apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 16:48:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

polish code with new pre-commit rule (#2923)

Browse Source
This commit is contained in:
Zero Rains
2025-07-19 23:19:27 +08:00
committed by GitHub
parent b8676d71a8
commit 25698d56d1
424 changed files with 14307 additions and 13518 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
fastdeploy/model_executor/models/tp_utils.py
View File

@@ -38,15 +38,14 @@ def check_tensor_parallel_prerequisites(
"""check_tensor_parallel_prerequisites"""
if fd_config.parallel_config.tensor_parallel_size > 1:
tensor_parallel_map = cls._get_tensor_parallel_mappings(
fd_config.model_config.pretrained_config, is_split=True)
fd_config.model_config.pretrained_config, is_split=True
)
if not tensor_parallel_map:
logger.error(
"filtered_quant_map should not be empty. \
parallel splitting required, but _get_tensor_parallel_mappings is not implemented."
)
filtered_tp_keys = cls._resolve_prefix_keys(
tensor_parallel_map.keys(), safetensor_keys
)
filtered_tp_keys = cls._resolve_prefix_keys(tensor_parallel_map.keys(), safetensor_keys)
for k, v in filtered_tp_keys.items():
tensor_parallel_filtered_map[v] = tensor_parallel_map.pop(k)
if not tensor_parallel_filtered_map:
@@ -176,9 +175,7 @@ def build_expanded_keys(
if start_layer < 0:
continue
for layer_id in range(start_layer, num_layers):
for export_id in range(
text_num_experts, text_num_experts + img_num_experts
):
for export_id in range(text_num_experts, text_num_experts + img_num_experts):
update_final_actions(
{
LayerIdPlaceholder.MOE_LAYER_ID.value: layer_id,
@@ -188,10 +185,7 @@ def build_expanded_keys(
key,
action,
)
elif (
LayerIdPlaceholder.MOE_LAYER_ID.value in placeholders
and len(placeholders) == 1
):
elif LayerIdPlaceholder.MOE_LAYER_ID.value in placeholders and len(placeholders) == 1:
if start_layer < 0:
continue
for layer_id in range(start_layer, num_layers):
@@ -222,8 +216,7 @@ def gqa_qkv_split_func(
def get_shape(tensor):
"""get_shape"""
return tensor.get_shape() if hasattr(tensor,
"get_shape") else tensor.shape
return tensor.get_shape() if hasattr(tensor, "get_shape") else tensor.shape
def slice_tensor(tensor, start, end):
"""slice_tensor"""
@@ -251,10 +244,7 @@ def gqa_qkv_split_func(
size = shape[-1] if is_column else shape[0]
block_size = size // degree
if hasattr(tensor, "get_shape"):
return [
slice_tensor(tensor, i * block_size, (i + 1) * block_size)
for i in range(degree)
]
return [slice_tensor(tensor, i * block_size, (i + 1) * block_size) for i in range(degree)]
else:
if isinstance(x, paddle.Tensor):
if is_column:
@@ -342,8 +332,8 @@ def gqa_qkv_merge_func(num_attention_heads, num_key_value_heads, head_dim):
def fn(weight_list, is_column=True):
"""fn"""
tensor_parallel_degree = len(weight_list)
num_attention_heads = num_attention_heads // tensor_parallel_degree # noqa: F823
num_key_value_heads = num_key_value_heads // tensor_parallel_degree # noqa: F823
local_num_attention_heads = num_attention_heads // tensor_parallel_degree
local_num_key_value_heads = num_key_value_heads // tensor_parallel_degree
is_paddle_tensor = not isinstance(weight_list[0], np.ndarray)
@@ -351,8 +341,7 @@ def gqa_qkv_merge_func(num_attention_heads, num_key_value_heads, head_dim):
"""
get_shape
"""
return tensor.get_shape() if hasattr(tensor,
"get_shape") else tensor.shape
return tensor.get_shape() if hasattr(tensor, "get_shape") else tensor.shape
def slice_tensor(tensor, start, end):
"""
@@ -368,9 +357,9 @@ def gqa_qkv_merge_func(num_attention_heads, num_key_value_heads, head_dim):
q_list, k_list, v_list = [], [], []
for weight in weight_list:
q_end = num_attention_heads * head_dim
k_end = q_end + num_key_value_heads * head_dim
v_end = k_end + num_key_value_heads * head_dim
q_end = local_num_attention_heads * head_dim
k_end = q_end + local_num_key_value_heads * head_dim
v_end = k_end + local_num_key_value_heads * head_dim
q = slice_tensor(weight, 0, q_end)
k = slice_tensor(weight, q_end, k_end)