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.

Keep Dia's audio pad token in fast generate (#13128)

Daniel Hancd79442dfc70
diff --git a/tests/test_dia_generate_pad_token.py b/tests/test_dia_generate_pad_token.py
new file mode 100644
index 00000000000..b0808f50658
--- /dev/null
+++ b/tests/test_dia_generate_pad_token.py
@@ -0,0 +1,103 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+"""Dia generation keeps its own audio pad token instead of EOS (#2560)."""
+
+import ast
+import inspect
+import os
+from contextlib import nullcontext
+from importlib.metadata import version as installed_version
+from pathlib import Path
+from types import SimpleNamespace
+
+from packaging.version import Version
+
+
+VISION_PATH = Path(__file__).parents[1] / "unsloth" / "models" / "vision.py"
+
+
+def _load_function(name, namespace):
+ tree = ast.parse(VISION_PATH.read_text(encoding = "utf-8"))
+ function = next(
+ node for node in tree.body if isinstance(node, ast.FunctionDef) and node.name == name
+ )
+ exec(compile(ast.Module(body = [function], type_ignores = []), str(VISION_PATH), "exec"), namespace)
+ return namespace[name]
+
+
+class _FakeTensor:
+ shape = (1, 3)
+
+ def to(self, dtype):
+ return self
+
+
+def _captured_generate_kwargs(config, **generate_kwargs):
+ # A FlashAttention text config takes the delegate branch, which hands kwargs straight to _old_generate.
+ config.architectures = ["DiaForConditionalGeneration"]
+ config._attn_implementation = "flash_attention_2"
+ namespace = {
+ "torch": SimpleNamespace(
+ Tensor = _FakeTensor,
+ bfloat16 = "bfloat16",
+ float16 = "float16",
+ inference_mode = nullcontext,
+ autocast = lambda **kwargs: nullcontext(),
+ ),
+ "os": os,
+ "inspect": inspect,
+ "FastBaseModel": SimpleNamespace(for_inference = lambda model: None),
+ "dtype_from_config": lambda config: "bfloat16",
+ "_get_dtype": lambda dtype: dtype,
+ "_unsloth_generate_accepts_kwarg": lambda model, name: False,
+ "NUM_LOGITS_TO_KEEP": {"DiaForConditionalGeneration": None},
+ "DEVICE_TYPE_TORCH": "cuda",
+ "Version": Version,
+ "transformers_version": installed_version("transformers"),
+ "_uses_flash_attention_for_generation": lambda config: True,
+ "_clear_generation_caches": lambda model: None,
+ "_is_text_seq2seq_config": lambda config: False,
+ }
+ fast_generate = _load_function("unsloth_base_fast_generate", namespace)
+ captured = {}
+
+ class Model:
+ def forward(self, input_ids = None):
+ return input_ids
+
+ def _old_generate(self, *args, **kwargs):
+ captured.update(kwargs)
+
+ Model.config = config
+ fast_generate(Model(), input_ids = _FakeTensor(), **generate_kwargs)
+ return captured
+
+
+# Not DiaConfig: Dia needs transformers >= 4.53, the floor is 4.52.4. Ids sit at the top level before 5, on decoder_config from 5.
+def _legacy_dia_config():
+ return SimpleNamespace(
+ model_type = "dia", pad_token_id = 1025, eos_token_id = 1024, decoder_config = SimpleNamespace()
+ )
+
+
+def _dia_config():
+ return SimpleNamespace(
+ model_type = "dia",
+ pad_token_id = None,
+ eos_token_id = None,
+ decoder_config = SimpleNamespace(pad_token_id = 1025, eos_token_id = 1024),
+ )
+
+
+def test_dia_defaults_to_its_audio_pad_token():
+ assert _captured_generate_kwargs(_legacy_dia_config())["pad_token_id"] == 1025
+ assert _captured_generate_kwargs(_dia_config())["pad_token_id"] == 1025
+
+
+def test_dia_explicit_pad_token_id_wins():
+ assert _captured_generate_kwargs(_dia_config(), pad_token_id = 7)["pad_token_id"] == 7
+
+
+def test_other_models_still_default_to_eos():
+ config = SimpleNamespace(model_type = "csm", eos_token_id = [2, 3], pad_token_id = 0)
+ assert _captured_generate_kwargs(config)["pad_token_id"] == 2
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 64be4931d58..069fd9c3b2b 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -1625,11 +1625,17 @@ def unsloth_base_fast_generate(self, *args, **kwargs):
# Encoder-decoders keep their own pad (T5: pad 0, EOS 1), used to infer the encoder mask and pad finished rows.
default_pad_token_id = model_eos_token_id
- if (
- _is_text_seq2seq_config(self.config)
- and getattr(self.config, "pad_token_id", None) is not None
- ):
- default_pad_token_id = self.config.pad_token_id
+ config_pad_token_id = getattr(self.config, "pad_token_id", None)
+ is_dia = getattr(self.config, "model_type", None) == "dia"
+ if is_dia:
+ # Dia's audio pad fills its delay pattern; EOS there breaks the DAC decode (#2560). transformers 5 moved it to decoder_config.
+ decoder_pad_token_id = getattr(
+ getattr(self.config, "decoder_config", None), "pad_token_id", None
+ )
+ if decoder_pad_token_id is not None:
+ config_pad_token_id = decoder_pad_token_id
+ if (is_dia or _is_text_seq2seq_config(self.config)) and config_pad_token_id is not None:
+ default_pad_token_id = config_pad_token_id
kwargs["pad_token_id"] = kwargs.pop("pad_token_id", default_pad_token_id)
try: