Repository navigation
Tags
Tags
test: opt-in GPU V-PROGRAM run on a confidential NVIDIA CRN, fixtures… … on aleph-vm 2.1.2 (#43) * ci: opt-in GPU run with a second static confidential CRN * feat(crn-up): gpu marker and its own IPv6 pool override * test: GPU V-PROGRAM fixtures (gpu runtime bundle, cuda-probe workload) * test: GPU V-PROGRAM end to end (create --gpu, placement, attested CUDA calls) * docs: opt-in GPU runs * ci: tidy GPU-run plumbing * docs: a static GPU host needs Docker for the vm-connector * test: pin non-GPU confidential tests to the TEE server during GPU runs Once the GPU host reports confidential capability, the scheduler is free to place non-GPU SEV-SNP workloads there too, breaking tests that depend on TEE-server specifics (routed IPv6 /64, confidential artifacts, TCB override flags). Add a tee_pin_args fixture that splats --crn <hash> into the affected create commands only when ALEPH_TESTNET_NVIDIA_CC_CRN_HOST is set, and export the TEE server's node hash (excluding the GPU-marked CRN dir, which carries both markers) alongside its host. Revert the earlier either-SNP-host loosening in test_vprograms.py now that pinning covers it. * test: plain punctuation in the pin fixture's failure message Review caught an em-dash in tee_pin_args' pytest.fail message. * chore: plain punctuation in the GPU fixture list * fix(tests): skip the GPU runtime upload on default runs * test(gpu): pin the H200 model and move to the fixtures built with requirement enforcement * test(gpu): expect the pinned model in the shown requirement * test: move every V-PROGRAM fixture to the aleph-vm dev-2.1.2 build Exec, compose and GPU runtimes and the cuda-probe workload rebuilt from aleph-vm 80e48a8e (dev-2.1.2, main at 0295bb08 plus the 2.2 work), published as vprogram-fixtures-3 and vprogram-fixtures-gpu-3. The fib workload is unchanged. * test(vprogram): the agent's 503 "upstream not ready" is a startup signal Since aleph-vm#1282 the attest agent answers a workload that is not listening yet with 503 "upstream not ready" instead of 502 "upstream unreachable"; the compose polls only knew the old body and failed the first call (run 35722004130). Both bodies now mean "still starting" in the compose tests and in the shared attested-call helper.
test: opt-in GPU V-PROGRAM run on a confidential NVIDIA CRN, fixtures… … on aleph-vm 2.1.2 (#43) * ci: opt-in GPU run with a second static confidential CRN * feat(crn-up): gpu marker and its own IPv6 pool override * test: GPU V-PROGRAM fixtures (gpu runtime bundle, cuda-probe workload) * test: GPU V-PROGRAM end to end (create --gpu, placement, attested CUDA calls) * docs: opt-in GPU runs * ci: tidy GPU-run plumbing * docs: a static GPU host needs Docker for the vm-connector * test: pin non-GPU confidential tests to the TEE server during GPU runs Once the GPU host reports confidential capability, the scheduler is free to place non-GPU SEV-SNP workloads there too, breaking tests that depend on TEE-server specifics (routed IPv6 /64, confidential artifacts, TCB override flags). Add a tee_pin_args fixture that splats --crn <hash> into the affected create commands only when ALEPH_TESTNET_NVIDIA_CC_CRN_HOST is set, and export the TEE server's node hash (excluding the GPU-marked CRN dir, which carries both markers) alongside its host. Revert the earlier either-SNP-host loosening in test_vprograms.py now that pinning covers it. * test: plain punctuation in the pin fixture's failure message Review caught an em-dash in tee_pin_args' pytest.fail message. * chore: plain punctuation in the GPU fixture list * fix(tests): skip the GPU runtime upload on default runs * test(gpu): pin the H200 model and move to the fixtures built with requirement enforcement * test(gpu): expect the pinned model in the shown requirement * test: move every V-PROGRAM fixture to the aleph-vm dev-2.1.2 build Exec, compose and GPU runtimes and the cuda-probe workload rebuilt from aleph-vm 80e48a8e (dev-2.1.2, main at 0295bb08 plus the 2.2 work), published as vprogram-fixtures-3 and vprogram-fixtures-gpu-3. The fib workload is unchanged. * test(vprogram): the agent's 503 "upstream not ready" is a startup signal Since aleph-vm#1282 the attest agent answers a workload that is not listening yet with 503 "upstream not ready" instead of 502 "upstream unreachable"; the compose polls only knew the old body and failed the first call (run 35722004130). Both bodies now mean "still starting" in the compose tests and in the shared attested-call helper.
