apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-06 00:57:33 +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

153
fastdeploy/model_executor/models/qwen3.py
View File

@@ -25,12 +25,12 @@ from paddleformers.utils.log import logger
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.graph_optimization.decorator import (
support_graph_optimization,
)
from fastdeploy.model_executor.layers.attention.attention import Attention
from fastdeploy.model_executor.layers.embeddings import VocabParallelEmbedding
from fastdeploy.model_executor.layers.linear import (QKVParallelLinear,
RowParallelLinear)
from fastdeploy.model_executor.layers.linear import QKVParallelLinear, RowParallelLinear
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
@@ -38,52 +38,51 @@ from fastdeploy.model_executor.models.qwen2 import Qwen2DecoderLayer, Qwen2MLP
class Qwen3MLP(Qwen2MLP):
"""
"""
""" """
pass
class Qwen3Attention(nn.Layer):
"""
"""
""" """
def __init__(self,
fd_config: FDConfig,
layer_id: int,
prefix: str = "") -> None:
def __init__(self, fd_config: FDConfig, layer_id: int, prefix: str = "") -> None:
super().__init__()
self.fd_config = fd_config
self.head_dim = fd_config.model_config.head_dim
self.qkv_proj = QKVParallelLinear(fd_config,
prefix=f"{prefix}.qkv_proj",
with_bias=False)
self.qkv_proj = QKVParallelLinear(fd_config, prefix=f"{prefix}.qkv_proj", with_bias=False)
nranks = fd_config.parallel_config.tensor_parallel_size
self.o_proj = RowParallelLinear(
fd_config,
prefix=f"{prefix}.o_proj",
input_size=fd_config.model_config.head_dim *
fd_config.model_config.num_attention_heads,
input_size=fd_config.model_config.head_dim * fd_config.model_config.num_attention_heads,
output_size=fd_config.model_config.hidden_size,
)
self.attn = Attention(fd_config,
layer_id=layer_id,
prefix=prefix,
use_neox_rotary_style=True)
self.attn = Attention(
fd_config,
layer_id=layer_id,
prefix=prefix,
use_neox_rotary_style=True,
)
self.q_norm = RMSNorm(fd_config,
hidden_size=self.head_dim,
eps=fd_config.model_config.rms_norm_eps,
prefix=f"{prefix}.q_norm",
begin_norm_axis=2)
self.k_norm = RMSNorm(fd_config,
hidden_size=self.head_dim,
eps=fd_config.model_config.rms_norm_eps,
prefix=f"{prefix}.k_norm",
begin_norm_axis=2)
self.q_norm = RMSNorm(
fd_config,
hidden_size=self.head_dim,
eps=fd_config.model_config.rms_norm_eps,
prefix=f"{prefix}.q_norm",
begin_norm_axis=2,
)
self.k_norm = RMSNorm(
fd_config,
hidden_size=self.head_dim,
eps=fd_config.model_config.rms_norm_eps,
prefix=f"{prefix}.k_norm",
begin_norm_axis=2,
)
nranks = fd_config.parallel_config.tensor_parallel_size
num_kv_heads_replicas = max(1, nranks // fd_config.model_config.num_key_value_heads)
@@ -91,8 +90,7 @@ class Qwen3Attention(nn.Layer):
self.kv_size = fd_config.model_config.num_key_value_heads * self.head_dim * num_kv_heads_replicas // nranks
def load_state_dict(self, state_dict):
"""
"""
""" """
self.qkv_proj.load_state_dict(state_dict)
self.o_proj.load_state_dict(state_dict)
self.q_norm.load_state_dict(state_dict)
@@ -103,20 +101,16 @@ class Qwen3Attention(nn.Layer):
forward_meta: ForwardMeta,
hidden_states: paddle.Tensor,
):
"""
"""
""" """
qkv_out = self.qkv_proj(hidden_states)
# origin_qkv_out = qkv_out
q, k, v = qkv_out.split([self.q_size, self.kv_size, self.kv_size],
axis=-1)
q, k, v = qkv_out.split([self.q_size, self.kv_size, self.kv_size], axis=-1)
q_by_head = q.reshape(
[*q.shape[:-1], q.shape[-1] // self.head_dim, self.head_dim])
q_by_head = q.reshape([*q.shape[:-1], q.shape[-1] // self.head_dim, self.head_dim])
q_by_head = self.q_norm(q_by_head)
q = q_by_head.reshape(q.shape)
k_by_head = k.reshape(
[*k.shape[:-1], k.shape[-1] // self.head_dim, self.head_dim])
k_by_head = k.reshape([*k.shape[:-1], k.shape[-1] // self.head_dim, self.head_dim])
k_by_head = self.k_norm(k_by_head)
k = k_by_head.reshape(k.shape)
@@ -131,8 +125,7 @@ class Qwen3Attention(nn.Layer):
class Qwen3DecoderLayer(Qwen2DecoderLayer):
"""
"""
""" """
def __init__(
self,
@@ -140,16 +133,13 @@ class Qwen3DecoderLayer(Qwen2DecoderLayer):
prefix: str = "",
) -> None:
super().__init__(fd_config, prefix)
layer_id = int(prefix.split(sep='.')[-1])
self.self_attn = Qwen3Attention(fd_config=fd_config,
layer_id=layer_id,
prefix=f"{prefix}.self_attn")
layer_id = int(prefix.split(sep=".")[-1])
