update for bge-m3 (#13138)
This commit is contained in:
parent
e88a2aa65b
commit
f2598b119e
1 changed files with 2 additions and 1 deletions
|
|
@ -65,13 +65,14 @@ def _model_sample_convert():
|
||||||
def _ipex_llm_convert(load_in_low_bit):
|
def _ipex_llm_convert(load_in_low_bit):
|
||||||
# import pdb
|
# import pdb
|
||||||
# pdb.set_trace()
|
# pdb.set_trace()
|
||||||
from vllm.worker.xpu_model_runner import XPUModelRunner
|
from vllm.worker.xpu_model_runner import XPUModelRunner, XPUModelRunnerBase
|
||||||
from ipex_llm.vllm.xpu.ipex_llm_wrapper import get_ipex_llm_wrapper
|
from ipex_llm.vllm.xpu.ipex_llm_wrapper import get_ipex_llm_wrapper
|
||||||
from ipex_llm.vllm.xpu.ipex_llm_v1_wrapper import get_ipex_llm_v1_wrapper
|
from ipex_llm.vllm.xpu.ipex_llm_v1_wrapper import get_ipex_llm_v1_wrapper
|
||||||
import vllm.executor.ray_utils as ray_utils_v0
|
import vllm.executor.ray_utils as ray_utils_v0
|
||||||
import vllm.v1.executor.ray_utils as ray_utils_v1
|
import vllm.v1.executor.ray_utils as ray_utils_v1
|
||||||
from vllm.v1.worker.gpu_model_runner import GPUModelRunner
|
from vllm.v1.worker.gpu_model_runner import GPUModelRunner
|
||||||
setattr(XPUModelRunner, "load_model", get_load_function(load_in_low_bit))
|
setattr(XPUModelRunner, "load_model", get_load_function(load_in_low_bit))
|
||||||
|
setattr(XPUModelRunnerBase, "load_model", get_load_function(load_in_low_bit))
|
||||||
setattr(GPUModelRunner, "load_model", get_load_function(load_in_low_bit))
|
setattr(GPUModelRunner, "load_model", get_load_function(load_in_low_bit))
|
||||||
setattr(ray_utils_v0, "RayWorkerWrapper", get_ipex_llm_wrapper(load_in_low_bit))
|
setattr(ray_utils_v0, "RayWorkerWrapper", get_ipex_llm_wrapper(load_in_low_bit))
|
||||||
setattr(ray_utils_v1, "RayWorkerWrapper", get_ipex_llm_v1_wrapper(load_in_low_bit))
|
setattr(ray_utils_v1, "RayWorkerWrapper", get_ipex_llm_v1_wrapper(load_in_low_bit))
|
||||||
|
|
|
||||||
Loading…
Reference in a new issue