Update app.py
Browse files
app.py
CHANGED
|
@@ -14,7 +14,6 @@ llm = Llama(
|
|
| 14 |
n_ctx=2048,
|
| 15 |
chat_format="llama-3",
|
| 16 |
n_gpu_layers=-1, # ensure all layers are on GPU
|
| 17 |
-
flash_attn=True,
|
| 18 |
)
|
| 19 |
|
| 20 |
# Placeholder responses for when context is empty
|
|
|
|
| 14 |
n_ctx=2048,
|
| 15 |
chat_format="llama-3",
|
| 16 |
n_gpu_layers=-1, # ensure all layers are on GPU
|
|
|
|
| 17 |
)
|
| 18 |
|
| 19 |
# Placeholder responses for when context is empty
|