QPY serialization

qiskit.qpy

QPY is a binary serialization format for QuantumCircuit and ScheduleBlock objects that is designed to be cross-platform, Python version agnostic, and backwards compatible moving forward. QPY should be used if you need a mechanism to save or copy between systems a QuantumCircuit or ScheduleBlock that preserves the full Qiskit object structure (except for custom attributes defined outside of Qiskit code). This differs from other serialization formats like OpenQASM (2.0 or 3.0) which has a different abstraction model and can result in a loss of information contained in the original circuit (or is unable to represent some aspects of the Qiskit objects) or Python’s pickle which will preserve the Qiskit object exactly but will only work for a single Qiskit version (it is also potentially insecure).

Basic Usage

Using QPY is defined to be straightforward and mirror the user API of the serializers in Python’s standard library, pickle and json. There are 2 user facing functions: qiskit.qpy.dump() and qiskit.qpy.load() which are used to dump QPY data to a file object and load circuits from QPY data in a file object respectively. For example:

from qiskit.circuit import QuantumCircuit
from qiskit import qpy
 
qc = QuantumCircuit(2, name='Bell', metadata={'test': True})
qc.h(0)
qc.cx(0, 1)
qc.measure_all()
 
with open('bell.qpy', 'wb') as fd:
    qpy.dump(qc, fd)
 
with open('bell.qpy', 'rb') as fd:
    new_qc = qpy.load(fd)[0]

The qiskit.qpy.dump() function also lets you include multiple circuits in a single QPY file:

with open('twenty_bells.qpy', 'wb') as fd:
    qpy.dump([qc] * 20, fd)

and then loading that file will return a list with all the circuits

with open('twenty_bells.qpy', 'rb') as fd:
    twenty_new_bells = qpy.load(fd)

API documentation

load

from qiskit import qpy
 
with open('bell.qpy', 'rb') as fd:
    circuits = qpy.load(fd)

import gzip
from qiskit import qpy
 
with gzip.open('bell.qpy.gz', 'rb') as fd:
    circuits = qpy.load(fd)

dump

from qiskit.circuit import QuantumCircuit
from qiskit import qpy
 
qc = QuantumCircuit(2, name='Bell', metadata={'test': True})
qc.h(0)
qc.cx(0, 1)
qc.measure_all()

with open('bell.qpy', 'wb') as fd:
    qpy.dump(qc, fd)

import gzip
 
with gzip.open('bell.qpy.gz', 'wb') as fd:
    qpy.dump(qc, fd)

These functions will raise a custom subclass of QiskitError if they encounter problems during serialization or deserialization.

QpyError

When a lower-than-maximum target QPY version is set for serialization, but the object to be serialized contains features that cannot be represented in that format, a subclass of QpyError is raised:

UnsupportedFeatureForVersion

qiskit.qpy.QPY_VERSION

qiskit.qpy.QPY_COMPATIBILITY_VERSION

QPY Compatibility

The QPY format is designed to be backwards compatible moving forward. This means you should be able to load a QPY with any newer Qiskit version than the one that generated it. However, loading a QPY file with an older Qiskit version is not supported and may not work.

For example, if you generated a QPY file using qiskit-terra 0.18.1 you could load that QPY file with qiskit-terra 0.19.0 and a hypothetical qiskit-terra 0.29.0. However, loading that QPY file with 0.18.0 is not supported and may not work.

If a feature being loaded is deprecated in the corresponding qiskit release, QPY will raise a QPYLoadingDeprecatedFeatureWarning informing of the deprecation period and how the feature will be internally handled.

QPYLoadingDeprecatedFeatureWarning

QPY format version history

If you’re planning to load a QPY file between different Qiskit versions knowing which versions were available in a given release are useful. As the QPY is backwards compatible but not forwards compatible you need to ensure a given QPY format version was released in the release you’re calling load() with. The following table lists the QPY versions that were supported in every Qiskit (and qiskit-terra prior to Qiskit 1.0.0) release going back to the introduction of QPY in qiskit-terra 0.18.0.

QPY Format

The QPY serialization format is a portable cross-platform binary serialization format for QuantumCircuit objects in Qiskit. The basic file format is as follows:

A QPY file (or memory object) always starts with the following 6 byte UTF8 string: QISKIT which is immediately followed by the overall file header. The contents of the file header as defined as a C struct are:

struct {
    uint8_t qpy_version;
    uint8_t qiskit_major_version;
    uint8_t qiskit_minor_version;
    uint8_t qiskit_patch_version;
    uint64_t num_circuits;
}

From V10 on, a new field is added to the file header struct to represent the encoding scheme used for symbolic expressions:

struct {
    uint8_t qpy_version;
    uint8_t qiskit_major_version;
    uint8_t qiskit_minor_version;
    uint8_t qiskit_patch_version;
    uint64_t num_circuits;
    char symbolic_encoding;
}

All values use network byte order [1] (big endian) for cross platform compatibility.

The file header is immediately followed by the circuit payloads. Each individual circuit is composed of the following parts:

HEADER | METADATA | REGISTERS | STANDALONE_VARS | CUSTOM_DEFINITIONS | INSTRUCTIONS

The STANDALONE_VARS are new in QPY version 12; before that, there was no data between REGISTERS and CUSTOM_DEFINITIONS.

There is a circuit payload for each circuit (where the total number is dictated by num_circuits in the file header). There is no padding between the circuits in the data.

Version 13

Version 13 added a native Qiskit serialization representation for ParameterExpression. Previous QPY versions relied on either sympy or symengine to serialize the underlying symbolic expression. Starting in Version 13, QPY now represents the sequence of API calls used to create the ParameterExpression.

The main change in the serialization format is in the PARAMETER_EXPR payload. The expr_size bytes following the head now contain an array of PARAM_EXPR_ELEM_V13 structs. The intent is for this array to be read one struct at a time, where each struct describes one of the calls to make to reconstruct the ParameterExpression.

PARAM_EXPR_ELEM_V13

The struct format is defined as:

struct {
    unsigned char op_code;
    char lhs_type;
    char lhs[16];
    char rhs_type;
    char rhs[16];
} PARAM_EXPR_ELEM_V13;

The op_code field is used to define the operation added to the ParameterExpression. The value can be:

The NULL value of 255 is only used to fill the op code field for entries that are not actual operations but indicate recursive definitions. Then the lhs_type and rhs_type fields are used to describe the operand types and can be one of the following UTF-8 encoded characters:

If the type value is f, c, or i, the corresponding lhs or rhs field widths are 128 bits each. In the case of floats, the literal value is encoded as a double with 0 padding, while complex numbers are encoded as real part followed by imaginary part, taking up 64 bits each. For i, the value is encoded as a 64 bit signed integer with 0 padding for the full 128 bit width. n is used to represent a None and typically isn’t directly used as it indicates an argument that’s not used. For p the data is the UUID for the Parameter which can be looked up in the symbol map described in the map_elements outer PARAMETER_EXPR payload. If the type value is s this marks the start of a a new recursive section for a nested ParameterExpression. For example, in the following snippet there is an inner expr contained in final_expr, constituting a nested expression:

from qiskit.circuit import Parameter
 
x = Parameter("x")
y = Parameter("y")
z = Parameter("z")
 
expr = (x + y) / 2
final_expr = z**2 + expr

When s is encountered, this indicates that until an e` struct is reached, the next structs are used for a recursive definition. For both ``s and e types, the data values are not used, and always set to 0. The type value of u is used to represent a substitution call. This is only used for lhs_type and is always paired with an rhs_type of n. The data value is the size in bytes of a MAPPING encoded mapping of Parameter names to their value for the subs() call. The mapping data is immediately following the struct, and the next struct starts immediately after the mapping data.

Version 12

Version 12 adds support for:

• circuits containing memory-owning expr.Var variables.

Changes to HEADER

The HEADER struct for an individual circuit has added three uint32_t counts of the input, captured and locally declared variables in the circuit. The new form looks like:

struct {
    uint16_t name_size;
    char global_phase_type;
    uint16_t global_phase_size;
    uint32_t num_qubits;
    uint32_t num_clbits;
    uint64_t metadata_size;
    uint32_t num_registers;
    uint64_t num_instructions;
    uint32_t num_vars;
} HEADER_V12;

The HEADER_V12 struct is followed immediately by the same name, global-phase, metadata and register information as the V2 version of the header. Immediately following the registers is num_vars instances of EXPR_VAR_STANDALONE that define the variables in this circuit. After that, the data continues with custom definitions and instructions as in prior versions of QPY.

EXPR_VAR_DECLARATION

An EXPR_VAR_DECLARATION defines an expr.Var instance that is standalone; that is, it represents a self-owned memory location rather than wrapping a Clbit or ClassicalRegister. The payload is a C struct:

struct {
    char uuid_bytes[16];
    char usage;
    uint16_t name_size;
}

which is immediately followed by an EXPR_TYPE payload and then name_size bytes of UTF-8 encoding string data containing the name of the variable.

The char usage type code takes the following values:

Changes to EXPR_VAR

The EXPR_VAR variable has gained a new type code and payload, in addition to the pre-existing ones:

Notably, this new type-code indexes into pre-defined variables from the circuit header, rather than redefining the variable again in each location it is used.

Changes to EXPRESSION

The EXPRESSION type code has a new possible entry, i, corresponding to expr.Index nodes.

Version 11

Version 11 is identical to Version 10 except for the following. First, the names in the CUSTOM_INSTRUCTION blocks have a suffix of the form "_{uuid_hex}" where uuid_hex is a uuid hexadecimal string such as returned by UUID.hex. For example: "b3ecab5b4d6a4eb6bc2b2dbf18d83e1e". Second, it adds support for AnnotatedOperation objects. The base operation of an annotated operation is stored using the INSTRUCTION block, and an additional type value 'a'``is added to indicate that the custom instruction is an annotated operation. The list of modifiers are stored as instruction parameters using INSTRUCTION_PARAM, with an additional value ``'m' is added to indicate that the parameter is of type Modifier. Each modifier is stored using the MODIFIER struct.

MODIFIER

This represents Modifier

struct {
    char type;
    uint32_t num_ctrl_qubits;
    uint32_t ctrl_state;
    double power;
}

This is sufficient to store different types of modifiers required for serializing objects of type AnnotatedOperation. The field type is either 'i', 'c' or 'p', representing whether the modifier is respectively an inverse modifier, a control modifier or a power modifier. In the second case, the fields num_ctrl_qubits and ctrl_state specify the control logic of the base operation, and in the third case the field power represents the power of the base operation.

Version 10

Version 10 adds support for:

• symengine-native serialization for objects of type ParameterExpression as well as symbolic expressions in Pulse schedule blocks.
• new fields in the TranspileLayout class added in the Qiskit 0.45.0 release.

The symbolic_encoding field is added to the file header, and a new encoding type char is introduced, mapped to each symbolic library as follows: p refers to sympy encoding and e refers to symengine encoding.

Changes to FILE_HEADER

The contents of FILE_HEADER after V10 are defined as a C struct as:

struct {
    uint8_t qpy_version;
    uint8_t qiskit_major_version;
    uint8_t qiskit_minor_version;
    uint8_t qiskit_patch_version;
    uint64_t num_circuits;
    char symbolic_encoding;
} FILE_HEADER_V10;

Changes to LAYOUT

The LAYOUT struct is updated to have an additional input_qubit_count field. With version 10 the LAYOUT struct is now:

struct {
    char exists;
    int32_t initial_layout_size;
    int32_t input_mapping_size;
    int32_t final_layout_size;
    uint32_t extra_registers;
    int32_t input_qubit_count;
}

The rest of the layout data after the LAYOUT struct is represented as in previous versions. If input qubit_count is < 0 that indicates that both _input_qubit_count and _output_qubit_list in the TranspileLayout object are None.

Version 9

Version 9 adds support for classical Expr nodes and their associated Types.

EXPRESSION

An Expr node is represented by a stream of variable-width data. A node itself is represented by (in order in the byte stream):

1. a one-byte type code discriminator;
2. an EXPR_TYPE object;
3. a type-code-specific additional payload;
4. a type-code-specific number of child EXPRESSION payloads (the number of these is implied by the type code and not explicitly stored).

Each of these are described in the following table:

EXPR_TYPE

A Type is encoded by a single-byte ASCII char that encodes the kind of type, followed by a payload that varies depending on the type. The defined codes are:

EXPR_VAR

This represents a runtime variable of a Var node. These are a type code, followed by a type-code-specific payload:

EXPR_VALUE

This represents a literal object in the classical type system, such as an integer. Currently there are very few such literals. These are encoded as a type code, followed by a type-code-specific payload.

Changes to INSTRUCTION

To support the use of Expr nodes in the fields IfElseOp.condition, WhileLoopOp.condition and SwitchCaseOp.target, the INSTRUCTION struct is changed in an ABI compatible-manner to its previous definition. The new struct is the C struct:

struct {
    uint16_t name_size;
    uint16_t label_size;
    uint16_t num_parameters;
    uint32_t num_qargs;
    uint32_t num_cargs;
    uint8_t conditional_key;
    uint16_t conditional_reg_name_size;
    int64_t conditional_value;
    uint32_t num_ctrl_qubits;
    uint32_t ctrl_state;
}

where the only change is that a uint8_t conditional_key entry has replaced _Bool has_conditional. This new conditional_key takes the following numeric values, with these effects:

Changes to INSTRUCTION_PARAM

A new type code x is added that defines an EXPRESSION parameter.

Version 8

Version 8 adds support for handling a TranspileLayout stored in the QuantumCircuit.layout attribute. In version 8 immediately following the calibrations block at the end of the circuit payload there is now the LAYOUT struct. This struct outlines the size of the three attributes of a TranspileLayout class.

LAYOUT

struct {
    char exists;
    int32_t initial_layout_size;
    int32_t input_mapping_size;
    int32_t final_layout_size;
    uint32_t extra_registers;
}

If any of the signed values are -1 this indicates the corresponding attribute is None.

Immediately following the LAYOUT struct there is a REGISTERS struct for extra_registers (specifically the format introduced in Version 4) standalone register definitions that aren’t present in the circuit. Then there are initial_layout_size INITIAL_LAYOUT_BIT structs to define the TranspileLayout.initial_layout attribute.

INITIAL_LAYOUT_BIT

struct {
    int32_t index;
    int32_t register_size;
}

Where a value of -1 indicates None (as in no register is associated with the bit). Following each INITIAL_LAYOUT_BIT struct is register_size bytes for a utf8 encoded string for the register name.

Following the initial layout there is input_mapping_size array of uint32_t integers representing the positions of the physical bit from the initial layout. This enables constructing a list of virtual bits where the array index is its input mapping position.

Finally, there is an array of final_layout_size uint32_t integers. Each element is an index in the circuit’s qubits attribute which enables building a mapping from qubit starting position to the output position at the end of the circuit.

Version 7

Version 7 adds support for Reference instruction and serialization of a ScheduleBlock program while keeping its reference to subroutines:

from qiskit import pulse
from qiskit import qpy
 
with pulse.build() as schedule:
    pulse.reference("cr45p", "q0", "q1")
    pulse.reference("x", "q0")
    pulse.reference("cr45p", "q0", "q1")
 
with open('template_ecr.qpy', 'wb') as fd:
    qpy.dump(schedule, fd)

The conventional SCHEDULE_BLOCK data model is preserved, but in version 7 it is immediately followed by an extra MAPPING utf8 bytes block representing the data of the referenced subroutines.

New type key character is added to the SCHEDULE_BLOCK_INSTRUCTIONS group for the Reference instruction.

• y: Reference instruction

New type key character is added to the SCHEDULE_BLOCK_OPERANDS group for the operands of Reference instruction, which is a tuple of strings, e.g. (“cr45p”, “q0”, “q1”).

• o: string (operand string)

Note that this is the same encoding with the built-in Python string, however, the standard value encoding in QPY uses s type character for string data, which conflicts with the SymbolicPulse in the scope of pulse instruction operands. A special type character o is reserved for the string data that appears in the pulse instruction operands.

In addition, version 7 adds two new type keys to the INSTRUCTION_PARM struct. "d" is followed by no data and represents the literal value CASE_DEFAULT for switch-statement support. "R" represents a ClassicalRegister or Clbit, and is followed by the same format as the description of register or classical bit as used in the first element of the condition of an INSTRUCTION field.

Version 6

Version 6 adds support for ScalableSymbolicPulse. These objects are saved and read like SymbolicPulse objects, and the class name is added to the data to correctly handle the class selection.

SymbolicPulse block now starts with SYMBOLIC_PULSE_V2 header:

struct {
    uint16_t class_name_size;
    uint16_t type_size;
    uint16_t envelope_size;
    uint16_t constraints_size;
    uint16_t valid_amp_conditions_size;
    _bool amp_limited;
}

The only change compared to Version 5 is the addition of class_name_size. The header is then immediately followed by class_name_size utf8 bytes with the name of the class. Currently, either SymbolicPulse or ScalableSymbolicPulse are supported. The rest of the data is then identical to Version 5.

Version 5

Version 5 changes from Version 4 by adding support for ScheduleBlock and changing two payloads the INSTRUCTION metadata payload and the CUSTOM_INSTRUCTION block. These now have new fields to better account for ControlledGate objects in a circuit. In addition, new payload MAP_ITEM is defined to implement the MAPPING block.

With the support of ScheduleBlock, now QuantumCircuit can be serialized together with calibrations, or Pulse Gates. In QPY version 5 and above, CIRCUIT_CALIBRATIONS payload is packed after the INSTRUCTIONS block.

In QPY version 5 and above,

struct {
    char type;
}

immediately follows the file header block to represent the program type stored in the file.

• When type==c, QuantumCircuit payload follows
• When type==s, ScheduleBlock payload follows

SCHEDULE_BLOCK

ScheduleBlock is first supported in QPY Version 5. This allows users to save pulse programs in the QPY binary format as follows:

from qiskit import pulse, qpy
 
with pulse.build() as schedule:
    pulse.play(pulse.Gaussian(160, 0.1, 40), pulse.DriveChannel(0))
 
with open('schedule.qpy', 'wb') as fd:
    qpy.dump(schedule, fd)
 
with open('schedule.qpy', 'rb') as fd:
    new_schedule = qpy.load(fd)[0]

Note that circuit and schedule block are serialized and deserialized through the same QPY interface. Input data type is implicitly analyzed and no extra option is required to save the schedule block.

SCHEDULE_BLOCK_HEADER

ScheduleBlock block starts with the following header:

struct {
    uint16_t name_size;
    uint64_t metadata_size;
    uint16_t num_element;
}

which is immediately followed by name_size utf8 bytes of schedule name and metadata_size utf8 bytes of the JSON serialized metadata dictionary attached to the schedule.

SCHEDULE_BLOCK_ALIGNMENTS

Then, alignment context of the schedule block starts with char representing the supported context type followed by the SEQUENCE block representing the parameters associated with the alignment context AlignmentKind._context_params. The context type char is mapped to each alignment subclass as follows:

• l: AlignLeft
• r: AlignRight
• s: AlignSequential
• e: AlignEquispaced

Note that AlignFunc context is not supported because of the callback function stored in the context parameters.

