Repository navigation
Conversation
dennis-upbound
left a comment
dennis-upbound
left a comment
There was a problem hiding this comment.
The vendor tag needs to cover the metrics exporter, not just the driver.
DCGM ships as an NVIDIA-stack component (nvidia-dcgm, per the allowlist in #456), so tagging it NVIDIA is correct — and it means an AMD-only cluster installs no GPU exporter at all. The telemetry design in #363 names DCGM as the source for temperature, throttling, ECC and interconnect errors, which is the fault taxonomy behind a drain. Those go silent on AMD, and nothing says so.
AMD's GPU Operator ships a Device Metrics Exporter with Prometheus-format temperature, utilization, memory and power, so the fix looks like the driver case:
Chart(key="amd-device-metrics-exporter", accelerator_vendor="AMD", ...)Worth a line in the design saying a vendor's stack includes its exporter. Keeping modelplane_gpu_* meaning the same thing whichever exporter produced it is #363's problem, not this one — I'll take that side.
The Upbound Inference environment runs this stack against real traffic and records what it found. Five things from it. The GPU exporter is vendor-dependent, which modelplaneai#432 makes true rather than hypothetical: DCGM on NVIDIA, the Device Metrics Exporter on AMD, chosen by the serving stack alongside the driver. This document had named DCGM as the source, which would have left an AMD cluster with no GPU series and no statement that it had. modelplane_gpu_* means the same thing whichever exporter produced it, which is the point of renaming at all. Three metrics that environment's dashboard has and this set did not. Termination reason is the important one: a response cut off at the token cap reports length and still returns 200 with a well-formed body, and every layer above reads it as success. A timeout severing a stream looks the same from the other side. That environment has now been truncated twice, once by a route timeout and once by a load balancer, and neither showed in a status code. Tool-call parse outcome fails the same quiet way, returning the unparsed call as text. Tensor-pipe activity is the third, and is closer to useful for inference than the graphics-engine figure beside it. Two traps, both observed rather than reasoned about. An engine container's port is unnamed by default, so selecting it by name matches nothing and reports nothing, which is how that environment's first PodMonitor scraped no engines at all. Modelplane has to name it, and matching by number instead would find the sidecar on a disaggregated pod. And vLLM's model_name is the LLM while DCGM's modelName is the card: one underscore and one capital apart, and a join across them pairs every model with every GPU type. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Dennis Ramdass <dennis@upbound.io>
3fe0dab to
ddd9fe7
Compare
Signed-off-by: Christopher Haar <christopher.haar@upbound.io>
ddd9fe7 to
1b11f26
Compare
negz
left a comment
negz
left a comment
There was a problem hiding this comment.
Thanks, this addresses most of my earlier comments. A few more below.
| acceptable: the stack components are per-cluster Helm releases holding no | ||
| state worth keeping, the nodes they served leave with the pools that claimed | ||
| them, and a stack reinstalls cleanly when a class returns. It is also the |
There was a problem hiding this comment.
On Existing the nodes stick around when their pool goes away. If someone removes the last NVIDIA pool while ModelDeployments are still running on it, nothing orders uninstalling the NVIDIA DRA driver against deleting the engine pods. If the driver goes first, the kubelet can't unprepare their claims, and the pods hang in Terminating until someone force deletes them.
I don't think this design needs to fix that, though. We should probably handle deleting a pool with running replicas better instead.
Signed-off-by: Christopher Haar <christopher.haar@upbound.io>
3 participants
Description of your changes
The serving stack we propose in #422 is NVIDIA-only by construction: every clouds component list installs the NVIDIA operator or DRA driver unconditionally, the accelerator taint and toleration are nvidia.com/gpu, and nothing checks that the devices an InferenceClass declares match the stack its cluster gets. AMD Instinct GPUs are reaching the clouds Modelplane targets, CPU inference (#509) adds a third device stack, and TPU and Trainium sit
behind the same gap: a cluster may carry devices served by more than one DRA driver, and the stack must install each driver's components exactly where that driver's devices are.
The design identifies the accelerator by the DRA driver name the class already carries (
gpu.nvidia.com,gpu.amd.com,dra.cpu), with no vendor enum or map. The cluster composition derives the claimed drivers onto ServingStackspec.accelerators, a component that is part of one driver's device stack is tagged with that driver's name, and the component join installs tagged components only where their driver is present. Every cloud participates, hand-written and generated:spec.acceleratorsis always sent, the generated halves tag through one more value pergenerate.pyALLOW entry, and each cloud's supported set is derived from its tags. Mismatches fail early in the composition functions: on the class, on the class/cluster pairing, and in the join. CPU is just another accelerator: dra-driver-cpu becomes one tagged chart, replacing #509'sspec.cpu.poolsfield and transform.Fixes #
I have:
nix flake check(or./nix.sh flake check) and made sure it passes.git commit -s.