[Quant][Kernel] Remove GPTQ Group/Dynamic Activation Ordering (#54809)

Signed-off-by: Roderick-Wu <roderickwu2003@gmail.com>
Signed-off-by: Roderick Wu <roderickwu2003@gmail.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Roderick Wu
2026-09-08 12:45:12 -07:00
committed by GitHub
co-authored by mgoin
parent 1a522b6949
commit 2c9d68f80b
99 changed files with 1182 additions and 3917 deletions
+3 -9
View File
@@ -11,9 +11,8 @@ def test_gptq_shuffle_opcheck():
weight = torch.randint(
-2000000, 2000000, (1792, 4096), device="cuda", dtype=torch.int32
)
perm = torch.empty((0,), device="cuda", dtype=torch.int32)
bit = 4
opcheck(torch.ops._C.gptq_shuffle, (weight, perm, bit))
opcheck(torch.ops._C.gptq_shuffle, (weight, bit))
def test_gptq_gemm_opcheck():
@@ -23,13 +22,8 @@ def test_gptq_gemm_opcheck():
)
zeros = torch.zeros((32, 768), device="cuda", dtype=torch.int32)
scales = torch.rand((32, 6144), device="cuda", dtype=torch.float16)
idx = torch.empty((0,), device="cuda", dtype=torch.int32)
use_exllama = True
bit = 4
# Test both GPTQv1 and GPTQv2 format
opcheck(
torch.ops._C.gptq_gemm, (a, weight, zeros, scales, idx, use_exllama, True, bit)
)
opcheck(
torch.ops._C.gptq_gemm, (a, weight, zeros, scales, idx, use_exllama, False, bit)
)
opcheck(torch.ops._C.gptq_gemm, (a, weight, zeros, scales, use_exllama, True, bit))
opcheck(torch.ops._C.gptq_gemm, (a, weight, zeros, scales, use_exllama, False, bit))