SCHEDULE_BLOCK_INSTRUCTIONS

This alignment block is further followed by num_element length of block elements which may consist of nested schedule blocks and schedule instructions. Each schedule instruction starts with char representing the instruction type followed by the SEQUENCE block representing the instruction operands. Note that the data structure of pulse Instruction is unified so that instance can be uniquely determined by the class and a tuple of operands. The mapping of type char to the instruction subclass is defined as follows:

• a: Acquire instruction
• p: Play instruction
• d: Delay instruction
• f: SetFrequency instruction
• g: ShiftFrequency instruction
• q: SetPhase instruction
• r: ShiftPhase instruction
• b: RelativeBarrier instruction
• t: TimeBlockade instruction
• y: Reference instruction (new in version 0.7)

SCHEDULE_BLOCK_OPERANDS

The operands of these instances can be serialized through the standard QPY value serialization mechanism, however there are special object types that only appear in the schedule operands. Since the operands are serialized as SEQUENCE, each element must be packed with the INSTRUCTION_PARAM pack struct, where each payload starts with a header block consists of the char type and uint64_t size. Special objects start with the following type key:

• c: Channel
• w: Waveform
• s: SymbolicPulse
• o: string (operand string, new in version 0.7)

CHANNEL

Channel block starts with channel subtype char that maps an object data to Channel subclass. Mapping is defined as follows:

• d: DriveChannel
• c: ControlChannel
• m: MeasureChannel
• a: AcquireChannel
• e: MemorySlot
• r: RegisterSlot

The key is immediately followed by the channel index serialized as the INSTRUCTION_PARAM.

Waveform

Waveform block starts with WAVEFORM header:

struct {
    double epsilon;
    uint32_t data_size;
    _bool amp_limited;
}

which is followed by data_size bytes of complex ndarray binary generated by numpy.save. This represents the complex IQ data points played on a quantum device. name is saved after the samples in the INSTRUCTION_PARAM pack struct, which can be string or None.

SymbolicPulse

SymbolicPulse block starts with SYMBOLIC_PULSE header:

struct {
    uint16_t type_size;
    uint16_t envelope_size;
    uint16_t constraints_size;
    uint16_t valid_amp_conditions_size;
    _bool amp_limited;
}

which is followed by type_size utf8 bytes of SymbolicPulse.pulse_type string that represents a class of waveform, such as “Gaussian” or “GaussianSquare”. Then, envelope_size, constraints_size, valid_amp_conditions_size utf8 bytes of serialized symbolic expressions are generated for SymbolicPulse.envelope, SymbolicPulse.constraints, and SymbolicPulse.valid_amp_conditions, respectively. Since string representation of these expressions are usually lengthy, the expression binary is generated by the python zlib module with data compression.

To uniquely specify a pulse instance, we also need to store the associated parameters, which consist of duration and the rest of parameters as a dictionary. Dictionary parameters are first dumped in the MAPPING form, and then duration is dumped with the INSTRUCTION_PARAM pack struct. Lastly, name is saved also with the INSTRUCTION_PARAM pack struct, which can be string or None.

MAPPING

The MAPPING is a representation for arbitrary mapping object. This is a fixed length SEQUENCE of key-value pair represented by the MAP_ITEM payload.

A MAP_ITEM starts with a header defined as:

struct {
    uint16_t key_size;
    char type;
    uint16_t size;
}

which is immediately followed by the key_size utf8 bytes representing the dictionary key in string and size utf8 bytes of arbitrary object data of QPY serializable type.

CIRCUIT_CALIBRATIONS

The CIRCUIT_CALIBRATIONS block is a dictionary to define pulse calibrations of the custom instruction set. This block starts with the following CALIBRATION header:

struct {
    uint16_t num_cals;
}

which is followed by the num_cals length of calibration entries, each starts with the CALIBRATION_DEF header:

struct {
    uint16_t name_size;
    uint16_t num_qubits;
    uint16_t num_params;
    char type;
}

The calibration definition header is then followed by name_size utf8 bytes of the gate name, num_qubits length of integers representing a sequence of qubits, and num_params length of INSTRUCTION_PARAM payload for parameters associated to the custom instruction. The type indicates the class of pulse program which is either, in principle, ScheduleBlock or Schedule. As of QPY Version 5, only ScheduleBlock payload is supported. Finally, SCHEDULE_BLOCK payload is packed for each CALIBRATION_DEF entry.

