xax 0.2.22__py3-none-any.whl → 0.2.23__py3-none-any.whl

xax/__init__.py

@@ -12,7 +12,7 @@ and running the update script:
     python -m scripts.update_api --inplace
 """
 
-__version__ = "0.2.22"
+__version__ = "0.2.23"
 
 # This list shouldn't be modified by hand; instead, run the update script.
 __all__ = [
@@ -23,6 +23,10 @@ __all__ = [
     "get_run_dir",
     "load_user_config",
     "State",
+    "CrossAttentionBlock",
+    "SelfAttentionBlock",
+    "Transformer",
+    "TransformerBlock",
     "FourierEmbeddings",
     "IdentityPositionalEmbeddings",
     "LearnedPositionalEmbeddings",
@@ -200,6 +204,10 @@ NAME_MAP: dict[str, str] = {
     "get_run_dir": "core.conf",
     "load_user_config": "core.conf",
     "State": "core.state",
+    "CrossAttentionBlock": "nn.attention",
+    "SelfAttentionBlock": "nn.attention",
+    "Transformer": "nn.attention",
+    "TransformerBlock": "nn.attention",
     "FourierEmbeddings": "nn.embeddings",
     "IdentityPositionalEmbeddings": "nn.embeddings",
     "LearnedPositionalEmbeddings": "nn.embeddings",
@@ -370,6 +378,7 @@ if IMPORT_ALL or TYPE_CHECKING:
         load_user_config,
     )
     from xax.core.state import Phase, State
+    from xax.nn.attention import CrossAttentionBlock, SelfAttentionBlock, Transformer, TransformerBlock
     from xax.nn.embeddings import (
         EmbeddingKind,
         FourierEmbeddings,

xax/core/state.py

@@ -27,11 +27,8 @@ def _int_to_phase(i: int) -> Phase:
 class StateDict(TypedDict, total=False):
     num_steps: NotRequired[int | Array]
     num_samples: NotRequired[int | Array]
-    num_valid_steps: NotRequired[int | Array]
-    num_valid_samples: NotRequired[int | Array]
     start_time_s: NotRequired[float | Array]
     elapsed_time_s: NotRequired[float | Array]
-    valid_elapsed_time_s: NotRequired[float | Array]
     phase: NotRequired[Phase]
     _phase: NotRequired[int | Array]
 
@@ -47,38 +44,26 @@ class State:
         return self._int32_arr[0]
 
     @property
-    def num_valid_steps(self) -> Array:
-        return self._int32_arr[1]
+    def phase(self) -> Phase:
+        return _int_to_phase(self._int32_arr[1].item())
 
     @property
     def num_samples(self) -> Array:
         return self._float32_arr[0]
 
-    @property
-    def num_valid_samples(self) -> Array:
-        return self._float32_arr[1]
-
     @property
     def start_time_s(self) -> Array:
-        return self._float32_arr[2]
+        return self._float32_arr[1]
 
     @property
     def elapsed_time_s(self) -> Array:
-        return self._float32_arr[3]
-
-    @property
-    def valid_elapsed_time_s(self) -> Array:
-        return self._float32_arr[4]
-
-    @property
-    def phase(self) -> Phase:
-        return _int_to_phase(self._int32_arr[2].item())
+        return self._float32_arr[2]
 
     @classmethod
     def init_state(cls) -> "State":
         return cls(
-            _int32_arr=jnp.array([0, 0, 0], dtype=jnp.int32),
-            _float32_arr=jnp.array([0.0, 0.0, time.time(), 0.0, 0.0], dtype=jnp.float32),
+            _int32_arr=jnp.array([0, 0], dtype=jnp.int32),
+            _float32_arr=jnp.array([0.0, time.time(), 0.0], dtype=jnp.float32),
         )
 
     @property
@@ -91,25 +76,19 @@ class State:
 
         if "num_steps" in kwargs:
             int32_arr = int32_arr.at[0].set(kwargs["num_steps"])
-        if "num_valid_steps" in kwargs:
-            int32_arr = int32_arr.at[1].set(kwargs["num_valid_steps"])
 
         if "phase" in kwargs:
-            int32_arr = int32_arr.at[2].set(_phase_to_int(kwargs["phase"]))
+            int32_arr = int32_arr.at[1].set(_phase_to_int(kwargs["phase"]))
         if "_phase" in kwargs:
-            int32_arr = int32_arr.at[2].set(kwargs["_phase"])
+            int32_arr = int32_arr.at[1].set(kwargs["_phase"])
 
         if "num_samples" in kwargs:
             float32_arr = float32_arr.at[0].set(kwargs["num_samples"])
-        if "num_valid_samples" in kwargs:
-            float32_arr = float32_arr.at[1].set(kwargs["num_valid_samples"])
 
         if "start_time_s" in kwargs:
-            float32_arr = float32_arr.at[2].set(kwargs["start_time_s"])
+            float32_arr = float32_arr.at[1].set(kwargs["start_time_s"])
         if "elapsed_time_s" in kwargs:
-            float32_arr = float32_arr.at[3].set(kwargs["elapsed_time_s"])
-        if "valid_elapsed_time_s" in kwargs:
-            float32_arr = float32_arr.at[4].set(kwargs["valid_elapsed_time_s"])
+            float32_arr = float32_arr.at[2].set(kwargs["elapsed_time_s"])
 
         return State(
             _int32_arr=int32_arr,
@@ -119,12 +98,9 @@ class State:
     def to_dict(self) -> dict[str, int | float | str]:
         return {
             "num_steps": int(self.num_steps.item()),
-            "num_valid_steps": int(self.num_valid_steps.item()),
             "num_samples": int(self.num_samples.item()),
-            "num_valid_samples": int(self.num_valid_samples.item()),
             "start_time_s": float(self.start_time_s.item()),
             "elapsed_time_s": float(self.elapsed_time_s.item()),
-            "valid_elapsed_time_s": float(self.valid_elapsed_time_s.item()),
             "phase": str(self.phase),
         }
 
@@ -136,7 +112,6 @@ class State:
         int32_arr = jnp.array(
             [
                 d.get("num_steps", 0),
-                d.get("num_valid_steps", 0),
                 d.get("_phase", 0),
             ],
             dtype=jnp.int32,
@@ -145,10 +120,8 @@ class State:
         float32_arr = jnp.array(
             [
                 d.get("num_samples", 0),
-                d.get("num_valid_samples", 0),
                 d.get("start_time_s", time.time()),
                 d.get("elapsed_time_s", 0.0),
-                d.get("valid_elapsed_time_s", 0.0),
             ],
             dtype=jnp.float32,
         )

xax/nn/attention.py

@@ -0,0 +1,738 @@
+"""Attention mechanisms for transformer models."""
+
+from typing import Literal, cast, overload
+
+import chex
+import equinox as eqx
+import jax
+import jax.numpy as jnp
+from jaxtyping import Array, PRNGKeyArray
+
+
+class SelfAttentionBlock(eqx.Module):
+    """Self-attention block using jax.nn.dot_product_attention."""
+
+    q_proj: eqx.nn.Linear
+    k_proj: eqx.nn.Linear
+    v_proj: eqx.nn.Linear
+    output_proj: eqx.nn.Linear
+    num_heads: int = eqx.static_field()
+    head_dim: int = eqx.static_field()
+    causal: bool = eqx.static_field()
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        *,
+        key: PRNGKeyArray,
+        causal: bool = False,
+    ) -> None:
+        keys = jax.random.split(key, 4)
+
+        self.num_heads = num_heads
+        self.head_dim = embed_dim // num_heads
+        assert self.head_dim * num_heads == embed_dim, "embed_dim must be divisible by num_heads"
+
+        self.q_proj = eqx.nn.Linear(embed_dim, embed_dim, key=keys[0])
+        self.k_proj = eqx.nn.Linear(embed_dim, embed_dim, key=keys[1])
+        self.v_proj = eqx.nn.Linear(embed_dim, embed_dim, key=keys[2])
+        self.output_proj = eqx.nn.Linear(embed_dim, embed_dim, key=keys[3])
+
+        self.causal = causal
+
+    def _reshape_for_multihead(self, x: Array) -> Array:
+        """Reshape from (seq_len, embed_dim) to (seq_len, num_heads, head_dim)."""
+        seq_len, _ = x.shape
+        return x.reshape(seq_len, self.num_heads, self.head_dim)
+
+    def _combine_heads(self, x: Array) -> Array:
+        """Reshape from (seq_len, num_heads, head_dim) to (seq_len, embed_dim)."""
+        seq_len, _, _ = x.shape
+        return x.reshape(seq_len, -1)
+
+    def __call__(
+        self,
+        x: Array,
+        *,
+        key: PRNGKeyArray | None = None,
+        mask: Array | None = None,
+        cache: dict[str, Array] | None = None,
+        update_cache: bool = False,
+    ) -> Array | tuple[Array, dict[str, Array]]:
+        """Apply self-attention to the input.
+
+        Args:
+            x: Input tensor of shape (seq_len, embed_dim)
+            key: PRNGKey for dropout randomness
+            mask: Optional mask tensor of shape (seq_len, seq_len) or broadcastable
+            cache: Optional dictionary containing cached key and value tensors
+            update_cache: Whether to update the cache and return it
+
+        Returns:
+            If update_cache is False: Output tensor of shape (seq_len, embed_dim)
+            If update_cache is True: Tuple of (output tensor, updated cache)
+        """
+        chex.assert_rank(x, 2)
+
+        # Project inputs to queries, keys, and values
+        q = jax.vmap(self.q_proj)(x)
+
+        # Use cached key/value if provided and not updating cache
+        if cache is not None and not update_cache:
+            k = cache["k"]
+            v = cache["v"]
+        else:
+            k = jax.vmap(self.k_proj)(x)
+            v = jax.vmap(self.v_proj)(x)
+
+            # Update cache if needed
+            if update_cache:
+                if cache is None:
+                    cache = {}
+                cache = {"k": k, "v": v}
+
+        # Reshape to multihead format
+        q = self._reshape_for_multihead(q)
+        k = self._reshape_for_multihead(k)
+        v = self._reshape_for_multihead(v)
+
+        # Apply dot product attention.
+        # Note that Apple Silicon struggles with this:
+        # https://github.com/jax-ml/jax/issues/20114
+        attn_output = jax.nn.dot_product_attention(
+            q,
+            k,
+            v,
+            mask=mask,
+            is_causal=self.causal and mask is None,
+        )
+
+        # Combine heads
+        attn_output = self._combine_heads(attn_output)
+
+        # Final projection
+        output = jax.vmap(self.output_proj)(attn_output)
+
+        if update_cache:
+            return output, cast(dict[str, Array], cache)
+        return output
+
+
+class CrossAttentionBlock(eqx.Module):
+    """Cross-attention block using jax.nn.dot_product_attention."""
+
+    q_proj: eqx.nn.Linear
+    k_proj: eqx.nn.Linear
+    v_proj: eqx.nn.Linear
+    output_proj: eqx.nn.Linear
+    num_heads: int = eqx.static_field()
+    head_dim: int = eqx.static_field()
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        *,
+        key: PRNGKeyArray,
+    ) -> None:
+        keys = jax.random.split(key, 4)
+
+        self.num_heads = num_heads
+        self.head_dim = embed_dim // num_heads
+        assert self.head_dim * num_heads == embed_dim, "embed_dim must be divisible by num_heads"
+
+        self.q_proj = eqx.nn.Linear(embed_dim, embed_dim, key=keys[0])
+        self.k_proj = eqx.nn.Linear(embed_dim, embed_dim, key=keys[1])
+        self.v_proj = eqx.nn.Linear(embed_dim, embed_dim, key=keys[2])
+        self.output_proj = eqx.nn.Linear(embed_dim, embed_dim, key=keys[3])
+
+    def _reshape_for_multihead(self, x: Array) -> Array:
+        """Reshape from (seq_len, embed_dim) to (seq_len, num_heads, head_dim)."""
+        seq_len, _ = x.shape
+        return x.reshape(seq_len, self.num_heads, self.head_dim)
+
+    def _combine_heads(self, x: Array) -> Array:
+        """Reshape from (seq_len, num_heads, head_dim) to (seq_len, embed_dim)."""
+        seq_len, _, _ = x.shape
+        return x.reshape(seq_len, -1)
+
+    def __call__(
+        self,
+        q_input: Array,
+        kv_input: Array,
+        *,
+        key: PRNGKeyArray | None = None,
+        mask: Array | None = None,
+        cache: dict[str, Array] | None = None,
+        update_cache: bool = False,
+    ) -> Array | tuple[Array, dict[str, Array]]:
+        """Apply cross-attention.
+
+        Args:
+            q_input: Query input tensor of shape (q_seq_len, embed_dim)
+            kv_input: Key/value input tensor of shape (kv_seq_len, embed_dim)
+            key: PRNGKey for dropout randomness
+            mask: Optional mask tensor
+            cache: Optional dictionary containing cached key and value tensors
+            update_cache: Whether to update the cache and return it
+
+        Returns:
+            If update_cache is False: Output tensor of shape (q_seq_len, embed_dim)
+            If update_cache is True: Tuple of (output tensor, updated cache)
+        """
+        chex.assert_rank(q_input, 2)
+        chex.assert_rank(kv_input, 2)
+
+        # Project inputs to queries, keys, and values
+        q = jax.vmap(self.q_proj)(q_input)
+
+        # Use cached key/value if provided and not updating cache
+        if cache is not None and not update_cache:
+            k = cache["k"]
+            v = cache["v"]
+        else:
+            k = jax.vmap(self.k_proj)(kv_input)
+            v = jax.vmap(self.v_proj)(kv_input)
+
+            # Update cache if needed
+            if update_cache:
+                if cache is None:
+                    cache = {}
+                cache = {"k": k, "v": v}
+
+        # Reshape to multihead format
+        q = self._reshape_for_multihead(q)
+        k = self._reshape_for_multihead(k)
+        v = self._reshape_for_multihead(v)
+
+        # Apply dot product attention
+        attn_output = jax.nn.dot_product_attention(
+            q,
+            k,
+            v,
+            mask=mask,
+            is_causal=False,
+        )
+
+        # Combine heads
+        attn_output = self._combine_heads(attn_output)
+
+        # Final projection
+        output = jax.vmap(self.output_proj)(attn_output)
+
+        if update_cache:
+            return output, cast(dict[str, Array], cache)
+        return output
+
+
+class TransformerBlock(eqx.Module):
+    self_attn: SelfAttentionBlock
+    cross_attn: CrossAttentionBlock | None
+    feed_forward: eqx.nn.MLP
+    layer_norm1: eqx.nn.LayerNorm
+    layer_norm2: eqx.nn.LayerNorm
+    layer_norm3: eqx.nn.LayerNorm | None
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        ff_dim: int,
+        *,
+        key: PRNGKeyArray,
+        causal: bool = False,
+        cross_attention: bool = False,
+    ) -> None:
+        keys = jax.random.split(key, 4)
+
+        self.self_attn = SelfAttentionBlock(
+            embed_dim=embed_dim,
+            num_heads=num_heads,
+            key=keys[0],
+            causal=causal,
+        )
+
+        if cross_attention:
+            self.cross_attn = CrossAttentionBlock(
+                embed_dim=embed_dim,
+                num_heads=num_heads,
+                key=keys[1],
+            )
+            self.layer_norm3 = eqx.nn.LayerNorm(embed_dim)
+        else:
+            self.cross_attn = None
+            self.layer_norm3 = None
+
+        self.layer_norm1 = eqx.nn.LayerNorm(embed_dim)
+        self.layer_norm2 = eqx.nn.LayerNorm(embed_dim)
+
+        self.feed_forward = eqx.nn.MLP(
+            in_size=embed_dim,
+            out_size=embed_dim,
+            width_size=ff_dim,
+            depth=1,
+            activation=jax.nn.gelu,
+            key=keys[2],
+        )
+
+    @overload
+    def __call__(
+        self,
+        x: Array,
+        *,
+        context: Array | None = None,
+        self_mask: Array | None = None,
+        cross_mask: Array | None = None,
+        key: PRNGKeyArray | None = None,
+        cache: dict[str, dict[str, Array]] | None = None,
+        update_cache: Literal[True],
+    ) -> tuple[Array, dict[str, dict[str, Array]]]: ...
+
+    @overload
+    def __call__(
+        self,
+        x: Array,
+        *,
+        context: Array | None = None,
+        self_mask: Array | None = None,
+        cross_mask: Array | None = None,
+        key: PRNGKeyArray | None = None,
+        cache: dict[str, dict[str, Array]] | None = None,
+        update_cache: Literal[False] = False,
+    ) -> Array: ...
+
+    def __call__(
+        self,
+        x: Array,
+        *,
+        context: Array | None = None,
+        self_mask: Array | None = None,
+        cross_mask: Array | None = None,
+        key: PRNGKeyArray | None = None,
+        cache: dict[str, dict[str, Array]] | None = None,
+        update_cache: bool = False,
+    ) -> Array | tuple[Array, dict[str, dict[str, Array]]]:
+        """Apply transformer block.
+
+        Args:
+            x: Input tensor
+            context: Optional context for cross-attention
+            self_mask: Mask for self-attention
+            cross_mask: Mask for cross-attention
+            key: Optional PRNG key for dropout
+            cache: Optional dictionary containing cached key and value tensors
+            update_cache: Whether to update the cache and return it
+
+        Returns:
+            If update_cache is False: Output tensor
+            If update_cache is True: Tuple of (output tensor, updated cache)
+        """
+        chex.assert_rank(x, 2)
+        if key is not None:
+            key1, key2 = jax.random.split(key)
+        else:
+            key1 = key2 = None
+
+        # Initialize cache if needed
+        updated_cache = {}
+        if cache is None:
+            cache = {}
+
+        # Self-attention block with pre-norm
+        norm_x = jax.vmap(self.layer_norm1)(x)
+
+        self_attn_cache = cache.get("self_attn")
+        if update_cache:
+            attn_output, self_attn_cache = self.self_attn(
+                norm_x, key=key1, mask=self_mask, cache=self_attn_cache, update_cache=True
+            )
+            updated_cache["self_attn"] = self_attn_cache
+        else:
+            attn_output = self.self_attn(norm_x, key=key1, mask=self_mask, cache=self_attn_cache)
+
+        x = x + attn_output
+
+        # Cross-attention block (if enabled) with pre-norm
+        if self.cross_attn is not None and context is not None:
+            assert self.layer_norm3 is not None
+
+            norm_x = jax.vmap(self.layer_norm3)(x)
+            cross_attn_cache = cache.get("cross_attn")
+
+            if update_cache:
+                cross_attn_output, cross_attn_cache = self.cross_attn(
+                    norm_x, context, key=key2, mask=cross_mask, cache=cross_attn_cache, update_cache=True
+                )
+                updated_cache["cross_attn"] = cross_attn_cache
+            else:
+                cross_attn_output = self.cross_attn(norm_x, context, key=key2, mask=cross_mask, cache=cross_attn_cache)
+
+            x = x + cross_attn_output
+
