Spaces:
Running
Running
Commit ·
944fb90
1
Parent(s): 5951bbe
fix: pass HF_API_KEY as HF_TOKEN to avoid Hub rate limits when loading models
Browse files- config/settings.py +2 -1
config/settings.py
CHANGED
|
@@ -78,9 +78,10 @@ TOP_K_RERANK = 5 # Keep top 5 after reranking
|
|
| 78 |
|
| 79 |
# ------------------------------------------
|
| 80 |
# LLM SETTINGS
|
| 81 |
-
# ------------------------------------------
|
| 82 |
GROQ_API_KEY = os.getenv('GROQ_API_KEY') # Loaded from .env
|
| 83 |
HF_API_KEY = os.getenv('HF_API_KEY')
|
|
|
|
|
|
|
| 84 |
LLM_MODEL_NAME = 'llama-3.3-70b-versatile' # Groq model ID
|
| 85 |
LLM_TEMPERATURE = 0.1 # Low = More factual/consistent
|
| 86 |
LLM_MAX_TOKENS = 2048 # Max response tokens
|
|
|
|
| 78 |
|
| 79 |
# ------------------------------------------
|
| 80 |
# LLM SETTINGS
|
|
|
|
| 81 |
GROQ_API_KEY = os.getenv('GROQ_API_KEY') # Loaded from .env
|
| 82 |
HF_API_KEY = os.getenv('HF_API_KEY')
|
| 83 |
+
if HF_API_KEY:
|
| 84 |
+
os.environ["HF_TOKEN"] = HF_API_KEY
|
| 85 |
LLM_MODEL_NAME = 'llama-3.3-70b-versatile' # Groq model ID
|
| 86 |
LLM_TEMPERATURE = 0.1 # Low = More factual/consistent
|
| 87 |
LLM_MAX_TOKENS = 2048 # Max response tokens
|