diff --git a/ffn/jax/accelerator_utils.py b/ffn/jax/accelerator_utils.py
new file mode 100644
index 0000000..2f22300
--- /dev/null
+++ b/ffn/jax/accelerator_utils.py
@@ -0,0 +1,59 @@
+# Copyright 2026 Google Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Accelerator topology and batch size utilities."""
+
+import dataclasses
+import jax
+
+
+@dataclasses.dataclass(frozen=True)
+class AcceleratorTopologyInfo:
+  global_batch_size: int
+  host_batch_size: int
+  cores_per_chip: int
+  num_chips: int
+  local_chips: int
+
+
+def get_accelerator_topology_info(
+    per_device_batch_size: int,
+) -> AcceleratorTopologyInfo:
+  """Computes topology info and batch sizes.
+
+  Args:
+    per_device_batch_size: Desired batch size per device (core).
+
+  Returns:
+    AcceleratorTopologyInfo with topology and batch size info.
+  """
+  devices = jax.local_devices()
+  if devices and hasattr(devices[0], 'core_on_chip'):
+    cores_per_chip = max(d.core_on_chip for d in devices) + 1
+  else:
+    cores_per_chip = 1
+
+  num_chips = jax.device_count() // cores_per_chip
+  local_chips = jax.local_device_count() // cores_per_chip
+
+  global_batch_size = per_device_batch_size * num_chips
+  host_batch_size = per_device_batch_size * local_chips
+
+  return AcceleratorTopologyInfo(
+      global_batch_size=global_batch_size,
+      host_batch_size=host_batch_size,
+      cores_per_chip=cores_per_chip,
+      num_chips=num_chips,
+      local_chips=local_chips,
+  )
diff --git a/ffn/jax/input_pipeline.py b/ffn/jax/input_pipeline.py
index f4fc4c9..36e6173 100644
--- a/ffn/jax/input_pipeline.py
+++ b/ffn/jax/input_pipeline.py
@@ -24,6 +24,7 @@
 from connectomics.common import bounding_box
 from connectomics.common import utils
 from ffn.input import volume
+from ffn.jax import accelerator_utils
 from ffn.jax import tracker
 from ffn.training import examples
 from ffn.training import inputs
@@ -151,7 +152,10 @@ def _add_ffn_data(ex: volume.Example) -> volume.Example:
         patches=(emt - config.image_mean) / config.image_stddev,
     )
 
-  batch_size = config.per_device_batch_size * jax.local_device_count()
+  topo_info = accelerator_utils.get_accelerator_topology_info(
+      config.per_device_batch_size
+  )
+  batch_size = topo_info.host_batch_size
 
   if cfg.sampling.vsi_coords:
     num_examples = getattr(config, 'train_num_coords', 100_000_000)
diff --git a/ffn/jax/train.py b/ffn/jax/train.py
index 788b5e7..2a897c4 100644
--- a/ffn/jax/train.py
+++ b/ffn/jax/train.py
@@ -27,6 +27,7 @@
 from clu import parameter_overview
 from connectomics.jax import training
 from etils import epath
+from ffn.jax import accelerator_utils
 from ffn.jax import input_pipeline
 from ffn.jax import tracker
 from ffn.training import examples
@@ -521,13 +522,27 @@ def train_and_evaluate(
     train_iter = checkpointed_state['train_iter']
   initial_step = int(state.step) + 1
 
-  global_batch_size = config.per_device_batch_size * jax.device_count()
-  host_batch_size = config.per_device_batch_size * jax.local_device_count()
+  # Compute batch sizes based on chip count, not core count. Multi-core
+  # chips (e.g., tpu7x with 2 cores/chip) expose each core as a
+  # separate JAX device, which would accidentally double the batch size.
+  topo_info = accelerator_utils.get_accelerator_topology_info(
+      config.per_device_batch_size
+  )
+
+  logging.info(
+      'cores_per_chip=%d, num_chips=%d (global), local_chips=%d, '
+      'global_batch_size=%d, host_batch_size=%d',
+      topo_info.cores_per_chip,
+      topo_info.num_chips,
+      topo_info.local_chips,
+      topo_info.global_batch_size,
+      topo_info.host_batch_size,
+  )
 
   # Upper bound. The real number will be lower as not all steps are
   # taken for every example.
   steps_per_epoch = (
-      num_total_examples // global_batch_size * (len(fov_shifts) + 1)
+      num_total_examples // topo_info.global_batch_size * (len(fov_shifts) + 1)
   )
   num_train_steps = steps_per_epoch * config.num_epochs
   logging.info(
@@ -593,7 +608,9 @@ def train_fn(state, batch, loss_scale):
   logging.info('Starting training loop at step %d.', initial_step)
   hooks = []
   report_progress = training.ReportProgress(
-      global_batch_size, num_train_steps=num_train_steps, writer=writer
+      topo_info.global_batch_size,
+      num_train_steps=num_train_steps,
+      writer=writer,
   )
   if jax.process_index() == 0:
     hooks.append(report_progress)
@@ -608,7 +625,7 @@ def train_fn(state, batch, loss_scale):
       info,
       config,
       seed_shape=tuple(train_canvas_size(info, config).tolist()[::-1]),
-      batch_size=host_batch_size,
+      batch_size=topo_info.host_batch_size,
       jmp_policy=jmp_policy,
   )
 
@@ -623,9 +640,9 @@ def _reshape(x):
       per_device_data = np.split(x, len(mesh.local_devices), axis=0)
 
       on_dev = jax.device_put(per_device_data, mesh.local_devices)
-      global_shape = (
-          len(batch_sharding.device_set) * config.per_device_batch_size,
-      ) + per_device_data[0].shape[1:]
+      global_shape = (topo_info.global_batch_size,) + per_device_data[0].shape[
+          1:
+      ]
       return jax.make_array_from_single_device_arrays(
           global_shape, batch_sharding, on_dev
       )
@@ -672,16 +689,12 @@ def _reshape(x):
           )
 
         with training.MeasureTime(timings, 'update_seed'):
-          host_local_seeds = []  # [b, z, y, x, 1] * num_devices
-          dev_to_slice = batch_sharding.addressable_devices_indices_map(
-              updated_seed.shape
-          )
-
           # Ensure device order is the same as that used to build the
           # global array in postprocess_batch().
-          assert list(dev_to_slice.keys()) == list(mesh.local_devices)
-          for slc in dev_to_slice.values():
-            host_local_seeds.append(updated_seed[slc])
+          shard_by_device = {
+              s.device: s.data for s in updated_seed.addressable_shards
+          }
+          host_local_seeds = [shard_by_device[d] for d in mesh.local_devices]
 
           batch_iter.update_seeds(host_local_seeds)
 
diff --git a/ffn/training/mask.py b/ffn/training/mask.py
index 06607d5..0d8093b 100644
--- a/ffn/training/mask.py
+++ b/ffn/training/mask.py
@@ -45,7 +45,8 @@ def crop(tensor, offset, crop_shape):
     off_y = shape[-3] // 2 - crop_shape[1] // 2 + offset[1]
     off_x = shape[-2] // 2 - crop_shape[0] // 2 + offset[0]
 
-    # Note: native indexing syntax not used below due to TPU compatibility.
+    # Note: native indexing syntax not used below due to accelerator
+    # compatibility.
     if len(offset) == 2:
       cropped = tf.slice(
           tensor,