Add configurable data_seed independent of training seed

kempnerforge/config/training.py

@@ -25,7 +25,8 @@ class TrainConfig:
     max_steps: int = 100000
     grad_accum_steps: int = 1
     grad_clip_norm: float = 1.0
-    seed: int = 42
+    seed: int = 42  # Seeds parameter init, dropout, and (by default) data order
+    data_seed: int | None = None  # Overrides data/batch order only; None falls back to seed
     compile_model: bool = True
     mixed_precision: Literal["bf16", "fp16", "fp32", "fp8"] = "bf16"
     activation_checkpointing: ActivationCheckpointing = ActivationCheckpointing.none
@@ -67,3 +68,15 @@ def param_dtype(self) -> torch.dtype:
     def is_fp8(self) -> bool:
         """Whether FP8 mixed precision is enabled."""
         return self.mixed_precision == "fp8"
+
+    @property
+    def effective_data_seed(self) -> int:
+        """Seed for data shuffling / batch composition.
+
+        Falls back to ``seed`` when ``data_seed`` is unset, so existing configs
+        reproduce their current trajectory. Kept independent from ``seed``
+        (parameter init) so stability studies can vary batch order while holding
+        initialization fixed. Must stay identical across data-parallel ranks so
+        the global shuffle is consistent before rank partitioning.
+        """
+        return self.data_seed if self.data_seed is not None else self.seed

scripts/train.py

@@ -319,7 +319,7 @@ def main() -> None:
             _pad_id = _tok.eos_token_id if _tok.eos_token_id is not None else 0
         collator = VLMCollator(pad_id=int(_pad_id), max_text_len=vlm_cfg.max_text_len)
         sampler = DistributedSampler(
-            dataset, num_replicas=dp_size, rank=dp_rank, shuffle=True, seed=tc.seed
+            dataset, num_replicas=dp_size, rank=dp_rank, shuffle=True, seed=tc.effective_data_seed
         )
         dataloader = StatefulDataLoader(
             dataset,
@@ -373,7 +373,7 @@ def main() -> None:
             num_replicas=dp_size,
             rank=dp_rank,
             shuffle=True,
-            seed=tc.seed,
+            seed=tc.effective_data_seed,
             temperature=config.data.mix_temperature,
         )
         dataloader = StatefulDataLoader(
@@ -401,7 +401,7 @@ def main() -> None:
             num_replicas=dp_size,
             rank=dp_rank,
             shuffle=True,
-            seed=tc.seed,
+            seed=tc.effective_data_seed,
         )
         dataloader = StatefulDataLoader(
             dataset,
@@ -429,7 +429,7 @@ def main() -> None:
                 dataset_config=config.data.hf_dataset_config,
                 rank=dp_rank,
                 world_size=dp_size,
-                seed=tc.seed,
+                seed=tc.effective_data_seed,
                 pack_sequences=config.data.pack_sequences,
             )
             dataloader = TorchDataLoader(
@@ -463,7 +463,7 @@ def main() -> None:
                 num_replicas=dp_size,
                 rank=dp_rank,
                 shuffle=True,
-                seed=tc.seed,
+                seed=tc.effective_data_seed,
             )
             dataloader = StatefulDataLoader(
                 dataset,

tests/unit/test_config.py

@@ -169,6 +169,18 @@ def test_rejects_zero_grad_accum(self):
         with pytest.raises(ValueError, match="grad_accum_steps must be positive"):
             TrainConfig(grad_accum_steps=0)
 
+    def test_data_seed_defaults_to_none(self):
+        assert TrainConfig().data_seed is None
+
+    def test_effective_data_seed_falls_back_to_seed(self):
+        t = TrainConfig(seed=123)
+        assert t.effective_data_seed == 123
+
+    def test_effective_data_seed_overrides_when_set(self):
+        t = TrainConfig(seed=123, data_seed=456)
+        assert t.effective_data_seed == 456
+        assert t.seed == 123  # init seed stays independent
+
 
 # ---------------------------------------------------------------------------
 # OptimizerConfig