@infersec/conduit 1.111.0 → 1.112.1

package/dist/cli.js

@@ -21567,6 +21567,90 @@ object$5({
 // Source of truth for curated recommended models. Consumed by @infersec/definitions
 // and the website model page generator (applications/website/scripts/generate-model-pages.ts).
 const modelsData = [
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-2.6B",
+        blurb: "Liquid's LFM2.5 2.6B hybrid at high fidelity - agentic post-training, 128K context, and 220 tok/s decode on an M5 Max. Fits a 4 GB card with room to spare.",
+        contextLength: 131072,
+        fileSizeBytes: 2874779648,
+        format: "gguf",
+        license: "lfm1.0",
+        parameterCount: "2.6B",
+        quantization: "Q8_0",
+        slug: "LiquidAI/LFM2.5-2.6B-GGUF",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF",
+        vramTierGB: 4
+    },
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-2.6B",
+        blurb: "Liquid's LFM2.5 2.6B hybrid at 4-bit - competitive with models 4x its size on tool use and multi-step agentic tasks. A 1.7 GB download that sips memory.",
+        contextLength: 131072,
+        fileSizeBytes: 1674455040,
+        format: "gguf",
+        license: "lfm1.0",
+        parameterCount: "2.6B",
+        quantization: "Q4_K_M",
+        slug: "LiquidAI/LFM2.5-2.6B-GGUF",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF",
+        vramTierGB: 4
+    },
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-2.6B",
+        blurb: "The native BF16 checkpoint of Liquid's LFM2.5 2.6B for vLLM and SGLang - hybrid conv + GQA architecture, 128K context window, and agentic reinforcement-learning post-training.",
+        contextLength: 131072,
+        fileSizeBytes: 5394427456,
+        format: "safetensors",
+        license: "lfm1.0",
+        parameterCount: "2.6B",
+        quantization: "BF16",
+        slug: "LiquidAI/LFM2.5-2.6B",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-2.6B",
+        vramTierGB: 8
+    },
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-230M",
+        blurb: "Liquid's most compact LFM2.5 at high fidelity - a 230M edge model distilled from the 350M, tuned for tool use and data extraction. 213 tok/s decode on a Galaxy S25 Ultra.",
+        contextLength: 32768,
+        fileSizeBytes: 246598496,
+        format: "gguf",
+        license: "lfm1.0",
+        parameterCount: "230M",
+        quantization: "Q8_0",
+        slug: "LiquidAI/LFM2.5-230M-GGUF",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-230M-GGUF",
+        vramTierGB: 2
+    },
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-230M",
+        blurb: "Liquid's most compact LFM2.5 at 4-bit - a 153 MB download that brings real tool-use capability to the tightest memory budgets, from Raspberry Pi to phones.",
+        contextLength: 32768,
+        fileSizeBytes: 153406304,
+        format: "gguf",
+        license: "lfm1.0",
+        parameterCount: "230M",
+        quantization: "Q4_K_M",
+        slug: "LiquidAI/LFM2.5-230M-GGUF",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-230M-GGUF",
+        vramTierGB: 2
+    },
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-230M",
+        blurb: "The native BF16 checkpoint of Liquid's LFM2.5 230M for vLLM - a 230M edge model with a 32K context window, suited to data extraction and lightweight on-device agent pipelines.",
+        contextLength: 32768,
+        fileSizeBytes: 459401112,
+        format: "safetensors",
+        license: "lfm1.0",
+        parameterCount: "230M",
+        quantization: "BF16",
+        slug: "LiquidAI/LFM2.5-230M",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-230M",
+        vramTierGB: 4
+    },
     {
         author: "Qwen",
         baseModel: "Qwen/Qwen3-8B",
@@ -21623,6 +21707,48 @@ const modelsData = [
         sourceURL: "https://huggingface.co/bartowski/deepreinforce-ai_Ornith-1.0-9B-GGUF",
         vramTierGB: 16
     },
+    {
+        author: "unsloth",
+        baseModel: "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
+        blurb: "NVIDIA's Nemotron 3.5 Lightning MoE squeezed into 2-bit - 30B total with only 3B active in a Mamba-2 + MoE + attention hybrid. Configurable thinking and a 256K context on a single 24 GB card.",
+        contextLength: 262144,
+        fileSizeBytes: 19432070208,
+        format: "gguf",
+        license: "openmdw-1.1",
+        parameterCount: "30B",
+        quantization: "UD-IQ2_M",
+        slug: "unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        sourceURL: "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        vramTierGB: 24
+    },
+    {
+        author: "unsloth",
+        baseModel: "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
+        blurb: "NVIDIA's Nemotron 3.5 Lightning at Unsloth's dynamic 4-bit - a Mamba-2 hybrid MoE with 3B active parameters, toggleable reasoning, and strong instruction following (IFBench 71.9). Needs a 32 GB GPU.",
+        contextLength: 262144,
+        fileSizeBytes: 25266255936,
+        format: "gguf",
+        license: "openmdw-1.1",
+        parameterCount: "30B",
+        quantization: "UD-Q4_K_M",
+        slug: "unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        sourceURL: "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        vramTierGB: 32
+    },
+    {
+        author: "unsloth",
+        baseModel: "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
+        blurb: "Nemotron 3.5 Lightning at 8-bit fidelity - near-reference quality from the 30B-A3B Mamba-2 hybrid, with configurable thinking and DSpark speculative decoding support. Needs a 48 GB card.",
+        contextLength: 262144,
+        fileSizeBytes: 35004643392,
+        format: "gguf",
+        license: "openmdw-1.1",
+        parameterCount: "30B",
+        quantization: "Q8_0",
+        slug: "unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        sourceURL: "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        vramTierGB: 48
+    },
     {
         author: "unsloth",
         baseModel: "openai/gpt-oss-20b",
@@ -21681,6 +21807,76 @@ const modelsData = [
         sourceURL: "https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF",
         vramTierGB: 128
     },
+    {
+        author: "Qwen",
+        baseModel: "Qwen/Qwen3.8-Flash-Next",
+        blurb: "The official BF16 release of Qwen's experimental Qwen4-preview architecture - 180B total with only 6B activated, plus 51B n-gram embeddings, a vision encoder, and a 262K native context. For vLLM/SGLang on multi-GPU server hardware.",
+        contextLength: 262144,
+        fileSizeBytes: 360000192888,
+        format: "safetensors",
+        license: "qwen-community-1.0",
+        parameterCount: "180B",
+        quantization: "BF16",
+        slug: "Qwen/Qwen3.8-Flash-Next",
+        sourceURL: "https://huggingface.co/Qwen/Qwen3.8-Flash-Next",
+        vramTierGB: 640
+    },
+    {
+        author: "RadixArk",
+        baseModel: "Qwen/Qwen3.8-Flash-Next",
+        blurb: "Qwen3.8-Flash-Next with its routed experts quantized to NVIDIA's NVFP4 W4A4 via Model Optimizer - 135 GB instead of 360 GB BF16, keeping GSM8K/AIME in-band with the reference. Serve with SGLang on Blackwell hardware.",
+        contextLength: 262144,
+        fileSizeBytes: 135195303851,
+        format: "safetensors",
+        license: "qwen-community-1.0",
+        parameterCount: "180B",
+        quantization: "NVFP4",
+        slug: "RadixArk/Qwen3.8-Flash-Next-NVFP4",
+        sourceURL: "https://huggingface.co/RadixArk/Qwen3.8-Flash-Next-NVFP4",
+        vramTierGB: 256
+    },
+    {
+        author: "deepseek-ai",
+        baseModel: "deepseek-ai/DeepSeek-V4-Flash",
+        blurb: "DeepSeek's V4-Flash - a 284B MoE with only 13B activated and a native 1M context window, shipped in mixed FP4/FP8 precision. Hybrid CSA/HCA attention keeps long-context inference cheap, with three reasoning-effort modes including Think Max.",
+        contextLength: 1048576,
+        fileSizeBytes: 159617149040,
+        format: "safetensors",
+        license: "mit",
+        parameterCount: "284B",
+        quantization: "FP4-FP8",
+        slug: "deepseek-ai/DeepSeek-V4-Flash",
+        sourceURL: "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash",
+        vramTierGB: 256
+    },
+    {
+        author: "unsloth",
+        baseModel: "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp",
+        blurb: "Unsloth's 2-bit cut of DeepSeek's experimental vision variant of V4-Flash - the same 284B MoE with 13B activated, plus image understanding via a separate mmproj projector. Fits 128 GB of unified memory or 2x 96 GB.",
+        contextLength: 1048576,
+        fileSizeBytes: 90692965818,
+        format: "gguf",
+        license: "mit",
+        parameterCount: "284B",
+        quantization: "UD-IQ2_XXS",
+        slug: "unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF",
+        sourceURL: "https://huggingface.co/unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF",
+        vramTierGB: 128
+    },
+    {
+        author: "unsloth",
+        baseModel: "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp",
+        blurb: "The 4-bit cut of DeepSeek's experimental vision variant of V4-Flash - 284B MoE with 13B activated at higher fidelity, with multimodal agent scores well clear of the text-only release. Image input currently needs a preview llama.cpp build; needs 256 GB.",
+        contextLength: 1048576,
+        fileSizeBytes: 136662447706,
+        format: "gguf",
+        license: "mit",
+        parameterCount: "284B",
+        quantization: "UD-IQ4_XS",
+        slug: "unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF",
+        sourceURL: "https://huggingface.co/unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF",
+        vramTierGB: 256
+    },
     {
         author: "unsloth",
         baseModel: "Qwen/Qwen3.6-27B",
@@ -137740,12 +137936,20 @@ async function detectVRAMViaSysfs(pciBusSuffix) {
             const deviceDir = path$1.join(DRM_PATH, entry, "device");
             const totalStr = await readSysfsFile(path$1.join(deviceDir, "mem_info_vram_total"));
             const usedStr = await readSysfsFile(path$1.join(deviceDir, "mem_info_vram_used"));
+            const gttTotalStr = await readSysfsFile(path$1.join(deviceDir, "mem_info_gtt_total"));
+            const gttUsedStr = await readSysfsFile(path$1.join(deviceDir, "mem_info_gtt_used"));
             const totalBytes = totalStr !== null ? parseInt(totalStr, 10) : null;
             const usedBytes = usedStr !== null ? parseInt(usedStr, 10) : null;
+            const gttTotalBytes = gttTotalStr !== null ? parseInt(gttTotalStr, 10) : null;
+            const gttUsedBytes = gttUsedStr !== null ? parseInt(gttUsedStr, 10) : null;
             const validTotal = totalBytes !== null && Number.isFinite(totalBytes) && totalBytes >= 0;
             const validUsed = usedBytes !== null && Number.isFinite(usedBytes) && usedBytes >= 0;
+            const validGttTotal = gttTotalBytes !== null && Number.isFinite(gttTotalBytes) && gttTotalBytes >= 0;
+            const validGttUsed = gttUsedBytes !== null && Number.isFinite(gttUsedBytes) && gttUsedBytes >= 0;
             if (validTotal) {
                 return {
+                    gttTotalBytes: validGttTotal ? gttTotalBytes : null,
+                    gttUsedBytes: validGttUsed ? gttUsedBytes : null,
                     memoryTotalBytes: totalBytes,
                     memoryUsedBytes: validUsed ? usedBytes : null
                 };
@@ -137755,36 +137959,74 @@ async function detectVRAMViaSysfs(pciBusSuffix) {
     catch {
         // sysfs not available
     }
-    return { memoryTotalBytes: null, memoryUsedBytes: null };
+    return {
+        gttTotalBytes: null,
+        gttUsedBytes: null,
+        memoryTotalBytes: null,
+        memoryUsedBytes: null
+    };
 }
-async function detectVRAMViaRocmSmi() {
+// rocm-smi accepts a single --showmeminfo type per invocation, so VRAM and GTT
+// arrive from separate calls; both share this per-card parser.
+function parseRocmSmiMemory({ keyPrefix, stdout }) {
+    const parsed = JSON.parse(stdout);
+    const results = [];
+    const cards = Object.entries(parsed).filter(([key]) => key.startsWith("card"));
+    for (const [, data] of cards) {
+        const bus = data["PCI Bus"] ?? null;
+        const totalStr = data[`${keyPrefix} Total Memory (B)`] ?? null;
+        const usedStr = data[`${keyPrefix} Total Used Memory (B)`] ?? null;
+        const totalBytes = totalStr !== null ? parseInt(totalStr, 10) : null;
+        const usedBytes = usedStr !== null ? parseInt(usedStr, 10) : null;
+        const validTotal = totalBytes !== null && Number.isFinite(totalBytes) && totalBytes >= 0;
+        const validUsed = usedBytes !== null && Number.isFinite(usedBytes) && usedBytes >= 0;
+        if (bus) {
+            results.push({
+                bus,
+                memoryTotalBytes: validTotal ? totalBytes : null,
+                memoryUsedBytes: validUsed ? usedBytes : null
+            });
+        }
+    }
+    return results;
+}
+const ROCM_SMI_TIMEOUT_MS = 10_000;
+async function detectVRAMViaRocmSmi({ logger }) {
     try {
-        const { stdout } = await execa("rocm-smi", [
-            "--showbus",
-            "--showmeminfo",
-            "vram",
-            "--json"
+        const [vramResult, gttResult] = await Promise.allSettled([
+            execa("rocm-smi", ["--showbus", "--showmeminfo", "vram", "--json"], {
+                timeout: ROCM_SMI_TIMEOUT_MS
+            }),
+            execa("rocm-smi", ["--showbus", "--showmeminfo", "gtt", "--json"], {
+                timeout: ROCM_SMI_TIMEOUT_MS
+            })
         ]);
-        const parsed = JSON.parse(stdout);
-        const results = [];
-        const cards = Object.entries(parsed).filter(([key]) => key.startsWith("card"));
-        for (const [, data] of cards) {
-            const bus = data["PCI Bus"] ?? null;
-            const totalStr = data["VRAM Total Memory (B)"] ?? null;
-            const usedStr = data["VRAM Total Used Memory (B)"] ?? null;
-            const totalBytes = totalStr !== null ? parseInt(totalStr, 10) : null;
-            const usedBytes = usedStr !== null ? parseInt(usedStr, 10) : null;
-            const validTotal = totalBytes !== null && Number.isFinite(totalBytes) && totalBytes >= 0;
-            const validUsed = usedBytes !== null && Number.isFinite(usedBytes) && usedBytes >= 0;
-            if (bus) {
-                results.push({
-                    bus,
-                    memoryTotalBytes: validTotal ? totalBytes : null,
-                    memoryUsedBytes: validUsed ? usedBytes : null
+        if (vramResult.status !== "fulfilled")
+            return [];
+        let gttEntries = [];
+        if (gttResult.status === "fulfilled") {
+            try {
+                gttEntries = parseRocmSmiMemory({
+                    keyPrefix: "GTT",
+                    stdout: gttResult.value.stdout
                 });
             }
+            catch (error) {
+                // Unusable GTT output (e.g. older rocm-smi) degrades to VRAM-only
+                logger.warn("rocm-smi GTT output parse failed", { error: asError(error) });
+            }
         }
-        return results;
+        const gttByBus = new Map(gttEntries.map(entry => [entry.bus, entry]));
+        return parseRocmSmiMemory({ keyPrefix: "VRAM", stdout: vramResult.value.stdout }).map(vram => {
+            const gtt = gttByBus.get(vram.bus);
+            return {
+                bus: vram.bus,
+                gttTotalBytes: gtt?.memoryTotalBytes ?? null,
+                gttUsedBytes: gtt?.memoryUsedBytes ?? null,
+                memoryTotalBytes: vram.memoryTotalBytes,
+                memoryUsedBytes: vram.memoryUsedBytes
+            };
+        });
     }
     catch {
         return [];
@@ -137849,8 +138091,9 @@ async function detectGPUsViaNvidiaSmi({ logger }) {
                 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.
+        // Shared/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. (AMD iGPUs get the same treatment via GTT merging.)
         if (gpus.some(gpu => gpu.memoryTotalBytes === null)) {
             const systemMemory = await readSystemMemoryBytes({ logger });
             const total = systemMemory.totalBytes;
@@ -137874,7 +138117,7 @@ async function detectGPUsViaNvidiaSmi({ logger }) {
     }
 }
 function buildMergedGPUs(options) {
-    const { lspciGPUs, nvidiaGPUs, rocmVRAM, siGPUs, sysfsVRAMMap } = options;
+    const { lspciGPUs, nvidiaGPUs, rocmVRAM, siGPUs, systemTotalBytes, sysfsVRAMMap } = options;
     const rocmByBus = new Map();
     for (const entry of rocmVRAM) {
         const key = normalizeBusAddress(entry.bus);
@@ -137915,6 +138158,7 @@ function buildMergedGPUs(options) {
                     gpu: existing,
                     key,
                     rocmByBus,
+                    systemTotalBytes,
                     sysfsVRAMMap
                 });
             }
@@ -137934,30 +138178,73 @@ function buildMergedGPUs(options) {
             gpu,
             key,
             rocmByBus,
+            systemTotalBytes,
             sysfsVRAMMap
         });
         byBus.set(key, gpu);
     }
     return [...byBus.values()];
 }
-function applySysfsOrRocmVRAM({ gpu, key, rocmByBus, sysfsVRAMMap }) {
+// Shared-memory GPUs (AMD iGPUs) expose a small dedicated VRAM carve-out via
+// mem_info_vram_total while the real usable pool - GTT - is carved dynamically
+// from system RAM. When GTT exceeds VRAM the device is treated as integrated
+// and the two pools combine (capped by installed RAM). Note that GTT defaults
+// to half of system RAM even on discrete GPUs, so a dGPU with less VRAM than
+// that also merges - acceptable, since GTT remains real addressable memory
+// for amdgpu compute (albeit slower over PCIe).
