updates

src/axolotl/utils/models.py

@@ -66,7 +66,6 @@ from axolotl.utils.distributed import (
 from axolotl.utils.gradient_checkpointing import hf_grad_checkpoint_offload_wrapper
 from axolotl.utils.lora_embeddings import get_linear_embedding_layers
 from axolotl.utils.model_shard_quant import load_sharded_model, load_sharded_model_quant
-from axolotl.utils.ring_attn import register_ring_attn
 
 LOG = logging.getLogger("axolotl")
 
@@ -549,6 +548,12 @@ class ModelLoader:
             patch_self_attn_lora(self.cfg)
 
         if self.cfg.sequence_parallel_size > 1:
+            from axolotl.utils.ring_attn import register_ring_attn
+
+            # Initialize ring attention for sequence parallelism if enabled.
+            # This must be done after model initialization but before the first forward pass,
+            # as it modifies the flash attention implementation to use ring communication
+            # patterns for efficient sequence-parallel training across multiple GPUs.
             register_ring_attn(self.cfg.sequence_parallel_size)
 
     def patch_attention(self) -> None:

src/axolotl/utils/ring_attn.py

@@ -1,6 +1,14 @@
+"""Ring attention group registration and utils."""
+
 import torch.distributed as dist
+from accelerate.logging import get_logger
 from ring_flash_attn import substitute_hf_flash_attn
 
+from axolotl.logging_config import configure_logging
+
+configure_logging()
+LOG = get_logger(__name__)
+
 RING_ATTN_GROUP = None
 
 
@@ -9,17 +17,21 @@ def get_ring_attn_group():
 
 
 def set_ring_attn_group(ring_attn_group):
-    global RING_ATTN_GROUP
+    global RING_ATTN_GROUP  # pylint: disable=global-statement
     RING_ATTN_GROUP = ring_attn_group
 
 
-def register_ring_attn(sequence_parallel_size):
+def register_ring_attn(sequence_parallel_size: int):
     """
-    Create ring attention group and substitute flash attention with ring flash
-    attention.
+    Create ring attention group and substitute flash attn with ring flash attn.
+
+    Args:
+        sequence_parallel_size: Sequence parallelism factor.
     """
-    if sequence_parallel_size == 1:
-        return
+    LOG.info(
+        "Enabling ring attention sequence parallelism: "
+        f"each sequence will be processed across {sequence_parallel_size} GPUs"
+    )
 
     world_size = dist.get_world_size()
     assert world_size % sequence_parallel_size == 0, (

src/axolotl/utils/schemas/config.py

@@ -1104,6 +1104,17 @@ class AxolotlInputConfig(
 
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_sequence_parallel_config(cls, data):
+        if data.get("sequence_parallel_size") > 1:
+            if not data.get("flash_attention"):
+                raise ValueError(
+                    "flash_attention: true must be set with sequence_parallel_size > 1"
+                )
+
+        return data
+
 
 class AxolotlConfigWCapabilities(AxolotlInputConfig):
     """wrapper to valdiate gpu capabilities with the configured options"""