doc

Author: OutisLi

deepmd/pt/model/model/sezm_model.py

@@ -2845,9 +2845,10 @@ def deserialize(cls, data: dict[str, Any]) -> SeZMModel:
     def tf32_precision_ctx(self) -> Generator[None, None, None]:
         """Context manager to temporarily set TF32 matmul precision.
 
-        Training follows ``enable_tf32``. Eval/inference follows
-        ``DP_TF32_INFER``: 0 keeps ``highest`` precision, 1 selects
-        ``high``, and 2 selects ``medium``.
+        Training follows ``enable_tf32`` independently of whether the current
+        forward uses the compile path. Eval/inference follows ``DP_TF32_INFER``:
+        0 keeps ``highest`` precision, 1 selects ``high``, and 2 selects
+        ``medium``.
         """
         if not torch.cuda.is_available():
             yield

deepmd/utils/argcheck.py

@@ -3170,7 +3170,12 @@ def sezm_model_args() -> Argument:
         "Requires torch==2.11. NVIDIA GPUs require CUDA >= 12.6. "
         "Apple Silicon Macs are also supported. Tested with Python 3.13."
     )
-    doc_enable_tf32 = "If True, enable TF32 matmul precision when use_compile=True."
+    doc_enable_tf32 = (
+        "If True, enable TF32 matmul precision for CUDA training forwards. "
+        "This training-time setting is independent of `use_compile`; eval-time "
+        "TF32 is controlled separately by `validating.tf32_infer` or "
+        "`DP_TF32_INFER`."
+    )
     doc_bridging_method = (
         "Short-range bridging method. Currently supports 'ZBL'. "
         "The value is case-insensitive; set it to 'None' to disable bridging."

doc/model/dpa4.md

@@ -456,7 +456,8 @@ During training validation, the input option
 `validating.tf32_infer: true` is translated into `DP_TF32_INFER=1` before
 model construction, again without overriding an explicitly exported
 environment variable. Training forwards are controlled separately by
-`model.enable_tf32`.
+`model.enable_tf32`, independently of whether `model.use_compile` selects the
+compiled or eager training path.
 
 For molecular dynamics and other workflows that are sensitive to potential
 energy surface smoothness, keep `DP_TF32_INFER=0`. Enabling TF32 inference may