Add `float8_e8m0_fnu` (E8M0) OCP MX scale format #166

balancap · 2024-08-09T12:18:29Z

Adding the OCP MX scale format E8M0, which has the following properties:

Unsigned format;
8 exponent bits;
Exponent range from -127 to 127;
No zero and infinity;
Single NaN value (0xFF);

The E8M0 format is used in all MX block format definitions for scale representation (see https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf and issue #116).

ml_dtypes float8_base C++ class is extended to support floating point formats which are unsigned and with no zero (i.e. additional kIsSigned and kHasZero Traits properties).

Base on these traits, float8_e8m0_fnu has been implemented using the existing functionalities (convert, unary/binary ops, ...). Float8 Python unit tests have been extended to be able to cover unsigned floating point formats.

Points to discuss on the PR:

Naming: not sure about float8_e8m0_fnu. Maybe ufloat8_e8m0_fn could be better?
Testing: not 100% satisfied with additional branching/checks when extending unit tests for E8M0. Is there a cleaner way?
(Separate PR?) Additional optimization, as same convertion + ops (mul/div) could be specialized with a more efficient implementation.

hawkinsp · 2024-08-20T18:56:05Z

This looks good to me. There's a small failure in the test suite: please fix?

hawkinsp · 2024-08-20T18:56:23Z

(Sorry about the slow review.)

balancap · 2024-08-21T14:12:12Z

Thanks for looking at it, I'll check the failing tests in the next few days. In the meantime, do you have an opinion on the naming? Should we stick to the name in the PR?

balancap · 2024-09-03T10:12:15Z

@hawkinsp Apologies it took me a bit longer than expected. The CI issue should be fixed now (was just a trivial skipTest issue with Python 3.12). I have rebased the branch as well on the latest main

hawkinsp · 2024-09-10T20:49:57Z

This LGTM, but I'd appreciate @sergey-kozub's review, since he's in the process of adding the (complementary) MX data types in #181

hawkinsp · 2024-09-10T20:50:25Z

There's a test failure on mac os (different integer overflow behavior?)

sergey-kozub · 2024-09-11T08:30:22Z

Overall looks fine to me.

This PR touches many files also affected by #181, so I'd like to upstream that one first. I'll do the rebase to fix the merge failures.

ml_dtypes/tests/custom_float_test.py

balancap · 2024-09-12T13:30:34Z

@sergey-kozub @hawkinsp Removing 0 fixes the MacOS test run, but I am still investigating a crash on the Windows build (stack overflow).

hawkinsp · 2024-09-12T13:34:02Z

Ideally we'd like to make a new ml_dtypes release in the next few days, and it'd be great to get both of the MX dtype PRs in.

balancap · 2024-09-12T14:13:44Z

@hawkinsp Just pushed a fix commit for MacOS and Windows. Hopefully should be the last one if @sergey-kozub is good with the PR as it is.

hawkinsp · 2024-09-12T14:16:56Z

I'm trying to merge #181 first, which will probably create a merge conflict and you'll need to rebase on top of that PR. Hopefully it will be in soon.

hawkinsp · 2024-09-12T14:46:37Z

Ok, #181 is merged.

Would you:
a) squash your commits, please, and
b) rebase on top of head?

hawkinsp · 2024-09-12T14:56:18Z

Oh, it looks like you have a test failure on Python 3.13t. I doubt the "free-threading" part is important, but we are trying to make a Python 3.13 release.

balancap · 2024-09-12T15:07:51Z

I'll rebase & squash commits. On the Python 3.13 failure, we should probably skip this test as float8_e8m0 does not have a zero (or negative zero!). Just surprising it triggers on error only on Python 3.13

balancap · 2024-09-12T16:04:53Z

@hawkinsp I should be hopefully the last round of rebase & fix.

I added a skip on the test testHashZero, as E8M0 does not have a zero representation. I can dig a bit more into why the test is failing on Python 3.13, but I don't think it should block the PR merge.

Note: there is a new failure on Python 3.9 introduced by the previous MX float4/float6 PR. I am looking into it.

Adding the OCP MX scale format `E8M0`, which has the following properties: * Unsigned format; * 8 exponent bits; * Exponent range from -127 to 127; * No zero and infinity; * Single NaN value (0xFF); `ml_dtypes` `float8_base` C++ class is extended to support floating point formats which are unsigned and with no zero (i.e. additional `kIsSigned` and `kHasZero` Traits properties). Base on these traits, `float8_e8m0_fnu` has been implemented using the existing functionalities (convert, unary/binary ops, ...). Float8 Python unit tests have been extended to be able to cover unsigned floating point formats.

balancap · 2024-09-12T16:45:04Z

The trick cls._finfo_cache[dtype] = init.__func__() seems to do it for Python 3.9. CI all green on my fork, finally!

hawkinsp · 2024-09-12T16:46:22Z

What went wrong without init.__func__()? @sergey-kozub had that in his PR, but i had to revert it because it hit a type checker error. But I'm not sure what problem it's solving.

balancap · 2024-09-12T16:47:32Z

Python 3.9 does not seem to support the direct call to staticmethod (getting this error without __func__(): https://stackoverflow.com/questions/41921255/staticmethod-object-is-not-callable)

The merge commit on the main branch has this issue: https://github.com/jax-ml/ml_dtypes/actions/runs/10833234151/job/30059511975

Imported from GitHub PR #19096 This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - jax-ml/ml_dtypes#181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - jax-ml/ml_dtypes#166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 The PR is split into multiple commits just to make the review easier, it is possible that some tests could fail if only some (i.e. not all) of these commits are applied. Copybara import of the project: -- f493e48 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: import mxfloat.h -- 87d0056 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: primitive type -- 70ca820 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: literal support -- c479f09 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: conversion codegen -- daaa3af by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: python interface -- 1f0e19f by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: FFI -- 999bf96 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: HLO evaluator -- d7d5af7 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: add tests -- 9e8c7bc by Sergey Kozub <skozub@nvidia.com>: Add F8E8M0FNU type -- 1e34417 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments -- d4de0a3 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments (round 2) Merging this change closes #19096 FUTURE_COPYBARA_INTEGRATE_REVIEW=#19096 from openxla:skozub/e2m1 d4de0a3 PiperOrigin-RevId: 707638099

Imported from GitHub PR openxla/xla#19096 This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](openxla/xla#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - jax-ml/ml_dtypes#181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - jax-ml/ml_dtypes#166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 The PR is split into multiple commits just to make the review easier, it is possible that some tests could fail if only some (i.e. not all) of these commits are applied. Copybara import of the project: -- f493e4803eaa5ff3da3ceb130e9348c014b4a2e8 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: import mxfloat.h -- 87d005630b310a355d7c30b22828c35237373f17 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: primitive type -- 70ca82093faeec98f2dc5e8b82f617d99ca96849 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: literal support -- c479f0940da490e9668e2f48e14a7466f0c4a97f by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: conversion codegen -- daaa3af3ce3af456f2ef44dbc291ebeb09e86d9b by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: python interface -- 1f0e19ff14733eff790726936b68ef0cf607a766 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: FFI -- 999bf96092e57c7b3039811f2887281f347ff17a by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: HLO evaluator -- d7d5af74c5f8a94522779a121c0a4a962156fb64 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: add tests -- 9e8c7bc02849f241d0f05941221d99f1d08d9e67 by Sergey Kozub <skozub@nvidia.com>: Add F8E8M0FNU type -- 1e344174b931cea4978770ab740dfed67186c2f4 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments -- d4de0a369d9dc853f34f3cf3bf7dcc5a47502106 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments (round 2) Merging this change closes #19096 FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#19096 from openxla:skozub/e2m1 d4de0a369d9dc853f34f3cf3bf7dcc5a47502106 PiperOrigin-RevId: 707638099

Imported from GitHub PR #19096 This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - jax-ml/ml_dtypes#181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - jax-ml/ml_dtypes#166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 The PR is split into multiple commits just to make the review easier, it is possible that some tests could fail if only some (i.e. not all) of these commits are applied. Copybara import of the project: -- f493e48 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: import mxfloat.h -- 87d0056 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: primitive type -- 70ca820 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: literal support -- c479f09 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: conversion codegen -- daaa3af by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: python interface -- 1f0e19f by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: FFI -- 999bf96 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: HLO evaluator -- d7d5af7 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: add tests -- 9e8c7bc by Sergey Kozub <skozub@nvidia.com>: Add F8E8M0FNU type -- 1e34417 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments -- d4de0a3 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments (round 2) Merging this change closes #19096 FUTURE_COPYBARA_INTEGRATE_REVIEW=#19096 from openxla:skozub/e2m1 d4de0a3 PiperOrigin-RevId: 707638099

Imported from GitHub PR openxla/xla#19096 This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](openxla/xla#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - jax-ml/ml_dtypes#181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - jax-ml/ml_dtypes#166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 The PR is split into multiple commits just to make the review easier, it is possible that some tests could fail if only some (i.e. not all) of these commits are applied. Copybara import of the project: -- f493e4803eaa5ff3da3ceb130e9348c014b4a2e8 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: import mxfloat.h -- 87d005630b310a355d7c30b22828c35237373f17 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: primitive type -- 70ca82093faeec98f2dc5e8b82f617d99ca96849 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: literal support -- c479f0940da490e9668e2f48e14a7466f0c4a97f by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: conversion codegen -- daaa3af3ce3af456f2ef44dbc291ebeb09e86d9b by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: python interface -- 1f0e19ff14733eff790726936b68ef0cf607a766 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: FFI -- 999bf96092e57c7b3039811f2887281f347ff17a by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: HLO evaluator -- d7d5af74c5f8a94522779a121c0a4a962156fb64 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: add tests -- 9e8c7bc02849f241d0f05941221d99f1d08d9e67 by Sergey Kozub <skozub@nvidia.com>: Add F8E8M0FNU type -- 1e344174b931cea4978770ab740dfed67186c2f4 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments -- d4de0a369d9dc853f34f3cf3bf7dcc5a47502106 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments (round 2) Merging this change closes #19096 FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#19096 from openxla:skozub/e2m1 d4de0a369d9dc853f34f3cf3bf7dcc5a47502106 PiperOrigin-RevId: 707638099

Imported from GitHub PR #19096 This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - jax-ml/ml_dtypes#181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - jax-ml/ml_dtypes#166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 The PR is split into multiple commits just to make the review easier, it is possible that some tests could fail if only some (i.e. not all) of these commits are applied. Copybara import of the project: -- f493e48 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: import mxfloat.h -- 87d0056 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: primitive type -- 70ca820 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: literal support -- c479f09 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: conversion codegen -- daaa3af by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: python interface -- 1f0e19f by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: FFI -- 999bf96 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: HLO evaluator -- d7d5af7 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: add tests -- 9e8c7bc by Sergey Kozub <skozub@nvidia.com>: Add F8E8M0FNU type -- 1e34417 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments -- d4de0a3 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments (round 2) Merging this change closes #19096 FUTURE_COPYBARA_INTEGRATE_REVIEW=#19096 from openxla:skozub/e2m1 d4de0a3 PiperOrigin-RevId: 707638099

Imported from GitHub PR openxla/xla#19096 This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](openxla/xla#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - jax-ml/ml_dtypes#181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - jax-ml/ml_dtypes#166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 The PR is split into multiple commits just to make the review easier, it is possible that some tests could fail if only some (i.e. not all) of these commits are applied. Copybara import of the project: -- f493e4803eaa5ff3da3ceb130e9348c014b4a2e8 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: import mxfloat.h -- 87d005630b310a355d7c30b22828c35237373f17 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: primitive type -- 70ca82093faeec98f2dc5e8b82f617d99ca96849 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: literal support -- c479f0940da490e9668e2f48e14a7466f0c4a97f by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: conversion codegen -- daaa3af3ce3af456f2ef44dbc291ebeb09e86d9b by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: python interface -- 1f0e19ff14733eff790726936b68ef0cf607a766 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: FFI -- 999bf96092e57c7b3039811f2887281f347ff17a by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: HLO evaluator -- d7d5af74c5f8a94522779a121c0a4a962156fb64 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: add tests -- 9e8c7bc02849f241d0f05941221d99f1d08d9e67 by Sergey Kozub <skozub@nvidia.com>: Add F8E8M0FNU type -- 1e344174b931cea4978770ab740dfed67186c2f4 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments -- d4de0a369d9dc853f34f3cf3bf7dcc5a47502106 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments (round 2) Merging this change closes #19096 FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#19096 from openxla:skozub/e2m1 d4de0a369d9dc853f34f3cf3bf7dcc5a47502106 PiperOrigin-RevId: 707638099

