Update inference-cache-config/llama-variants.json
Browse files
inference-cache-config/llama-variants.json
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
{
|
2 |
-
|
3 |
{
|
4 |
"batch_size": 1,
|
5 |
"sequence_length": 4096,
|
@@ -13,7 +13,7 @@
|
|
13 |
"auto_cast_type": "fp16"
|
14 |
}
|
15 |
],
|
16 |
-
|
17 |
{
|
18 |
"batch_size": 1,
|
19 |
"sequence_length": 4096,
|
@@ -27,37 +27,37 @@
|
|
27 |
"auto_cast_type": "fp16"
|
28 |
}
|
29 |
],
|
30 |
-
"
|
31 |
{
|
32 |
"batch_size": 1,
|
33 |
"sequence_length": 4096,
|
34 |
-
"num_cores":
|
35 |
"auto_cast_type": "fp16"
|
36 |
},
|
37 |
{
|
38 |
"batch_size": 4,
|
39 |
"sequence_length": 4096,
|
40 |
-
"num_cores":
|
41 |
-
"auto_cast_type": "fp16"
|
42 |
-
},
|
43 |
-
{
|
44 |
-
"batch_size": 8,
|
45 |
-
"sequence_length": 4096,
|
46 |
-
"num_cores": 8,
|
47 |
"auto_cast_type": "fp16"
|
48 |
}
|
49 |
],
|
50 |
-
"
|
51 |
{
|
52 |
"batch_size": 1,
|
53 |
"sequence_length": 4096,
|
54 |
-
"num_cores":
|
55 |
"auto_cast_type": "fp16"
|
56 |
},
|
57 |
{
|
58 |
"batch_size": 4,
|
59 |
"sequence_length": 4096,
|
60 |
-
"num_cores":
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
"auto_cast_type": "fp16"
|
62 |
}
|
63 |
],
|
@@ -103,21 +103,21 @@
|
|
103 |
"auto_cast_type": "fp16"
|
104 |
}
|
105 |
],
|
106 |
-
|
107 |
{
|
108 |
"batch_size": 1,
|
109 |
"sequence_length": 4096,
|
110 |
-
"num_cores":
|
111 |
"auto_cast_type": "fp16"
|
112 |
},
|
113 |
{
|
114 |
"batch_size": 4,
|
115 |
"sequence_length": 4096,
|
116 |
-
"num_cores":
|
117 |
"auto_cast_type": "fp16"
|
118 |
}
|
119 |
],
|
120 |
-
"
|
121 |
{
|
122 |
"batch_size": 1,
|
123 |
"sequence_length": 4096,
|
|
|
1 |
{
|
2 |
+
"princeton-nlp/Sheared-LLaMA-1.3B": [
|
3 |
{
|
4 |
"batch_size": 1,
|
5 |
"sequence_length": 4096,
|
|
|
13 |
"auto_cast_type": "fp16"
|
14 |
}
|
15 |
],
|
16 |
+
"defog/sqlcoder-7b-2": [
|
17 |
{
|
18 |
"batch_size": 1,
|
19 |
"sequence_length": 4096,
|
|
|
27 |
"auto_cast_type": "fp16"
|
28 |
}
|
29 |
],
|
30 |
+
"m-a-p/OpenCodeInterpreter-DS-6.7B": [
|
31 |
{
|
32 |
"batch_size": 1,
|
33 |
"sequence_length": 4096,
|
34 |
+
"num_cores": 2,
|
35 |
"auto_cast_type": "fp16"
|
36 |
},
|
37 |
{
|
38 |
"batch_size": 4,
|
39 |
"sequence_length": 4096,
|
40 |
+
"num_cores": 2,
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
"auto_cast_type": "fp16"
|
42 |
}
|
43 |
],
|
44 |
+
"ibm/labradorite-13b": [
|
45 |
{
|
46 |
"batch_size": 1,
|
47 |
"sequence_length": 4096,
|
48 |
+
"num_cores": 8,
|
49 |
"auto_cast_type": "fp16"
|
50 |
},
|
51 |
{
|
52 |
"batch_size": 4,
|
53 |
"sequence_length": 4096,
|
54 |
+
"num_cores": 8,
|
55 |
+
"auto_cast_type": "fp16"
|
56 |
+
},
|
57 |
+
{
|
58 |
+
"batch_size": 8,
|
59 |
+
"sequence_length": 4096,
|
60 |
+
"num_cores": 8,
|
61 |
"auto_cast_type": "fp16"
|
62 |
}
|
63 |
],
|
|
|
103 |
"auto_cast_type": "fp16"
|
104 |
}
|
105 |
],
|
106 |
+
"01-ai/Yi-34B-200K": [
|
107 |
{
|
108 |
"batch_size": 1,
|
109 |
"sequence_length": 4096,
|
110 |
+
"num_cores": 24,
|
111 |
"auto_cast_type": "fp16"
|
112 |
},
|
113 |
{
|
114 |
"batch_size": 4,
|
115 |
"sequence_length": 4096,
|
116 |
+
"num_cores": 24,
|
117 |
"auto_cast_type": "fp16"
|
118 |
}
|
119 |
],
|
120 |
+
"abacusai/Smaug-72B-v0.1": [
|
121 |
{
|
122 |
"batch_size": 1,
|
123 |
"sequence_length": 4096,
|