INSTRUCTION

The INSTRUCTION block was modified to add two new fields num_ctrl_qubits and ctrl_state which are used to model the ControlledGate.num_ctrl_qubits and ControlledGate.ctrl_state attributes. The new payload packed struct format is:

struct {
    uint16_t name_size;
    uint16_t label_size;
    uint16_t num_parameters;
    uint32_t num_qargs;
    uint32_t num_cargs;
    _Bool has_conditional;
    uint16_t conditional_reg_name_size;
    int64_t conditional_value;
    uint32_t num_ctrl_qubits;
    uint32_t ctrl_state;
}

The rest of the instruction payload is the same. You can refer to INSTRUCTIONS for the details of the full payload.

CUSTOM_INSTRUCTION

The CUSTOM_INSTRUCTION block in QPY version 5 adds a new field base_gate_size which is used to define the size of the qiskit.circuit.Instruction object stored in the ControlledGate.base_gate attribute for a custom ControlledGate object. With this change the CUSTOM_INSTRUCTION metadata block becomes:

struct {
    uint16_t name_size;
    char type;
    uint32_t num_qubits;
    uint32_t num_clbits;
    _Bool custom_definition;
    uint64_t size;
    uint32_t num_ctrl_qubits;
    uint32_t ctrl_state;
    uint64_t base_gate_size
}

Immediately following the CUSTOM_INSTRUCTION struct is the utf8 encoded name of size name_size.

If custom_definition is True that means that the immediately following size bytes contains a QPY circuit data which can be used for the custom definition of that gate. If custom_definition is False then the instruction can be considered opaque (ie no definition). The type field determines what type of object will get created with the custom definition. If it’s 'g' it will be a Gate object, 'i' it will be a Instruction object.

Following this the next base_gate_size bytes contain the INSTRUCTION payload for the ControlledGate.base_gate.

Additionally an addition value for type is added 'c' which is used to indicate the custom instruction is a custom ControlledGate.

Version 4

Version 4 is identical to Version 3 except that it adds 2 new type strings to the INSTRUCTION_PARAM struct, z to represent None (which is encoded as no data), q to represent a QuantumCircuit (which is encoded as a QPY circuit), r to represent a range of integers (which is encoded as a RANGE), and t to represent a sequence (which is encoded as defined by SEQUENCE). Additionally, version 4 changes the type of register index mapping array from uint32_t to int64_t. If the values of any of the array elements are negative they represent a register bit that is not present in the circuit.

The REGISTERS header format has also been updated to

struct {
    char type;
    _Bool standalone;
    uint32_t size;
    uint16_t name_size;
    _bool in_circuit;
}

which just adds the in_circuit field which represents whether the register is part of the circuit or not.

RANGE

A RANGE is a representation of a range object. It is defined as:

struct {
    int64_t start;
    int64_t stop;
    int64_t step;
}

SEQUENCE

A SEQUENCE is a representation of an arbitrary sequence object. As sequence are just fixed length containers of arbitrary python objects their QPY can’t fully represent any sequence, but as long as the contents in a sequence are other QPY serializable types for the INSTRUCTION_PARAM payload the sequence object can be serialized.

A sequence instruction parameter starts with a header defined as:

struct {
    uint64_t size;
}

followed by size elements that are INSTRUCTION_PARAM payloads, where each of these define an element in the sequence. The sequence object will be typecasted into proper type, e.g. tuple, afterwards.

Version 3

Version 3 of the QPY format is identical to Version 2 except that it defines a struct format to represent a PauliEvolutionGate natively in QPY. To accomplish this the CUSTOM_DEFINITIONS struct now supports a new type value 'p' to represent a PauliEvolutionGate. Enties in the custom instructions tables have unique name generated that start with the string "###PauliEvolutionGate_" followed by a uuid string. This gate name is reservered in QPY and if you have a custom Instruction object with a definition set and that name prefix it will error. If it’s of type 'p' the data payload is defined as follows:

PAULI_EVOLUTION

This represents the high level PauliEvolutionGate

struct {
    uint64_t operator_count;
    _Bool standalone_op;
    char time_type;
    uint64_t time_size;
    uint64_t synthesis_size;
}

This is immediately followed by operator_count elements defined by the SPARSE_PAULI_OP_LIST_ELEM payload. Following that we have time_size bytes representing the time attribute. If standalone_op is True then there must only be a single operator. The encoding of these bytes is determined by the value of time_type. Possible values of time_type are 'f', 'p', and 'e'. If time_type is 'f' it’s a double, 'p' defines a Parameter object which is represented by a PARAMETER, e defines a ParameterExpression object (that’s not a Parameter) which is represented by a PARAMETER_EXPR. Following that is synthesis_size bytes which is a utf8 encoded json payload representing the EvolutionSynthesis class used by the gate.

SPARSE_PAULI_OP_LIST_ELEM

This represents an instance of SparsePauliOp.

struct {
    uint32_t pauli_op_size;
}

which is immediately followed by pauli_op_size bytes which are .npy format [2] data which represents the SparsePauliOp.

Version 3 of the QPY format also defines a struct format to represent a ParameterVectorElement as a distinct subclass from a Parameter. This adds a new parameter type char 'v' to represent a ParameterVectorElement which is now supported as a type string value for an INSTRUCTION_PARAM. The payload for these parameters are defined below as PARAMETER_VECTOR_ELEMENT.

PARAMETER_VECTOR_ELEMENT

A PARAMETER_VECTOR_ELEMENT represents a ParameterVectorElement object the data for a INSTRUCTION_PARAM. The contents of the PARAMETER_VECTOR_ELEMENT are defined as:

struct {
    uint16_t vector_name_size;
    uint64_t vector_size;
    char uuid[16];
    uint64_t index;
}

which is immediately followed by vector_name_size utf8 bytes representing the parameter’s vector name.

PARAMETER_EXPR

Additionally, since QPY format version v3 distinguishes between a Parameter and ParameterVectorElement the payload for a ParameterExpression needs to be updated to distinguish between the types. The following is the modified payload format which is mostly identical to the format in Version 1 and Version 2 but just modifies the map_elements struct to include a symbol type field.

A PARAMETER_EXPR represents a ParameterExpression object that the data for an INSTRUCTION_PARAM. The contents of a PARAMETER_EXPR are defined as:

struct {
    uint64_t map_elements;
    uint64_t expr_size;
}

Immediately following the header is expr_size bytes of utf8 data containing the expression string, which is the sympy srepr of the expression for the parameter expression. Following that is a symbol map which contains map_elements elements with the format

struct {
    char symbol_type;
    char type;
    uint64_t size;
}

The symbol_type key determines the payload type of the symbol representation for the element. If it’s p it represents a Parameter and if it’s v it represents a ParameterVectorElement. The map element struct is immediately followed by the symbol map key payload, if symbol_type is p then it is followed immediately by a PARAMETER object (both the struct and utf8 name bytes) and if symbol_type is v then the struct is imediately followed by PARAMETER_VECTOR_ELEMENT (both the struct and utf8 name bytes). That is followed by size bytes for the data of the symbol. The data format is dependent on the value of type. If type is p then it represents a Parameter and size will be 0, the value will just be the same as the key. Similarly if the type is v then it represents a ParameterVectorElement and size will be 0 as the value will just be the same as the key. If type is f then it represents a double precision float. If type is c it represents a double precision complex, which is represented by the COMPLEX. Finally, if type is i it represents an integer which is an int64_t.

Version 2

Version 2 of the QPY format is identical to version 1 except for the HEADER section is slightly different. You can refer to the Version 1 section for the details on the rest of the payload format.

HEADER

The contents of HEADER are defined as a C struct are:

struct {
    uint16_t name_size;
    char global_phase_type;
    uint16_t global_phase_size;
    uint32_t num_qubits;
    uint32_t num_clbits;
    uint64_t metadata_size;
    uint32_t num_registers;
    uint64_t num_instructions;
}

This is immediately followed by name_size bytes of utf8 data for the name of the circuit. Following this is immediately global_phase_size bytes representing the global phase. The content of that data is dictated by the value of global_phase_type. If it’s 'f' the data is a float and is the size of a double. If it’s 'p' defines a Parameter object which is represented by a PARAM struct (see below), e defines a ParameterExpression object (that’s not a Parameter) which is represented by a PARAM_EXPR struct (see below).

Version 1

HEADER

The contents of HEADER as defined as a C struct are:

struct {
    uint16_t name_size;
    double global_phase;
    uint32_t num_qubits;
    uint32_t num_clbits;
    uint64_t metadata_size;
    uint32_t num_registers;
    uint64_t num_instructions;
}

This is immediately followed by name_size bytes of utf8 data for the name of the circuit.

METADATA

The METADATA field is a UTF8 encoded JSON string. After reading the HEADER (which is a fixed size at the start of the QPY file) and the name string you then read the metadata_size number of bytes and parse the JSON to get the metadata for the circuit.

REGISTERS

The contents of REGISTERS is a number of REGISTER object. If num_registers is > 0 then after reading METADATA you read that number of REGISTER structs defined as:

struct {
    char type;
    _Bool standalone;
    uint32_t size;
    uint16_t name_size;
}

type can be 'q' or 'c'.

Immediately following the REGISTER struct is the utf8 encoded register name of size name_size. After the name utf8 bytes there is then an array of int64_t values of size size that contains a map of the register’s index to the circuit’s qubit index. For example, array element 0’s value is the index of the register[0]’s position in the containing circuit’s qubits list.

The standalone boolean determines whether the register is constructed as a standalone register that was added to the circuit or was created from existing bits. A register is considered standalone if it has bits constructed solely as part of it, for example:

qr = QuantumRegister(2)
qc = QuantumCircuit(qr)

the register qr would be a standalone register. While something like:

bits = [Qubit(), Qubit()]
qr2 = QuantumRegister(bits=bits)
qc = QuantumCircuit(qr2)

qr2 would have standalone set to False.

CUSTOM_DEFINITIONS

This section specifies custom definitions for any of the instructions in the circuit.

CUSTOM_DEFINITION_HEADER contents are defined as:

struct {
    uint64_t size;
}

If size is greater than 0 that means the circuit contains custom instruction(s). Each custom instruction is defined with a CUSTOM_INSTRUCTION block defined as:

struct {
    uint16_t name_size;
    char type;
    uint32_t num_qubits;
    uint32_t num_clbits;
    _Bool custom_definition;
    uint64_t size;
}

Immediately following the CUSTOM_INSTRUCTION struct is the utf8 encoded name of size name_size.

If custom_definition is True that means that the immediately following size bytes contains a QPY circuit data which can be used for the custom definition of that gate. If custom_definition is False then the instruction can be considered opaque (ie no definition). The type field determines what type of object will get created with the custom definition. If it’s 'g' it will be a Gate object, 'i' it will be a Instruction object.

INSTRUCTIONS

The contents of INSTRUCTIONS is a list of INSTRUCTION metadata objects

struct {
    uint16_t name_size;
    uint16_t label_size;
    uint16_t num_parameters;
    uint32_t num_qargs;
    uint32_t num_cargs;
    _Bool has_conditional;
    uint16_t conditional_reg_name_size;
    int64_t conditional_value;
}

This metadata object is immediately followed by name_size bytes of utf8 bytes for the name. name here is the Qiskit class name for the Instruction class if it’s defined in Qiskit. Otherwise it falls back to the custom instruction name. Following the name bytes there are label_size bytes of utf8 data for the label if one was set on the instruction. Following the label bytes if has_conditional is True then there are conditional_reg_name_size bytes of utf8 data for the name of the conditional register name. In case of single classical bit conditions the register name utf8 data will be prefixed with a null character “x00” and then a utf8 string integer representing the classical bit index in the circuit that the condition is on.

This is immediately followed by the INSTRUCTION_ARG structs for the list of arguments of that instruction. These are in the order of all quantum arguments (there are num_qargs of these) followed by all classical arguments (num_cargs of these).

The contents of each INSTRUCTION_ARG is:

struct {
    char type;
    uint32_t index;
}

type can be 'q' or 'c'.

After all arguments for an instruction the parameters are specified with num_parameters INSTRUCTION_PARAM structs.

The contents of each INSTRUCTION_PARAM is:

struct {
    char type;
    uint64_t size;
}

After each INSTRUCTION_PARAM the next size bytes are the parameter’s data. The type field can be 'i', 'f', 'p', 'e', 's', 'c' or 'n' which dictate the format. For 'i' it’s an integer, 'f' it’s a double, 's' if it’s a string (encoded as utf8), 'c' is a complex and the data is represented by the struct format in the PARAMETER_EXPR section. 'p' defines a Parameter object which is represented by a PARAMETER struct, e defines a ParameterExpression object (that’s not a Parameter) which is represented by a PARAMETER_EXPR struct (on QPY format Version 3 the format is tweak slightly see: PARAMETER_EXPR), 'n' represents an object from numpy (either an ndarray or a numpy type) which means the data is .npy format [2] data, and in QPY Version 3 'v' represents a ParameterVectorElement which is represented by a PARAMETER_VECTOR_ELEMENT struct.

PARAMETER

A PARAMETER represents a Parameter object the data for a INSTRUCTION_PARAM. The contents of the PARAMETER are defined as:

struct {
    uint16_t name_size;
    char uuid[16];
}

which is immediately followed by name_size utf8 bytes representing the parameter name.

PARAMETER_EXPR

A PARAMETER_EXPR represents a ParameterExpression object that the data for an INSTRUCTION_PARAM. The contents of a PARAMETER_EXPR are defined as:

The PARAMETER_EXPR data starts with a header:

struct {
    uint64_t map_elements;
    uint64_t expr_size;
}

Immediately following the header is expr_size bytes of utf8 data containing the expression string, which is the sympy srepr of the expression for the parameter expression. Follwing that is a symbol map which contains map_elements elements with the format

struct {
    char type;
    uint64_t size;
}

Which is followed immediately by PARAMETER object (both the struct and utf8 name bytes) for the symbol map key. That is followed by size bytes for the data of the symbol. The data format is dependent on the value of type. If type is p then it represents a Parameter and size will be 0, the value will just be the same as the key. If type is f then it represents a double precision float. If type is c it represents a double precision complex, which is represented by COMPLEX. Finally, if type is i it represents an integer which is an int64_t.

COMPLEX

When representing a double precision complex value in QPY the following struct is used:

struct {
    double real;
    double imag;
}

this matches the internal C representation of Python’s complex type. [3]

References

[1]

https://tools.ietf.org/html/rfc1700

[2] (1,2)

https://numpy.org/doc/stable/reference/generated/numpy.lib.format.html

[3]

https://docs.python.org/3/c-api/complex.html#c.Py_complex