Imported from GitHub PR openxla/xla#19096 This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](openxla/xla#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - #181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - #166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 The PR is split into multiple commits just to make the review easier, it is possible that some tests could fail if only some (i.e. not all) of these commits are applied. Copybara import of the project: -- f493e4803eaa5ff3da3ceb130e9348c014b4a2e8 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: import mxfloat.h -- 87d005630b310a355d7c30b22828c35237373f17 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: primitive type -- 70ca82093faeec98f2dc5e8b82f617d99ca96849 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: literal support -- c479f0940da490e9668e2f48e14a7466f0c4a97f by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: conversion codegen -- daaa3af3ce3af456f2ef44dbc291ebeb09e86d9b by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: python interface -- 1f0e19ff14733eff790726936b68ef0cf607a766 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: FFI -- 999bf96092e57c7b3039811f2887281f347ff17a by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: HLO evaluator -- d7d5af74c5f8a94522779a121c0a4a962156fb64 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: add tests -- 9e8c7bc02849f241d0f05941221d99f1d08d9e67 by Sergey Kozub <skozub@nvidia.com>: Add F8E8M0FNU type -- 1e344174b931cea4978770ab740dfed67186c2f4 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments -- d4de0a369d9dc853f34f3cf3bf7dcc5a47502106 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments (round 2) Merging this change closes #19096 PiperOrigin-RevId: 708390061

Imported from GitHub PR openxla/xla#19096 This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](openxla/xla#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - jax-ml/ml_dtypes#181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - jax-ml/ml_dtypes#166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 The PR is split into multiple commits just to make the review easier, it is possible that some tests could fail if only some (i.e. not all) of these commits are applied. Copybara import of the project: -- f493e4803eaa5ff3da3ceb130e9348c014b4a2e8 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: import mxfloat.h -- 87d005630b310a355d7c30b22828c35237373f17 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: primitive type -- 70ca82093faeec98f2dc5e8b82f617d99ca96849 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: literal support -- c479f0940da490e9668e2f48e14a7466f0c4a97f by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: conversion codegen -- daaa3af3ce3af456f2ef44dbc291ebeb09e86d9b by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: python interface -- 1f0e19ff14733eff790726936b68ef0cf607a766 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: FFI -- 999bf96092e57c7b3039811f2887281f347ff17a by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: HLO evaluator -- d7d5af74c5f8a94522779a121c0a4a962156fb64 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: add tests -- 9e8c7bc02849f241d0f05941221d99f1d08d9e67 by Sergey Kozub <skozub@nvidia.com>: Add F8E8M0FNU type -- 1e344174b931cea4978770ab740dfed67186c2f4 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments -- d4de0a369d9dc853f34f3cf3bf7dcc5a47502106 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments (round 2) Merging this change closes #19096 PiperOrigin-RevId: 708390061

Imported from GitHub PR #19096 This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - jax-ml/ml_dtypes#181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - jax-ml/ml_dtypes#166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 The PR is split into multiple commits just to make the review easier, it is possible that some tests could fail if only some (i.e. not all) of these commits are applied. Copybara import of the project: -- f493e48 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: import mxfloat.h -- 87d0056 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: primitive type -- 70ca820 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: literal support -- c479f09 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: conversion codegen -- daaa3af by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: python interface -- 1f0e19f by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: FFI -- 999bf96 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: HLO evaluator -- d7d5af7 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: add tests -- 9e8c7bc by Sergey Kozub <skozub@nvidia.com>: Add F8E8M0FNU type -- 1e34417 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments -- d4de0a3 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments (round 2) Merging this change closes #19096 COPYBARA_INTEGRATE_REVIEW=#19096 from openxla:skozub/e2m1 d4de0a3 PiperOrigin-RevId: 708390061

Imported from GitHub PR openxla/xla#19096 This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](openxla/xla#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - jax-ml/ml_dtypes#181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - jax-ml/ml_dtypes#166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 The PR is split into multiple commits just to make the review easier, it is possible that some tests could fail if only some (i.e. not all) of these commits are applied. Copybara import of the project: -- f493e4803eaa5ff3da3ceb130e9348c014b4a2e8 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: import mxfloat.h -- 87d005630b310a355d7c30b22828c35237373f17 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: primitive type -- 70ca82093faeec98f2dc5e8b82f617d99ca96849 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: literal support -- c479f0940da490e9668e2f48e14a7466f0c4a97f by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: conversion codegen -- daaa3af3ce3af456f2ef44dbc291ebeb09e86d9b by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: python interface -- 1f0e19ff14733eff790726936b68ef0cf607a766 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: FFI -- 999bf96092e57c7b3039811f2887281f347ff17a by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: HLO evaluator -- d7d5af74c5f8a94522779a121c0a4a962156fb64 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN type: add tests -- 9e8c7bc02849f241d0f05941221d99f1d08d9e67 by Sergey Kozub <skozub@nvidia.com>: Add F8E8M0FNU type -- 1e344174b931cea4978770ab740dfed67186c2f4 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments -- d4de0a369d9dc853f34f3cf3bf7dcc5a47502106 by Sergey Kozub <skozub@nvidia.com>: Addressing PR#19096 review comments (round 2) Merging this change closes #19096 PiperOrigin-RevId: 708390061

Imported from GitHub PR #21380 Previous PR #19096 was rolled back, re-trying. This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - jax-ml/ml_dtypes#181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - jax-ml/ml_dtypes#166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 Copybara import of the project: -- d7e00c4 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN and F8E8M0FNU types Merging this change closes #21380 FUTURE_COPYBARA_INTEGRATE_REVIEW=#21380 from openxla:skozub/e2m1_e8m0 d7e00c4 PiperOrigin-RevId: 715070992

Imported from GitHub PR openxla/xla#21380 Previous PR openxla/xla#19096 was rolled back, re-trying. This PR adds F4E2M1FN primitive type (4-bit float with 2 bits exponent and 1 bit mantissa), F8E8M0FNU primitive type (8-bit float with 8 bits exponent, no mantissa and no sign) and enables loads/stores in the same way S4/U4 type is implemented. This will enable using microscaling (MX) formats ([RFC](openxla/xla#18085)), such as MXFP4. ```c F4E2M1FN - Exponent bias: 1 - Maximum stored exponent value: 3 (binary 11) - Maximum unbiased exponent value: 3 - 1 = 2 - Minimum stored exponent value: 1 (binary 01) - Minimum unbiased exponent value: 1 − 1 = 0 - Has Positive and Negative zero - Doesn't have infinity - Doesn't have NaNs Additional details: - Zeros (+/-): S.00.0 - Max normal number: S.11.1 = ±2^(2) x (1 + 0.5) = ±6.0 - Min normal number: S.01.0 = ±2^(0) = ±1.0 - Min subnormal number: S.00.1 = ±2^(0) x 0.5 = ±0.5 F8E8M0FNU - Exponent bias: 127 - Maximum stored exponent value: 254 (binary 1111'1110) - Maximum unbiased exponent value: 254 - 127 = 127 - Minimum stored exponent value: 0 (binary 0000'0000) - Minimum unbiased exponent value: 0 − 127 = -127 - Doesn't have zero - Doesn't have infinity - NaN is encoded as binary 1111'1111 Additional details: - Zeros cannot be represented - Negative values cannot be represented - Mantissa is always 1 ``` Related PRs: - openxla/stablehlo#2582 - jax-ml/ml_dtypes#181 - llvm/llvm-project#95392 - llvm/llvm-project#108877 - jax-ml/ml_dtypes#166 - llvm/llvm-project#107127 - llvm/llvm-project#111028 Copybara import of the project: -- d7e00c49a4b4f26c06266d6bb941275e67464c01 by Sergey Kozub <skozub@nvidia.com>: Add F4E2M1FN and F8E8M0FNU types Merging this change closes #21380 FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#21380 from openxla:skozub/e2m1_e8m0 d7e00c49a4b4f26c06266d6bb941275e67464c01 PiperOrigin-RevId: 715070992