dspark

Up to 4× faster LLM decoding on Apple Silicon, lossless. Native MLX port of DeepSeek's DSpark & z-lab's DFlash speculative decoding — Gemma-4, Qwen3.8, Muse-Glimmer, Nemotron, LFM2.5, Ornith-1.0, ternary Bonsai-27B.

A PyTorch native library for training speculative decoding models

DeepSeek-V4-Flash-0731 on 8x RTX 3090 (SM86) with vLLM — verified FP8 serving, benchmarks, build guide, and reproducible release.

Anemll + MiaAI-Lab dual DGX Spark DSV4F+DSpark on vLLM 0.25.2.dev0 — 100% abliterated weights, Spec-Kit one-shot agent package. Full credit Anemll/MiaAI.

DeepSeek-V4-Flash + DSpark speculative decoding on a pair of NVIDIA DGX Sparks (vLLM TP=2 over RoCE) — tuned recipe, overlays that halve multi-turn TTFT, contamination-guarded benchmarks, ops runbook

Lossless inference speedup benchmarking suite for local LLMs using DeepSeek DSpark speculative draft heads.

Offline DeepSeek-V4.1-Flash deployment for 8x A100: DSpark k5, 256K context, local APIs and measured benchmarks.

DSpark-style speculative decoding for OCR VLMs on Apple Silicon (MLX). Faster local OCR for DeepSeek-OCR-2, GLM-OCR, Unlimited-OCR.

Verified IQ2_XXS plus native DSpark recipe for DeepSeek V4 Vision on one DGX Spark.

Universal Dual-Engine & Speculative Curation Configuration Kit for Antigravity, Cursor, Claude Code, Windsurf & Grok (FastMCP)

Exactness-first fast iteration lab for improving DSpark inference speed

Verified two-node GB10 DeepSeek-V4-Flash-0731 DSpark Graph-8 deployment and benchmarks

Running Cogni-Brain on DGX Spark · Nemotron-3.5-Lightning-30B-A3B-NVFP4 + DSpark (1M context)

SGLang patches for RTX 4090/SM89: GLM-5.2 FP8/256K, NVFP4 Marlin BF16, and Qwen3.8 DSpark.

Making AI research available to everyone. Screw NVIDIA and their overpriced GPUs

Complete GB10 (DGX Spark, sm_121) block-FP8 Triton kernel configs for DeepSeek-V4-Flash-0731 at TP=2, including the DSpark main_proj shape missing from every earlier run

Train workload-specific speculative decoding drafts for LLM inference — DSpark, DFlash2, EAGLE3 and more.

Multi-Ensemble Memory-Elastic Token Prediction — a numpy implementation of DeepSeek's DSpark speculative decoding + the MEMTP elastic-ensemble extension, on AMD Strix Halo (gfx1151).

Speculative-draft training pipeline (DFLASH/DSPARK/DFLASH2), orchestrating the vllm-project/speculators engine end to end, with preset configs for Qwen-family models