Update einsum layer for Gemma example

PiperOrigin-RevId: 718420528

examples/gemma/layers.py

@@ -16,8 +16,8 @@
 
 from __future__ import annotations
 
-from typing import Any, Union
 from collections.abc import Sequence
+from typing import Any, Union
 
 from flax import nnx
 import flax.linen as nn
@@ -31,11 +31,12 @@
 class Einsum(nnx.Module):
   """Einsum is a convenience module for parameterized tensor multiplication."""
 
-  def __init__(self, shape: Shape, *, rngs: nnx.Rngs):
+  def __init__(self, einsum_str: str, shape: Shape, *, rngs: nnx.Rngs):
+    self.einsum_str = einsum_str
     self.w = nnx.Param(nn.initializers.normal()(rngs.params(), shape))
 
-  def __call__(self, eqn: str, x: ArrayLike) -> Array:
-    return jnp.einsum(eqn, x, self.w.value)
+  def __call__(self, x: ArrayLike) -> Array:
+    return jnp.einsum(self.einsum_str, x, self.w.value)
 
   @property
   def shape(self) -> Shape:
@@ -48,7 +49,7 @@ class RMSNorm(nnx.Module):
   def __init__(self, dim: int, *, rngs: nnx.Rngs):
     self.scale = nnx.Param(nn.initializers.zeros_init()(rngs.params(), dim))
 
-  def __call__(self, x):
+  def __call__(self, x: Array) -> Array:
     var = jnp.mean(jnp.square(x), axis=-1, keepdims=True)
     normed_inputs = jnp.asarray(x * jnp.reciprocal(jnp.sqrt(var + 1e-06)))
     # normed_inputs is a rank-K tensor, K > 1 (K is typically 2 or 3). scale is

examples/gemma/layers_test.py

@@ -38,9 +38,8 @@ class EinsumTest(parameterized.TestCase):
       ),
   )
   def test_einsum(self, inputs_shape, params_shape, eqn, expected_shape):
-    einsum = layers.Einsum(params_shape, rngs=nnx.Rngs(params=0))
+    einsum = layers.Einsum(eqn, params_shape, rngs=nnx.Rngs(params=0))
     output = einsum(
-        eqn,
         jnp.ones(inputs_shape),
     )
     self.assertEqual(output.shape, expected_shape)
@@ -54,7 +53,7 @@ def test_einsum(self, inputs_shape, params_shape, eqn, expected_shape):
       ),
   )
   def test_shape(self, shape):
-    einsum = layers.Einsum(shape, rngs=nnx.Rngs(params=0))
+    einsum = layers.Einsum('ij->ji', shape, rngs=nnx.Rngs(params=0))
     self.assertEqual(einsum.shape, shape)
 
 

examples/gemma/modules.py

@@ -16,14 +16,14 @@
 
 from __future__ import annotations
 
+from collections.abc import Sequence
 import enum
 from typing import Any, Union
-from collections.abc import Sequence
 
 from flax import nnx
-import flax.linen as nn
 import layers
 import positional_embeddings
+import flax.linen as nn
 import jax
 import jax.numpy as jnp
 from jaxtyping import Array, ArrayLike  # pylint: disable=g-importing-member,g-multiple-import
@@ -94,21 +94,25 @@ def __init__(
     self.sliding_window_size = sliding_window_size
     self.attn_logits_soft_cap = attn_logits_soft_cap
     self.attn_vec_einsum = layers.Einsum(
+        einsum_str='BTNH,NHD->BTD',
         shape=(num_heads, head_dim, features),
         rngs=rngs,
     )
 
     if num_heads == num_kv_heads:
       self.qkv_einsum = layers.Einsum(
+          einsum_str='BTD,SNDH->SBTNH',
           shape=(3, num_heads, features, head_dim),
           rngs=rngs,
       )
     else:
       self.q_einsum = layers.Einsum(
+          einsum_str='BTD,NDH->BTNH',
           shape=(num_heads, features, head_dim),
           rngs=rngs,
       )
       self.kv_einsum = layers.Einsum(
+          einsum_str='BSD,CKDH->CBSKH',
           shape=(2, num_kv_heads, features, head_dim),
           rngs=rngs,
       )
@@ -123,10 +127,10 @@ def __call__(
     seq_len = x.shape[1]
 
     if self.use_qkv_einsum:
-      query_proj, key_proj, value_proj = self.qkv_einsum('BTD,SNDH->SBTNH', x)
+      query_proj, key_proj, value_proj = self.qkv_einsum(x)
     else:
-      query_proj = self.q_einsum('BTD,NDH->BTNH', x)
-      key_proj, value_proj = self.kv_einsum('BSD,CKDH->CBSKH', x)
+      query_proj = self.q_einsum(x)
+      key_proj, value_proj = self.kv_einsum(x)
 
     query_proj = positional_embeddings.apply_rope(
         query_proj,
@@ -173,7 +177,7 @@ def __call__(
     padded_logits = jnp.where((jnp.expand_dims(attn_mask, -2)), logits, K_MASK)
     probs = jax.nn.softmax(padded_logits, axis=-1).astype(key_proj.dtype)
     encoded = jnp.einsum('BTNS,BSNH->BTNH', probs, value_proj)
-    attn_output = self.attn_vec_einsum('BTNH,NHD->BTD', encoded)
+    attn_output = self.attn_vec_einsum(encoded)
 
     if cache is not None:
       new_cache = {