Update experimental Megatron Lite package

[ISEEKYAN]

Summary

• Update the experimental Megatron Lite package from the latest BB Lite cleanup.
• Keep the runtime backend name as mlite while keeping model implementations named lite.
• Replace Qwen3.5 with the native lite implementation, add FSDP2 optimizer primitives, and add a runnable VERL SFT example.
• Pin the VERL example requirement to the latest GitHub release tag, verl-project/verl@v0.8.0.
• Support both FSDP2 offload modes:
  • runtime context offload through PARAM_OFFLOAD / OPTIMIZER_OFFLOAD, which moves model parameters and optimizer state between CPU and GPU when VERL switches modes;
  • optimizer update-state offload through offload_fraction=1.0, which keeps FSDP2 optimizer update state on CPU during forward/backward.

Implementation Notes

• All new user-facing code remains under experimental/lite, with imports exposed through megatron.lite when users add experimental/lite to PYTHONPATH.
• The VERL example uses backend=mlite; model implementations remain impl=lite.
• experimental/lite/examples/verl/REQUIRED_VERL.txt is now requirements-style and pins verl @ git+https://github.com/verl-project/verl.git@v0.8.0.
• FSDP2 model offload uses module-level device movement instead of direct param.data reassignment, which avoids invalid DTensor/FSDP2 storage transitions.
• FSDP2 optimizer runtime offload now includes DTensor optimizer state by default; offload_fraction=1.0 remains a separate CPU-update path.

Validation

• Compile:
  • PYTHONPATH=mlite/experimental/lite python -m compileall -q mlite/experimental/lite/megatron/lite/runtime/megatron_utils.py mlite/experimental/lite/examples/verl/verl_mlite/engine/mlite_engine.py mlite/experimental/lite/tests/unit/primitive/test_fsdp2_offload_gpu.py
  • Earlier package compile/import checks for experimental/lite/megatron/lite and experimental/lite/examples.
• VERL release pin:
  • GitHub verl-project/verl latest release checked on 2026-06-06: v0.8.0.
• FSDP2 offload unit tests:
  • Local no-CUDA run skips as expected.
  • Slurm single-node single-GPU container job 12579939: 2 passed
    • test_fsdp2_runtime_model_and_optimizer_offload_roundtrip_single_gpu
    • test_fsdp2_offload_fraction_keeps_optimizer_update_state_on_cpu_single_gpu
• 100-step VERL SFT validation from this branch before the offload follow-up:
  • Qwen3-30B-A3B, engine=mlite, impl=lite, optimizer=fsdp2, 4 nodes x 8 GPUs, job 12537468, rc=0, step100 train/loss 0.0958933, val/loss 0.4291660.
  • Qwen3.5-35B-A3B, engine=mlite, impl=lite, optimizer=fsdp2, 4 nodes x 8 GPUs, job 12545533, rc=0, step100 train/loss 0.2139087, val/loss 0.3420881.

Merge Readiness

• Ready for review on devlite.
• The FSDP2 offload change is covered by focused single-GPU tests for both offload modes.
• If reviewers want full system coverage for runtime PARAM_OFFLOAD=True, the next validation should be a 100-step Qwen3.5 VERL SFT run with both PARAM_OFFLOAD=True and OPTIMIZER_OFFLOAD=True.