+function mergeGTTMemory({ gttTotalBytes, gttUsedBytes, systemTotalBytes, vramTotalBytes, vramUsedBytes }) {
+    if (vramTotalBytes === null || !Number.isFinite(vramTotalBytes)) {
+        return { memoryTotalBytes: null, memoryUsedBytes: null };
+    }
+    const isIntegrated = gttTotalBytes !== null && gttTotalBytes > vramTotalBytes;
+    if (!isIntegrated) {
+        return { memoryTotalBytes: vramTotalBytes, memoryUsedBytes: vramUsedBytes };
+    }
+    const combined = vramTotalBytes + gttTotalBytes;
+    const capped = systemTotalBytes !== null && systemTotalBytes > 0
+        ? Math.min(combined, systemTotalBytes)
+        : combined;
+    const hasCompleteUsage = vramUsedBytes !== null &&
+        Number.isFinite(vramUsedBytes) &&
+        gttUsedBytes !== null &&
+        Number.isFinite(gttUsedBytes);
+    const memoryUsedBytes = hasCompleteUsage
+        ? Math.min(vramUsedBytes + gttUsedBytes, capped)
+        : null;
+    return { memoryTotalBytes: capped, memoryUsedBytes };
+}
+function applySysfsOrRocmVRAM({ gpu, key, rocmByBus, systemTotalBytes, sysfsVRAMMap }) {
     const sysfs = sysfsVRAMMap.get(key);
-    let totalBytes = sysfs?.memoryTotalBytes ?? null;
-    let usedBytes = sysfs?.memoryUsedBytes ?? null;
-    if (totalBytes === null) {
+    let vramTotalBytes = sysfs?.memoryTotalBytes ?? null;
+    let vramUsedBytes = sysfs?.memoryUsedBytes ?? null;
+    let gttTotalBytes = sysfs?.gttTotalBytes ?? null;
+    let gttUsedBytes = sysfs?.gttUsedBytes ?? null;
+    if (vramTotalBytes === null) {
         const rocm = rocmByBus.get(key);
         if (rocm) {
-            totalBytes = rocm.memoryTotalBytes;
-            usedBytes = rocm.memoryUsedBytes;
-        }
-    }
-    if (totalBytes === null || !Number.isFinite(totalBytes))
+            vramTotalBytes = rocm.memoryTotalBytes;
+            vramUsedBytes = rocm.memoryUsedBytes;
+            gttTotalBytes = rocm.gttTotalBytes;
+            gttUsedBytes = rocm.gttUsedBytes;
+        }
+    }
+    const merged = mergeGTTMemory({
+        gttTotalBytes,
+        gttUsedBytes,
+        systemTotalBytes,
+        vramTotalBytes,
+        vramUsedBytes
+    });
+    if (merged.memoryTotalBytes === null || !Number.isFinite(merged.memoryTotalBytes))
         return;
-    gpu.memoryTotalBytes = totalBytes;
-    gpu.memoryUsedBytes = usedBytes !== null && Number.isFinite(usedBytes) ? usedBytes : null;
+    gpu.memoryTotalBytes = merged.memoryTotalBytes;
+    gpu.memoryUsedBytes =
+        merged.memoryUsedBytes !== null && Number.isFinite(merged.memoryUsedBytes)
+            ? merged.memoryUsedBytes
+            : null;
     gpu.memoryFreeBytes =
-        usedBytes !== null && Number.isFinite(usedBytes)
-            ? Math.max(totalBytes - usedBytes, 0)
+        gpu.memoryUsedBytes !== null
+            ? Math.max(gpu.memoryTotalBytes - gpu.memoryUsedBytes, 0)
             : null;
 }
 async function collectMachineMetadata({ logger }) {
@@ -137968,7 +138255,7 @@ async function collectMachineMetadata({ logger }) {
         si.graphics(),
         detectGPUsViaLspci(),
         detectGPUsViaNvidiaSmi({ logger }),
-        detectVRAMViaRocmSmi()
+        detectVRAMViaRocmSmi({ logger })
     ]);
     const cpuInfo = cpuResult.status === "fulfilled" ? cpuResult.value : null;
     const memInfo = memResult.status === "fulfilled" ? memResult.value : null;
@@ -138023,6 +138310,7 @@ async function collectMachineMetadata({ logger }) {
         nvidiaGPUs: resolvedNvidiaGPUs,
         rocmVRAM: resolvedRocmVRAM,
         siGPUs,
+        systemTotalBytes: memInfo?.total ?? null,
         sysfsVRAMMap
     });
     const machineMetadata = {
@@ -342696,12 +342984,13 @@ async function runModelFit(options) {
     console.log();
     const feasible = filterFeasibleModels({ detection, models: recommendedModels });
     if (feasible.length === 0) {
+        const smallestTierGB = Math.min(...recommendedModels.map(model => model.vramTierGB));
         console.error("No recommended models fit this hardware. " +
             `Budget: ${formatBytes$1(detection.gpus.some(gpu => gpu.memoryTotalBytes)
                 ? Math.max(...detection.gpus
                     .map(gpu => gpu.memoryTotalBytes ?? 0)
                     .filter(bytes => bytes > 0))
-                : detection.memory.totalBytes)}. Smallest tier starts at 16 GB.`);
+                : detection.memory.totalBytes)}. Smallest tier starts at ${smallestTierGB} GB.`);
         process.exitCode = 1;
         return;
     }

package/dist/cli.sea.cjs

@@ -21581,6 +21581,90 @@ object$5({
 // Source of truth for curated recommended models. Consumed by @infersec/definitions
 // and the website model page generator (applications/website/scripts/generate-model-pages.ts).
