docs: serving one model across two DGX Sparks via llama.cpp RPC - #1620
Open
Defilan wants to merge 3 commits into
Open
docs: serving one model across two DGX Sparks via llama.cpp RPC#1620Defilan wants to merge 3 commits into
Defilan wants to merge 3 commits into
Conversation
…pp RPC The pattern is proven (DeepSeek V4-Flash, 144GiB, both Sparks) and the runtime half is built and guarded (llmkube-runtimes defilantech#32: GGML_RPC=ON, ggml-rpc-server shipped in cuda-gb10), but the cluster-side wiring lived in no repo: the worker pod placement, the fabric-address security model, the --rpc extraArgs shape, and the wire verification that distinguishes a real two-node split from a Ready that quietly serves from one GPU. Written now because the next use is immediate (Ornith-1.5-397B Q4_K_M at 224GiB, which no single Spark can hold) and because the pattern is the two-node rehearsal for paired DGX-class hardware generally. Every manual step is annotated as a requirement on defilantech#1423, which is the operator feature this runbook papers over. Refs defilantech#1423 Signed-off-by: Christopher Maher <chris@mahercode.io>
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
Caught by executing the runbook: the object form is rejected by the apiserver with a strict decoding error. The manifest now matches what the live serve actually accepted. Signed-off-by: Christopher Maher <chris@mahercode.io>
…lure modes Attempt one on Ornith-1.5-397B Q4_K_M (224GiB) OOM-killed the unbounded worker mid-tensor-upload, which surfaced on the main side as a network- looking RPC send failure, and the main node's mmap page cache stacked on its UMA weight share drove ahazidgx1 unreachable. Three additions, all paid for: worker memory bounds are mandatory, --no-mmap is mandatory at this scale on unified memory (the Strix >64GB rule at twice the size), and the real per-node budget is ~118GiB allocatable, not 128. Also documents spec.resources.gpu as the taint-toleration trigger and the Terminating-hang when a worker's peer dies mid-transfer. Signed-off-by: Christopher Maher <chris@mahercode.io>
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.
What
Runbook for serving one model across two DGX Sparks via llama.cpp RPC:
docs/dgx-spark-two-node-rpc.md, cross-linked from the existing Spark guide.Why
Refs #1423
The pattern works and is in use (DeepSeek V4-Flash, 144GiB, spanning both Sparks), and its runtime half is built and CI-guarded (llmkube-runtimes #32:
GGML_RPC=ON,ggml-rpc-servershipped incuda-gb10, the deployed image). But the cluster-side wiring existed in no repository. Reconstructing it this week required archaeology across four repos, the NVIDIA clustering docs, and a commit message. The next use is immediate: Ornith-1.5-397B Q4_K_M is 224GiB, which no single Spark can hold.How
The runbook covers the full chain, each layer sourced:
dgx-rdma.ymland its verify play (RoCE /30s on the ConnectX-7 link)extraArgs: ["--rpc", ...], with the Model staging multi-artifact from MinIO viaspec.filesReadyalone has fooled this fleet beforeEvery manual step is annotated as a requirement on #1423, which is the operator feature this runbook papers over: the worker is unplaced, unmonitored, and unhealthchecked by the operator today.
Checklist
make testpasses locally - no code changedmake lintpasses locally - no Go changedgit commit -s) per DCOAssisted-by: Claude Opus 5(reconstructed the wiring from the fabric role, the runtimes commit, the NVIDIA clustering doc, and the surviving cluster artifacts; wrote the runbook).