Skip to content

Build Linux Wheels (x86): rocm10.1 leg fails — "No matching distribution found for torch" (nightly ROCm 10.1 wheels missing from index) #5003

Open
Open
Build Linux Wheels (x86): rocm10.1 leg fails — "No matching distribution found for torch" (nightly ROCm 10.1 wheels missing from index)#5003
Labels
ci-failureAutomated report of CI failure on mainmodule: rocm

Description

@github-actions
github-actions
bot
opened on Oct 15, 2026

Commits affected: b4dcda49cb54ed9bf14de14df61bac684ca7a787
Run: https://github.com/pytorch/ao/actions/runs/38818567230

Summary

The rocm10.1 wheel-build leg fails because PyTorch's nightly ROCm 10.1 index serves no torch wheels, so pip install torch cannot resolve a candidate. This is a dependency-availability problem in the upstream nightly channel, not a code regression in the triggering commit.

Failure details

  • Failed job(s) (one root cause, two cascading jobs):
    • build / build-manywheel-py3_10-rocm10_1 — step Install torch dependency (exit code 1).
    • build / upload / upload-manywheel-py3_10-rocm10_1 — step Download the artifact (cascade: no wheel was produced).
  • Representative error (build job): 
    + conda run --no-capture-output -p /__w/_temp/conda_environment_38818567230 \
    pip install torch --pre --index-url https://download.pytorch.org/whl/nightly/rocm10.1
ERROR: Could not find a version that satisfies the requirement torch (from versions: none)
ERROR: No matching distribution found for torch
Looking in indexes: https://download.pytorch.org/whl/nightly/rocm10.1
  • Cascade error (upload job): 
    ##[error]Unable to download artifact(s): Artifact not found for name: pytorch_ao__3.10_rocm10.1_x86_64
  • Contrast: the sibling rocm10.0 leg installed torch-2.15.0.dev20260923+rocm10.0 from .../whl/nightly/rocm10.0 without issue. All CPU, CUDA 13.0/13.2/13.4, and XPU legs passed.

Root cause

The build matrix includes a rocm10.1 entry whose install command points at https://download.pytorch.org/whl/nightly/rocm10.1. pip reads that index but finds zero torch candidates (from versions: none), i.e. no nightly torch wheels have been published for ROCm 10.1 yet (the rocm10.0 index, by contrast, is populated). The build therefore cannot install its torch dependency, and the downstream upload job fails only because the build produced no artifact to download.

The triggering commit b4dcda4 ("[pat] Add opt-in global group-count score correction", #4983) is a pure quantization-code change and does not touch the build matrix, ROCm configuration, or the torch install logic. The failure is driven entirely by the state of the external nightly package index.

Note: the runner emitted a generic line stating "This is a script/workflow error, not an infrastructure issue." That is a canned runner message and is contradicted by the actual evidence — the empty upstream rocm10.1 index is an environment/dependency-availability condition, not a defect in this repo's code or the triggering commit.

Classification

infra failure — a missing/unpublished upstream dependency (no nightly torch wheels on the ROCm 10.1 index) external to the repository and unrelated to the triggering commit. It is deterministic (so not a per-run flaky test) and not a code regression (the changed commit does not touch build/ROCm logic, and the parallel ROCm 10.0 leg succeeds).

Suggested fix

  • Short term: re-run once PyTorch nightly publishes ROCm 10.1 torch wheels, or temporarily drop/disable the rocm10.1 matrix entry until upstream wheels exist (mirroring how rocm10.0 is handled).
  • Medium term: gate ROCm matrix versions on wheel availability so a newly-added ROCm version is only enabled after the corresponding nightly index is populated, to avoid redundant red builds on main.
  • The upload-leg failure needs no separate action — it resolves automatically once the build leg can install torch and produce the artifact.