self.self_attn = Qwen3Attention(fd_config=fd_config, layer_id=layer_id, prefix=f"{prefix}.self_attn")
@support_graph_optimization
class Qwen3Model(nn.Layer):
"""
"""
""" """
def __init__(
self,
@@ -174,12 +164,15 @@ class Qwen3Model(nn.Layer):
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.pretrained_config.prefix_name}.layers.{i}")
for i in range(self.num_layers)
])
self.layers = nn.LayerList(
[
Qwen3DecoderLayer(
fd_config=fd_config,
prefix=f"{fd_config.model_config.pretrained_config.prefix_name}.layers.{i}",
)
for i in range(self.num_layers)
]
)
self.norm = RMSNorm(
fd_config,
@@ -208,15 +201,13 @@ class Qwen3Model(nn.Layer):
ids_remove_padding: paddle.Tensor,
forward_meta: ForwardMeta,
):
"""
"""
""" """
hidden_states = self.embed_tokens(ids_remove_padding=ids_remove_padding)
residual = None
for i in range(self.num_layers):
hidden_states, residual = self.layers[i](forward_meta,
hidden_states, residual)
hidden_states, residual = self.layers[i](forward_meta, hidden_states, residual)
hidden_states = hidden_states + residual
@@ -250,8 +241,7 @@ class Qwen3ForCausalLM(ModelForCasualLM):
@classmethod
def name(self):
"""
"""
""" """
return "Qwen3ForCausalLM"
@paddle.no_grad()
@@ -266,17 +256,15 @@ class Qwen3ForCausalLM(ModelForCasualLM):
"""
self.model.load_state_dict(state_dict)
if self.tie_word_embeddings:
self.lm_head.linear.weight.set_value(
self.model.embed_tokens.embeddings.weight.transpose([1, 0]))
self.lm_head.linear.weight.set_value(self.model.embed_tokens.embeddings.weight.transpose([1, 0]))
else:
self.lm_head.load_state_dict(state_dict)
def compute_logits(self, hidden_states: paddle.Tensor):
"""
"""
""" """
logits = self.lm_head(hidden_states)
logits = paddle.cast(logits, paddle.float32)
logits[:, self.ori_vocab_size:] = -float("inf")
logits[:, self.ori_vocab_size :] = -float("inf")
return logits
@@ -285,10 +273,8 @@ class Qwen3ForCausalLM(ModelForCasualLM):
ids_remove_padding: paddle.Tensor,
forward_meta: ForwardMeta,
):
"""
"""
hidden_states = self.model(ids_remove_padding=ids_remove_padding,
forward_meta=forward_meta)
""" """
hidden_states = self.model(ids_remove_padding=ids_remove_padding, forward_meta=forward_meta)
return hidden_states
@@ -309,8 +295,7 @@ class Qwen3PretrainedModel(PretrainedModel):
@classmethod
def _get_tensor_parallel_mappings(cls, config, is_split=True):
from paddleformers.transformers.conversion_utils import \
split_or_merge_func
from paddleformers.transformers.conversion_utils import split_or_merge_func
fn = split_or_merge_func(
is_split=is_split,
@@ -326,34 +311,26 @@ class Qwen3PretrainedModel(PretrainedModel):
# Row Linear
"lm_head.weight": partial(fn, is_column=True),
"embed_tokens.weight": partial(fn, is_column=False),
"layers.0.self_attn.o_proj.weight": partial(fn,
is_column=False),
"layers.0.self_attn.o_proj.weight": partial(fn, is_column=False),
"layers.0.mlp.down_proj.weight": partial(fn, is_column=False),
}
# Column Linear
base_actions["layers.0.self_attn.q_proj.weight"] = partial(
fn, is_column=True)
base_actions["layers.0.self_attn.q_proj.bias"] = partial(
fn, is_column=True)
base_actions["layers.0.self_attn.q_proj.weight"] = partial(fn, is_column=True)
base_actions["layers.0.self_attn.q_proj.bias"] = partial(fn, is_column=True)
# if we have enough num_key_value_heads to split, then split it.
if config.num_key_value_heads % config.tensor_parallel_degree == 0:
base_actions["layers.0.self_attn.k_proj.weight"] = partial(
fn, is_column=True)
base_actions["layers.0.self_attn.v_proj.weight"] = partial(
fn, is_column=True)
base_actions["layers.0.self_attn.k_proj.weight"] = partial(fn, is_column=True)
base_actions["layers.0.self_attn.v_proj.weight"] = partial(fn, is_column=True)
base_actions["layers.0.mlp.gate_proj.weight"] = partial(
fn, is_column=True)
base_actions["layers.0.mlp.up_proj.weight"] = partial(
fn, is_column=True)
base_actions["layers.0.mlp.gate_proj.weight"] = partial(fn, is_column=True)
base_actions["layers.0.mlp.up_proj.weight"] = partial(fn, is_column=True)
for key, action in base_actions.items():
if "layers.0." in key:
for i in range(num_layers):
final_actions[key.replace("layers.0.",
f"layers.{i}.")] = action
final_actions[key.replace("layers.0.", f"layers.{i}.")] = action
final_actions[key] = action
return final_actions