Sequence parallelism quick follow-ups; remove ModelCallback (#2450)

* guard return if ring attn alrady registered

* add docs link, bits in multi-gpu docs, remove save model callback (subsumed by HF trainers)

* configurable heads_k_stride from ring-flash-attn hf adapter

src/axolotl/core/trainer_builder.py

@@ -69,7 +69,6 @@ from axolotl.utils.callbacks import (
     LossWatchDogCallback,
     SaveAxolotlConfigtoWandBCallback,
     SaveBetterTransformerModelCallback,
-    SaveModelCallback,
     bench_eval_callback_factory,
     causal_lm_bench_eval_callback_factory,
     log_prediction_callback_factory,
@@ -249,7 +248,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
         if self.cfg.gc_steps:
             callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
-        callbacks.append(SaveModelCallback())
 
         return callbacks
 
@@ -937,7 +935,6 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
 
     def get_callbacks(self):
         callbacks = super().get_callbacks()
-        callbacks.append(SaveModelCallback())
 
         return callbacks
 

src/axolotl/monkeypatch/attention/ring_attn.py

@@ -38,13 +38,19 @@ def set_ring_attn_group(ring_attn_group: dist.ProcessGroup | None):
     RING_ATTN_GROUP = ring_attn_group
 
 
-def register_ring_attn(sequence_parallel_degree: int):
+def register_ring_attn(sequence_parallel_degree: int, heads_k_stride: int | None):
     """
     Create ring attention group and substitute flash attn with ring flash attn.
 
     Args:
         sequence_parallel_degree: Sequence parallelism factor.
+        heads_k_stride: Sequence parallelism K head stride size. Passed
+            through to `ring_flash_attn.substitute_hf_flash_attn`.
     """
+    if get_ring_attn_group() is not None:
+        LOG.info("Ring attention already registered, exiting early...")
+        return
+
     LOG.info(
         "Enabling ring attention sequence parallelism: "
         f"each sequence will be processed across {sequence_parallel_degree} GPUs"
@@ -84,6 +90,11 @@ def register_ring_attn(sequence_parallel_degree: int):
     if rank == 0:
         LOG.info(f"Sequence parallel group assignments: {group_assignments}")
 
+    if heads_k_stride is None:
+        heads_k_stride = 1
+
     from ring_flash_attn import substitute_hf_flash_attn
 
-    substitute_hf_flash_attn(get_ring_attn_group(), sequence_parallel_degree)
+    substitute_hf_flash_attn(
+        process_group=get_ring_attn_group(), heads_k_stride=heads_k_stride
+    )

src/axolotl/utils/callbacks/__init__.py

@@ -816,27 +816,6 @@ class SaveAxolotlConfigtoWandBCallback(TrainerCallback):
         return control
 
 
-class SaveModelCallback(TrainerCallback):
-    """Callback to save model on train end"""
-
-    def on_step_end(  # pylint: disable=unused-argument
-        self,
-        args: TrainingArguments,
-        state: TrainerState,
-        control: TrainerControl,
-        **kwargs,
-    ):
-        # Save
-        if state.global_step >= state.max_steps:
-            control.should_save = True
-
-    def on_train_end(  # pylint: disable=unused-argument
-        self, args, state, control, **kwargs
-    ):
-        control.should_save = True
-        return control
-
-
 class GCCallback(TrainerCallback):
     """Callback to garbage collect torch cache"""
 

src/axolotl/utils/models.py

@@ -609,7 +609,10 @@ class ModelLoader:
             # Initialize ring attn for sequence parallelism. This must be done after
             # model init but before the first forward pass, since it modifies flash
             # attn to use ring comm for SP training across multiple GPUs.
-            register_ring_attn(self.cfg.sequence_parallel_degree)
+            register_ring_attn(
+                sequence_parallel_degree=self.cfg.sequence_parallel_degree,
+                heads_k_stride=self.cfg.heads_k_stride,
+            )
 
     def patch_attention(self) -> None:
         if hasattr(self.model_config, "model_type"):

src/axolotl/utils/schemas/config.py

@@ -248,6 +248,7 @@ class AxolotlInputConfig(
     val_set_size: float | None = Field(default=0.0)
 
     sequence_parallel_degree: int | None = None
+    heads_k_stride: int | None = None
 
     special_tokens: SpecialTokensConfig | None = None
     tokens: list[str] | None = None
@@ -1108,7 +1109,7 @@ class AxolotlInputConfig(
 
     @field_validator("sequence_parallel_degree", mode="before")
     @classmethod
-    def check_sequence_parallel_config(cls, value, info):
+    def check_sequence_parallel_degree(cls, value, info):
         if not value:
             value = 1