Merge pull request #1 from TimDettmers/main

Update main branch

9 files changed, 20 insertions, 165 deletions

diff --git a/bitsandbytes/__init__.py b/bitsandbytes/__init__.py
index 7901f96..6d1177f 100644
--- a/bitsandbytes/__init__.py
+++ b/bitsandbytes/__init__.py
@@ -12,7 +12,7 @@ from .autograd._functions import (
 )
 from .cextension import COMPILED_WITH_CUDA
 from .nn import modules
-from . import cuda_setup
+from . import cuda_setup, utils
 
 if COMPILED_WITH_CUDA:
     from .optim import adam
diff --git a/bitsandbytes/__main__.py b/bitsandbytes/__main__.py
index 5f11875..175a30e 100644
--- a/bitsandbytes/__main__.py
+++ b/bitsandbytes/__main__.py
@@ -3,8 +3,9 @@
 # cli()
 import os
 import sys
-import torch
+from warnings import warn
 
+import torch
 
 HEADER_WIDTH = 60
 
@@ -32,8 +33,6 @@ print()
 from . import COMPILED_WITH_CUDA, PACKAGE_GITHUB_URL
 from .cuda_setup.main import get_compute_capabilities, get_cuda_lib_handle
 from .cuda_setup.env_vars import to_be_ignored
-from .utils import print_stderr
-
 
 print_header("POTENTIALLY LIBRARY-PATH-LIKE ENV VARS")
 for k, v in os.environ.items():
@@ -84,7 +83,7 @@ try:
 
 except ImportError:
     print()
-    print_stderr(
+    warn(
         f"WARNING: {__package__} is currently running as CPU-only!\n"
         "Therefore, 8-bit optimizers and GPU quantization are unavailable.\n\n"
         f"If you think that this is so erroneously,\nplease report an issue!"
diff --git a/bitsandbytes/autograd/_functions.py b/bitsandbytes/autograd/_functions.py
index 3bd39a9..226cbb5 100644
--- a/bitsandbytes/autograd/_functions.py
+++ b/bitsandbytes/autograd/_functions.py
@@ -1,6 +1,5 @@
 import operator
 import torch
-import bitsandbytes as bnb
 import bitsandbytes.functional as F
 
 from dataclasses import dataclass
@@ -368,9 +367,6 @@ class MatMul8bitLt(torch.autograd.Function):
         return grad_A, grad_B, None, grad_bias, None
 
 
-matmul = MatMul8bitLt.apply
-
-
 def matmul(
     A: tensor,
     B: tensor,
diff --git a/bitsandbytes/cuda_setup/compute_capability.py b/bitsandbytes/cuda_setup/compute_capability.py
deleted file mode 100644
index 7a3f463..0000000
--- a/bitsandbytes/cuda_setup/compute_capability.py
+++ /dev/null
@@ -1,79 +0,0 @@
-import ctypes
-from dataclasses import dataclass, field
-
-
-@dataclass
-class CudaLibVals:
-    # code bits taken from
-    # https://gist.github.com/f0k/63a664160d016a491b2cbea15913d549
-
-    nGpus: ctypes.c_int = field(default=ctypes.c_int())
-    cc_major: ctypes.c_int = field(default=ctypes.c_int())
-    cc_minor: ctypes.c_int = field(default=ctypes.c_int())
-    device: ctypes.c_int = field(default=ctypes.c_int())
-    error_str: ctypes.c_char_p = field(default=ctypes.c_char_p())
-    cuda: ctypes.CDLL = field(init=False, repr=False)
-    ccs: List[str, ...] = field(init=False)
-
-    def _initialize_driver_API(self):
-        self.check_cuda_result(self.cuda.cuInit(0))
-
-    def _load_cuda_lib(self):
-        """
-        1. find libcuda.so library (GPU driver) (/usr/lib)
-           init_device -> init variables -> call function by reference
-        """
-        libnames = "libcuda.so"
-        for libname in libnames:
-            try:
-                self.cuda = ctypes.CDLL(libname)
-            except OSError:
-                continue
-            else:
-                break
-        else:
-            raise OSError("could not load any of: " + " ".join(libnames))
-
-    def call_cuda_func(self, function_obj, **kwargs):
-        CUDA_SUCCESS = 0  # constant taken from cuda.h
-        pass
-        # if (CUDA_SUCCESS := function_obj(
-
-    def _error_handle(cuda_lib_call_return_value):
-        """
-        2. call extern C function to determine CC
-           (see https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__DEVICE__DEPRECATED.html)
-        """
-        CUDA_SUCCESS = 0  # constant taken from cuda.h
-
-        if cuda_lib_call_return_value != CUDA_SUCCESS:
-            self.cuda.cuGetErrorString(
-                cuda_lib_call_return_value,
-                ctypes.byref(self.error_str),
-            )
-            print("Count not initialize CUDA - failure!")
-            raise Exception("CUDA exception!")
-        return cuda_lib_call_return_value
-
-    def __post_init__(self):
-        self._load_cuda_lib()
-        self._initialize_driver_API()
-        self.check_cuda_result(
-            self.cuda, self.cuda.cuDeviceGetCount(ctypes.byref(self.nGpus))
-        )
-        tmp_ccs = []
-        for gpu_index in range(self.nGpus.value):
-            check_cuda_result(
-                self.cuda,
-                self.cuda.cuDeviceGet(ctypes.byref(self.device), gpu_index),
-            )
-            check_cuda_result(
-                self.cuda,
-                self.cuda.cuDeviceComputeCapability(
-                    ctypes.byref(self.cc_major),
-                    ctypes.byref(self.cc_minor),
-                    self.device,
-                ),
-            )
-            tmp_ccs.append(f"{self.cc_major.value}.{self.cc_minor.value}")
-        self.ccs = sorted(tmp_ccs, reverse=True)
diff --git a/bitsandbytes/cuda_setup/main.py b/bitsandbytes/cuda_setup/main.py
index 975b772..78a2844 100644
--- a/bitsandbytes/cuda_setup/main.py
+++ b/bitsandbytes/cuda_setup/main.py
@@ -17,9 +17,7 @@ evaluation:
 """
 
 import ctypes
-from pathlib import Path
 
-from ..utils import execute_and_return
 from .paths import determine_cuda_runtime_lib_path
 
 
@@ -28,7 +26,7 @@ def check_cuda_result(cuda, result_val):
     if result_val != 0:
         error_str = ctypes.c_char_p()
         cuda.cuGetErrorString(result_val, ctypes.byref(error_str))
-        raise Exception(f"CUDA exception! Error code: {error_str.value.decode()}")
+        print(f"CUDA exception! Error code: {error_str.value.decode()}")
 
 def get_cuda_version(cuda, cudart_path):
     # https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART____VERSION.html#group__CUDART____VERSION
@@ -57,7 +55,7 @@ def get_cuda_lib_handle():
         cuda = ctypes.CDLL("libcuda.so")
     except OSError:
         # TODO: shouldn't we error or at least warn here?
-        raise Exception('CUDA SETUP: ERROR! libcuda.so not found! Do you have a CUDA driver installed? If you are on a cluster, make sure you are on a CUDA machine!')
+        print('CUDA SETUP: WARNING! libcuda.so not found! Do you have a CUDA driver installed? If you are on a cluster, make sure you are on a CUDA machine!')
         return None
     check_cuda_result(cuda, cuda.cuInit(0))
 
@@ -80,7 +78,6 @@ def get_compute_capabilities(cuda):
     cc_major = ctypes.c_int()
     cc_minor = ctypes.c_int()
 
-    result = ctypes.c_int()
     device = ctypes.c_int()
 
     check_cuda_result(cuda, cuda.cuDeviceGetCount(ctypes.byref(nGpus)))
@@ -119,6 +116,10 @@ def evaluate_cuda_setup():
     print('For effortless bug reporting copy-paste your error into this form: https://docs.google.com/forms/d/e/1FAIpQLScPB8emS3Thkp66nvqwmjTEgxp8Y9ufuWTzFyr9kJ5AoI47dQ/viewform?usp=sf_link')
     print('='*80)
     binary_name = "libbitsandbytes_cpu.so"
+    #if not torch.cuda.is_available():
+        #print('No GPU detected. Loading CPU library...')
+        #return binary_name
+
     cudart_path = determine_cuda_runtime_lib_path()
     if cudart_path is None:
         print(
diff --git a/bitsandbytes/cuda_setup/paths.py b/bitsandbytes/cuda_setup/paths.py
index 9c565c7..6f6425f 100644
--- a/bitsandbytes/cuda_setup/paths.py
+++ b/bitsandbytes/cuda_setup/paths.py
@@ -2,23 +2,11 @@ from pathlib import Path
 from typing import Set, Union
 from warnings import warn
 
-from ..utils import print_stderr
 from .env_vars import get_potentially_lib_path_containing_env_vars
 
-
 CUDA_RUNTIME_LIB: str = "libcudart.so"
 
 
-def purge_unwanted_semicolon(tentative_path: Path) -> Path:
-    """
-    Special function to handle the following exception:
-    __LMOD_REF_COUNT_PATH=/sw/cuda/11.6.2/bin:2;/mmfs1/home/dettmers/git/sched/bin:1;/mmfs1/home/dettmers/data/anaconda3/bin:1;/mmfs1/home/dettmers/data/anaconda3/condabin:1;/mmfs1/home/dettmers/.local/bin:1;/mmfs1/home/dettmers/bin:1;/usr/local/bin:1;/usr/bin:1;/usr/local/sbin:1;/usr/sbin:1;/mmfs1/home/dettmers/.fzf/bin:1;/mmfs1/home/dettmers/data/local/cuda-11.4/bin:1
-    """
-    # if ';' in str(tentative_path):
-    #     path_as_str, _ = str(tentative_path).split(';')
-    pass
-
-
 def extract_candidate_paths(paths_list_candidate: str) -> Set[Path]:
     return {Path(ld_path) for ld_path in paths_list_candidate.split(":") if ld_path}
 
@@ -29,7 +17,7 @@ def remove_non_existent_dirs(candidate_paths: Set[Path]) -> Set[Path]:
     }
 
     if non_existent_directories:
-        print_stderr(
+        warn(
             "WARNING: The following directories listed in your path were found to "
             f"be non-existent: {non_existent_directories}"
         )
@@ -117,8 +105,6 @@ def determine_cuda_runtime_lib_path() -> Union[Path, None]:
         if env_var not in {"CONDA_PREFIX", "LD_LIBRARY_PATH"}
     }
 
-
-
     cuda_runtime_libs = set()
     for env_var, value in remaining_candidate_env_vars.items():
         cuda_runtime_libs.update(find_cuda_lib_in(value))
diff --git a/bitsandbytes/functional.py b/bitsandbytes/functional.py
index 6637554..22200f2 100644
--- a/bitsandbytes/functional.py
+++ b/bitsandbytes/functional.py
@@ -5,7 +5,6 @@
 import ctypes as ct
 import operator
 import random
-import math
 import torch
 
 from typing import Tuple
@@ -185,14 +184,9 @@ def create_dynamic_map(signed=True, n=7):
 
 
 def get_special_format_str():
+    if not torch.cuda.is_available(): return 'col_turing'
     major, minor = torch.cuda.get_device_capability()
-    if major < 7:
-        print(
-            f"Device with CUDA capability of {major} not supported for 8-bit matmul. Device has no tensor cores!"
-        )
-        assert major >= 7
-
-    if major == 7:
+    if major <= 7:
         return "col_turing"
     elif major == 8:
         return "col_ampere"
@@ -248,23 +242,6 @@ def get_transform_func(dtype, orderA, orderOut, transpose=False):
         return getattr(lib, name)
 
 
-class GlobalData(object):
-    _instance = None
-
-    def __init__(self):
-        raise RuntimeError("Call get_instance() instead")
-
-    def initialize(self):
-        self.data = {}
-
-    @classmethod
-    def get_instance(cls):
-        if cls._instance is None:
-            cls._instance = cls.__new__(cls)
-            cls._instance.initialize()
-        return cls._instance
-
-
 def get_transform_buffer(
     shape, dtype, device, to_order, from_order="row", transpose=False
 ):
@@ -1685,21 +1662,6 @@ def double_quant(
     return out_row, out_col, row_stats, col_stats, coo_tensor
 
 
-def get_special_format_str():
-    major, minor = torch.cuda.get_device_capability()
-    if major < 7:
-        print(
-            f"Device with CUDA capability of {major} not supported for 8-bit matmul. Device has no tensor cores!"
-        )
-        assert major >= 7
-
-    if major == 7: return 'col_turing'
-    elif major == 8: return 'col_ampere'
-    else: return 'col_turing'
-
-
-
-
 def transform(A, to_order, from_order='row', out=None, transpose=False, state=None, ld=None):
     prev_device = pre_call(A.device)
     if state is None: state = (A.shape, from_order)
diff --git a/bitsandbytes/optim/__init__.py b/bitsandbytes/optim/__init__.py
index a76d717..d18f1d1 100644
--- a/bitsandbytes/optim/__init__.py
+++ b/bitsandbytes/optim/__init__.py
@@ -5,13 +5,12 @@
 
 from bitsandbytes.cextension import COMPILED_WITH_CUDA
 
-if COMPILED_WITH_CUDA:
-    from .adam import Adam, Adam8bit, Adam32bit
-    from .adamw import AdamW, AdamW8bit, AdamW32bit
-    from .sgd import SGD, SGD8bit, SGD32bit
-    from .lars import LARS, LARS8bit, LARS32bit, PytorchLARS
-    from .lamb import LAMB, LAMB8bit, LAMB32bit
-    from .rmsprop import RMSprop, RMSprop8bit, RMSprop32bit
-    from .adagrad import Adagrad, Adagrad8bit, Adagrad32bit
+from .adam import Adam, Adam8bit, Adam32bit
+from .adamw import AdamW, AdamW8bit, AdamW32bit
+from .sgd import SGD, SGD8bit, SGD32bit
+from .lars import LARS, LARS8bit, LARS32bit, PytorchLARS
+from .lamb import LAMB, LAMB8bit, LAMB32bit
+from .rmsprop import RMSprop, RMSprop8bit, RMSprop32bit
+from .adagrad import Adagrad, Adagrad8bit, Adagrad32bit
 
 from .optimizer import GlobalOptimManager
diff --git a/bitsandbytes/utils.py b/bitsandbytes/utils.py
index 4256a87..1cd90e3 100644
--- a/bitsandbytes/utils.py
+++ b/bitsandbytes/utils.py
@@ -1,6 +1,5 @@
 import shlex
 import subprocess
-import sys
 from typing import Tuple
 
 
@@ -22,11 +21,3 @@ def execute_and_return(command_string: str) -> Tuple[str, str]:
 
     std_out, std_err = execute_and_return_decoded_std_streams(command_string)
     return std_out, std_err
-
-
-def print_stderr(s: str) -> None:
-    print(s, file=sys.stderr)
-
-
-def warn_of_missing_prerequisite(s: str) -> None:
-    print_stderr("WARNING, missing pre-requisite: " + s)