Set mem cache args on inference
Browse files- scripts/finetune.py +6 -0
scripts/finetune.py
CHANGED
|
@@ -77,6 +77,11 @@ def do_inference(cfg, model, tokenizer, prompter="AlpacaPrompter"):
|
|
| 77 |
importlib.import_module("axolotl.prompters"), prompter
|
| 78 |
)
|
| 79 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 80 |
while True:
|
| 81 |
print("=" * 80)
|
| 82 |
# support for multiline inputs
|
|
@@ -90,6 +95,7 @@ def do_inference(cfg, model, tokenizer, prompter="AlpacaPrompter"):
|
|
| 90 |
else:
|
| 91 |
prompt = instruction.strip()
|
| 92 |
batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
|
|
|
|
| 93 |
print("=" * 40)
|
| 94 |
model.eval()
|
| 95 |
with torch.no_grad():
|
|
|
|
| 77 |
importlib.import_module("axolotl.prompters"), prompter
|
| 78 |
)
|
| 79 |
|
| 80 |
+
if cfg.landmark_attention:
|
| 81 |
+
model.set_mem_cache_args(
|
| 82 |
+
max_seq_len=255, mem_freq=50, top_k=5, max_cache_size=None
|
| 83 |
+
)
|
| 84 |
+
|
| 85 |
while True:
|
| 86 |
print("=" * 80)
|
| 87 |
# support for multiline inputs
|
|
|
|
| 95 |
else:
|
| 96 |
prompt = instruction.strip()
|
| 97 |
batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
|
| 98 |
+
|
| 99 |
print("=" * 40)
|
| 100 |
model.eval()
|
| 101 |
with torch.no_grad():
|