chore: pin aleph-vm 2.1.0 and aleph-cli 0.19.0 (#42) aleph-vm 2.1.0 (tag = main 3579c430, the build the last green run of the 2.1 chain used) instead of the main branch's CI artifacts, and aleph-cli 0.19.0 instead of 0.18.0. Since aleph-cli 0.19.0, `instance create --confidential` means SEV-SNP by default and rejects --confidential-firmware; the session-based legacy SEV tests now ask for that flavor explicitly with --tee sev. Run 35213971275: 29 passed, 3 skipped.
chore: pin aleph-vm 2.1.0 and aleph-cli 0.19.0 (#42) aleph-vm 2.1.0 (tag = main 3579c430, the build the last green run of the 2.1 chain used) instead of the main branch's CI artifacts, and aleph-cli 0.19.0 instead of 0.18.0. Since aleph-cli 0.19.0, `instance create --confidential` means SEV-SNP by default and rejects --confidential-firmware; the session-based legacy SEV tests now ask for that flavor explicitly with --tee sev. Run 35213971275: 29 passed, 3 skipped.
chore: pin aleph-vm 2.1.0 and aleph-cli 0.19.0 (#42) aleph-vm 2.1.0 (tag = main 3579c430, the build the last green run of the 2.1 chain used) instead of the main branch's CI artifacts, and aleph-cli 0.19.0 instead of 0.18.0. Since aleph-cli 0.19.0, `instance create --confidential` means SEV-SNP by default and rejects --confidential-firmware; the session-based legacy SEV tests now ask for that flavor explicitly with --tee sev. Run 35213971275: 29 passed, 3 skipped.
chore: pin aleph-vm 2.1.0 and aleph-cli 0.19.0 (#42) aleph-vm 2.1.0 (tag = main 3579c430, the build the last green run of the 2.1 chain used) instead of the main branch's CI artifacts, and aleph-cli 0.19.0 instead of 0.18.0. Since aleph-cli 0.19.0, `instance create --confidential` means SEV-SNP by default and rejects --confidential-firmware; the session-based legacy SEV tests now ask for that flavor explicitly with --tee sev. Run 35213971275: 29 passed, 3 skipped.
fix: retry flaky SSH connections in confidential-artifacts.sh (#34) PR CI failed in 'Prepare confidential VM artifacts' with the known fresh-droplet flake (kex_exchange_identification: Connection reset by peer) while scp-ing sevctl from CRN 0. crn-up.sh already mitigates this with ControlMaster multiplexing plus a retry-on-exit-255 wrapper; port the same mitigation to confidential-artifacts.sh (tee_ssh, sevctl_ssh, and the three scp call sites). The remote commands are idempotent, so retrying connection-level failures is safe.
fix: switch CRN droplets to ubuntu-24.04, DO retired debian-12-x64 (#33) DigitalOcean no longer offers the debian-12-x64 image (only debian-13-x64 remains), so PR CI fails at droplet creation with 422 'invalid image'. Move the DO CRN droplets to ubuntu-24-04-x64, the same distro as the CCN droplet. This keeps the sevctl binary copied from CRN 0 guaranteed to run on the CCN (debian-13's glibc 2.41 is newer than ubuntu-24.04's 2.39 and would break that). Also make the branch (CI artifact) install path pick the aleph-vm .deb per CRN distro via /etc/os-release, like the release path already does: aleph-vm CI builds all four distro debs, and a static TEE server may not run the same distro as the DO droplets.