g

Atlas Inference Engine

Pure Rust & CUDA LLM Inference
Universal Inference At Unimaginable Speeds

---

⚡ What is Atlas?

Atlas is a high-performance, pure Rust & CUDA LLM inference engine purpose-built for prosumer workstations (NVIDIA DGX Spark / GB10 SM121 and AMD Strix Halo). No Python, no PyTorch, no bloated dependency trees—just one compact binary with hand-tuned micro-kernels.

• Sub-90s First Token: Boots in seconds with cached weights; zero JIT compile or Python startup lag.
• Default Flagship Qwen 3.8 27B: Dense hybrid GDN + Attention running at 23.59 tok/s single-stream with MTP speculative decoding on a single GB10.
• Nemotron 3.5 Lightning + DSpark: Full bring-up of hybrid Mamba-2 SSM + MoE paired with DSpark speculative decoding drafters for sub-10ms token decode latencies.
• Qwen 3.8 Flash-Next Support: Stream massive ~180B hybrid MoE models inside ~90 GB resident VRAM using direct parallel pread NVMe offloading.
• Turnkey Sparkrun Integration: Launch any verified model recipe instantly with sparkrun run @atlas/<recipe>.
• OpenAI & Anthropic Compatible: Drop-in API endpoint supporting streaming, tool calling, and reasoning traces.
• MLPerf Proven: Official contributor to the MLPerf Inference v6.1 Edge Agentic benchmark.

---

🚀 Quick Start

1. Default Flagship: Qwen 3.8 27B (Dense) via Sparkrun

The default flagship recipe deploys Qwen 3.8 27B in NVFP4 on a single GB10 with native MTP speculative decoding:

# Step 1: Install sparkrun
pip install sparkrun  # or: uvx sparkrun setup install

# Step 2: Download weights (or let sparkrun fetch automatically)
huggingface-cli download unsloth/Qwen3.8-27B-NVFP4 \
  --local-dir ~/.cache/huggingface/hub/models--unsloth--Qwen3.8-27B-NVFP4

# Step 3: Launch service (port 8888)
sparkrun run @atlas/qwen3.8-27b-nvfp4 --hosts localhost

Prefer the one-line quickstart script?

curl -fsSL https://atlasinference.dev/quickstart.sh | sh

2. Nemotron 3.5 Lightning 30B + DSpark Speculative Drafter

Pairs the hybrid Mamba-2 + Attention + MoE backbone with NVIDIA's 6-layer DSpark drafter (gamma=4 / K=3 verify) for sub-10ms token generation:

# Export recommended performance environment
export ATLAS_DFLASH_OPTION_B=1
export ATLAS_NO_TOOL_INJECT=1   # +15.58 BFCL accuracy boost

# Launch Nemotron 3.5 Lightning with DSpark
sparkrun run @atlas/nemotron-3.5-lightning-30b-a3b-nvfp4-dspark --hosts localhost

3. Qwen 3.8 Flash-Next (~180B Hybrid MoE)

Atlas dynamically streams the 47.7 GB PLE n-gram table off NVMe using parallel pread workers, keeping peak resident memory under 90 GB on a single 128 GB GB10:

huggingface-cli download RadixArk/Qwen3.8-Flash-Next-NVFP4 \
  --local-dir ~/.cache/huggingface/hub/models--RadixArk--Qwen3.8-Flash-Next-NVFP4

sparkrun run @atlas/qwen3.8-flash-next-nvfp4 --hosts localhost

4. AMD Strix Halo (gfx1151) Quick Start

Atlas converts CUDA sources natively to AMD RDNA 3.5. Bring-up is active across two dedicated branches:

• Linux Leg (Ubuntu 24.04 / ROCm 6.2+): Branch port/qwen3.8-strix-linux (PR #8)
• Windows Leg (DirectX 12 / native MSVC): Branch port/qwen3.8-windows (PR #9)

# Clone the AMD Strix Halo port branch
git clone -b port/qwen3.8-strix-linux https://github.com/Atlas-Inf/atlas.git
cd atlas

# Build native binary targeting AMD gfx1151 silicon
cargo build --release --features scale,rocm

# Launch Qwen 3.8 27B on Strix Halo
./target/release/spark serve unsloth/Qwen3.8-27B-NVFP4 --bind 127.0.0.1 --port 8888

5. Querying the Endpoint

Atlas serves an OpenAI-compatible API on the designated port:

curl http://localhost:8888/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "atlas",
    "messages": [{"role": "user", "content": "Explain quantum computing in three sentences."}],
    "max_tokens": 256
  }'

---

📦 Verified Model Recipes

Every recipe is maintained in the sparkrun-recipes SSOT repository and verified against committed gate baselines:

Vendor	Family	Model Recipe	Quant	Topology	Highlights
Qwen	Qwen3.8	@atlas/qwen3.8-27b-nvfp4	NVFP4	Single GB10	Default Flagship. Dense hybrid GDN + Attn, MTP spec decode, FP8 KV, 23.59 tok/s
Qwen	Qwen3.8	@atlas/qwen3.8-27b-nvfp4-latency	NVFP4	Single GB10	Low-concurrency / interactive profile tuned for minimal single-stream latency
Qwen	Qwen3.8	@atlas/qwen3.8-27b-nvfp4-throughput	NVFP4	Single GB10	Concurrency profile beating vLLM from 1 to 128 streams on GB10
Nemotron	Nemotron-3.5	@atlas/nemotron-3.5-lightning-30b-a3b-nvfp4-dspark	NVFP4	Single GB10	New. Hybrid Mamba-2 SSM + MoE with 1.3 GB DSpark drafter (K=3 verify)
Nemotron	Nemotron-3	@atlas/nemotron-3-nano-30b-a3b-nvfp4	NVFP4	Single GB10	30B / 3B active Mamba-2 + MoE
Nemotron	Nemotron-3	@atlas/nemotron-3-super-120b-a12b-nvfp4	NVFP4	Single GB10	120B / 12B active hybrid architecture
Qwen	Qwen3.8	@atlas/qwen3.8-flash-next-nvfp4	NVFP4	Single GB10	~180B hybrid MoE, 8K context, parallel pread NVMe offload (750–800 tok/s prefill, 36.7 tok/s decode, ~90 GB resident)
Qwen	Qwen3.8	@atlas/qwen3.8-flash-next-nvfp4-throughput	NVFP4	Single GB10	Throughput-tuned 8K context profile
Qwen	Qwen3.6	@atlas/qwen3.6-35b-a3b-fp8-mtp	FP8	Single GB10	35B / 3B active GDN + MoE + vision, MTP speculative
Gemma	Gemma-4	@atlas/gemma-4-26b-a4b-nvfp4	NVFP4	Single GB10	26B / 4B active MoE with GeGLU
DeepSeek	DeepSeek-V4	@atlas/deepseek-v4-flash-nvfp4-ep2	NVFP4	EP=2 (2 Sparks)	Dual-node Expert Parallelism

Browse the interactive recipe browser at atlasinference.dev/#models.

---

🏛️ Engine Architecture & Innovations

• Double-Buffered Mamba-2 Chunked Scans: Hand-tuned SM121 PTX kernels delivering an 8.4x prefill latency reduction over generic vLLM implementations.
• Native FP4 Tensor Core Prefill GEMMs: Direct execution in Blackwell NVFP4 precision without dequantization overhead.
• PLE N-Gram NVMe Streaming (Direct Parallel pread vs. mmap): Traditional engines (like baseline llama.cpp) suffer from thousands of scattered 4KB mmap page faults for tiny ~90-byte rows, stalling prefill at ~300 tok/s. Atlas implements an asynchronous O_DIRECT worker pool (ATLAS_PLE_FAULT_THREADS=32) using parallel pread directly off NVMe storage (similar to the optimization in llama.cpp PR #28136). This delivers 750–800 tok/s cold prefill on DGX Spark (2.5x faster) and +20–32% on Strix Halo 128GB, completely bypassing OS page-cache faults while keeping the entire 47.7 GB n-gram table off RAM/VRAM.
• Recurrent State Checkpoint & Rollback: Enables multi-token speculative decoding with DSpark on recurrent state models (Mamba-2 / GDN) without state divergence.
• TurboQuant+ KV Cache: Symmetric and asymmetric KV quantization (bf16, fp8, nvfp4, turbo4) with Randomized Hadamard Rotation and Lloyd-Max codebooks.

KV Cache Options

Flag	Bits/elem	Storage	Description
--kv-cache-dtype bf16	16	BF16	Uncompressed baseline. Recommended for short-context or high-precision needs.
--kv-cache-dtype fp8	8	FP8 E4M3	Default. Halves memory with minimal quality degradation across all benchmarks.
--kv-cache-dtype nvfp4	4	E2M1	4x compression vs BF16. Excellent for long context windows.
--kv-cache-dtype turbo4	4	E2M1 + WHT	~2x lower reconstruction MSE than standard NVFP4 via Lloyd-Max codebooks.

---

🤝 Community & Support

• Website: atlasinference.dev
• Discord: Join our Discord — Active daily development, live kernel tuning, and model requests.
• Recipes Repository: Atlas-Inf/sparkrun-recipes
• Deployment Guide: GB10 Deployment Guide

---

⚖️ Dual License

• Community Edition: Licensed under AGPLv3. Free and open for personal use, research, and non-commercial local deployments.
• Enterprise Edition: Commercial licensing for proprietary applications, SaaS hosting without AGPLv3 copyleft obligations, dedicated support, and custom hardware/kernel porting. Contact debaterishaqui@gmail.com.

_{Continuity notice. Atlas is continuing. This repository, the Atlas-Inf GitHub organization, and atlasinference.dev are the replacement official Atlas channels. The existing website and GitHub repository remain disputed Atlas assets that have not been relinquished.}