Skip to content

support npu multinomial #10668

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
ShawnXuan merged 2 commits into from
Aug 20, 2025
+18 −11
Merged

support npu multinomial #10668

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
a3df590
support npu multinomial
ShawnXuan Aug 15, 2025
2cacce3
auto format by CI
oneflow-ci-bot Aug 15, 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
26 changes: 16 additions & 10 deletions oneflow/core/functional/impl/random_functor.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -801,6 +801,8 @@ class MultinomialFunctor {
.Input("prefix_sum")
.Output("out")
.Build());
op_npu_ =
CHECK_JUST(one::OpBuilder("multinomial_with_replacement").Input("x").Output("out").Build());
}

Maybe<Tensor> operator()(const std::shared_ptr<one::Tensor>& x, const int& num_samples,
Expand All @@ -823,10 +825,17 @@ class MultinomialFunctor {
CHECK_OR_RETURN(num_categories <= FLOAT32_MAX_CONSECUTIVE_INT)
<< "number of categories cannot exceed 2^24";

DeviceType input_device = DeviceType::kCPU;
if (x->is_global()) {
JUST(CheckDeviceIdsIsValid(JUST(x->parallel_desc())));
input_device = JUST(x->parallel_desc())->device_type();
} else {
input_device = JUST(x->device())->enum_type();
}
// Fast-path for no replacement.
// Reference:
// https://github.com/pytorch/pytorch/issues/11931#issuecomment-625882503
if (!replacement) {
if (!replacement && input_device != DeviceType::kNPU) {
// The algorithm is from gumbel softmax.
// s = argmax( logp - log(-log(eps)) ) where eps ~ U(0, 1)
// Here we can apply exp to the formula which will not affect result of
Expand All @@ -847,6 +856,7 @@ class MultinomialFunctor {
std::shared_ptr<Tensor> result;
if (num_samples == 1) {
result = JUST(functional::ArgMax(q, -1, true, JUST(DType::Get(DataType::kInt64))));
} else if (input_device == DeviceType::kNPU) {
} else {
std::shared_ptr<TensorTuple> temp =
JUST(functional::TopK(q, num_samples, -1,
Expand All @@ -856,23 +866,18 @@ class MultinomialFunctor {
return result;
}

DeviceType input_device = DeviceType::kCPU;
if (x->is_global()) {
JUST(CheckDeviceIdsIsValid(JUST(x->parallel_desc())));
input_device = JUST(x->parallel_desc())->device_type();
} else {
input_device = JUST(x->device())->enum_type();
}
auto gen = generator.value_or(JUST(one::DefaultAutoGenerator()));
gen = JUST(GetGeneratorForLazyOrGlobal(gen, LazyMode::is_enabled(), x));
auto& attrs = THREAD_CACHED_MUTABLE_ATTR_MAP("seed", "num_samples");
attrs.SetAllAttrs(static_cast<int64_t>(gen->current_seed()), num_samples);
auto& attrs = THREAD_CACHED_MUTABLE_ATTR_MAP("seed", "num_samples", "replacement");
attrs.SetAllAttrs(static_cast<int64_t>(gen->current_seed()), num_samples, replacement);

const auto& distribution_state = std::make_shared<DistributionKernelState>(gen);
OpExprInterpContext ctx(attrs, distribution_state);

if (input_device == DeviceType::kCPU) {
return OpInterpUtil::Dispatch<Tensor>(*op_cpu_, {x}, ctx);
} else if (input_device == DeviceType::kNPU) {
return OpInterpUtil::Dispatch<Tensor>(*op_npu_, {x}, ctx);
} else {
std::shared_ptr<Tensor> sum_last_dim = JUST(functional::ReduceSum(x, {-1}, true, NullOpt));
std::shared_ptr<Tensor> norm_dist = JUST(functional::Div(x, sum_last_dim));
Expand All @@ -884,6 +889,7 @@ class MultinomialFunctor {
private:
std::shared_ptr<OpExpr> op_cpu_;
std::shared_ptr<OpExpr> op_gpu_;
std::shared_ptr<OpExpr> op_npu_;
};

} // namespace impl
Expand Down
3 changes: 2 additions & 1 deletion oneflow/ir/include/OneFlow/OneFlowUserOps.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6355,7 +6355,8 @@ def OneFlow_MultinomialWithReplacementOp : OneFlow_BaseOp<"multinomial_with_repl
);
let attrs = (ins
DefaultValuedAttr<SI64Attr, "0">:$seed,
DefaultValuedAttr<SI32Attr, "1">:$num_samples
DefaultValuedAttr<SI32Attr, "1">:$num_samples,
DefaultValuedAttr<BoolAttr, "true">:$replacement
);
let has_logical_tensor_desc_infer_fn = 1;
let has_physical_tensor_desc_infer_fn = 1;
Expand Down
Loading