Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

remove ark support for cpu (#1206)

Wenhua Cheng083b635105c4
diff --git a/auto_round/inference/backend.py b/auto_round/inference/backend.py
index 057ecf5bd..3b243bff0 100644
--- a/auto_round/inference/backend.py
+++ b/auto_round/inference/backend.py
@@ -439,21 +439,21 @@ BackendInfos["auto_awq:gemm"] = BackendInfo(
requirements=["autoawq", "transformers"],
)
-BackendInfos["auto_round_kernel"] = BackendInfo(
- device=["cpu"],
- sym=[True, False],
- packing_format=GPTQ_FORMAT_NO_ZP,
- bits=[2, 4, 8],
- group_size=None,
- priority=6,
- checkers=[ark_feature_checker],
- alias=["ark"],
- compute_dtype=["float32", "float16"],
- data_type=["int"],
- act_bits=WOQ_DEFAULT_ACT_BITS,
- requirements=["torch>=2.9.0", "auto_round_kernel"],
- systems=["linux"],
-)
+# BackendInfos["auto_round_kernel"] = BackendInfo(
+# device=["cpu"],
+# sym=[True, False],
+# packing_format=GPTQ_FORMAT_NO_ZP,
+# bits=[2, 4, 8],
+# group_size=None,
+# priority=6,
+# checkers=[ark_feature_checker],
+# alias=["ark"],
+# compute_dtype=["float32", "float16"],
+# data_type=["int"],
+# act_bits=WOQ_DEFAULT_ACT_BITS,
+# requirements=["torch>=2.9.0", "auto_round_kernel"],
+# systems=["linux"],
+# )
BackendInfos["auto_round_kernel_xpu"] = BackendInfo(
device=["xpu"],
@@ -471,11 +471,27 @@ BackendInfos["auto_round_kernel_xpu"] = BackendInfo(
systems=["linux"],
)
-BackendInfos["auto_round_kernel_zp"] = BackendInfo(
- device=["cpu"],
- sym=[True, False],
+# BackendInfos["auto_round_kernel_zp"] = BackendInfo(
+# device=["cpu"],
+# sym=[True, False],
+# packing_format=GPTQ_FORMAT,
+# bits=[2, 4, 8],
+# group_size=None,
+# priority=6,
+# checkers=[ark_feature_checker],
+# alias=["ark"],
+# compute_dtype=["float32", "float16"],
+# data_type=["int"],
+# act_bits=WOQ_DEFAULT_ACT_BITS,
+# requirements=["torch>=2.9.0", "auto_round_kernel"],
+# systems=["linux"],
+# )
+
+BackendInfos["auto_round_kernel_zp_xpu"] = BackendInfo(
+ device=["xpu"],
+ sym=[True],
packing_format=GPTQ_FORMAT,
- bits=[2, 4, 8],
+ bits=[4, 8],
group_size=None,
priority=6,
checkers=[ark_feature_checker],
@@ -487,10 +503,26 @@ BackendInfos["auto_round_kernel_zp"] = BackendInfo(
systems=["linux"],
)
-BackendInfos["auto_round_kernel_zp_xpu"] = BackendInfo(
+# BackendInfos["auto_round_kernel_awq"] = BackendInfo(
+# device=["cpu"],
+# sym=[True, False],
+# packing_format=AWQ_FORMAT,
+# bits=[2, 4, 8],
+# group_size=None,
+# priority=6,
+# checkers=[ark_feature_checker],
+# alias=["ark"],
+# compute_dtype=["float32", "float16"],
+# data_type=["int"],
+# act_bits=WOQ_DEFAULT_ACT_BITS,
+# requirements=["torch>=2.9.0", "auto_round_kernel"],
+# systems=["linux"],
+# )
+
+BackendInfos["auto_round_kernel_awq_xpu"] = BackendInfo(
device=["xpu"],
sym=[True],
- packing_format=GPTQ_FORMAT,
+ packing_format=AWQ_FORMAT,
bits=[4, 8],
group_size=None,
priority=6,
@@ -503,42 +535,40 @@ BackendInfos["auto_round_kernel_zp_xpu"] = BackendInfo(
systems=["linux"],
)
-BackendInfos["auto_round_kernel_awq"] = BackendInfo(
+BackendInfos["ipex_gptq_cpu"] = BackendInfo(
device=["cpu"],
sym=[True, False],
- packing_format=AWQ_FORMAT,
- bits=[2, 4, 8],
+ packing_format=GPTQ_FORMAT,
+ bits=[4],
group_size=None,
- priority=6,
- checkers=[ark_feature_checker],
- alias=["ark"],
- compute_dtype=["float32", "float16"],
+ priority=5,
+ checkers=[],
+ compute_dtype=["float16", "bfloat16"],
data_type=["int"],
act_bits=WOQ_DEFAULT_ACT_BITS,
- requirements=["torch>=2.9.0", "auto_round_kernel"],
- systems=["linux"],
+ alias=["ipex"],
+ requirements=["torch<2.9", "intel-extension-for-pytorch>=2.5"],
)
-BackendInfos["auto_round_kernel_awq_xpu"] = BackendInfo(
+BackendInfos["ipex_gptq"] = BackendInfo(
device=["xpu"],
- sym=[True],
- packing_format=AWQ_FORMAT,
- bits=[4, 8],
+ sym=[True, False],
+ packing_format=GPTQ_FORMAT,
+ bits=[4],
group_size=None,
- priority=6,
- checkers=[ark_feature_checker],
- alias=["ark"],
- compute_dtype=["float32", "float16"],
+ priority=5,
+ checkers=[],
+ compute_dtype=["float16", "bfloat16"],
data_type=["int"],
act_bits=WOQ_DEFAULT_ACT_BITS,
- requirements=["torch>=2.9.0", "auto_round_kernel"],
- systems=["linux"],
+ alias=["ipex"],
+ requirements=["intel-extension-for-pytorch>=2.5"],
)
-BackendInfos["ipex_gptq"] = BackendInfo(
- device=["cpu", "xpu"],
+BackendInfos["ipex_awq_cpu"] = BackendInfo(
+ device=["cpu"],
sym=[True, False],
- packing_format=GPTQ_FORMAT,
+ packing_format=AWQ_FORMAT,
bits=[4],
group_size=None,
priority=5,
@@ -547,11 +577,12 @@ BackendInfos["ipex_gptq"] = BackendInfo(
data_type=["int"],
act_bits=WOQ_DEFAULT_ACT_BITS,
alias=["ipex"],
- requirements=["intel-extension-for-pytorch>=2.5"],
+ requirements=["torch<2.9", "intel-extension-for-pytorch>=2.5"],
)
+
BackendInfos["ipex_awq"] = BackendInfo(
- device=["cpu", "xpu"],
+ device=["cpu"],
sym=[True, False],
packing_format=AWQ_FORMAT,
bits=[4],