From: Jérôme Benoit <jerome.benoit@piment-noir.org>
Date: Sun, 15 Feb 2026 12:31:58 +0000 (+0100)
Subject: feat(ReforceXY): add gpu_memory_fraction tunable
X-Git-Url: https://git.piment-noir.org/?a=commitdiff_plain;h=939f1343e3f217d622aeca17f250762aee0400ed;p=freqai-strategies.git

feat(ReforceXY): add gpu_memory_fraction tunable
---

diff --git a/ReforceXY/user_data/freqaimodels/ReforceXY.py b/ReforceXY/user_data/freqaimodels/ReforceXY.py
index baeed3b..b62f707 100644
--- a/ReforceXY/user_data/freqaimodels/ReforceXY.py
+++ b/ReforceXY/user_data/freqaimodels/ReforceXY.py
@@ -109,6 +109,10 @@ class ReforceXY(BaseReinforcementLearningModel):
         ...
         "freqai": {
             ...
+            "model_training_parameters": {
+                "device": "auto",                   // PyTorch device (auto|cpu|cuda|cuda:0)
+                "gpu_memory_fraction": null,        // GPU VRAM fraction limit per process (0.0, 1.0], null disables
+            },
             "rl_config": {
                 ...
                 "n_envs": 1,                        // Number of DummyVecEnv or SubProcVecEnv training environments
@@ -328,6 +332,35 @@ class ReforceXY(BaseReinforcementLearningModel):
             ],
         ] = {}
         self.unset_unsupported()
+        self._configure_gpu_memory()
+
+    def _configure_gpu_memory(self) -> None:
+        """
+        Configure GPU memory fraction limit from model_training_parameters.
+        Called after config validation, before any CUDA operations.
+        """
+        gpu_memory_fraction: Optional[float] = self.model_training_parameters.get(
+            "gpu_memory_fraction"
+        )
+        if gpu_memory_fraction is None:
+            return
+        if not th.cuda.is_available():
+            logger.warning(
+                "Config [global]: gpu_memory_fraction=%.2f ignored; CUDA not available",
+                gpu_memory_fraction,
+            )
+            return
+        if not 0.0 < gpu_memory_fraction <= 1.0:
+            logger.warning(
+                "Config [global]: gpu_memory_fraction=%.2f invalid; must be in (0.0, 1.0]",
+                gpu_memory_fraction,
+            )
+            return
+        th.cuda.set_per_process_memory_fraction(gpu_memory_fraction, device=0)
+        logger.info(
+            "Config [global]: gpu_memory_fraction=%.2f applied",
+            gpu_memory_fraction,
+        )
 
     @staticmethod
     def _normalize_position(position: Any) -> Positions:
@@ -356,7 +389,7 @@ class ReforceXY(BaseReinforcementLearningModel):
         )
         if cache_key in ReforceXY._action_masks_cache:
             logger.debug(
-                "ActionMask: cache hit for can_short=%s position=%s",
+                "ActionMask [global]: cache hit for can_short=%s position=%s",
                 can_short,
                 position.name,
             )
@@ -376,7 +409,7 @@ class ReforceXY(BaseReinforcementLearningModel):
 
         ReforceXY._action_masks_cache[cache_key] = action_masks
         logger.debug(
-            "ActionMask: cache miss for can_short=%s position=%s, computed=%s",
+            "ActionMask [global]: cache miss for can_short=%s position=%s; computed=%s",
             can_short,
             position.name,
             action_masks,