mirror of
https://github.com/mudler/LocalAI.git
synced 2024-12-18 20:27:57 +00:00
feat: Openvino runtime for transformer backend and streaming support for Openvino and CUDA (#1892)
* fixes #1775 and #1774 Add BitsAndBytes Quantization and fixes embedding on CUDA devices * Manage 4bit and 8 bit quantization Manage different BitsAndBytes options with the quantization: parameter in yaml * fix compilation errors on non CUDA environment * OpenVINO draft First draft of OpenVINO integration in transformer backend * first working implementation * Streaming working * Small fix for regression on CUDA and XPU * use pip version of optimum[openvino] * Update backend/python/transformers/transformers_server.py Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com> --------- Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com> Co-authored-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
This commit is contained in:
fakezeta
GitHub
parent
b500ceaf73
commit
e7cbe32601
@ -34,6 +34,7 @@ dependencies:
|
||||
- boto3==1.28.61
|
||||
- botocore==1.31.61
|
||||
- certifi==2023.7.22
|
||||
- coloredlogs==15.0.1
|
||||
- TTS==0.22.0
|
||||
- charset-normalizer==3.3.0
|
||||
- datasets==2.14.5
|
||||
@ -48,6 +49,7 @@ dependencies:
|
||||
- funcy==2.0
|
||||
- grpcio==1.59.0
|
||||
- huggingface-hub
|
||||
- humanfriendly==10.0
|
||||
- idna==3.4
|
||||
- jinja2==3.1.2
|
||||
- jmespath==1.0.1
|
||||
@ -57,7 +59,10 @@ dependencies:
|
||||
- multiprocess==0.70.15
|
||||
- networkx
|
||||
- numpy==1.26.0
|
||||
- optimum==1.17.1
|
||||
- onnx==1.15.0
|
||||
- openvino==2024.0.0
|
||||
- openvino-telemetry==2023.2.1
|
||||
- optimum[openvino]==1.17.1
|
||||
- packaging==23.2
|
||||
- pandas
|
||||
- peft==0.5.0
|
||||
|
||||
@ -8,6 +8,7 @@ import argparse
|
||||
import signal
|
||||
import sys
|
||||
import os
|
||||
from threading import Thread
|
||||
|
||||
import time
|
||||
import backend_pb2
|
||||
@ -17,13 +18,16 @@ import grpc
|
||||
import torch
|
||||
import torch.cuda
|
||||
|
||||
|
||||
XPU=os.environ.get("XPU", "0") == "1"
|
||||
if XPU:
|
||||
import intel_extension_for_pytorch as ipex
|
||||
from intel_extension_for_transformers.transformers.modeling import AutoModelForCausalLM
|
||||
from transformers import AutoTokenizer, AutoModel, set_seed
|
||||
from transformers import AutoTokenizer, AutoModel, set_seed, TextIteratorStreamer
|
||||
from optimum.intel.openvino import OVModelForCausalLM
|
||||
from openvino.runtime import Core
|
||||
else:
|
||||
from transformers import AutoTokenizer, AutoModel, AutoModelForCausalLM, set_seed, BitsAndBytesConfig
|
||||
from transformers import AutoTokenizer, AutoModel, AutoModelForCausalLM, set_seed, BitsAndBytesConfig, TextIteratorStreamer
|
||||
|
||||
|
||||
_ONE_DAY_IN_SECONDS = 60 * 60 * 24
|
||||
@ -81,6 +85,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
|
||||
compute=torch.bfloat16
|
||||
|
||||
self.CUDA = request.CUDA
|
||||
self.OV=False
|
||||
|
||||
device_map="cpu"
|
||||
|
||||
@ -105,23 +110,55 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
|
||||
bnb_4bit_compute_dtype = None,
|
||||
load_in_8bit=True,
|
||||
)
|
||||
|
||||
|
||||
|
||||
try:
|
||||
if request.Type == "AutoModelForCausalLM":
|
||||
if XPU:
|
||||
if quantization == "xpu_4bit":
|
||||
device_map="xpu"
|
||||
compute=torch.float16
|
||||
if request.Quantization == "xpu_4bit":
|
||||
xpu_4bit = True
|
||||
self.model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=request.TrustRemoteCode,
|
||||
device_map="xpu", load_in_4bit=xpu_4bit)
|
||||
xpu_8bit = False
|
||||
elif request.Quantization == "xpu_8bit":
|
||||
xpu_4bit = False
|
||||
xpu_8bit = True
|
||||
else:
|
||||
xpu_4bit = False
|
||||
xpu_8bit = False
|
||||
self.model = AutoModelForCausalLM.from_pretrained(model_name,
|
||||
trust_remote_code=request.TrustRemoteCode,
|
||||
use_safetensors=True,
|
||||
device_map=device_map,
|
||||
load_in_4bit=xpu_4bit,
|
||||
load_in_8bit=xpu_8bit,
|
||||
torch_dtype=compute)
|
||||
else:
|
||||
self.model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=request.TrustRemoteCode, use_safetensors=True, quantization_config=quantization, device_map=device_map, torch_dtype=compute)
|
||||
self.model = AutoModelForCausalLM.from_pretrained(model_name,
|
||||
trust_remote_code=request.TrustRemoteCode,
|
||||
use_safetensors=True,
|
||||
quantization_config=quantization,
|
||||
device_map=device_map,
|
||||
torch_dtype=compute)
|
||||
elif request.Type == "OVModelForCausalLM":
|
||||
if "GPU" in Core().available_devices:
|
||||
device_map="GPU"
|
||||
else:
|
||||
device_map="CPU"
|
||||
self.model = OVModelForCausalLM.from_pretrained(model_name,
|
||||
compile=True,
|
||||
device=device_map)
|
||||
self.OV = True
|
||||
else:
|
||||
self.model = AutoModel.from_pretrained(model_name, trust_remote_code=request.TrustRemoteCode, use_safetensors=True, quantization_config=quantization, device_map=device_map, torch_dtype=compute)
|
||||
self.model = AutoModel.from_pretrained(model_name,
|
||||
trust_remote_code=request.TrustRemoteCode,
|
||||
use_safetensors=True,
|
||||
quantization_config=quantization,
|
||||
device_map=device_map,
|
||||
torch_dtype=compute)
|
||||
self.tokenizer = AutoTokenizer.from_pretrained(model_name, use_safetensors=True)
|
||||
self.XPU = False
|
||||
|
||||
if XPU:
|
||||
if XPU and self.OV == False:
|
||||
self.XPU = True
|
||||
try:
|
||||
print("Optimizing model", model_name, "to XPU.", file=sys.stderr)
|
||||
@ -130,6 +167,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
|
||||
print("Not using XPU:", err, file=sys.stderr)
|
||||
|
||||
except Exception as err:
|
||||
print("Error:", err, file=sys.stderr)
|
||||
return backend_pb2.Result(success=False, message=f"Unexpected {err=}, {type(err)=}")
|
||||
# Implement your logic here for the LoadModel service
|
||||
# Replace this with your desired response
|
||||
@ -167,7 +205,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
|
||||
print("Embeddings:", sentence_embeddings, file=sys.stderr)
|
||||
return backend_pb2.EmbeddingResult(embeddings=sentence_embeddings[0])
|
||||
|
||||
def Predict(self, request, context):
|
||||
def Predict(self, request, context, streaming=False):
|
||||
"""
|
||||
Generates text based on the given prompt and sampling parameters.
|
||||
|
||||
@ -186,15 +224,42 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
|
||||
if request.Tokens > 0:
|
||||
max_tokens = request.Tokens
|
||||
|
||||
inputs = self.tokenizer(request.Prompt, return_tensors="pt").input_ids
|
||||
inputs = self.tokenizer(request.Prompt, return_tensors="pt")
|
||||
if self.CUDA:
|
||||
inputs = inputs.to("cuda")
|
||||
if XPU:
|
||||
if XPU and self.OV == False:
|
||||
inputs = inputs.to("xpu")
|
||||
streaming = False
|
||||
|
||||
outputs = self.model.generate(inputs,max_new_tokens=max_tokens, temperature=request.Temperature, top_p=request.TopP, do_sample=True, pad_token_id=self.tokenizer.eos_token_id)
|
||||
generated_text = self.tokenizer.batch_decode(outputs[:, inputs.shape[1]:], skip_special_tokens=True)[0]
|
||||
|
||||
if streaming:
|
||||
streamer=TextIteratorStreamer(self.tokenizer,
|
||||
skip_prompt=True,
|
||||
skip_special_tokens=True)
|
||||
config=dict(inputs,
|
||||
max_new_tokens=max_tokens,
|
||||
temperature=request.Temperature,
|
||||
top_p=request.TopP,
|
||||
top_k=request.TopK,
|
||||
do_sample=True,
|
||||
attention_mask=inputs["attention_mask"],
|
||||
eos_token_id=self.tokenizer.eos_token_id,
|
||||
pad_token_id=self.tokenizer.eos_token_id,
|
||||
streamer=streamer)
|
||||
thread=Thread(target=self.model.generate, kwargs=config)
|
||||
thread.start()
|
||||
generated_text = ""
|
||||
for new_text in streamer:
|
||||
generated_text += new_text
|
||||
yield backend_pb2.Reply(message=bytes(new_text, encoding='utf-8'))
|
||||
else:
|
||||
outputs = self.model.generate(inputs["input_ids"],
|
||||
max_new_tokens=max_tokens,
|
||||
temperature=request.Temperature,
|
||||
top_p=request.TopP,
|
||||
top_k=request.TopK,
|
||||
do_sample=True,
|
||||
pad_token=self.tokenizer.eos_token_id)
|
||||
generated_text = self.tokenizer.batch_decode(outputs[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0]
|
||||
return backend_pb2.Reply(message=bytes(generated_text, encoding='utf-8'))
|
||||
|
||||
def PredictStream(self, request, context):
|
||||
@ -208,7 +273,9 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
|
||||
Returns:
|
||||
backend_pb2.Result: The predict stream result.
|
||||
"""
|
||||
yield self.Predict(request, context)
|
||||
iterations = self.Predict(request, context, streaming=True)
|
||||
for iteration in iterations:
|
||||
yield iteration
|
||||
|
||||
|
||||
def serve(address):
|
||||
|
||||
Loading…
Reference in New Issue
Block a user