Groq doesn't even have any true deepseek models --- I thought they only had `deepseek-r1-distill-llama-70b` which was distilled onto llama 70b [1].
Groq doesn't even have any true deepseek models --- I thought they only had `deepseek-r1-distill-llama-70b` which was distilled onto llama 70b [1].
Groq has a weak selection of models, which is frustrating because their inference speed is insane. I get it though, selection + optimization = performance.
From conversation with someone from Groq, they have a custom compiler and runtime for the models to run on their custom hardware, which is why the selection is poor. For every model type they need to port the architecture to run on their compiler beforehand.
They can't host DeepSeek because it's too big. Their chips have 230mb of memory, so it will take them ~3000 chips to host the model + (possible large) number of chips to keep kv cache. I bet it's just too hard to bring such topology online at all, and impossible to make even near to be profitable.
the only reason they are fast is because the models they host are severely quantized so i've heard.
Huh. I heard a podcast with the founder talking about their custom hardware, but quantization would explain it.
Quantization alone does not explain it. It's mostly custom hardware[0].
[0] https://groq.com/the-groq-lpu-explained/
Why repeat this nonsense when it’s so trivial to just check. The reason Groq is fast is because they employ absolutely ludicrous amounts of SRAM. (Which is 10 times faster than the fastest VRAM.)
they responded to my tweet last year and said they didn't quantize the models.
It's very hard to find right now but I'm sure they said they don't quantize KV cache, but their weights are in fp8.