[aien-spark]

Why

Since #170 the GB10 cannot be opened: nvrm_open refuses it with unsupported SM 0xa04 or missing target classes. At driver 580.173.02 the GB10 reports the GR SM-version field as 10_04 (0xa04), and admission only accepted 12_01.

What changed

NVIDIA's ctrl2080gr.h:323-328 keeps 10_04 only "until 12_* is tested and validated", so 10_04 alone does not identify the GB10. The admission rule in seeded Omega (om/encoder/amp_runtime.om op 1) now accepts sm_121 when:

• the field is 12_01, or
• the field is 10_04 and the PCI identity is 0x10de:0x2e12 (the GB10).

The complete Blackwell-B class set is still required in both cases. nvrm/nvrm.c only passes the vendor constant and the PCI ids to Omega; no admission logic is in C.

The rule is the one in #251, which keeps it inside a much larger change. This PR is today's main plus that rule alone. The first version of this PR accepted any 10_04 device and conflicted with main; it has been replaced. Generated units come from om/units.sh with the pinned omc (2c824805…) and omega-bridge 3dc4a00e. The generated src/cudapp_om_amp_runtime.c is byte-identical to #251's. docs/PORT-AMPERE.md line references move with the edit.

Evidence

• Host: make test-ampere-runtime PASS. The GB10 identity is admitted. A wrong device id, a wrong vendor id, a partial class set and the Ampere class set are refused.
• GB10, scratch build of this head: build/hello_gpu PASS, 1048576 of 1048576 elements correct, 0 never written.
• Negative control on the same chip: origin/main hello_gpu prints GPU not available (stage: nvrm_open: unsupported SM 0xa04 or missing target classes).

Not covered

No speed claim. No other GB10 kernels were run on this head.

Unblocks #254, which needs the device to open. After this lands, #251 should drop its copy of the rule when it rebases.

🤖 Generated with Claude Code

[aien-spark]

Chip result (EXERCISED on GB10, not merged/qualified). spark-window 20261101T004827Z-b1ae7a, --gb10 exclusive --quiet none, 18091 and Flint down; window verdict: quiet machine. Head c5f6ec2d61e1da33919ff1c8ebb7668b286b15be, prebuilt from /tmp/prfix/c170-gb10fix.

Run	rc	Result
make hello-gpu	0	checked 1048576 of 1048576 elements: 0 wrong, 0 never written, PASS
gpu_elementwise_test --out	0	OMEGA_GPU_ELEMENTWISE_PASS, device_opens=1, 35 checks, 0 failed, mutants CAUGHT
gpu_attention_test --out	0	OMEGA_GPU_ATTENTION_PASS, device_opens=1, 219 checks, 0 failed
make hello-gpu-control (negative)	0 (target)	binary exit 1, RED as it must (the one unwritten answer caught)
unfixed main tree (99ec266, #170) make hello-gpu	2	nvrm_open: unsupported SM 0xa04 or missing target classes, NOT_RUN

So the 0xA04 admission refusal on main reproduces on the chip, and this commit clears it with the Blackwell class check still in place. Host: make test rc 0 (74 targets), new test red on main/green here, two mutants caught.

Disclosure: one unauthorized pre-window run of make hello-gpu-control from this tree happened at 2026-11-01T00:45:25Z during another value-only window (one 1 MiB vecadd); it is not counted here.

Logs: /tmp/prfix/window-20261101T004827Z/ on the Spark. No labels added.

[cursor]

Risk: medium. Approved. Cursor Security Agent passed with no findings that need human review; Bugbot was not present after the first poll and was skipped. The change is a bounded GB10 SM/PCI admission rule with negative-control tests, so it stays within the medium approval threshold.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}