@infersec/conduit 1.90.2 → 1.90.3

package/dist/cli.js

@@ -20042,16 +20042,16 @@ const InferenceAgentMachineMetadataSchema = object$1({
         model: string$1().nullable(),
         physicalCores: number$1().int().positive().nullable()
     }),
-    exllamav3Version: string$1().nullable(),
+    exllamav3Version: string$1().nullable().default(null),
     gpus: array(InferenceAgentMachineGPUSchema),
     hostname: string$1(),
-    llamaCppVersion: string$1().nullable(),
+    llamaCppVersion: string$1().nullable().default(null),
     machineID: string$1(),
     memory: object$1({
         availableBytes: number$1().int().nonnegative().nullable(),
         totalBytes: number$1().int().nonnegative().nullable()
     }),
-    mlxlmVersion: string$1().nullable(),
+    mlxlmVersion: string$1().nullable().default(null),
     os: object$1({
         arch: string$1(),
         platform: string$1(),
@@ -20059,9 +20059,9 @@ const InferenceAgentMachineMetadataSchema = object$1({
         type: string$1().nullable(),
         version: string$1().nullable()
     }),
-    sglangVersion: string$1().nullable(),
-    tensorrtLlmVersion: string$1().nullable(),
-    vllmVersion: string$1().nullable()
+    sglangVersion: string$1().nullable().default(null),
+    tensorrtLlmVersion: string$1().nullable().default(null),
+    vllmVersion: string$1().nullable().default(null)
 });
 const InferenceAgentMachineReportPayloadSchema = object$1({
     machine: InferenceAgentMachineMetadataSchema
@@ -135290,7 +135290,12 @@ function normalizeMegabytes(value) {
 function normalizeBusAddress(bus) {
     if (!bus)
         return null;
-    return bus.toLowerCase().replace(/^0000:/, "");
+    // PCI addresses arrive with varying domain widths across tools:
+    // nvidia-smi "00000000:09:00.0", lspci/sysfs "0000:09:00.0",
+    // systeminformation "09:00.0". Reduce all to the trailing "bus:device.function"
+    // so the same GPU matches across sources.
+    const match = bus.toLowerCase().match(/([0-9a-f]{1,2}:[0-9a-f]{2}\.[0-9a-f])$/);
+    return match ? match[1] : bus.toLowerCase();
 }
 function resolveCpuValue(value) {
     if (typeof value === "number" && Number.isFinite(value)) {
@@ -135438,76 +135443,184 @@ async function detectVRAMViaRocmSmi() {
         return [];
     }
 }
-function buildMergedGPUs(options) {
-    const { lspciGPUs, rocmVRAM, siGPUs, sysfsVRAMMap } = options;
-    const siByBus = new Map();
-    for (const gpu of siGPUs) {
-        const key = normalizeBusAddress(gpu.bus);
-        if (key)
-            siByBus.set(key, gpu);
+const NVIDIA_SMI_TIMEOUT_MS$1 = 10_000;
+function parseNvidiaSmiNumber(value) {
+    if (!value)
+        return null;
+    // Unified-memory devices (e.g. NVIDIA GB10) report "[N/A]"; treat as unknown
+    const parsed = parseInt(value, 10);
+    return Number.isFinite(parsed) && parsed >= 0 ? parsed : null;
+}
+function parseNvidiaSmiMemoryBytes(value) {
+    const mib = parseNvidiaSmiNumber(value);
+    return mib !== null ? normalizeMegabytes(mib) : null;
+}
+async function readSystemMemoryBytes({ logger }) {
+    try {
+        const contents = await readFile("/proc/meminfo", "utf8");
+        let availableBytes = null;
+        let totalBytes = null;
+        for (const line of contents.split("\n")) {
+            const match = line.match(/^(MemTotal|MemAvailable):\s+(\d+)\s*kB/i);
+            if (!match)
+                continue;
+            const bytes = parseInt(match[2], 10) * 1024;
+            if (match[1].toLowerCase() === "memtotal") {
+                totalBytes = bytes;
+            }
+            else {
+                availableBytes = bytes;
+            }
+        }
+        return { availableBytes, totalBytes };
+    }
+    catch (error) {
+        logger.debug("Failed to read system memory", { error: asError(error) });
+        return { availableBytes: null, totalBytes: null };
+    }
+}
+async function detectGPUsViaNvidiaSmi({ logger }) {
+    try {
+        const { stdout } = await execa("nvidia-smi", [
+            "--query-gpu=name,pci.bus_id,driver_version,memory.total,memory.used,memory.free,temperature.gpu",
+            "--format=csv,noheader,nounits"
+        ], { timeout: NVIDIA_SMI_TIMEOUT_MS$1 });
+        const lines = stdout.split("\n").filter(line => line.trim().length > 0);
+        const gpus = [];
+        for (const line of lines) {
+            const [name, bus, driverVersion, memTotal, memUsed, memFree, temperature] = line
+                .split(",")
+                .map(part => part.trim());
+            gpus.push({
+                bus: normalizeBusAddress(bus),
+                driverVersion: driverVersion || null,
+                memoryFreeBytes: parseNvidiaSmiMemoryBytes(memFree),
+                memoryTotalBytes: parseNvidiaSmiMemoryBytes(memTotal),
+                memoryUsedBytes: parseNvidiaSmiMemoryBytes(memUsed),
+                model: name || null,
+                temperatureCelsius: parseNvidiaSmiNumber(temperature),
+                vendor: "NVIDIA"
+            });
+        }
+        // Unified-memory devices (e.g. NVIDIA GB10 on DGX Spark) report "[N/A]" for
+        // memory. Their compute pool is system RAM, so fall back to /proc/meminfo.
+        if (gpus.some(gpu => gpu.memoryTotalBytes === null)) {
+            const systemMemory = await readSystemMemoryBytes({ logger });
+            const total = systemMemory.totalBytes;
+            if (total !== null) {
+                const free = systemMemory.availableBytes;
+                const used = free !== null ? total - free : null;
+                for (const gpu of gpus) {
+                    if (gpu.memoryTotalBytes !== null)
+                        continue;
+                    gpu.memoryFreeBytes = free;
+                    gpu.memoryTotalBytes = total;
+                    gpu.memoryUsedBytes = used;
+                }
+            }
+        }
+        return gpus;
     }
+    catch (error) {
+        logger.debug("nvidia-smi GPU query failed", { error: asError(error) });
+        return [];
+    }
+}
+function buildMergedGPUs(options) {
+    const { lspciGPUs, nvidiaGPUs, rocmVRAM, siGPUs, sysfsVRAMMap } = options;
     const rocmByBus = new Map();
     for (const entry of rocmVRAM) {
         const key = normalizeBusAddress(entry.bus);
         if (key)
             rocmByBus.set(key, entry);
     }
-    const seen = new Set();
-    const merged = [...siGPUs];
+    const byBus = new Map();
+    // Priority 1 — nvidia-smi: authoritative for NVIDIA (name, driver, VRAM, temp, bus)
+    for (const gpu of nvidiaGPUs) {
+        const key = normalizeBusAddress(gpu.bus);
+        if (key)
+            byBus.set(key, gpu);
+    }
+    // Priority 2 — systeminformation: catches GPUs nvidia-smi doesn't see and
+    // enriches existing entries with temperature
+    for (const gpu of siGPUs) {
+        const key = normalizeBusAddress(gpu.bus);
+        if (!key)
+            continue;
+        const existing = byBus.get(key);
+        if (existing) {
+            if (existing.temperatureCelsius === null && gpu.temperatureCelsius !== null) {
+                existing.temperatureCelsius = gpu.temperatureCelsius;
+            }
+            continue;
+        }
+        byBus.set(key, gpu);