+        # Feed-forward block with pre-norm
+        norm_x = jax.vmap(self.layer_norm2)(x)
+        ff_output = jax.vmap(self.feed_forward)(norm_x)
+        x = x + ff_output
+
+        if update_cache:
+            return x, updated_cache
+        return x
+
+
+class Transformer(eqx.Module):
+    token_embedding: eqx.nn.Embedding
+    position_embedding: eqx.nn.Embedding | None
+    layers: list[TransformerBlock]
+    output_layer: eqx.nn.Linear | None
+    layer_norm: eqx.nn.LayerNorm
+    max_seq_len: int = eqx.static_field()
+    embed_dim: int = eqx.static_field()
+
+    def __init__(
+        self,
+        vocab_size: int,
+        embed_dim: int,
+        num_heads: int,
+        ff_dim: int,
+        num_layers: int,
+        max_seq_len: int,
+        output_size: int | None = None,
+        *,
+        key: PRNGKeyArray,
+        causal: bool = False,
+        cross_attention: bool = False,
+        use_absolute_position: bool = True,
+    ) -> None:
+        keys = jax.random.split(key, num_layers + 3)
+
+        self.token_embedding = eqx.nn.Embedding(vocab_size, embed_dim, key=keys[0])
+
+        # Position embeddings can be disabled
+        if use_absolute_position:
+            self.position_embedding = eqx.nn.Embedding(max_seq_len, embed_dim, key=keys[1])
+        else:
+            self.position_embedding = None
+
+        self.layers = [
+            TransformerBlock(
+                embed_dim=embed_dim,
+                num_heads=num_heads,
+                ff_dim=ff_dim,
+                key=keys[i + 2],
+                causal=causal,
+                cross_attention=cross_attention,
+            )
+            for i in range(num_layers)
+        ]
+
+        self.layer_norm = eqx.nn.LayerNorm(embed_dim)
+
+        if output_size is not None:
+            self.output_layer = eqx.nn.Linear(embed_dim, output_size, key=keys[-1])
+        else:
+            self.output_layer = None
+
+        self.max_seq_len = max_seq_len
+        self.embed_dim = embed_dim
+
+    def _add_positional_embedding(self, x_embedded: Array, positions: Array | None = None) -> Array:
+        """Add positional embeddings to the token embeddings."""
+        if self.position_embedding is None:
+            return x_embedded
+
+        seq_len, _ = x_embedded.shape
+
+        if positions is None:
+            positions = jnp.arange(seq_len)
+        pos_embedded = jax.vmap(self.position_embedding)(positions)
+
+        return x_embedded + pos_embedded
+
+    def encode(
+        self,
+        x: Array,
+        mask: Array | None = None,
+        positions: Array | None = None,
+        key: PRNGKeyArray | None = None,
+        cache: dict[str, dict[str, dict[str, Array]]] | None = None,
+        update_cache: bool = False,
+    ) -> Array | tuple[Array, dict[str, dict[str, dict[str, Array]]]]:
+        """Encode the input sequence.
+
+        Args:
+            x: Input token indices of shape (seq_len)
+            mask: Optional attention mask
+            positions: Optional positions
+            key: Optional PRNG key for dropout
+            cache: Optional dictionary containing cached key and value tensors
+            update_cache: Whether to update the cache and return it
+
+        Returns:
+            If update_cache is False: Encoded representation
+            If update_cache is True: Tuple of (encoded representation, updated cache)
+        """
+        # Token embedding
+        x_embedded = jax.vmap(self.token_embedding)(x)
+
+        # Add positional embedding
+        x_embedded = self._add_positional_embedding(x_embedded, positions)
+
+        # Initialize layer caches
+        if cache is None and update_cache:
+            cache = {f"layer_{i}": {} for i in range(len(self.layers))}
+
+        # Updated cache will be built if needed
+        updated_cache = {}
+
+        # Apply transformer layers
+        keys: Array | list[None] = [None] * len(self.layers)
+        if key is not None:
+            keys = jax.random.split(key, len(self.layers))
+
+        for i, (layer, layer_key) in enumerate(zip(self.layers, keys, strict=False)):
+            layer_cache = None if cache is None else cache.get(f"layer_{i}")
+
+            if update_cache:
+                x_embedded, layer_updated_cache = layer.__call__(
+                    x_embedded,
+                    self_mask=mask,
+                    key=layer_key,
+                    cache=layer_cache,
+                    update_cache=True,
+                )
+                updated_cache[f"layer_{i}"] = layer_updated_cache
+            else:
+                x_embedded = layer.__call__(
+                    x_embedded,
+                    self_mask=mask,
+                    key=layer_key,
+                    cache=layer_cache,
+                )
+
+        # Apply final layer norm
+        output = jax.vmap(self.layer_norm)(x_embedded)
+
+        if update_cache:
+            return output, updated_cache
+        return output
+
+    def decode(
+        self,
+        x: Array,
+        context: Array,
+        self_mask: Array | None = None,
+        cross_mask: Array | None = None,
+        positions: Array | None = None,
+        key: PRNGKeyArray | None = None,
+        cache: dict[str, dict[str, dict[str, Array]]] | None = None,
+        update_cache: bool = False,
+    ) -> Array | tuple[Array, dict[str, dict[str, dict[str, Array]]]]:
+        """Decode with self-attention and cross-attention.
+
+        Args:
+            x: Input token indices
+            context: Context from encoder
+            self_mask: Optional self-attention mask
+            cross_mask: Optional cross-attention mask
+            positions: Optional positions
+            key: Optional PRNG key for dropout
+            cache: Optional dictionary containing cached key and value tensors
+            update_cache: Whether to update the cache and return it
+
+        Returns:
+            If update_cache is False: Decoded representation
+            If update_cache is True: Tuple of (decoded representation, updated cache)
+        """
+        # Token embedding
+        x_embedded = jax.vmap(lambda x_seq: jax.vmap(self.token_embedding)(x_seq))(x)
+
+        # Add positional embedding
+        x_embedded = self._add_positional_embedding(x_embedded, positions)
+
+        # Initialize layer caches
+        if cache is None and update_cache:
+            cache = {f"layer_{i}": {} for i in range(len(self.layers))}
+
+        # Updated cache will be built if needed
