English | 中文

A C# inference engine for running large language models (LLMs) locally using GGUF model files. TensorSharp provides a console application, a web-based chatbot interface, and Ollama/OpenAI-compatible HTTP APIs for programmatic access.

• Multi-architecture support -- Gemma 4, Gemma 3, Qwen 3, Qwen 3.5
• Multimodal inference -- image, video, and audio inputs (Gemma 4); images for Gemma 3 / Qwen 3.5
• Thinking / reasoning mode -- structured chain-of-thought output with <think> / <|channel>thought tags (Qwen 3, Qwen 3.5, Gemma 4)
• Tool calling / function calling -- models can invoke user-defined tools; multi-turn tool-call conversations supported across all three API styles
• Quantized model support -- loads GGUF files with Q4_K_M, Q8_0, F16, and other quantization formats; performs native quantized matmul without dequantizing to FP32
• GPU-accelerated -- Apple Metal backend via GGML with fused whole-model GPU dispatch for Gemma 4 decode (~2.6x speedup over per-op dispatch)
• Ollama & OpenAI API compatibility -- drop-in replacement endpoints for existing tooling
• Configurable sampling -- temperature, top-k, top-p, min-p, repetition/presence/frequency penalties, seed, stop sequences
• Chat templates -- auto-loaded from GGUF metadata (Jinja2), with hardcoded fallbacks per architecture
• Request queue -- FIFO inference queue ensures single-request execution for KV cache stability, with real-time position tracking for clients
• Batch processing -- JSONL input support in the console application
• Streaming -- token-by-token output via SSE (web) or stdout (console)
• Mixture of Experts -- Gemma 4 MoE variants (e.g. gemma-4-26B-A4B)
• Large file uploads -- supports video/audio uploads up to 500 MB in the web interface

TensorSharp/
├── TensorSharp/                 # Core tensor library (CPU operations, SIMD)
├── TensorSharp.GGML/            # GGML backend bindings (Metal/CPU via native library)
├── TensorSharp.GGML.Native/     # Native C++ bridge to ggml (builds libGgmlOps)
├── AdvUtils/                    # Utility library
├── InferenceEngine/             # Model loading, tokenization, and inference logic
│   ├── Models/
│   │   ├── Gemma3/
│   │   ├── Gemma4/              # Vision encoder, audio encoder, MoE, fused GPU decode
│   │   ├── Qwen3/
│   │   └── Qwen35/
│   ├── GgufReader.cs            # GGUF file parser
│   ├── ModelBase.cs             # Base class for all model architectures
│   ├── ChatTemplate.cs          # Chat template rendering (hardcoded + Jinja2 from GGUF)
│   ├── Jinja2Template.cs        # Jinja2 template renderer
│   ├── OutputParser.cs          # Extracts thinking, content, and tool calls from model output
│   ├── SamplingConfig.cs        # Sampling parameter configuration
│   ├── TokenSampler.cs          # Token sampling (greedy, top-k, top-p, min-p, penalties)
│   └── MediaHelper.cs           # Video frame extraction, audio decoding
├── InferenceConsole/            # CLI application
├── InferenceWeb/                # Web chatbot + API server (ASP.NET Core)
│   ├── ModelService.cs          # Model lifecycle management
│   ├── InferenceQueue.cs        # FIFO request queue with position tracking
│   ├── wwwroot/index.html       # Chat UI
│   ├── testdata/                # Integration test suites (bash + Python)
│   └── API_EXAMPLES.md          # Detailed API documentation
└── ExternalProjects/            # Third-party dependencies (ggml)

• .NET 8.0 SDK or later (.NET 10 for InferenceWeb)
• macOS (Metal backend): CMake 3.20+ and Xcode command-line tools for building the native GGML library
• GGUF model files (e.g., from Hugging Face)

dotnet build TensorSharp.slnx

# Console application
dotnet build InferenceConsole/InferenceConsole.csproj

# Web application
dotnet build InferenceWeb/InferenceWeb.csproj

The native library is built automatically during the first dotnet build if it doesn't exist. To build it manually:

cd TensorSharp.GGML.Native

macOS:

bash build-macos.sh

Linux:

bash build-linux.sh

On macOS this compiles libGgmlOps.dylib with Metal GPU support. On Linux it compiles libGgmlOps.so with the GGML CPU backend. The build output is automatically copied to the application's output directory.

cd InferenceConsole/bin

# Text inference
./InferenceConsole --model <model.gguf> --input prompt.txt --output result.txt \
    --max-tokens 200 --backend ggml_metal

# Image inference (Gemma 3/4, Qwen 3.5)
./InferenceConsole --model <model.gguf> --image photo.png --backend ggml_metal

# Video inference (Gemma 4)
./InferenceConsole --model <model.gguf> --video clip.mp4 --backend ggml_metal

# Audio inference (Gemma 4)
./InferenceConsole --model <model.gguf> --audio speech.wav --backend ggml_metal

# Thinking / reasoning mode
./InferenceConsole --model <model.gguf> --input prompt.txt --backend ggml_metal --think

# Tool calling
./InferenceConsole --model <model.gguf> --input prompt.txt --backend ggml_metal \
    --tools tools.json

# With sampling parameters
./InferenceConsole --model <model.gguf> --input prompt.txt --backend ggml_metal \
    --temperature 0.7 --top-p 0.9 --top-k 40 --repeat-penalty 1.2 --seed 42

# Batch processing (JSONL)
./InferenceConsole --model <model.gguf> --input-jsonl requests.jsonl \
    --output results.txt --backend ggml_metal

Command-line options:

The multimodal projector file is auto-detected if placed alongside the model file with a recognized name (e.g., gemma-4-mmproj-F16.gguf).

JSONL input format:

Each line is a JSON object with messages, optional prompt, and optional sampling parameters:

{"id": "q1", "messages": [{"role": "user", "content": "What is 2+3?"}], "max_tokens": 50}
{"id": "q2", "messages": [{"role": "user", "content": "Write a haiku."}], "max_tokens": 100, "temperature": 0.8}

cd InferenceWeb/bin

# Set environment variables and run
MODEL_DIR=./models BACKEND=ggml_metal ./InferenceWeb

Open http://localhost:5000 in your browser. The web interface supports:

• Multi-turn chat conversations
• Model selection from available GGUF files in MODEL_DIR
• Image, video, and audio uploads for multimodal inference (up to 500 MB)
• Thinking/reasoning mode toggle
• Tool calling with function definitions
• Streaming token generation via Server-Sent Events
• Request queue with real-time position feedback

Environment variables:

InferenceWeb exposes three API styles. See API_EXAMPLES.md for full documentation with curl and Python examples.

Ollama-compatible API:

# List models
curl http://localhost:5000/api/tags

# Generate text
curl -X POST http://localhost:5000/api/generate \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3-4B-Q8_0.gguf", "prompt": "Hello!", "stream": false}'

# Chat
curl -X POST http://localhost:5000/api/chat/ollama \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3-4B-Q8_0.gguf", "messages": [{"role": "user", "content": "Hi"}], "stream": false}'

# Chat with thinking mode
curl -X POST http://localhost:5000/api/chat/ollama \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3-4B-Q8_0.gguf", "messages": [{"role": "user", "content": "Solve 17*23"}], "think": true, "stream": false}'

# Chat with tool calling
curl -X POST http://localhost:5000/api/chat/ollama \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3-4B-Q8_0.gguf", "messages": [{"role": "user", "content": "What is the weather?"}], "tools": [{"function": {"name": "get_weather", "description": "Get current weather", "parameters": {"properties": {"city": {"type": "string"}}, "required": ["city"]}}}], "stream": false}'

OpenAI-compatible API:

# Chat completions
curl -X POST http://localhost:5000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3-4B-Q8_0.gguf", "messages": [{"role": "user", "content": "Hi"}], "max_tokens": 50}'

OpenAI Python SDK:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:5000/v1", api_key="not-needed")
response = client.chat.completions.create(
    model="Qwen3-4B-Q8_0.gguf",
    messages=[{"role": "user", "content": "What is 2+3?"}],
    max_tokens=50
)
print(response.choices[0].message.content)

Queue status:

curl http://localhost:5000/api/queue/status
# {"busy":false,"pending_requests":0,"total_processed":42}

Models that support thinking mode (Qwen 3, Qwen 3.5, Gemma 4) can produce structured chain-of-thought reasoning before generating the final answer. The thinking content is separated from the main response and can be displayed or hidden by the client.

• Qwen 3 / Qwen 3.5: uses <think>...</think> tags
• Gemma 4: uses <|channel>thought\n...<channel|> tags

Enable via --think (console), "think": true (Ollama API), or the thinking toggle in the web UI.

Models can invoke user-defined tools and participate in multi-turn tool-call conversations. Define tools as JSON and pass them via --tools (console) or the tools parameter in the API.

Each architecture uses its own wire format for tool calls:

• Qwen 3 / Qwen 3.5: <tool_call>{"name": "...", "arguments": {...}}</tool_call>
• Gemma 4: <|tool_call>call:function_name{args}<tool_call|>

The output parser (OutputParser.cs) automatically extracts tool calls from the model's raw output regardless of architecture.

Gemma 4 models support image, video, and audio inputs. Place the multimodal projector (gemma-4-mmproj-F16.gguf) in the same directory as the model file for automatic loading.

• Images: PNG, JPEG
• Video: MP4 (extracts up to 8 frames at 1 fps using OpenCV)
• Audio: WAV (16kHz mono), MP3, OGG Vorbis

These models support image inputs with their respective multimodal projector files.

TensorSharp is structured as a layered system:

1. TensorSharp provides the core Tensor type, storage abstraction, and an extensible operation registry (Ops). CPU implementations use System.Numerics.Vectors for SIMD acceleration.

2. TensorSharp.GGML registers GPU-accelerated implementations of the same operations via a native C++ bridge (libGgmlOps) that links against ggml. On macOS, this provides Metal GPU compute. Operations include native quantized matmul (Q4_K_M, Q8_0, etc.) without dequantizing to FP32.

3. InferenceEngine implements model-specific logic: GGUF parsing, tokenization (SentencePiece BPE), chat template rendering (Jinja2 from GGUF metadata with hardcoded fallbacks), configurable token sampling, output parsing (thinking extraction, tool-call extraction), and the forward pass for each architecture. Models are loaded via ModelBase.Create() which auto-detects the architecture from GGUF metadata.

4. InferenceConsole and InferenceWeb are application layers that handle I/O and user interaction. InferenceWeb provides Ollama-compatible and OpenAI-compatible REST APIs alongside a browser-based chat UI, with a FIFO inference queue to serialize concurrent requests.

• Fused GPU decode (Gemma 4): all transformer layers are executed in a single GGML compute graph dispatch on Metal, reducing CPU-GPU round-trips from hundreds per token to one. This achieves ~2.6x speedup over per-operation dispatch.
• Fused weight projections: Q/K/V projections are fused into a single QKV matmul; gate and up projections are fused into a single gate_up matmul.
• Native quantized compute: quantized weights (Q4_K_M, Q6_K, Q8_0, etc.) are used directly in matmul without expanding to FP32, saving memory and bandwidth.
• Circular KV cache: sliding-window attention layers use a fixed-size circular buffer, bounding memory usage regardless of sequence length.
• Memory-efficient model loading: large tensors are streamed directly to native memory without intermediate managed allocations.

Integration tests for InferenceWeb are in InferenceWeb/testdata/. They cover all three API styles (Web UI SSE, Ollama, OpenAI), multi-turn conversations, thinking mode, tool calling, queue behavior, concurrent requests, and abort support.

# Start InferenceWeb, then run:
python3 InferenceWeb/testdata/test_multiturn.py
# or
bash InferenceWeb/testdata/test_multiturn.sh

See InferenceWeb/testdata/README.md for the full test matrix.

Zhongkai Fu

See LICENSE for details.