[BUG] Incorrect Model Outputs When Using Beam Search

Describe the bug
When I use kernel injection I get worse generation results than when using transformers without DeepSpeed.
I don't know if the results should be the same, but they are not only not the same, but even worse.

I saw this two issues marked as closed:
https://github.com/microsoft/DeepSpeed/issues/2048
https://github.com/microsoft/DeepSpeed/issues/2230

But I use version after this fix(https://github.com/microsoft/DeepSpeed/pull/2489) and still have a problem

To Reproduce

Huggingface:

import random
import os
import numpy as np
import torch

def set_random_seed(seed):
    random.seed(seed)
    np.random.seed(seed)
    os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:2"
    os.environ["PL_GLOBAL_SEED"] = str(seed)
    os.environ["PYTHONHASHSEED"] = str(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.benchmark = False
    torch.backends.cudnn.deterministic = True

set_random_seed(42)

params = {
 'num_beams': 2,
 'do_sample': False,
 'max_new_tokens': 65,
 'use_cache': True,
 'no_repeat_ngram_size': 5,
 'num_return_sequences': 1}

from transformers import AutoTokenizer, AutoModelForCausalLM

DEVICE = torch.device("cuda:0")
name = "EleutherAI/gpt-j-6B"
model = AutoModelForCausalLM.from_pretrained(name).to(DEVICE).eval().half()
tokenizer = AutoTokenizer.from_pretrained(name)

prompt = "Quantum computers are"

inputs = tokenizer(prompt, return_tensors="pt").to(DEVICE)
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        **params
    )

print(prompt)
print()
print(tokenizer.decode(outputs[0])[len(prompt):].strip())

output:

Quantum computers are

the holy grail of computing. They promise to solve problems that are intractable on today’s supercomputers, and they could be the key to solving some of the world’s most pressing problems.

But they’re not quite there yet.

Quantum computers are still in

DeepSpeed

import random
import os
import numpy as np
import torch

def set_random_seed(seed):
    random.seed(seed)
    np.random.seed(seed)
    os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:2"
    os.environ["PL_GLOBAL_SEED"] = str(seed)
    os.environ["PYTHONHASHSEED"] = str(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.benchmark = False
    torch.backends.cudnn.deterministic = True

set_random_seed(42)

params = {
 'num_beams': 2,
 'do_sample': False,
 'max_new_tokens': 65,
 'use_cache': True,
 'no_repeat_ngram_size': 5,
 'num_return_sequences': 1}

from transformers import AutoTokenizer, AutoModelForCausalLM
import deepspeed

DEVICE = torch.device("cuda:0")
name = "EleutherAI/gpt-j-6B"
model = AutoModelForCausalLM.from_pretrained(name).to(DEVICE).eval().half()
tokenizer = AutoTokenizer.from_pretrained(name)

model = deepspeed.init_inference(
    model=model,     
    mp_size=1,       
    dtype=torch.float16, 
    replace_method="auto", 
    replace_with_kernel_inject=True,
)

prompt = "Quantum computers are"

inputs = tokenizer(prompt, return_tensors="pt").to(DEVICE)
with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        **params
    )

print(prompt)
print()
print(tokenizer.decode(outputs[0])[len(prompt):].strip())

output:

Quantum computers are

the holy grail of modern science. Physics World War II, but they’s.


Quantum computers are a holy grail of quantum computers are a holy

of modern.
quantum-computers are a holy gra
of.

and

are.

.

ds_report output

--------------------------------------------------
DeepSpeed C++/CUDA extension op report
--------------------------------------------------
NOTE: Ops not installed will be just-in-time (JIT) compiled at
      runtime if needed. Op compatibility means that your system
      meet the required dependencies to JIT install the op.
--------------------------------------------------
JIT compiled ops requires ninja
ninja .................. [OKAY]
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
fused_adam ............. [NO] ....... [OKAY]
fused_lamb ............. [NO] ....... [OKAY]
 [WARNING]  please install triton==1.0.0 if you want to use sparse attention
sparse_attn ............ [NO] ....... [NO]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
 [WARNING]  async_io requires the dev libaio .so object and headers but these were not found.
 [WARNING]  async_io: please install the libaio-dev package with apt
 [WARNING]  If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
async_io ............... [NO] ....... [NO]
utils .................. [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
transformer_inference .. [NO] ....... [OKAY]
spatial_inference ...... [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/home/mikhail/venv/lib/python3.7/site-packages/torch']
torch version .................... 1.13.0+cu116
torch cuda version ............... 11.6
torch hip version ................ None
nvcc version ..................... 11.3
deepspeed install path ........... ['/home/mikhail/venv/lib/python3.7/site-packages/deepspeed']
deepspeed info ................... 0.7.6+a4ceabb6, a4ceabb6, master
deepspeed wheel compiled w. ...... torch 1.13, cuda 11.6

System info (please complete the following information):

• OS: Debian GNU/Linux 10 (buster) (GNU/Linux 4.19.0-21-cloud-amd64 x86_64\n)
• CUDA Version: 11.6
• x1 A100 40Gb
• DeepSpeed 0.7.6+a4ceabb6
• Hugging Face Transformers 4.18.0
• Python 3.7