Fix residual GPU memory usage in TrainingWorker in RL workflow

[hhaAndroid]

No description provided.

[hhaAndroid]

@claude review

[claude]

Claude: ## Summary

This PR fixes activation offload memory leaks by expanding OffloadManager.clear() to also clean may_npu_tensors and getcnt, adds a group-scoped clear mode, and introduces cuBLAS workspace clearing during model offload. It also updates the example config to enable EP=4, FP8, and reduce MTP layers/draft tokens.

Issues

Warning

• xtuner/v1/utils/activation_offload.py:321-325 device_item is not cleared in the full-clear path. If it accumulates GPU tensor references across iterations this could still leak.
• xtuner/v1/rl/trainer/worker.py:953-963 _clear_cublas_workspaces uses an internal PyTorch API (torch._C._cuda_clearCublasWorkspaces) that may disappear across versions.
• xtuner/v1/utils/activation_offload.py:298-303 Behavioral change: clear(key=...) now silently discards missing keys instead of raising via assert_exist.

Nit

• examples/v1/config/reasoning_rl_qwen3p5vl_mtp_ep.py:64 Missing spaces around = in train_ep_size=4.
• examples/v1/config/reasoning_rl_qwen3p5vl_mtp_ep.py:220 Dead commented-out code (# model_cfg.float8_cfg = None).

Verdict

COMMENT — the core fix is correct and well-structured, but please clarify whether device_item should also be cleared, and whether the silent-discard change is intentional.