vLLM Blog

Agree that routing is becoming the critical layer here. Vllm iris is really promising for this https://blog.vllm.ai/2026/01/05/vllm-sr-iris.html There's already some good work on router benchmarking which is pretty interesting

AMD GPUs are becoming a serious contender for LLM inference. vLLM is already showing impressive performance on AMD [1], even with consumer-grade Radeon cards (even support GGUF) [2]. This could be a game-changer for folks who want to run LLMs without shelling out for expensive NVIDIA hardware. [1] https://blog.vllm.ai/2024/10/23/vllm-serving-amd.html [2] https://embeddedllm.com/blog/vllm-now-supports-running-gguf-on-amd-rad…

1x MI300x has 192GB HBM3. 1x MI325x has 256GB HBM3e. They cost less, you can fit more into a rack and you can buy/deploy at least the 300's today and 325's early next year. AMD and library software performance for AI is improving daily [0]. I'm still trying to wrap my head around how these companies think they are going to do well in this market without more memory. [0] https://blog.vllm.ai/2024/10/23/vllm-serving-amd.html

Basically no one uses FP32 at inference time. BF16/FP16 is typically considered unquantized, whereas FP8 is lightly quantized. That being said there's pretty minimal quality loss at FP8 compared to 16-bit typically; Llama 3.1 405b, for example, only benchmarks around ~1% worse when run at FP8: https://blog.vllm.ai/2024/07/23/llama31.html Every major inference provider other than Hyperbolic Labs runs Llama 3.1 405b at FP8, FWIW (e.g. Together, Fireworks, Lepton), so to compare against FP32 is misleading to say the least. Even Hyperbolic runs it at BF16. Pretraining is typically done in FP32, although some labs (e.g. Character AI, RIP) apparently train in INT8: https://research.character.ai/optimizing-inference/

While Llama 4 had a pretty bad launch (the LM Arena gaming in particular is terrible), having run my own evals on it (using the April 5 v0.8.3 vLLM release - https://blog.vllm.ai/2025/04/05/llama4.html , so before the QKNorm fix https://github.com/vllm-project/vllm/pull/16311) - it seemed pretty decent to me. For English, on a combination of MixEval, LiveBench, IFEval, and EvalPlus Maverick FP8 (17B/400B) was about on par with DeepSeek V3 FP8 (37B/671B) and Scout (17B/109B) was punching in the ballpark of Gemma 3 27B, but not too far off Llama 3.3 70B and Mistral Large 2411 (123B). Llama 4 claimed to be trained on 10X more multilingual tokens than Llama 3 and testing on Japanese (including with some new, currently unreleased evals) the models did perform better than Llama 3 (although I'd characterize their overall Japanese performance as "middle of the pack": https://shisa.ai/posts/llama4-japanese-performance/ I think a big part of the negative reaction is that in terms of memory footprint, Llama 4 looks more built for Meta (large scale inference provider) than home users, although with the move to APUs and more efficient CPU offloading, there's still something to be said for strong capabilities at 17B of inference. I think people are quick to forget that Llama 3, while not so disastrous, was much improved with 3.1. Also the competitive landscape is pretty different now. And I think the visual capabilities are being a bit slept upon, but I think that's also the case of…