Activity

  1. github-actions
    added
    ci-failureAutomated report of CI failure on main
    on Oct 15, 2026
  2. pytorch-bot
    added
    module: rocm
    on Oct 15, 2026

  3. github-actions commented on Oct 15, 2026

    @github-actions
    github-actionsBot
    on Oct 15, 2026 – with GitHub Actions
    ContributorAuthor

    ⚠️ Run 4xH100 tests also failed on this commit (b4dcda49cb54ed9bf14de14df61bac684ca7a787) — https://github.com/pytorch/ao/actions/runs/38818566659

  4. github-actions commented on Oct 15, 2026

    @github-actions
    github-actionsBot
    on Oct 15, 2026 – with GitHub Actions
    ContributorAuthor

    Recurs on commit 7b0382e13cb0184010c9395fed6f0218741c627b — Run: https://github.com/pytorch/ao/actions/runs/38819718512

    Matches the existing diagnosis exactly. The build / build-manywheel-py3_10-rocm10_1 leg fails at the Install torch dependency step:

    pip install torch --pre --index-url https://download.pytorch.org/whl/nightly/rocm10.1
ERROR: Could not find a version that satisfies the requirement torch (from versions: none)
ERROR: No matching distribution found for torch

    and build / upload / upload-manywheel-py3_10-rocm10_1 cascades with Artifact not found for name: pytorch_ao__3.10_rocm10.1_x86_64. All other legs (CPU, CUDA 13.0/13.2/13.4, ROCm 10.0, XPU) are unaffected.

    Still an infra/upstream dependency-availability issue: no nightly torch wheels published on the ROCm 10.1 index. The triggering commit ([pat] Add direct projected-weight optimization mode, #4984) is a quantization-code change and does not touch the build matrix or ROCm config. No new diagnosis needed; this confirms the failure is persistent on main until upstream ROCm 10.1 nightlies are published (or the matrix entry is gated/disabled).

    Note: the runner's canned line "This is a script/workflow error, not an infrastructure issue." is contradicted by the evidence — the empty upstream index is an environment condition, not a repo code defect.

  5. github-actions commented on Oct 15, 2026

    @github-actions
    github-actionsBot
    on Oct 15, 2026 – with GitHub Actions
    ContributorAuthor

    ⚠️ Build Linux Wheels (AArch64) also failed on this commit (7b0382e13cb0184010c9395fed6f0218741c627b) — https://github.com/pytorch/ao/actions/runs/38819718595

  6. github-actions commented on Oct 15, 2026

    @github-actions
    github-actionsBot
    on Oct 15, 2026 – with GitHub Actions
    ContributorAuthor

    ⚠️ Run 4xH100 tests also failed on this commit (7b0382e13cb0184010c9395fed6f0218741c627b) — https://github.com/pytorch/ao/actions/runs/38819718193

  7. github-actions commented on Oct 15, 2026

    @github-actions
    github-actionsBot
    on Oct 15, 2026 – with GitHub Actions
    ContributorAuthor

    Recurred on main at commit aa2521284b3beb7d35c1154b145eba694f390220 ("[pat] Add coupled shared-axis pruning", #4985).
    Run: https://github.com/pytorch/ao/actions/runs/38821023471

    Same root cause as this issue's diagnosis — matches exactly:

    • build / build-manywheel-py3_10-rocm10_1 fails at the torch-install step: 
      + conda run --no-capture-output -p /__w/_temp/conda_environment_38821023471 pip install torch --pre --index-url https://download.pytorch.org/whl/nightly/rocm10.1
ERROR: Could not find a version that satisfies the requirement torch (from versions: none)
Looking in indexes: https://download.pytorch.org/whl/nightly/rocm10.1
ERROR: No matching distribution found for torch
    • build / upload / upload-manywheel-py3_10-rocm10_1 cascades: Unable to download artifact(s): Artifact not found for name: pytorch_ao__3.10_rocm10.1_x86_64.

    The nightly ROCm 10.1 index still serves no torch wheels, so the build cannot resolve its dependency. The triggering commit touches only quantization code (shared-axis pruning) and does not modify the build matrix, ROCm config, or torch-install logic — consistent with an unpopulated upstream index rather than a code regression. (The runner's canned "script/workflow error, not an infrastructure issue" line is contradicted by the empty upstream index.)

    Still an infra failure; no new action beyond the existing suggestions (re-run once the rocm10.1 nightly index is populated, or gate the rocm10.1 matrix entry on wheel availability).

  8. github-actions commented on Oct 15, 2026

    @github-actions
    github-actionsBot
    on Oct 15, 2026 – with GitHub Actions
    ContributorAuthor

    ⚠️ Build Linux Wheels (AArch64) also failed on this commit (7b0382e13cb0184010c9395fed6f0218741c627b) — https://github.com/pytorch/ao/actions/runs/38819718595

  9. github-actions commented on Oct 15, 2026

    @github-actions
    github-actionsBot
    on Oct 15, 2026 – with GitHub Actions
    ContributorAuthor

    ⚠️ Run 4xH100 tests also failed on this commit (aa2521284b3beb7d35c1154b145eba694f390220) — https://github.com/pytorch/ao/actions/runs/38821023076

  10. github-actions commented on Oct 15, 2026

    @github-actions
    github-actionsBot
    on Oct 15, 2026 – with GitHub Actions
    ContributorAuthor

    Recurrence on commit 5d763d6e0d07a97d164295431d8dfa64208caa1b ("[pat] Accelerate coupled pruning for sharded DTensors", #4986).
    Run: https://github.com/pytorch/ao/actions/runs/38822682369

    Same root cause as this issue, matching the existing diagnosis exactly:

    • build / build-manywheel-py3_10-rocm10_1 fails at the Install torch dependency step: 
      conda run pip install torch --pre --index-url https://download.pytorch.org/whl/nightly/rocm10.1
ERROR: Could not find a version that satisfies the requirement torch (from versions: none)
ERROR: No matching distribution found for torch
    • build / upload / upload-manywheel-py3_10-rocm10_1 cascades: Unable to download artifact(s): Artifact not found for name: pytorch_ao__3.10_rocm10.1_x86_64.

    As before, the runner prints the canned line "This is a script/workflow error, not an infrastructure issue" — contradicted by the evidence: the upstream nightly ROCm 10.1 index still has zero torch wheels. The triggering commit is a quantization/pruning change (torchao/) and does not touch the build matrix, ROCm config, or torch-install logic. Still an infra failure (missing upstream dependency), not a regression in this commit. The ROCm 10.1 nightly index remains unpopulated days after this was first filed.

  11. github-actions commented on Oct 15, 2026

    @github-actions
    github-actionsBot
    on Oct 15, 2026 – with GitHub Actions
    ContributorAuthor

    ⚠️ Run 4xH100 tests also failed on this commit (5d763d6e0d07a97d164295431d8dfa64208caa1b) — https://github.com/pytorch/ao/actions/runs/38822682048

  12. github-actions commented on Oct 16, 2026

    @github-actions
    github-actionsBot
    on Oct 16, 2026 – with GitHub Actions
    ContributorAuthor

    Recurs on main at commit 861156873fdc084c03be6e99512312a6e3cb0909 (run https://github.com/pytorch/ao/actions/runs/38945305724).

    The build / build-manywheel-py3_10-rocm10_1 leg fails identically to the diagnosis in this issue — pip install torch --pre --index-url https://download.pytorch.org/whl/nightly/rocm10.1 returns ERROR: Could not find a version that satisfies the requirement torch (from versions: none) / No matching distribution found for torch, i.e. the upstream nightly ROCm 10.1 index still has no torch wheels. The upload-manywheel-py3_10-rocm10_1 job fails only as a cascade (artifact not produced). Still infra/dependency-availability, unrelated to the triggering commit (a qnn_test numerics backout that touches no build/ROCm logic).

    Additional note for this run (NOT the same root cause): the build / build-manywheel-py3_10-xpu leg also failed, but with a transient SIGBUS (exit 128) during pip install torch while downloading intel-openmp, with upload-manywheel-py3_10-xpu cascading. This was leg-specific (the rocm leg kept running and hit its own error ~2 min later, so it was not a run-level cancel) and looks like a one-off infra/bus-error blip rather than a reproducible failure — likely clears on re-run. Flagging it here rather than opening a separate issue; worth a re-run to confirm it's transient.

  13. github-actions commented on Oct 16, 2026

    @github-actions
    github-actionsBot
    on Oct 16, 2026 – with GitHub Actions
    ContributorAuthor

    ⚠️ Run 4xH100 tests also failed on this commit (861156873fdc084c03be6e99512312a6e3cb0909) — https://github.com/pytorch/ao/actions/runs/38945305156

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

Metadata

Metadata

Assignees

No one assigned

    Labels

    ci-failureAutomated report of CI failure on mainmodule: rocm

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions