Quant.h File

Quant Node — CVU kernel that quantizes FP32 to INT8 (scale + zero-point). More...

Included Headers

#include "builder/NodeContractConfigurable.h" #include "builder/NodeContractProvider.h" #include "builder/Node.h" #include <nlohmann/json.hpp> #include <memory> #include <optional> #include <string> #include <vector>

Namespaces Index

namespace	simaai
namespace	neat
namespace	nodes

Classes Index

struct	QuantOptions
	Construction options for a Quant Node. More...
class	Quant
	CVU kernel Node that quantizes FP32 tensors to INT8 using scale + zero-point. More...

Description

Quant Node — CVU kernel that quantizes FP32 to INT8 (scale + zero-point).

Reads FP32 tensors and emits INT8 using the per-tensor scale and zero-point bound to the model. Inserted at the MLA boundary on the INT8 path before the MLA when the upstream stage emits FP32.

See Also

"The dtype contract" page in /concepts/dtype_contract

File Listing

The file content with the documentation metadata removed is:

1

12#pragma once

13

14#include "builder/NodeContractConfigurable.h"

15#include "builder/NodeContractProvider.h"

16#include "builder/Node.h"

17#ifdef SIMA_NEAT_INTERNAL

18#include "model/internal/ModelRouteRetarget.h"

19#endif

20

21#include <nlohmann/json.hpp>

22

23#include <memory>

24#include <optional>

25#include <string>

26#include <vector>

27

28namespace simaai::neat {

29class Model;

30struct CompiledProcessCvuContract;

31

37struct QuantOptions {

39 QuantOptions() = default;

41 explicit QuantOptions(const simaai::neat::Model& model);

42

43 std::string config_path;

44 std::optional<nlohmann::json>

45 config_json;

46 std::string element_name;

47 std::shared_ptr<const CompiledProcessCvuContract>

48 compiled_contract;

49 int num_buffers = 0;

50 int num_buffers_model = 0;

51 bool num_buffers_locked = false;

52#ifdef SIMA_NEAT_INTERNAL

53 std::shared_ptr<const simaai::neat::internal::ModelLineageBinding> model_lineage;

54#endif

55};

56

68class Quant final : public Node, public NodeContractProvider, public NodeContractConfigurable {

69public:

71 explicit Quant(QuantOptions opt = {});

72 struct ConfigHolder;

73

75 std::string kind() const override {

76 return "Quant";

77 }

79 NodeCapsBehavior caps_behavior() const override {

80 return NodeCapsBehavior::Static;

81 }

83 NodeContractDefinition contract_definition() const override;

85 bool compile_node_contract(const ContractCompileInput& input, CompiledNodeContract* out,

86 std::string* err) const override;

88 void apply_compiled_contract(const CompiledNodeContract& contract, std::string* err) override;

90 std::string backend_fragment(int node_index) const override;

92 std::vector<std::string> element_names(int node_index) const override;

93

95 const nlohmann::json* config_json() const;

96

98 const QuantOptions& options() const {

99 return opt_;

100 }

102 const std::string& config_path() const {

103 return config_path_;

104 }

105

106#ifdef SIMA_NEAT_INTERNAL

107 const std::optional<CompiledProcessCvuContract>& compiled_contract_internal() const;

108#endif

109

110private:

111 QuantOptions opt_;

112 std::shared_ptr<ConfigHolder> config_holder_;

113 std::string config_path_;

114};

115

116} // namespace simaai::neat

117

118namespace simaai::neat::nodes {

120std::shared_ptr<simaai::neat::Node> Quant(QuantOptions opt = {});

121} // namespace simaai::neat::nodes

---

Generated via doxygen2docusaurus 2.0.0 by Doxygen 1.9.8.