Skip to content

[TorchAO] Float8 training support on XPU #2325

New issue
New issue
Open
Feature
Open
[TorchAO] Float8 training support on XPU#2325
Feature
Assignees
arlesniak
Labels
module: aoQuantization: int4/int8/fp8/PT2E/torchaotest: utUnit-test failure surface
Milestone
PT2.16

Description

@liangan1
liangan1
opened on Nov 11, 2025

🚀 The feature, motivation and pitch

The scaled_mm(pytorch/pytorch#165978) and inference(pytorch/ao#3332) is WIP, You can enable the realted training UTs with emulate path and comments your PR in pytorch/ao#2917 to help us to track the total status.
Reference: https://github.com/pytorch/ao/tree/main/torchao/float8

Alternatives

No response

Additional context

No response

Activity

  1. liangan1
    added
    module: aoQuantization: int4/int8/fp8/PT2E/torchao
    on Nov 11, 2025
  2. chuanqi129
    added the
    Feature
    issue type on Nov 12, 2025
  3. liangan1
    assigned
    riverliuintel
    on Nov 18, 2025
  4. riverliuintel
    added
    bug_fix_stage3
    on Nov 19, 2025
  5. arlesniak
    self-assigned this
    on Nov 26, 2025

  6. LifengWang commented on Jan 21, 2026

    @LifengWang
    LifengWang
    on Jan 21, 2026

    Hi @arlesniak , are there any updates on this feature?

  7. astachowiczhabana
    removed
    bug_fix_stage3
    on Feb 19, 2026

  8. karol-brejna-i commented on Apr 14, 2026

    @karol-brejna-i
    karol-brejna-i
    on Apr 14, 2026

    @LifengWang detailed progress is tracked in the following Jira ticket: https://jira.devtools.intel.com/browse/PYTORCHDGQ-7737
    Let me know if you have access to it.

  9. chuanqi129
    unassigned
    riverliuintel
    on Apr 15, 2026
  10. chuanqi129
    added this to the PT2.15 milestone on Apr 15, 2026
  11. karol-brejna-i
    modified the milestones: PT2.15, PT2.16 on Sep 22, 2026

  12. karol-brejna-i commented on Sep 22, 2026

    @karol-brejna-i
    karol-brejna-i
    on Sep 22, 2026

    All single-card UT have been enabled.
    There are open PRs for multi-card UTs. Waiting for multi-card GH runners to prove they work.

  13. daisyden
    added
    test: utUnit-test failure surface
    on Nov 11, 2026

  14. daisyden commented on Nov 11, 2026

    @daisyden
    daisyden
    on Nov 11, 2026
    Contributor
    label-issue: #2325

    Root cause: not a defect. Feature request to add TorchAO Float8 training support on
    XPU via the emulate path (scaled_mm/inference WIP upstream). No failure or test case reported.

    axis value reason
    type Feature Native Type is Feature; new-capability enablement request.
    test test: ut No e2e/oob/UT surface named; default test surface.
    module module: ao Human-set label preserved; TorchAO float8 quantization.
    priority Low Native Priority unset; feature request, no failing case.

    Applied to this issue. Generated by a local label-issue run.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

module: aoQuantization: int4/int8/fp8/PT2E/torchaotest: utUnit-test failure surface

Type

Feature

Projects

No projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions