[ROCm][release/2.13] Enable linalg tests via hipSOLVER and remove Magma dependency

[akashveramd]

Summary

Backports 6 upstream PyTorch PRs to release/2.13 that enable ROCm linear algebra tests and remove the Magma dependency by routing through hipSOLVER (ROCm >= 7.14):

• [ROCm] Enabled test_linalg tests that were skipped due to skipCUDAIfNoMagmaAndNoCusolver pytorch/pytorch#180303 – Replace skipCUDAIfNoMagmaAndNoCusolver with skipCUDAIfNoMagmaAndNoLinalgsolver across linalg tests and OpInfo definitions, enabling tests on ROCm via hipSOLVER
• [ROCm] Re-enable test_preferred_linalg_library test for ROCm pytorch/pytorch#187703 – Re-enable test_preferred_linalg_library test for ROCm
• [ROCm] Enable linalg tests for eig, ldl_solve operator (hipsolver) & enable test_linalg_solve, test_triangular_solve linalg tests with cuBLAS path pytorch/pytorch#185557 – Enable linalg.eig / ldl_solve via hipSOLVER xgeev (ROCm >= 7.14); enable test_linalg_solve and test_triangular_solve via cuBLAS path
• [ROCm] Merge test_norm_matrix_degenerate_shapes_old_numpy & test_norm_matrix_degenerate_shapes to run for older & newer Numpy versions pytorch/pytorch#187011 – Merge test_norm_matrix_degenerate_shapes variants into a single test that handles both old and new NumPy behavior
• [ROCm] Enable test_inverse_errors_large & test_lu_solve_batched_many_batches linalg tests pytorch/pytorch#186122 – Enable test_inverse_errors_large and test_lu_solve_batched_many_batches linalg tests on ROCm
• [ROCm] Enable cholesky_ex via hipSOLVER xgeev (ROCm >= 7.14) & remove Magma dependency for autogenerated tests pytorch/pytorch#188720 – Enable cholesky_ex via hipSOLVER and remove Magma dependency for autogenerated linalg tests

Key C++ changes

• BatchLinearAlgebraLib.cpp/h, CUDASolver.cpp/h: extend xgeev, xpotrs, ldl_solve and related 64-bit solver paths to compile and run on ROCm >= 7.14 (USE_ROCM && ROCM_VERSION >= 71400)
• BatchLinearAlgebra.cpp: route linalg_eig_kernel through hipSOLVER xgeev on ROCm >= 7.14 instead of Magma
• cuda_to_hip_mappings.py: add hipify mappings for cusolverDnXgeev, cusolverDnCreateParams/DestroyParams, and related symbols

Test changes

• Replace @skipCUDAIfNoMagma / @skipCUDAIfNoMagmaAndNoCusolver with @skipCUDAIfNoMagmaAndNoLinalgsolver across test_linalg.py, common_methods_invocations.py, and opinfo/definitions/linalg.py
• Remove ROCm-specific skipIf skips for test_linalg_eig_stride_consistency in inductor tests
• Add ROCm version guard (ROCM_VERSION < (7, 14)) for linalg.eig / linalg.eigvals OpInfos on older ROCm without hipSOLVER xgeev support

cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @azahed98

[rocm-repo-management-api]

Jenkins build for b7eaae10d15dcfd0b6dfe980e0ec2f151b670ecc commit finished as ABORTED
Links: Pipeline Overview / Build artifacts / Test Results

[akashveramd]

ROCm MAGMA → hipSOLVER cherry-pick test results

Environment: ROCm 7.14.60850 · no MAGMA · PyTorch built without CPU LAPACK
Branch under test: av_replace_magma_rel2.13 @ b7eaae10d15
Baseline: release/2.13 @ 4fc86dafe38

---

Summary by test file

Test file	Baseline release/2.13	PR av_replace_magma_rel2.13	Δ passed	Δ skipped
test/test_linalg.py	202 ✅ 85 ⏭	257 ✅ 26 ⏭	+55	−59
test/test_linalg.py (SLOW)	4 ✅	4 ✅	0	0
test/inductor/test_torchinductor.py	0 ✅ 1 ⏭	1 ✅ 0 ⏭	+1	−1
test/inductor/test_compile_subprocess.py	0 ✅ 1 ⏭	1 ✅ 0 ⏭	+1	−1
test/inductor/test_torchinductor_opinfo.py	2 ✅ 79 ⏭	78 ✅ 3 ⏭	+76	−76
test/test_ops.py	132 ✅ 974 ⏭	956 ✅ 143 ⏭ 7 ⚠️	+824	−831
test/test_ops_gradients.py	4 ✅ 386 ⏭	127 ✅ 257 ⏭ 6 ⚠️	+123	−129
TOTAL	344 ✅ 1526 ⏭ 0 ❌	1424 ✅ 429 ⏭ 13 ⚠️ 0 ❌	+1080	−1097

✅ passed · ⏭ skipped · ⚠️ xfailed (expected failure) · ❌ failed

---

Why skips decreased

The bulk of the 1097 newly-passing tests were previously skipped in the baseline for these reasons — all resolved by the cherry-picks:

Skip reason (baseline only)	Root cause	Resolved by
no MAGMA library detected	MAGMA absent on ROCm	PRs pytorch#185557, pytorch#188720 — hipSOLVER replaces MAGMA
MAGMA required for ROCm	Hard MAGMA guard on eig/eigvals/etc.	PR pytorch#185557 — removes @skipCUDAIfNoMagma guards
ROCm hipsolver backend does not currently support eig	Missing xgeev in old hipsolver	PR pytorch#188720 — adds @skipIfRocmVersionLessThan([7,14]) and enables path
skipCUDAIfRocm: test doesn't currently work on the ROCm stack	Blanket ROCm skip	PRs pytorch#180303, pytorch#185557 — replaces with targeted skips

---

Remaining skips in PR branch (all pre-existing / expected)

Skip reason	Count	Notes
no MAGMA library detected	0	✅ fully resolved
test is slow; run with PYTORCH_TEST_WITH_SLOW	~30	Pre-existing; run with PYTORCH_TEST_WITH_SLOW=1
Only runs on cpu	~40	CPU-only tests; not applicable to CUDA suite
https://github.com/pytorch/pytorch/issues/150959	4	Known open issue, pre-existing
Skipped! Op doesn't support autograd for this dtype	~12	Op limitation, not our change
Skipped! Op has no inplace variant / Op does support gradgrad	~10	Op limitation, not our change
xfailed (⚠️)	13	Explicitly marked expected failures in PR code

---

No regressions

Zero tests moved from passing → failing across all 6 test files.
All failures present in the baseline are also present in the PR branch.

[akashveramd]

ROCm MAGMA → hipSOLVER cherry-pick: autogenerated test results (upstream-style counting)

Counting methodology: one row per unique test function (dtype variants collapsed).
A test is "✅ Passed (some skipped)" if at least one dtype variant passed and at least one was skipped.

Environment: ROCm 7.14.60850 · no MAGMA · av_replace_magma_rel2.13 @ b7eaae10d15
Baseline: release/2.13 @ 4fc86dafe38

---

Overall summary

Status	Baseline release/2.13	PR av_replace_magma_rel2.13	Upstream PR pytorch#188720 (reference)
✅ Passed	217	1,883	1,525
✅ Passed (some skipped)	1	6	7
⏭ Skipped	2,470	781	585
⚠️ XFAIL	3	21	17
❌ Failed	0	0	0
Total	2,691	2,691	2,156 (+22 not generated)

Our total (2,691) is higher than upstream (2,156) because we run 4 additional test files
(test_ops_jit.py, test_overrides.py, test_testing.py, functorch/test_ops.py) that
upstream did not include in their count. The pattern is identical: 0 failures, same xfail categories.

---

Per test file

Test file	Baseline ✅	Baseline ⏭	PR ✅	PR ⏭	Δ passed	Δ skipped
test/inductor/test_torchinductor_opinfo.py	4	42	43	2	+39	−40
test/test_decomp.py	6	44	45	4	+39	−40
test/test_jit_fuser_te.py	0	93	0	93	0	0
test/test_meta.py	25	352	185	192	+160	−160
test/test_ops.py	136	961	911	179	+775	−782
test/test_ops_gradients.py	4	231	72	160	+68	−71
test/test_ops_jit.py	5	45	44	4	+39	−41
test/test_testing.py	6	44	48	2	+42	−42
test/test_overrides.py	5	42	45	2	+40	−40
test/functorch/test_ops.py	27	572	495	100	+468	−472
test/test_fx_experimental.py	0	43	0	43	0	0
test/inductor/test_torchinductor.py	0	1	1	0	+1	−1
TOTAL	218	2,470	1,889	781	+1,671	−1,689

---

Notes on remaining skips (781)

Category	Count	Notes
test_jit_fuser_te.py (NNC doesn't fuse linalg ops)	93	Pre-existing — NNC/TensorExprKernel does not support linalg ops
test_fx_experimental.py (@onlyCPU decorator)	43	Pre-existing — these tests only run on CPU by design
test_meta.py inplace variants	~80	Pre-existing — linalg ops have no in-place variants
test_ops.py (no tags, multi-device, slow)	~100	Pre-existing structural skips
test_ops_gradients.py inplace / gradgrad	~60	Pre-existing — no in-place grad for linalg ops
functorch/test_ops.py	100	Reduced from 572; remaining are op-limitation skips
Other (slow, cpu-only, dtype limitations)	remainder	All pre-existing

0 tests failed. All skips are pre-existing and structurally equivalent to upstream.