draft: hardware partition and XCP capacity - #17
Draft
fishman wants to merge 12 commits into
Draft
Conversation
fishman
commented
Aug 18, 2026
fishman
commented
- testdata: add MI355X node fixtures for hardware-free regression tests
- amdgpu: configurable sysfs root, amdsmi memory-partition and profile lookups
- plugin: injectable AMD SMI seams, fixture-backed registration test
- README: document partition profiles and mode-change semantics
- allocator, utils, labeller: nil-guard stale device ids, dual allocation keys
- plugin, amdgpu: advertise per-XCP capacity as an even share of the GPU
- testdata, README: xgmi/topology fixtures, per-XCP capacity, memory-partition rationale
|
Important Review skippedDraft detected. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
[APPROVALNOTIFIER] This PR is NOT APPROVED This pull-request has been approved by: fishman The full list of commands accepted by this bot can be found here. DetailsNeeds approval from an approver in each of these files:Approvers can indicate their approval by writing |
added 7 commits
August 19, 2026 11:19
Captured from nodepool-gpu-7b3016af2911 (kernel 6.8.0-136): KFD topology, amdgpu_xcp devices, per-GPU partition state, and amd-smi static, partition, memory-partition and per-GPU partition JSON. SPI captures document the zero-registration behavior on this kernel (XCP render minors absent from KFD topology). Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
…lookups GetAMDGPUs takes a sysfs root (defaults to /sys) so discovery can run against captured fixtures. GetAMDSCurrentMemoryPartitions and GetAMDGPUPartitionProfiles extend the cgo AMD SMI layer: the first resolves the current NPS mode per BDF, the second the accelerator partition profiles (SPX/DPX/QPX/CPX with partition geometry) from amdsmi_get_gpu_accelerator_partition_profile_config. Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
WithSysfsRoot, WithAmdSMI and WithAMDSPartitionProfiles constructor options let registration run without hardware or AMD SMI; nil means the real /sys and cgo implementations. Registration overrides the sysfs memory partition with the amdsmi value and advertises available partition profiles per whole GPU in customInfo. TestRegistrationFromFixture drives the full flow against the captured MI355X tree and pins the kernel 6.8 zero-registration behavior; TestPartitionProfilesFromFixture pins the fixture profile truth. Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
Explains the per-GPU profile table (SPX/DPX/QPX/CPX with partition x XCC geometry), profile_index semantics, the idle requirement for partition changes, and the per-GPU compute vs node-wide memory reload distinction. Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
…on keys Best-effort policy skips stale kubelet device ids across plugin restarts instead of panicking in sort.Slice. Allocation reads the committed hami.io/amd-devices-allocated payload and falls back to the fork's own to-allocate annotation. Labeller Dockerfile picks up the glibc patch. Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
XCP partitions share their parent's PCI BDF. Their registered capacity is now the whole-GPU VRAM and CU count divided by the number of XCP partitions, computed once at start from the libdrm whole-GPU read. Floor division under-advertises rather than over-committing. Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
…rtition rationale Store amd-smi xgmi and topology node captures next to the existing fixtures. Document the profile storage model (register annotation, mirroring HAMi's MIG storage), the per-XCP capacity derivation, and why NPS memory partitions are not an allocation mechanism for HAMi. Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
added 2 commits
August 19, 2026 11:23
Generate a per-GPU catalog of valid compute partition x NPS combinations with the NPS->physical memory count table at plugin Start() (gfx950: spx/dpx/qpx x nps1/4/8). Held in memory only, never written to disk (immutable-OS friendly); complementable later by HAMi default templates. Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
MI355X (gfx950) admission fixes: - isSchedulableTopologyKey: spx/dpx/qpx whole GPUs have no amdgpu_xcp_ children in KFD topology and keep their own DRI card, so they register. Only XCP parents (BDF with amdgpu_xcp_ children) are skipped. Previously all 8 GPUs were skipped, leaving zero registered devices, an empty sortedBDFs, and Allocate unable to resolve upstream AMDGPU-<i> ids. - GetPreferredAllocation: kubelet passes checkpoint-persisted ids (bare BDFs from the pre-fork plugin) that the split-keyed allocator cannot map. GPA is advisory and Allocate resolves the pod annotation, so echo kubelet's own ids instead of failing admission. Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
added 3 commits
August 19, 2026 15:03
…ition flip Per-node operating mode (NVIDIA mig.config-style), defaulting to cu: - dp.operatingMode chart value / OPERATING_MODE env, overridable per node via the hami.io/amd-operating-mode annotation: "cu" registers soft whole-GPU devices (Count 10, unchanged behavior); "partition" registers hard compute partitions (id#mode, Count 1) and publishes only hard entries to kubelet. - dp.computePartition / COMPUTE_PARTITION env, overridable via hami.io/amd-compute-partition: in partition mode the plugin flips every GPU to the target profile through amdsmi_set_gpu_accelerator_partition_profile at startup; no manual amd-smi on the node. Busy GPUs fail (status 30) and keep their current mode. - XCP capacity divisor now uses the AMD SMI profile NumPartitions of the current compute type instead of amdgpu_xcp_ child counts (7 XCD chiplets per GPU regardless of partition mode). - Docs: configuration.md operating-mode section, new mi355x-partitioning.md (kernel caveats, verification), QPX example at example/vllm-qpx-serve/ (deployment with gpumem/gpucores requests, HAMi device-config amd section, README). Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
A busy GPU refuses the compute-partition flip (AMD SMI status 30) and stays in its old mode, leaving the node heterogeneous. Previously getResourceList aborted the plugin before Start() ran, so the flip never executed and the node crash-looped forever. - main.go: heterogeneous node with single strategy logs a warning and reports all devices under "gpu" instead of exiting - plugin.go: ListAndWatch publishes every device when the resource name is not a per-partition-type key - plugin.go: log flip duration at glog V(2) (measured 435ms for 7 GPUs) Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
Skipping unknown ids prevents the panic, but a silent fix hides the cause: another device plugin running on the same node merged its devices into kubelet's list. Log a warning naming the likely cause and the remedy (disable the other plugin, restart this one). Signed-off-by: Reza Jelveh <fishmangit@dynamia.ai>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.