Created
August 24, 2024 15:57
-
-
Save aredden/a92434a98b794ca13becf3d607f5650b to your computer and use it in GitHub Desktop.
quantize fp8 using quantile vs absmax for very ood values
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
| import torch | |
| torch.set_printoptions(precision=4, sci_mode=False) | |
| import triton | |
| import triton.language as tl | |
| from torch import Tensor | |
| def quanitze_fp8_tensorwise(x: torch.Tensor, dtype=torch.float8_e4m3fn): | |
| scale = x.abs().max() / torch.finfo(dtype).max | |
| x = x.float() / scale | |
| return x.to(dtype), scale.float() | |
| def quanitze_fp8_tensorwise_quantile(weight: torch.Tensor, dtype=torch.float8_e4m3fn): | |
| finfo = torch.finfo(dtype) | |
| quanti = torch.quantile( | |
| weight.abs().float(), 0.999, dim=0, interpolation="lower" | |
| ).max() | |
| scale = quanti / finfo.max | |
| q_weight = (weight.float() / scale).clamp(min=-finfo.max, max=finfo.max).to(dtype) | |
| return q_weight, scale.float() | |
| def fp8_linear_torch( | |
| x: torch.Tensor, | |
| x_scale: torch.Tensor, | |
| weight_fp8: torch.Tensor, | |
| weight_scale: torch.Tensor, | |
| ): | |
| out = torch._scaled_mm( | |
| x, weight_fp8.T, scale_a=x_scale, scale_b=weight_scale, out_dtype=torch.bfloat16 | |
| ) | |
| return out | |
| if __name__ == "__main__": | |
| from triton.testing import do_bench | |
| act_bf16 = torch.randn(1024, 2048).bfloat16().cuda() | |
| weight_bf16 = torch.randn(4096, 2048).bfloat16().cuda() | |
| act_bf16[0, 0] = 62320 | |
| ref = act_bf16 @ weight_bf16.T | |
| q_weight, q_w_scale = quanitze_fp8_tensorwise_quantile(weight_bf16) | |
| q_act, q_a_scale = quanitze_fp8_tensorwise_quantile(act_bf16) | |
| abs_weight, abs_w_scale = quanitze_fp8_tensorwise(weight_bf16) | |
| abs_act, abs_a_scale = quanitze_fp8_tensorwise(act_bf16) | |
| out_quantile = fp8_linear_torch(q_act, q_a_scale, q_weight, q_w_scale) | |
| out_abs = fp8_linear_torch(abs_act, abs_a_scale, abs_weight, abs_w_scale) | |
| print( | |
| "Median abs diff for fp8 quantile: ", | |
| (out_quantile - ref).abs().median(), | |
| ) | |
| print( | |
| "Median abs diff for fp8 absmax: ", | |
| (out_abs - ref).abs().median(), | |
| ) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment