nits

conversion/qwen4exp.py

@@ -15,7 +15,7 @@ from .qwen3vl import Qwen3VLVisionModel
 
 
 @ModelBase.register("Qwen4ExpForConditionalGeneration", "Qwen4ExpForCausalLM")
-@ModelBase.example("unsloth/Qwen3.8-Flash-Next")
+@ModelBase.example("Qwen/Qwen3.8-Flash-Next")
 class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
     """Qwen3.8-Flash-Next.
 
@@ -98,7 +98,6 @@ class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
             self._read_hash_constants("ple_embedding.ngram_heads_vocab_sizes"))
 
     def _image_token_id(self) -> int | None:
-        # base.py merges text_config into the root of hparams, where image_token_id already is
         img = self.hparams.get("image_token_id")
         return None if img is None else int(img)
 
@@ -242,6 +241,6 @@ class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
 
 
 @ModelBase.register("Qwen4ExpForConditionalGeneration")
-@ModelBase.example("unsloth/Qwen3.8-Flash-Next")
+@ModelBase.example("Qwen/Qwen3.8-Flash-Next")
 class Qwen4ExpVisionModel(Qwen3VLVisionModel):
     """The vision tower is an unmodified Qwen3-VL ViT."""

src/models/qwen4exp.cpp

@@ -115,7 +115,7 @@ void llama_model_qwen4exp::load_arch_tensors(llama_model_loader & ml) {
         GGML_ASSERT(ple_w != nullptr && "qwen4exp is missing the PLE n-gram table");
         const int64_t ple_rows = ple_w->tensor->ne[1];
         per_layer_tok_embd = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"),
-                                           { hparams.ple_head_dim, ple_rows }, 0);
+                                           { hparams.ple_head_dim, ple_rows }, TENSOR_READ_LAZY);
     }
 
     for (int il = 0; il < n_layer; ++il) {