Skip to content

feat(superdex): restore TorchEnv CPU host-bridge owners - #2070

Merged
TATP-233 merged 2 commits into
develop/tensor-runtimefrom
TATP-233/superdex-torchenv-runtime
Oct 8, 2026
Merged

TATP-233 merged 2 commits into
develop/tensor-runtimefrom
TATP-233/superdex-torchenv-runtime

Conversation

@TATP-233

@TATP-233 TATP-233 commented Oct 7, 2026

Copy link
Copy Markdown
Collaborator

Summary

  • Restore SuperDex as an active tensor-only Manager backend through the existing public packed CPU HOST_BRIDGE lifecycle.
  • Restore only the narrow Go2 and FR3 research owners, keeping their support grade at Configured rather than claiming full-training or cross-platform support.
  • Adapt FR3 Manager terms to the Torch read-plan and selected tensor joint-reset boundaries.
  • Add an explicit task-owned state-only packed-read declaration so task terms can consume public qpos/qvel without backend-name dispatch or private layout probing.
  • Update registry/CLI scope, support matrix, tests, and bilingual backend docs.
  • Preserve fail-closed behavior for reset randomization, fixed variants, and host pre-step callbacks.

Linked Work

Declared base

  • Base branch: develop/tensor-runtime

Validation

  • make test-all passed on the final local head before this PR was created or updated
  • Additional task-specific validation listed below

Commands actually run on the final local head:

make check
# Ruff format/check, mypy, pyright, focused test lint: passed

UNILAB_LOCAL_UNISIM=/home/user/orca/workspaces/UniLab/unisim-superdex-torchenv \
  uv run --no-sync pytest -m "not slow" -q
# 1844 passed, 45 skipped

uv run --no-sync pytest \
  tests/base/test_superdex_backend_options.py \
  tests/base/test_registry.py \
  tests/assets/test_superdex_assets.py \
  tests/test_cli_runtime_requirements.py \
  tests/envs/test_fr3_superdex.py \
  tests/envs/test_go2_superdex.py -q
# 71 passed, 1 skipped

uv run --no-sync pytest \
  tests/envs/test_fr3_superdex.py \
  tests/envs/test_go2_superdex.py -q
# 13 passed, 1 skipped

uv run --no-sync pytest -m slow \
  tests/scripts/test_support_matrix.py \
  tests/envs/test_fr3_superdex.py \
  tests/envs/test_go2_superdex.py -q
# 16 passed

uv run --no-sync pytest tests/architecture/test_backend_public_contract_boundary.py -q
# passed

UNILAB_LOCAL_UNISIM=/home/user/orca/workspaces/UniLab/unisim-superdex-torchenv \
  uv run --no-sync pytest tests/base/test_backend_imports.py -q
# passed

UniSim prerequisite contract validation:

# in unisim PR #342 checkout
uv run --no-sync pytest tests/adapters/superdex/test_tensor_host_bridge.py -q
# 15 passed

Local environment preparation (not a repository change):

uv sync --locked --extra superdex --extra uni_rl
uv pip install mujoco~=3.11.0 mjbatch-uni~=0.2.4
uv run unilab-pull-assets --robot allegro_hand

Remote CI route:

  • Base is develop/tensor-runtime, not main: remote CI is not the merge gate; local final-head evidence and review are the gate. The later integration PR to main runs remote CI.

Required gate limitation

make test-all was not fully green locally. Full slow-test execution currently reports 48 failed, 372 passed, 119 skipped, 10 errors; the failures/errors are in unrelated optional-backend and data fixtures requiring mujoco-warp, Newton, Motrix, motion assets, and visualization fixtures. They are not caused by this SuperDex change, but the complete required final-head gate remains unresolved and this PR must not merge until that evidence is complete.

Impact

  • Backend impact: SuperDex only
  • Platform impact: SuperDex runtime is CPython 3.12/3.13 Linux x86_64; the UniLab contract code is platform-neutral
  • Training effect expected: no change for existing owners; adds the two configured SuperDex research owners

Artifacts

  • W&B: n/a
  • benchmark result: n/a
  • video / screenshot: n/a
  • ONNX / checkpoint: n/a

Checklist

  • Added or updated tests where needed
  • Updated bilingual docs and generated support matrix
  • Linked the driving issue and required cross-repo contract PR
  • Noted the unresolved full-gate limitation explicitly

No Publication

  • This PR does not build or publish a PyPI distribution.

@TATP-233

TATP-233 commented Oct 7, 2026

Copy link
Copy Markdown
Collaborator Author

Depends on unilabsim/unisim#342 (SuperDex public tensor reset widths). I will monitor both current-head CI runs; the UniLab integration base is develop/tensor-runtime, so local final-head validation plus review is the direct gate.

@TATP-233

TATP-233 commented Oct 8, 2026

Copy link
Copy Markdown
Collaborator Author

Resolved conflicts with the current develop/tensor-runtime head (ef62660f), which added scoped Drake and CUDA MPS work.

Merge resolution preserves both scopes:

  • Active tensor backends now include superdex and scoped drake alongside the existing five.
  • Go2 registers MuJoCo + SuperDex + scoped Drake.
  • Support-matrix scope and generated bilingual matrices include both SuperDex and Drake; SuperDex stays Configured and Drake keeps its scoped Configured evidence test.
  • CLI keeps both backend runtime diagnostics and their fail-closed gates.
  • Backend overview documents both CPU host-bridge research profiles.
  • Shelved scope remains only isaacgym/isaacsim.

Final merged-head validation:

make check
# passed

CUDA_MPS_PIPE_DIRECTORY=/tmp/unilab-test-no-mps \
UNILAB_LOCAL_UNISIM=/home/user/orca/workspaces/UniLab/unisim-superdex-torchenv \
  uv run --no-sync pytest -m "not slow" -q
# 1867 passed, 45 skipped

The temporary MPS pipe path isolates the deterministic “invalid MPS topology” tests from the live host daemon, as those tests expect no reachable daemon. The Allegro binary assets were also prefetched before the final full non-slow rerun.

@TATP-233
TATP-233 merged commit 113271c into develop/tensor-runtime Oct 8, 2026
2 checks passed
@TATP-233
TATP-233 deleted the TATP-233/superdex-torchenv-runtime branch October 8, 2026 05:39
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.

1 participant