Follow up fix custom flash block sizes fallback in Ulysses attention

Ensure that block sizes and heads_per_tile fall back to default values
when resolved as None from CustomFlashBlockSizes dataclass. This fixes
a TypeError in ulysses_custom attention when heads_per_tile is
not specified.

Author: ninatu

src/maxdiffusion/models/attention_flax.py

@@ -363,24 +363,25 @@ def _extract_custom_block_sizes(flash_block_sizes):
   if flash_block_sizes is not None:
     if isinstance(flash_block_sizes, dict):
       get = flash_block_sizes.get
-      bq = get("block_q", bq)
-      bkv = get("block_kv", bkv)
-      bkv_compute = get("block_kv_compute", bkv_compute)
-      bkv_compute_in = get("block_kv_compute_in", bkv_compute_in)
-      heads_per_tile = get("heads_per_tile", heads_per_tile)
-      vmem_limit_bytes = get("vmem_limit_bytes", vmem_limit_bytes)
+      bq = get("block_q", None) or bq
+      bkv = get("block_kv", None) or bkv
+      bkv_compute = get("block_kv_compute", None) or bkv_compute
+      bkv_compute_in = get("block_kv_compute_in", None) or bkv_compute_in
+      # A BlockSizes object carries heads_per_tile=None when the config dict omitted
+      # it; getattr then returns that None instead of the default, so coerce it back
+      # to 1 (the custom-kernel default) to keep the `heads_per_tile > 1` guards safe.
+      heads_per_tile = get("heads_per_tile", None) or heads_per_tile
+      vmem_limit_bytes = get("vmem_limit_bytes", None) or vmem_limit_bytes
     else:
-      bq = getattr(flash_block_sizes, "block_q", bq)
-      bkv = getattr(flash_block_sizes, "block_kv", bkv)
-      bkv_compute = getattr(flash_block_sizes, "block_kv_compute", bkv_compute)
-      bkv_compute_in = getattr(flash_block_sizes, "block_kv_compute_in", bkv_compute_in)
-      heads_per_tile = getattr(flash_block_sizes, "heads_per_tile", heads_per_tile)
-      vmem_limit_bytes = getattr(flash_block_sizes, "vmem_limit_bytes", vmem_limit_bytes)
-  # A BlockSizes object carries heads_per_tile=None when the config dict omitted
-  # it; getattr then returns that None instead of the default, so coerce it back
-  # to 1 (the custom-kernel default) to keep the `heads_per_tile > 1` guards safe.
-  if heads_per_tile is None:
-    heads_per_tile = 1
+      bq = getattr(flash_block_sizes, "block_q", None) or bq
+      bkv = getattr(flash_block_sizes, "block_kv", None) or bkv
+      bkv_compute = getattr(flash_block_sizes, "block_kv_compute", None) or bkv_compute
+      bkv_compute_in = getattr(flash_block_sizes, "block_kv_compute_in", None) or bkv_compute_in
+      # A BlockSizes object carries heads_per_tile=None when the config dict omitted
+      # it; getattr then returns that None instead of the default, so coerce it back
+      # to 1 (the custom-kernel default) to keep the `heads_per_tile > 1` guards safe.
+      heads_per_tile = getattr(flash_block_sizes, "heads_per_tile", None) or heads_per_tile
+      vmem_limit_bytes = getattr(flash_block_sizes, "vmem_limit_bytes", None) or vmem_limit_bytes
   return bq, bkv, bkv_compute, bkv_compute_in, heads_per_tile, vmem_limit_bytes
 
 
@@ -1113,9 +1114,7 @@ def wrap_ulysses_ring_attention(query, key, value):
           use_fixed_m=use_fixed_m,
       )
       if use_fixed_m:
-        attention_output = jnp.swapaxes(
-            jax.vmap(splash_kernel, in_axes=(0, 0, 0, None))(query, key, value, mk_arr), 2, 3
-        )
+        attention_output = jnp.swapaxes(jax.vmap(splash_kernel, in_axes=(0, 0, 0, None))(query, key, value, mk_arr), 2, 3)
       else:
         attention_output = jnp.swapaxes(jax.vmap(splash_kernel, in_axes=(0, 0, 0))(query, key, value), 2, 3)
     else: