From 3356d8ea2b3ef5307ff73a31a54ebaf7327ddc8e Mon Sep 17 00:00:00 2001 From: Tarek Dakhran Date: Wed, 23 Sep 2026 21:48:34 +0200 Subject: [PATCH] convert : allow vision target for DFlash/Dspark Resolve the target arch with get_model_architecture so vision targets (e.g. Lfm2VlForConditionalGeneration) map to their text model for the vocab. Fix double rope reorder for LFM2/LFM2.5 DSpark drafters --- conversion/qwen.py | 17 ++--------------- 1 file changed, 2 insertions(+), 15 deletions(-) diff --git a/conversion/qwen.py b/conversion/qwen.py index ef5504f3de5b..64d606176eb2 100644 --- a/conversion/qwen.py +++ b/conversion/qwen.py @@ -10,7 +10,7 @@ if TYPE_CHECKING: from torch import Tensor -from .base import LazyTorchTensor, ModelBase, TextModel, gguf, logger +from .base import LazyTorchTensor, ModelBase, ModelType, TextModel, get_model_architecture, gguf, logger @ModelBase.register("QWenLMHeadModel") @@ -666,7 +666,7 @@ def set_vocab(self): from . import get_model_class with open(self.target_model_dir / "config.json", "r", encoding="utf-8") as f: target_hparams = json.load(f) - target_arch = target_hparams["architectures"][0] + target_arch = get_model_architecture(target_hparams, ModelType.TEXT) target_cls = get_model_class(target_arch) if target_cls is not type(self): @@ -841,13 +841,6 @@ def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Ca return None return super().filter_tensors(item) - _ROPE_PERMUTE_SUFFIXES = ( - "self_attn.q_proj.weight", - "self_attn.k_proj.weight", - "self_attn.q_norm.weight", - "self_attn.k_norm.weight", - ) - def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: if name == "model.d2t": self._d2t = data_torch @@ -856,12 +849,6 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter if self._n_vocab_draft == self.hparams["vocab_size"] and name.endswith("lm_head.weight"): return - # interleaved-rope checkpoints (rope_is_neox_style = false) -> NeoX layout: per head, even dims first then odd - if not self.hparams.get("rope_is_neox_style", True) and name.endswith(self._ROPE_PERMUTE_SUFFIXES): - head_dim = self.hparams["head_dim"] - shape = data_torch.shape - data_torch = data_torch.reshape(-1, head_dim // 2, 2, *shape[1:]).transpose(1, 2).reshape(shape) - yield from super().modify_tensors(data_torch, name, bid) def prepare_tensors(self):