Add support for I2 element type by anikulk · Pull Request #34956 · openvinotoolkit/openvino

anikulk · 2026-03-26T12:36:45Z

Add i2 (signed 2-bit integer) element type to core registration
- Add Type_t::i2 to the enum in element_type.hpp
- Add inline constexpr Type i2(Type_t::i2) constant
- Add i2_aliases and type_info entry (2-bit, signed, quantized)
- Add i2 to the EnumNames string lookup map
Add i2 element type iterator and template backend support
- element_type_traits.hpp: add element_type_traits with value_type int8_t
- element_iterator.hpp: add i2 to is_bit_type(), bit_width() = 2 specialization, sign-extension in operator value_type() for i2 alongside i4, and exclude i2 from the unsigned raw-cast operator
- template backend convert.cpp: add i2 cases to evaluate_by_input_type and evaluate_by_output_type dispatch switches
- template backend evaluate_node.hpp: add Type_t::i2 case to generic dispatch
- template backend subtract.cpp: add evaluate specialization (unpack i2->i8, subtract, repack i8->i2) and dispatch case, matching u2 pattern
Changes related to int2 unpacking in python bindings
Add support for in tflite FE
- Updated schema.fbs for int2

Details:

item1
...

Tickets:

ticket-id

AI Assistance:

AI assistance used: no / yes
If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).

- Add i2 (signed 2-bit integer) element type to core registration - Add Type_t::i2 to the enum in element_type.hpp - Add inline constexpr Type i2(Type_t::i2) constant - Add i2_aliases and type_info entry (2-bit, signed, quantized) - Add i2 to the EnumNames string lookup map - Add i2 element type iterator and template backend support - element_type_traits.hpp: add element_type_traits<i2> with value_type int8_t - element_iterator.hpp: add i2 to is_bit_type(), bit_width<i2>() = 2 specialization, sign-extension in operator value_type() for i2 alongside i4, and exclude i2 from the unsigned raw-cast operator - template backend convert.cpp: add i2 cases to evaluate_by_input_type and evaluate_by_output_type dispatch switches - template backend evaluate_node.hpp: add Type_t::i2 case to generic dispatch - template backend subtract.cpp: add evaluate<i2> specialization (unpack i2->i8, subtract, repack i8->i2) and dispatch case, matching u2 pattern - Changes related to int2 unpacking in python bindings - Add support for in tflite FE - Updated schema.fbs for int2 Signed-off-by: Anisha Kulkarni <anisha.dattatraya.kulkarni@intel.com>

Copilot

Pull request overview

Adds a new signed 2-bit integer element type (i2) across OpenVINO core type registration and multiple integration points (template backend, CPU plugin conversion, Python/C bindings, and TFLite FE), enabling model import, conversion, and basic evaluation paths to recognize and operate on i2.

Changes:

Register ov::element::i2 in core (Type_t, type info/aliases, string ↔ enum mapping, traits, iterator support, Constant/Convert handling).
Extend execution/evaluation support for i2 (template backend Convert/Subtract dispatch; CPU plugin cpu_convert 2-bit unpack path).
Wire i2 into external interfaces (Python helpers/bindings, C API mapping, TFLite schema + type mapping).

