-
Notifications
You must be signed in to change notification settings - Fork 19
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
gushiqiao
committed
May 30, 2024
1 parent
8177310
commit 4513e93
Showing
2 changed files
with
96 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,41 @@ | ||
## Alignment with the Original Paper | ||
|
||
### The conda environment is consistent with the requirements.txt file | ||
### All other configurations are aligned with the original paper/code: | ||
|
||
| | calib_data | seq_len | num_data | seed | | ||
|-------------|------------|---------|----------|------| | ||
| GPTQ | c4 | 2048 | 128 | 0 | | ||
| AWQ | pileval | 512 | 128 | 42 | | ||
| Omniquant | wikitext2 | 2048 | 128 | 2 | | ||
| Smoothquant | pileval | 512 | 128 | 42 | | ||
| Os_plus | pileval | 512 | 128 | 42 | | ||
|
||
### Results | ||
|
||
#### Weight-Only Asymmetric Quantization Results | ||
|
||
| | w4a16g128 | w3a16g128 | w2a16g64 | | ||
|--------------|--------|--------|--------| | ||
| GPTQ | 5.623 | 6.318 | 14.968 | | ||
| GPTQ-LLMC | 5.623 | 6.318 | 14.968 | | ||
| AWQ | | | | | ||
| AWQ-LLMC | | | | | ||
| Omniquant | 5.590 | 6.092 | 9.525 | | ||
| Omniquant-LLMC | 5.590 | 6.092 | 9.525 | | ||
|
||
#### Weight-Activation Asymmetric Quantization Results | ||
|
||
| | w8a8 | w6a6 | w4a4 | | ||
|---------------|--------|--------|--------| | ||
| Omniquant | 5.491 | 5.703 | 12.212 | | ||
| Omniquant-LLMC | 5.490 | 5.703 | 12.239 | | ||
|
||
#### Weight-Activation Symmetric Quantization Results | ||
|
||
| | w8A8 | | ||
|---------------|-------| | ||
| SmoothQuant | | | ||
| SmoothQuant-LLMC | | | ||
| Os_plus | | | ||
| Os_plus-LLMC | | |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,55 @@ | ||
## Impact of calibration data | ||
|
||
### Setting 1: w4a16g128 llama2-7b seq_len=512 | ||
|
||
#### Calibrate with wikitext2 | ||
|
||
| | wikitext2 | c4 | ptb | | ||
|--------|-----------|------|--------| | ||
| GPTQ | **5.575** | 7.470 | 63.575 | | ||
| AWQ | **5.595** | 7.444 | 35.167 | | ||
| OmniQuant | **5.586** | 7.455 | 34.192 | | ||
|
||
#### Calibrate with c4 | ||
|
||
| | wikitext2 | c4 | ptb | | ||
|--------|-----------|-------|---------| | ||
| GPTQ | 5.615 | **7.443** | 122.070 | | ||
| AWQ | 5.596 | **7.436** | 33.148 | | ||
| OmniQuant | 5.620 | 7.457 | 34.001 | | ||
|
||
#### Calibrate with pileval | ||
|
||
| | wikitext2 | c4 | ptb | | ||
|--------|-----------|-------|---------| | ||
| GPTQ | 5.610 | 7.477 | 136.84 | | ||
| AWQ | 5.613 | 7.438 | 33.18 | | ||
| OmniQuant | 5.618 | 7.458 | 34.526 | | ||
|
||
|
||
### Setting 2: w3a16g128 llama2-7b seq_len=512 | ||
|
||
#### Calibrate with wikitext2 | ||
|
||
| | wikitext2 | c4 | ptb | | ||
|--------|-----------|------|----------| | ||
| GPTQ | **6.133** | 8.696 | 234.977 | | ||
| AWQ | **6.138** | 8.272 | 38.86 | | ||
| OmniQuant | **6.096** | 8.325 | 40.667 | | ||
|
||
#### Calibrate with c4 | ||
|
||
| | wikitext2 | c4 | ptb | | ||
|--------|-----------|-------|----------| | ||
| GPTQ | 6.324 | **8.385** | 358.013 | | ||
| AWQ | 6.181 | **8.249** | 39.27 | | ||
| OmniQuant | 6.259 | **8.317** | 41.835 | | ||
|
||
|
||
#### Calibrate with pileval | ||
|
||
| | wikitext2 | c4 | ptb | | ||
|--------|-----------|-------|---------| | ||
| GPTQ | 6.330 | 8.534 | 263.279 | | ||
| AWQ | 6.217 | 8.284 | 37.117 | | ||
| OmniQuant | 6.214 | 8.320 | 42.335 | |