Snapdragon 노트북 GPU에서 Llama를 실행하기 위해 순수 WGSL LLM 엔진을 구축했습니다.

A minimalist Llama inference engine written in Rust and WGSL. Runs locally on any GPU — no CUDA required. A from-scratch LLM inference engine that uses wgpu v29 to dispatch WGSL compute shaders for the full Transformer forward pass. No CUDA. No Python. No framework dependencies. Just Rust, raw shaders, and your GPU. Target architecture: Llama (TinyLlama, Llama 2/3, and compatible fine-tunes). Tested with: TinyLlama-1.1B-Chat-v1.0 | GPU | Mode | VRAM (weights) | tok/s | |---|---|---|---| | RTX 3090 (Vulkan) | f16-weights | ~2.05 GiB | 66+ | | Adreno X1-85 (Vulkan) | INT8-b64 | 1.27 GiB | 32.8 | | Adreno X1-85 (Vulkan) | f16-weights | 2.05 GiB | 25.5 | - The Copilot+ Hardware Gap — I recently got a Snapdragon X Elite Copilot+ laptop and quickly realized my GPU was effectively a paperweight for local AI. Standard tools like LM Studio and the massive PyTorch ecosystem didn't support the integrated Adreno GPU, forcing everything onto the CPU. I didn't want to wait for the ecosystem to catch up, so I bypassed it entirely. - Portability — runs on Windows, macOS, Linux, and anything with a Vulkan/Metal/DX12 driver. - Educational — every layer of the Transformer is visible as a standalone WGSL shader. No framework magic. - Hackable — small codebase, clear data flow, easy to experiment with. wgpu-llm is an experiment in highly leveraged, AI-accelerated engineering. Starting from a detailed, human-designed architectural blueprint, the engine was scaffolded into working code in under 16 hours using a custom LLM orchestration pipeline with a human-in-the-loop approach. To ensure the AI could write complex, low-level code without hallucinating or losing track of the architecture, the orchestrator was augmented with my own open-source tools: - MCP_WGSL_Docs: Fed the LLM the exact, up-to-date WGSL specification so it could accurately implement features like tiled GEMM and compile-time f32/f16 switching across 12 standalone compute shaders. - mcp-local-memory: Maintained persistent, strict architectural context across the entire 16-hour development sprint, ensuring the Rust host code and WGSL shaders stayed perfectly aligned. Prerequisites: Rust and Cargo + a GPU with Vulkan, Metal, or DX12 support. cargo install wgpu-llm git clone https://github.com/Beledarian/wgpu-llm.git cd wgpu-llm cargo build --release Download a compatible model (e.g., TinyLlama-1.1B-Chat) — the model directory needs config.json , tokenizer.json , and *.safetensors . # If installed via cargo install: wgpu-llm --model-dir /path/to/model --prompt "The overarching philosophy of stoicism teaches us" --max-tokens 256 --f16-weights # If built from source: cargo run --release --bin wgpu-llm -- --model-dir /path/to/model --prompt "The overarching philosophy of stoicism teaches us" --max-tokens 256 --f16-weights The engine auto-detects your GPU capabilities, loads weights, and streams generated text to stdout. After generation it prints a telemetry summary (VRAM breakdown, tok/s, timing). INT8 mode (~2× VRAM reduction): # 1. Quantize your model (one-time step, requires Python + safetensors) python scripts/quantize_int8.py /path/to/model /path/to/model-int8 --block-size 64 # 2. Run with quantized weights wgpu-llm --model-dir /path/to/model-int8 --prompt "..." --max-tokens 256 --gemm-int8-block-size 64 For the full list of CLI flags (sampling, INT8 GEMM, KV spill, memory budget, dry-run), see the CLI Reference. - Alpha quality — functional but not production-hardened - Decode-only — prompt tokens are processed sequentially (no batched prefill yet) - Single-sequence — no batching or concurrent requests - INT8 lightly tested — verified with TinyLlama 1.1B; quality may vary with larger models - CPU-side sampling — logits read back to CPU for top-k/top-p/temperature; fast enough for single-sequence but not optimal - No chat template — raw text completion only; no multi-turn conversation support yet - Timing is wall-clock — reported tok/s includes CPU overhead, not isolated GPU kernel time Text → Tokenizer → Embedding → [N × Transformer Layers] → Logits → Sampler → Token │ ┌────┴────┐ │ Attn │ ← KV Cache (paged) │ FFN │ └─────────┘ Each box is one or more WGSL compute shader dispatches orchestrated by Rust. See docs/architecture.md for the full system design. | Decision | Detail | |---|---| | 12 WGSL compute shaders | Every Transformer op is a standalone shader (GEMM, MATVEC, RMSNorm, RoPE, SiLU, softmax, etc.) | | f32/f16 compile-time switching | WGSL string injection — no runtime branching | | Paged KV Cache | Lazy GPU page allocation with per-page sequence-offset writes | | Single CommandEncoder | All dispatches per token in one GPU submission — no mid-pipeline sync | | CPU-side sampling | Temperature → top-k → top-p → softmax → weighted sample; GPU argmax at temp=0 | | Row-sharding | Large embedding/LM-head tensors auto-split to fit max_storage_buffer_binding_size | | GQA support | Grouped-Query Attention for Llama 3 models | - Batched prompt prefill — process all prompt tokens in one dispatch - GPU timestamp queries — accurate kernel-level timing - 7B+ model support — optimize for Llama 3.1 8B and larger INT8 GEMM activation✅ — ~2× VRAM reduction viascripts/quantize_int8.py +--gemm-int8-block-size Streaming output✅ — tokens are emitted as generated- Multi-turn chat — KV cache reuse with chat template wrapper See docs/plan.html for the full interactive roadmap. | Document | Purpose | |---|---| | Architecture | System design, data flow, buffer layouts | | CLI Reference | All flags, sampling config, advanced modes | | Roadmap | Interactive implementation plan | | Doc Maintenance | Documentation rules for contributors | | Guides | wgpu & WGSL pitfalls and best practices | | Contributing | How to contribute | Dual-licensed under MIT or Apache-2.0 at your option.