Download scripts/verify_quant_linear.py from jp2023213625/nlp-chapter6-nextgpt: direct link, hf CLI and curl.
- Browser
- Download file 1.11 kB
-
https://huggingface.co/spaces/jp2023213625/nlp-chapter6-nextgpt/resolve/main/scripts/verify_quant_linear.py
- Command line
-
hf download hf://spaces/jp2023213625/nlp-chapter6-nextgpt/scripts/verify_quant_linear.py
-
curl -L -o verify_quant_linear.py https://huggingface.co/spaces/jp2023213625/nlp-chapter6-nextgpt/resolve/main/scripts/verify_quant_linear.py
1.11 kB
| """Numerically check portable GPTQ unpacking + LoRA against a dense layer.""" | |
| import torch | |
| from minimal_inference import GPTQLinear | |
| torch.manual_seed(13) | |
| qin=torch.randint(0,16,(32,16),dtype=torch.int32) | |
| zero=torch.full((2,16),8,dtype=torch.int32) | |
| groups=torch.arange(32,dtype=torch.int32)//16 | |
| # Non-monotonic grouping checks activation-order g_idx indexing. | |
| groups=groups[torch.randperm(32)] | |
| scales=torch.rand(2,16,dtype=torch.float16) | |
| qw=torch.zeros(4,16,dtype=torch.int32) | |
| qz=torch.zeros(2,2,dtype=torch.int32) | |
| for j in range(8): | |
| qw|=qin[j::8]<<(4*j) | |
| qz|=(zero[:,j::8]-1)<<(4*j) | |
| a=torch.randn(4,32,dtype=torch.float16)/20 | |
| b=torch.randn(16,4,dtype=torch.float16)/20 | |
| layer=GPTQLinear({'qweight':qw,'qzeros':qz,'scales':scales,'g_idx':groups},a,b) | |
| x=torch.randn(1,3,32,device='cuda',dtype=torch.float16) | |
| dense=((qin-zero[groups]).half()*scales[groups]).cuda() | |
| expected=x@dense + (x@a.cuda().T)@b.cuda().T | |
| with torch.no_grad():actual=layer(x) | |
| torch.testing.assert_close(actual,expected,atol=.01,rtol=.01) | |
| print('PASS: GPTQ 4bit unpacking, activation-order groups and LoRA match dense reference.') | |