happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

convert: minor fix

This commit is contained in:
bluebread 2025-12-17 03:03:17 +00:00
parent 87e4a00c4c
commit f629d02ee1
1 changed files with 7 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
convert_hf_to_gguf.py
View File

@ -7189,9 +7189,10 @@ class DeepseekV2Model(TextModel):
def __init__(self, *args, **kwargs): def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs) super().__init__(*args, **kwargs)
vision_config = self.hparams.get('vision_config', {}).get('width', {}) hparams: dict = ModelBase.load_hparams(self.dir_model, is_mistral_format=False)
self.origin_hf_arch = hparams.get('architectures', [None])[0]
if 'clip-l-14-224' in vision_config and 'sam_vit_b' in vision_config: if self.origin_hf_arch == "DeepseekOCRForCausalLM":
self.model_arch = gguf.MODEL_ARCH.DEEPSEEK2OCR self.model_arch = gguf.MODEL_ARCH.DEEPSEEK2OCR
self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch] self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
self.gguf_writer.add_architecture() self.gguf_writer.add_architecture()
@ -7260,6 +7261,9 @@ class DeepseekV2Model(TextModel):
# note: deepseek2 using MLA converts into MQA (ie: GQA with 1 group) # note: deepseek2 using MLA converts into MQA (ie: GQA with 1 group)
self.hparams["num_key_value_heads"] = 1 self.hparams["num_key_value_heads"] = 1
if (rope_mscale_all := self.rope_parameters.get("mscale_all_dim")) is not None:
self.hparams["num_key_value_heads"] = self.hparams.get("rms_norm_eps", 1e-6)
super().set_gguf_parameters() super().set_gguf_parameters()
hparams = self.hparams hparams = self.hparams
kv_lora_rank = hparams["kv_lora_rank"] if hparams.get("kv_lora_rank") is not None else 512 kv_lora_rank = hparams["kv_lora_rank"] if hparams.get("kv_lora_rank") is not None else 512
@ -7290,10 +7294,9 @@ class DeepseekV2Model(TextModel):
if (rope_mscale_all := self.rope_parameters.get("mscale_all_dim")) is not None: if (rope_mscale_all := self.rope_parameters.get("mscale_all_dim")) is not None:
# [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX] # [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
# note: for legacy reasons, this is not consistent with the other usages of self.gguf_writer.add_rope_scaling_yarn_log_mul
# ref https://github.com/ggml-org/llama.cpp/pull/17945 # ref https://github.com/ggml-org/llama.cpp/pull/17945
self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * rope_mscale_all) self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * rope_mscale_all)
self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-6))
_experts: list[dict[str, Tensor]] | None = None _experts: list[dict[str, Tensor]] | None = None