From 460c40d8abb07bda636fc0d26af025e790bef3bf Mon Sep 17 00:00:00 2001
From: Gennadij <berkut1@users.noreply.github.com>
Date: Tue, 12 Sep 2023 23:02:42 +0300
Subject: [PATCH] Read more GGUF metadata (scale_linear and freq_base) (#3877)

---
 models/config.yaml         | 2 +-
 modules/metadata_gguf.py   | 3 +++
 modules/models_settings.py | 5 +++++
 3 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/models/config.yaml b/models/config.yaml
index 8783c49b..d98e95fb 100644
--- a/models/config.yaml
+++ b/models/config.yaml
@@ -226,4 +226,4 @@ llama-65b-gptq-3bit:
 .*codellama:
   rope_freq_base: 1000000
 .*codellama.*instruct:
-  instruction_template: 'Llama-v2'
+  instruction_template: 'Llama-v2'
\ No newline at end of file
diff --git a/modules/metadata_gguf.py b/modules/metadata_gguf.py
index 42f3a190..f633d70c 100644
--- a/modules/metadata_gguf.py
+++ b/modules/metadata_gguf.py
@@ -70,6 +70,9 @@ def load_metadata(fname):
         GGUF_VERSION = struct.unpack("<I", file.read(4))[0]
         ti_data_count = struct.unpack("<Q", file.read(8))[0]
         kv_data_count = struct.unpack("<Q", file.read(8))[0]
+        
+        if GGUF_VERSION == 1: 
+            raise Exception('You are using an outdated GGUF, please download a new one.')
 
         for i in range(kv_data_count):
             key_length = struct.unpack("<Q", file.read(8))[0]
diff --git a/modules/models_settings.py b/modules/models_settings.py
index e30772ec..e3f86c31 100644
--- a/modules/models_settings.py
+++ b/modules/models_settings.py
@@ -17,6 +17,7 @@ def get_fallback_settings():
         'truncation_length': shared.settings['truncation_length'],
         'n_ctx': 2048,
         'rope_freq_base': 0,
+        'compress_pos_emb': 1,
     }
 
 
@@ -48,6 +49,10 @@ def get_model_metadata(model):
         metadata = metadata_gguf.load_metadata(model_file)
         if 'llama.context_length' in metadata:
             model_settings['n_ctx'] = metadata['llama.context_length']
+        if 'llama.rope.scale_linear' in metadata:
+            model_settings['compress_pos_emb'] = metadata['llama.rope.scale_linear']
+        if 'llama.rope.freq_base' in metadata:
+            model_settings['rope_freq_base'] = metadata['llama.rope.freq_base']    
 
     return model_settings