CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

model : support LiquidAI LFM2 hybrid family (#14620)

Browse Source 
**Important**
LFM2 was [merged ](https://github.com/huggingface/transformers/pull/39340)into transformers, but has not yet been released.
To convert into gguf, install transformers from source
```shell
pip install "transformers @ git+https://github.com/huggingface/transformers.git@main"
```
This commit is contained in:
Tarek Dakhran
2025-07-11 20:27:01 +02:00
committed by GitHub
parent 756aa1020a
commit f5e96b368f
14 changed files with 373 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
convert_hf_to_gguf.py
View File

@@ -300,6 +300,7 @@ class ModelBase:
gguf.MODEL_TENSOR.POS_EMBD,
gguf.MODEL_TENSOR.TOKEN_TYPES,
gguf.MODEL_TENSOR.SSM_CONV1D,
gguf.MODEL_TENSOR.SHORTCONV_CONV,
gguf.MODEL_TENSOR.TIME_MIX_FIRST,
gguf.MODEL_TENSOR.TIME_MIX_W1,
gguf.MODEL_TENSOR.TIME_MIX_W2,
@@ -836,6 +837,9 @@ class TextModel(ModelBase):
if chkhsh == "f6791d196f87ce6b56a7d234be618e0d58f8cda3549416635b2bebcd22cd95c4":
# ref: https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct
res = "midm-2.0"
if chkhsh == "169bf0296a13c4d9b7672313f749eb36501d931022de052aad6e36f2bf34dd51":
# ref: https://huggingface.co/LiquidAI/LFM2-Tokenizer
res = "lfm2"
if res is None:
logger.warning("\n")
@@ -7073,6 +7077,50 @@ class SmolLM3Model(LlamaModel):
chat_template = tokenizer.chat_template.replace("[:]", "")
self.gguf_writer.add_chat_template(chat_template)
@ModelBase.register("Lfm2ForCausalLM")
@ModelBase.register("LFM2ForCausalLM")
class LFM2Model(TextModel):
model_arch = gguf.MODEL_ARCH.LFM2
def _add_feed_forward_length(self):
ff_dim = self.hparams["block_ff_dim"]
auto_adjust_ff_dim = self.hparams["block_auto_adjust_ff_dim"]
ff_dim = self.hparams["block_ff_dim"]
ffn_dim_multiplier = self.hparams["block_ffn_dim_multiplier"]
multiple_of = self.hparams["block_multiple_of"]
if auto_adjust_ff_dim:
ff_dim = int(2 * ff_dim / 3)
# custom dim factor multiplier
if ffn_dim_multiplier is not None:
ff_dim = int(ffn_dim_multiplier * ff_dim)
ff_dim = multiple_of * ((ff_dim + multiple_of - 1) // multiple_of)
self.gguf_writer.add_feed_forward_length(ff_dim)
def set_gguf_parameters(self):
# set num_key_value_heads only for attention layers
self.hparams["num_key_value_heads"] = [
self.hparams["num_key_value_heads"] if layer_type == "full_attention" else 0
for layer_type in self.hparams["layer_types"]
]
super().set_gguf_parameters()
self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
self.gguf_writer.add_shortconv_l_cache(self.hparams["conv_L_cache"])
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["norm_eps"])
self._add_feed_forward_length()
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# conv op requires 2d tensor
if 'conv.conv' in name:
data_torch = data_torch.squeeze(1)
return [(self.map_tensor_name(name), data_torch)]
###### CONVERSION LOGIC ######