Merge pull request #3108 from AI-Hypercomputer:qinwen/add_checkpoint

PiperOrigin-RevId: 866687753

Author: Google-ML-Automation

src/MaxText/configs/base.yml

@@ -311,6 +311,7 @@ qkv_proj: 'remat'
 out_proj: 'remat'
 mla_q: 'remat'
 mla_kv: 'remat'
+attention_out: 'remat'
 
 optimizer_memory_host_offload: False
 parameter_memory_host_offload: False

src/MaxText/configs/types.py

@@ -876,6 +876,11 @@ class RematAndOffload(BaseModel):
       RematLocation.REMAT,
       description="Remat policy for the mla's key and value projection.",
   )
+  attention_out: RematLocation = Field(
+      RematLocation.REMAT,
+      description="Remat policy for the attention output.",
+  )
+
   optimizer_memory_host_offload: bool = Field(False, description="Offload optimizer state to host memory.")
   parameter_memory_host_offload: bool = Field(False, description="Offload parameters to host memory.")
 
@@ -2064,6 +2069,7 @@ def calculate_global_batch_sizes(per_device_batch_size, expansion_factor, num_de
           "mla_kv",
           "mla_q",
           "qkv_proj",
+          "attention_out",
           "out_proj",
       ]
       self.tensors_on_device = [t for t in tensors if getattr(self, t) == "device"]

src/MaxText/layers/attention_mla.py

@@ -1038,6 +1038,7 @@ def __call__(
       # Pass the index_mask to the Attention Op
       out = self.attention_op(query, key, value, decoder_segment_ids, model_mode, cached_values, index_mask=index_mask)
 
+    out = jax.ad_checkpoint.checkpoint_name(out, "attention_out")
     if model_mode == MODEL_MODE_TRAIN and self.config.expert_shard_attention_option == EP_AS_CONTEXT:
       out = self._maybe_shard_with_logical(out, self.ep_out_axis_names)
     else:

src/MaxText/layers/attentions.py

@@ -1132,6 +1132,7 @@ def __call__(
           bidirectional_mask,
           self.sinks,
       )
+    out = jax.ad_checkpoint.checkpoint_name(out, "attention_out")
     if model_mode == MODEL_MODE_PREFILL:
       out = self._maybe_shard_with_logical(out, self.prefill_out_axis_names)
     elif model_mode == MODEL_MODE_TRAIN and self.config.expert_shard_attention_option == EP_AS_CONTEXT:

src/MaxText/layers/deepseek_batchsplit.py

@@ -339,6 +339,7 @@ def mla(
       qk_nope_head_dim=qk_nope_head_dim,
       mscale=mscale,
   )
+  query = jax.ad_checkpoint.checkpoint_name(query, "query_proj")
   key, value = kv_projection(
       inputs,
       positions,
@@ -358,6 +359,8 @@ def mla(
       qk_nope_head_dim=qk_nope_head_dim,
       num_query_heads=num_query_heads,
   )
+  key = jax.ad_checkpoint.checkpoint_name(key, "key_proj")
+  value = jax.ad_checkpoint.checkpoint_name(value, "value_proj")
   out = attention_op_fn(
       query,
       key,
@@ -366,7 +369,9 @@ def mla(
       model_mode,
       cached_values=[None, None],
   )
+  out = jax.ad_checkpoint.checkpoint_name(out, "attention_out")
   out = dot(out, out_weights, axes=2)
+  out = jax.ad_checkpoint.checkpoint_name(out, "out_proj")
   return out
 
 
@@ -405,6 +410,7 @@ def query_projection(
       epsilon=epsilon,
       dtype=dtype,
   )
+  low_rank_q = jax.ad_checkpoint.checkpoint_name(low_rank_q, "mla_q")
   q = dot(low_rank_q, wq_b_weights)
 
   # Split into non-positional and rotary parts.
@@ -454,6 +460,7 @@ def kv_projection(
       epsilon=kv_norm_epsilon,
       dtype=dtype,
   )
+  low_rank_main = jax.ad_checkpoint.checkpoint_name(low_rank_main, "mla_kv")
   key_rope = jnp.expand_dims(low_rank_rope, axis=2)
   key_rope = yarn(
       key_rope,
@@ -693,6 +700,8 @@ def compute(x, w0, w1, wo, group_sizes, weights, *, wi_tile_size, wo_tile_size,
   )
   layer_w0 = gmm_fn(x, w0, tiling=wi_tile_size)
   layer_w1 = gmm_fn(x, w1, tiling=wi_tile_size)
+  layer_w0 = jax.ad_checkpoint.checkpoint_name(layer_w0, "mlpwi_0")
+  layer_w1 = jax.ad_checkpoint.checkpoint_name(layer_w1, "mlpwi_1")
   intermediate_layer = jax.nn.silu(layer_w0) * layer_w1
   intermediate_layer *= weights[:, None]
   return gmm_fn(intermediate_layer, wo, tiling=wo_tile_size)