Reviewed changes

Copilot reviewed 20 out of 20 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
src/plugins/template/backend/ops/subtract.cpp	Adds `i2` specialization and dispatch for Subtract evaluation via i8 intermediate conversion.
src/plugins/template/backend/ops/evaluate_node.hpp	Adds `Type_t::i2` to generic evaluate dispatch.
src/plugins/template/backend/ops/convert.cpp	Adds `i2` to Convert evaluation dispatch switches.
src/plugins/intel_cpu/src/nodes/common/cpu_convert.cpp	Extends 2-bit conversion path to handle `i2` (signed unpack) alongside existing `u2`.
src/frontends/tensorflow_lite/src/utils.cpp	Maps TFLite `INT2` tensor type to `ov::element::i2`.
src/frontends/tensorflow_lite/src/schema/schema.fbs	Extends TFLite schema with `INT2` enum value.
src/frontends/tensorflow_lite/src/frontend.cpp	Adds new includes (context not shown indicating functional usage).
src/core/src/type/element_type.cpp	Registers `i2` aliases/type_info and adds string-to-enum name mapping.
src/core/src/pass/visualize_tree.cpp	Includes `i2` in Constant visualization handling set.
src/core/src/op/convert.cpp	Adds `i2` to Convert op supported element type lists.
src/core/src/op/constant.cpp	Adds `i2` to supported Constant types and range validation logic.
src/core/include/openvino/core/type/element_type_traits.hpp	Adds `element_type_traits<i2>` with `int8_t` value type.
src/core/include/openvino/core/type/element_type.hpp	Adds `Type_t::i2` and `inline constexpr Type i2`.
src/core/dev_api/openvino/core/type/element_iterator.hpp	Extends bit-type iterator utilities and signed conversion proxy to support `i2`.
src/common/transformations/src/transformations/convert_precision.cpp	Allows low-precision integer conversion utilities to include `i2`.
src/bindings/python/src/pyopenvino/graph/types/element_type.cpp	Exposes `Type.i2` in Python bindings.
src/bindings/python/src/pyopenvino/core/common.cpp	Maps `i2` to NumPy dtype and updates packed-stride handling; retains a duplicate `nf4` map entry.
src/bindings/python/src/openvino/helpers/packing.py	Adds `i2` to Python pack/unpack helpers (signed handling).
src/bindings/c/src/ov_tensor.cpp	Adds C API enum → OpenVINO type mapping for `I2`.
src/bindings/c/include/openvino/c/ov_common.h	Adds `I2` to C API element type enum.

Copilot · 2026-03-30T12:21:14Z

src/plugins/intel_cpu/src/nodes/common/cpu_convert.cpp

 #define INTEL_CPU_CVT_FROM_2BIT_LIST                                                                 \
-    INTEL_CPU_CVT(u2, f32), INTEL_CPU_CVT(u2, f16), INTEL_CPU_CVT(u2, bf16), INTEL_CPU_CVT(u2, i32), \
-        INTEL_CPU_CVT(u2, u8), INTEL_CPU_CVT(u2, i8)
+    INTEL_CPU_CVT(u2, f32), INTEL_CPU_CVT(u2, f16), INTEL_CPU_CVT(u2, bf16), INTEL_CPU_CVT(u2, u8), INTEL_CPU_CVT(u2, i8),                                                 \


[BLOCKER] INTEL_CPU_CVT_FROM_2BIT_LIST no longer includes a u2 -> i32 conversion entry, so cpu_convert() will hit OPENVINO_ASSERT(ctx.converted, ...) for that conversion path. Please restore the missing u2 conversion(s) (at least u2 -> i32, as previously supported) or otherwise ensure u2 can still be converted to i32 via the 2-bit dispatch.

Suggested change

INTEL_CPU_CVT(u2, f32), INTEL_CPU_CVT(u2, f16), INTEL_CPU_CVT(u2, bf16), INTEL_CPU_CVT(u2, u8), INTEL_CPU_CVT(u2, i8), \

INTEL_CPU_CVT(u2, f32), INTEL_CPU_CVT(u2, f16), INTEL_CPU_CVT(u2, bf16), INTEL_CPU_CVT(u2, u8), INTEL_CPU_CVT(u2, i8), INTEL_CPU_CVT(u2, i32), \

Copilot · 2026-03-30T12:21:14Z

src/bindings/python/src/openvino/helpers/packing.py

@@ -24,7 +24,7 @@ def pack_data(array: np.ndarray, type: Type) -> np.ndarray:

    :param array: numpy array with values to pack.
    :type array: numpy array
-    :param type: Type to interpret the array values. Type must be u1, u2, u3, u4, u6, i4, nf4 or f4e2m1.
+    :param type: Type to interpret the array values. Type must be u1, u2, u3, u4, u6, i2, i4, nf4 or f4e2m1.
    :type type: openvino.Type
    """
    # Handle u3 and u6 with special transposed packing
@@ -33,9 +33,9 @@ def pack_data(array: np.ndarray, type: Type) -> np.ndarray:
    elif type == Type.u6:
        return _pack_u6(array)

-    assert type in [Type.u1, Type.u2, Type.u4, Type.i4, Type.nf4, Type.f4e2m1], "Packing algorithm for the" "data types stored in 1, 2 or 4 bits"
+    assert type in [Type.u1, Type.u2, Type.u4, Type.i2, Type.i4, Type.nf4, Type.f4e2m1], "Packing algorithm for the" "data types stored in 1, 2 or 4 bits"

-    minimum_regular_dtype = np.int8 if type == Type.i4 else np.uint8
+    minimum_regular_dtype = np.int8 if type in (Type.i2, Type.i4) else np.uint8
    casted_to_regular_type = array.astype(dtype=minimum_regular_dtype, casting="unsafe")
    if not np.array_equal(casted_to_regular_type, array):
        raise RuntimeError(f'The conversion of array "{array}" to dtype' f' "{casted_to_regular_type}" results in rounding')


[HIGH] pack_data/unpack_data now support Type.i2, but the Python test suite doesn’t appear to cover i2 yet (existing parametrized packing/constant tests cover u1/u2/u3/u4/u6/i4/nf4). Please extend the relevant tests (e.g. test_constant_direct_packing and runtime tensor packing tests) to include Type.i2, including negative values to validate sign-extension and round-trip correctness against ops.constant(..., dtype=Type.i2) / Tensor(Type.i2, ...).

Copilot · 2026-03-30T12:21:15Z

src/bindings/python/src/pyopenvino/core/common.cpp

+        {ov::element::boolean, py::dtype("bool")},  {ov::element::u1, py::dtype("uint8")},
+        {ov::element::u2, py::dtype("uint8")},      {ov::element::u3, py::dtype("uint8")},
+        {ov::element::u4, py::dtype("uint8")},      {ov::element::u6, py::dtype("uint8")},
+        {ov::element::nf4, py::dtype("uint8")},     {ov::element::nf4, py::dtype("uint8")},


The ov_type_to_dtype_mapping initializer contains ov::element::nf4 twice. This duplicate key is redundant (only one entry will be kept) and makes the mapping harder to audit; please remove the duplicate entry.

Suggested change

{ov::element::nf4, py::dtype("uint8")}, {ov::element::nf4, py::dtype("uint8")},

{ov::element::nf4, py::dtype("uint8")},

Copilot · 2026-03-30T12:21:15Z

src/bindings/c/include/openvino/c/ov_common.h

 typedef enum {
    DYNAMIC = 0U,  //!< Dynamic element type
    BOOLEAN,       //!< boolean element type
    BF16,          //!< bf16 element type
    F16,           //!< f16 element type
    F32,           //!< f32 element type
    F64,           //!< f64 element type
+    I2,            //!< i2 element type
    I4,            //!< i4 element type
    I8,            //!< i8 element type


[BLOCKER] ov_element_type_e is documented as being aligned with ov::element::Type_t, but inserting I2 in the middle changes the underlying numeric values of all subsequent enumerators (I4, I8, …). This is an ABI/API break for any C/C++ user code (or binaries) that persists/passes these enum values by integer. If numeric stability is required, please preserve existing values by appending I2 at the end or by explicitly assigning fixed values to all enumerators while keeping the old ones unchanged (and make the same adjustment in ov::element::Type_t).

praasz self-assigned this Mar 27, 2026

anikulk force-pushed the i2 branch from 633aa03 to 8c4e8f2 Compare March 30, 2026 10:54

Merge branch 'master' into i2

f7332f4

anikulk marked this pull request as ready for review March 30, 2026 12:00

anikulk requested review from a team as code owners March 30, 2026 12:00

anikulk requested a review from Copilot March 30, 2026 12:11

Copilot started reviewing on behalf of anikulk March 30, 2026 12:12 View session

Copilot AI reviewed Mar 30, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add support for I2 element type#34956

Add support for I2 element type#34956
anikulk wants to merge 2 commits intoopenvinotoolkit:masterfrom
anikulk:i2

anikulk commented Mar 26, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Mar 30, 2026

Uh oh!

Copilot AI Mar 30, 2026

Uh oh!

Copilot AI Mar 30, 2026

Uh oh!

Copilot AI Mar 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

	INTEL_CPU_CVT(u2, f32), INTEL_CPU_CVT(u2, f16), INTEL_CPU_CVT(u2, bf16), INTEL_CPU_CVT(u2, u8), INTEL_CPU_CVT(u2, i8), \
	INTEL_CPU_CVT(u2, f32), INTEL_CPU_CVT(u2, f16), INTEL_CPU_CVT(u2, bf16), INTEL_CPU_CVT(u2, u8), INTEL_CPU_CVT(u2, i8), INTEL_CPU_CVT(u2, i32), \

	{ov::element::nf4, py::dtype("uint8")}, {ov::element::nf4, py::dtype("uint8")},
	{ov::element::nf4, py::dtype("uint8")},

Conversation

anikulk commented Mar 26, 2026

Details:

Tickets:

AI Assistance:

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Mar 30, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 30, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 30, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 30, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants