You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
GPU matching by device-name keyword is fragile — match by VID:DID instead #51
Follow-up to #49 / #50. PR #50 correctly routed GPUs to the GPU guide (program=sptg) but matches by device-model keyword, which is not robust. Testing against a customer's real GPU fleet (target ESXi 9.1) exposed two failure modes:
Device (vSphere name → tool keyword)
Keyword verdict
Correct?
NVIDIA A40
TRUE
✓
NVIDIA H100 NVL PCIe 94GB
FALSE
✗ false negative
NVIDIA H200 NVL
TRUE
✓
NVIDIA L4
TRUE
✓ but fragile (also matched L40S/L40)
NVIDIA L40S
TRUE
✓
iLO5 VGA
FALSE
✓ (Matrox BMC VGA, not a compute GPU)
Failure modes
False negative from vSphere name noise. The host reports NVIDIA H100 NVL PCIe 94GB; the guide lists NVIDIA H100 NVL. The sptg keyword search ANDs tokens, so the extra PCIe 94GB yields count=0 → FALSE, even though the card is certified.
Latent false positive from loose matching. Keyword L4 also returns L40S/L40 (prefix match). The verdict logic (count>0 AND body contains the release) can then be satisfied by a sibling model, certifying a device that itself may not be supported for the target release.
Fix
Match GPUs by exact VID:DID instead of by name. The sptg response already carries VID/DID per row in hoverData, and the whole GPU guide is only ~63 rows. Fetch it once per run, build map["vid:did"] → [supported ESXi releases], and verify each GPU by its PCI VID:DID against its own row's esxVersion list (exact release match).
Validated against the live guide — every device above resolves correctly (H100 NVL PCIe 94GB → TRUE, iLO5 VGA → FALSE), and all resolve FALSE for ESXi 7.0 (these cards' guide rows don't list it). GPUs stay SKIPPED under -offline; a failed guide fetch yields ERROR (undetermined), not a false FALSE. Note: ~8 guide rows carry no VID/DID (SXM/NVSwitch form factors); those aren't PCI-pluggable discrete cards and are out of scope for PCI-ID matching.
Activity
Never report a CPU as uncertified when its CPUID is unavailable (#62)
Never report a CPU as uncertified when its CPUID is unavailable (#62)