logging improvements (#1808) [skip ci]

* logging improvements

* fix sort

src/axolotl/utils/data/sft.py

@@ -187,7 +187,12 @@ def load_tokenized_prepared_datasets(
     else:
         if cfg.push_dataset_to_hub:
             LOG.info("Unable to find prepared dataset in Huggingface hub")
-        LOG.info(f"Unable to find prepared dataset in {prepared_ds_path}")
+        if cfg.is_preprocess:
+            LOG.info(
+                f"Skipping prepared dataset in {prepared_ds_path} for pre-processing..."
+            )
+        else:
+            LOG.info(f"Unable to find prepared dataset in {prepared_ds_path}")
         LOG.info("Loading raw datasets...")
         if not cfg.is_preprocess:
             LOG.warning(

src/axolotl/utils/distributed.py

@@ -153,11 +153,11 @@ def compute_and_broadcast(fn):  # pylint: disable=invalid-name
     if is_main_process():
         value_scalar = fn()
         value_tensor = torch.tensor(
-            value_scalar, device=torch.cuda.current_device()
-        ).float()
+            value_scalar, device=torch.cuda.current_device(), dtype=torch.float32
+        )
     else:
         value_tensor = torch.tensor(
-            0.0, device=torch.cuda.current_device()
+            0.0, device=torch.cuda.current_device(), dtype=torch.float32
         )  # Placeholder tensor
 
     # Broadcast the tensor to all processes.

src/axolotl/utils/models.py

@@ -1022,7 +1022,7 @@ def load_lora(model, cfg, inference=False, config_only=False):
 
     if cfg.lora_target_linear:
         linear_names = find_all_linear_names(model)
-        LOG.info(f"found linear modules: {repr(linear_names)}")
+        LOG.info(f"found linear modules: {repr(sorted(linear_names))}")
         lora_target_modules = list(set(lora_target_modules + linear_names))
 
     lora_config_kwargs = {}