mirror of
https://github.com/PaddlePaddle/FastDeploy.git
synced 2025-10-04 08:16:42 +08:00
qwen loader (#3057)
This commit is contained in:
@@ -41,7 +41,7 @@ from fastdeploy.model_executor.layers.attention.base_attention_backend import (
|
||||
from fastdeploy.model_executor.layers.rotary_embedding import get_rope
|
||||
from fastdeploy.model_executor.layers.sample.meta_data import SamplingMetadata
|
||||
from fastdeploy.model_executor.layers.sample.sampler import Sampler, SpeculativeSampler
|
||||
from fastdeploy.model_executor.model_loader import get_model_from_loader
|
||||
from fastdeploy.model_executor.model_loader import get_model_loader
|
||||
from fastdeploy.model_executor.ops.gcu import set_value_by_flags_and_idx
|
||||
from fastdeploy.model_executor.pre_and_post_process import (
|
||||
post_process,
|
||||
@@ -547,10 +547,9 @@ class GCUModelRunner(ModelRunnerBase):
|
||||
|
||||
def load_model(self) -> None:
|
||||
"""load or download model"""
|
||||
logger.info(f"Starting to load model {self.model_config.architectures[0]}")
|
||||
time_before_load = time.perf_counter()
|
||||
# 1. Load original model
|
||||
self.model = get_model_from_loader(fd_config=self.fd_config)
|
||||
model_loader = get_model_loader(load_config=self.fd_config.load_config)
|
||||
self.model = model_loader.load_model(fd_config=self.fd_config)
|
||||
# 1.1 Load RL dynamic model
|
||||
if self.fd_config.load_config.dynamic_load_weight:
|
||||
from fastdeploy.rl.dynamic_weight_manager import DynamicWeightManager
|
||||
@@ -561,9 +560,6 @@ class GCUModelRunner(ModelRunnerBase):
|
||||
|
||||
# 3. Load drafter model(for speculative decoding)
|
||||
|
||||
time_after_load = time.perf_counter()
|
||||
logger.info(f"Model loading took {time_after_load - time_before_load} seconds")
|
||||
|
||||
# 4. Init proposer for speculative method
|
||||
self._init_speculative_proposer()
|
||||
|
||||
|
Reference in New Issue
Block a user