Merge pull request #3727 from AI-Hypercomputer:aireen/vocab_dtype

PiperOrigin-RevId: 904656787

Author: Google-ML-Automation

src/maxtext/models/gemma4.py

@@ -299,7 +299,7 @@ def __init__(
     else:
       self.post_ffw_norm = None
 
-    self.layer_scalar = nnx.Param(jnp.ones((1,), dtype=config.dtype), sharding=(None,))
+    self.layer_scalar = nnx.Param(jnp.ones((1,), dtype=config.weight_dtype), sharding=(None,))
 
     if model_mode == MODEL_MODE_PREFILL:
       self.activation_axis_names = ("activation_batch", "prefill_activation_norm_length", "activation_embed")

src/maxtext/utils/vocabulary_tiling.py

@@ -224,10 +224,10 @@ 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)
+    # Cast cotangents back to each primal's dtype; custom_vjp requires dtype match.
+    grad_params = jax.tree_util.tree_map(lambda x, y: y.astype(x.dtype), gathered_params, grad_params)
     # Give back sharding constraint
     grad_reshaped_hidden_states = _reshape(grad_reshaped_hidden_states, (batch_size, seq_len, emb_dim), hidden_spec)
     return (

tests/unit/train_compile_test.py

@@ -955,3 +955,22 @@ def test_qk_clip(self):
             "qk_clip_threshold=100",
         )
     )
+
+  @pytest.mark.cpu_only
+  def test_vocab_tiling_bf16(self):
+    """test vocab_tiling when weight_dtype=bfloat16"""
+    compiled_trainstep_file = "/tmp/test_vocab_tiling_bf16.pickle"
+    train_compile_main(
+        (
+            "",
+            get_test_config_path(),
+            f"compiled_trainstep_file={compiled_trainstep_file}",
+            "compile_topology=v5p-8",
+            "compile_topology_num_slices=1",
+            "base_num_decoder_layers=2",
+            "per_device_batch_size=2",
+            "max_target_length=1024",
+            "num_vocab_tiling=4",
+            "weight_dtype=bfloat16",
+        )
+    )