Skip to content

LLM inference in C/C++

C++ ggml

convert_hf_to_gguf: support split MTP export for HY V3 (#25641)

Thiago Padilhacb489bc0fb78
diff --git a/conversion/base.py b/conversion/base.py
index 0421aa4bc4..1b85ef0a1b 100644
--- a/conversion/base.py
+++ b/conversion/base.py
@@ -109,7 +109,9 @@ class ModelBase:
sentence_transformers_dense_modules: bool = False
# MTP (multi-token prediction) export modes; set by main() before instantiation.
- # Architectures opt in by overriding the handling (see _Qwen35MtpMixin).
+ # Architectures that implement the filtering/export behavior opt in by
+ # setting supports_mtp_export = True on their model class or a mixin.
+ supports_mtp_export: bool = False
mtp_only: bool = False
no_mtp: bool = False
diff --git a/conversion/hunyuan.py b/conversion/hunyuan.py
index 4d2545f8b4..65d294fbe9 100644
--- a/conversion/hunyuan.py
+++ b/conversion/hunyuan.py
@@ -361,6 +361,7 @@ class HunyuanVLTextModel(HunYuanModel):
@ModelBase.register("HYV3ForCausalLM")
class HYV3Model(TextModel):
model_arch = gguf.MODEL_ARCH.HY_V3
+ supports_mtp_export = True
# Trunk layer count, stashed before indexing so the classmethod
# filter_tensors can identify the appended MTP block(s) (mirrors
diff --git a/conversion/qwen.py b/conversion/qwen.py
index 0356bd2da7..82d42fcc16 100644
--- a/conversion/qwen.py
+++ b/conversion/qwen.py
@@ -541,6 +541,7 @@ class _Qwen35MtpMixin:
`mtp.*` to the standard layer-indexed nextn naming so the existing
tensor_map handles them."""
+ supports_mtp_export = True
hparams: dict[str, Any]
model_arch: gguf.MODEL_ARCH
gguf_writer: gguf.GGUFWriter
diff --git a/conversion/step3.py b/conversion/step3.py
index 49bb5244a6..f7cdc997e5 100644
--- a/conversion/step3.py
+++ b/conversion/step3.py
@@ -98,6 +98,7 @@ class Step3VLTextModel(Qwen3Model):
@ModelBase.register("Step3p5ForCausalLM", "Step3p7ForConditionalGeneration")
class Step35Model(TextModel):
model_arch = gguf.MODEL_ARCH.STEP35
+ supports_mtp_export = True
# The --mtp / --no-mtp toggles are ModelBase.mtp_only / no_mtp (set in
# convert_hf_to_gguf.py main()). Unlike Qwen3.5, which stores MTP under a
diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 3b23d5ebc0..2c5e62a16f 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -259,10 +259,8 @@ def main() -> None:
sys.exit(1)
if args.mtp or args.no_mtp:
- from conversion.qwen import _Qwen35MtpMixin
- from conversion.step3 import Step35Model
- if not (issubclass(model_class, _Qwen35MtpMixin) or issubclass(model_class, Step35Model)):
- logger.error("--mtp / --no-mtp are only supported for Qwen3.5/3.6 and Step3.5 text variants today")
+ if not model_class.supports_mtp_export:
+ logger.error("--mtp / --no-mtp are not supported for %s", model_architecture)
sys.exit(1)
if args.no_mtp:
model_class.no_mtp = True