apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 08:37:06 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[NewFeatures] support eplb (#3547)

Browse Source 
* [NewFeatures] support eplb

* fix eplb
This commit is contained in:
xiaoxiaohehe001
2025-08-26 16:19:30 +08:00
committed by GitHub
parent 56e2d7e668
commit 9afa236e39
17 changed files with 174 additions and 67 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
fastdeploy/model_executor/layers/backends/dcu/fused_moe_triton_backends.py
View File

@@ -38,7 +38,7 @@ class DCUTritonWeightOnlyMoEMethod(QuantMethodBase):
"down_proj_weight_scale",
]
def process_prequanted_weights(self, layer: nn.Layer, state_dict) -> None:
def process_prequanted_weights(self, layer: nn.Layer, state_dict, is_rearrange: bool = False) -> None:
"""process_prequanted_weights"""
pass
@@ -46,7 +46,7 @@ class DCUTritonWeightOnlyMoEMethod(QuantMethodBase):
"""
Triton MoE create weight process.
"""
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
assert len(up_gate_proj_weights) == layer.num_local_experts
assert len(down_proj_weights) == layer.num_local_experts
assert self.quant_method.name() == "wint8"

6
fastdeploy/model_executor/layers/backends/gcu/moe/fused_moe_method_gcu_backend.py
View File

@@ -49,7 +49,7 @@ class GCUFusedMoeMethod(UnquantizedFusedMoEMethod):
self.group_size = -1
def process_loaded_weights(self, layer: nn.Layer, state_dict):
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
stacked_up_gate_proj_weights = paddle.stack(up_gate_proj_weights, axis=0)
stacked_down_proj_weights = paddle.stack(down_proj_weights, axis=0)
layer.up_gate_proj_weight.set_value(paddle.transpose(stacked_up_gate_proj_weights, [0, 2, 1]))
@@ -254,7 +254,7 @@ class GCUWeightOnlyMoEMethod(GCUFusedMoeMethod):
self.quant_multi_process_group_size = int(os.getenv("FD_MOE_QUANT_MULTI_PROCESS_GROUP_SIZE", 8))
logger.info(f"GCUWeightOnlyMoEMethod quant_multi_process_group_size: {self.quant_multi_process_group_size}")
def process_prequanted_weights(self, layer: nn.Layer, state_dict):
def process_prequanted_weights(self, layer: nn.Layer, state_dict, is_rearrange: bool = False):
"""
Paddle gcu process prequanted weights.
"""
@@ -299,7 +299,7 @@ class GCUWeightOnlyMoEMethod(GCUFusedMoeMethod):
"""
Paddle cutlass create weight process.
"""
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
self.check(layer, up_gate_proj_weights, down_proj_weights)
def quant_worker(p_group_idx, shared_dict, weights, moe_quant_type, group_size):

2
fastdeploy/model_executor/layers/backends/gcu/quantization/weight_only.py
View File

@@ -59,7 +59,7 @@ class GCUWeightOnlyLinearMethod(WeightOnlyLinearMethod):
is_bias=False,
)
def process_prequanted_weights(self, layer, state_dict) -> None:
def process_prequanted_weights(self, layer, state_dict, is_rearrange: bool = False) -> None:
"""
Process pre-quantized weights before applying them to the model
Args:

4
fastdeploy/model_executor/layers/backends/metax/moe/fused_moe_triton_metax_backend.py
View File

@@ -41,7 +41,7 @@ class MetaxTritonWeightOnlyMoEMethod(QuantMethodBase):
"down_proj_weight_scale",
]
def process_prequanted_weights(self, layer: nn.Layer, state_dict) -> None:
def process_prequanted_weights(self, layer: nn.Layer, state_dict, is_rearrange: bool = False) -> None:
"""process_prequanted_weights"""
pass
@@ -50,7 +50,7 @@ class MetaxTritonWeightOnlyMoEMethod(QuantMethodBase):
"""
Triton MoE create weight process.
"""
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
assert len(up_gate_proj_weights) == layer.num_local_experts
assert len(down_proj_weights) == layer.num_local_experts

109
fastdeploy/model_executor/layers/moe/fused_moe_cutlass_backend.py
View File

@@ -74,7 +74,9 @@ class CutlassMoEMethod(UnquantizedFusedMoEMethod):
"""
def process_loaded_weights(self, layer: nn.Layer, state_dict):
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, logical_expert_ids, ep_rank_to_expert_id_list = (
layer.extract_moe_ffn_weights(state_dict)
)
stacked_up_gate_proj_weights = paddle.stack(up_gate_proj_weights, axis=0)
stacked_down_proj_weights = paddle.stack(down_proj_weights, axis=0)
@@ -333,7 +335,7 @@ class CutlassW4A8MoEMethod(CutlassMoEMethod):
self.moe_quant_type = "w4a8"
self.pack_num = 2
def process_prequanted_weights(self, layer: nn.Layer, state_dict):
def process_prequanted_weights(self, layer: nn.Layer, state_dict, is_rearrange: bool = False):
"""
Paddle cutlass process prequanted weights.
"""
@@ -349,6 +351,7 @@ class CutlassW4A8MoEMethod(CutlassMoEMethod):
state_dict,
up_gate_proj_expert_weight_key,
down_proj_expert_weight_key,
is_rearrange,
)
)
@@ -358,22 +361,62 @@ class CutlassW4A8MoEMethod(CutlassMoEMethod):
up_gate_proj_in_scale = []
down_proj_in_scale = []
if isinstance(state_dict, list):
state_dict = dict(state_dict)
if layer.ep_size > 1:
for expert_idx in ep_rank_to_expert_id_list:
scale_tensor = get_tensor(state_dict[up_gate_proj_expert_in_scale_key.format(expert_idx)])
scale_tensor = get_tensor(
(
state_dict[up_gate_proj_expert_in_scale_key.format(expert_idx)]
if up_gate_proj_expert_in_scale_key.format(expert_idx) in state_dict
else up_gate_proj_expert_in_scale_key.format(expert_idx)
),
layer.fd_config.model_config.model,
)
up_gate_proj_in_scale_all_experts.append(scale_tensor)
for expert_idx in logical_expert_ids:
up_gate_proj_weight_scale.append(
get_tensor(state_dict.pop(up_gate_proj_expert_weight_scale_key.format(expert_idx)))
get_tensor(
(
state_dict.pop(up_gate_proj_expert_weight_scale_key.format(expert_idx))
if up_gate_proj_expert_weight_scale_key.format(expert_idx) in state_dict
else up_gate_proj_expert_weight_scale_key.format(expert_idx)
),
layer.fd_config.model_config.model,
)
)
down_proj_weight_scale.append(
get_tensor(state_dict.pop(down_proj_expert_weight_scale_key.format(expert_idx)))
get_tensor(
(
state_dict.pop(down_proj_expert_weight_scale_key.format(expert_idx))
if down_proj_expert_weight_scale_key.format(expert_idx) in state_dict
else down_proj_expert_weight_scale_key.format(expert_idx)
),
layer.fd_config.model_config.model,
)
)
up_gate_proj_in_scale.append(
get_tensor(state_dict.pop(up_gate_proj_expert_in_scale_key.format(expert_idx)))
get_tensor(
(
state_dict.pop(up_gate_proj_expert_in_scale_key.format(expert_idx))
if up_gate_proj_expert_in_scale_key.format(expert_idx) in state_dict
else up_gate_proj_expert_in_scale_key.format(expert_idx)
),
layer.fd_config.model_config.model,
)
)
down_proj_in_scale.append(
get_tensor(
(
state_dict.pop(down_proj_expert_in_scale_key.format(expert_idx))
if down_proj_expert_in_scale_key.format(expert_idx) in state_dict
else down_proj_expert_in_scale_key.format(expert_idx)
),
layer.fd_config.model_config.model,
)
)
down_proj_in_scale.append(get_tensor(state_dict.pop(down_proj_expert_in_scale_key.format(expert_idx))))
up_gate_proj_weight = paddle.stack(up_gate_proj_weights, axis=0)
down_proj_weight = paddle.stack(down_proj_weights, axis=0)
@@ -435,7 +478,9 @@ class CutlassW4A8MoEMethod(CutlassMoEMethod):
"""
Paddle cutlass load weight process.
"""
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, logical_expert_ids, ep_rank_to_expert_id_list = (
layer.extract_moe_ffn_weights(state_dict)
)
self.check(layer, up_gate_proj_weights, down_proj_weights)
for idx, weight_tensor in enumerate([up_gate_proj_weights, down_proj_weights]):
weight_name = self.added_weight_attrs[idx]
@@ -446,7 +491,9 @@ class CutlassW4A8MoEMethod(CutlassMoEMethod):
quanted_weight = paddle.stack(weight_list, axis=0)
getattr(layer, weight_name).set_value(quanted_weight)
self.load_w4a8_scale_weights(layer, layer.weight_key_map, state_dict)
self.load_w4a8_scale_weights(
layer, layer.weight_key_map, state_dict, logical_expert_ids, ep_rank_to_expert_id_list
)
def create_w4a8_scale_weights(self, layer: nn.Layer, weight_key_map: dict):
"""
@@ -499,7 +546,14 @@ class CutlassW4A8MoEMethod(CutlassMoEMethod):
),
)
def load_w4a8_scale_weights(self, layer: nn.Layer, weight_key_map: dict, state_dict: dict):
def load_w4a8_scale_weights(
self,
layer: nn.Layer,
weight_key_map: dict,
state_dict: dict,
logical_expert_ids: paddle.Tensor,
ep_rank_to_expert_id_list: list,
):
"""
Get w4a8 weights from state dict and process them.
Args:
@@ -508,8 +562,15 @@ class CutlassW4A8MoEMethod(CutlassMoEMethod):
state_dict (dict): The state dict.
"""
def _extract_scale_tensor(state_dict, key_template, expert_idx):
return get_tensor(state_dict.pop(key_template.format(expert_idx)))
def _extract_scale_tensor(layer: nn.Layer, state_dict, key_template, expert_idx):
return get_tensor(
(
state_dict.pop(key_template.format(expert_idx))
if key_template.format(expert_idx) in state_dict
else key_template.format(expert_idx)
),
layer.fd_config.model_config.model,
)
def _process_in_scale(name: str, in_scales: list[paddle.Tensor]):
processed_in_scale = 1 / paddle.concat(in_scales)
@@ -551,17 +612,23 @@ class CutlassW4A8MoEMethod(CutlassMoEMethod):
# 2. Extract scale tensor from state dict
if layer.ep_size > 1:
for expert_idx in range(layer.num_experts):
scale_tensor = get_tensor(state_dict[scale_key_map["up_gate_proj_in_scale"].format(expert_idx)])
for expert_idx in ep_rank_to_expert_id_list:
scale_tensor = get_tensor(
(
state_dict[scale_key_map["up_gate_proj_in_scale"].format(expert_idx)]
if scale_key_map["up_gate_proj_in_scale"].format(expert_idx) in state_dict
else scale_key_map["up_gate_proj_in_scale"].format(expert_idx)
),
layer.fd_config.model_config.model,
)
up_gate_proj_in_scales_all_experts.append(1 / scale_tensor)
getattr(layer, "up_gate_proj_in_scale_all_experts").set_value(
paddle.concat(up_gate_proj_in_scales_all_experts)
)
for local_expert_idx in range(layer.num_local_experts):
expert_idx = local_expert_idx + layer.expert_id_offset
for expert_idx in logical_expert_ids:
for name, scale_key_template in scale_key_map.items():
scale_tensor = _extract_scale_tensor(state_dict, scale_key_template, expert_idx)
scale_tensor = _extract_scale_tensor(layer, state_dict, scale_key_template, expert_idx)
scale_weight_map[name].append(scale_tensor)
# 3. Process scale tensor and set to layer
@@ -845,7 +912,7 @@ class CutlassWeightOnlyMoEMethod(CutlassMoEMethod):
self.moe_quant_type = self.quant_config.algo
self.pack_num = 1
def process_prequanted_weights(self, layer: nn.Layer, state_dict):
def process_prequanted_weights(self, layer: nn.Layer, state_dict, is_rearrange: bool = False):
"""
Paddle cutlass process prequanted weights.
"""
@@ -855,9 +922,7 @@ class CutlassWeightOnlyMoEMethod(CutlassMoEMethod):
down_proj_expert_weight_scale_key = layer.weight_key_map.get("down_proj_expert_weight_scale_key", None)
up_gate_proj_weights, down_proj_weights, logical_expert_ids, _ = layer.load_experts_weight(
state_dict,
up_gate_proj_expert_weight_key,
down_proj_expert_weight_key,
state_dict, up_gate_proj_expert_weight_key, down_proj_expert_weight_key, is_rearrange
)
# self.check(layer, up_gate_proj_weights, down_proj_weights)
up_gate_proj_weight_scale = []
@@ -1065,7 +1130,7 @@ class CutlassWeightOnlyMoEMethod(CutlassMoEMethod):
"""
Paddle cutlass load weight process.
"""
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
self.check(layer, up_gate_proj_weights, down_proj_weights)
for idx, weight_tensor in enumerate([up_gate_proj_weights, down_proj_weights]):
weight_name = self.added_weight_attrs[idx]

8
fastdeploy/model_executor/layers/moe/fused_moe_deepgemm_backend.py
View File

@@ -99,7 +99,7 @@ class DeepGemmFusedMoeMethod(MoEMethodBase):
"""
deepgemm create weight process.
"""
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
self.check(layer, up_gate_proj_weights, down_proj_weights)
@@ -124,7 +124,7 @@ class DeepGemmFusedMoeMethod(MoEMethodBase):
quanted_weight_scale = quanted_weight_scale.transpose([0, 2, 1]).contiguous()
getattr(layer, scale_name).set_value(quanted_weight_scale)
def process_prequanted_weights(self, layer: nn.Layer, state_dict):
def process_prequanted_weights(self, layer: nn.Layer, state_dict, is_rearrange: bool = False):
"""
Paddle cutlass process prequanted weights.
"""
@@ -134,9 +134,7 @@ class DeepGemmFusedMoeMethod(MoEMethodBase):
down_proj_expert_weight_scale_key = layer.weight_key_map.get("down_proj_expert_weight_scale_key", None)
up_gate_proj_weights, down_proj_weights, logical_expert_ids, _ = layer.load_experts_weight(
state_dict,
up_gate_proj_expert_weight_key,
down_proj_expert_weight_key,
state_dict, up_gate_proj_expert_weight_key, down_proj_expert_weight_key, is_rearrange
)
# self.check(layer, up_gate_proj_weights, down_proj_weights)
up_gate_proj_weight_scale = []

2
fastdeploy/model_executor/layers/moe/fused_moe_marlin_backend.py
View File

@@ -197,7 +197,7 @@ class MarlinWeightOnlyMoEMethod(QuantMethodBase):
"""
Marlin MoE load weight process.
"""
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
assert len(up_gate_proj_weights) == layer.num_local_experts
assert len(down_proj_weights) == layer.num_local_experts
assert up_gate_proj_weights[0].shape == [

10
fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py
View File

@@ -48,7 +48,7 @@ class TritonWeightOnlyMoEMethod(QuantMethodBase):
"down_proj_weight_scale",
]
def process_prequanted_weights(self, layer: nn.Layer, state_dict) -> None:
def process_prequanted_weights(self, layer: nn.Layer, state_dict, is_rearrange: bool = False) -> None:
"""process_prequanted_weights"""
pass
@@ -112,7 +112,7 @@ class TritonWeightOnlyMoEMethod(QuantMethodBase):
"""
Triton MoE load weight process.
"""
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
assert len(up_gate_proj_weights) == layer.num_local_experts
assert len(down_proj_weights) == layer.num_local_experts
@@ -311,7 +311,7 @@ class TensorWiseFP8MoEMethod(QuantMethodBase):
"down_proj_in_scale",
]
def process_prequanted_weights(self, layer: nn.Layer, state_dict) -> None:
def process_prequanted_weights(self, layer: nn.Layer, state_dict, is_rearrange: bool = False) -> None:
"""process_prequanted_weights"""
up_gate_proj_tensor, down_proj_tensor = layer.extract_moe_ffn_weights(state_dict)
@@ -595,7 +595,7 @@ class BlockWiseFP8MoEMethod(QuantMethodBase):
"down_proj_weight_scale",
]
def process_prequanted_weights(self, layer: nn.Layer, state_dict) -> None:
def process_prequanted_weights(self, layer: nn.Layer, state_dict, is_rearrange: bool = False) -> None:
"""process_prequanted_weights"""
raise NotImplementedError
@@ -667,7 +667,7 @@ class BlockWiseFP8MoEMethod(QuantMethodBase):
"""
Triton MoE create weight process.
"""
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
self.check(layer, up_gate_proj_weights, down_proj_weights)
for idx, weight_tensor in enumerate([up_gate_proj_weights, down_proj_weights]):

2
fastdeploy/model_executor/layers/moe/fused_moe_wint2_backend.py
View File

@@ -73,7 +73,7 @@ class CutlassWint2FusedMoeMethod(Wint2MoeMethod):
"""
pass
def process_prequanted_weights(self, layer: nn.Layer, state_dict):
def process_prequanted_weights(self, layer: nn.Layer, state_dict, is_rearrange: bool = False):
"""
Paddle cutlass process prequanted weights.
"""

4
fastdeploy/model_executor/layers/moe/fused_moe_xpu_backend.py
View File

@@ -34,7 +34,7 @@ class XPUMoEMethod(UnquantizedFusedMoEMethod):
def process_loaded_weights(self, layer: nn.Layer, state_dict):
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
for weights in [up_gate_proj_weights, down_proj_weights]:
for idx, weight in enumerate(weights):
weights[idx] = weight.transpose([1, 0])
@@ -119,7 +119,7 @@ class XPUWeightOnlyMoEMethod(QuantMethodBase):
"""
Paddle cutlass create weight process.
"""
up_gate_proj_weights, down_proj_weights = layer.extract_moe_ffn_weights(state_dict)
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
assert len(up_gate_proj_weights) == layer.num_local_experts
assert len(down_proj_weights) == layer.num_local_experts
assert up_gate_proj_weights[0].shape == [

34
fastdeploy/model_executor/layers/moe/moe.py
View File

@@ -80,6 +80,7 @@ class FusedMoE(nn.Layer):
layer_idx: int = -1,
moe_tag: str = "",
gate_correction_bias=None,
redundant_table_manger: RedundantExpertManger = None,
weight_key_map: dict = {},
):
"""
@@ -147,15 +148,8 @@ class FusedMoE(nn.Layer):
self.moe_quant_type = moe_quant_config.name()
else:
self.quant_method = get_moe_method()
self.redundant_table_manger = None
self.redundant_table_manger = redundant_table_manger
if self.ep_size > 1:
if fd_config.model_config.enable_redundant_experts is True:
self.redundant_table_manger = RedundantExpertManger(
n_routed_experts=fd_config.model_config.moe_num_experts,
num_hidden_layers=fd_config.model_config.num_hidden_layers,
redundant_experts_num=fd_config.model_config.redundant_experts_num,
ep_size=self.ep_size,
)
self.quant_method.init_ep(self)
if fd_config.load_config.dynamic_load_weight:
@@ -423,6 +417,7 @@ class FusedMoE(nn.Layer):
state_dict: dict,
up_gate_proj_expert_weight_key: str,
down_proj_expert_weight_key: str,
is_rearrange: bool = False,
):
"""
Load experts weight from state_dict.
@@ -451,7 +446,12 @@ class FusedMoE(nn.Layer):
]
up_gate_proj_weights = []
down_proj_weights = []
is_ffn_merged = up_gate_proj_expert_weight_key.format(self.expert_id_offset) in state_dict
if isinstance(state_dict, list):
state_dict = dict(state_dict)
is_ffn_merged = (
up_gate_proj_expert_weight_key.format(logical_expert_ids[0] if is_rearrange else self.expert_id_offset)
in state_dict
)
if is_ffn_merged:
for expert_idx in logical_expert_ids:
down_proj_expert_weight_key_name = down_proj_expert_weight_key.format(expert_idx)
@@ -533,10 +533,12 @@ class FusedMoE(nn.Layer):
assert up_gate_proj_expert_weight_key is not None, "up_gate_proj_expert_weight_key should not be none."
assert down_proj_expert_weight_key is not None, "down_proj_expert_weight_key should not be none."
up_gate_proj_weights, down_proj_weights, logical_expert_ids, _ = self.load_experts_weight(
state_dict,
up_gate_proj_expert_weight_key,
down_proj_expert_weight_key,
up_gate_proj_weights, down_proj_weights, logical_expert_ids, ep_rank_to_expert_id_list = (
self.load_experts_weight(
state_dict,
up_gate_proj_expert_weight_key,
down_proj_expert_weight_key,
)
)
assert (
len(up_gate_proj_weights) == self.num_local_experts
@@ -545,7 +547,7 @@ class FusedMoE(nn.Layer):
len(down_proj_weights) == self.num_local_experts
), "down_proj_weights length should be equal to num_local_experts."
return up_gate_proj_weights, down_proj_weights
return up_gate_proj_weights, down_proj_weights, logical_expert_ids, ep_rank_to_expert_id_list
def extract_gate_correction_bias(self, gate_correction_bias_key, state_dict):
"""
@@ -561,7 +563,7 @@ class FusedMoE(nn.Layer):
if is_supported_moe_backend is not None and is_supported_moe_backend(self.quant_method):
if self.fd_config.model_config.is_quantized:
if getattr(self.fd_config.quant_config, "is_permuted", True):
self.quant_method.process_prequanted_weights(self, state_dict)
self.quant_method.process_prequanted_weights(self, state_dict, is_rearrange)
else:
self.quant_method.process_loaded_weights(self, state_dict)
else:
@@ -569,7 +571,7 @@ class FusedMoE(nn.Layer):
else:
if self.fd_config.model_config.is_quantized:
if getattr(self.fd_config.quant_config, "is_permuted", True):
self.quant_method.process_prequanted_weights(self, state_dict)
self.quant_method.process_prequanted_weights(self, state_dict, is_rearrange)
else:
self.quant_method.create_weights(self, state_dict)
else:

2
fastdeploy/model_executor/layers/quantization/block_wise_fp8.py
View File

@@ -108,7 +108,7 @@ class BlockWiseFP8LinearMethod(QuantMethodBase):
layer.weight.copy_(quanted_weight_tensor, False)
layer.weight_scale.set_value(weight_block_scale_tensor)
def process_prequanted_weights(self, layer, state_dict):
def process_prequanted_weights(self, layer, state_dict, is_rearrange: bool = False):
"""
process_prequanted_weights
"""

2
fastdeploy/model_executor/layers/quantization/tensor_wise_fp8.py
View File

@@ -90,7 +90,7 @@ class TensorWiseFP8LinearMethod(QuantMethodBase):
default_initializer=paddle.nn.initializer.Constant(0),
)
def process_prequanted_weights(self, layer, state_dict) -> None:
def process_prequanted_weights(self, layer, state_dict, is_rearrange: bool = False) -> None:
"""
Process pre-quantized weights before applying them to the model
Args:

2
fastdeploy/model_executor/layers/quantization/weight_only.py
View File

@@ -305,7 +305,7 @@ class GPUWeightOnlyLinearMethod(WeightOnlyLinearMethod):
) -> None:
super().__init__(quant_config)
def process_prequanted_weights(self, layer, state_dict) -> None:
def process_prequanted_weights(self, layer, state_dict, is_rearrange: bool = False) -> None:
"""
Process pre-quantized weights before applying them to the model
Args: