Skip to content

[None][chore] GB300 development workflow smoke test - #21451

Draft
litaotju wants to merge 4 commits into
NVIDIA:mainfrom
litaotju:codex/omnistation-gb300-workflow-smoke-20261102
Draft

litaotju wants to merge 4 commits into
NVIDIA:mainfrom
litaotju:codex/omnistation-gb300-workflow-smoke-20261102

Conversation

@litaotju

@litaotju litaotju commented Nov 2, 2026 •

Copy link
Copy Markdown
Collaborator

Description

Draft smoke test of the TensorRT-LLM development workflow on an Omnistation and GB300: local edit, ARM64 source build targeting SM103, focused GPU validation, fork push, and draft PR creation.

This branch is based on #21376 at 66d808b2b2ec83439b225f8decc23a37d71d2d90, so the comparison to main also includes that parent PR's changes. The additional workflow change is limited to two files: an invalid native-loader package error now tells the caller to pass __name__ from the target's __init__.py, and the existing assertion checks the hint.

This is a draft workflow test. No reviewers are requested. GitHub currently reports conflicts against main; this stacked draft is not ready to merge.

Test Coverage

  • Applicable repository pre-commit and commit-message checks passed, including DCO sign-off.
  • ARM64 source build targeting SM103 completed successfully.
  • All 9 selected tests passed with no skips on GB300:
    • tests/unittest/_torch/modeling_v2/test_modeling_v2_native_loader.py
    • tests/unittest/_torch/modeling_v2/test_modeling_v2_ci_selection.py
    • tests/unittest/_torch/modeling_v2/models/gpt_oss/gpt_oss_120b__sm_10x__default/test_modeling_v2_gpt_oss_swizzle_block_scales.py
  • A single-GPU PyTorch runtime smoke generated 8 tokens using the existing Qwen2-0.5B checkpoint. This is a basic execution check, not an accuracy or performance result.

The build/test source inputs were the parent commit plus the two-file patch. The patch is byte-identical to the diff committed in 47003af897b974265b427101fa9061b2d6608b57 (SHA-256 2111bf414fa275b989377e979c4f5eda8893eefea9c5631dfd5f4b129a1ce645).

PR Checklist

  • Reviewed the applicable checklist items. The additional two-file change introduces no dependency, public API, or ownership change.
  • Kept this workflow PR as a draft without requesting reviewers.

Signed-off-by: Tao Li <litaotju@users.noreply.github.com>
Signed-off-by: Tao Li <litaotju@users.noreply.github.com>
…guides

Modeling V2 conventions are encouraged through docs and examples rather
than enforced by tests. Keep only checks the code or CI depends on.

- Docs: rewrite both Modeling V2 READMEs in plainer style, fix facts
  (resolver name, incremental rebuild of family libraries, build
  options), and describe target naming and family layout as conventions.
- _native.py: find a target by its position in the package path rather
  than by "__" in its name.
- modeling_v2.cmake: target names only need to be Python package names;
  family names may use uppercase; family torch namespaces join the
  collision check.
- CBTS: select Modeling V2 tests per family from the changed path alone.
- Tests: delete test_modeling_v2_claims.py and
  test_modeling_v2_native_claims.py. Keep the three checks per-family
  selection relies on in test_modeling_v2_ci_selection.py, and keep
  test_modeling_v2_native_loader.py to loader behavior.

Signed-off-by: Tao Li <litaotju@users.noreply.github.com>
Signed-off-by: Tao Li <litaotju@users.noreply.github.com>
@ruocheng-nv

Copy link
Copy Markdown
Collaborator

Hi, heads-up: #22286 has merged, so ModelingV2 now lives at tensorrt_llm/_torch/modeling_v2 (module tensorrt_llm._torch.modeling_v2), and the old _torch/_experimental/modeling_v2 path no longer exists.

Since this PR touches modeling_v2, please rebase it onto the latest main. You can use these migration commands to move any files you added under the old directory and update any remaining references to the old path.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants