Integrating b/v blocking for linear softmax cross entropy loss in maxtext

PiperOrigin-RevId: 886387786

Author: Google-ML-Automation

src/maxtext/configs/base.yml

@@ -567,10 +567,12 @@ num_slices: -1
 
 # Vocab Tiling Configs
 # Enables a memory-saving optimization by computing the cross-entropy loss in chunks.
-# The logits are tiled into `num_vocab_tiling` parts along the batch-sequence axis,
-# reducing peak memory usage. This is highly recommended for models with large
-# vocabularies (e.g., Gemma). Set to a value greater than 1 to enable.
+# The logits are tiled into `num_vocab_tiling` parts along the vocabulary axis,
+# and `num_of_batch_tiling` parts along the batch-sequence axis, reducing peak memory usage.
+# This is highly recommended for models with large vocabularies (e.g., Gemma).
+# Set to a value greater than 1 to enable.
 num_vocab_tiling: 1
+num_of_batch_tiling: 1
 
 # Tokenizer
 vocab_size: 32_000 # powers of 2 for sharding

src/maxtext/configs/pyconfig_deprecated.py

@@ -194,9 +194,20 @@ def validate_expert_shard_attention_option(expert_shard_attention_option: str) -
     )
 
 
-def validate_vocab_tiling(num_vocab_tiling: int, per_device_batch_size: int, max_target_length: int, enable_nnx: bool):
+def validate_vocab_tiling(
+    num_vocab_tiling: int,
+    num_of_batch_tiling: int,
+    per_device_batch_size: int,
+    max_target_length: int,
+    enable_nnx: bool,
+):
   if (per_device_batch_size * max_target_length) % num_vocab_tiling != 0:
     raise ValueError("Per device batch size times sequence length should be divisible by the number of vocab tiles.")
+  if (per_device_batch_size * max_target_length) % num_of_batch_tiling != 0:
+    raise ValueError(
+        "Per device batch size times sequence length should be divisible by the"
+        " number of batch tiles."
+    )
   if num_vocab_tiling > 1 and enable_nnx:  # TODO (chengnuojin) enable vocab tiling on NNX after NNX migration
     raise ValueError("We currently don't support vocab tiling on NNX module.")
 

src/maxtext/configs/types.py

@@ -953,7 +953,17 @@ class Tokenizer(BaseModel):
   )
   num_vocab_tiling: int = Field(
       1,
-      description="Enables memory-saving optimization by tiling cross-entropy loss computation. >1 to enable.",
+      description=(
+          "Enables memory-saving optimization by tiling cross-entropy loss"
+          " computation along the vocabulary axis. >1 to enable."
+      ),
+  )
+  num_of_batch_tiling: int = Field(
+      1,
+      description=(
+          "Enables memory-saving optimization by tiling cross-entropy loss"
+          " computation along the batch-sequence axis. >1 to enable."
+      ),
   )
 
 
@@ -2461,6 +2471,16 @@ def calculate_global_batch_sizes(per_device_batch_size, expansion_factor, num_de
         and (self.per_device_batch_size * self.max_target_length) % self.num_vocab_tiling != 0
     ):
       raise ValueError("Per device batch size times sequence length should be divisible by the number of vocab tiles.")
+    if (
+        self.per_device_batch_size > 0
+        and (self.per_device_batch_size * self.max_target_length)
+        % self.num_of_batch_tiling
+        != 0
+    ):
+      raise ValueError(
+          "Per device batch size times sequence length should be divisible by"
+          " the number of batch tiles."
+      )
     if self.num_vocab_tiling > 1 and self.enable_nnx:
       raise ValueError("We currently don't support vocab tiling on NNX module.")
     if self.context_parallel_size > 1 and self.context_parallel_strategy.lower() == "ring":

src/maxtext/layers/decoders.py

@@ -678,9 +678,8 @@ def _apply_embedding(
     return y
 
   @nn.compact
-  def apply_output_head(self, shared_embedding: nn.Module | nnx.Module, y, deterministic, model_mode):
-    """Applies final normalization and projects hidden states to logits."""
-
+  def normalize_hidden_states(self, y, deterministic, model_mode):
+    """Applies final normalization and dropout to hidden states."""
     cfg = self.config
     if cfg.shard_mode == ShardMode.EXPLICIT:
       norm_out_sharding = create_sharding(self.mesh, ("activation_batch", "activation_length_no_exp", "activation_embed"))
@@ -696,6 +695,20 @@ def apply_output_head(self, shared_embedding: nn.Module | nnx.Module, y, determi
         parameter_memory_host_offload=cfg.parameter_memory_host_offload,
     )(y, out_sharding=norm_out_sharding)
     y = nn.Dropout(rate=cfg.dropout_rate, broadcast_dims=(-2,))(y, deterministic=deterministic)
+    return y
+
+  @nn.compact
+  def apply_output_head(
+      self,
+      shared_embedding: nn.Module | nnx.Module,
+      y,
+      deterministic,
+      model_mode,
+  ):
+    """Applies final normalization and projects hidden states to logits."""
+
+    cfg = self.config
+    y = self.normalize_hidden_states(y, deterministic, model_mode)
 
     if model_mode in (MODEL_MODE_PREFILL, MODEL_MODE_AUTOREGRESSIVE):
       out_sharding = create_sharding(self.mesh, (None, None, "activation_vocab"))

src/maxtext/models/models.py

@@ -118,6 +118,14 @@ def logits_from_hidden_states(self, hidden_states, deterministic, model_mode):
     )
     return logits
 
+  def normalize_hidden_states(self, hidden_states, deterministic, model_mode):
+    """Normalize hidden states (wrapping decoder.normalize_hidden_states)."""
+    return self.decoder.normalize_hidden_states(
+        y=hidden_states,
+        deterministic=deterministic,
+        model_mode=model_mode,
+    )
+
   def __call__(
       self,
       decoder_input_tokens: jnp.ndarray,

src/maxtext/utils/vocabulary_tiling.py

@@ -115,6 +115,153 @@ def chunked_cross_entropy_loss(gathered_params, hidden_states, labels, segmentat
     (total_loss, total_z_loss), _ = _chunked_cross_entropy_loss_fwd(gathered_params, hidden_states, labels, segmentation)
     return total_loss, total_z_loss
 
