https://conselara.dev/tags/gpt-oss/ Recent content in Gpt-Oss on Conselara Labs Hugo en Mon, 15 Jun 2026 00:00:00 +0000 https://conselara.dev/notes/gpt-oss-120b-single-dgx-spark/ Mon, 15 Jun 2026 00:00:00 +0000 https://conselara.dev/notes/gpt-oss-120b-single-dgx-spark/ gpt-oss-120b runs at ~38 tok/s on a single DGX Spark in MXFP4. The SM121 config is unforgiving: marlin is mandatory, FlashInfer is invalid, and the Harmony tokenizer needs an offline pre-stage. Here's the working setup and the traps.