[jinyouzhi]

Adds training/deepspeed_finetune_demo/autoep_rollout_e2e.py and run_autoep_rollout_e2e.sh, which check precision along HF checkpoint -> AutoEP ZeRO training -> universal checkpoint or AffineIR transfer -> rollout.

1. Load a truncated HF checkpoint (Qwen3-30B-A3B, 4 layers) into an AutoEP training engine (EP4). Every weight is compared bitwise with the safetensors tensors.
2. Train 3 steps.
3. Move the trained weights to NaN-initialized rollout engines at EP8:
   • ucp: save_checkpoint -> ds_to_universal -> load_checkpoint(load_universal);
   • affine: AutoEP storage maps -> plan_transfer -> execute_transfers (experts, device to device); dense weights copied on device.
4. Generate greedily with HybridEngineRollout; teacher-force the other engines on the same sequences.

The README adds a section with the weight-flow diagrams for both paths and the in-place HybridEngineRollout, a comparison table, and measured times.

Depends on deepspeedai/DeepSpeed#8800 (AutoEP storage maps) and deepspeedai/DeepSpeed#8809 (AffineIR transfer executor).

Results (8x Intel Arc Pro B70, Qwen3-30B-A3B truncated to 4 layers, EP4 -> EP8)

Check	ZeRO-2	ZeRO-3
HF weights vs training engine	bitwise, 10.4e9 elements	bitwise, 10.4e9 elements
UCP rollout weights vs trained	bitwise, 8.0e9 elements	not run (1)
Affine rollout weights vs trained	bitwise, 8.0e9 elements	bitwise, 8.0e9 elements
UCP vs Affine logits	bitwise	-
Rollout vs training engine logits	max diff 0.68, 16/1240 argmax differ	max diff 0.67, 32/1248 argmax differ

The last row is reported, not required: EP4 and EP8 group tokens differently in the grouped GEMM, so the bf16 accumulation order differs even with identical weights.

Step	UCP, ZeRO-2	Affine, ZeRO-2	Affine, ZeRO-3
save / plan	7.1 s	3.2 s	0.05 s
convert / execute	107.7 s	0.27 s	0.23 s
load / dense copy	25.1 s	0.01 s	0.18 s

(1) A universal checkpoint converted from ZeRO-3 keeps ZeRO-3 model file names, which the ZeRO-1 rollout engine cannot load.

• ZeRO-3 training with AutoEP can hang from the third step: the ZeRO-3 prefetcher passes a set to all_gather_coalesced, which issues the dense and expert-group all-gathers in an order that can differ across ranks. The ZeRO-3 numbers above were taken with a local fix (sort by ds_id before grouping); ([AutoEP] Ensure consistent order of partition groups in all-gather when ZeRO-3  DeepSpeed#8876).