NeuroBrix 0.6.10

NeuroBrix 0.6.10

With nbxtensor 0.1.4. The engine now leaves to its tensor library what is the library's own: the replay of decode
steps, the expert store of mixture-of-experts models (its device slots, copies and the routing over several cards), and
the engine's remaining tensor helpers live in nbxtensor, and the engine calls them. Measured on the outputs: byte-identical
to 0.6.9 on the gate's models, one card and two cards, both engines.

• Long prompts are tokenized in linear time, with the same tokens. A 12 265-token prompt to a Llama-tokenizer model
  took 25.7 s before the model saw it; now 0.03 s. A model with thousands of added tokens no longer stalls on a long
  prompt (orpheus, 34 880 characters: 43.6 s -> 0.02 s). serve no longer keeps every long prompt in memory.
• A new server process no longer measures again the kernels this machine already measured. The settings a previous
  run measured were given only to the kernels loaded before the first computation; the ones a server loads at its first
  request were measured again by every new process.
• neurobrix run refuses by name a prompt longer than the model's context, before anything loads, as serve does.
  It used to fail in a GPU kernel (--triton) or abort the process (--compiled).
• A run whose card reports its error late exits with that error named, rc 1, instead of aborting.
• /v1/audio/speech without a voice, for a model that ships several and declares no default, is a 400 naming
  voice and the voices, no longer a 500.
• neurobrix run --turns states each turn's own timings.
• Removed: the terminal coding agent (neurobrix chat --agent, run --mode agent and their flags). Coding agents
  become their own project, speaking to neurobrix serve through the OpenAI-compatible API.

Install: pip install -U neurobrix (the Triton engine, no PyTorch) or pip install -U "neurobrix[pytorch]"; on a Mac,
python -m nbxtensor install-metal. nbxtensor 0.1.4, nbxgrammar 0.1.1 (unchanged).