Added TinyLlama as requested by Jim burtoft
Browse files
inference-cache-config/llama.json
CHANGED
@@ -82,5 +82,13 @@
|
|
82 |
"num_cores": 2,
|
83 |
"auto_cast_type": "bf16"
|
84 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
]
|
86 |
}
|
|
|
82 |
"num_cores": 2,
|
83 |
"auto_cast_type": "bf16"
|
84 |
}
|
85 |
+
],
|
86 |
+
"TinyLlama/TinyLlama-1.1B-Chat-v1.0": [
|
87 |
+
{
|
88 |
+
"batch_size": 1,
|
89 |
+
"sequence_length": 2048,
|
90 |
+
"num_cores": 2,
|
91 |
+
"auto_cast_type": "bf16"
|
92 |
+
}
|
93 |
]
|
94 |
}
|