Back
RCreddit.com
12
·19 hr ago·Dev community · RSS

(NInfer Fork) I wanted to have a 1M context Qwen-3.8 27B, tp2, dual 5090s

View original
QwenModel release

Heat trend

Collecting trend data

The percentage is based on available heat signal, not comment count or independent people.

Why it matters

Qwen model activity is surfacing — worth tracking for capability changes, ecosystem impact, and availability.

AI summary

A developer forked NInfer, a C++20/CUDA inference engine for Qwen models, to enable a 1M context for Qwen-3.8 27B. This was achieved by adding tensor-parallel processing across two GPUs and YaRN ×4 rope scaling. This setup allows the Qwen3.8-27B NVFP4 model to run a 1,048,576-token context on two consumer 5090 GPUs, utilizing 27.4 GB per card without NVLink.

Hey! I forked NInfer (a from-scratch C++20/CUDA inference engine for Qwen models) and added two things: tensor-parallel across two GPUs, and YaRN ×4 rope scaling. Together they let Qwen3.8-27B NVFP4 run a 1,048,576-token context on two consumer 5090s — 27.4 GB per card, no NVLink.

- Decode at 653k context: 119 tok/s with MTP speculative decoding, 57 without. vLLM on the same prompts: 42 tok/s.

- Why: past its native 262k window, vLLM's MTP acceptance drops to exactly zero (0 of 1,533 drafts accepted) — it keeps paying for the drafter. NInfer's acceptance stays ~55–60% out to 1M.

- Prefill is where vLLM wins: 1.2–1.3× faster. A full 1M prompt takes ~18 minutes on NInfer. That's the untuned part.

- vLLM's fp8 KV cache tops out at ~759k tokens on this hardware; NInfer's INT8 KV fits 1,048,576 in less memory.

Two GPUs are also just faster than one: 75 vs 54 tok/s at 250k, because weights and KV traffic halve per card and the ~128 cross-GPU reductions per token cost only ~0.2 ms under CUDA graphs.

(NInfer Fork) I wanted to have a 1M context Qwen-3.8 27B, tp2, dual 5090s · BuzzRadr