 const modelsData = [
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-2.6B",
+        blurb: "Liquid's LFM2.5 2.6B hybrid at high fidelity - agentic post-training, 128K context, and 220 tok/s decode on an M5 Max. Fits a 4 GB card with room to spare.",
+        contextLength: 131072,
+        fileSizeBytes: 2874779648,
+        format: "gguf",
+        license: "lfm1.0",
+        parameterCount: "2.6B",
+        quantization: "Q8_0",
+        slug: "LiquidAI/LFM2.5-2.6B-GGUF",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF",
+        vramTierGB: 4
+    },
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-2.6B",
+        blurb: "Liquid's LFM2.5 2.6B hybrid at 4-bit - competitive with models 4x its size on tool use and multi-step agentic tasks. A 1.7 GB download that sips memory.",
+        contextLength: 131072,
+        fileSizeBytes: 1674455040,
+        format: "gguf",
+        license: "lfm1.0",
+        parameterCount: "2.6B",
+        quantization: "Q4_K_M",
+        slug: "LiquidAI/LFM2.5-2.6B-GGUF",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF",
+        vramTierGB: 4
+    },
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-2.6B",
+        blurb: "The native BF16 checkpoint of Liquid's LFM2.5 2.6B for vLLM and SGLang - hybrid conv + GQA architecture, 128K context window, and agentic reinforcement-learning post-training.",
+        contextLength: 131072,
+        fileSizeBytes: 5394427456,
+        format: "safetensors",
+        license: "lfm1.0",
+        parameterCount: "2.6B",
+        quantization: "BF16",
+        slug: "LiquidAI/LFM2.5-2.6B",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-2.6B",
+        vramTierGB: 8
+    },
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-230M",
+        blurb: "Liquid's most compact LFM2.5 at high fidelity - a 230M edge model distilled from the 350M, tuned for tool use and data extraction. 213 tok/s decode on a Galaxy S25 Ultra.",
+        contextLength: 32768,
+        fileSizeBytes: 246598496,
+        format: "gguf",
+        license: "lfm1.0",
+        parameterCount: "230M",
+        quantization: "Q8_0",
+        slug: "LiquidAI/LFM2.5-230M-GGUF",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-230M-GGUF",
+        vramTierGB: 2
+    },
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-230M",
+        blurb: "Liquid's most compact LFM2.5 at 4-bit - a 153 MB download that brings real tool-use capability to the tightest memory budgets, from Raspberry Pi to phones.",
+        contextLength: 32768,
+        fileSizeBytes: 153406304,
+        format: "gguf",
+        license: "lfm1.0",
+        parameterCount: "230M",
+        quantization: "Q4_K_M",
+        slug: "LiquidAI/LFM2.5-230M-GGUF",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-230M-GGUF",
+        vramTierGB: 2
+    },
+    {
+        author: "LiquidAI",
+        baseModel: "LiquidAI/LFM2.5-230M",
+        blurb: "The native BF16 checkpoint of Liquid's LFM2.5 230M for vLLM - a 230M edge model with a 32K context window, suited to data extraction and lightweight on-device agent pipelines.",
+        contextLength: 32768,
+        fileSizeBytes: 459401112,
+        format: "safetensors",
+        license: "lfm1.0",
+        parameterCount: "230M",
+        quantization: "BF16",
+        slug: "LiquidAI/LFM2.5-230M",
+        sourceURL: "https://huggingface.co/LiquidAI/LFM2.5-230M",
+        vramTierGB: 4
+    },
     {
         author: "Qwen",
         baseModel: "Qwen/Qwen3-8B",
@@ -21637,6 +21721,48 @@ const modelsData = [
         sourceURL: "https://huggingface.co/bartowski/deepreinforce-ai_Ornith-1.0-9B-GGUF",
         vramTierGB: 16
     },
+    {
+        author: "unsloth",
+        baseModel: "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
+        blurb: "NVIDIA's Nemotron 3.5 Lightning MoE squeezed into 2-bit - 30B total with only 3B active in a Mamba-2 + MoE + attention hybrid. Configurable thinking and a 256K context on a single 24 GB card.",
+        contextLength: 262144,
+        fileSizeBytes: 19432070208,
+        format: "gguf",
+        license: "openmdw-1.1",
+        parameterCount: "30B",
+        quantization: "UD-IQ2_M",
+        slug: "unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        sourceURL: "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        vramTierGB: 24
+    },
+    {
+        author: "unsloth",
+        baseModel: "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
+        blurb: "NVIDIA's Nemotron 3.5 Lightning at Unsloth's dynamic 4-bit - a Mamba-2 hybrid MoE with 3B active parameters, toggleable reasoning, and strong instruction following (IFBench 71.9). Needs a 32 GB GPU.",
+        contextLength: 262144,
+        fileSizeBytes: 25266255936,
+        format: "gguf",
+        license: "openmdw-1.1",
+        parameterCount: "30B",
+        quantization: "UD-Q4_K_M",
+        slug: "unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        sourceURL: "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        vramTierGB: 32
+    },
+    {
+        author: "unsloth",
+        baseModel: "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
+        blurb: "Nemotron 3.5 Lightning at 8-bit fidelity - near-reference quality from the 30B-A3B Mamba-2 hybrid, with configurable thinking and DSpark speculative decoding support. Needs a 48 GB card.",
+        contextLength: 262144,
+        fileSizeBytes: 35004643392,
+        format: "gguf",
+        license: "openmdw-1.1",
+        parameterCount: "30B",
+        quantization: "Q8_0",
+        slug: "unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        sourceURL: "https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF",
+        vramTierGB: 48
+    },
     {
         author: "unsloth",
         baseModel: "openai/gpt-oss-20b",
@@ -21695,6 +21821,76 @@ const modelsData = [
         sourceURL: "https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF",
