Upload 3 files
Browse files- hexstate_requantize.py +9 -2
hexstate_requantize.py
CHANGED
|
@@ -277,14 +277,14 @@ TYPE_NAME = {
|
|
| 277 |
13: "Q5_K", 14: "Q6_K", 15: "Q8_K", 30: "BF16",
|
| 278 |
}
|
| 279 |
|
| 280 |
-
# Block sizes and byte sizes for each type
|
| 281 |
TYPE_BLOCK_SIZE = {
|
| 282 |
0: 1, 1: 1, 2: 32, 3: 32, 6: 32, 7: 32,
|
| 283 |
8: 32, 9: 32, 10: 256, 11: 256, 12: 256,
|
| 284 |
13: 256, 14: 256, 15: 256, 30: 1,
|
| 285 |
}
|
| 286 |
TYPE_BLOCK_BYTES = {
|
| 287 |
-
0: 4, 1: 2, 2: 18, 3: 20, 6:
|
| 288 |
8: 34, 9: 36, 10: 84, 11: 110, 12: 144,
|
| 289 |
13: 176, 14: 210, 15: 292, 30: 2,
|
| 290 |
}
|
|
@@ -967,6 +967,13 @@ def main():
|
|
| 967 |
out_data_offset += out_size
|
| 968 |
out_data_offset = align_offset(out_data_offset)
|
| 969 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 970 |
# ββ Update KV pairs ββ
|
| 971 |
updated_kv = []
|
| 972 |
if keep_metadata:
|
|
|
|
| 277 |
13: "Q5_K", 14: "Q6_K", 15: "Q8_K", 30: "BF16",
|
| 278 |
}
|
| 279 |
|
| 280 |
+
# Block sizes and byte sizes for each type (from ggml.c)
|
| 281 |
TYPE_BLOCK_SIZE = {
|
| 282 |
0: 1, 1: 1, 2: 32, 3: 32, 6: 32, 7: 32,
|
| 283 |
8: 32, 9: 32, 10: 256, 11: 256, 12: 256,
|
| 284 |
13: 256, 14: 256, 15: 256, 30: 1,
|
| 285 |
}
|
| 286 |
TYPE_BLOCK_BYTES = {
|
| 287 |
+
0: 4, 1: 2, 2: 18, 3: 20, 6: 22, 7: 24,
|
| 288 |
8: 34, 9: 36, 10: 84, 11: 110, 12: 144,
|
| 289 |
13: 176, 14: 210, 15: 292, 30: 2,
|
| 290 |
}
|
|
|
|
| 967 |
out_data_offset += out_size
|
| 968 |
out_data_offset = align_offset(out_data_offset)
|
| 969 |
|
| 970 |
+
# ββ Detect Architecture ββ
|
| 971 |
+
arch = 'llama'
|
| 972 |
+
for key, vtype, val in kv_pairs:
|
| 973 |
+
if key == 'general.architecture' and vtype == 8:
|
| 974 |
+
arch = val.decode('utf-8', errors='ignore')
|
| 975 |
+
break
|
| 976 |
+
|
| 977 |
# ββ Update KV pairs ββ
|
| 978 |
updated_kv = []
|
| 979 |
if keep_metadata:
|