Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

model: correctly support input vision for deepseek4 (#28154)

Browse Source 
* model: correctly support input vision for deepseek4

* nits
This commit is contained in:
Xuan-Son Nguyen
2026-09-02 19:14:46 +02:00
committed by GitHub
parent d5fec32a87
commit 9400c8946e
9 changed files with 34 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
conversion/deepseek.py
+6 -2
View File
@@ -578,8 +578,7 @@ class DeepseekV4Model(TextModel):
@classmethod
def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
name, gen = item
if (name.startswith(("aligner.", "image_"))
or name.endswith(".ffn.gate.bias_vl")):
if name.startswith(("aligner.", "image_")):
return None
if name.startswith("mtp."):
if not cls.mtp_only:
@@ -856,6 +855,7 @@ class DeepseekV4Model(TextModel):
"ffn_norm.weight": (gguf.MODEL_TENSOR.FFN_NORM, ".weight"),
"ffn.gate.weight": (gguf.MODEL_TENSOR.FFN_GATE_INP, ".weight"),
"ffn.gate.bias": (gguf.MODEL_TENSOR.FFN_EXP_PROBS_B, ".bias"),
"ffn.gate.bias_vl": (gguf.MODEL_TENSOR.FFN_EXP_PROBS_B_VL, ".bias"),
"ffn.gate.tid2eid": (gguf.MODEL_TENSOR.FFN_GATE_TID2EID, ".weight"),
"ffn.shared_experts.w1.weight": (gguf.MODEL_TENSOR.FFN_GATE_SHEXP, ".weight"),
"ffn.shared_experts.w2.weight": (gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, ".weight"),
@@ -881,6 +881,10 @@ class DeepseekV4Model(TextModel):
if re.match(r"layers\.\d+\.ffn\.experts\.\d+\.w[123]\.(weight|scale)$", name):
return []
# hash layers route text tokens via tid2eid and image tokens via bias_vl; gate.bias is unused
if name.endswith(".ffn.gate.bias") and bid is not None and bid < self.hparams["num_hash_layers"]:
return []
tensor_key, suffix = self._map_dsv4_tensor_name(name, bid)
if tensor_key == gguf.MODEL_TENSOR.FFN_GATE_TID2EID:
return []