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

Simplify the Config code (#2770)

Browse Source 
* simplify the code

* fix vl

* delete config

* fix

* perfect code

* fix ci

* fix xpu

* fix xpu

* fix server

* resolve conflict

* fix mtp

* resolve conflict

* fix xpu

* fix xpu

* fix vl

* fix log

* fix qwen moe

* fix qwen moe

* fix qwen moe
This commit is contained in:
YuanRisheng
2025-07-14 19:50:05 +08:00
committed by GitHub
parent 2e81792d64
commit 4c7b8bc458
34 changed files with 551 additions and 911 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
fastdeploy/model_executor/models/qwen3.py
View File

@@ -23,7 +23,8 @@ from paddle import nn
from paddleformers.transformers import PretrainedModel
from paddleformers.utils.log import logger
from fastdeploy.config import FDConfig, ModelConfig
from fastdeploy.config import FDConfig
from fastdeploy.model_executor.forward_meta import ForwardMeta
from fastdeploy.model_executor.graph_optimization.decorator import \
support_graph_optimization
from fastdeploy.model_executor.layers.attention.attention import Attention
@@ -34,7 +35,6 @@ from fastdeploy.model_executor.layers.lm_head import ParallelLMHead
from fastdeploy.model_executor.layers.normalization import RMSNorm
from fastdeploy.model_executor.models.model_base import ModelForCasualLM
from fastdeploy.model_executor.models.qwen2 import Qwen2DecoderLayer, Qwen2MLP
from fastdeploy.model_executor.forward_meta import ForwardMeta
class Qwen3MLP(Qwen2MLP):
@@ -59,7 +59,7 @@ class Qwen3Attention(nn.Layer):
self.qkv_proj = QKVParallelLinear(fd_config,
prefix=f"{prefix}.qkv_proj",
with_bias=False)
nranks = fd_config.parallel_config.tensor_parallel_degree
nranks = fd_config.parallel_config.tensor_parallel_size
self.o_proj = RowParallelLinear(
fd_config,
@@ -85,7 +85,7 @@ class Qwen3Attention(nn.Layer):
prefix=f"{prefix}.k_norm",
begin_norm_axis=2)
nranks = fd_config.parallel_config.tensor_parallel_degree
nranks = fd_config.parallel_config.tensor_parallel_size
num_kv_heads_replicas = max(1, nranks // fd_config.model_config.num_key_value_heads)
self.q_size = fd_config.model_config.num_attention_heads * self.head_dim // nranks
self.kv_size = fd_config.model_config.num_key_value_heads * self.head_dim * num_kv_heads_replicas // nranks
@@ -163,21 +163,21 @@ class Qwen3Model(nn.Layer):
"""
super().__init__()
self.num_layers = fd_config.model_config.num_layers
fd_config.model_config.prefix_name = "model"
self.num_layers = fd_config.model_config.num_hidden_layers
fd_config.model_config.pretrained_config.prefix_name = "model"
self.embeddings = VocabParallelEmbedding(
fd_config=fd_config,
num_embeddings=fd_config.model_config.vocab_size,
embedding_dim=fd_config.model_config.hidden_size,
params_dtype=paddle.get_default_dtype,
prefix=(f"{fd_config.model_config.prefix_name}.embed_tokens"),
prefix=(f"{fd_config.model_config.pretrained_config.prefix_name}.embed_tokens"),
)
self.layers = nn.LayerList([
Qwen3DecoderLayer(
fd_config=fd_config,
prefix=f"{fd_config.model_config.prefix_name}.layers.{i}")
prefix=f"{fd_config.model_config.pretrained_config.prefix_name}.layers.{i}")
for i in range(self.num_layers)
])
@@ -185,7 +185,7 @@ class Qwen3Model(nn.Layer):
fd_config,
hidden_size=fd_config.model_config.hidden_size,
eps=fd_config.model_config.rms_norm_eps,
prefix=f"{fd_config.model_config.prefix_name}.norm",
prefix=f"{fd_config.model_config.pretrained_config.prefix_name}.norm",
)
def load_state_dict(self, state_dict):
@@ -307,7 +307,7 @@ class Qwen3PretrainedModel(PretrainedModel):
return None
@classmethod
def _get_tensor_parallel_mappings(cls, config: ModelConfig, is_split=True):
def _get_tensor_parallel_mappings(cls, config, is_split=True):
from paddleformers.transformers.conversion_utils import \
split_or_merge_func
@@ -358,5 +358,5 @@ class Qwen3PretrainedModel(PretrainedModel):
return final_actions
mappings = get_tensor_parallel_split_mappings(config.num_layers)
mappings = get_tensor_parallel_split_mappings(config.num_hidden_layers)
return mappings