Add expert-parallelism example: SFT of 100B-753B MoE models

[qgallouedec]

Companion to #6846's 1M-context example, for the other axis: model size. Two things:

• a "Training MoE Models at the 100B–753B Scale" section in docs/source/distributing_training.md: expert parallelism via DistributedConfig(tp_size=E, fsdp_size=D, enable_expert_parallel=True), the verified configurations, and what to expect operationally (loading, saving, memory);
• examples/sft_glm_4_5_air_full_finetune/ (110B, every parameter) and examples/sft_glm_5_2_expert_parallel/ (753B, LoRA): one folder per training, each with a fixed-config script and its SLURM launcher.

Measured

H100 nodes, seq 2048, bf16, per-device batch 1, default chunked_nll loss, gradient checkpointing:

(Step times are update cadence, not 64-GPU throughput: the expert-parallel group shares one batch, so an ep=32 x fsdp=2 step processes 2 sequences / ~4k tokens.)

Model	Training	GPUs	Config	Step time	Peak GPU memory
Qwen3-30B-A3B	full FT	8 (1 node)	ep=8	0.9 s	34 GB
GLM-4.5-Air (110B)	full FT	64 (8 nodes)	ep=32 × fsdp=2	3.7 s	41 GB
GLM-4.6 (357B)	LoRA	16 (2 nodes)	ep=16	4.4 s	73 GB
GLM-5.2 (753B)	LoRA	64 (8 nodes)	ep=32 × fsdp=2	3.1 s	56 GB

Real training, not just "it fits": GLM-5.2 loss 3.3 → 2.3 in 50 steps on tulu-3; Air full FT 3.35 → 2.13 in 20 steps, including the full 206 GiB model save.

Assumptions

The example needs branches that have not landed yet:

• [Umbrella] Enable FSDP2 + expert parallelism via a 2-D (fsdp, tp) device mesh transformers#48204 — FSDP2 × expert parallelism (the 2-D mesh itself)
• [Umbrella] FSDP2 × expert-parallelism fixes (do not merge) accelerate#4178 — umbrella of three accelerate fixes
• Fix LoRA on FSDP2-sharded models: shape inference and module hooks peft#3578 — LoRA on FSDP2-sharded models (LoRA configs only)

Until they do, install:

pip install git+https://github.com/huggingface/transformers@ep-fsdp-2d-mesh
pip install git+https://github.com/huggingface/accelerate@fsdp2-ep-integration
pip install git+https://github.com/huggingface/peft@fsdp2-dtensor-fixes

TRL itself needs nothing beyond main. Draft until the branches above land or the install block is confirmed.