A Rust-native LLM inference engine. Load models from Hugging Face, chat locally or serve via OpenAI-compatible API. Single binary, no Python, no runtime dependencies.

中文说明

# From crates.io
cargo install ferrum-cli

# Or build from source
cargo build --release -p ferrum-cli --bin ferrum

# CUDA (Linux + NVIDIA)
CUDA_HOME=/usr/local/cuda cargo build --release --features cuda -p ferrum-cli --bin ferrum

For gated models (e.g. Llama 3.2), set your Hugging Face token first:

export HF_TOKEN=hf_your_token_here

# Download a model
ferrum pull qwen3:0.6b

# Chat
ferrum run qwen3:0.6b

# Or start an API server
ferrum serve --model qwen3:0.6b --port 8000

Any Hugging Face model using a supported architecture works out of the box:

# Text generation
ferrum run Qwen/Qwen3-4B
ferrum run llama3.2:3b

# Speech-to-Text (supports WAV/M4A/MP3/FLAC — auto ffmpeg conversion)
ferrum transcribe whisper-turbo recording.m4a -l zh
ferrum transcribe whisper-turbo meeting.wav -l en

# Text-to-Speech
ferrum tts qwen3-tts "Hello, welcome to Ferrum TTS" -o output.wav
ferrum tts qwen3-tts "你好欢迎使用语音合成系统" -o output.wav

# Voice clone (ICL mode — clone any voice from 5s reference audio)
ferrum tts qwen3-tts "你好" --ref-audio ref.wav --ref-text "参考文本" -o clone.wav

# Streaming TTS (first audio chunk in ~2.5s)
ferrum tts qwen3-tts "你好世界" --streaming -o output.wav

# TTS API server (OpenAI-compatible)
ferrum serve qwen3-tts
curl localhost:8000/v1/audio/speech -d '{"input":"你好","language":"chinese"}' -o speech.wav

# Whisper API server (OpenAI-compatible)
ferrum serve whisper-turbo
curl localhost:8000/v1/audio/transcriptions -F "[email protected]" -F "language=zh"

# Embeddings (text + image)
ferrum embed OFA-Sys/chinese-clip-vit-base-patch16 --text "sunset at the beach"
ferrum embed google/siglip-base-patch16-224 --image photo.jpg

# Embedding API server
ferrum serve --model OFA-Sys/chinese-clip-vit-base-patch16
curl localhost:8000/v1/embeddings -d '{"model":"clip","input":"hello"}'
curl localhost:8000/v1/embeddings -d '{"model":"clip","input":{"image":"/path/to/photo.jpg"}}'

# Chat completions (OpenAI-compatible)
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3:0.6b","messages":[{"role":"user","content":"Hello"}]}'

# Audio transcription (OpenAI-compatible, multipart form)
curl http://localhost:8000/v1/audio/transcriptions \
  -F "[email protected]" -F "language=zh"

# Text-to-speech (OpenAI-compatible)
curl http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input":"Hello world","language":"english"}' -o speech.wav

# Embeddings
curl http://localhost:8000/v1/embeddings \
  -d '{"model":"clip","input":"hello"}'

# List models
curl http://localhost:8000/v1/models

# Health check
curl http://localhost:8000/health

Benchmarked on RTX PRO 6000 (Blackwell):

CUDA graph replay is automatic after 3-step warmup; eliminates per-step launch overhead and sits on the Blackwell + CUDA 13 path we hardened in this cycle (see docs/phase-e-cuda-status.md).

TP decode uses persistent per-rank threads with NCCL all-reduce. Current bottleneck is PCIe interconnect latency (~0.44ms × 72 NCCL calls/step). TP is most beneficial for models that don't fit on a single GPU, or with NVLink interconnect.

Custom Whisper forward pass with rustfft STFT. Full decode pipeline: timestamp-based sequential decode, temperature fallback, compression ratio check. Mel precision matches Python whisper exactly.

All-Metal fused transformer pipeline: custom GEMM (64×32 simdgroup tiles), fused residual+norm, flash attention with layer_scale. Full Mimi-based vocoder with 8-layer pre-transformer. Zero-copy on Apple Silicon unified memory.

