llama 3.1 / 3.2 / 3.3 JSON tool-call parser.
|
|
|
|
Meta's Llama 3.1+ instruct chat templates emit tool calls in two broadly
|
|
compatible shapes:
|
|
|
|
1. With the `<|python_tag|>` lead-in:
|
|
<|python_tag|>{"name": "get_weather", "parameters": {"city": "Paris"}}
|
|
2. As a bare JSON object (or list of objects) at the end of the turn.
|
|
|
|
We also handle multi-call shapes where the model emits several JSON objects
|
|
separated by `;` or newlines, and JSON arrays `[{...}, {...}]`. The key field
|
|
for Llama 3 is historically `parameters` (older docs) but recent checkpoints
|
|
also emit `arguments` — accept either.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import re
|
|
from dataclasses import dataclass
|
|
|
|
from .base import ToolCall, ToolParser, register
|
|
|
|
_PYTHON_TAG = "<|python_tag|>"
|
|
_JSON_OBJECT_RE = re.compile(r"\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}", re.DOTALL)
|
|
|
|
|
|
def _coerce_call(obj: object, index: int) -> ToolCall | None:
|
|
if not isinstance(obj, dict):
|
|
return None
|
|
name = obj.get("name")
|
|
if not isinstance(name, str):
|
|
return None
|
|
args = obj.get("arguments", obj.get("parameters", {}))
|
|
args_str = args if isinstance(args, str) else json.dumps(args, ensure_ascii=False)
|
|
return ToolCall(index=index, name=name, arguments=args_str)
|
|
|
|
|
|
@register
|
|
class Llama3JsonToolParser(ToolParser):
|
|
name = "llama3_json"
|
|
|
|
def parse(self, text: str) -> tuple[str, list[ToolCall]]:
|
|
calls: list[ToolCall] = []
|
|
|
|
# Strip <|python_tag|> segments first — each segment is one tool call
|
|
# body. The content after the final python_tag (if any) is the call.
|
|
remaining = text
|
|
if _PYTHON_TAG in text:
|
|
head, *tails = text.split(_PYTHON_TAG)
|
|
remaining = head
|
|
for tail in tails:
|
|
parsed = _try_parse(tail.strip(), len(calls))
|
|
calls.extend(parsed)
|
|
|
|
# Any JSON objects / arrays left in `remaining` count as tool calls too
|
|
# if they parse to a {"name": ..., "arguments": ...} shape.
|
|
for match in _JSON_OBJECT_RE.finditer(remaining):
|
|
parsed = _try_parse(match.group(0), len(calls))
|
|
if parsed:
|
|
calls.extend(parsed)
|
|
remaining = remaining.replace(match.group(0), "", 1)
|
|
|
|
content = remaining.strip()
|
|
return content, calls
|
|
|
|
|
|
def _try_parse(blob: str, start_index: int) -> list[ToolCall]:
|
|
"""Parse a fragment that may be a JSON object or a JSON array of objects."""
|
|
blob = blob.strip().rstrip(";")
|
|
if not blob:
|
|
return []
|
|
try:
|
|
obj = json.loads(blob)
|
|
except json.JSONDecodeError:
|
|
return []
|
|
if isinstance(obj, dict):
|
|
call = _coerce_call(obj, start_index)
|
|
return [call] if call else []
|
|
if isinstance(obj, list):
|
|
calls: list[ToolCall] = []
|
|
for i, item in enumerate(obj):
|
|
c = _coerce_call(item, start_index + i)
|
|
if c:
|
|
calls.append(c)
|
|
return calls
|
|
return []
|