+  def _b_v_chunked_cross_entropy_loss_fwd(
+      gathered_params, hidden_states, labels, segmentation
+  ):
+    batch_size, seq_len, emb_dim = hidden_states.shape
+    v_dim = config.vocab_size
+
+    b_dim = batch_size * seq_len
+    b_block_sz = b_dim // config.num_of_batch_tiling
+    v_block_sz = v_dim // config.num_vocab_tiling
+
+    if b_dim % b_block_sz != 0 or v_dim % v_block_sz != 0:
+      raise ValueError(
+          "Batch/sequence dimension and vocab dimension must be divisible by"
+          " their block sizes."
+      )
+
+    num_b_blocks = b_dim // b_block_sz
+    num_v_blocks = v_dim // v_block_sz
+
+    flat_hidden = _reshape(
+        hidden_states,
+        (b_dim, emb_dim),
+        create_sharding(
+            model.mesh,
+            ("activation_embed_and_logits_batch_sequence", "activation_embed"),
+        ),
+    )
+    flat_labels = _reshape(
+        labels,
+        (b_dim,),
+        create_sharding(
+            model.mesh, ("activation_embed_and_logits_batch_sequence",)
+        ),
+    )
+    flat_segmentation = _reshape(
+        segmentation,
+        (b_dim,),
+        create_sharding(
+            model.mesh, ("activation_embed_and_logits_batch_sequence",)
+        ),
+    )
+
+    if config.logits_via_embedding:
+      w = gathered_params["params"]["shared_embedding"]["embedding"]
+    else:
+      w = gathered_params["params"]["decoder"]["logits_dense"]["kernel"]
+
+    def b_loop_body(i, carry):
+      total_loss, total_z_loss = carry
+      b_start = i * b_block_sz
+
+      def v_loop_body(j, v_carry):
+        lse_b_, b_loss_sum_neg_logits_ = v_carry
+        v_start = j * v_block_sz
+        labels_b = jax.lax.dynamic_slice(flat_labels, (b_start,), (b_block_sz,))
+        x_b = jax.lax.dynamic_slice(
+            flat_hidden, (b_start, 0), (b_block_sz, emb_dim)
+        )
+
+        # Apply normalization to the batch block
+        x_b_norm = model.apply(
+            {"params": gathered_params["params"]},
+            x_b,
+            deterministic=deterministic,
+            method="normalize_hidden_states",
+        )
+        x_b_norm = _maybe_shard_with_name(x_b_norm, chunked_hidden_spec)
+
+        # Extract w_j
+        if config.logits_via_embedding:
+          # Attend on embedding table. Table is (vocab_size, emb_dim)
+          # Transpose to (emb_dim, vocab_size)
+          w_j = jax.lax.dynamic_slice(w.T, (0, v_start), (emb_dim, v_block_sz))
+        else:
+          w_j = jax.lax.dynamic_slice(w, (0, v_start), (emb_dim, v_block_sz))
+
+        # Compute logits for the block
+        logits_bv = jnp.dot(x_b_norm, w_j)
+
+        if config.logits_via_embedding and config.normalize_embedding_logits:
+          logits_bv = logits_bv / jnp.sqrt(emb_dim)
+        if config.final_logits_soft_cap:
+          logits_bv = logits_bv / config.final_logits_soft_cap
+          logits_bv = jnp.tanh(logits_bv) * config.final_logits_soft_cap
+
+        if config.cast_logits_to_fp32:
+          logits_bv = logits_bv.astype(jnp.float32)
+
+        lse_b__ = jnp.logaddexp(lse_b_, jax.nn.logsumexp(logits_bv, axis=-1))
+
+        labels_one_hot = jax.nn.one_hot(
+            labels_b - v_start, v_block_sz, dtype=logits_bv.dtype
+        )
+        b_loss_sum_neg_logits__ = b_loss_sum_neg_logits_ - jnp.sum(
+            logits_bv * labels_one_hot, axis=-1
+        )
+        return lse_b__, b_loss_sum_neg_logits__
+
+      lse_b, b_loss_sum_neg_logits = jax.lax.fori_loop(
+          0,
+          num_v_blocks,
+          v_loop_body,
+          (
+              jnp.full((b_block_sz,), -jnp.inf, dtype=jnp.float32),
+              jnp.zeros((b_block_sz,), dtype=jnp.float32),
+          ),
+      )
+
+      segmentation_b = jax.lax.dynamic_slice(
+          flat_segmentation, (b_start,), (b_block_sz,)
+      )
+      mask = (segmentation_b != 0).astype(jnp.float32)
+
+      # Z-loss
+      z_loss_b = config.z_loss_multiplier * jnp.square(lse_b) * mask
+      total_z_loss += jnp.sum(z_loss_b)
+
+      b_loss_sum_neg_logits = b_loss_sum_neg_logits * mask
+      lse_b_masked = lse_b * mask
+
+      total_loss += jnp.sum(b_loss_sum_neg_logits) + jnp.sum(lse_b_masked)
+
+      return total_loss, total_z_loss
+
+    initial_acc = (0.0, 0.0)
+    total_loss, total_z_loss = jax.lax.fori_loop(
+        0,
+        num_b_blocks,
+        b_loop_body,
+        initial_acc,
+    )
+
+    # For drop-in replacement, we return residuals as the current method does.
+    # We pack necessary values for the backward pass.
+    # Note that the backward pass would also need to be implemented for this method
+    # to be fully compatible with jax.custom_vjp.
+    residuals = (
+        gathered_params,
+        flat_hidden,
+        flat_labels,
+        flat_segmentation,
+        batch_size,
+        seq_len,
+        emb_dim,
+    )
+    return (total_loss, total_z_loss), residuals
+
   def _chunked_cross_entropy_loss_fwd(gathered_params, hidden_states, labels, segmentation):
     batch_size, seq_len, emb_dim = hidden_states.shape
     vocab_tile_size = (batch_size * seq_len) // config.num_vocab_tiling