+    }
+    // Priority 3 — lspci: enumerates AMD/other GPUs and supplies VRAM via sysfs/rocm-smi
     for (const lspciGPU of lspciGPUs) {
         const key = normalizeBusAddress(lspciGPU.bus);
-        if (!key || seen.has(key))
+        if (!key)
             continue;
-        seen.add(key);
-        const existing = siByBus.get(key);
+        const existing = byBus.get(key);
         if (existing) {
             if (existing.memoryTotalBytes === null) {
-                const sysfs = sysfsVRAMMap.get(key);
-                const sysfsTotal = sysfs?.memoryTotalBytes ?? null;
-                const sysfsUsed = sysfs?.memoryUsedBytes ?? null;
-                const validSysfsTotal = sysfsTotal !== null && Number.isFinite(sysfsTotal);
-                const validSysfsUsed = sysfsUsed !== null && Number.isFinite(sysfsUsed);
-                if (validSysfsTotal) {
-                    existing.memoryTotalBytes = sysfsTotal;
-                    existing.memoryUsedBytes = validSysfsUsed ? sysfsUsed : null;
-                    existing.memoryFreeBytes =
-                        validSysfsUsed && sysfsUsed !== null ? sysfsTotal - sysfsUsed : null;
-                }
+                applySysfsOrRocmVRAM({
+                    gpu: existing,
+                    key,
+                    rocmByBus,
+                    sysfsVRAMMap
+                });
             }
             continue;
         }
-        const sysfs = sysfsVRAMMap.get(key);
-        let totalBytes = sysfs?.memoryTotalBytes ?? null;
-        let usedBytes = sysfs?.memoryUsedBytes ?? null;
-        if (totalBytes === null) {
-            const rocm = rocmByBus.get(key);
-            if (rocm) {
-                totalBytes = rocm.memoryTotalBytes;
-                usedBytes = rocm.memoryUsedBytes;
-            }
-        }
-        merged.push({
-            bus: lspciGPU.bus,
+        const gpu = {
+            bus: normalizeBusAddress(lspciGPU.bus),
             driverVersion: null,
-            memoryFreeBytes: totalBytes !== null && usedBytes !== null && Number.isFinite(usedBytes)
-                ? totalBytes - usedBytes
-                : null,
-            memoryTotalBytes: totalBytes,
-            memoryUsedBytes: usedBytes !== null && Number.isFinite(usedBytes) ? usedBytes : null,
+            memoryFreeBytes: null,
+            memoryTotalBytes: null,
+            memoryUsedBytes: null,
             model: lspciGPU.model,
             temperatureCelsius: null,
             vendor: lspciGPU.vendor
+        };
+        applySysfsOrRocmVRAM({
+            gpu,
+            key,
+            rocmByBus,
+            sysfsVRAMMap
         });
+        byBus.set(key, gpu);
     }
-    return merged;
+    return [...byBus.values()];
 }
-async function collectMachineMetadata() {
-    const [cpuResult, memResult, osResult, graphicsResult, lspciGPUs, rocmVRAM] = await Promise.allSettled([
+function applySysfsOrRocmVRAM({ gpu, key, rocmByBus, sysfsVRAMMap }) {
+    const sysfs = sysfsVRAMMap.get(key);
+    let totalBytes = sysfs?.memoryTotalBytes ?? null;
+    let usedBytes = sysfs?.memoryUsedBytes ?? null;
+    if (totalBytes === null) {
+        const rocm = rocmByBus.get(key);
+        if (rocm) {
+            totalBytes = rocm.memoryTotalBytes;
+            usedBytes = rocm.memoryUsedBytes;
+        }
+    }
+    if (totalBytes === null || !Number.isFinite(totalBytes))
+        return;
+    gpu.memoryTotalBytes = totalBytes;
+    gpu.memoryUsedBytes = usedBytes !== null && Number.isFinite(usedBytes) ? usedBytes : null;
+    gpu.memoryFreeBytes =
+        usedBytes !== null && Number.isFinite(usedBytes)
+            ? Math.max(totalBytes - usedBytes, 0)
+            : null;
+}
+async function collectMachineMetadata({ logger }) {
+    const [cpuResult, memResult, osResult, graphicsResult, lspciGPUs, nvidiaGPUs, rocmVRAM] = await Promise.allSettled([
         si.cpu(),
         si.mem(),
         si.osInfo(),
         si.graphics(),
         detectGPUsViaLspci(),
+        detectGPUsViaNvidiaSmi({ logger }),
         detectVRAMViaRocmSmi()
     ]);
     const cpuInfo = cpuResult.status === "fulfilled" ? cpuResult.value : null;
@@ -135517,13 +135630,14 @@ async function collectMachineMetadata() {
         ? graphicsResult.value
         : { controllers: [] };
     const resolvedLspciGPUs = lspciGPUs.status === "fulfilled" ? lspciGPUs.value : [];
+    const resolvedNvidiaGPUs = nvidiaGPUs.status === "fulfilled" ? nvidiaGPUs.value : [];
     const resolvedRocmVRAM = rocmVRAM.status === "fulfilled" ? rocmVRAM.value : [];
     const siGPUs = (graphicsInfo.controllers ?? []).map((controller) => {
         const totalBytes = normalizeMegabytes(controller.memoryTotal ?? null);
         const freeBytes = normalizeMegabytes(controller.memoryFree ?? null);
         const usedBytes = totalBytes !== null && freeBytes !== null ? totalBytes - freeBytes : null;
         return {
-            bus: controller.bus ?? null,
+            bus: normalizeBusAddress(controller.pciBus ?? controller.busAddress ?? controller.bus ?? null),
             driverVersion: controller.driverVersion ?? null,
             memoryFreeBytes: freeBytes,
             memoryTotalBytes: totalBytes,
@@ -135537,6 +135651,10 @@ async function collectMachineMetadata() {
     for (const lspciGPU of resolvedLspciGPUs) {
         busSources.add(lspciGPU.bus);
     }
+    for (const nvidiaGPU of resolvedNvidiaGPUs) {
+        if (nvidiaGPU.bus)
+            busSources.add(nvidiaGPU.bus);
+    }
     for (const siGPU of siGPUs) {
         if (siGPU.bus)
             busSources.add(siGPU.bus);
@@ -135555,6 +135673,7 @@ async function collectMachineMetadata() {
     }
     const gpus = buildMergedGPUs({
         lspciGPUs: resolvedLspciGPUs,
+        nvidiaGPUs: resolvedNvidiaGPUs,
         rocmVRAM: resolvedRocmVRAM,
         siGPUs,
         sysfsVRAMMap
@@ -135612,7 +135731,7 @@ async function createApplication({ abortController, apiClient, configuration, lo
     });
     let machine = null;
     try {
-        machine = await collectMachineMetadata();
+        machine = await collectMachineMetadata({ logger });
         logger.info("Detected machine hardware", {
             gpuCount: machine.gpus.length
         });
@@ -135645,7 +135764,7 @@ async function createApplication({ abortController, apiClient, configuration, lo
     let modelManager = createModelManagerFromConfig(conduitConfiguration, configuration, logger);
     const conduitStateReportManager = new ConduitStateReportManager({
         apiClient,
-        collectMachineMetadata: collectMachineMetadata,
+        collectMachineMetadata: () => collectMachineMetadata({ logger }),
         conduitStateManager,
         downloadProgressReportIntervalMs: 5000,
         logger,

package/dist/cli.sea.cjs

@@ -20056,16 +20056,16 @@ const InferenceAgentMachineMetadataSchema = object$1({
         model: string$1().nullable(),
         physicalCores: number$1().int().positive().nullable()
     }),
-    exllamav3Version: string$1().nullable(),
+    exllamav3Version: string$1().nullable().default(null),
     gpus: array(InferenceAgentMachineGPUSchema),
     hostname: string$1(),
-    llamaCppVersion: string$1().nullable(),
+    llamaCppVersion: string$1().nullable().default(null),
     machineID: string$1(),
     memory: object$1({
         availableBytes: number$1().int().nonnegative().nullable(),
         totalBytes: number$1().int().nonnegative().nullable()
     }),
-    mlxlmVersion: string$1().nullable(),
+    mlxlmVersion: string$1().nullable().default(null),
     os: object$1({
         arch: string$1(),
         platform: string$1(),
@@ -20073,9 +20073,9 @@ const InferenceAgentMachineMetadataSchema = object$1({
         type: string$1().nullable(),
         version: string$1().nullable()
     }),
-    sglangVersion: string$1().nullable(),
-    tensorrtLlmVersion: string$1().nullable(),
-    vllmVersion: string$1().nullable()
+    sglangVersion: string$1().nullable().default(null),
+    tensorrtLlmVersion: string$1().nullable().default(null),
+    vllmVersion: string$1().nullable().default(null)
 });
 const InferenceAgentMachineReportPayloadSchema = object$1({
     machine: InferenceAgentMachineMetadataSchema
@@ -155519,7 +155519,12 @@ function normalizeMegabytes(value) {
 function normalizeBusAddress(bus) {
     if (!bus)
         return null;
-    return bus.toLowerCase().replace(/^0000:/, "");
+    // PCI addresses arrive with varying domain widths across tools:
+    // nvidia-smi "00000000:09:00.0", lspci/sysfs "0000:09:00.0",
+    // systeminformation "09:00.0". Reduce all to the trailing "bus:device.function"
+    // so the same GPU matches across sources.
+    const match = bus.toLowerCase().match(/([0-9a-f]{1,2}:[0-9a-f]{2}\.[0-9a-f])$/);
+    return match ? match[1] : bus.toLowerCase();
 }
 function resolveCpuValue(value) {
     if (typeof value === "number" && Number.isFinite(value)) {
@@ -155667,76 +155672,184 @@ async function detectVRAMViaRocmSmi() {
         return [];
     }
 }
-function buildMergedGPUs(options) {
-    const { lspciGPUs, rocmVRAM, siGPUs, sysfsVRAMMap } = options;
-    const siByBus = new Map();
-    for (const gpu of siGPUs) {
-        const key = normalizeBusAddress(gpu.bus);
-        if (key)
-            siByBus.set(key, gpu);
+const NVIDIA_SMI_TIMEOUT_MS$1 = 10_000;
+function parseNvidiaSmiNumber(value) {
+    if (!value)
+        return null;
+    // Unified-memory devices (e.g. NVIDIA GB10) report "[N/A]"; treat as unknown
+    const parsed = parseInt(value, 10);
+    return Number.isFinite(parsed) && parsed >= 0 ? parsed : null;
+}
+function parseNvidiaSmiMemoryBytes(value) {
+    const mib = parseNvidiaSmiNumber(value);
+    return mib !== null ? normalizeMegabytes(mib) : null;
+}
+async function readSystemMemoryBytes({ logger }) {
+    try {
+        const contents = await require$$0$d.readFile("/proc/meminfo", "utf8");
+        let availableBytes = null;
+        let totalBytes = null;
+        for (const line of contents.split("\n")) {
+            const match = line.match(/^(MemTotal|MemAvailable):\s+(\d+)\s*kB/i);
+            if (!match)
+                continue;
+            const bytes = parseInt(match[2], 10) * 1024;
+            if (match[1].toLowerCase() === "memtotal") {
+                totalBytes = bytes;
+            }
+            else {
+                availableBytes = bytes;
+            }
+        }
+        return { availableBytes, totalBytes };
+    }
+    catch (error) {
+        logger.debug("Failed to read system memory", { error: asError(error) });
+        return { availableBytes: null, totalBytes: null };
+    }
+}
+async function detectGPUsViaNvidiaSmi({ logger }) {
+    try {
+        const { stdout } = await execa("nvidia-smi", [
+            "--query-gpu=name,pci.bus_id,driver_version,memory.total,memory.used,memory.free,temperature.gpu",
+            "--format=csv,noheader,nounits"
+        ], { timeout: NVIDIA_SMI_TIMEOUT_MS$1 });
+        const lines = stdout.split("\n").filter(line => line.trim().length > 0);
+        const gpus = [];
+        for (const line of lines) {
+            const [name, bus, driverVersion, memTotal, memUsed, memFree, temperature] = line
+                .split(",")
+                .map(part => part.trim());
+            gpus.push({
+                bus: normalizeBusAddress(bus),
+                driverVersion: driverVersion || null,
+                memoryFreeBytes: parseNvidiaSmiMemoryBytes(memFree),
+                memoryTotalBytes: parseNvidiaSmiMemoryBytes(memTotal),
+                memoryUsedBytes: parseNvidiaSmiMemoryBytes(memUsed),
+                model: name || null,
+                temperatureCelsius: parseNvidiaSmiNumber(temperature),
+                vendor: "NVIDIA"
+            });
+        }
+        // Unified-memory devices (e.g. NVIDIA GB10 on DGX Spark) report "[N/A]" for
+        // memory. Their compute pool is system RAM, so fall back to /proc/meminfo.
+        if (gpus.some(gpu => gpu.memoryTotalBytes === null)) {
+            const systemMemory = await readSystemMemoryBytes({ logger });
+            const total = systemMemory.totalBytes;
+            if (total !== null) {
+                const free = systemMemory.availableBytes;
+                const used = free !== null ? total - free : null;
+                for (const gpu of gpus) {
+                    if (gpu.memoryTotalBytes !== null)
+                        continue;
+                    gpu.memoryFreeBytes = free;
+                    gpu.memoryTotalBytes = total;
+                    gpu.memoryUsedBytes = used;
+                }
+            }
+        }
+        return gpus;
     }
+    catch (error) {
+        logger.debug("nvidia-smi GPU query failed", { error: asError(error) });
+        return [];
+    }
+}
+function buildMergedGPUs(options) {
+    const { lspciGPUs, nvidiaGPUs, rocmVRAM, siGPUs, sysfsVRAMMap } = options;
     const rocmByBus = new Map();
     for (const entry of rocmVRAM) {
         const key = normalizeBusAddress(entry.bus);
         if (key)
             rocmByBus.set(key, entry);
     }
-    const seen = new Set();
-    const merged = [...siGPUs];
+    const byBus = new Map();
+    // Priority 1 — nvidia-smi: authoritative for NVIDIA (name, driver, VRAM, temp, bus)
+    for (const gpu of nvidiaGPUs) {
+        const key = normalizeBusAddress(gpu.bus);
+        if (key)
+            byBus.set(key, gpu);
+    }
+    // Priority 2 — systeminformation: catches GPUs nvidia-smi doesn't see and
+    // enriches existing entries with temperature
+    for (const gpu of siGPUs) {
+        const key = normalizeBusAddress(gpu.bus);
+        if (!key)
+            continue;
+        const existing = byBus.get(key);
+        if (existing) {
+            if (existing.temperatureCelsius === null && gpu.temperatureCelsius !== null) {
+                existing.temperatureCelsius = gpu.temperatureCelsius;
+            }
+            continue;
+        }
+        byBus.set(key, gpu);
+    }
+    // Priority 3 — lspci: enumerates AMD/other GPUs and supplies VRAM via sysfs/rocm-smi
     for (const lspciGPU of lspciGPUs) {
         const key = normalizeBusAddress(lspciGPU.bus);
-        if (!key || seen.has(key))
+        if (!key)
             continue;
-        seen.add(key);
-        const existing = siByBus.get(key);
+        const existing = byBus.get(key);
         if (existing) {
             if (existing.memoryTotalBytes === null) {
