600 Toks/Second Gemma4-26B —The Setting That Actually Wins (vLLM + Dflash Speculative Decoding)

600t/s ? it feels illegal. I swept every DFlash speculative decoding setting from n=0 to n=15 on Gemma 26B running on a single RTX 5090. Baseline was 228 output tokens per second. The winning setting hit 578 — a 2.56x speedup. But the right answer wasn't the highest number, and the batch scheduling budget mattered just as much as the token count. #dflash #vllm #5090 #gpu #llm #gemma4
👉ⓢⓤⓑⓢⓒⓡⓘⓑⓔ
👉 !! try Wan Video online at https://agireact.com/wan-t2v !!

If you found this useful:
👍 Like if the results surprised you
🔔 Subscribe for more local AI benchmarks and hardware deep-dives
💬 Drop your setup in the comments — curious what you're running models on

Whether you're running local AI on older hardware or wondering if the 5090 is actually worth it — this one's for you. #qwen #LocalAI #LlamaCPP #RTX5090 #RTX4090 #RTX3090 #MacBook #AIBenchmark #LLM

🖥️ Hardware Tested:
- NVIDIA RTX 5090 (32GB VRAM)

🤖 Models Benchmarked:
- Gemma4 26B (Q4)

For Gemma4 model comparison, see https://youtu.be/VYc47oqBnqI

Please join the discord server at https://discord.gg/SgmBydQ2Mn where you developed free chatgpt bot and stable diffusion bot!
If you would like to support me, here is my Kofi link: https://ko-fi.com/techpractice and Patreon page: https://www.patreon.com/user?u=89548519
Thank you for watching!

Tutorial links:
For python virtualenv install, see https://youtu.be/uOCL6h9fuVc
ComfyUI for more advanced workflows
ComfyUI on Macbook tutorial: https://youtu.be/ZCswfm0dBYY
FLUX on Macbook: https://youtu.be/asngm4s_9Ho
The ComfyUI workflow can be downloaded from https://github.com/ttio2tech/ComfyUI_workflows_collection (Pulid_flux_workflow.json)

Affiliate links: buy hardware on Amazon
Mac-Mini M4: https://amzn.to/4emPxrB (also has coupon)
AMD GPU: https://amzn.to/3vCp6h1
4600G https://amzn.to/45LhGFa
5600G: https://amzn.to/3LgnFtC (same iGPU, better CPU)
5700G: https://amzn.to/3Z9gUiM (better iGPU, and better CPU)
ssd drive: https://amzn.to/3MVJdg2
DDR4 drive: https://amzn.to/3sKNufi
AM4 motherboard: https://amzn.to/3GfrPit
PSU (power supply unit): https://amzn.to/3Gd87UA
PC Case: https://amzn.to/3QPDNnF
if you are interested in discrete GPU: https://amzn.to/3QT1wDp

Видео 600 Toks/Second Gemma4-26B —The Setting That Actually Wins (vLLM + Dflash Speculative Decoding) канала Tech-Practice