apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-04 08:16:42 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

fix bf16 and add comments (#4106)

Browse Source
This commit is contained in:
bukejiyu
2025-09-15 17:23:07 +08:00
committed by GitHub
parent 69aa2781a1
commit 113e330030
4 changed files with 8 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
fastdeploy/model_executor/layers/linear.py
View File

@@ -498,6 +498,7 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
if weight_need_transpose: if weight_need_transpose:
loaded_weight = get_tensor(loaded_weight) loaded_weight = get_tensor(loaded_weight)
loaded_weight = loaded_weight.transpose([1, 0]) loaded_weight = loaded_weight.transpose([1, 0])
# Avoid redundant transpose of fused weights when weight_loader is called iteratively
param.weight_need_transpose = False param.weight_need_transpose = False
# Loaded weight is already fused on disk. # Loaded weight is already fused on disk.
shard_offsets = [ shard_offsets = [
@@ -638,6 +639,7 @@ class QKVParallelLinear(ColumnParallelLinear):
if weight_need_transpose: if weight_need_transpose:
loaded_weight = get_tensor(loaded_weight) loaded_weight = get_tensor(loaded_weight)
loaded_weight = loaded_weight.transpose([1, 0]) loaded_weight = loaded_weight.transpose([1, 0])
# Avoid redundant transpose of fused weights when weight_loader is called iteratively
param.weight_need_transpose = False param.weight_need_transpose = False
# Loaded weight is already fused on disk # Loaded weight is already fused on disk
shard_offsets = [ shard_offsets = [

4
fastdeploy/model_executor/layers/quantization/__init__.py
View File

@@ -16,6 +16,8 @@ quantization module
""" """
from typing import Dict, List, Type from typing import Dict, List, Type
from fastdeploy.utils import parse_quantization
from .quant_base import QuantConfigBase from .quant_base import QuantConfigBase
QUANTIZATION_METHODS: List[str] = [ QUANTIZATION_METHODS: List[str] = [
@@ -35,6 +37,8 @@ QUANTIZATION_METHODS: List[str] = [
def parse_quant_config(args, model_config, is_ernie, is_v1_loader): def parse_quant_config(args, model_config, is_ernie, is_v1_loader):
if args.quantization is not None and isinstance(args.quantization, str):
args.quantization = parse_quantization(args.quantization)
# 1.model_config.is_quantized # 1.model_config.is_quantized
# TODO(bukejiyu) model_config.is_quantized is v0 only need to be removed in future # TODO(bukejiyu) model_config.is_quantized is v0 only need to be removed in future
if model_config.model_format == "torch": if model_config.model_format == "torch":

2
fastdeploy/model_executor/utils.py
View File

@@ -240,7 +240,7 @@ def rename_offline_ckpt_suffix_to_fd_suffix(
} }
moe_quant_type = "" moe_quant_type = ""
dense_quant_type = "" dense_quant_type = ""
if fd_config.quant_config is None: if fd_config.quant_config is not None:
if fd_config.quant_config.name() == "mix_quant": if fd_config.quant_config.name() == "mix_quant":
moe_quant_type = fd_config.quant_config.moe_quant_type moe_quant_type = fd_config.quant_config.moe_quant_type
dense_quant_type = fd_config.quant_config.dense_quant_type dense_quant_type = fd_config.quant_config.dense_quant_type

4
fastdeploy/worker/worker_process.py
View File

@@ -44,7 +44,7 @@ from fastdeploy.inter_communicator import EngineWorkerQueue as TaskQueue
from fastdeploy.inter_communicator import IPCSignal from fastdeploy.inter_communicator import IPCSignal
from fastdeploy.model_executor.layers.quantization import parse_quant_config from fastdeploy.model_executor.layers.quantization import parse_quant_config
from fastdeploy.platforms import current_platform from fastdeploy.platforms import current_platform
from fastdeploy.utils import get_logger, parse_quantization from fastdeploy.utils import get_logger
from fastdeploy.worker.worker_base import WorkerBase from fastdeploy.worker.worker_base import WorkerBase
logger = get_logger("worker_process", "worker_process.log") logger = get_logger("worker_process", "worker_process.log")
@@ -655,8 +655,6 @@ def initialize_fd_config(args, ranks: int = 1, local_rank: int = 0) -> FDConfig:
FDConfig: Initialized FastDeploy configuration object FDConfig: Initialized FastDeploy configuration object
""" """
# RL rollout # RL rollout
if args.quantization is not None and isinstance(args.quantization, str):
args.quantization = parse_quantization(args.quantization)
paddle.set_default_dtype(args.dtype) paddle.set_default_dtype(args.dtype)
model_config = ModelConfig(vars(args)) model_config = ModelConfig(vars(args))
device_config = DeviceConfig(vars(args)) device_config = DeviceConfig(vars(args))