• Custom CUDA decode runner: bypasses candle for the decode hot path (Qwen3 + LLaMA)
• INT4 quantization: GPTQ models auto-detected, Marlin fused INT4×FP16 kernel
• Tensor parallelism: persistent per-rank threads, barrier sync, NCCL all-reduce (Megatron-LM pattern)
• Batched attention kernel: single launch for all batch items (SM utilization 17%→67%)
• Batched RoPE: per-item positions in single kernel launch
• Custom CUDA kernels: fused RmsNorm, SiLU×mul, RoPE, decode attention (all on single stream)
• Flash Decoding: split-K for long-context decode (auto at KV > 256)
• Batch decode: batched cuBLAS GEMM + batched attention for concurrent requests
• Metal TTS pipeline: all-Metal fused transformer for talker (28 layers) + SubTalker (5 layers) + vocoder (8 layers), GPU-side RMSNorm, cached projection weights
• TTS streaming: chunk-by-chunk audio generation (~800ms chunks), first audio in ~2.5s
• TTS voice clone: ICL prompting with speaker encoder (ECAPA-TDNN) + speech tokenizer (Mimi RVQ), sinc resampling
• TTS HTTP API: OpenAI-compatible /v1/audio/speech with streaming support
• Paged KV attention: GPU block pool with block-table indirection
• Double-buffered residual: cross-layer norm fusion (-108 kernel launches)

What works:

• CLI chat, HTTP serving with streaming, benchmarking
• Qwen3, Qwen2/2.5, LLaMA 3.x, TinyLlama architectures
• Custom CUDA decode runner for Qwen3 and LLaMA (2x speedup)
• Metal GPU acceleration (macOS), CUDA (NVIDIA), CPU
• INT4 GPTQ quantization with Marlin fused kernel (Blackwell compatible)
• FlashAttention-2 prefill + custom CUDA decode runner
• Paged KV cache with block reclamation
• Continuous batching with batch decode
• Tensor parallelism (multi-GPU NCCL, auto-detects GPU count)
• CLIP/Chinese-CLIP/SigLIP embeddings (text + image, /v1/embeddings API)
• Whisper ASR (speech-to-text, Metal accelerated, /v1/audio/transcriptions API)
• Multi-format audio support (WAV/M4A/MP3/FLAC via ffmpeg)
• Top-k/top-p/temperature/repetition-penalty sampling

• Speculative decoding — draft model verification
• More model architectures — Mistral, Phi, DeepSeek, etc.
• Qwen2 CUDA runner — same pattern as LLaMA

See docs/ROADMAP.md for full details.

# CPU only (default)
cargo install ferrum-cli

# With Metal acceleration (macOS)
cargo install ferrum-cli --features metal

# With CUDA acceleration (NVIDIA, requires CUDA toolkit + nvcc)
cargo install ferrum-cli --features cuda

Or build from source:

cargo build --release -p ferrum-cli                    # CPU
cargo build --release -p ferrum-cli --features metal   # Metal (macOS)
cargo build --release -p ferrum-cli --features cuda    # CUDA (NVIDIA)
cargo build --release -p ferrum-cli --features cuda    # Multi-GPU auto-detected when available

Prerequisites: Rust stable toolchain.

crates/
├── ferrum-types          # Shared type definitions
├── ferrum-interfaces     # Core trait contracts (ComputeBackend, KernelOps, ModelExecutor)
├── ferrum-runtime        # Backend implementations (Candle, CPU)
├── ferrum-engine         # Metal kernels, model orchestration
├── ferrum-models         # Model architectures (LLaMA, Qwen2, Qwen3, BERT, Whisper)
├── ferrum-kernels   # Custom CUDA kernels + decode runner
├── ferrum-tokenizer      # Tokenization
├── ferrum-sampler        # Sampling strategies
├── ferrum-scheduler      # Request scheduling
├── ferrum-kv             # KV cache management
├── ferrum-server         # HTTP API server
├── ferrum-cli            # CLI binary
└── ferrum-testkit        # Testing utilities

MIT