This model requires you to accept the Apache 2.0 license terms before download.
Description
IBM Granite 3.1 1B-A400M Instruct — a 32-expert mixture-of-experts language model (8 experts per token, 24 layers), the first MoE container through the fused expert dispatch; traced and validated on NVIDIA V100 in sequential, compiled and Triton modes
Tags
At a glance
| Kind | language model |
|---|---|
| Precision | bfloat16 |
| Container size | 2.6 GiB |
| Weights inside | 2.6 GiB |
| Components | 2 |
| Licence | Apache 2.0 |
What is inside the container
A .nbx container packs every part of granite-3.1-1b-a400m-instruct — 2 components with weights, graphs and runtime settings — so one file is all the runtime needs.
| Component | Blocks | Layers | Weights |
|---|---|---|---|
| model | 0 | 24 | 2.5 GiB |
| lm_head | 0 | — | 96.0 MiB |
Default settings
The values the container ships with. They apply when you do not pass your own.
- Max length
- 8192
- Temperature
- 0
- Top-p
- 1
- Top-k
- 50
Memory needed
The weights of granite-3.1-1b-a400m-instruct take 2.6 GiB in bfloat16. That is the floor for the memory the runtime has to place, spread over the devices it finds; activations come on top and grow with the size and length you ask for. A measured peak per device class is not published for this model yet.
Verify what you downloaded
Every download carries its SHA-256, and the file endpoint sends it in the X-NBX-SHA256 header. The container is 2.6 GiB (2,771,829,873 bytes).
$ sha256sum granite-3.1-1b-a400m-instruct.nbx
58cd33fee67e4cde7df86039b07eb743a2eead731b93c4c4270cb1b3c08bcf67
Questions
What is granite-3.1-1b-a400m-instruct?
A language model by ibm-granite, distributed as a bfloat16 .nbx container and licensed Apache 2.0.
How big is the download?
2.6 GiB, of which 2.6 GiB are weights.
How do I run it?
Install the runtime from the download page, then use the commands under “CLI Usage” below: neurobrix import ibm-granite/granite-3.1-1b-a400m-instruct, then neurobrix serve --model granite-3.1-1b-a400m-instruct.
CLI Usage
$ neurobrix import ibm-granite/granite-3.1-1b-a400m-instruct
$ neurobrix serve --model granite-3.1-1b-a400m-instruct
$ neurobrix run --prompt "your prompt here"
$ neurobrix stop