feat: configure chunked Ulysses all-to-all

Add a ulysses_attention_chunks attention config to split the Ulysses all-to-all into head-group passes. The chunked path lets XLA overlap all-to-all collectives with head-parallel local attention compute while preserving the existing single-shot path by default.

Apply the same chunking to plain Ulysses and Ulysses+Ring, and allow the final chunk to carry the remainder when the requested chunk count does not divide the Ulysses head groups evenly.

Add mocked attention tests for numerical and layout equivalence across chunk counts.

Author: csgoogle

src/maxdiffusion/configs/base_wan_27b.yml

@@ -94,6 +94,8 @@ use_base2_exp: True
 use_experimental_scheduler: True
 # For attention=ulysses_ring, hidden Ulysses shard count; ring shards are context / this.
 ulysses_shards: -1
+# Splits Ulysses all-to-all into head-group chunks. The last chunk carries any remainder.
+ulysses_attention_chunks: 1
 flash_min_seq_length: 4096
 dropout: 0.0
 

src/maxdiffusion/models/attention_flax.py

@@ -413,6 +413,44 @@ def _build_padding_segment_ids(
   return segment_ids_cls(q=q_segment_ids, kv=kv_segment_ids)
 
 
+def _ulysses_head_chunk_ranges(num_heads: int, ulysses_shards: int, num_chunks: int):
+  """Build head-axis ranges for chunked Ulysses all-to-all.
+
+  The Ulysses all-to-all splits each local chunk's head axis over
+  `ulysses_shards`, so every returned range length is a multiple of
+  `ulysses_shards`. When `num_chunks` does not evenly divide the number of
+  Ulysses-sized head groups, earlier chunks get the floor-sized range and the
+  final chunk carries the remainder.
+
+  Returns:
+    A list of `(start, end)` half-open ranges over the head axis. Concatenating
+    tensors sliced with these ranges along the head axis restores the original
+    head layout. For `num_chunks <= 1`, returns `[(0, num_heads)]`, which is the
+    unchunked all-to-all path.
+  """
+  if num_chunks <= 1:
+    return [(0, num_heads)]
+  if num_heads % ulysses_shards != 0:
+    raise ValueError(
+        "Ulysses attention requires the number of heads to be divisible by the Ulysses shard count, "
+        f"got heads={num_heads} and ulysses_shards={ulysses_shards}."
+    )
+
+  head_groups = num_heads // ulysses_shards
+  num_chunks = min(num_chunks, head_groups)
+  regular_groups_per_chunk = max(1, head_groups // num_chunks)
+
+  ranges = []
+  start_group = 0
+  for chunk_idx in range(num_chunks):
+    end_group = head_groups if chunk_idx == num_chunks - 1 else min(start_group + regular_groups_per_chunk, head_groups)
+    if start_group >= end_group:
+      break
+    ranges.append((start_group * ulysses_shards, end_group * ulysses_shards))
+    start_group = end_group
+  return ranges
+
+
 def _tpu_flash_attention(
     query: jax.Array,
     key: jax.Array,
@@ -640,6 +678,7 @@ def _ulysses_attention(
     use_custom_kernel: bool = False,
     use_base2_exp: bool = True,
     use_experimental_scheduler: bool = False,
+    ulysses_attention_chunks: int = 1,
 ) -> jax.Array:
   """Ulysses sequence-parallel attention.
 
@@ -662,6 +701,7 @@ def _ulysses_attention(
         "Ulysses attention requires the number of heads to be divisible by the context shard count, "
         f"got heads={num_heads} and context_shards={num_shards}."
     )
+
   if not use_custom_kernel:
     block_sizes = _select_flash_block_sizes(query, key, flash_block_sizes, dtype, "flash")
 
@@ -762,7 +802,21 @@ def wrap_ulysses_attention(query, key, value):
         "Warning, batch dimension should be shardable among the devices in data and fsdp"
         f" axis, batch dimension: {query.shape[0]}, devices_in_batch_sharding: {devices_in_batch_sharding}"
     )
-  x = wrap_ulysses_attention(query, key, value)
+  head_chunk_ranges = _ulysses_head_chunk_ranges(num_heads, num_shards, ulysses_attention_chunks)
+  if len(head_chunk_ranges) > 1:
+    # Run Ulysses all-to-all per head group so XLA can overlap one group's
+    # collective with another group's head-parallel local attention compute.
+    chunk_outputs = [
+        wrap_ulysses_attention(
+            query[:, start:end],
+            key[:, start:end],
+            value[:, start:end],
+        )
+        for start, end in head_chunk_ranges
+    ]
+    x = jnp.concatenate(chunk_outputs, axis=1)
+  else:
+    x = wrap_ulysses_attention(query, key, value)
   x = x[:, :, :orig_q_seq_len, :]
   x = _reshape_heads_to_head_dim(x)
 
@@ -787,6 +841,7 @@ def _ulysses_ring_attention(
     use_base2_exp: bool = False,
     use_experimental_scheduler: bool = False,
     ulysses_shards: int = -1,
+    ulysses_attention_chunks: int = 1,
 ) -> jax.Array:
   """2D context-parallel attention using a private Ulysses x ring mesh.
 
@@ -916,7 +971,21 @@ def wrap_ulysses_ring_attention(query, key, value):
         "Warning, batch dimension should be shardable among the devices in data and fsdp"
         f" axis, batch dimension: {query.shape[0]}, devices_in_batch_sharding: {devices_in_batch_sharding}"
     )
-  x = wrap_ulysses_ring_attention(query, key, value)
+  head_chunk_ranges = _ulysses_head_chunk_ranges(num_heads, num_ulysses_shards, ulysses_attention_chunks)
+  if len(head_chunk_ranges) > 1:
+    # Run the Ulysses phase per head group before the ring phase so XLA can
+    # overlap all-to-all collectives with head-parallel local attention compute.
+    chunk_outputs = [
+        wrap_ulysses_ring_attention(
+            query[:, start:end],
+            key[:, start:end],
+            value[:, start:end],
+        )
+        for start, end in head_chunk_ranges
+    ]
+    x = jnp.concatenate(chunk_outputs, axis=1)
+  else:
+    x = wrap_ulysses_ring_attention(query, key, value)
   x = jax.lax.with_sharding_constraint(x, q_axis_names)
   x = x[:, :, :orig_q_seq_len, :]
   x = _reshape_heads_to_head_dim(x)
@@ -1207,6 +1276,7 @@ def ulysses_custom_kernel(q, k, v, context):
       use_custom_kernel=True,
       use_base2_exp=context.get("use_base2_exp", True),
       use_experimental_scheduler=context.get("use_experimental_scheduler", False),
+      ulysses_attention_chunks=context["ulysses_attention_chunks"],
   )
 
 
@@ -1271,6 +1341,7 @@ def ulysses_kernel(q, k, v, context):
       mask_padding_tokens=context["mask_padding_tokens"],
       residual_checkpoint_name=context["residual_checkpoint_name"],
       attention_mask=context["attention_mask"],
+      ulysses_attention_chunks=context["ulysses_attention_chunks"],
   )
 
 
@@ -1292,6 +1363,7 @@ def ulysses_ring_kernel(q, k, v, context):
       use_base2_exp=context["use_base2_exp"],
       use_experimental_scheduler=context["use_experimental_scheduler"],
       ulysses_shards=context["ulysses_shards"],
+      ulysses_attention_chunks=context["ulysses_attention_chunks"],
   )
 
 
@@ -1404,6 +1476,7 @@ def _apply_attention(
     use_base2_exp: bool = False,
     use_experimental_scheduler: bool = False,
     ulysses_shards: int = -1,
+    ulysses_attention_chunks: int = 1,
 ):
   """Routes to different attention kernels using a module-level registry."""
 
@@ -1435,6 +1508,7 @@ def _apply_attention(
       "use_base2_exp": use_base2_exp,
       "use_experimental_scheduler": use_experimental_scheduler,
       "ulysses_shards": ulysses_shards,
+      "ulysses_attention_chunks": ulysses_attention_chunks,
       "dim_head": dim_head,
       "split_head_dim": split_head_dim,
       "float32_qk_product": float32_qk_product,
@@ -1648,11 +1722,13 @@ def __init__(
       use_base2_exp: bool = False,
       use_experimental_scheduler: bool = False,
       ulysses_shards: int = -1,
+      ulysses_attention_chunks: int = 1,
   ):
     self.dpa_layer = None
     self.use_base2_exp = use_base2_exp
     self.use_experimental_scheduler = use_experimental_scheduler
     self.ulysses_shards = ulysses_shards
+    self.ulysses_attention_chunks = ulysses_attention_chunks
     if attention_kernel == "cudnn_flash_te":
       from transformer_engine.jax.flax.transformer import DotProductAttention  # pytype: disable=import-error
 
