I kept seeing people ask "Which model i can run on my gpu", "will model X fit on my GPU". Thats why I built a filter on whichllmmodel that lets you search models by what will actually fit on your hardware (8GB, 16GB, 24GB, etc.) at a given quantization level.
handy, but the gap most of these filters have is that "fits in VRAM" doesn't mean usable.
context length blows up the KV cache fast, a 7B that fits at 2k tokens will OOM at 32k.
factoring context len + quant into the estimate is where it'd actually save people from getting burned.
i think you did not check app properly, it is actually taking required context window from the user and then caluclate kv cache size and then count it along with size of model itself. it also reserves some more memory to avoid oom....
very nice idea. Would be nice if you could also keep desired context as a free parameter and let the models tell you what maximum context you could have.
actually, currently it is not open source, but I am thinking about making it open source so that other developers can also contribute in it(espeically data layer). what do you think?