Add out_sharding to dot_product_attention

Author: samanklesaria

jax/_src/nn/functions.py

@@ -19,6 +19,7 @@
 from collections.abc import Sequence
 from functools import partial
 import operator
+import itertools
 import numpy as np
 from typing import Any, Literal, overload
 import warnings
@@ -903,7 +904,7 @@ def _apply_masks(logits, mask, is_causal, q_seqlen, kv_seqlen,
 
 def _dot_product_attention_core(query, key, value, bias, mask, is_causal,
                                 scale, q_seqlen, kv_seqlen, local_window_size,
-                                return_residual):
+                                return_residual, out_sharding = None, s_sharding = None):
   logits_dtype = jnp.promote_types(query.dtype, np.float32)
 
   # If the query and logits dtypes are different, then the default precision
@@ -920,13 +921,15 @@ def _dot_product_attention_core(query, key, value, bias, mask, is_causal,
   # Explicit precision will fail on platforms that don't support it. For example,
   # some GPUs do not support BF16_BF16_F32, and TPU does not support F16_F16_F32.
   # Use the default precision as a fallback in these cases.
+  out_sharding1 = permute_shard(out_sharding, "BTNH", "BNTS", S=s_sharding)
   try:
     logits = jnp_einsum.einsum(
         "BTNH,BSNH->BNTS",
         query,
         key,
         precision=precision,
         preferred_element_type=logits_dtype,
+        out_sharding=out_sharding1
     )
   except:  # pylint: disable=bare-except
     logits = jnp_einsum.einsum(
@@ -935,12 +938,13 @@ def _dot_product_attention_core(query, key, value, bias, mask, is_causal,
         key,
         precision=None,
         preferred_element_type=logits_dtype,
+        out_sharding=out_sharding1
     )
 
   logits *= jnp.array(scale, dtype=logits.dtype)
 
   if bias is not None:
-    logits = (logits + bias).astype(logits.dtype)
+    logits = lax.add(logits, bias).astype(logits.dtype)
 
   padded_logits = _apply_masks(logits, mask, is_causal, q_seqlen, kv_seqlen,
                                local_window_size)
@@ -949,7 +953,7 @@ def _dot_product_attention_core(query, key, value, bias, mask, is_causal,
   padded_logits = padded_logits.astype(np.float32)
   probs = softmax(padded_logits, axis=-1).astype(key.dtype)
 
-  encoded = jnp_einsum.einsum('BNTS,BSNH->BTNH', probs, value)
+  encoded = jnp_einsum.einsum('BNTS,BSNH->BTNH', probs, value, out_sharding=out_sharding)
   if q_seqlen is not None:
     mask = _get_padding_mask_encoded(encoded.shape[1], q_seqlen)
     encoded *= mask.astype(encoded.dtype)
@@ -972,7 +976,9 @@ def _dot_product_attention_xla(
     q_seqlen: Array | None,
     kv_seqlen: Array | None,
     local_window_size: tuple[int, int] | None,
-    return_residual: bool = False):
+    return_residual: bool = False,
+    out_sharding: Any = None,
+    s_sharding : Any = None):
 
   B, T, N, H = query.shape
   _, S, K, _ = key.shape
@@ -991,7 +997,7 @@ def _reshape_to_grouped(t):
   bias = _reshape_to_grouped(bias)
   mask = _reshape_to_grouped(mask)
   vmapped_fn = api.vmap(
-      _dot_product_attention_core,
+      partial(_dot_product_attention_core, out_sharding=out_sharding, s_sharding=s_sharding),
       in_axes=(3, None, None, 2, 2, None, None, None, None, None, None),
       out_axes=3,
   )
@@ -1056,6 +1062,8 @@ def dot_product_attention(
     local_window_size: int | tuple[int, int] | None = None,
     implementation: Literal['xla', 'cudnn'] | None = None,
     return_residual: bool = False,
+    out_sharding=None,
+    s_sharding=None,
 ):
   r"""Scaled dot product attention function.
 
@@ -1185,6 +1193,8 @@ def _check_shape_and_dtype(t: Array | None, shape: Sequence[int],
           kv_seqlen=key_value_seq_lengths,
           local_window_size=local_window_size,
           return_residual=return_residual,
+          out_sharding=out_sharding,
+          s_sharding=s_sharding,
       )
     case 'cudnn':
       use_padding = (
@@ -1490,3 +1500,45 @@ def log1mexp(x: ArrayLike) -> Array:
   )
 
 log1mexp.defjvps(lambda g, ans, x: g / jnp.expm1(x))
+
+def permute_shard(outer_t, in_spec: str, out_spec: str, **bindings):
+    """
+    Permute a tuple according to einops-like dimension specifications.
+
+    Args:
+        t: Input sharding or partition spec or tuple to permute
+        in_spec: Input specification string (e.g., "BTNH")
+        out_spec: Output specification string (e.g., "BNTH")
+        **bindings: Values for dimensions in out_spec not present in in_spec
+
+    Returns:
+        Permuted tuple according to out_spec
+
+    Examples:
+        >>> permute_shard(P('A', 'B', 'C'), "ABC", "ACB")
+        PartitionSpec('A', 'C', 'B')
+
+        >>> permute_shard(P('A', 'B', 'C'), "XC", "CX")
+        PartitionSpec('C', 'A', 'B')
+
+        >>> permute_shard(P('A', 'B', 'C'), "ABC", "ANZ", N="D")
+        PartitionSpec('A', 'D', None)
+    """
+    if outer_t is None:
+      return None
+    t = outer_t.spec if isinstance(outer_t, NamedSharding) else outer_t
+    if t is None:
+      return outer_t
+    d = len(t) - len(in_spec)
+    assert d >= 0, f"Tuple length {len(t)} is less than named dimensions {len(in_spec)}"
+    np_t = np.array(t)
+    dim_map : dict = {dim: [i+d] for (i, dim) in enumerate(in_spec)}
+    dim_map[in_spec[0]] = range(0, d+1)
+    new_spec = P(*list(itertools.chain.from_iterable(
+      np_t[dim_map[dim_name]] if dim_name in dim_map else [bindings.get(dim_name, None)]
+      for dim_name in out_spec
+    )))
+    if isinstance(outer_t, NamedSharding):
+        return NamedSharding(outer_t.mesh, new_spec)
+    else:
+        return new_spec