[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
+1 -6
View File
@@ -220,10 +220,8 @@ def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
bt.w_g_s, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.group_size
)
sort_indices = torch.empty(0, dtype=torch.int, device=device)
g_idx = torch.empty(0, dtype=torch.int, device=device)
w_q = ops.gptq_marlin_repack(
bt.w_q, sort_indices, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.wtype.size_bits
bt.w_q, bt.w_ref.shape[0], bt.w_ref.shape[1], bt.wtype.size_bits
)
if bt.a.dtype.is_floating_point:
@@ -240,14 +238,11 @@ def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
a_scales=None,
global_scale=None,
b_zeros=w_zp,
g_idx=g_idx,
perm=sort_indices,
workspace=workspace.scratch,
b_q_type=bt.wtype,
size_m=bt.a.shape[0],
size_n=bt.w_ref.shape[1],
size_k=bt.w_ref.shape[0],
is_k_full=True,
is_zp_float=False,
)
else: