Flash Attention is not installed?

[ObliviousDonkey]

    raise NotImplementedError("flash attention is not installed")
NotImplementedError: flash attention is not installed
2024-09-06T05:28:11.268655Z ERROR shard-manager: lorax_launcher: Shard complete stan
dard error output:
/opt/conda/lib/python3.10/site-packages/torch/cuda/__init__.py:128: UserWarning: CUD
A initialization: Unexpected error from cudaGetDeviceCount(). Did you run some cuda
functions before calling NumCudaDevices() that might have already set an error? Erro
r 803: system has unsupported display driver / cuda driver combination (Triggered in
ternally at /opt/conda/conda-bld/pytorch_1720538438429/work/c10/cuda/CUDAFunctions.c
pp:108.)
  return torch._C._cuda_getDeviceCount() > 0
The installed version of bitsandbytes was compiled without GPU support. 8-bit optimi
zers, 8-bit multiplication, and GPU quantization are unavailable.
2024-09-06 05:28:02.507 | INFO     | lorax_server.utils.state:<module>:11 - Prefix c
aching = False
Traceback (most recent call last):
  File "/opt/conda/bin/lorax-server", line 8, in <module>
    sys.exit(app())
  File "/opt/conda/lib/python3.10/site-packages/lorax_server/cli.py", line 87, in se
rve
    server.serve(
  File "/opt/conda/lib/python3.10/site-packages/lorax_server/server.py", line 408, i
n serve
    asyncio.run(
  File "/opt/conda/lib/python3.10/asyncio/runners.py", line 44, in run
    return loop.run_until_complete(main)
  File "/opt/conda/lib/python3.10/asyncio/base_events.py", line 649, in run_until_co
mplete
    return future.result()
  File "/opt/conda/lib/python3.10/site-packages/lorax_server/server.py", line 274, i
n serve_inner
    model = get_model(
  File "/opt/conda/lib/python3.10/site-packages/lorax_server/models/__init__.py", li
ne 165, in get_model
    from lorax_server.models.flash_llama import FlashLlama
  File "/opt/conda/lib/python3.10/site-packages/lorax_server/models/flash_llama.py",
 line 9, in <module>
    from lorax_server.models.custom_modeling.flash_llama_modeling import (
  File "/opt/conda/lib/python3.10/site-packages/lorax_server/models/custom_modeling/
flash_llama_modeling.py", line 32, in <module>
    from lorax_server.utils import flash_attn, paged_attention
  File "/opt/conda/lib/python3.10/site-packages/lorax_server/utils/flash_attn.py", l
ine 330, in <module>
    raise NotImplementedError("flash attention is not installed")
NotImplementedError: flash attention is not installed
 rank=0
2024-09-06T05:28:11.324488Z ERROR lorax_launcher: Shard 0 failed to start
2024-09-06T05:28:11.324520Z  INFO lorax_launcher: Shutting down shards
Error: ShardCannotStart

What might be causing the issue? Im on rtx 3060

[poddarabhinav]

try using this image as they have updated to cuda 12.4 there is some driver issue
docker image : ghcr.io/predibase/lorax:07addea

[2022dc04025]

When using mistralai/Mistral-7B-Instruct-v0.1, I am getting the below error even if changing the docker image to ghcr.io/predibase/lorax:07addea

ImportError: Mistral model requires flash attn v2
rank=0
Error: ShardCannotStart

[ObliviousDonkey]

Any solutions?

[poddarabhinav]

I tried this with Llama 3.1, and the issue was that the Nvidia driver didn't support the CUDA version in the Lorax Docker image. In my case, when I executed commands in the Docker image, CUDA was not available, so I tried different Docker images provided by them. Luckily, ghcr.io/predibase/lorax:07addea worked. Apparently, CUDA 12.4 had a mismatch with Nvidia driver version 550.9.

[ObliviousDonkey]

I tried this with Llama 3.1, and the issue was that the Nvidia driver didn't support the CUDA version in the Lorax Docker image. In my case, when I executed commands in the Docker image, CUDA was not available, so I tried different Docker images provided by them. Luckily, ghcr.io/predibase/lorax:07addea worked. Apparently, CUDA 12.4 had a mismatch with Nvidia driver version 550.9.

Thanks for the help. But after switching to the docker image you gave, I'm facing new errors when running unsloth/Meta-Llama-3.1-8B-bnb-4bit

Driver version: 535.183.01, Cuda Version: 12.2

Did you succeed running this one?

[dumbPy]

I too had this issue with that particular model.
You may try other models like hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4 which gives me 55 tokens/sec on A100
or meta-llama/Meta-Llama-3.1-8B-Instruct that gives 60 tokens/sec on A100.

[vsokolovskii]

Tried the image you suggested, base model is unsloth 70B Llama3.1

getting some dimensions mismatches:

AssertionError: [41943040, 1] != [10240, 8192]
 rank=0
Error: ShardCannotStart

[ObliviousDonkey]

Alright, looks like theres a problem with the specific model