apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-04 16:22:57 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[Bug fix] Fixed the garbled text issues in Qwen3-8B (#2737)

Browse Source 
* fix qwen3.py

* update

* update lm_head tie_word_embeddings

* update tie_word_embeddings

* fix

* fix tie_word_embedding not in config.json

---------

Co-authored-by: lizexu <lizexu@baidu.com>
This commit is contained in:
lizexu123
2025-07-08 14:15:27 +08:00
committed by GitHub
parent d0f4d6ba3a
commit 525be243e7
2 changed files with 10 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
fastdeploy/model_executor/models/qwen3.py
View File

@@ -164,7 +164,6 @@ class Qwen3Model(nn.Layer):
self.num_layers = fd_config.model_config.num_layers
fd_config.model_config.prefix_name = "model"
fd_config.model_config.tie_word_embeddings = True
self.embeddings = VocabParallelEmbedding(
fd_config=fd_config,
@@ -240,14 +239,13 @@ class Qwen3ForCausalLM(ModelForCasualLM):
self.model = Qwen3Model(fd_config=fd_config)
self.ori_vocab_size = fd_config.model_config.ori_vocab_size
self.tie_word_embeddings = fd_config.model_config.tie_word_embeddings
self.lm_head = ParallelLMHead(
fd_config=fd_config,
embedding_dim=fd_config.model_config.hidden_size,
num_embeddings=fd_config.model_config.vocab_size,
prefix=(f"{fd_config.model_config.prefix_name}.embed_tokens"),
prefix="lm_head",
)
self.tie_word_embeddings = fd_config.model_config.tie_word_embeddings
@classmethod
def name(self):
@@ -269,6 +267,7 @@ class Qwen3ForCausalLM(ModelForCasualLM):
if self.tie_word_embeddings:
self.lm_head.out_linear.weight.set_value(
self.model.embeddings.word_embeddings.weight.transpose([1, 0]))
else:
self.lm_head.load_state_dict(state_dict)
def compute_logits(self, hidden_states: paddle.Tensor):
@@ -324,6 +323,7 @@ class Qwen3PretrainedModel(PretrainedModel):
base_actions = {
# 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),

5
fastdeploy/worker/worker_process.py
View File

@@ -577,12 +577,15 @@ def initialize_fd_config(config_or_args) -> FDConfig:
# Get model config from model directory
model_config_dict, _ = ModelConfig.get_config_dict(config_or_args.model_name_or_path)
# Handle MoE related configs
if 'num_experts' in model_config_dict:
model_config_dict['moe_num_experts'] = model_config_dict.pop('num_experts')
if 'num_experts_per_tok' in model_config_dict:
model_config_dict['moe_topk'] = model_config_dict.pop('num_experts_per_tok')
# Set default values for model config
model_config_dict["head_dim"] = model_config_dict.get(
"head_dim", model_config_dict["hidden_size"] // model_config_dict["num_attention_heads"])
@@ -592,6 +595,8 @@ def initialize_fd_config(config_or_args) -> FDConfig:
model_config = ModelConfig.from_dict(model_config_dict)
model_config.head_dim = model_config_dict["head_dim"]
paddle.set_default_dtype(config_or_args.dtype)
if 'tie_word_embeddings' in model_config_dict:
model_config_dict['tie_word_embeddings'] = model_config_dict.pop('tie_word_embeddings')
# Initialize all config components
device_config = DeviceConfig()