Posts
All the articles I've posted.
-
How we got Qwen3.8-27B to 92.9 tok/s and a 360K window on one RTX 5090
The engineering behind our NVFP4 build of Qwen3.8-27B: why decode is a bandwidth problem, why lm_head was the tensor that mattered, and the exact environment it was built and measured in.
-
830K downloads later: what quantizing Qwen3.8-27B to NVFP4 for one RTX 5090 actually taught me
Our NVFP4 build of Qwen3.8-27B passed 830K monthly downloads. Then NVIDIA and two others published their own. I benchmarked all three on one 5090 — quality was a tie, and what separated them was precision allocation, a drafter trained against the checkpoint, and the runtime underneath.
-
18.7× in two days: what sparkinfer-k3 taught me about shipping fast on Gittensor
sparkinfer-k3 took Kimi K3 (2.8T) from 1 tok/s to 3× llama.cpp on 8× H200 in days. The climb, its bugs, and what it taught me about racing on Gittensor's SN74.
-
Proof you can't fake: shipping dual-vendor Intel TDX + NVIDIA CC attestation for AI training claims
I closed every trust gap in a real proof-of-training system: TDX measured-VM quotes, Intel DCAP signature verification, and NVIDIA JWKS-verified GPU tokens — live, not claimed.
-
Distilling frontier reasoning into an edge model — Unsloth + W&B in Colab, Axolotl on Blackwell
Distill Opus 4.6 / Qwen3.5 reasoning into ~14k SFT samples; prototype on Llama 3.2 3B in Colab with Unsloth + W&B, then Axolotl on RTX PRO 6000 Blackwell.
-
TRT-LLM vs vLLM vs llama.cpp on Blackwell — sparkinfer v0.3.7 owns decode on the RTX 5090
sparkinfer v0.3.7: ~30% faster decode than llama.cpp on RTX 5090 at the same LLM quality — 2.5 MB native binary. I develop and maintain it. Plus TRT-LLM vs vLLM vs llama.cpp on PRO 6000.