From bb8bf2c42e37189ec94378f5c837f4080a46c1ad Mon Sep 17 00:00:00 2001
From: Xuan Son Nguyen <son@huggingface.co>
Date: Fri, 2 Oct 2026 19:08:26 +0200
Subject: [PATCH] nits

---
 common/common.cpp      |  2 +-
 conversion/__init__.py |  1 +
 conversion/clef.py     | 11 +++++++++--
 src/llama-batch.cpp    |  2 +-
 src/llama-ext.h        |  3 ++-
 5 files changed, 14 insertions(+), 5 deletions(-)

diff --git a/common/common.cpp b/common/common.cpp
index 0eb4878468..0269bc895c 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -2215,7 +2215,7 @@ llama_batch_ext * common_batch::get_sub_batch(int32_t off, int32_t n) {
             llama_batch_ext_set_output_logits(res, idx, true);
         }
         if (t.decision_order != 0) {
-            llama_batch_ext_set_decision_order(res, idx, t.decision_order);
+            llama_batch_ext_set_decision_order(res, idx, (llama_decision_order) t.decision_order);
         }
     }
 
diff --git a/conversion/__init__.py b/conversion/__init__.py
index 612bd5f189..6d8ae9c1a5 100644
--- a/conversion/__init__.py
+++ b/conversion/__init__.py
@@ -297,6 +297,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
 
 MMPROJ_MODEL_MAP: dict[str, str] = {
     "AudioFlamingo3ForConditionalGeneration": "ultravox",
+    "ClefModel": "clef",
     "CogVLMForCausalLM": "cogvlm",
     "DeepseekOCR2ForCausalLM": "deepseek",
     "DeepseekOCRForCausalLM": "deepseek",
diff --git a/conversion/clef.py b/conversion/clef.py
index 3cc57a4e35..a6f7243945 100644
--- a/conversion/clef.py
+++ b/conversion/clef.py
@@ -11,7 +11,7 @@ import torch
 if TYPE_CHECKING:
     from torch import Tensor
 
-from .base import ModelBase, gguf, logger
+from .base import MmprojModel, ModelBase, gguf, logger
 from .qwen import Qwen3_5TextModel
 
 
@@ -29,7 +29,6 @@ def _load_clef_hparams(dir_model: Path) -> dict[str, Any]:
     return hparams
 
 
-# TODO: image input needs token and embedding entries in the same batch, see https://github.com/ggml-org/llama.cpp/pull/29622
 @ModelBase.register("ClefModel")
 class ClefModel(Qwen3_5TextModel):
     model_arch = gguf.MODEL_ARCH.CLEF
@@ -141,3 +140,11 @@ class ClefModel(Qwen3_5TextModel):
                 return
 
         yield self.map_tensor_name(name), data_torch
+
+
+@ModelBase.register("ClefModel")
+class ClefVisionModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        del args, kwargs
+        raise NotImplementedError(
+            "multimodal input is not supported yet for Clef, requires https://github.com/ggml-org/llama.cpp/pull/29622 to be merged first")
diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp
index 6e58af8727..1b3d70627a 100644
--- a/src/llama-batch.cpp
+++ b/src/llama-batch.cpp
@@ -1268,7 +1268,7 @@ bool llama_batch_ext_set_output_logits(llama_batch_ext * batch, int32_t idx, boo
     return batch->set_output(idx, value);
 }
 
-bool llama_batch_ext_set_decision_order(llama_batch_ext * batch, int32_t idx, int32_t order) {
+bool llama_batch_ext_set_decision_order(llama_batch_ext * batch, int32_t idx, llama_decision_order order) {
     return batch->set_decision_order(idx, order);
 }
 
diff --git a/src/llama-ext.h b/src/llama-ext.h
index 3db728645d..fcc78f47e4 100644
--- a/src/llama-ext.h
+++ b/src/llama-ext.h
@@ -101,6 +101,7 @@ LLAMA_API void llama_set_embeddings_nextn(struct llama_context * ctx, bool value
 LLAMA_API void llama_set_nextn_layer_offset(struct llama_context * ctx, int32_t offset);
 
 // Marks the entries that a joint decision head (clef) reads, the default is 0
+// See https://github.com/ggml-org/llama.cpp/pull/29831 for details
 // A run of entries with the same value is one span, spans must be separated by entries with value 0
 // An option belongs to the last question before it
 enum llama_decision_order {
@@ -111,7 +112,7 @@ enum llama_decision_order {
     LLAMA_DECISION_ORDER_OPTION          = 4, // text of an option
 };
 // The embeddings output has one value per entry: row i is the score of option i
-LLAMA_API bool llama_batch_ext_set_decision_order(struct llama_batch_ext * batch, int32_t idx, int32_t order);
+LLAMA_API bool llama_batch_ext_set_decision_order(struct llama_batch_ext * batch, int32_t idx, enum llama_decision_order order);
 
 // mirrors:
 // LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);