Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

convert.py : Get rope scale from HuggingFace models #2772

Merged
merged 4 commits into from
Aug 25, 2023
Merged

convert.py : Get rope scale from HuggingFace models #2772

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
8d0d83b
Get rope scale from HF models
pnb Aug 24, 2023
8ac33ce
Save rope scale only for linear scaling
pnb Aug 24, 2023
aa896e7
Rewrite for clarity
pnb Aug 24, 2023
4950b2d
Merge branch 'master' into convert_rope_scale
pnb Aug 25, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 10 additions & 0 deletions convert.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -105,6 +105,7 @@ class Params:
f_norm_eps: float

f_rope_freq_base: Optional[float] = None
f_rope_scale: Optional[float] = None

ftype: Optional[GGMLFileType] = None

Expand Down Expand Up @@ -169,6 +170,11 @@ def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
f_norm_eps = config["rms_norm_eps"]
f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None

if "rope_scaling" in config and config["rope_scaling"].get("type") == "linear":
f_rope_scale = config["rope_scaling"].get("factor")
else:
f_rope_scale = None

n_mult = Params.find_n_mult(n_ff, n_embd)

if "max_sequence_length" in config:
Expand All @@ -190,6 +196,7 @@ def loadHFTransformerJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
n_head_kv = n_head_kv,
f_norm_eps = f_norm_eps,
f_rope_freq_base = f_rope_freq_base,
f_rope_scale = f_rope_scale,
)

# LLaMA v2 70B params.json
Expand Down Expand Up @@ -773,6 +780,9 @@ def add_meta_arch(self, params: Params) -> None:
if params.f_rope_freq_base:
self.gguf.add_rope_freq_base(params.f_rope_freq_base)

if params.f_rope_scale:
self.gguf.add_rope_scale_linear(params.f_rope_scale)

if params.ftype:
self.gguf.add_file_type(params.ftype)

Expand Down