+        updated_cache = {}
+
+        # Apply transformer layers with cross-attention
+        keys: Array | list[None] = [None] * len(self.layers)
+        if key is not None:
+            keys = jax.random.split(key, len(self.layers))
+
+        for i, (layer, layer_key) in enumerate(zip(self.layers, keys, strict=False)):
+            layer_cache = None if cache is None else cache.get(f"layer_{i}")
+
+            if update_cache:
+                x_embedded, layer_updated_cache = layer.__call__(
+                    x_embedded,
+                    context=context,
+                    self_mask=self_mask,
+                    cross_mask=cross_mask,
+                    key=layer_key,
+                    cache=layer_cache,
+                    update_cache=True,
+                )
+                updated_cache[f"layer_{i}"] = layer_updated_cache
+            else:
+                x_embedded = layer(
+                    x_embedded,
+                    context=context,
+                    self_mask=self_mask,
+                    cross_mask=cross_mask,
+                    key=layer_key,
+                    cache=layer_cache,
+                )
+
+        # Apply final layer norm
+        output = jax.vmap(self.layer_norm)(x_embedded)
+
+        if update_cache:
+            return output, updated_cache
+        return output
+
+    @overload
+    def __call__(
+        self,
+        x: Array,
+        *,
+        mask: Array | None = None,
+        positions: Array | None = None,
+        key: PRNGKeyArray | None = None,
+        cache: dict[str, dict[str, dict[str, Array]]] | None = None,
+        update_cache: Literal[True],
+    ) -> tuple[Array, dict[str, dict[str, dict[str, Array]]]]: ...
+
+    @overload
+    def __call__(
+        self,
+        x: Array,
+        *,
+        mask: Array | None = None,
+        positions: Array | None = None,
+        key: PRNGKeyArray | None = None,
+        cache: dict[str, dict[str, dict[str, Array]]] | None = None,
+        update_cache: Literal[False] = False,
+    ) -> Array: ...
+
+    def __call__(
+        self,
+        x: Array,
+        *,
+        mask: Array | None = None,
+        positions: Array | None = None,
+        key: PRNGKeyArray | None = None,
+        cache: dict[str, dict[str, dict[str, Array]]] | None = None,
+        update_cache: bool = False,
+    ) -> Array | tuple[Array, dict[str, dict[str, dict[str, Array]]]]:
+        """Forward pass for encoder-only or decoder-only transformers.
+
+        Args:
+            x: Input token indices of shape (seq_len)
+            mask: Optional attention mask
+            positions: Optional positions
+            key: Optional PRNG key for dropout
+            cache: Optional dictionary containing cached key and value tensors
+            update_cache: Whether to update the cache and return it
+
+        Returns:
+            If update_cache is False: Output representation
+            If update_cache is True: Tuple of (output representation, updated cache)
+        """
+        chex.assert_rank(x, 1)
+
+        if update_cache:
+            output, updated_cache = self.encode(
+                x, mask=mask, positions=positions, key=key, cache=cache, update_cache=True
+            )
+        else:
+            output = self.encode(x, mask=mask, positions=positions, key=key, cache=cache)
+
+        # Apply output layer if it exists
+        if self.output_layer is not None:
+            output = jax.vmap(self.output_layer)(output)
+
+        if update_cache:
+            return output, updated_cache
+        return output
+
+    def predict_sequence(self, x_seq: Array) -> Array:
+        return self(x=x_seq)
+
+    def generate_sequence(
+        self,
+        prompt_seq: Array,
+        max_len: int,
+        temperature: float = 1.0,
+        top_k: int | None = None,
+        key: PRNGKeyArray | None = None,
+    ) -> Array:
+        """Generate a sequence autoregressively with KV caching.
+
+        Args:
+            prompt_seq: Input token indices of shape (prompt_len,)
+            max_len: Maximum length of generated sequence
+            temperature: Sampling temperature
+            top_k: Optional top-k sampling parameter
+            key: PRNG key for sampling
+
+        Returns:
+            Generated sequence of shape (prompt_len + max_len,)
+        """
+        if key is None:
+            key = jax.random.PRNGKey(0)
+
+        prompt_len = prompt_seq.shape[0]
+        sequence = prompt_seq
+
+        # Create causal mask for generation
+        causal_mask = jnp.tril(jnp.ones((self.max_seq_len, self.max_seq_len), dtype=jnp.bool_))
+
+        # Initialize cache with the prompt
+        _, cache = self(x=prompt_seq, mask=causal_mask[:prompt_len, :prompt_len], update_cache=True)
+
+        # Define decode step function (for clarity)
+        def decode_step(seq: Array, pos: int, cur_cache: dict, rng: PRNGKeyArray) -> tuple[Array, dict, PRNGKeyArray]:
+            # Get the next position and last token
+            pos_tensor = jnp.array([pos])
+            last_token = seq[-1:]
+
+            # Get logits for next token
+            rng, subrng = jax.random.split(rng)
+            logits, new_cache = self(
+                x=last_token,
+                positions=pos_tensor,
+                key=subrng,
+                cache=cur_cache,
+                update_cache=True,
+            )
+
+            # Extract final logits and apply temperature
+            logits = logits[-1] / temperature
+
+            # Apply top-k sampling if specified
+            if top_k is not None:
+                top_logits, top_indices = jax.lax.top_k(logits, top_k)
+                logits = jnp.full_like(logits, float("-inf"))
+                logits = logits.at[top_indices].set(top_logits)
+
+            # Sample next token
+            rng, subrng = jax.random.split(rng)
+            next_token = jax.random.categorical(subrng, logits[None, ...])[0]
+
+            # Add token to sequence
+            new_seq = jnp.concatenate([seq, next_token[None]], axis=0)
+            return new_seq, new_cache, rng
+
+        # Generate tokens one by one
+        for _ in range(max_len):
+            # Break if max sequence length reached
+            if sequence.shape[0] >= self.max_seq_len:
+                break
+
+            # Decode next token
+            sequence, cache, key = decode_step(seq=sequence, pos=sequence.shape[0] - 1, cur_cache=cache, rng=key)
+
+        return sequence

xax/task/logger.py

@@ -526,7 +526,7 @@ class LoggerImpl(ABC):
         Returns:
             If the logger should log the current step.
         """
-        elapsed_time = state.elapsed_time_s.item() if state.phase == "train" else state.valid_elapsed_time_s.item()
+        elapsed_time = state.elapsed_time_s.item()
         return self.tickers[state.phase].tick(elapsed_time)
 
 

xax/task/mixins/train.py

@@ -121,7 +121,7 @@ class ValidStepTimer:
         self.last_valid_step = state.num_steps.item()
 
     def __call__(self, state: State) -> bool:
-        if state.num_steps < self.valid_first_n_steps and state.num_valid_steps < self.valid_first_n_steps:
+        if state.num_steps < self.valid_first_n_steps:
             return True
 
         if self.last_valid_time is None or self.last_valid_step is None:
@@ -130,18 +130,15 @@ class ValidStepTimer:
 
         # Step-based validation.
         valid_every_n_steps = self.valid_every_n_steps
-        if valid_every_n_steps is not None and (
-            state.num_steps >= valid_every_n_steps + self.last_valid_step
-            or state.num_valid_steps >= valid_every_n_steps + self.last_valid_step
-        ):
+        if valid_every_n_steps is not None and state.num_steps >= valid_every_n_steps + self.last_valid_step:
             self._reset(state)
             return True
 
         # Time-based validation.
         valid_every_n_seconds = self.valid_every_n_seconds
-        if valid_every_n_seconds is not None and (
-            state.elapsed_time_s.item() - self.last_valid_time >= valid_every_n_seconds
-            or state.valid_elapsed_time_s.item() - self.last_valid_time >= valid_every_n_seconds
+        if (
+            valid_every_n_seconds is not None
+            and state.elapsed_time_s.item() - self.last_valid_time >= valid_every_n_seconds
         ):
             self._reset(state)
             return True
@@ -149,10 +146,7 @@ class ValidStepTimer:
         # Time-based validation for first validation step.
         if self.first_valid_step_flag:
             valid_first_n_seconds = self.valid_first_n_seconds
-            if valid_first_n_seconds is not None and (
-                state.elapsed_time_s.item() >= valid_first_n_seconds
-                or state.valid_elapsed_time_s.item() >= valid_first_n_seconds
-            ):
+            if valid_first_n_seconds is not None and state.elapsed_time_s.item() >= valid_first_n_seconds:
                 self._reset(state)
                 self.first_valid_step_flag = False
                 return True
@@ -777,12 +771,12 @@ class TrainMixin(
                     self.log_step(eqx.combine(model_arr, model_static), valid_batch, output, metrics, state)
 
                     state = state.replace(
-                        num_valid_steps=state.num_valid_steps + 1,
-                        num_valid_samples=state.num_valid_samples + (self.get_size_of_batch(valid_batch) or 0),
+                        num_steps=state.num_steps + 1,
+                        num_samples=state.num_samples + (self.get_size_of_batch(valid_batch) or 0),
                     )
 
                 state = state.replace(
-                    valid_elapsed_time_s=state.valid_elapsed_time_s + timer.elapsed_time,
+                    elapsed_time_s=state.elapsed_time_s + timer.elapsed_time,
                 )
 
             with ContextTimer() as timer:
@@ -882,7 +876,7 @@ class TrainMixin(
             key, model_key = jax.random.split(key)
             models, optimizers, opt_states, state = self.load_initial_state(model_key, load_optimizer=True)
             logger.info("Model size: %s", f"{get_pytree_param_count(models):,}")
-            logger.info("Optimizer size: %s", f"{get_pytree_param_count(optimizers):,}")
+            logger.info("Optimizer size: %s", f"{get_pytree_param_count(opt_states):,}")
 
             state = self.on_training_start(state)
 

xax/utils/experiments.py

@@ -111,8 +111,8 @@ class StateTimer:
 
     def step(self, state: State) -> None:
         cur_time = time.time()
-        num_steps = int((state.num_steps if state.phase == "train" else state.num_valid_steps).item())
-        num_samples = int((state.num_samples if state.phase == "train" else state.num_valid_samples).item())
+        num_steps = int(state.num_steps.item())
+        num_samples = int(state.num_samples.item())
         self.step_timer.step(num_steps, cur_time)
         self.sample_timer.step(num_samples, cur_time)
         self.iter_timer.step(cur_time)

{xax-0.2.22.dist-info → xax-0.2.23.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: xax
-Version: 0.2.22
+Version: 0.2.23
 Summary: A library for fast Jax experimentation
 Home-page: https://github.com/kscalelabs/xax
 Author: Benjamin Bolte

{xax-0.2.22.dist-info → xax-0.2.23.dist-info}/RECORD

@@ -1,4 +1,4 @@
-xax/__init__.py,sha256=Wh6x1Nohprb7ZxS_Y1aHPSo2xD7rAFSbmz31HLRl5og,15293
+xax/__init__.py,sha256=Q0boKxPtEUiiJ9j7Cdx51bLLFtYx3fPfCTG-o8o2Chk,15653
 xax/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 xax/requirements-dev.txt,sha256=qkscNkFzWd1S5fump-AKH53rR65v2x5FmboFdy_kKvs,128
 xax/requirements.txt,sha256=6qY-84e-sTmlfJNrSjwONQKqzAn5h8G_oGIhnhmfSr4,302
@@ -6,8 +6,9 @@ xax/cli/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 xax/cli/edit_config.py,sha256=LQUIlOS6hvPZyVEaMme3FP-62M0BKQPYavCwVDWuBLw,2600
 xax/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 xax/core/conf.py,sha256=d7Dp_GwKnaxtkztlSrJSM_LR0UYJX_FWTtceIWCBkxc,5138
-xax/core/state.py,sha256=KsNMnM_RgsZ2Ntc2pp4Fi6zG4rZb_89-kqmyGxDvyRg,4974
+xax/core/state.py,sha256=F9Tj3FfCw8zFKaDEoEGiThZE2ntYEtzNjnBX3pQ1g60,3826
 xax/nn/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+xax/nn/attention.py,sha256=0essK90OO3x9FxnUqU0DhufwXKRMN41zMtRCki5iKzQ,24742
 xax/nn/embeddings.py,sha256=bQGxBFxkLwi2MQLkRfGaHPH5P_KKB21HdI7VNWTKIOQ,11847
 xax/nn/functions.py,sha256=bA5kJYzMtFM8eUqBC086i355zJMAO7k_vPFNSDBI9-s,2814
 xax/nn/geom.py,sha256=A7WPefMvgwUNReZC7_HX1GmvHPASyghbaXaKsuhwDrE,7382
@@ -17,7 +18,7 @@ xax/nn/parallel.py,sha256=fnTiT7MsG7eQrJvqwjIz2Ifo3P27TuxIJzmpGYSa_dQ,4608
 xax/nn/ssm.py,sha256=8dLAcQ1hBaMT-kkHvwGu_ecxJeTY32WeMYmd4T4KtxA,10745
 xax/task/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 xax/task/base.py,sha256=TYANmjNcce4_V5ZSYLnE91PXRn7Nn0nT7hN8plW_Au0,8117
-xax/task/logger.py,sha256=W_BpluYvQai1lh1dDCAj-2_mWUC1buhwJncHygDffjc,41125
+xax/task/logger.py,sha256=Bmhl4mv08Aq49ZyX6BdjPIsPJK28e8s3mVFatM4IY2Q,41060
 xax/task/script.py,sha256=bMMIJoUtpSBvPp6-7bejTrajTXvSg0794sYLKdPIToE,972
 xax/task/task.py,sha256=UHMpnv__gqMcfbC_L-Hhk-DCnUYlFVsgbNf-v8o8B7U,1424
 xax/task/launchers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -41,10 +42,10 @@ xax/task/mixins/logger.py,sha256=6oXsJJyNUx6YT3q58FVXMZBUpMgjVkGre6BXFN20cVI,280
 xax/task/mixins/process.py,sha256=hqDEsMp_SL6ee97iq26-G0g49OcWZZaX82JD4F22eJU,1781
 xax/task/mixins/runnable.py,sha256=IYIsLd2k09g-_y6o44EhJqT7E6BpsyEMmsyLSuzqjtc,1979
 xax/task/mixins/step_wrapper.py,sha256=-Yu5Nft2CRw1JvZt6J_94SM1vqX8fk08IDK95Pmd2ew,1648
-xax/task/mixins/train.py,sha256=eueQc6P15Gkc9_lU7sp7fIHt4qrqOmhc4Xt6pCYZPkw,33636
+xax/task/mixins/train.py,sha256=TZatz5QwTfrNhQTiO2IqrmQY9P4Lay6FAD2VsQpWa54,33245
 xax/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 xax/utils/debugging.py,sha256=OtUdu-3tQsQtik0Q9UM-SNV46IbPjwrAfZcywzoB5d4,1940
-xax/utils/experiments.py,sha256=bj8BftSHT3fFzfiJ0Co0WvqWo0rUS8kQnQYpVvH8FTM,29942
+xax/utils/experiments.py,sha256=5k5hPYSaVjzoR_nm2Q3DAHMMYi3Bcp3N3PAQbwZq7Gg,29830
 xax/utils/jax.py,sha256=6cP95-rcjkRt1fefkZWJQhJhH0uUYWJB3w4NP1-aDp0,10136
 xax/utils/jaxpr.py,sha256=H7pWl48ROXIB1-ZPWYfOn-ou3EBMxYWIwc_A0reJQoo,2333
 xax/utils/logging.py,sha256=GAhTne2rdB4Fa1lzk06DMO15U8MTejn6XTClShC-ZtU,6622
@@ -58,9 +59,9 @@ xax/utils/data/collate.py,sha256=Rd9vMomr_S_zCa_Hi4dO-8ntzAfVwndIUtuXFA3iNcc,706
 xax/utils/types/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 xax/utils/types/frozen_dict.py,sha256=ebtHENhyUzSjyJTlbMaLtcckQIJ7EtgJiok_40TJZpo,4689
 xax/utils/types/hashable_array.py,sha256=l5iIcFmkYzfGeaZmcSoeFkthFASqM8xJYK3AXhZQYwc,992
-xax-0.2.22.dist-info/licenses/LICENSE,sha256=HCN2bImAzUOXldAZZI7JZ9PYq6OwMlDAP_PpX1HnuN0,1071
-xax-0.2.22.dist-info/METADATA,sha256=FtyVr4ve7FYrZCkDUWqneCAszYI-QSNs_ZTPrdbXUxg,1247
-xax-0.2.22.dist-info/WHEEL,sha256=QZxptf4Y1BKFRCEDxD4h2V0mBFQOVFLFEpvxHmIs52A,91
-xax-0.2.22.dist-info/entry_points.txt,sha256=uRC6rx5ce0bf-FblJaZSBMxxKFfMyoWTf8OWbBmLSe8,61
-xax-0.2.22.dist-info/top_level.txt,sha256=g4Au_r2XhvZ-lTybviH-Fh9g0zF4DAYHYxPue1-xbs8,4
-xax-0.2.22.dist-info/RECORD,,
+xax-0.2.23.dist-info/licenses/LICENSE,sha256=HCN2bImAzUOXldAZZI7JZ9PYq6OwMlDAP_PpX1HnuN0,1071
+xax-0.2.23.dist-info/METADATA,sha256=mA98vsIjdfb8XM2mN1vUb2VRVEPU4xf10IWLxxFJjmY,1247
+xax-0.2.23.dist-info/WHEEL,sha256=Nw36Djuh_5VDukK0H78QzOX-_FQEo6V37m3nkm96gtU,91
+xax-0.2.23.dist-info/entry_points.txt,sha256=uRC6rx5ce0bf-FblJaZSBMxxKFfMyoWTf8OWbBmLSe8,61
+xax-0.2.23.dist-info/top_level.txt,sha256=g4Au_r2XhvZ-lTybviH-Fh9g0zF4DAYHYxPue1-xbs8,4
+xax-0.2.23.dist-info/RECORD,,

{xax-0.2.22.dist-info → xax-0.2.23.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (80.6.0)
+Generator: setuptools (80.7.1)
 Root-Is-Purelib: true
 Tag: py3-none-any