Skip to content

feat(inference): materialize verified llama.cpp runtime #8338

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ericksoa merged 3 commits into from
Aug 5, 2026
Merged

feat(inference): materialize verified llama.cpp runtime #8338

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
b047f89
feat(inference): materialize verified llama.cpp runtime
ericksoa Aug 5, 2026
e28c88e
fix(inference): omit unenforced llama.cpp limits
ericksoa Aug 5, 2026
07ad2d4
Merge remote-tracking branch 'origin/main' into feat/llamacpp-verifie…
ericksoa Aug 5, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
100 changes: 18 additions & 82 deletions scripts/checks/run-llama-cpp-dgx-spark-qualification.mts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,10 @@ import net from "node:net";
import path from "node:path";
import { fileURLToPath } from "node:url";

import {
buildLlamaCppHostLocalDockerArgv,
type VerifiedLocalModelArtifact,
} from "../../src/lib/inference/llama-cpp/host-local-runtime.ts";
import {
LLAMA_CPP_DGX_SPARK_MODEL_PATH_PATTERN,
LLAMA_CPP_DGX_SPARK_QUALIFICATION_IMAGE_REPOSITORY,
Expand All @@ -27,8 +31,6 @@ const registryImage =
const registryOwnerLabel = "io.nvidia.nemoclaw.llama-cpp-qualification-owner";
const localImageRepository = LLAMA_CPP_DGX_SPARK_QUALIFICATION_IMAGE_REPOSITORY;
const expectedWorkflowRef = "NVIDIA/NemoClaw/.github/workflows/e2e.yaml@refs/heads/main";
const containerModelPath = "/models/Nemotron-3-Nano-30B-A3B-UD-Q4_K_XL.gguf";
const containerApiKeyPath = "/run/secrets/llama-cpp-api-key";
const maximumPlanBytes = 1024 * 1024;
const maximumDockerfileBytes = 1024 * 1024;
const maximumModelBytes = 128 * 1024 * 1024 * 1024;
Expand Down Expand Up @@ -303,86 +305,23 @@ export function buildServerContainerArgv(
apiKeyHostPath: string;
containerName: string;
imageReference: string;
modelHostPath: string;
model: VerifiedLocalModelArtifact;
networkName: string;
registryOwner: string;
runtimeGid: number;
runtimeUid: number;
},
): string[] {
const { resources } = plan.recipe.runtime;
const { serve } = plan.recipe;
const runtimeUid = requiredInteger(options.runtimeUid, "runtime uid", 1, 2_147_483_647);
const runtimeGid = requiredInteger(options.runtimeGid, "runtime gid", 1, 2_147_483_647);
return [
"run",
"--detach",
"--name",
options.containerName,
"--label",
`${registryOwnerLabel}=${options.registryOwner}`,
"--network",
options.networkName,
"--user",
`${runtimeUid}:${runtimeGid}`,
"--publish",
`127.0.0.1::${serve.port}`,
"--read-only",
"--cap-drop",
"ALL",
"--security-opt",
"no-new-privileges=true",
"--memory",
`${resources.memoryBytes}b`,
"--memory-swap",
`${resources.memoryBytes}b`,
"--pids-limit",
String(resources.pidsLimit),
"--gpus",
"1",
"--tmpfs",
`/tmp:rw,noexec,nosuid,nodev,size=${resources.writableStorageBytes},uid=${runtimeUid},gid=${runtimeGid},mode=1777`,
"--mount",
`type=bind,source=${options.modelHostPath},target=${containerModelPath},readonly`,
"--mount",
`type=bind,source=${options.apiKeyHostPath},target=${containerApiKeyPath},readonly`,
options.imageReference,
"--model",
containerModelPath,
"--alias",
plan.recipe.model.servedName,
"--host",
"0.0.0.0",
"--port",
String(serve.port),
"--gpu-layers",
"all",
"--ctx-size",
String(serve.contextSize),
"--parallel",
String(serve.slots),
"--sleep-idle-seconds",
String(serve.idleSleepSeconds),
"--batch-size",
String(serve.batchSize),
"--ubatch-size",
String(serve.microBatchSize),
"--cache-type-k",
serve.kvCache.key,
"--cache-type-v",
serve.kvCache.value,
"--flash-attn",
"on",
"--timeout",
String(serve.limits.requestTimeoutSeconds),
"--api-key-file",
containerApiKeyPath,
"--metrics",
"--no-ui",
"--no-slots",
"--no-mmproj",
"--no-agent",
];
return buildLlamaCppHostLocalDockerArgv(plan.recipe, {
apiKeyHostPath: options.apiKeyHostPath,
containerName: options.containerName,
imageReference: options.imageReference,
model: options.model,
network: { isolation: "docker-internal", name: options.networkName },
ownerLabel: { name: registryOwnerLabel, value: options.registryOwner },
runtimeGid: options.runtimeGid,
runtimeUid: options.runtimeUid,
});
}

export function validateStartupLog(log: string): { offloadedLayers: number; totalLayers: number } {
Expand Down Expand Up @@ -560,10 +499,7 @@ function validateCandidateCheckout(invocation: QualificationInvocation): string
return root;
}

function hashModelFile(
modelHostPath: string,
plan: QualificationPlan,
): { digest: string; sizeBytes: number } {
function hashModelFile(modelHostPath: string, plan: QualificationPlan): VerifiedLocalModelArtifact {
if (path.basename(modelHostPath) !== plan.recipe.model.file.path) {
throw new Error("model file name does not match the qualification plan");
}
Expand Down Expand Up @@ -607,7 +543,7 @@ function hashModelFile(
if (digest !== plan.recipe.model.file.digest) {
throw new Error("model file digest does not match the qualification plan");
}
return { digest, sizeBytes: after.size };
return { digest, hostPath: modelHostPath, sizeBytes: after.size };
} finally {
fs.closeSync(descriptor);
}
Expand Down Expand Up @@ -943,7 +879,7 @@ async function runQualification(
apiKeyHostPath,
containerName: names.containerName,
imageReference: image.reference,
modelHostPath: invocation.modelHostPath,
model,
networkName: names.networkName,
registryOwner: names.registryOwner,
runtimeGid,
Expand Down
Loading
Loading