apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-06 17:17:14 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

【Sync develop】support vl model name_mapping and ori_vocab_size (#2915)

Browse Source 
* support vl ori_vacab_size

* support trainer_degree in name_mapping

* fix
This commit is contained in:
gaoziyuan
2025-07-21 14:14:15 +08:00
committed by GitHub
parent f941124402
commit 4ead15822c
6 changed files with 167 additions and 134 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
fastdeploy/worker/worker_process.py
View File

@@ -25,7 +25,8 @@ import paddle.distributed.fleet as fleet
from fastdeploy.config import (DecodingConfig, DeviceConfig, FDConfig,
GraphOptimizationConfig, LoadConfig,
ModelConfig, ParallelConfig, SpeculativeConfig)
ModelConfig, ParallelConfig, SpeculativeConfig,
ErnieArchitectures)
from fastdeploy.input.ernie_tokenizer import ErnieBotTokenizer
from fastdeploy.inter_communicator import EngineWorkerQueue as TaskQueue
from fastdeploy.inter_communicator import IPCSignal
@@ -641,9 +642,7 @@ def initialize_fd_config(args, ranks: int = 1, local_rank: int = 0) -> FDConfig:
quant_config_name = args.quantization
quantization_config["quantization"] = quant_config_name
# Special handling for Ernie models
is_ernie = "Ernie4_5_ForCausalLM" in model_config.architectures or \
"Ernie4_5_MoeForCausalLM" in model_config.architectures or \
"Ernie4_5_VLMoeForConditionalGeneration" in model_config.architectures
is_ernie = ErnieArchitectures.contains_ernie_arch(model_config.architectures)
if quant_config_name == "wint4" and is_ernie:
quantization_config["dense_quant_type"] = "wint8"
quantization_config["moe_quant_type"] = "wint4"