It worth not hiding exceptions (#21432)

* Worth not to hide exceptions

What do you think about logging of exceptions, instead of hiding them? + There is an example of such hidden exception when head_shards and q_seq_shards were just not initialized because get_dist()==None.

    And despite hardware supports flash attention - this code silently falls back to the standard with O(N^2) memory.

* pre-commit run

Author: WIgor

keras/src/backend/jax/nn.py

@@ -5,6 +5,7 @@
 import jax
 import jax.experimental.sparse as jax_sparse
 import jax.numpy as jnp
+from absl import logging
 from jax import lax
 from jax import nn as jnn
 from jax.experimental.pallas.ops.tpu.splash_attention import (
@@ -1256,6 +1257,9 @@ def dot_product_attention(
     # TPU-specific flash attention path
     if is_tpu and flash_attention:
         # Get sharding parameters from distribution context
+        head_shards = 1
+        # Typically keep q_seq_shards=1 for best performance
+        q_seq_shards = 1
         try:
             from keras.src.distribution.distribution_lib import ModelParallel
             from keras.src.distribution.distribution_lib import (
@@ -1270,12 +1274,12 @@ def dot_product_attention(
                     model_dim_index = mesh.axis_names.index("model")
                     # Set head_shards based on the model dimension of the mesh
                     head_shards = mesh.shape[model_dim_index]
-                    # Typically keep q_seq_shards=1 for best performance
-                    q_seq_shards = 1
         except (ImportError, ValueError, AttributeError):
             # Use default values if detection fails
-            head_shards = 1
-            q_seq_shards = 1
+            logging.exception(
+                "Failed to determine distribution context for sharding. "
+                "Using default head_shards=1 and q_seq_shards=1."
+            )
         # Transpose to ('batch', 'heads', 'length', 'head_dim')
         query_tpu_layout = jnp.transpose(query, axes=(0, 2, 1, 3))
         key_tpu_layout = jnp.transpose(key, axes=(0, 2, 1, 3))
@@ -1328,24 +1332,17 @@ def dot_product_attention(
             # Transpose output back to Keras layout
             return jnp.transpose(output, axes=(0, 2, 1, 3))
         except Exception:
+            logging.exception(
+                "Failed to apply Splash kernel for flash attention. "
+                "Falling back to JAX native dot_product_attention."
+            )
             flash_attention = False
 
     # JAX native dot_product_attention for GPU or fallback for TPU
     if hasattr(jax.nn, "dot_product_attention"):
-        try:
-            return jax.nn.dot_product_attention(
-                query,
-                key,
-                value,
-                bias=bias,
-                mask=mask,
-                scale=scale,
-                is_causal=is_causal,
-                implementation="cudnn" if flash_attention else "xla",
-            )
-        except Exception:
-            # If flash attention fails, fall back to XLA implementation
-            if flash_attention:
+        impls = ["cudnn", "xla"] if flash_attention else ["xla"]
+        for impl in impls:
+            try:
                 return jax.nn.dot_product_attention(
                     query,
                     key,
@@ -1354,9 +1351,13 @@ def dot_product_attention(
                     mask=mask,
                     scale=scale,
                     is_causal=is_causal,
-                    implementation="xla",
+                    implementation=impl,
+                )
+            except Exception:
+                logging.exception(
+                    f"Failed to apply {impl} implementation of "
+                    "jax.nn.dot_product_attention."
                 )
-            raise
 
     if flash_attention:
         raise RuntimeError(