[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
@@ -439,7 +439,6 @@ class AutoAWQMarlinLinearMethod(LinearMethodBase):
act_type=params_dtype if self.input_dtype is None else self.input_dtype,
group_size=self.quant_config.group_size,
zero_points=self.quant_config.zero_point,
has_g_idx=False,
)
kernel_type = choose_mp_linear_kernel(mp_linear_kernel_config)
@@ -558,11 +557,6 @@ class AutoAWQMoEMethod(FusedMoEMethodBase):
}
)
intermediate_size_full = extra_weight_attrs.pop(
"intermediate_size_full", intermediate_size_per_partition
)
self.is_k_full = intermediate_size_per_partition == intermediate_size_full
w13_qweight = Parameter(
torch.empty(
num_experts,
@@ -672,10 +666,6 @@ class AutoAWQMoEMethod(FusedMoEMethodBase):
w2,
w13_scale,
w2_scale,
w13_g_idx,
w2_g_idx,
w13_g_idx_sort_indices,
w2_g_idx_sort_indices,
w13_qzeros,
w2_qzeros,
w13_input_global_scale,
@@ -696,14 +686,6 @@ class AutoAWQMoEMethod(FusedMoEMethodBase):
replace_parameter(layer, "w13_scales", w13_scale)
replace_parameter(layer, "w2_scales", w2_scale)
_replace_or_register_parameter(
layer, "w13_g_idx_sort_indices", w13_g_idx_sort_indices
)
_replace_or_register_parameter(
layer, "w2_g_idx_sort_indices", w2_g_idx_sort_indices
)
_replace_or_register_parameter(layer, "w13_g_idx", w13_g_idx)
_replace_or_register_parameter(layer, "w2_g_idx", w2_g_idx)
_replace_or_register_parameter(layer, "w13_qzeros", w13_qzeros)
_replace_or_register_parameter(layer, "w2_qzeros", w2_qzeros)
_replace_or_register_parameter(
@@ -726,11 +708,6 @@ class AutoAWQMoEMethod(FusedMoEMethodBase):
moe_config=self.moe,
experts_cls=self.experts_cls,
backend=self.wna16_moe_backend,
is_k_full=self.is_k_full,
w13_g_idx=getattr(layer, "w13_g_idx", None),
w2_g_idx=getattr(layer, "w2_g_idx", None),
w13_g_idx_sort_indices=getattr(layer, "w13_g_idx_sort_indices", None),
w2_g_idx_sort_indices=getattr(layer, "w2_g_idx_sort_indices", None),
routing_tables=layer._expert_routing_tables(),
)