Rypo commited on
Commit
82fa202
·
1 Parent(s): 34dc9d8

skip norm quantization for numerical stability

Browse files
Files changed (2) hide show
  1. config.json +3 -1
  2. model.safetensors +2 -2
config.json CHANGED
@@ -25,7 +25,9 @@
25
  "bnb_4bit_use_double_quant": false,
26
  "llm_int8_enable_fp32_cpu_offload": false,
27
  "llm_int8_has_fp16_weight": false,
28
- "llm_int8_skip_modules": null,
 
 
29
  "llm_int8_threshold": 6.0,
30
  "load_in_4bit": true,
31
  "load_in_8bit": false,
 
25
  "bnb_4bit_use_double_quant": false,
26
  "llm_int8_enable_fp32_cpu_offload": false,
27
  "llm_int8_has_fp16_weight": false,
28
+ "llm_int8_skip_modules": [
29
+ "norm"
30
+ ],
31
  "llm_int8_threshold": 6.0,
32
  "load_in_4bit": true,
33
  "load_in_8bit": false,
model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:765aa6cf02507cfdd01a6e18d20991973886e9b145fc88bf9223bb69228f693d
3
- size 2484808522
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:138531c3da36391ade90542592b0ac60fc4970caefa14578199c10ca575d3b43
3
+ size 2484808656