Open dvrogozh opened 2 days ago
Another case which breaks is importing bitsandbytes
:
$ pip3 install bitsandbytes
$ pip3 list | grep bitsandbytes
bitsandbytes 0.43.3
$ python3 -c 'import bitsandbytes'
The installed version of bitsandbytes was compiled without GPU support. 8-bit optimizers, 8-bit multiplication, and GPU quantization are unavailable.
Traceback (most recent call last):
File "<string>", line 1, in <module>
File "/home/dvrogozh/pytorch.xpu/lib/python3.10/site-packages/bitsandbytes/__init__.py", line 15, in <module>
from .nn import modules
File "/home/dvrogozh/pytorch.xpu/lib/python3.10/site-packages/bitsandbytes/nn/__init__.py", line 17, in <module>
from .triton_based_modules import (
File "/home/dvrogozh/pytorch.xpu/lib/python3.10/site-packages/bitsandbytes/nn/triton_based_modules.py", line 7, in <module>
from bitsandbytes.triton.int8_matmul_mixed_dequantize import (
File "/home/dvrogozh/pytorch.xpu/lib/python3.10/site-packages/bitsandbytes/triton/int8_matmul_mixed_dequantize.py", line 12, in <module>
from triton.ops.matmul_perf_model import early_config_prune, estimate_matmul_time
ModuleNotFoundError: No module named 'triton.ops'
Thus, any project which imports bitsandbytes might get broken if intel triton version is installed. I noticed that using Huggingface peft and text-generation-inference projects.
With:
release/3.1.x
)Intel's triton built from main branch (versioned 3.0.0) or built from release/3.1.x branch (versioned 3.1.1) don't expose
triton.ops
- see error from pytorch benchmark below. Actually Intel's triton version corresponds to some later state of triton (probably 3.2.x?).triton.ops
were dropped from triton main in this PR between 3.1 and 3.2:You can see that Triton 3.1.x release branch does have
ops
folder in there:and Intel 3.1.x does not:
Can Intel triton builds be correctly versioned? I believe that build for pytorch 2.5 is actually 3.2.x not 3.1.x
@vlad-penkin