fix(inference): bound Spark llama.cpp request bodies

managed-inference/images/llama-cpp/Dockerfile

@@ -4,7 +4,7 @@
 ARG CUDA_DEV_IMAGE
 ARG CUDA_RUNTIME_IMAGE
 
 FROM ${CUDA_DEV_IMAGE} AS build
 
 ARG LLAMA_CPP_REVISION
 ARG LLAMA_CPP_ARCHIVE_SHA256
@@ -29,11 +29,11 @@
       build-essential=12.10ubuntu1 \
       ca-certificates=20260601~24.04.1 \
       cmake=3.28.3-1build7 \
-      curl=8.5.0-2ubuntu10.11 \
+      curl=8.5.0-2ubuntu10.12 \
       g++-14=14.2.0-4ubuntu2~24.04.1 \
       gcc-14=14.2.0-4ubuntu2~24.04.1 \
       golang-go=2:1.22~2build1 \
-      libcurl4-openssl-dev=8.5.0-2ubuntu10.11 \
+      libcurl4-openssl-dev=8.5.0-2ubuntu10.12 \
       libssl-dev=3.0.13-0ubuntu3.12 \
     && rm -rf /var/lib/apt/lists/*
 
@@ -98,7 +98,7 @@
     && chmod 0555 /opt/llama.cpp/licenses/go \
     && chmod 0444 /opt/llama.cpp/licenses/go/copyright
 
 FROM ${CUDA_RUNTIME_IMAGE} AS runtime
 
 ARG CUDA_DEV_IMAGE
 ARG CUDA_RUNTIME_IMAGE
@@ -123,7 +123,7 @@
 RUN apt-get update \
     && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
       ca-certificates=20260601~24.04.1 \
-      libcurl4t64=8.5.0-2ubuntu10.11 \
+      libcurl4t64=8.5.0-2ubuntu10.12 \
       libgomp1=14.2.0-4ubuntu2~24.04.1 \
     && groupadd --gid "${RUNTIME_GID}" nemoclaw-llama \
     && useradd --uid "${RUNTIME_UID}" --gid "${RUNTIME_GID}" \

managed-inference/images/llama-cpp/image.yaml

@@ -72,6 +72,7 @@ spec:
         - context-window
         - authentication
         - malformed-request
+        - request-body-limit
         - cancellation
         - client-timeout
         - log-redaction
@@ -89,8 +90,8 @@ spec:
 
   source:
     repository: https://github.com/ggml-org/llama.cpp
-    revision: 22dc605c4ead20e36f447cc67b55ef87e523bd55
-    archiveSha256: sha256:975f70723e053785e894f4e1d9cf770f2f1a7bc762fd3af174ff5635014108b6
+    revision: 8e7f22b67ef4667b4ddd50230771287f328cfb3f
+    archiveSha256: sha256:45a24299e7a24410624489d19924d492bc71a120fa17d9b7cb32f6d5c4f1aed0
 
   cuda:
     developmentBase: docker.io/nvidia/cuda@sha256:ef2203909e80b8b976cfc672f7e2ae2b00bc0e25c404ee86d89e10a3802f1c52
@@ -115,11 +116,11 @@ spec:
       build-essential: 12.10ubuntu1
       ca-certificates: 20260601~24.04.1
       cmake: 3.28.3-1build7
-      curl: 8.5.0-2ubuntu10.11
+      curl: 8.5.0-2ubuntu10.12
       g++-14: 14.2.0-4ubuntu2~24.04.1
       gcc-14: 14.2.0-4ubuntu2~24.04.1
       golang-go: 2:1.22~2build1
-      libcurl4-openssl-dev: 8.5.0-2ubuntu10.11
+      libcurl4-openssl-dev: 8.5.0-2ubuntu10.12
       libssl-dev: 3.0.13-0ubuntu3.12
     cmake:
       ggmlBackendDl: true
@@ -162,7 +163,7 @@ spec:
       - /usr/bin/sh
     packages:
       ca-certificates: 20260601~24.04.1
-      libcurl4t64: 8.5.0-2ubuntu10.11
+      libcurl4t64: 8.5.0-2ubuntu10.12
       libgomp1: 14.2.0-4ubuntu2~24.04.1
     writablePaths:
       - /tmp

managed-inference/images/llama-cpp/request-guard/main.go

@@ -97,27 +97,29 @@ func requireExactCommandMarker(command []string, option string) error {
 
 func validateSupportedCommandOptions(command []string) error {
 	allowed := map[string]bool{
-		"--alias":              true,
-		"--api-key-file":       true,
-		"--batch-size":         true,
-		"--cache-type-k":       true,
-		"--cache-type-v":       true,
-		"--ctx-size":           true,
-		"--flash-attn":         true,
-		"--gpu-layers":         true,
-		"--host":               true,
-		"--metrics":            false,
-		"--model":              true,
-		"--no-agent":           false,
-		"--no-mmproj":          false,
-		"--no-slots":           false,
-		"--no-ui":              false,
-		"--n-predict":          true,
-		"--parallel":           true,
-		"--port":               true,
-		"--sleep-idle-seconds": true,
-		"--timeout":            true,
-		"--ubatch-size":        true,
+		"--alias":                true,
+		"--api-key-file":         true,
+		"--batch-size":           true,
+		"--cache-type-k":         true,
+		"--cache-type-v":         true,
+		"--chat-template-kwargs": true,
+		"--ctx-size":             true,
+		"--flash-attn":           true,
+		"--gpu-layers":           true,
+		"--host":                 true,
+		"--jinja":                false,
+		"--metrics":              false,
+		"--model":                true,
+		"--no-agent":             false,
+		"--no-mmproj":            false,
+		"--no-slots":             false,
+		"--no-ui":                false,
+		"--n-predict":            true,
+		"--parallel":             true,
+		"--port":                 true,
+		"--sleep-idle-seconds":   true,
+		"--timeout":              true,
+		"--ubatch-size":          true,
 	}
 	seen := make(map[string]bool, len(allowed))
 	for index := 0; index < len(command); index++ {

managed-inference/images/llama-cpp/request-guard/main_test.go

@@ -119,6 +119,8 @@ func TestParseConfigRequiresEveryDeclaredValue(t *testing.T) {
 		"--port", "8082",
 		"--api-key-file", llamaServerAPIKeyPath,
 		"--n-predict", "4096",
+		"--jinja",
+		"--chat-template-kwargs", `{"reasoning_strength":"low"}`,
 		"--no-ui",
 		"--no-slots",
 		"--no-mmproj",

managed-inference/recipes/llama-cpp.muse-glimmer-30b.spark-single.v1.yaml

@@ -88,7 +88,7 @@ spec:
       value: f16
     speculativeDecoding: disabled
     limits:
-      maxRequestBodyBytes: 1048576
+      maxRequestBodyBytes: 16384
       maxRequestHeaderBytes: 32768
       maxOutputTokens: 4096
       requestTimeoutSeconds: 900

managed-inference/recipes/llama-cpp.nemotron-3-nano-30b-a3b.spark-single.v1.yaml

@@ -16,7 +16,7 @@ spec:
     technology: llama.cpp
     source:
       repository: ggml-org/llama.cpp
-      revision: 22dc605c4ead20e36f447cc67b55ef87e523bd55
+      revision: 8e7f22b67ef4667b4ddd50230771287f328cfb3f
 
   model:
     id: unsloth/Nemotron-3-Nano-30B-A3B-GGUF
@@ -86,7 +86,7 @@ spec:
       value: f16
     speculativeDecoding: disabled
     limits:
-      maxRequestBodyBytes: 1048576
+      maxRequestBodyBytes: 32768
       maxRequestHeaderBytes: 32768
       maxOutputTokens: 4096
       requestTimeoutSeconds: 900

scripts/checks/export-llama-cpp-image-config.mts

@@ -625,11 +625,11 @@ export function loadLlamaCppImageConfig(
     "build-essential": "12.10ubuntu1",
     "ca-certificates": "20260601~24.04.1",
     cmake: "3.28.3-1build7",
-    curl: "8.5.0-2ubuntu10.11",
+    curl: "8.5.0-2ubuntu10.12",
     "g++-14": "14.2.0-4ubuntu2~24.04.1",
     "gcc-14": "14.2.0-4ubuntu2~24.04.1",
     "golang-go": "2:1.22~2build1",
-    "libcurl4-openssl-dev": "8.5.0-2ubuntu10.11",
+    "libcurl4-openssl-dev": "8.5.0-2ubuntu10.12",
     "libssl-dev": "3.0.13-0ubuntu3.12",
   };
   const expectedCompiler = {
@@ -639,7 +639,7 @@ export function loadLlamaCppImageConfig(
   };
   const expectedRuntimePackages = {
     "ca-certificates": "20260601~24.04.1",
-    libcurl4t64: "8.5.0-2ubuntu10.11",
+    libcurl4t64: "8.5.0-2ubuntu10.12",
     libgomp1: "14.2.0-4ubuntu2~24.04.1",
   };
   const expectedRequiredPaths = [

scripts/checks/llama-cpp-dgx-spark-protocol-qualification.mts

@@ -3,6 +3,7 @@
 
 import {
   LLAMA_CPP_DGX_SPARK_PROTOCOL_PROBES,
+  LLAMA_CPP_DGX_SPARK_REJECTED_REQUEST_BODY_BYTES,
   LLAMA_CPP_DGX_SPARK_REQUIRED_METRIC_SERIES,
   type LlamaCppDgxSparkExecutionPlan,
   type LlamaCppDgxSparkQualificationReceipt,
@@ -164,6 +165,51 @@ function jsonRequest(
   };
 }
 
+function exactSizeChatRequest(model: string, targetBytes: number): string {
+  const request = {
+    max_tokens: 1,
+    messages: [{ content: "", role: "user" }],
+    model,
+    temperature: 0,
+  };
+  const emptyBody = JSON.stringify(request);
+  const contentBytes = targetBytes - new TextEncoder().encode(emptyBody).byteLength;
+  if (contentBytes < 0) throw new Error("request-body probe target is too small");
+  request.messages[0].content = "x".repeat(contentBytes);
+  const body = JSON.stringify(request);
+  if (new TextEncoder().encode(body).byteLength !== targetBytes) {
+    throw new Error("request-body probe did not construct the exact declared size");
+  }
+  return body;
+}
+
+function exactSizeJsonRequest(
+  authorization: string,
+  body: string,
+  timeoutMilliseconds: number,
+): RequestInit {
+  return {
+    body,
+    headers: {
+      Authorization: authorization,
+      "Content-Type": "application/json",
+    },
+    method: "POST",
+    signal: requestSignal(timeoutMilliseconds),
+  };
+}
+
+function validateRequestBodyLimitError(value: unknown): void {
+  if (
+    !isRecord(value) ||
+    !isRecord(value.error) ||
+    value.error.code !== "request_body_too_large" ||
+    value.error.type !== "invalid_request_error"
+  ) {
+    throw new Error("request-body limit probe did not return the declared error contract");
+  }
+}
+
 function usageFrom(value: unknown): ProtocolEvidence["usage"] {
   if (!isRecord(value)) throw new Error("chat usage was not returned");
   const promptTokens = value.prompt_tokens;
@@ -727,6 +773,73 @@ export async function runLlamaCppDgxSparkProtocolQualification(options: {
   await expectStatus(malformedResponse, 400, bounds.maxResponseBytes, "malformed-request probe");
   executedProbes.add("malformed-request");
 
+  const acceptedRequestBytes = plan.recipe.serve.limits.maxRequestBodyBytes;
+  const acceptedResponse = await fetchImpl(
+    chatUrl,
+    exactSizeJsonRequest(
+      authorization,
+      exactSizeChatRequest(model, acceptedRequestBytes),
+      timeoutMilliseconds,
+    ),
+  );
+  await expectStatus(
+    acceptedResponse,
+    200,
+    bounds.maxResponseBytes,
+    "request-body boundary probe",
+  );
+
+  const rejectedResponse = await fetchImpl(
+    chatUrl,
+    exactSizeJsonRequest(
+      authorization,
+      exactSizeChatRequest(model, LLAMA_CPP_DGX_SPARK_REJECTED_REQUEST_BODY_BYTES),
+      timeoutMilliseconds,
+    ),
+  );
+  validateRequestBodyLimitError(
+    await readJson(
+      rejectedResponse,
+      413,
+      bounds.maxResponseBytes,
+      "oversized request-body probe",
+    ),
+  );
+
+  const continuationHealthResponse = await fetchImpl(`${baseUrl}/health`, {
+    headers: { Authorization: authorization },
+    signal: requestSignal(bounds.clientTimeoutMilliseconds),
+  });
+  await expectStatus(
+    continuationHealthResponse,
+    200,
+    bounds.maxResponseBytes,
+    "request-body continuation health probe",
+  );
+  const bodyLimitContinuationResponse = await fetchImpl(
+    chatUrl,
+    jsonRequest(
+      authorization,
+      {
+        max_tokens: bounds.maxTokens.synchronousChat,
+        messages: [{ content: "Return one short continuation token.", role: "user" }],
+        model,
+        temperature: 0,
+      },
+      timeoutMilliseconds,
+    ),
+  );
+  validateChatCompletionResponse(
+    await readJson(
+      bodyLimitContinuationResponse,
+      200,
+      bounds.maxResponseBytes,
+      "request-body continuation completion probe",
+    ),
+    model,
+  );
+  executedProbes.add("request-body-limit");
+
   const synchronousResponse = await fetchImpl(
     chatUrl,
     jsonRequest(
@@ -914,6 +1027,17 @@ export async function runLlamaCppDgxSparkProtocolQualification(options: {
     },
     health: { httpStatus: 200, ok: true },
     malformedRequest: { httpStatus: 400, ok: true },
+    requestBodyLimit: {
+      acceptedBytes: acceptedRequestBytes,
+      acceptedHttpStatus: 200,
+      continuationHealthHttpStatus: 200,
+      continuationHttpStatus: 200,
+      errorCode: "request_body_too_large",
+      errorType: "invalid_request_error",
+      ok: true,
+      rejectedBytes: LLAMA_CPP_DGX_SPARK_REJECTED_REQUEST_BODY_BYTES,
+      rejectedHttpStatus: 413,
+    },
     models: { httpStatus: 200, model, ok: true },
     metrics,
     properties: propertiesEvidence.properties,