Commit
·
bd9ae91
1
Parent(s):
17ff8ed
Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
## This is a 4bit quant of https://huggingface.co/MetaIX/GPT4-X-Alpasta-30b
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
|
| 5 |
+
# My secret sauce:
|
| 6 |
+
* Using comit <a href="https://github.com/0cc4m/GPTQ-for-LLaMa/tree/3c16fd9c7946ebe85df8d951cb742adbc1966ec7">3c16fd9</a> of 0cc4m's GPTQ fork
|
| 7 |
+
* Using C4 as the calibration dataset
|
| 8 |
+
* Act-order, True-sequential, percdamp 0.1
|
| 9 |
+
(<i>the default percdamp is 0.01</i>)
|
| 10 |
+
* No groupsize
|
| 11 |
+
* Will run with CUDA, does not need triton.
|
| 12 |
+
* Quant completed on a 'Premium GPU' and 'High Memory' Google Colab.
|
| 13 |
+
|
| 14 |
+
## Benchmark results
|
| 15 |
+
|
| 16 |
+
|<b>Model<b>|<b>C4<b>|<b>WikiText2<b>|<b>PTB<b>|
|
| 17 |
+
|:---:|---|---|---|
|
| 18 |
+
|MetaIX's FP16|6.98400259|4.607768536|9.414786339|
|
| 19 |
+
|This Quant|7.292364597|4.954069614|9.754593849|
|