Philippe Bich

Research Scientist @ Huawei

[Philippe Bich @pbicho96]

Small KVarN update 🚀

Thanks to everyone who tested it and shared feedback! The vLLM PR is coming soon.

At the same KV-cache budget:
• ~2× faster decode than TurboQuant
• Better accuracy
• MLA, hybrid & speculative decoding

[Philippe Bich @pbicho96]

@so_sthbryan Thanks for sharing our work :)

[Philippe Bich @pbicho96]

@MorelMatth66161 Thanks for sharing our work! https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@LivingInHarmony Ahah, thanks for sharing :) here our code: https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@vintcessun @mobailabs Thanks for sharing! Here the code https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@sztlink @no_stp_on_snek Thanks, happy everything works. I am still fixing many things on the repo so this kind of posts help me a lot :) put a star if you want to support the project :) https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@rascazzione @jordimash Thanks for talking about KVarN! Code to test it is here https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@vintcessun Thanks for sharing our work!

[Philippe Bich @pbicho96]

@sztlink @no_stp_on_snek Author here. Can you provide some info about your setup? It should not break, I tested it on math and it was great 🥲

[Philippe Bich @pbicho96]

@andre_banandre Thanks! Code available here https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@Marco_Ramilli Thanks for sharing our work!

[Philippe Bich @pbicho96]

@MorelMatth66161 Thanks for sharing. Here the code for benchmarking :) https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@ai_paper_jp Thanks for sharing! Here the code: https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@fukuronomoride Thanks for sharing our work! :)

[Philippe Bich @pbicho96]

@devaxsha Thanks for sharing. You can test using our code: https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@Memoirs Thanks for sharing :) https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@dr_cintas Our Huawei version that beats TurboQuant https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@ai_hakase_ Thanks! Code: https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@ai_hakase_ Thanks for talking about KVarN :) We also have the code available here: https://t.co/sBv78XjTBs

[Philippe Bich @pbicho96]

@jaga_prasanna @PavloMolchanov tbh, this is a trend I struggle to understand as well. In terms of accuracy, INT4 was very strong (especially compared to MXFP4). The main reason we're now dealing with all these new microscaling formats is that hardware is being built around them. Not a huge fan, personally