[Kuriharamio]

Summary

• Enable conservative torch.compile support for FastSAC and FlashSAC training hot paths.
• Add algo.algo_params.amp_dtype for SAC/FlashSAC with auto/fp16/bf16 support.
• Allow FlashSAC double-buffer training to run with explicit training.use_amp=true and algo.algo_params.use_compile=true.

Changes

• FastSAC:

  • Compile only _critic_loss_tensors and _actor_loss_tensors.
  • Disable CUDA Graph capture via triton.cudagraphs=false.
  • Add configurable AMP dtype; CUDA bf16 skips GradScaler.

• FlashSAC:

  • Remove double-buffer guards that rejected AMP and compile.
  • Compile actor.get_mean_and_std, _critic_loss_tensors, and _actor_loss_tensors.
  • Add configurable AMP dtype; CUDA bf16 skips GradScaler.

• Config/tests:

  • Add amp_dtype: auto to SAC and FlashSAC algo params.
  • Add config, compile-target, AMP dtype, and runner pass-through tests.

Validation

make test-all

CUDA/MuJoCo smoke:

uv run train --algo=sac --task=g1_walk_flat --sim mujoco \
  training.use_amp=true \
  algo.algo_params.use_compile=true \
  algo.algo_params.amp_dtype=bf16 \
  algo.max_iterations=1 \
  training.no_play=true

uv run train --algo=flashsac --task=g1_walk_flat --sim mujoco \
  algo.max_iterations=1 \
  training.use_amp=true \
  algo.algo_params.use_compile=true \
  algo.algo_params.amp_dtype=bf16 \
  training.no_play=true

[TATP-233]

请等我验证