update grad dtype

Author: NuojCheng

src/MaxText/vocabulary_tiling.py

@@ -21,9 +21,11 @@
 import jax
 import jax.numpy as jnp
 from MaxText import max_utils
+from MaxText import max_logging
 from MaxText.sharding import maybe_shard_with_name, all_gather_over_fsdp
 from MaxText.common_types import ShardMode
 
+max_logging.log("Using new new vocab tiling file!!!")
 
 def vocab_tiling_linen_loss(
     hidden_states,
@@ -198,15 +200,13 @@ def _bwd_scan_body(grad_params_acc, chunk_data):
         _bwd_scan_body, initial_grad_params_acc, (reshaped_hidden_states, reshaped_labels, reshaped_segmentation)
     )
     grad_reshaped_hidden_states = _maybe_shard_with_name(grad_reshaped_hidden_states, reshaped_hidden_spec)
-    # TODO (chengnuojin): we may want to convert grad_params to bf16 to save memory
-    # grad_params = jax.tree_util.tree_map(lambda x, y: y.astype(x.dtype), gathered_params, grad_params)
     # Chain-rule to accumulate gradients
-    grad_params = jax.tree_util.tree_map(lambda g: g * loss_cotangent, grad_params)
+    grad_params = jax.tree_util.tree_map(lambda g, p: (g * loss_cotangent).astype(p.dtype), grad_params, gathered_params)
     # Give back sharding constraint
-    grad_reshaped_hidden_states = _reshape(grad_reshaped_hidden_states, (batch_size, seq_len, emb_dim), hidden_spec)
+    grad_hidden_states = _reshape(grad_reshaped_hidden_states, (batch_size, seq_len, emb_dim), hidden_spec)
     return (
         grad_params,  # grad for params
-        grad_reshaped_hidden_states.astype(reshaped_hidden_states.dtype),
+        grad_hidden_states.astype(config.weight_dtype),  # Enforce activation gradients dtype as 
         None,  # grad for reshaped_labels
         None,  # grad for reshaped_segmentation
     )