         vramTierGB: 128
     },
+    {
+        author: "Qwen",
+        baseModel: "Qwen/Qwen3.8-Flash-Next",
+        blurb: "The official BF16 release of Qwen's experimental Qwen4-preview architecture - 180B total with only 6B activated, plus 51B n-gram embeddings, a vision encoder, and a 262K native context. For vLLM/SGLang on multi-GPU server hardware.",
+        contextLength: 262144,
+        fileSizeBytes: 360000192888,
+        format: "safetensors",
+        license: "qwen-community-1.0",
+        parameterCount: "180B",
+        quantization: "BF16",
+        slug: "Qwen/Qwen3.8-Flash-Next",
+        sourceURL: "https://huggingface.co/Qwen/Qwen3.8-Flash-Next",
+        vramTierGB: 640
+    },
+    {
+        author: "RadixArk",
+        baseModel: "Qwen/Qwen3.8-Flash-Next",
+        blurb: "Qwen3.8-Flash-Next with its routed experts quantized to NVIDIA's NVFP4 W4A4 via Model Optimizer - 135 GB instead of 360 GB BF16, keeping GSM8K/AIME in-band with the reference. Serve with SGLang on Blackwell hardware.",
+        contextLength: 262144,
+        fileSizeBytes: 135195303851,
+        format: "safetensors",
+        license: "qwen-community-1.0",
+        parameterCount: "180B",
+        quantization: "NVFP4",
+        slug: "RadixArk/Qwen3.8-Flash-Next-NVFP4",
+        sourceURL: "https://huggingface.co/RadixArk/Qwen3.8-Flash-Next-NVFP4",
+        vramTierGB: 256
+    },
+    {
+        author: "deepseek-ai",
+        baseModel: "deepseek-ai/DeepSeek-V4-Flash",
+        blurb: "DeepSeek's V4-Flash - a 284B MoE with only 13B activated and a native 1M context window, shipped in mixed FP4/FP8 precision. Hybrid CSA/HCA attention keeps long-context inference cheap, with three reasoning-effort modes including Think Max.",
+        contextLength: 1048576,
+        fileSizeBytes: 159617149040,
+        format: "safetensors",
+        license: "mit",
+        parameterCount: "284B",
+        quantization: "FP4-FP8",
+        slug: "deepseek-ai/DeepSeek-V4-Flash",
+        sourceURL: "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash",
+        vramTierGB: 256
+    },
+    {
+        author: "unsloth",
+        baseModel: "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp",
+        blurb: "Unsloth's 2-bit cut of DeepSeek's experimental vision variant of V4-Flash - the same 284B MoE with 13B activated, plus image understanding via a separate mmproj projector. Fits 128 GB of unified memory or 2x 96 GB.",
+        contextLength: 1048576,
+        fileSizeBytes: 90692965818,
+        format: "gguf",
+        license: "mit",
+        parameterCount: "284B",
+        quantization: "UD-IQ2_XXS",
+        slug: "unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF",
+        sourceURL: "https://huggingface.co/unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF",
+        vramTierGB: 128
+    },
+    {
+        author: "unsloth",
+        baseModel: "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp",
+        blurb: "The 4-bit cut of DeepSeek's experimental vision variant of V4-Flash - 284B MoE with 13B activated at higher fidelity, with multimodal agent scores well clear of the text-only release. Image input currently needs a preview llama.cpp build; needs 256 GB.",
+        contextLength: 1048576,
+        fileSizeBytes: 136662447706,
+        format: "gguf",
+        license: "mit",
+        parameterCount: "284B",
+        quantization: "UD-IQ4_XS",
+        slug: "unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF",
+        sourceURL: "https://huggingface.co/unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF",
+        vramTierGB: 256
+    },
     {
         author: "unsloth",
         baseModel: "Qwen/Qwen3.6-27B",
@@ -157969,12 +158165,20 @@ async function detectVRAMViaSysfs(pciBusSuffix) {
             const deviceDir = path$1.join(DRM_PATH, entry, "device");
             const totalStr = await readSysfsFile(path$1.join(deviceDir, "mem_info_vram_total"));
             const usedStr = await readSysfsFile(path$1.join(deviceDir, "mem_info_vram_used"));
+            const gttTotalStr = await readSysfsFile(path$1.join(deviceDir, "mem_info_gtt_total"));
+            const gttUsedStr = await readSysfsFile(path$1.join(deviceDir, "mem_info_gtt_used"));
             const totalBytes = totalStr !== null ? parseInt(totalStr, 10) : null;
             const usedBytes = usedStr !== null ? parseInt(usedStr, 10) : null;
+            const gttTotalBytes = gttTotalStr !== null ? parseInt(gttTotalStr, 10) : null;
+            const gttUsedBytes = gttUsedStr !== null ? parseInt(gttUsedStr, 10) : null;
             const validTotal = totalBytes !== null && Number.isFinite(totalBytes) && totalBytes >= 0;
             const validUsed = usedBytes !== null && Number.isFinite(usedBytes) && usedBytes >= 0;
+            const validGttTotal = gttTotalBytes !== null && Number.isFinite(gttTotalBytes) && gttTotalBytes >= 0;
+            const validGttUsed = gttUsedBytes !== null && Number.isFinite(gttUsedBytes) && gttUsedBytes >= 0;
             if (validTotal) {
                 return {
+                    gttTotalBytes: validGttTotal ? gttTotalBytes : null,
+                    gttUsedBytes: validGttUsed ? gttUsedBytes : null,
                     memoryTotalBytes: totalBytes,
                     memoryUsedBytes: validUsed ? usedBytes : null
                 };
@@ -157984,36 +158188,74 @@ async function detectVRAMViaSysfs(pciBusSuffix) {
     catch {
         // sysfs not available
     }
-    return { memoryTotalBytes: null, memoryUsedBytes: null };
+    return {
+        gttTotalBytes: null,
+        gttUsedBytes: null,
+        memoryTotalBytes: null,
+        memoryUsedBytes: null
+    };
 }
-async function detectVRAMViaRocmSmi() {
+// rocm-smi accepts a single --showmeminfo type per invocation, so VRAM and GTT
+// arrive from separate calls; both share this per-card parser.
