kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-13 22:18:36 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

Merge 019ab04531 into 38296a472b

Browse Source
This commit is contained in:
muse-coder 2026-01-13 03:17:11 -08:00 committed by GitHub
commit 366b9b0c7c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
tests/unittest/trt/functional/test_fp4_gemm.py
View File

@ -83,8 +83,8 @@ def random_fp4_tensor_and_sf_v2(shape, sf_vec_size):
float_tensor = torch.randn(shape, dtype=torch.float32)
half_tensor = float_tensor.to(torch.float16).cuda()
# global scale trick for int4 quantization.
alpha = 448.0 / (torch.max(float_tensor) / 6.0)
# global scale trick for fp4 quantization.
alpha = (448 * 6) / float_tensor.abs().max().float()
sf_scale_tensor = torch.FloatTensor([alpha]).cuda()
gemm_alpha_tensor = torch.FloatTensor([1.0 / alpha])
@ -282,9 +282,8 @@ class TestFunctional(unittest.TestCase):
input_fp16 = input_fp32.to(torch.float16).cuda()
# global scale trick for int4 quantization.
alpha = 448.0 / (torch.max(input_fp32) / 6.0)
# global scale trick for fp4 quantization.
alpha = (448 * 6) / input_fp32.abs().max().float()
weights_fp32_transposed = torch.transpose(weights_fp32, 0, 1)
sf_scale_tensor = torch.FloatTensor([alpha]).cuda()
act_unscale_tensor = torch.FloatTensor([1.0 / alpha]).cuda()