mrq/bitsandbytes-rocm
1
0
Fork 1
Code Issues 1 Pull Requests Packages Projects Releases Wiki Activity

Fixed bugs in cuda setup.

Browse Source
This commit is contained in:
Tim Dettmers 2022-08-04 09:16:00 -07:00
parent 758c7175a2
commit 8f84674d67
4 changed files with 19 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
bitsandbytes/cextension.py
View File

@ -17,12 +17,13 @@ class CUDALibrary_Singleton(object):
binary_path = package_dir / binary_name binary_path = package_dir / binary_name
if not binary_path.exists(): if not binary_path.exists():
print(f"TODO: compile library for specific version: {binary_name}") print(f"CUDA_SETUP: TODO: compile library for specific version: {binary_name}")
legacy_binary_name = "libbitsandbytes.so" legacy_binary_name = "libbitsandbytes.so"
print(f"Defaulting to {legacy_binary_name}...") print(f"CUDA_SETUP: Defaulting to {legacy_binary_name}...")
self.lib = ct.cdll.LoadLibrary(package_dir / legacy_binary_name) self.lib = ct.cdll.LoadLibrary(package_dir / legacy_binary_name)
else: else:
self.lib = ct.cdll.LoadLibrary(package_dir / binary_name) print(f"CUDA_SETUP: Loading binary {binary_path}...")
self.lib = ct.cdll.LoadLibrary(binary_path)
@classmethod @classmethod
def get_instance(cls): def get_instance(cls):

2
bitsandbytes/cuda_setup/__init__.py
View File

@ -0,0 +1,2 @@
from .paths import CUDA_RUNTIME_LIB, extract_candidate_paths, determine_cuda_runtime_lib_path
from .main import evaluate_cuda_setup

14
bitsandbytes/cuda_setup/main.py
View File

@ -47,6 +47,7 @@ def get_compute_capabilities():
cuda = ctypes.CDLL("libcuda.so") cuda = ctypes.CDLL("libcuda.so")
except OSError: except OSError:
# TODO: shouldn't we error or at least warn here? # TODO: shouldn't we error or at least warn here?
print('ERROR: libcuda.so not found!')
return None return None
nGpus = ctypes.c_int() nGpus = ctypes.c_int()
@ -70,7 +71,7 @@ def get_compute_capabilities():
) )
ccs.append(f"{cc_major.value}.{cc_minor.value}") ccs.append(f"{cc_major.value}.{cc_minor.value}")
return ccs.sort() return ccs
# def get_compute_capability()-> Union[List[str, ...], None]: # FIXME: error # def get_compute_capability()-> Union[List[str, ...], None]: # FIXME: error
@ -80,7 +81,8 @@ def get_compute_capability():
capabilities are downwards compatible. If no GPUs are detected, it returns capabilities are downwards compatible. If no GPUs are detected, it returns
None. None.
""" """
if ccs := get_compute_capabilities() is not None: ccs = get_compute_capabilities()
if ccs is not None:
# TODO: handle different compute capabilities; for now, take the max # TODO: handle different compute capabilities; for now, take the max
return ccs[-1] return ccs[-1]
return None return None
@ -92,8 +94,7 @@ def evaluate_cuda_setup():
cc = get_compute_capability() cc = get_compute_capability()
binary_name = "libbitsandbytes_cpu.so" binary_name = "libbitsandbytes_cpu.so"
# FIXME: has_gpu is still unused if cc == '':
if not (has_gpu := bool(cc)):
print( print(
"WARNING: No GPU detected! Check your CUDA paths. Processing to load CPU-only library..." "WARNING: No GPU detected! Check your CUDA paths. Processing to load CPU-only library..."
) )
@ -115,6 +116,7 @@ def evaluate_cuda_setup():
ls_output.split(" ")[-1].replace("libcudart.so.", "").split(".") ls_output.split(" ")[-1].replace("libcudart.so.", "").split(".")
) )
cuda_version_string = f"{major}{minor}" cuda_version_string = f"{major}{minor}"
print(f'CUDA_SETUP: Detected CUDA version {cuda_version_string}')
def get_binary_name(): def get_binary_name():
"if not has_cublaslt (CC < 7.5), then we have to choose _nocublaslt.so" "if not has_cublaslt (CC < 7.5), then we have to choose _nocublaslt.so"
@ -122,6 +124,8 @@ def evaluate_cuda_setup():
if has_cublaslt: if has_cublaslt:
return f"{bin_base_name}{cuda_version_string}.so" return f"{bin_base_name}{cuda_version_string}.so"
else: else:
return f"{bin_base_name}_nocublaslt.so" return f"{bin_base_name}{cuda_version_string}_nocublaslt.so"
binary_name = get_binary_name()
return binary_name return binary_name

8
tests/test_autograd.py
View File

@ -351,9 +351,9 @@ def test_matmullt(
err = torch.abs(out_bnb - out_torch).mean().item() err = torch.abs(out_bnb - out_torch).mean().item()
# print(f'abs error {err:.4f}') # print(f'abs error {err:.4f}')
idx = torch.isclose(out_bnb, out_torch, atol=0.01, rtol=0.1) idx = torch.isclose(out_bnb, out_torch, atol=0.01, rtol=0.1)
assert (idx == 0).sum().item() < n * 0.0175 assert (idx == 0).sum().item() <= n * 0.0175
idx = torch.isclose(out_bnb, out_torch, atol=0.035, rtol=0.2) idx = torch.isclose(out_bnb, out_torch, atol=0.035, rtol=0.2)
assert (idx == 0).sum().item() < n * 0.001 assert (idx == 0).sum().item() <= n * 0.001
if has_fp16_weights: if has_fp16_weights:
if any(req_grad): if any(req_grad):
@ -391,9 +391,9 @@ def test_matmullt(
assert torch.abs(gradB2).sum() == 0.0 assert torch.abs(gradB2).sum() == 0.0
idx = torch.isclose(gradB1, gradB2, atol=0.06, rtol=0.3) idx = torch.isclose(gradB1, gradB2, atol=0.06, rtol=0.3)
assert (idx == 0).sum().item() < n * 0.1 assert (idx == 0).sum().item() <= n * 0.1
idx = torch.isclose(gradB1, gradB2, atol=0.10, rtol=0.3) idx = torch.isclose(gradB1, gradB2, atol=0.10, rtol=0.3)
assert (idx == 0).sum().item() < n * 0.02 assert (idx == 0).sum().item() <= n * 0.02
torch.testing.assert_allclose( torch.testing.assert_allclose(
gradB1, gradB2, atol=0.18, rtol=0.3 gradB1, gradB2, atol=0.18, rtol=0.3
) )