[dev] Megatron Lite (1/4): shared primitive core

[ISEEKYAN]

Summary

First of a 4-PR foundation series that extends the experimental/lite (Megatron Lite) stack on dev. This PR lands the shared primitive core that the rest of the series — and the Lite model implementations that follow — build on.

What's included

• Primitive package: primitive/utils/ (packaged from utils.py, plus rotary/rope/packed_seq/moe helpers), primitive/modules/ (dispatcher, experts, router, moe, mrope, mtp, gqa, gated_delta_net, mlp), primitive/parallel/ (cp, thd, pp, pipeline, state), primitive/optimizers/fsdp2, primitive/ckpt/ (dcp, distckpt), and bundle/data/train_step.
• Runtime contracts (runtime/contracts/: config, data, loss) — the model-agnostic config/data/loss contracts the primitive layer depends on.
• Foundation kernel shims used directly by primitives: primitive/kernels/{__init__, swiglu, jit}.
• model/protocol_utils.py shared scaffolding, plus the corresponding primitive unit/smoke tests.

Series

Megatron Lite foundation (4 PRs), merged in order:

• 1/4 shared primitive core ← this PR
• 2/4 runtime layering guard (variable-leakage governance)
• 3/4 fused kernels
• 4/4 shared attention primitives

Model additions (Kimi K2 / GLM5 / DeepSeek-V4) follow as separate PRs on top of this foundation.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.