[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:
@@ -32,6 +32,9 @@ from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig,
|
||||
QuantizeMethodBase,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.gptq_utils import (
|
||||
normalize_and_validate_gptq_desc_act,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
||||
INT4_DTYPE,
|
||||
INT8_DTYPE,
|
||||
@@ -69,7 +72,11 @@ class MoeWNA16Config(QuantizationConfig):
|
||||
from vllm.model_executor.layers.quantization.auto_awq import AutoAWQConfig
|
||||
|
||||
if self.linear_quant_method == "gptq":
|
||||
pass
|
||||
normalize_and_validate_gptq_desc_act(
|
||||
full_config.get("desc_act", False),
|
||||
group_size,
|
||||
full_config.get("dynamic") or {},
|
||||
)
|
||||
elif self.linear_quant_method in ("awq", "awq_marlin"):
|
||||
capability_tuple = current_platform.get_device_capability()
|
||||
device_capability = (
|
||||
@@ -148,7 +155,16 @@ class MoeWNA16Config(QuantizationConfig):
|
||||
# Extract data from quant config.
|
||||
quant_method = quant_config.get("quant_method", "").lower()
|
||||
num_bits = quant_config.get("bits")
|
||||
desc_act = quant_config.get("desc_act")
|
||||
desc_act = quant_config.get("desc_act", False)
|
||||
try:
|
||||
normalize_and_validate_gptq_desc_act(
|
||||
desc_act,
|
||||
quant_config.get("group_size"),
|
||||
quant_config.get("dynamic") or {},
|
||||
)
|
||||
supported_act_order = True
|
||||
except ValueError:
|
||||
supported_act_order = False
|
||||
|
||||
capability_tuple = current_platform.get_device_capability()
|
||||
device_capability = (
|
||||
@@ -159,7 +175,9 @@ class MoeWNA16Config(QuantizationConfig):
|
||||
|
||||
awq_min_capability = AutoAWQConfig.get_min_capability()
|
||||
|
||||
gptq_compatible = quant_method == "gptq" and not desc_act and num_bits in [4, 8]
|
||||
gptq_compatible = (
|
||||
quant_method == "gptq" and supported_act_order and num_bits in [4, 8]
|
||||
)
|
||||
awq_compatible = (
|
||||
quant_method == "awq"
|
||||
and num_bits == 4
|
||||
@@ -356,7 +374,7 @@ class MoeWNA16Method(FusedMoEMethodBase):
|
||||
if self.quant_config.linear_quant_method == "gptq":
|
||||
# some param are unused, but we need to init them in order to
|
||||
# load weights
|
||||
invalid_param_keys = ["w13_g_idx", "w2_g_idx"]
|
||||
invalid_param_keys: list[str] = []
|
||||
if not self.quant_config.has_zp:
|
||||
invalid_param_keys += ["w13_qzeros", "w2_qzeros"]
|
||||
for key in invalid_param_keys:
|
||||
@@ -414,8 +432,6 @@ class MoeWNA16Method(FusedMoEMethodBase):
|
||||
w2=layer.w2_qweight,
|
||||
w13_scale=layer.w13_scales,
|
||||
w2_scale=layer.w2_scales,
|
||||
w13_g_idx=None,
|
||||
w2_g_idx=None,
|
||||
w13_qzeros=layer.w13_qzeros if has_zp else None,
|
||||
w2_qzeros=layer.w2_qzeros if has_zp else None,
|
||||
)
|
||||
@@ -430,10 +446,6 @@ class MoeWNA16Method(FusedMoEMethodBase):
|
||||
w2_qweight,
|
||||
w13_scales,
|
||||
w2_scales,
|
||||
_,
|
||||
_,
|
||||
_,
|
||||
_,
|
||||
w13_qzeros,
|
||||
w2_qzeros,
|
||||
w13_input_global_scale,
|
||||
@@ -576,8 +588,6 @@ class MoeWNA16Method(FusedMoEMethodBase):
|
||||
expert_id: int,
|
||||
return_success: bool = False,
|
||||
):
|
||||
if "g_idx" in weight_name:
|
||||
return False if return_success else None
|
||||
if not layer.quant_config.has_zp and "qzeros" in weight_name:
|
||||
return False if return_success else None
|
||||
|
||||
|
||||
Reference in New Issue
Block a user