@@ -1716,6 +1792,7 @@ def apply_attention(self, query: Array, key: Array, value: Array, attention_mask
         use_base2_exp=self.use_base2_exp if hasattr(self, "use_base2_exp") else False,
         use_experimental_scheduler=self.use_experimental_scheduler if hasattr(self, "use_experimental_scheduler") else False,
         ulysses_shards=(self.ulysses_shards if hasattr(self, "ulysses_shards") else -1),
+        ulysses_attention_chunks=(self.ulysses_attention_chunks if hasattr(self, "ulysses_attention_chunks") else 1),
     )
 
 
@@ -1737,6 +1814,7 @@ class AttentionOp(nn.Module):
   use_base2_exp: bool = False
   use_experimental_scheduler: bool = False
   ulysses_shards: int = -1
+  ulysses_attention_chunks: int = 1
 
   def setup(self):
     self.dpa_layer = None
@@ -1785,6 +1863,7 @@ def apply_attention(self, query: Array, key: Array, value: Array, attention_mask
         use_base2_exp=self.use_base2_exp,
         use_experimental_scheduler=self.use_experimental_scheduler,
         ulysses_shards=self.ulysses_shards,
+        ulysses_attention_chunks=self.ulysses_attention_chunks,
     )
 
 
@@ -1827,6 +1906,7 @@ def __init__(
         "use_base2_exp": False,
         "use_experimental_scheduler": False,
         "ulysses_shards": -1,
+        "ulysses_attention_chunks": 1,
         **(attention_config or {}),
     }
 
@@ -1878,6 +1958,7 @@ def __init__(
         use_base2_exp=attention_config["use_base2_exp"],
         use_experimental_scheduler=attention_config["use_experimental_scheduler"],
         ulysses_shards=attention_config["ulysses_shards"],
+        ulysses_attention_chunks=attention_config["ulysses_attention_chunks"],
     )
     # None axes corresponds to the stacked weights across all blocks
     # because of the use of nnx.vmap and nnx.scan.

src/maxdiffusion/models/wan/transformers/transformer_wan.py

@@ -360,6 +360,7 @@ def __init__(
         "use_base2_exp": False,
         "use_experimental_scheduler": False,
         "ulysses_shards": -1,
+        "ulysses_attention_chunks": 1,
         **(attention_config or {}),
     }
 
@@ -584,6 +585,7 @@ def __init__(
         "use_base2_exp": False,
         "use_experimental_scheduler": False,
         "ulysses_shards": -1,
+        "ulysses_attention_chunks": 1,
         **(attention_config or {}),
     }
 

src/maxdiffusion/models/wan/transformers/transformer_wan_vace.py

@@ -100,6 +100,7 @@ def __init__(
         "use_base2_exp": False,
         "use_experimental_scheduler": False,
         "ulysses_shards": -1,
+        "ulysses_attention_chunks": 1,
         **(attention_config or {}),
     }
 
@@ -348,6 +349,7 @@ def __init__(
         "use_base2_exp": False,
         "use_experimental_scheduler": False,
         "ulysses_shards": -1,
+        "ulysses_attention_chunks": 1,
         **(attention_config or {}),
     }
 

src/maxdiffusion/pipelines/wan/wan_pipeline.py

@@ -174,6 +174,7 @@ def create_model(rngs: nnx.Rngs, wan_config: dict):
       "use_base2_exp": config.use_base2_exp,
       "use_experimental_scheduler": config.use_experimental_scheduler,
       "ulysses_shards": getattr(config, "ulysses_shards", -1),
+      "ulysses_attention_chunks": getattr(config, "ulysses_attention_chunks", 1),
   }
 
   # 2. eval_shape - will not use flops or create weights on device

src/maxdiffusion/pipelines/wan/wan_vace_pipeline_2_1.py

@@ -90,6 +90,7 @@ def create_model(rngs: nnx.Rngs, wan_config: dict):
       "use_base2_exp": config.use_base2_exp,
       "use_experimental_scheduler": config.use_experimental_scheduler,
       "ulysses_shards": getattr(config, "ulysses_shards", -1),
+      "ulysses_attention_chunks": getattr(config, "ulysses_attention_chunks", 1),
   }
 
   wan_config["scan_layers"] = False