Update bitsandbytes, hivemind, transformers (#290)

- new bitsandbytes supports newer *and* older GPUs
- new hivemind supports a better bfloat16 codec

Co-authored-by: Alexander Borzunov <borzunov.alexander@gmail.com>

setup.cfg

@@ -32,12 +32,12 @@ packages = find:
 python_requires = >=3.7
 install_requires =
     torch>=1.12
-    bitsandbytes==0.34.0
+    bitsandbytes==0.37.1
     accelerate==0.15.0
     huggingface-hub==0.11.1
-    transformers==4.25.1
+    transformers>=4.25.1,<5.0.0
     speedtest-cli==2.1.3
-    hivemind==1.1.5
+    hivemind==1.1.6
     tensor_parallel==1.0.23
     humanfriendly
     async-timeout>=4.0.2

src/petals/bloom/block.py

@@ -8,10 +8,13 @@ from typing import Optional, Tuple
 
 import torch.nn.quantized.dynamic.modules.linear
 import transformers
+from packaging import version
 from transformers.models.bloom.modeling_bloom import BloomBlock, _expand_mask, _make_causal_mask, build_alibi_tensor
 
 if not os.getenv("PETALS_IGNORE_DEPENDENCY_VERSION"):
-    assert transformers.__version__.startswith("4.25."), "Please install transformers 4.25.1"
+    assert (
+        version.parse("4.26.0") < version.parse(transformers.__version__) < version.parse("5.0.0")
+    ), "Please install a proper transformers version: pip install transformers>=4.26.0,<5.0.0"
 
 
 class WrappedBloomBlock(BloomBlock):