[NEW Model] Add mamba

[JunnYu]

PR types

New features

PR changes

Models

Description

Mamba: Linear-Time Sequence Modeling with Selective State Spaces
作者仓库
huggingface仓库

需要develop版本的paddlepaddle
如何使用：
1.（可选）编译自定义cuda算子，https://github.com/JunnYu/mamba/tree/paddle-v2.2.2
2. 运行下面的代码。

import paddle
from paddlenlp.transformers import MambaForCausalLM, MambaTokenizer
from paddlenlp.generation import GenerationConfig

name = "state-spaces/mamba-2.8b-hf"

tokenizer = MambaTokenizer.from_pretrained(name)
model = MambaForCausalLM.from_pretrained(name, dtype="float16", low_cpu_mem_usage=True)
model.eval()

generation_config = GenerationConfig(
    decode_strategy="sampling",
    top_k=50,
    num_return_sequences=4,
    use_cache=True,
)

with paddle.no_grad():
    prompt = "Hello, it is"
    inputs = tokenizer.encode(prompt, return_tensors="pd")
    outputs = model.generate(**inputs, generation_config=generation_config, max_length=256)
    for e in tokenizer.batch_decode(outputs[0], skip_special_tokens=True):
        print(prompt + e)
        print('-'*100)

[paddle-bot]

Thanks for your contribution!

[JunnYu]

更新

[codecov]

Codecov Report

Attention: Patch coverage is 85.42056% with 78 lines in your changes missing coverage. Please review.

Project coverage is 55.00%. Comparing base (e0d2809) to head (1244675).
Report is 225 commits behind head on develop.

Files with missing lines	Patch %	Lines
paddlenlp/transformers/mamba/modeling.py	79.87%	65 Missing ⚠️
paddlenlp/transformers/mamba/tokenizer.py	92.35%	13 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #8513      +/-   ##
===========================================
+ Coverage    54.79%   55.00%   +0.20%     
===========================================
  Files          636      640       +4     
  Lines        99876   100543     +667     
===========================================
+ Hits         54732    55307     +575     
- Misses       45144    45236      +92     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.