root dependency our slime fork rl trainer with custom speed-up and algorithmic improvements, built on top of slime
megatron-lm nvidia's framework for training neural nets
pinned to specific commit (Dockerfile#L7)
patches applied at runtime (megatron.patch)
sglang inference engine — samples model outputs during rl
pinned to specific version (Dockerfile#L1)
patches applied at runtime (sglang.patch)
megatron-bridge enables efficient lora finetuning on top of megatron-lm