Hilarious read. I laugh out loud multiple times during the read. In the end, I think amd should pay the author simply for the will of debugging for a broken software written by amd, as well as the sheer amounts of loose ends this exploration leads to.
Compared to autoregressive decoding, diffusion is huge for local MoE inference because of the improved token generation efficiency, especially for normal GPU + ram offload setting.
However, there are models which are better positioned on the performance vs memory pareto front, i.e. dense models, so I'll just wait.
P.S. QAT is really something as it reduces the performance fluctuations compared to the normal one. Thanks again.
From the perspective of a local llm user, I think the qat doesn't solve the major problem of the gemma models.
Gemma family (gen 1 to gen 4) is consistent with extreme range of activations, i.e., 600000, essentially forcing people to use bf16 kv cache and accept a short context window, e.g., 31b, iq4_xs quantization, 100k context window on 32gb memory. Or, people use q8 kv cache, 200k context window, and accept a large performance penalty.
In contrast, for qwen 3.5 family, the largest activation is below 2000, making q8 or even lower-precision kv cache essentially free estates. Together with linear attention, which doesn't require kv cache, full 262k context window can be easily reached.
Qat training with w4a16 target, while improving performance on inference with low-precision weighs, doesn't solve kv cache problem at all.
In the end, a qat is a qat, and there are unseen efforts behind qat checkpoints. Thank you gemma team for releasing qat checkpoints.
More rants about local inference, consider yourself warned.
Together with bf16 related deliberate hardward degrades on consumer-level nvidia gpus, i.e., gtx 10, rtx 20, 30, 40, 50 series, things gets sour really quickly.
You can get SXM V100s for like $100 off ebay, if you're willing to do the troubleshooting work to get em running with adapters you can build a computer capable of fitting a Q4 quant of a 120b model in VRAM for something like fifteen hundred dollars. (assuming you already have some RAM sticks laying around T___T)
Wow. Just like using ungoogled-chromium instead of chrome, lineage os instead of oem android, using vscodium instead of vscode is again justified. These decisions really are the ones that I'll never regret.
In addition, using the word microslop instead of microsoft is again justified, too.
One thing that makes me wonder is that there are 4 security issues raised and all of them were automatically commented and closed by some bot called `pl-ghost` [1][2][3][4]. In the end, only this one [4] properly handled, and all bot comments are deleted. You can see the bot comments in another report [5], which is more informative than the OP one.
Andy from Lightning here. Yeah, the PyPi credentials were stolen through the compromised pl-ghost bot account. The attacker used this account to create a new actions workflow, which was ran and parsed out secrets for PyPi. After releasing the package, the attacker then used that account to troll us a bit with those comments.
reply