drop upper bounds in dependencies (#2022)

Author: t-vi

litgpt/pretrain.py

@@ -228,6 +228,19 @@ def main(
         fabric.load(resume, state)
 
     train_time = time.perf_counter()
+
+    # work around PyTorch issue https://github.com/pytorch/pytorch/issues/152162
+    # which does not like the lazy initialization to be called in dynamo.
+    # Happens with PyTorch 2.7.
+    if (
+        torch.__version__.startswith("2.7.")
+        and (model._forward_module.__class__.__name__ == "OptimizedModule")
+        and (model._forward_module._orig_mod.__class__.__name__ == "FullyShardedDataParallel")
+    ):
+        from torch.distributed.fsdp._runtime_utils import _root_pre_forward
+
+        _root_pre_forward(model._forward_module._orig_mod, model._forward_module._orig_mod, [], {})
+
     fit(
         fabric=fabric,
         devices=devices,

pyproject.toml

@@ -29,13 +29,13 @@ dependencies = [
   "huggingface-hub>=0.23.5",
   "jsonargparse[signatures]>=4.30.1,<=4.32.1; python_version<='3.9'", # 4.33 does not seem to be compatible with Python 3.9
   "jsonargparse[signatures]>=4.37; python_version>'3.9'",             # required to work with python3.12+
-  "lightning>=2.5,<2.6",
+  "lightning>=2.5",
   "numpy<2",                                                          # for older Torch versions
   "psutil==7",
   "safetensors>=0.4.3",
   # tokenization in most models:
   "tokenizers>=0.15.2",
-  "torch>=2.5,<2.7",
+  "torch>=2.5",
   # convert_hf_checkpoint
   "tqdm>=4.66",
 ]