-                const sysfs = sysfsVRAMMap.get(key);
-                const sysfsTotal = sysfs?.memoryTotalBytes ?? null;
-                const sysfsUsed = sysfs?.memoryUsedBytes ?? null;
-                const validSysfsTotal = sysfsTotal !== null && Number.isFinite(sysfsTotal);
-                const validSysfsUsed = sysfsUsed !== null && Number.isFinite(sysfsUsed);
-                if (validSysfsTotal) {
-                    existing.memoryTotalBytes = sysfsTotal;
-                    existing.memoryUsedBytes = validSysfsUsed ? sysfsUsed : null;
-                    existing.memoryFreeBytes =
-                        validSysfsUsed && sysfsUsed !== null ? sysfsTotal - sysfsUsed : null;
-                }
+                applySysfsOrRocmVRAM({
+                    gpu: existing,
+                    key,
+                    rocmByBus,
+                    sysfsVRAMMap
+                });
             }
             continue;
         }
-        const sysfs = sysfsVRAMMap.get(key);
-        let totalBytes = sysfs?.memoryTotalBytes ?? null;
-        let usedBytes = sysfs?.memoryUsedBytes ?? null;
-        if (totalBytes === null) {
-            const rocm = rocmByBus.get(key);
-            if (rocm) {
-                totalBytes = rocm.memoryTotalBytes;
-                usedBytes = rocm.memoryUsedBytes;
-            }
-        }
-        merged.push({
-            bus: lspciGPU.bus,
+        const gpu = {
+            bus: normalizeBusAddress(lspciGPU.bus),
             driverVersion: null,
-            memoryFreeBytes: totalBytes !== null && usedBytes !== null && Number.isFinite(usedBytes)
-                ? totalBytes - usedBytes
-                : null,
-            memoryTotalBytes: totalBytes,
-            memoryUsedBytes: usedBytes !== null && Number.isFinite(usedBytes) ? usedBytes : null,
+            memoryFreeBytes: null,
+            memoryTotalBytes: null,
+            memoryUsedBytes: null,
             model: lspciGPU.model,
             temperatureCelsius: null,
             vendor: lspciGPU.vendor
+        };
+        applySysfsOrRocmVRAM({
+            gpu,
+            key,
+            rocmByBus,
+            sysfsVRAMMap
         });
+        byBus.set(key, gpu);
     }
-    return merged;
+    return [...byBus.values()];
 }
-async function collectMachineMetadata() {
-    const [cpuResult, memResult, osResult, graphicsResult, lspciGPUs, rocmVRAM] = await Promise.allSettled([
+function applySysfsOrRocmVRAM({ gpu, key, rocmByBus, sysfsVRAMMap }) {
+    const sysfs = sysfsVRAMMap.get(key);
+    let totalBytes = sysfs?.memoryTotalBytes ?? null;
+    let usedBytes = sysfs?.memoryUsedBytes ?? null;
+    if (totalBytes === null) {
+        const rocm = rocmByBus.get(key);
+        if (rocm) {
+            totalBytes = rocm.memoryTotalBytes;
+            usedBytes = rocm.memoryUsedBytes;
+        }
+    }
+    if (totalBytes === null || !Number.isFinite(totalBytes))
+        return;
+    gpu.memoryTotalBytes = totalBytes;
+    gpu.memoryUsedBytes = usedBytes !== null && Number.isFinite(usedBytes) ? usedBytes : null;
+    gpu.memoryFreeBytes =
+        usedBytes !== null && Number.isFinite(usedBytes)
+            ? Math.max(totalBytes - usedBytes, 0)
+            : null;
+}
+async function collectMachineMetadata({ logger }) {
+    const [cpuResult, memResult, osResult, graphicsResult, lspciGPUs, nvidiaGPUs, rocmVRAM] = await Promise.allSettled([
         si.cpu(),
         si.mem(),
         si.osInfo(),
         si.graphics(),
         detectGPUsViaLspci(),
+        detectGPUsViaNvidiaSmi({ logger }),
         detectVRAMViaRocmSmi()
     ]);
