Dio Fu

@VceOfReason Staat „raubt“ dir 79€? Stell dir vor, du hättest sie noch: Du würdest kettenrauchend, völlig besoffen und mit vollem Tank als rollende Brandbombe durchs Land pflügen.

Du hast 100€. Dir bleiben nach Steuern und Abgaben durchschnittlich 47€. Du kaufst dir eine Schachtel Zigaretten für 9€. Davon gehen 6€ an den Staat. Bleiben 38€. Du kaufst anschließend eine Flasche Sekt für 12€. Davon gehen ca. 5€ an den Staat. Bleiben 26€. Du kaufst von diesem Geld etwa 13 Liter Benzin. 15€ gehen an den Staat. In diesem Szenario gehen 21€ an dich und 79€ an den Staat. Es ist nichts anderes als Raub. Schlimmer als die Mafia.

@xYurimax @spiritbuun i am on Windows 10, similar args. also same performance :-S

@spiritbuun this seems really huge improvement ,i tested it myself with latest build but i only get 45-50t/s ,on rtx 3090 windows 11 with these args jpst.it/4_8dE what should i change @spiritbuun

Now that the dust has settled, is DFlash real or hype? Today I was able to break the 200 tok/s barrier with 27B on a single RTX 3090: run 1: 194.6 t/s, accept=155/180 run 2: 205.1 t/s, accept=155/180 run 3: 206.3 t/s, accept=155/180 But what about real world usage? Well,

@spiritbuun maybe I can get it up to 0.2-0.3 , if I tell the model to repeat stuff.

@spiritbuun Nice! My acceptance rate hovers around 0.1 . Any ideas? here is how I run it: buun-llama-cpp/build/bin/Release/llama-server.exe -m Qwen3.6-27B-Q4_K_M.gguf -md dflash-draft-3.6-q8_0.gguf --spec-type dflash -c 196608 -cd 16536 --cache-type-k turbo3_tcq --cache-type-v turbo3_tcq

Pushed: DFlash implementation for llama-cpp. buun-llama-cpp/llama-server -m Qwen3.6-27B.gguf -md dflash-draft-q4_k_m.gguf --spec-type dflash

@pupposandro Megakernel 27B + DFlash + TurboCache, and we feast in Valhalla.

The new Qwen3.6-27B now runs on Luce DFlash. Up to 2x throughput on a single RTX 3090. Qwen3.6-27B ships the same Qwen35 architecture string and identical layer/head dims as 3.5, so the existing DFlash draft + DDTree stack loads it as-is. Throughput is lower than on 3.5. Looking forward for the updated version from the DFlash team to implement it as well! Repo in the first comment ⬇️

@pupposandro turboquant in there too someday? can the draft model also get quantized to make room for more context?

@Ali_TongyiLab anything audio related?

(•_•) < New Year gift <) > is on the way! / \

@RedCollie1 What if the rod was curved with its two ends coming together forming a ring?

@trutharmyfilms 182,687,495

How many views do you think the Tucker X Putin interview will get in the first 24hrs? Place your guesses in the comment section. Winner gets a Truth Army hoodie.

@PhoenixWallet @SatCastle21 Splicing funds in seems to require 2 on-chain transactions (the original transaction from another wallet and the actual splice-in). How is this better than doing a swap-in (in case there is enough incoming liquidity) ? Could that be improved ?

@SatCastle21 Suppose that your wallet is a bucket, and your balance is the water in the bucket. Receiving = adding more water Spending = pouring water out Any time the bucket needs to be resized to allow for more water, that's an on-chain operation.

Seriously @PhoenixWallet , WHEN does an incoming payment require an on chain transaction and when not? Please explain it to me like I‘m 5