Skip to content

[BF16] Support ndarray.asnumpy() to bfloat16 tensor natively using ml_dtypes #17739

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Mar 12, 2025
Merged

[BF16] Support ndarray.asnumpy() to bfloat16 tensor natively using ml_dtypes #17739

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6ad7e11
Update ndarray.py
spectrometerHBH Mar 11, 2025
1223e29
Fix llvm codegen bfloat16 test
MasterJH5574 Mar 12, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 6 additions & 1 deletion python/tvm/runtime/ndarray.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -233,7 +233,12 @@ def numpy(self):
if dtype == "int4":
dtype = "int8"
if dtype == "bfloat16":
dtype = "uint16"
if ml_dtypes is not None:
dtype = ml_dtypes.bfloat16
else:
raise RuntimeError(
"ml_dtypes is not installed, cannot convert bfloat16 array to numpy."
)
if dtype == "float8_e4m3fn":
if ml_dtypes is not None:
dtype = ml_dtypes.float8_e4m3fn
Expand Down
43 changes: 8 additions & 35 deletions tests/python/codegen/test_target_codegen_llvm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -758,33 +758,6 @@ def check_llvm_ir():
check_llvm_ir()


def np_float2np_bf16(arr):
"""Convert a numpy array of float to a numpy array
of bf16 in uint16"""
orig = arr.view("<u4")
bias = np.bitwise_and(np.right_shift(orig, 16), 1) + 0x7FFF
return np.right_shift(orig + bias, 16).astype("uint16")


def np_float2tvm_bf16(arr):
"""Convert a numpy array of float to a TVM array
of bf16"""
nparr = np_float2np_bf16(arr)
return tvm.nd.empty(nparr.shape, "bfloat16").copyfrom(nparr)


def np_bf162np_float(arr):
"""Convert a numpy array of bf16 (uint16) to a numpy array
of float"""
u32 = np.left_shift(arr.astype("uint32"), 16)
return u32.view("<f4")


def np_bf16_cast_and_cast_back(arr):
"""Convert a numpy array of float to bf16 and cast back"""
return np_bf162np_float(np_float2np_bf16(arr))


@tvm.testing.requires_llvm
def test_llvm_bf16():
def dotest(do_vectorize):
Expand All @@ -806,16 +779,16 @@ def dotest(do_vectorize):
sch.vectorize(loop)

module = tvm.compile(sch.mod, target="llvm")
npa = np.random.rand(32).astype("float32")
npb = np.random.rand(32).astype("float32")
va = np_bf16_cast_and_cast_back(npa)
vb = np_bf16_cast_and_cast_back(npb)
res = np_bf16_cast_and_cast_back(va + vb)
a_ = np_float2tvm_bf16(npa)
b_ = np_float2tvm_bf16(npb)
npa = np.random.rand(32).astype("bfloat16")
npb = np.random.rand(32).astype("bfloat16")
res = npa + npb
a_ = tvm.nd.array(npa)
b_ = tvm.nd.array(npb)
c_ = tvm.nd.empty((32,), "bfloat16")
module(a_, b_, c_)
tvm.testing.assert_allclose(np_bf162np_float(c_.numpy()), res)
# Note: directly compare without casting to float32 should work with the
# latest numpy version.
tvm.testing.assert_allclose(c_.numpy().astype("float32"), res.astype("float32"))

dotest(True)
dotest(False)
Expand Down
Loading