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

llama : add support for llama 3.1 rope scaling factors (#8676)

Browse source 
* Add llama 3.1 rope scaling factors to llama conversion and inference

This commit generates the rope factors on conversion and adds them to the resulting model as a tensor. At inference time, these factors are passed to the `ggml_rope_ext` rope oepration, improving results for context windows above 8192

* Update convert_hf_to_gguf.py

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

* address comments

* address comments

* Update src/llama.cpp

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

* Update convert_hf_to_gguf.py

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

---------

Co-authored-by: compilade <git@compilade.net>
This commit is contained in:
Jeffrey Morgan 2024-07-27 05:03:45 -07:00 • committed by GitHub
parent 92090eca21
commit b5e95468b1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 40 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

28
convert_hf_to_gguf.py
View file

@ -1570,6 +1570,34 @@ class LlamaModel(Model):
return [(self.map_tensor_name(name), data_torch)]
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()
if self._experts is not None: