Quantized Qwen3.8 draft model cuts VRAM 2.6x, boosts context from 90K to 130K on one RTX 5090
QuixiAI · x · 2026-08-30
- Developer nb4ld's first HF upload: calibrated NVFP4 (W4A4) quantization of the DFlash 2 block-diffusion speculative decoding draft for Qwen3.8-27B, in NVIDIA ModelOpt layout that SGLang loads natively.
- Built for a single RTX 5090 (32GB): the draft shrinks from 3.53GB (BF16) to 1.37GB, turning into +44% KV-cache context (90K → 130K tokens) at the same decode speed (215 → 228 tok/s end-to-end) and same acceptance rate (3.71 → 3.60, within ±0.2 noise).
- Output quality is unchanged by construction since the target model verifies every drafted token; uncalibrated RTN quantization drops acceptance to 3.26, showing calibration matters.
- Also benchmarked a coding-agent session with tool calls on the SGLang source tree, with decode matching BF16 across context sizes.
More from Infra
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16
- Google Raises 2026 Capital Spending to Up to $205 Billion — emmanuelvivier · 2026-10-16
- Google open-sources AX, a Kubernetes-style orchestrator for billions of AI agent tasks — thisdudelikesAI · 2026-10-16