Skip to content

[mps] Add offsets to enable aoti #2484

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

Open
wants to merge 3 commits into
base: gh/angelayi/1/base
Choose a base branch
from
Open

[mps] Add offsets to enable aoti #2484

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
05c8d04
Update
angelayi Jul 3, 2025
0aefecb
Update
angelayi Jul 3, 2025
6f8aff7
Update
angelayi Jul 3, 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
40 changes: 20 additions & 20 deletions torchao/experimental/kernels/mps/src/lowbit.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,11 +73,11 @@ using DispatchFn =
void (*)(id<MTLComputeCommandEncoder>, int32_t, int32_t, int32_t, int32_t);

inline void linear_lowbit_quant_weights_mps_impl(
id<MTLBuffer> a_buf,
id<MTLBuffer> b_buf,
id<MTLBuffer> s_buf,
id<MTLBuffer> z_buf,
id<MTLBuffer> out_buf,
std::pair<id<MTLBuffer>, size_t> a_buf_offset,
std::pair<id<MTLBuffer>, size_t> b_buf_offset,
std::pair<id<MTLBuffer>, size_t> s_buf_offset,
std::pair<id<MTLBuffer>, size_t> z_buf_offset,
std::pair<id<MTLBuffer>, size_t> out_buf_offset,
int32_t M,
int32_t K,
int32_t N,
Expand All @@ -97,11 +97,11 @@ inline void linear_lowbit_quant_weights_mps_impl(
metal_lowbit_quantized_lib.getPipelineStateForFunc(shader_func);
const auto maxThreadsPerGroup = [cpl maxTotalThreadsPerThreadgroup];
[computeEncoder setComputePipelineState:cpl];
[computeEncoder setBuffer:a_buf offset:0 atIndex:0];
[computeEncoder setBuffer:b_buf offset:0 atIndex:1];
[computeEncoder setBuffer:s_buf offset:0 atIndex:2];
[computeEncoder setBuffer:z_buf offset:0 atIndex:3];
[computeEncoder setBuffer:out_buf offset:0 atIndex:4];
[computeEncoder setBuffer:a_buf_offset.first offset:a_buf_offset.second atIndex:0];
[computeEncoder setBuffer:b_buf_offset.first offset:b_buf_offset.second atIndex:1];
[computeEncoder setBuffer:s_buf_offset.first offset:s_buf_offset.second atIndex:2];
[computeEncoder setBuffer:z_buf_offset.first offset:z_buf_offset.second atIndex:3];
[computeEncoder setBuffer:out_buf_offset.first offset:out_buf_offset.second atIndex:4];
[computeEncoder setBytes:sizes.data()
length:sizeof(uint32_t) * sizes.size()
atIndex:5];
Expand Down Expand Up @@ -133,12 +133,12 @@ std::tuple<const std::string, DispatchFn> get_shader_func_and_dispatch(
// LowBit Quantized Weights Linear on Metal
template <int nbit>
void linear_lowbit_quant_weights_mps(
id<MTLBuffer> a_buf,
id<MTLBuffer> b_buf,
std::pair<id<MTLBuffer>, size_t> a_buf_offset,
std::pair<id<MTLBuffer>, size_t> b_buf_offset,
int64_t qGroupSize,
id<MTLBuffer> s_buf,
id<MTLBuffer> z_buf,
id<MTLBuffer> out_buf,
std::pair<id<MTLBuffer>, size_t> s_buf_offset,
std::pair<id<MTLBuffer>, size_t> z_buf_offset,
std::pair<id<MTLBuffer>, size_t> out_buf_offset,
int32_t M,
int32_t K,
int32_t N,
Expand All @@ -154,11 +154,11 @@ void linear_lowbit_quant_weights_mps(
const DispatchFn dispatch_fn = std::get<1>(shader_func_and_dispatch);

return linear_lowbit_quant_weights_mps_impl(
a_buf,
b_buf,
s_buf,
z_buf,
out_buf,
a_buf_offset,
b_buf_offset,
s_buf_offset,
z_buf_offset,
out_buf_offset,
M,
K,
N,
Expand Down
10 changes: 5 additions & 5 deletions torchao/experimental/kernels/mps/test/test_lowbit.mm
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -118,12 +118,12 @@ void pack() {

void linear() {
LowBitQuantWeights<nbit>::linear(
buf_A,
buf_B,
{buf_A, 0},
{buf_B, 0},
qGroupSize,
buf_S,
buf_Z,
buf_C,
{buf_S, 0},
{buf_Z, 0},
{buf_C, 0},
M,
K,
N,
Expand Down
10 changes: 5 additions & 5 deletions torchao/experimental/ops/mps/linear_fp_act_xbit_weight_aten.mm
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -97,12 +97,12 @@ Tensor linear_mps_kernel_out(
auto K = A.size(1);

LowBitQuantWeights<nbit>::linear(
getMTLBufferStorage(A),
getMTLBufferStorage(B),
{getMTLBufferStorage(A), A.storage_offset() * A.element_size()},
{getMTLBufferStorage(B), B.storage_offset() * B.element_size()},
group_size,
getMTLBufferStorage(S),
getMTLBufferStorage(Z),
getMTLBufferStorage(C),
{getMTLBufferStorage(S), S.storage_offset() * S.element_size()},
{getMTLBufferStorage(Z), Z.storage_offset() * Z.element_size()},
{getMTLBufferStorage(C), C.storage_offset() * C.element_size()},
M,
K,
N,
Expand Down
10 changes: 5 additions & 5 deletions torchao/experimental/ops/mps/linear_fp_act_xbit_weight_executorch.mm
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,12 +95,12 @@ bool check_linear_mps_args(
auto K = A.size(1);

torchao::kernels::mps::lowbit::LowBitQuantWeights<nbit>::linear(
getMTLBufferStorage(A),
getMTLBufferStorage(B),
{getMTLBufferStorage(A), A.storage_offset() * A.element_size()},
{getMTLBufferStorage(B), B.storage_offset() * B.element_size()},
group_size,
getMTLBufferStorage(S),
getMTLBufferStorage(Z),
getMTLBufferStorage(out),
{getMTLBufferStorage(S), S.storage_offset() * S.element_size()},
{getMTLBufferStorage(Z), Z.storage_offset() * Z.element_size()},
{getMTLBufferStorage(out), out.storage_offset() * out.element_size()},
M,
K,
N,
Expand Down
36 changes: 36 additions & 0 deletions torchao/experimental/ops/mps/test/test_quantizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,6 +86,42 @@ def test_export(self, nbit):
== f"torchao._linear_fp_act_{nbit}bit_weight.default"
)

@parameterized.expand(BITWIDTHS)
def test_export_accuracy(self, nbit):
group_size = 32
m = 3
n = 12
k = 64
with torch.no_grad():
activations = torch.rand(m, k, dtype=torch.float32, device="mps")
model = torch.nn.Sequential(*[torch.nn.Linear(k, n, bias=False)])

# Compute expected result
weight_cpu = model[0].weight.data
weight_qvals_cpu, weight_scales_cpu, weight_zeros_cpu = _quantize(
weight_cpu, group_size, nbit, True, torch.uint8
)
weight_zeros_cpu = -weight_zeros_cpu * weight_scales_cpu
expected = self._reference_linear_lowbit_quant_weights(
activations.cpu(),
weight_qvals_cpu,
group_size,
weight_scales_cpu,
weight_zeros_cpu,
)

quantized_model = self._quantize_model(
model, torch.float32, nbit, group_size
)

ep = torch.export.export(quantized_model, (activations,), strict=True)
path = torch._inductor.aoti_compile_and_package(ep)
compiled_model = torch._inductor.aoti_load_package(path)
result = compiled_model(activations)

# Compare results
torch.testing.assert_close(result.cpu(), expected, rtol=0.001, atol=0.001)

@parameterized.expand(BITWIDTHS)
def test_2d_output_device_and_shape(self, nbit):
model, group_size, k0, n = self._model_setup()
Expand Down
Loading