CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-09 10:17:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

optimize deepstack structure

Browse Source
This commit is contained in:
JJJYmmm
2025-10-28 21:58:46 +08:00
parent 0443a098f3
commit 3271877207
5 changed files with 73 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
convert_hf_to_gguf.py
View File

@@ -4046,7 +4046,9 @@ class Qwen3VLVisionModel(MmprojModel):
self.hparams_vision["num_attention_heads"] = self.hparams_vision.get("num_heads")
self.hparams_vision["num_hidden_layers"] = self.hparams_vision.get("depth")
self.deepstack_layers: list[int] = list(self.hparams_vision.get("deepstack_visual_indexes", []))
self.is_deepstack_layers = [False] * int(self.hparams_vision["num_hidden_layers"] or 0)
for idx in self.hparams_vision.get("deepstack_visual_indexes", []):
self.is_deepstack_layers[idx] = True
def set_gguf_parameters(self):
super().set_gguf_parameters()
@@ -4062,10 +4064,11 @@ class Qwen3VLVisionModel(MmprojModel):
rms_norm_eps = self.global_config.get("text_config", {}).get("rms_norm_eps", 1e-6)
self.gguf_writer.add_vision_attention_layernorm_eps(rms_norm_eps)
if self.deepstack_layers:
self.gguf_writer.add_vision_deepstack_layers(self.deepstack_layers)
if self.is_deepstack_layers:
self.gguf_writer.add_vision_is_deepstack_layers(self.is_deepstack_layers)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
assert self.hparams_vision is not None
# Skip text model tensors - they go in the text model file
if name.startswith("model.language_model.") or name.startswith("lm_head."):
return []
@@ -4075,7 +4078,8 @@ class Qwen3VLVisionModel(MmprojModel):
if name.startswith("visual.deepstack_merger_list."):
prefix, rest = name.split(".", maxsplit=3)[2:]
idx = int(prefix)
# prefix is the layer index, convert to absolute clip layer index!
idx = self.hparams_vision.get("deepstack_visual_indexes", [])[int(prefix)]
target = rest
tensor_type: gguf.MODEL_TENSOR