mlboydaisuke commited on
Commit
bb4fa29
·
verified ·
1 Parent(s): 1e553b4

Add the blockwise-128 int4 recipe (reproduce with stock litert-torch)

Browse files
Files changed (1) hide show
  1. falcon_int4_block128.json +36 -0
falcon_int4_block128.json ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "regex": ".*",
4
+ "operation": "*",
5
+ "algorithm_key": "min_max_uniform_quantize",
6
+ "op_config": {
7
+ "weight_tensor_config": {
8
+ "num_bits": 4,
9
+ "symmetric": true,
10
+ "granularity": "BLOCKWISE_128",
11
+ "dtype": "INT"
12
+ },
13
+ "compute_precision": "INTEGER",
14
+ "explicit_dequantize": false,
15
+ "skip_checks": false,
16
+ "min_weight_elements": 0
17
+ }
18
+ },
19
+ {
20
+ "regex": ".*",
21
+ "operation": "EMBEDDING_LOOKUP",
22
+ "algorithm_key": "min_max_uniform_quantize",
23
+ "op_config": {
24
+ "weight_tensor_config": {
25
+ "num_bits": 8,
26
+ "symmetric": true,
27
+ "granularity": "CHANNELWISE",
28
+ "dtype": "INT"
29
+ },
30
+ "compute_precision": "INTEGER",
31
+ "explicit_dequantize": false,
32
+ "skip_checks": false,
33
+ "min_weight_elements": 0
34
+ }
35
+ }
36
+ ]