Merge branch 'master' into patch-3

Author: amorehead

docs/source-pytorch/common/precision_basic.rst

@@ -39,6 +39,14 @@ However, this setting can sometimes lead to unstable training.
 
     Trainer(precision="16-true")
 
+.. warning::
+
+    Float16 cannot represent values smaller than ~6e-5. Values like Adam's default ``eps=1e-8`` become zero, which can cause
+    NaN during training. Increase ``eps`` to 1e-4 or higher, and avoid extremely small values in your model weights and data.
+
+.. note::
+
+    BFloat16 (``"bf16-mixed"`` or ``"bf16-true"``) has better numerical stability with a wider dynamic range.
 
 ----
 

requirements/doctests.txt

@@ -1,2 +1,2 @@
 pytest ==8.4.2
-pytest-doctestplus ==1.4.0
+pytest-doctestplus ==1.5.0

requirements/pytorch/extra.txt

@@ -5,7 +5,7 @@
 matplotlib>3.1, <3.11.0
 omegaconf >=2.2.3, <2.4.0
 hydra-core >=1.2.0, <1.4.0
-jsonargparse[signatures,jsonnet] >=4.39.0, <4.42.0
+jsonargparse[signatures,jsonnet] >=4.39.0, <4.43.0
 rich >=12.3.0, <14.2.0
 tensorboardX >=2.2, <2.7.0  # min version is set by torch.onnx missing attribute
 bitsandbytes >=0.45.2,<0.47.0; platform_system != "Darwin"

requirements/pytorch/test.txt

@@ -13,7 +13,7 @@ numpy >1.20.0, <1.27.0
 onnx >1.12.0, <1.20.0
 onnxruntime >=1.12.0, <1.24.0
 onnxscript >= 0.1.0, < 0.5.0
-psutil <7.1.1 # for `DeviceStatsMonitor`
+psutil <7.1.2 # for `DeviceStatsMonitor`
 pandas >2.0, <2.4.0  # needed in benchmarks
 fastapi  # for `ServableModuleValidator`  # not setting version as re-defined in App
 uvicorn  # for `ServableModuleValidator`  # not setting version as re-defined in App