feat(flashsac): expose learner CUDA Graph controls

docs/sphinx/source/en/2-user_guide/2-algorithms/4-flash_sac.md

@@ -30,6 +30,28 @@ playback video. See {doc}`/en/1-getting_started/3-evaluation_and_playback`.
 - `algo.algo_params.actor_num_blocks=2`
 - `algo.algo_params.critic_num_blocks=2`
 
+## Learner acceleration controls
+
+The default single-NVIDIA learner enables all three acceleration layers. They
+remain separate configuration choices:
+
+- `algo.algo_params.use_compile` controls CUDA Inductor compilation. Set it to
+  `false` for the eager per-update compatibility path; this also disables the
+  dependent whole-cycle graph.
+- `algo.algo_params.compile_full_objectives` selects complete actor/critic
+  objectives instead of the smaller loss-tensor compile regions. It does not
+  control CUDA Graph capture.
+- `algo.algo_params.use_whole_cycle_cuda_graph` controls learner-owned
+  whole-cycle CUDA Graph capture. Set only this field to `false` to retain
+  Inductor compilation without whole-cycle capture.
+
+For example, run the eager debugging path with:
+
+```bash
+uv run train --algo flashsac --task g1_walk_flat --sim mujoco \
+  algo.algo_params.use_compile=false
+```
+
 FlashSAC requires synchronized collection and the same sole replay path as SAC:
 bounded host ingress plus one complete replay ring on a CUDA or Apple
 MPS learner device. CPU and XPU training are unsupported.

docs/sphinx/source/zh_CN/2-user_guide/2-algorithms/4-flash_sac.md

@@ -28,6 +28,24 @@ uv run train --algo flashsac --task go2_joystick_flat --sim mujoco training.no_p
 - `algo.algo_params.actor_num_blocks=2`
 - `algo.algo_params.critic_num_blocks=2`
 
+## Learner 加速开关
+
+单 NVIDIA learner 默认开启全部三层加速，但它们是三个独立的配置选择：
+
+- `algo.algo_params.use_compile` 控制 CUDA Inductor 编译。设为 `false` 后走
+  eager 的逐 update 兼容路径，同时也会关闭依赖编译的 whole-cycle graph。
+- `algo.algo_params.compile_full_objectives` 选择编译完整 actor/critic objective，
+  还是较小的 loss-tensor 区域；它不控制 CUDA Graph 捕获。
+- `algo.algo_params.use_whole_cycle_cuda_graph` 控制 learner 自己捕获并回放整轮
+  CUDA Graph。只把此字段设为 `false`，可以保留 Inductor 编译而关闭整轮捕获。
+
+例如，使用 eager 调试路径：
+
+```bash
+uv run train --algo flashsac --task g1_walk_flat --sim mujoco \
+  algo.algo_params.use_compile=false
+```
+
 FlashSAC 要求同步采集，并与 SAC 共用唯一 replay 路径：有界 host ingress 加
 一个驻留在 CUDA 或 Apple MPS learner device 上的完整 replay ring。CPU 与 XPU
 training 不受支持。

src/unilab/conf/flashsac/config.yaml

@@ -46,6 +46,7 @@ algo:
     amp_dtype: auto
     use_compile: true
     compile_full_objectives: true
+    use_whole_cycle_cuda_graph: true
 
 training:
   task_name: G1WalkFlat

src/unilab/structured_configs.py

@@ -83,6 +83,8 @@ class FlashSACAlgoParams:
     n_step: int = 1
     amp_dtype: str = "auto"
     use_compile: bool = True
+    compile_full_objectives: bool = True
+    use_whole_cycle_cuda_graph: bool = True
 
 
 @dataclass

tests/config/test_locomotion_params.py

@@ -46,6 +46,8 @@ def test_flashsac_config_defaults():
     assert cfg.algo_params.normalize_reward is True
     assert cfg.algo_params.amp_dtype == "auto"
     assert cfg.algo_params.use_compile is True
+    assert cfg.algo_params.compile_full_objectives is True
+    assert cfg.algo_params.use_whole_cycle_cuda_graph is True
 
 
 def test_ppo_config_defaults():
@@ -125,6 +127,9 @@ def test_offpolicy_flashsac_g1_task_overrides():
             overrides=["task=g1_walk_flat/mujoco"],
         )
     assert cfg.algo.algo == "flashsac"
+    assert cfg.algo.algo_params.use_compile is True
+    assert cfg.algo.algo_params.compile_full_objectives is True
+    assert cfg.algo.algo_params.use_whole_cycle_cuda_graph is True
 
 
 def test_offpolicy_flashsac_go2_task_overrides():