segmentation fault illegal instruction

setup

ubuntu 16.04 tvm 0.7 dev1 pytorch 1.4.0 transformer 2.11.0 other same as requirements.txt

issue

I uncomment the line in diagonaled_mm_tvm.py DiagonaledMM._get_function('float32', 'cuda')

After that, When I run the code , it show Loading tvm binary from :./longformer/lib/lib_diagonaled_mm_float32_cuda.so … segmentation fault (core dump) or show Loading tvm binary from :./longformer/lib/lib_diagonaled_mm_float32_cuda.so … illegal instruction (core dump)

other

I test the tvm, tensorflow and pytorch, there are fine. And I follow the scripts/cheatsheet.txt to regenerate the lib_diagonaled_mm_float32_cuda.so, it can generate succeed.

Any idea or suggestion?

the code is below

import torch
from longformer.longformer import Longformer, LongformerConfig
from longformer.sliding_chunks import pad_to_window_size
from transformers import RobertaTokenizer

config = LongformerConfig.from_pretrained('longformer-base-4096/') 
# choose the attention mode 'n2', 'tvm' or 'sliding_chunks'
# 'n2': for regular n2 attantion
# 'tvm': a custom CUDA kernel implementation of our sliding window attention
# 'sliding_chunks': a PyTorch implementation of our sliding window attention
config.attention_mode = 'tvm'

model = Longformer.from_pretrained('longformer-base-4096/', config=config)
tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
tokenizer.model_max_length = model.config.max_position_embeddings

SAMPLE_TEXT = ' '.join(['Hello world! '] * 1000)  # long input document

input_ids = torch.tensor(tokenizer.encode(SAMPLE_TEXT)).unsqueeze(0)  # batch of size 1

# TVM code doesn't work on CPU. Uncomment this if `config.attention_mode = 'tvm'`
model = model.cuda(); input_ids = input_ids.cuda()

# Attention mask values -- 0: no attention, 1: local attention, 2: global attention
attention_mask = torch.ones(input_ids.shape, dtype=torch.long, device=input_ids.device) # initialize to local attention
attention_mask[:, [1, 4, 21,]] =  2  # Set global attention based on the task. For example,
                                     # classification: the <s> token
                                     # QA: question tokens

# padding seqlen to the nearest multiple of 512. Needed for the 'sliding_chunks' attention
input_ids, attention_mask = pad_to_window_size(
        input_ids, attention_mask, config.attention_window[0], tokenizer.pad_token_id)

output = model(input_ids, attention_mask=attention_mask)[0]

[ibeltagy]

Very interesting. Maybe a fused self-attention function or something. I will be curious to see how this goes.

Depending on how familiar you are with TVM, you might find the following discussions useful, https://discuss.tvm.ai/t/optimizing-matrix-multiplication-for-gpu/4212/24 https://discuss.tvm.ai/t/competitive-gemm-matmul-example/5478 https://discuss.tvm.ai/t/developing-a-faster-schedule-for-longformers-kernel/6367

[ibeltagy]

Another suggestion; can you try running it from inside the docker container that we use to compile the cuda kernel? Follow the instructions here: https://github.com/allenai/longformer/blob/master/scripts/cheatsheet.txt#L6 to build and run the docker image, then try to run it. You don’t need to recompile the binaries, it is enough to load the existing one.

I am curious, what are you using it for, and would the sliding_chunks implementation be enough for your use case?