Merge pull request #30 from ilyalasy/bfloat_fix

Bfloat fix

Author: curt-tigges

clt/activation_generation/generator.py

@@ -41,6 +41,8 @@
 from collections import defaultdict
 import psutil
 
+from clt.training.utils import torch_bfloat16_to_numpy_uint16
+
 try:
     import GPUtil
 except ImportError:
@@ -764,13 +766,13 @@ def _write_chunk(
 
                                 # Convert to numpy
                                 with self._conditional_measure(f"chunk_{chunk_idx}_layer_{lid}_convert_numpy"):
-                                    inp_np = inp_perm.to(self.torch_dtype).numpy()
-                                    tgt_np = tgt_perm.to(self.torch_dtype).numpy()
-
-                                # Handle bfloat16 conversion
-                                if h5py_dtype_str == "uint16" and inp_np.dtype == np.dtype("bfloat16"):
-                                    inp_np = inp_np.view(np.uint16)
-                                    tgt_np = tgt_np.view(np.uint16)
+                                    # Handle bfloat16 conversion
+                                    if h5py_dtype_str == "uint16":
+                                        inp_np = torch_bfloat16_to_numpy_uint16(inp_perm)
+                                        tgt_np = torch_bfloat16_to_numpy_uint16(tgt_perm)
+                                    else:
+                                        inp_np = inp_perm.to(self.torch_dtype).numpy()
+                                        tgt_np = tgt_perm.to(self.torch_dtype).numpy()
 
                                 # Store prepared data
                                 layer_data[lid] = (inp_np, tgt_np)

clt/training/utils.py

@@ -1,5 +1,8 @@
 import datetime
 
+import numpy as np
+import torch
+
 
 # Helper function to format elapsed time
 def _format_elapsed_time(seconds: float) -> str:
@@ -11,3 +14,7 @@ def _format_elapsed_time(seconds: float) -> str:
         return f"{td.days * 24 + hours:02d}:{minutes:02d}:{seconds:02d}"
     else:
         return f"{minutes:02d}:{seconds:02d}"
+
+
+def torch_bfloat16_to_numpy_uint16(x: torch.Tensor) -> np.ndarray:
+    return np.frombuffer(x.float().numpy().tobytes(), dtype=np.uint16)[1::2].reshape(x.shape)