452 StableHLO modules crash without producing a meaningful stack trace. The diagnostic output from each crash is either nothing or a message that does not pin the failure down. Several do not produce the same diagnostic output twice. Grouping them by the last HLO pass to start before the crash and the set of backends that crashed yields the 13 groups below. All modules are included in ir.zip.
Groups
001-000644 — 14 modules — fails on GPU (failed (rc=-6, SIGABRT)) — last HLO pass entered: latency-hiding-scheduler — terminate called after throwing an instance of 'std::out_of_range'
002-000720 — 157 modules — fails on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered: latency-hiding-scheduler
003-000894 — 19 modules — fails on CPU (failed (rc=-6, SIGABRT)) — last HLO pass entered: reduce-window-rewriter — terminate called after throwing an instance of 'std::bad_alloc' or hlo_computation.cc:1022 instruction_count() == post_order.size() (…) n…
004-001019 — 123 modules — fails on Interpreter (failed (rc=-11, SIGSEGV)) — no pass log
005-001225 — 18 modules — fails on CPU (failed (rc=-6, SIGABRT)) — last HLO pass entered: apply-xla-transforms — free(): invalid next size (fast)
006-001450 — 9 modules — fails on CPU (crashed) — last HLO pass entered: reduce-window-rewriter — no diagnostic output or terminate called after throwing an instance of 'std::bad_alloc'
007-002593 — 46 modules — fails on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered: cse
008-007867 — 39 modules — fails on CPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered: apply-xla-transforms
009-010606 — dies on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered: reduce-window-rewriter
010-017624 — 15 modules — fails on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered: constant_folding
011-025224 — 3 modules — fails on Interpreter (failed (rc=-6, SIGABRT)) — last HLO pass entered: ConvolutionTypeCanonicalizer — terminate called recursively
012-025324 — 7 modules — fails on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered: priority-fusion
013-032176 — fails on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered: dynamic-slice-annotator
What is attached
ir.zip contains one folder for each group above. Each folder contains the module and a README with the environment, the exact command that crashes the process and its output, the reference command that completes successfully, and the per-backend outcome for that module. The group-members/ directory contains the remaining modules in that group.
Environment
- CPU:
Intel(R) Core(TM) i9-14900HX
- GPU:
NVIDIA GeForce RTX 4060 Laptop GPU
- CUDA Driver:
580.126.09
- XLA commit:
6f1f2d916dcef0428bfa7ea66765be539126c4f2
- StableHLO commit:
e6f81ebd06b3509f2c7fa6175430aadbd4d724ca
Contact
- Email:
ch395@njit.edu, zhihao.yao@njit.edu, benquike@gmail.com
452 StableHLO modules crash without producing a meaningful stack trace. The diagnostic output from each crash is either nothing or a message that does not pin the failure down. Several do not produce the same diagnostic output twice. Grouping them by the last HLO pass to start before the crash and the set of backends that crashed yields the 13 groups below. All modules are included in ir.zip.
Groups
001-000644— 14 modules — fails on GPU (failed (rc=-6, SIGABRT)) — last HLO pass entered:latency-hiding-scheduler— terminate called after throwing an instance of 'std::out_of_range'002-000720— 157 modules — fails on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered:latency-hiding-scheduler003-000894— 19 modules — fails on CPU (failed (rc=-6, SIGABRT)) — last HLO pass entered:reduce-window-rewriter— terminate called after throwing an instance of 'std::bad_alloc' or hlo_computation.cc:1022 instruction_count() == post_order.size() (…) n…004-001019— 123 modules — fails on Interpreter (failed (rc=-11, SIGSEGV)) — no pass log005-001225— 18 modules — fails on CPU (failed (rc=-6, SIGABRT)) — last HLO pass entered:apply-xla-transforms— free(): invalid next size (fast)006-001450— 9 modules — fails on CPU (crashed) — last HLO pass entered:reduce-window-rewriter— no diagnostic output or terminate called after throwing an instance of 'std::bad_alloc'007-002593— 46 modules — fails on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered:cse008-007867— 39 modules — fails on CPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered:apply-xla-transforms009-010606— dies on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered:reduce-window-rewriter010-017624— 15 modules — fails on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered:constant_folding011-025224— 3 modules — fails on Interpreter (failed (rc=-6, SIGABRT)) — last HLO pass entered:ConvolutionTypeCanonicalizer— terminate called recursively012-025324— 7 modules — fails on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered:priority-fusion013-032176— fails on GPU (failed (rc=-11, SIGSEGV)) — last HLO pass entered:dynamic-slice-annotatorWhat is attached
ir.zip contains one folder for each group above. Each folder contains the module and a README with the environment, the exact command that crashes the process and its output, the reference command that completes successfully, and the per-backend outcome for that module. The group-members/ directory contains the remaining modules in that group.
Environment
Intel(R) Core(TM) i9-14900HXNVIDIA GeForce RTX 4060 Laptop GPU580.126.096f1f2d916dcef0428bfa7ea66765be539126c4f2e6f81ebd06b3509f2c7fa6175430aadbd4d724caContact
ch395@njit.edu, zhihao.yao@njit.edu, benquike@gmail.com