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.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
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.pyindex 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 copyimport inspectimport loggingfrom 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 = Trueself.return_position_ids = Falseself.calls = 0+ self.seen = Nonedef torch_call(self, examples):self.calls += 1+ self.seen = examplesreturn {"input_ids": torch.tensor([[0]], dtype = torch.long),"examples_seen": self.calls,diff --git a/unsloth/utils/packing.py b/unsloth/utils/packing.pyindex 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 = examplesif 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:continuelengths = [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.