That's cause dude used DS4.1 Flash on SGLANG, which corrupts outputs. That's why I used vllm for this one. Jun hey remember when we were testing SGLANG on deepseek v4.1 and it was no go because you…

3/ Three things worth knowing:· DSpark k=5 with a greedy draft: +30% single stream (4.36 tokens accepted/step, 67% per-position) · SGLang's DSpark path corrupts output on GB10; vLLM's is clean on the…

Proof of Agent work. Built by Depths, tested and approved for our infrastructure work by Eva. But need to test for stability... we shall see in the next 24 hours.

Used Fable 5.1 repeatedly to squeeze more performance on 4x DGX Spark + GLM 5.3 Flash EXL3. Special thanks to MiaAI_lab and Reederey for their recent updates, which my repo uses in various upgrades…

The bad guys already have hacking tools. The good guys need AI that will actually help defend their code, not refuse the job. Grab GLM 5.3 DFlash2 Uncensored in EXL3. It runs great on DGX Sparks.

100 tok/s achieved on c1 structured GLM 5.3 Flash on 4x spark. EXL3 4bpw C4 performance 253 tok/s. Bench at 420k holds: structured 93.7 / code 37.8 / math 74.5 / prose 36.2 tok/s — no regression vs…

OK finally got it, my contribution: the best combo for GLM 5.3 Flash on 4x sparks: **SGLang + NVFP4 + DFlash2, TP4, NCCL over RoCE**: 88 tok/s for c1, 90 for c4. Was 10-38 tok/s before optimization.…

Reached a new peak 159 tok/s and record 290 tok/s on good o Deepseek v4 Flash 0731 abliterated, and it was still doing the same memory work. I'm going back to this setup on the 4x sparks (of the…

lyf/Qwen3.8-27B-Heretic-ARA-NVFP4-MTP-VL Why it matters: this is a compressed-tensors NVFP4 W4A4 Qwen3.8 derivative with the BF16 vision/video tower and all 15 native BF16 MTP tensors retained. It is…

quick topology connection - 1st (left) 400g connector splits to, - bottom left spark + top left spark - 2nd (right) 400g connector splits to, - bottom right spark + top right spark The github repo…