+function parseRocmSmiMemory({ keyPrefix, stdout }) {
+    const parsed = JSON.parse(stdout);
+    const results = [];
+    const cards = Object.entries(parsed).filter(([key]) => key.startsWith("card"));
+    for (const [, data] of cards) {
+        const bus = data["PCI Bus"] ?? null;
+        const totalStr = data[`${keyPrefix} Total Memory (B)`] ?? null;
+        const usedStr = data[`${keyPrefix} Total Used Memory (B)`] ?? null;
+        const totalBytes = totalStr !== null ? parseInt(totalStr, 10) : null;
+        const usedBytes = usedStr !== null ? parseInt(usedStr, 10) : null;
+        const validTotal = totalBytes !== null && Number.isFinite(totalBytes) && totalBytes >= 0;
+        const validUsed = usedBytes !== null && Number.isFinite(usedBytes) && usedBytes >= 0;
+        if (bus) {
+            results.push({
+                bus,
+                memoryTotalBytes: validTotal ? totalBytes : null,
+                memoryUsedBytes: validUsed ? usedBytes : null
+            });
+        }
+    }
+    return results;
+}
+const ROCM_SMI_TIMEOUT_MS = 10_000;
+async function detectVRAMViaRocmSmi({ logger }) {
     try {
-        const { stdout } = await execa("rocm-smi", [
-            "--showbus",
-            "--showmeminfo",
-            "vram",
-            "--json"
+        const [vramResult, gttResult] = await Promise.allSettled([
+            execa("rocm-smi", ["--showbus", "--showmeminfo", "vram", "--json"], {
+                timeout: ROCM_SMI_TIMEOUT_MS
+            }),
+            execa("rocm-smi", ["--showbus", "--showmeminfo", "gtt", "--json"], {
+                timeout: ROCM_SMI_TIMEOUT_MS
+            })
         ]);
-        const parsed = JSON.parse(stdout);
-        const results = [];
-        const cards = Object.entries(parsed).filter(([key]) => key.startsWith("card"));
-        for (const [, data] of cards) {
-            const bus = data["PCI Bus"] ?? null;
-            const totalStr = data["VRAM Total Memory (B)"] ?? null;
-            const usedStr = data["VRAM Total Used Memory (B)"] ?? null;
-            const totalBytes = totalStr !== null ? parseInt(totalStr, 10) : null;
-            const usedBytes = usedStr !== null ? parseInt(usedStr, 10) : null;
-            const validTotal = totalBytes !== null && Number.isFinite(totalBytes) && totalBytes >= 0;
-            const validUsed = usedBytes !== null && Number.isFinite(usedBytes) && usedBytes >= 0;
-            if (bus) {
-                results.push({
-                    bus,
-                    memoryTotalBytes: validTotal ? totalBytes : null,
-                    memoryUsedBytes: validUsed ? usedBytes : null
+        if (vramResult.status !== "fulfilled")
+            return [];
+        let gttEntries = [];
+        if (gttResult.status === "fulfilled") {
+            try {
+                gttEntries = parseRocmSmiMemory({
+                    keyPrefix: "GTT",
+                    stdout: gttResult.value.stdout
                 });
             }
+            catch (error) {
+                // Unusable GTT output (e.g. older rocm-smi) degrades to VRAM-only
+                logger.warn("rocm-smi GTT output parse failed", { error: asError(error) });
+            }
         }
-        return results;
+        const gttByBus = new Map(gttEntries.map(entry => [entry.bus, entry]));
+        return parseRocmSmiMemory({ keyPrefix: "VRAM", stdout: vramResult.value.stdout }).map(vram => {
+            const gtt = gttByBus.get(vram.bus);
+            return {
+                bus: vram.bus,
+                gttTotalBytes: gtt?.memoryTotalBytes ?? null,
+                gttUsedBytes: gtt?.memoryUsedBytes ?? null,
+                memoryTotalBytes: vram.memoryTotalBytes,
+                memoryUsedBytes: vram.memoryUsedBytes
+            };
+        });
     }
     catch {
         return [];
@@ -158078,8 +158320,9 @@ async function detectGPUsViaNvidiaSmi({ logger }) {
                 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.
+        // Shared/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. (AMD iGPUs get the same treatment via GTT merging.)
         if (gpus.some(gpu => gpu.memoryTotalBytes === null)) {
             const systemMemory = await readSystemMemoryBytes({ logger });
             const total = systemMemory.totalBytes;
@@ -158103,7 +158346,7 @@ async function detectGPUsViaNvidiaSmi({ logger }) {
     }
 }
 function buildMergedGPUs(options) {
-    const { lspciGPUs, nvidiaGPUs, rocmVRAM, siGPUs, sysfsVRAMMap } = options;
+    const { lspciGPUs, nvidiaGPUs, rocmVRAM, siGPUs, systemTotalBytes, sysfsVRAMMap } = options;
     const rocmByBus = new Map();
     for (const entry of rocmVRAM) {
         const key = normalizeBusAddress(entry.bus);
@@ -158144,6 +158387,7 @@ function buildMergedGPUs(options) {
                     gpu: existing,
                     key,
                     rocmByBus,
+                    systemTotalBytes,
                     sysfsVRAMMap
                 });
             }
@@ -158163,30 +158407,73 @@ function buildMergedGPUs(options) {
             gpu,
             key,
             rocmByBus,
+            systemTotalBytes,
             sysfsVRAMMap
         });
         byBus.set(key, gpu);
     }
     return [...byBus.values()];
 }
-function applySysfsOrRocmVRAM({ gpu, key, rocmByBus, sysfsVRAMMap }) {
+// Shared-memory GPUs (AMD iGPUs) expose a small dedicated VRAM carve-out via
+// mem_info_vram_total while the real usable pool - GTT - is carved dynamically
+// from system RAM. When GTT exceeds VRAM the device is treated as integrated
+// and the two pools combine (capped by installed RAM). Note that GTT defaults
+// to half of system RAM even on discrete GPUs, so a dGPU with less VRAM than
+// that also merges - acceptable, since GTT remains real addressable memory
+// for amdgpu compute (albeit slower over PCIe).
