Fast LLM speculative inference server for consumer hardware.

Model compression toolkit engineered for enhanced usability, comprehensiveness, and efficiency.

DFlash & TurboQuant in llama.cpp with up to 3x faster generation and 7.5x more KV cache in same VRAM

Fully uncensored, capability-enhanced abliteration of Qwen3.6-27B. NVFP4 + z-lab DFlash speculative decoding (n=12) on the unified ghcr.io/aeon-7/aeon-vllm-ultimate:latest container, tuned for long-context draft acceptance on DGX Spark. 6 HF variants (BF16/NVFP4/MTP/MTP-XS), docker-compose, and QuickStart.

DFlash vLLM for DGX Spark — Plug & Play Block-Diffusion Speculative Decoding

vLLM Qwen 3.6-27B (AWQ-INT4) + DFlash speculative decoding on AMD Strix Halo (gfx1151 iGPU, 128 GB UMA, ROCm 7.13). 24.8 t/s single-stream, vision, tool calling, 256K context, OpenAI-compatible, Docker. Matches DGX Spark FP8+DFlash+MTP at a third of the cost. No CUDA.

llama.cpp fork optimized for NVIDIA DGX Spark / GB10 (Blackwell, SM 12.1) — TurboQuant weights + KV, NVFP4, DFlash MTP

Local AI workstation — discover, run, chat, benchmark, and generate images from open-weight models. DFlash/DDTree speculative decoding, TurboQuant & TriAttention cache compression strategies, MLX + llama.cpp + vLLM + MTPLX backends.

Qwen3.6 27B × DFlash — 30-35 tok/s on NVIDIA DGX Spark (GB10) - LLama.Cpp

Experimental DDTree-on-vLLM research track for Qwen3.6 AEON Ultimate on DGX Spark / GB10.

Run Qwen 3.6-27B AWQ-INT4 models with DFlash speculative decoding on AMD Strix Halo hardware using vLLM for high-throughput inference.

Qwen3.6-27B BF16+DFlash 13-config parameter sweep on repne/vllm:v2. Stage A (3x3 buffer/graph) + Stage B (num_speculative_tokens) + Quality (HumanEval/MBPP). 7h, 421 problems, 195 cells.

GGUF-native DFlash speculative decoding runtime for local models

vLLM v0.21 + DFlash + thinking_token_budget for Gemma 4 & Qwen 3.6 on Blackwell GB10 (sm_121a / sm_120)

CLI for building and testing DFlash-style speculative decoding draft models.

Study #4: FP8+MTP{3,5} speed on repne/vllm:v2 + max_tokens=8192 quality re-runs for BF16+DFlash n=8 and FP8+MTP=3. Follow-up to studies #2 and #3.

Reproducible efficient-inference stack for Qwen3.5-4B (AdaptFM Efficient Qwen Competition): GPTQ W4A16 g128, untied W8 lm_head, DFlash speculative decoding, and per-step vLLM latency optimizations. 7.745× average latency speedup with all quality gates passing.

Same BF16+DFlash config on Repne fork vs upstream vLLM v0.20.1, dual RTX PRO 6000 Blackwell. Upstream's dflash collapses at long context (5-6x slower at 131K).

Cogni-Brain on DGX Spark: Qwen3.5-122B-A10B INT4+FP8 hybrid, DFlash speculative decoding, 262K context, ~54 tok/s, 100/100 Tool-Eval, vLLM.