wd 🇵🇹🔺's banner
wd 🇵🇹🔺's profile picture

wd 🇵🇹🔺

@populartourist • 1,231 subscribers

RUSH | Jack of all trades | Trigger warning

Videos

LIVE
1.2k

Anya Rossi

sweetdream.ai

SweetDream.aiSponsored•Livecam
Streaming Now

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

HD live stream
Exclusive private shows
1.2k viewers online

Free preview available • Premium content

Qwen3.8 Next Flash UD-IQ3_XXS ~1 hour + 117k total tokens RTX5090 + R9 9950X on 96GB DDR5 Decode falls from 50 tps to low 20's above 40k ctx window whilst prefill holds around 400-700 tps - no drafters. Never got to hit reasoning budget (max set 107k). Flags below.
2:23
populartourist's profile picture

Qwen3.8 Next Flash UD-IQ3_XXS ~1 hour + 117k total tokens RTX5090 + R9 9950X on 96GB DDR5 Decode falls from 50 tps to low 20's above 40k ctx window whilst prefill holds around 400-700 tps - no drafters. Never got to hit reasoning budget (max set 107k). Flags below.

wd 🇵🇹🔺

62,911 Aufrufe • vor 1 Monat

Qwen3.6 35B-A3B MTP is a devil performing around 1,000 tok/s on code with RTX 5090 once the prediction gets easier. Been a while I've tried 35B-A3B. Seems I might be sticking around for the fun.
0:43
populartourist's profile picture

Qwen3.6 35B-A3B MTP is a devil performing around 1,000 tok/s on code with RTX 5090 once the prediction gets easier. Been a while I've tried 35B-A3B. Seems I might be sticking around for the fun.

wd 🇵🇹🔺

49,211 Aufrufe • vor 2 Monaten

Qwen3.6 27B Q6_K (Unsloth) 199 tok/s average throughput on code + text combo on RTX 5090. Power capped 400W and clock 2200 MHz. Keeping --spec-draft-n-max 12 gives nice bump for 125k context with symmetric q8_0 KV with vision. High ceiling compression q8_0/q5_1 still possible but noticeable loss above 128k. Better lower quant at this rate or reduce --spec-draft-n-max 9 to bump closer to 140k context whilst maintaining full q8_0 KV. Mainline llama.cpp
5:32
populartourist's profile picture

Qwen3.6 27B Q6_K (Unsloth) 199 tok/s average throughput on code + text combo on RTX 5090. Power capped 400W and clock 2200 MHz. Keeping --spec-draft-n-max 12 gives nice bump for 125k context with symmetric q8_0 KV with vision. High ceiling compression q8_0/q5_1 still possible but noticeable loss above 128k. Better lower quant at this rate or reduce --spec-draft-n-max 9 to bump closer to 140k context whilst maintaining full q8_0 KV. Mainline llama.cpp

wd 🇵🇹🔺

13,972 Aufrufe • vor 2 Monaten

Keine weiteren Inhalte verfügbar