+function mergeGTTMemory({ gttTotalBytes, gttUsedBytes, systemTotalBytes, vramTotalBytes, vramUsedBytes }) {
+    if (vramTotalBytes === null || !Number.isFinite(vramTotalBytes)) {
+        return { memoryTotalBytes: null, memoryUsedBytes: null };
+    }
+    const isIntegrated = gttTotalBytes !== null && gttTotalBytes > vramTotalBytes;
+    if (!isIntegrated) {
+        return { memoryTotalBytes: vramTotalBytes, memoryUsedBytes: vramUsedBytes };
+    }
+    const combined = vramTotalBytes + gttTotalBytes;
+    const capped = systemTotalBytes !== null && systemTotalBytes > 0
+        ? Math.min(combined, systemTotalBytes)
+        : combined;
+    const hasCompleteUsage = vramUsedBytes !== null &&
+        Number.isFinite(vramUsedBytes) &&
+        gttUsedBytes !== null &&
+        Number.isFinite(gttUsedBytes);
+    const memoryUsedBytes = hasCompleteUsage
+        ? Math.min(vramUsedBytes + gttUsedBytes, capped)
+        : null;
+    return { memoryTotalBytes: capped, memoryUsedBytes };
+}
+function applySysfsOrRocmVRAM({ gpu, key, rocmByBus, systemTotalBytes, sysfsVRAMMap }) {
     const sysfs = sysfsVRAMMap.get(key);
-    let totalBytes = sysfs?.memoryTotalBytes ?? null;
-    let usedBytes = sysfs?.memoryUsedBytes ?? null;
-    if (totalBytes === null) {
+    let vramTotalBytes = sysfs?.memoryTotalBytes ?? null;
+    let vramUsedBytes = sysfs?.memoryUsedBytes ?? null;
+    let gttTotalBytes = sysfs?.gttTotalBytes ?? null;
+    let gttUsedBytes = sysfs?.gttUsedBytes ?? null;
+    if (vramTotalBytes === null) {
         const rocm = rocmByBus.get(key);
         if (rocm) {
-            totalBytes = rocm.memoryTotalBytes;
-            usedBytes = rocm.memoryUsedBytes;
-        }
-    }
-    if (totalBytes === null || !Number.isFinite(totalBytes))
+            vramTotalBytes = rocm.memoryTotalBytes;
+            vramUsedBytes = rocm.memoryUsedBytes;
+            gttTotalBytes = rocm.gttTotalBytes;
+            gttUsedBytes = rocm.gttUsedBytes;
+        }
+    }
+    const merged = mergeGTTMemory({
+        gttTotalBytes,
+        gttUsedBytes,
+        systemTotalBytes,
+        vramTotalBytes,
+        vramUsedBytes
+    });
+    if (merged.memoryTotalBytes === null || !Number.isFinite(merged.memoryTotalBytes))
         return;
-    gpu.memoryTotalBytes = totalBytes;
-    gpu.memoryUsedBytes = usedBytes !== null && Number.isFinite(usedBytes) ? usedBytes : null;
+    gpu.memoryTotalBytes = merged.memoryTotalBytes;
+    gpu.memoryUsedBytes =
+        merged.memoryUsedBytes !== null && Number.isFinite(merged.memoryUsedBytes)
+            ? merged.memoryUsedBytes
+            : null;
     gpu.memoryFreeBytes =
-        usedBytes !== null && Number.isFinite(usedBytes)
-            ? Math.max(totalBytes - usedBytes, 0)
+        gpu.memoryUsedBytes !== null
+            ? Math.max(gpu.memoryTotalBytes - gpu.memoryUsedBytes, 0)
             : null;
 }
 async function collectMachineMetadata({ logger }) {
@@ -158197,7 +158484,7 @@ async function collectMachineMetadata({ logger }) {
         si.graphics(),
         detectGPUsViaLspci(),
         detectGPUsViaNvidiaSmi({ logger }),
-        detectVRAMViaRocmSmi()
+        detectVRAMViaRocmSmi({ logger })
     ]);
     const cpuInfo = cpuResult.status === "fulfilled" ? cpuResult.value : null;
     const memInfo = memResult.status === "fulfilled" ? memResult.value : null;
@@ -158252,6 +158539,7 @@ async function collectMachineMetadata({ logger }) {
         nvidiaGPUs: resolvedNvidiaGPUs,
         rocmVRAM: resolvedRocmVRAM,
         siGPUs,
+        systemTotalBytes: memInfo?.total ?? null,
         sysfsVRAMMap
     });
     const machineMetadata = {
@@ -362925,12 +363213,13 @@ async function runModelFit(options) {
     console.log();
     const feasible = filterFeasibleModels({ detection, models: recommendedModels });
     if (feasible.length === 0) {
+        const smallestTierGB = Math.min(...recommendedModels.map(model => model.vramTierGB));
         console.error("No recommended models fit this hardware. " +
             `Budget: ${formatBytes$1(detection.gpus.some(gpu => gpu.memoryTotalBytes)
                 ? Math.max(...detection.gpus
                     .map(gpu => gpu.memoryTotalBytes ?? 0)
                     .filter(bytes => bytes > 0))
-                : detection.memory.totalBytes)}. Smallest tier starts at 16 GB.`);
+                : detection.memory.totalBytes)}. Smallest tier starts at ${smallestTierGB} GB.`);
         process.exitCode = 1;
         return;
     }

package/dist/utils/machineInfo.d.ts

@@ -1,8 +1,26 @@
 import type { InferenceAgentMachineMetadata } from "@infersec/definitions";
 import type { Logger } from "@infersec/logger";
 export declare function normalizeBusAddress(bus: string | null | undefined): string | null;
+export declare function parseRocmSmiMemory({ keyPrefix, stdout }: {
+    keyPrefix: "GTT" | "VRAM";
+    stdout: string;
+}): Array<{
+    bus: string;
+    memoryTotalBytes: number | null;
+    memoryUsedBytes: number | null;
+}>;
 export declare function parseNvidiaSmiNumber(value: string | undefined): number | null;
 export declare function parseNvidiaSmiMemoryBytes(value: string | undefined): number | null;
+export declare function mergeGTTMemory({ gttTotalBytes, gttUsedBytes, systemTotalBytes, vramTotalBytes, vramUsedBytes }: {
+    gttTotalBytes: number | null;
+    gttUsedBytes: number | null;
+    systemTotalBytes: number | null;
+    vramTotalBytes: number | null;
+    vramUsedBytes: number | null;
+}): {
+    memoryTotalBytes: number | null;
+    memoryUsedBytes: 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": "1.111.0",
+  "version": "1.112.1",
   "bin": {
     "infersec-conduit": "./dist/cli.js"
   },