Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Stop enable_padding_free_metadata writing seq_lengths into the caller's examples (#8049)

Vineeth Sai Varikuntla5d58a2c847e6
diff --git a/tests/utils/test_packing.py b/tests/utils/test_packing.py
index 35a9b55df5c..70b5a36d784 100644
--- a/tests/utils/test_packing.py
+++ b/tests/utils/test_packing.py
@@ -27,6 +27,7 @@ from unsloth.utils.packing import (
patch_hybrid_linear_attention_varlen,
)
+import copy
import inspect
import logging
from contextlib import ExitStack
@@ -160,6 +161,68 @@ def test_mask_packed_sequence_boundaries_across_multiple_rows():
assert torch.any(flat != -100)
+def test_enable_padding_free_metadata_does_not_mutate_examples():
+ collator = _PaddingFreeCollator()
+ trainer = SimpleNamespace(
+ data_collator = collator,
+ args = SimpleNamespace(remove_unused_columns = True),
+ )
+ enable_padding_free_metadata(_DummyModel(), trainer)
+
+ examples = [
+ {"input_ids": [1, 2, 3], "labels": [1, 2, 3]},
+ {"input_ids": [4, 5], "labels": [4, 5]},
+ ]
+ before = copy.deepcopy(examples)
+
+ batch = trainer.data_collator.torch_call(examples)
+
+ assert examples == before, "collator wrapper mutated the caller's examples"
+ assert torch.equal(batch["packed_seq_lengths"], torch.tensor([3, 2], dtype = torch.int32))
+
+ explicit = [{"input_ids": [1, 2, 3], "seq_lengths": [2, 1]}]
+ explicit_before = copy.deepcopy(explicit)
+ batch = trainer.data_collator.torch_call(explicit)
+ assert explicit == explicit_before
+ assert torch.equal(batch["packed_seq_lengths"], torch.tensor([2, 1], dtype = torch.int32))
+
+
+def test_enable_padding_free_metadata_still_hands_derived_lengths_to_the_collator():
+ collator = _PaddingFreeCollator()
+ trainer = SimpleNamespace(
+ data_collator = collator,
+ args = SimpleNamespace(remove_unused_columns = True),
+ )
+ enable_padding_free_metadata(_DummyModel(), trainer)
+
+ examples = [
+ {"input_ids": [1, 2, 3], "labels": [1, 2, 3]},
+ {"input_ids": [4, 5], "labels": [4, 5]},
+ ]
+ before = copy.deepcopy(examples)
+
+ trainer.data_collator.torch_call(examples)
+
+ assert examples == before, "collator wrapper mutated the caller's examples"
+ assert [row["seq_lengths"] for row in collator.seen] == [[3], [2]]
+ assert [row["labels"] for row in collator.seen] == [[1, 2, 3], [4, 5]]
+
+ # seq_lengths=None counts as missing: TRL would sum(None)
+ nulled = [{"input_ids": [1, 2], "seq_lengths": None}]
+ nulled_before = copy.deepcopy(nulled)
+
+ trainer.data_collator.torch_call(nulled)
+
+ assert nulled == nulled_before
+ assert [row["seq_lengths"] for row in collator.seen] == [[2]]
+
+ explicit = [{"input_ids": [1, 2, 3], "seq_lengths": [2, 1]}]
+
+ trainer.data_collator.torch_call(explicit)
+
+ assert collator.seen is explicit
+
+
def test_configure_sample_packing():
config = SimpleNamespace()
configure_sample_packing(config)
@@ -978,9 +1041,11 @@ class _PaddingFreeCollator:
self.padding_free = True
self.return_position_ids = False
self.calls = 0
+ self.seen = None
def torch_call(self, examples):
self.calls += 1
+ self.seen = examples
return {
"input_ids": torch.tensor([[0]], dtype = torch.long),
"examples_seen": self.calls,
diff --git a/unsloth/utils/packing.py b/unsloth/utils/packing.py
index f0a72b47969..419a863ae0c 100644
--- a/unsloth/utils/packing.py
+++ b/unsloth/utils/packing.py
@@ -255,18 +255,22 @@ def enable_padding_free_metadata(model, trainer):
def torch_call_with_padding_free_metadata(examples: Sequence[dict]):
seq_lengths: list[int] = []
+ collated = examples
if examples and isinstance(examples[0], dict):
- for example in examples:
+ for index, example in enumerate(examples):
lengths = example.get("seq_lengths")
if lengths is None:
ids = example.get("input_ids")
if ids is None:
continue
lengths = [len(ids)]
- example["seq_lengths"] = lengths
+ # TRL's collator keys seq_lengths off examples[0] and reads every row: pass a copy, not the caller's row.
+ if collated is examples:
+ collated = list(examples)
+ collated[index] = {**example, "seq_lengths": lengths}
seq_lengths.extend(lengths)
- batch = original_torch_call(examples)
+ batch = original_torch_call(collated)
if seq_lengths:
# Labels left alone for the same reason as enable_sample_packing: num_items_in_batch is counted off
# this batch and the zoo's discount of the boundary targets is idempotent.