chore(offpolicy): remove inference request timeout config

docs/sphinx/source/en/2-user_guide/3-backends/7-newton.md

@@ -36,9 +36,9 @@ Genesis quirk) is needed; the rank-local device reaches spawn collectors as
 a `newton_device="cuda:N"` env override, and uni_rl's collector-side
 process binding is injection-based — UniLab injects
 `bind_backend_process_device_for_backend`, which covers both mjwarp and
-newton — while the SAC owner raises the
-collector tick-0 timeout to 180 s to cover Warp kernel compilation on the
-cold path.
+newton — while SAC handles learner-owned warmup before collector startup and
+cold materialization through the collector-ready handshake instead of using a
+collector tick timeout.
 
 ## Installation
 

docs/sphinx/source/zh_CN/2-user_guide/3-backends/7-newton.md

@@ -33,8 +33,8 @@ CUDA 设备语义，无需 Genesis 那样的 `CUDA_VISIBLE_DEVICES` 钉卡；ran
 本地设备通过 env override 以 `newton_device="cuda:N"` 传入 spawn
 collector，且 uni_rl 的 collector 进程绑定为注入式——UniLab 注入的
 `bind_backend_process_device_for_backend` 同时覆盖 mjwarp 与
-newton；SAC owner 将 collector tick-0 超时提升到 180 s 以覆盖 Warp
-内核编译的冷路径。
+newton；SAC 会在 collector 启动前执行 learner 侧预热，并通过
+collector-ready 握手覆盖冷物化，不再使用 collector tick 超时。
 
 ## 安装
 

src/unilab/conf/flashsac/config.yaml

@@ -90,7 +90,6 @@ training:
   env_steps_per_sync: 1
   # Seconds the learner waits for each collector inference tick after the
   # collector-ready handshake.
-  inference_request_timeout_sec: 30.0
   trace_enabled: false
   trace_output_dir: null
   trace_thread_time: false

src/unilab/conf/flashsac/task/g1_motion_tracking/genesis.yaml

@@ -18,7 +18,6 @@ training:
   play_render_mode: auto
   # First tick includes collector env construction; genesis JIT startup exceeds
   # the 30s default (same override as sac g1_walk_flat/genesis).
-  inference_request_timeout_sec: 180.0
 env:
   # GPU device id for the process-wide Genesis session. Multi-GPU training
   # overrides this cold-path field from the rank topology; zero preserves the

src/unilab/conf/flashsac/task/g1_motion_tracking/newton.yaml

@@ -15,7 +15,6 @@ training:
   sim_backend: newton
   # First tick includes collector env construction; newton startup exceeds the
   # 30s default (same override as sac g1_walk_flat/newton).
-  inference_request_timeout_sec: 180.0
 env:
   # A null device is intentional: the off-policy collector binder injects the
   # rank-local CUDA device before materialization. 320/512 leaves headroom

src/unilab/conf/sac/config.yaml

@@ -77,9 +77,6 @@ training:
   log_root: null
   log_dir: null
   env_steps_per_sync: 1
-  # Coordination window for collector inference ticks. The first learner update
-  # can include ROCm/CUDA compilation, so keep this above measured cold-start times.
-  inference_request_timeout_sec: 90.0
   trace_enabled: false
   trace_output_dir: null
   trace_thread_time: false

src/unilab/conf/sac/task/g1_motion_tracking/genesis.yaml

@@ -22,7 +22,6 @@ training:
   play_render_mode: auto
   # First tick includes collector env construction; genesis JIT startup exceeds
   # the 30s default (same override as sac g1_walk_flat/genesis).
-  inference_request_timeout_sec: 180.0
 env:
   # GPU device id for the process-wide Genesis session. Multi-GPU training
   # overrides this cold-path field from the rank topology; zero preserves the

src/unilab/conf/sac/task/g1_motion_tracking/newton.yaml

@@ -17,10 +17,6 @@ defaults:
 training:
   task_name: G1MotionTrackingSAC
   sim_backend: newton
-  # Newton compiles its Warp kernels on the materialize cold path, which
-  # exceeds the runner's default 30 s collector tick-0 budget; raise it for
-  # this backend (same precedent as the Genesis owner).
-  inference_request_timeout_sec: 180.0
 env:
   # A null device is intentional: the off-policy collector binder injects the
   # rank-local CUDA device before materialization. 320/512 leaves headroom

src/unilab/conf/sac/task/g1_walk_flat/genesis.yaml

@@ -22,10 +22,6 @@ training:
   task_name: G1WalkFlat
   sim_backend: genesis
   play_render_mode: auto
-  # Genesis compiles its kernels on the materialize cold path (~35 s at
-  # num_envs=2048 on an RTX 4090, no cross-process cache), which exceeds the
-  # runner's default 30 s collector tick-0 budget; raise it for this backend.
-  inference_request_timeout_sec: 180.0
 algo:
   num_envs: 2048
   learning_starts: 10

src/unilab/conf/sac/task/g1_walk_flat/newton.yaml

@@ -17,10 +17,6 @@ defaults:
 training:
   task_name: G1WalkFlat
   sim_backend: newton
-  # Newton compiles its Warp kernels on the materialize cold path, which
-  # exceeds the runner's default 30 s collector tick-0 budget; raise it for
-  # this backend (same precedent as the Genesis owner).
-  inference_request_timeout_sec: 180.0
 algo:
   num_envs: 2048
   learning_starts: 10

src/unilab/conf/warpsac/config.yaml

@@ -89,7 +89,6 @@ training:
   log_root: null
   log_dir: null
   env_steps_per_sync: 1
-  inference_request_timeout_sec: 30.0
   trace_enabled: false
   trace_output_dir: null
   trace_thread_time: false

tests/algos/test_offpolicy_double_buffer_runner.py

@@ -153,13 +153,6 @@ def test_sac_dispatch_constructs_unique_runner(monkeypatch: pytest.MonkeyPatch):
         "nvtx_profile_ranges": cfg.training.nvtx_profile_ranges,
         "critic_obs_dim": 6,
     }
-    assert runner.kwargs["inference_request_timeout_sec"] == 90.0
-
-
-def test_sac_genesis_owner_raises_inference_request_timeout():
-    """Genesis kernel compilation exceeds the 30 s default tick-0 budget."""
-    cfg = _offpolicy_cfg(["task=g1_walk_flat/genesis"])
-    assert cfg.training.inference_request_timeout_sec == 180.0
 
 
 def test_sac_owner_custom_runtime_can_override_base_learner_kwargs(

tests/config/test_g1_motion_tracking_owners.py

@@ -79,7 +79,6 @@ def test_sac_g1_motion_tracking_genesis_inherits_mujoco_parity() -> None:
     assert cfg.training.task_name == "G1MotionTrackingSAC"
     assert cfg.training.sim_backend == "genesis"
     assert cfg.training.play_render_mode == "auto"
-    assert cfg.training.inference_request_timeout_sec == 180.0
     assert cfg.env.genesis_device_id == 0
     assert cfg.env.genesis_integrator == "implicitfast"
     # Genesis legacy scenes declare no model-field reset terms and no interval
@@ -112,7 +111,6 @@ def test_sac_g1_motion_tracking_newton_keeps_full_dr() -> None:
     cfg = _compose_sac("g1_motion_tracking/newton")
     assert cfg.training.task_name == "G1MotionTrackingSAC"
     assert cfg.training.sim_backend == "newton"
-    assert cfg.training.inference_request_timeout_sec == 180.0
     assert cfg.env.newton_device is None
     assert cfg.env.newton_nconmax == 320
     assert cfg.env.newton_njmax == 512