conversion : fix Qwen3.5 MTP config lookup from text_config (fixes unslothai/unsloth#8443)

conversion/qwen.py

@@ -289,12 +289,20 @@ class _QwenMtpMixin:
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.block_count = self.hparams["num_hidden_layers"]
+        # text_config may nest the text hyperparameters; merge them like index_tensors does
+        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
+        self.block_count = hparams["num_hidden_layers"]
         if not self.no_mtp:
-            n_mtp = self.hparams.get("mtp_num_hidden_layers", 0)
-            # Qwen-3-Next doesn't include `mtp_num_hidden_layers` in config.
+            n_mtp = hparams.get("mtp_num_hidden_layers", 0)
+            # Qwen-3-Next doesn't include mtp_num_hidden_layers in config.
             if n_mtp == 0:
-                assert self.opt_num_mtp_layers != 0
+                # The count is recovered from tensor names later; fail clearly instead of asserting.
+                if self.opt_num_mtp_layers == 0:
+                    raise ValueError(
+                        "MTP layer count not found in config (checked top level and "
+                        "text_config) and not yet recovered from tensor names. "
+                        "Re-export with --no-mtp, or add mtp_num_hidden_layers to config."
+                    )
                 n_mtp = self.opt_num_mtp_layers
             self.block_count += n_mtp
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)

conversion/tests/test_qwen_mtp.py

@@ -0,0 +1,72 @@
+from __future__ import annotations
+
+import pytest
+
+import gguf
+from conversion.qwen import _QwenMtpMixin
+
+
+class _DummyParent:
+    def __init__(self, *, hparams: dict, no_mtp: bool = False, model_arch: gguf.MODEL_ARCH, **kwargs):
+        self.hparams = hparams
+        self.no_mtp = no_mtp
+        self.model_arch = model_arch
+
+
+class _MtpModel(_QwenMtpMixin, _DummyParent):
+    pass
+
+
+@pytest.fixture(autouse=True)
+def _reset_mtp_class_state():
+    """The mixin uses class attributes that persist across conversions; isolate tests."""
+    _MtpModel._original_block_count = None
+    _MtpModel.opt_num_mtp_layers = 0
+    yield
+    _MtpModel._original_block_count = None
+    _MtpModel.opt_num_mtp_layers = 0
+
+
+def test_mtp_count_read_from_text_config():
+    model = _MtpModel(
+        hparams={
+            "num_hidden_layers": 24,
+            "text_config": {"num_hidden_layers": 24, "mtp_num_hidden_layers": 1},
+        },
+        model_arch=gguf.MODEL_ARCH.QWEN3NEXT,
+    )
+    assert model.block_count == 25
+
+
+def test_mtp_count_read_from_top_level_config():
+    model = _MtpModel(
+        hparams={"num_hidden_layers": 24, "mtp_num_hidden_layers": 2},
+        model_arch=gguf.MODEL_ARCH.QWEN3NEXT,
+    )
+    assert model.block_count == 26
+
+
+def test_mtp_count_falls_back_to_recovered_class_attribute():
+    _MtpModel.opt_num_mtp_layers = 1
+    model = _MtpModel(
+        hparams={"num_hidden_layers": 24},
+        model_arch=gguf.MODEL_ARCH.QWEN3NEXT,
+    )
+    assert model.block_count == 25
+
+
+def test_missing_mtp_count_raises_actionable_error():
+    with pytest.raises(ValueError, match="MTP layer count not found"):
+        _MtpModel(
+            hparams={"num_hidden_layers": 24},
+            model_arch=gguf.MODEL_ARCH.QWEN3NEXT,
+        )
+
+
+def test_no_mtp_skips_mtp_count_lookup():
+    model = _MtpModel(
+        hparams={"num_hidden_layers": 24},
+        no_mtp=True,
+        model_arch=gguf.MODEL_ARCH.QWEN3NEXT,
+    )
+    assert model.block_count == 24