Unlock 250k Context on RTX 4090: Optimizing Qwen 3.8 with Custom Drafters
alexcovo_eth · x · 2026-08-23
The author achieved a 250,000 token context window for Qwen 3.8-27B on a single RTX 4090 (24GB) running at 75 tokens/s through specific optimizations.
Key Optimization Techniques:
- Custom 2-bit Drafter: Quantized the DFlash 2 drafter to Q2K, saving 450MB of VRAM while maintaining 100% draft acceptance and decode speed (76 t/s) in benchmarks.
- Hidden VRAM Tax: Discovered that llama-server reserves massive chunks of VRAM for multi-user batching by default. By passing the --parallel 1 flag, the engine dedicates the full 24GB buffer to a single user session.
Results:
Combining the VRAM savings from the Q2K drafter with the --parallel 1 flag caused the context ceiling to "absolutely explode," enabling the massive 250k context length on consumer hardware.
More from Infra
- Fluidstack hires Electrical Engineer for Modular R&D to build frontier AI infrastructure — MxMnr · 2026-10-16
- Jensen Huang joins BG2 podcast: does AI revenue justify capex, '27 bottlenecks — elonmusk · 2026-10-16
- Mistral's on-device infra rack revealed: DGX Spark, Mac Studio and more — gregmushen · 2026-10-16
- NVIDIA and Sweden's Wallenberg launch Sferical One, a shared AI supercomputer — nvidia · 2026-10-16
- Next infra frontier: self-evolving systems that adapt and self-verify — _shreya_s · 2026-10-16
- Xiaomi's MiMo RL dashboard reveals $3.47M spent on just 30 training steps — bycloud · 2026-10-16