vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : add EXAONE model support (#9025)

Browse source 
* add exaone model support

* add chat template

* fix whitespace

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* add ftype

* add exaone pre-tokenizer in `llama-vocab.cpp`

Co-Authored-By: compilade <113953597+compilade@users.noreply.github.com>

* fix lint

Co-Authored-By: compilade <113953597+compilade@users.noreply.github.com>

* add `EXAONE` to supported models in `README.md`

* fix space

Co-authored-by: compilade <git@compilade.net>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: compilade <113953597+compilade@users.noreply.github.com>
Co-authored-by: compilade <git@compilade.net>
This commit is contained in:
Minsoo Cheong 2024-08-16 15:35:18 +09:00 • committed by GitHub
parent fb487bb567
commit c679e0cb5c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
8 changed files with 320 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

74
convert_hf_to_gguf.py
View file

@ -596,6 +596,9 @@ class Model:
if chkhsh == "bc01ce58980e1db43859146dc51b1758b3b88729b217a74792e9f8d43e479d21":
# ref: https://huggingface.co/TurkuNLP/gpt3-finnish-small
res = "gpt3-finnish"
if chkhsh == "4e2b24cc4770243d65a2c9ec19770a72f08cffc161adbb73fcbb6b7dd45a0aae":
# ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct
res = "exaone"
if res is None:
logger.warning("\n")
@ -3781,6 +3784,77 @@ class NemotronModel(Model):
return [(self.map_tensor_name(name), data_torch)]
@Model.register("ExaoneForCausalLM")
class ExaoneModel(Model):
model_arch = gguf.MODEL_ARCH.EXAONE
def set_gguf_parameters(self):
hparams = self.hparams
assert(hparams["activation_function"] == "silu")
max_position_embeddings = hparams["max_position_embeddings"]
embed_dim = hparams["hidden_size"]
num_heads = hparams["num_attention_heads"]
num_kv_heads = hparams.get("num_key_value_heads", num_heads)
layer_norm_eps = hparams["layer_norm_epsilon"]
intermediate_size = hparams["intermediate_size"] if "intermediate_size" in hparams else 4 * embed_dim
num_layers = hparams["num_layers"]
# ignore for now as EXAONE-3.0-7.8B-Instruct attentino_dropout is 0.0
# attention_dropout_rate = hparams["attention_dropout"]
# ignore for now as EXAONE-3.0-7.8B-Instruct embed_dropout is 0.0
# embed_dropout_rate = hparams["embed_dropout"]
self.gguf_writer.add_embedding_length(embed_dim)
self.gguf_writer.add_head_count(num_heads)
self.gguf_writer.add_head_count_kv(num_kv_heads)
self.gguf_writer.add_context_length(max_position_embeddings)
self.gguf_writer.add_layer_norm_rms_eps(layer_norm_eps)
self.gguf_writer.add_feed_forward_length(intermediate_size)
self.gguf_writer.add_block_count(num_layers)
self.gguf_writer.add_file_type(self.ftype)
if (rope_theta := self.hparams.get("rope_theta")) is not None:
self.gguf_writer.add_rope_freq_base(rope_theta)
rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"], optional=True)
rotary_factor = rotary_factor if rotary_factor is not None else 1.0
self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
if hparams.get("rope_scaling") is not None and "factor" in hparams["rope_scaling"]:
if hparams["rope_scaling"].get("type") == "linear":
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
def prepare_tensors(self):
if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
if rope_scaling.get("rope_type", '').lower() == "llama3":
base = self.hparams.get("rope_theta", 10000.0)
dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
factor = rope_scaling.get("factor", 8.0)
low_freq_factor = rope_scaling.get("low_freq_factor", 1.0)
high_freq_factor = rope_scaling.get("high_freq_factor", 4.0)
old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
low_freq_wavelen = old_context_len / low_freq_factor
high_freq_wavelen = old_context_len / high_freq_factor
assert low_freq_wavelen != high_freq_wavelen
rope_factors = []
for freq in freqs:
wavelen = 2 * math.pi / freq
if wavelen < high_freq_wavelen:
rope_factors.append(1)
elif wavelen > low_freq_wavelen:
rope_factors.append(factor)
else:
smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
rope_factors.append(1 / ((1 - smooth) / factor + smooth))
self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
super().prepare_tensors()
###### CONVERSION LOGIC ######