Issue with integrating with lm-eval harness

[sriyachakravarthy]

Hi! I tried evaluating 1bitLLM/bitnet_b1_58-3B from hugging face. i am getting the error ValueError: Tokenizer class BitnetTokenizer does not exist or is not currently imported.
Kindly help!

### Tasks
- [ ] Pull request to the 1bitLLM Repositry

[LeiWang1999]

Hi @sriyachakravarthy , Would you mind provide the scripts to reproduce?

[LeiWang1999]

you may also want to checkout https://github.com/LeiWang1999/vllm-bitblas/tree/bitblas-intg with

with VllmRunner(
    "BitBLASModel/open_llama_3b_1.58bits_bitblas",
    dtype="half",
    quantization="bitblas",
    enforce_eager=False,
) as bitnet_model:
    prompt = ""
    for i in range(0, in_seq_len):
        prompt += "a "
 
    prompts = [prompt] * batch_size
    from vllm import SamplingParams
 
    sampling_params = SamplingParams(max_tokens=out_seq_len)
    torch.cuda.profiler.start()
    bitbnet_outputs = bitnet_model.generate(
        prompts, sampling_params=sampling_params
    )
    torch.cuda.profiler.stop()

Which is much more faster than the naive integration implementation.

[sriyachakravarthy]

Hi @sriyachakravarthy , Would you mind provide the scripts to reproduce?

Sure, here is the script.

%pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@big-refactor

!lm_eval --model hf
--model_args pretrained=BitBLASModel/open_llama_3b_1.58bits_bitblas
--tasks hellaswag
--device cuda:0
--batch_size 8

and when i am trying to use instructions from model card(https://huggingface.co/1bitLLM/bitnet_b1_58-3B), i am getting the following:
$ python3 eval_ppl.py --hf_path 1bitLLM/bitnet_b1_58-3B --seqlen 2048
Traceback (most recent call last):
File "eval_ppl.py", line 7, in
from modeling_bitnet import BitnetForCausalLM
File "/home/sriyar/bitnet/modeling_bitnet.py", line 51, in
from .configuration_bitnet import BitnetConfig
ImportError: attempted relative import with no known parent package
$

[sriyachakravarthy]

you may also want to checkout https://github.com/LeiWang1999/vllm-bitblas/tree/bitblas-intg with

with VllmRunner(
    "BitBLASModel/open_llama_3b_1.58bits_bitblas",
    dtype="half",
    quantization="bitblas",
    enforce_eager=False,
) as bitnet_model:
    prompt = ""
    for i in range(0, in_seq_len):
        prompt += "a "
 
    prompts = [prompt] * batch_size
    from vllm import SamplingParams
 
    sampling_params = SamplingParams(max_tokens=out_seq_len)
    torch.cuda.profiler.start()
    bitbnet_outputs = bitnet_model.generate(
        prompts, sampling_params=sampling_params
    )
    torch.cuda.profiler.stop()

Which is much more faster than the naive integration implementation.

Sure, will do

[LeiWang1999]

Hi @sriyachakravarthy , Would you mind provide the scripts to reproduce?

Sure, here is the script.

%pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@big-refactor

!lm_eval --model hf --model_args pretrained=BitBLASModel/open_llama_3b_1.58bits_bitblas --tasks hellaswag --device cuda:0 --batch_size 8

and when i am trying to use instructions from model card(https://huggingface.co/1bitLLM/bitnet_b1_58-3B), i am getting the following: $ python3 eval_ppl.py --hf_path 1bitLLM/bitnet_b1_58-3B --seqlen 2048 Traceback (most recent call last): File "eval_ppl.py", line 7, in from modeling_bitnet import BitnetForCausalLM File "/home/sriyar/bitnet/modeling_bitnet.py", line 51, in from .configuration_bitnet import BitnetConfig ImportError: attempted relative import with no known parent package $

The code there was not provided by bitblas, checkout the integration under https://github.com/microsoft/BitBLAS/tree/main/integration/BitNet .

btw, some benchmark numbers of 1.58bits vllm

<style> </style>

		Token Per Second(tok/s)
model	framework	BS16IN32OUT128	BS1IN512OUT1024
openllama-3b-1.58bits	pytorch	106.83	49.34
openllama-3b-1.58bits	pytorch-bitblas	240.33	103.09
openllama-3b-1.58bits	vllm-bitblas	379.25	117.43
openllama-3b-1.58bits	vllm-bitblas-cuda-graph	2543.58	1621.08

[sriyachakravarthy]

Thanks! Also, is the transformer trainer package compatible for fine tuning the model?

[LeiWang1999]

@sriyachakravarthy Sry, I have no experience with that