[elibol]

Tile and SIMT currently expose separate cuda-core contexts, streams, launch configurations, and allocation owners. A buffer moving between the models consequently needs raw-pointer adaptation, and the separate runtimes cannot coordinate lifetime or conflicting accesses.

This removes cuda_core::simt and gives both models the same host handles and allocation owner. Device identifies a physical GPU; Context owns or borrows the actual CUDA context. Typed buffers, contiguous sliced views, foreign owners, and Tile tensors share allocation identity and access tracking. Tensor::from_buffer, into_buffer, and as_buffer transfer or borrow storage without copying device memory. Ownership transfer rejects live typed aliases, preserving the boolean representation guarantees in #1516.

Safe submissions retain allocations and access leases until GPU completion, including after a borrowed launch future is dropped or forgotten. Shared references to device atomics are conservatively tracked as writable. Buffer and stream destructors retire work without waiting on the dropping thread; unknown completion retains resources instead of freeing active memory. The existing SIMT async launcher receives the small retention bridge needed to make this core migration safe before its separate consolidation.

Allocation uniqueness counts user-visible byte views independently of completion leases. This removes the retirement window where a decremented lease count still had a live Arc and made a unique buffer appear aliased. The same separation at the Tensor layer comes from #1516; both checks remain required for zero-copy ownership transfer.

Nihal's fixes preserve ordinary shared-read permissions on empty and peeled buffers and wait for retirement before checking reclaimed VRAM.

Consumer code, generated launchers, examples, templates, documentation, tests, and CI paths use the shared API. Tile numeric storage wrappers also implement DeviceCopy; DType and DeviceCopy remain independent traits.

Stack: #1516 merged into main as dfea11698aa63b9a42d9f8181a711d27774d0722. This PR now targets main and contains only the core migration. Merge this before the follow-up #1604, which removes cuda_async::simt.

Validation:

• Full cuda-core suite: 111 passed; full cuda-async suite: 151 passed.
• Selected Tile library/GPU/interop/submission-lifetime suites: 156 passed, including graph replay, offset preservation, alias rejection, and forgotten futures.
• SIMT host/macro suites: 261 passed; cargo-oxide: 268 passed. Three additional GPU launch-lifetime regressions pass explicitly with --ignored.
• Clippy with warnings denied for host workspace, Tile, and affected SIMT packages; strict cuda-core rustdoc; formatting, host-path and SPDX guards. The Tile workspace is also checked with every example feature enabled, including Candle CPU references.
• Dependency checks pass for the host, SIMT, backend, device-only fixture, and Tile workspaces.

GPU validation used a single RTX 5090. Multi-GPU/NVLink cases self-skip on this machine; this does not claim multi-GPU coverage. Existing deliberately ignored tests remain ignored.

Closes #1460. Closes #1504. The global-context re-entry/state-machine follow-up remains tracked separately in #1403.

Earlier follow-up validation: all 50 cuda-core unit tests and all nine targeted shared-allocation/retirement GPU tests pass, as does cuda-core Clippy with warnings denied. The new concurrency regression failed 3,308 uniqueness checks on the previous implementation and passes with the fix.

Earlier main integration: Merged the updated #1516 prerequisite and migrated the new CuTe host consumers to shared LaunchConfig and the explicit event-flags API. Affected-package Clippy and all 16 CuTe example/model workspace Clippy checks pass. CI passed before this refresh; current-head CI is tracked below.

Current update (2026-10-30), head 98d0820debc6f3994cb45c041ac182e174bff364: Plain completion leases now retain allocation access independently of user-view tokens, so tensor-to-buffer conversion preserves unique ownership without weakening cross-stream exclusion or recorded graph ownership. The combined stack passes 1,427 CPU/assembler tests, 750 GPU tests, 27 supported examples and eight benchmark smoke tests. Initial disk exhaustion was resolved and the affected suites rerun successfully. The expanded atomic example is migrated to owned conversions; all 24 atomic GPU cases also pass on the final shared-runtime head. Fresh CI is running; #1516 still needs renewed approval before this stack can merge into main.

Merge status (2026-10-30): current head ba76442278ad1d6c67070ef5fde3dd63efcddb85 records the accepted #1516 squash merge. The squash-merge tree exactly matches the accepted #1516 head; the ancestry-only refresh leaves this PR's entire file tree unchanged from its reviewed predecessor. After copy-pr-bot deleted the old temporary bases and automatically closed the stacked PRs, they were reopened against main; temporary restoration refs were removed. Merge order is #1603 then #1604. Fresh CI is running, and retargeting caused GitHub to require review again. Issue/release tracking: #1606.

[nihalpasham]

Thanks @elibol, this looks good. I pushed a small commit (eec8776ff) for two minor things:

• peel_elem and the num_bytes == 0 early returns in DeviceBuffer were leaving shared_access_writes = true (from relabel / from_raw_parts), which blocked concurrent read-only launches on peeled or empty buffers.
• Added wait_for_retirement to vram_returns_to_baseline_after_buffer_cycles in device_buffer_leaks.rs so cuMemGetInfo_v2 doesn't race the background retirement thread.

[elibol]

@nihalpasham Thanks for the review and fixes. Your commit eec8776ff is preserved in the updated branch: peeled and empty buffers retain ordinary shared-read permissions, and the VRAM reclamation test waits for retirement before checking the baseline.

The follow-up work also fixes allocation uniqueness accounting by counting actual buffer views separately from GPU completion leases, and incorporates the corresponding tensor ownership fix from #1516. The regression fails on the earlier implementation and passes with the fix.

All of your feedback is incorporated. All 63 active CI checks pass at 36f9072030, including CodeQL; the targeted shared-allocation and retirement GPU tests also pass. This remains stacked on #1516.

[elibol]

Addressed the completion-retention follow-up in e6bb3ea. Plain completion leases now retain the access tracker and its allocation leases, rather than the Storage containing a user-view token. tensor.into_buffer() can therefore report unique ownership while completion retirement is pending. Cross-stream access exclusion and allocation lifetime remain enforced; recorded graph ownership deliberately still counts as an alias because the graph can access the storage again.

The new regression reproduced the false alias before the fix. All 32 tensor unit tests pass afterward, including allocation-lifetime, access-exclusion, real-alias and graph-alias checks. The combined stack also passes the shared-buffer and shared-async GPU interop suites (four tests). The full validation run is being completed separately after local disk exhaustion interrupted some suites.

Refreshed against current main and #1516, migrated the new examples' imports, and registered the shared-buffer test before the GPU runner exits, using its archive-aware adapter. No cuda_core::simt imports remain in Rust source.

@nihalpasham, GitHub dismisses approvals after code updates; please reapprove the follow-up once checked. This PR remains stacked on #1516 and will target main after that parent merges.

[elibol]

/ok to test dd7eb32

[elibol]

/ok to test 90191e4

[elibol]

Validation is complete on the combined stack: all CPU/assembler and GPU suites, supported examples and benchmark smoke tests pass. The detailed counts, tested commit, initial disk-space failure and CUDA 13.3 limits are recorded in the latest validation comment on #1604. This PR remains stacked on #1516, whose updated source-cast fix needs renewed approval.

[elibol]

/ok to test 98d0820

The base branch was changed.