     const cpuInfo = cpuResult.status === "fulfilled" ? cpuResult.value : null;
@@ -155746,13 +155859,14 @@ async function collectMachineMetadata() {
         ? graphicsResult.value
         : { controllers: [] };
     const resolvedLspciGPUs = lspciGPUs.status === "fulfilled" ? lspciGPUs.value : [];
+    const resolvedNvidiaGPUs = nvidiaGPUs.status === "fulfilled" ? nvidiaGPUs.value : [];
     const resolvedRocmVRAM = rocmVRAM.status === "fulfilled" ? rocmVRAM.value : [];
     const siGPUs = (graphicsInfo.controllers ?? []).map((controller) => {
         const totalBytes = normalizeMegabytes(controller.memoryTotal ?? null);
         const freeBytes = normalizeMegabytes(controller.memoryFree ?? null);
         const usedBytes = totalBytes !== null && freeBytes !== null ? totalBytes - freeBytes : null;
         return {
-            bus: controller.bus ?? null,
+            bus: normalizeBusAddress(controller.pciBus ?? controller.busAddress ?? controller.bus ?? null),
             driverVersion: controller.driverVersion ?? null,
             memoryFreeBytes: freeBytes,
             memoryTotalBytes: totalBytes,
@@ -155766,6 +155880,10 @@ async function collectMachineMetadata() {
     for (const lspciGPU of resolvedLspciGPUs) {
         busSources.add(lspciGPU.bus);
     }
+    for (const nvidiaGPU of resolvedNvidiaGPUs) {
+        if (nvidiaGPU.bus)
+            busSources.add(nvidiaGPU.bus);
+    }
     for (const siGPU of siGPUs) {
         if (siGPU.bus)
             busSources.add(siGPU.bus);
@@ -155784,6 +155902,7 @@ async function collectMachineMetadata() {
     }
     const gpus = buildMergedGPUs({
         lspciGPUs: resolvedLspciGPUs,
+        nvidiaGPUs: resolvedNvidiaGPUs,
         rocmVRAM: resolvedRocmVRAM,
         siGPUs,
         sysfsVRAMMap
@@ -155841,7 +155960,7 @@ async function createApplication({ abortController, apiClient, configuration, lo
     });
     let machine = null;
     try {
-        machine = await collectMachineMetadata();
+        machine = await collectMachineMetadata({ logger });
         logger.info("Detected machine hardware", {
             gpuCount: machine.gpus.length
         });
@@ -155874,7 +155993,7 @@ async function createApplication({ abortController, apiClient, configuration, lo
     let modelManager = createModelManagerFromConfig(conduitConfiguration, configuration, logger);
     const conduitStateReportManager = new ConduitStateReportManager({
         apiClient,
-        collectMachineMetadata: collectMachineMetadata,
+        collectMachineMetadata: () => collectMachineMetadata({ logger }),
         conduitStateManager,
         downloadProgressReportIntervalMs: 5000,
         logger,

package/dist/utils/machineInfo.d.ts

@@ -1,2 +1,8 @@
 import type { InferenceAgentMachineMetadata } from "@infersec/definitions";
-export declare function collectMachineMetadata(): Promise<InferenceAgentMachineMetadata>;
+import type { Logger } from "@infersec/logger";
+export declare function normalizeBusAddress(bus: string | null | undefined): string | null;
+export declare function parseNvidiaSmiNumber(value: string | undefined): number | null;
+export declare function parseNvidiaSmiMemoryBytes(value: string | undefined): number | null;
+export declare function collectMachineMetadata({ logger }: {
+    logger: Logger;
+}): Promise<InferenceAgentMachineMetadata>;

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@infersec/conduit",
   "description": "End user conduit agent for connecting local LLMs to the cloud.",
-  "version": "1.90.2",
+  "version": "1.90.3",
   "bin": {
     "infersec-conduit": "./dist/cli.js"
   },