llmman 0.1.689

Run any agent on any model, models stored as OCI images
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
//! The dynamically-loaded CUDA/HIP/OpenCL/Vulkan probes behind
//! [`super::detect_gpu_api_uncontained`]. Its own file purely so the
//! whole section hangs off the one `#[cfg]` on its `mod` declaration:
//! macOS has none of these four runtimes (it uses Metal — see
//! [`super::detect_macos`]), so gating them one attribute at a time is
//! what let ~22 `dead_code` warnings accumulate there unnoticed.
//!
//! Mirrors `cuda/probe.c`, `rocm/probe.cc` and `vulkan/probe.c` from
//! ggml-org/llama-install.sh — see `hostgpu.rs` for why these are
//! reached by runtime dynamic loading rather than as probe binaries.

use std::ffi::c_void;
use std::os::raw::c_char;

use libloading::{Library, Symbol};

use super::{igpu_enabled, HostGpu};

// ---------------------------------------------------------------------------
// Dynamic library loading — one candidate list per backend/OS, mirroring
// which shared library cuda/probe.c, rocm/probe.cc, and vulkan/probe.c
// each link against.
// ---------------------------------------------------------------------------

#[cfg(target_os = "windows")]
fn open_cuda_lib() -> Option<Library> {
    unsafe { Library::new("nvcuda.dll").ok() }
}

#[cfg(target_os = "linux")]
fn open_cuda_lib() -> Option<Library> {
    unsafe {
        Library::new("libcuda.so.1")
            .or_else(|_| Library::new("libcuda.so"))
            .ok()
    }
}

#[cfg(target_os = "windows")]
fn open_hip_lib() -> Option<Library> {
    unsafe {
        Library::new("amdhip64_6.dll")
            .or_else(|_| Library::new("amdhip64.dll"))
            .ok()
    }
}

#[cfg(target_os = "linux")]
fn open_hip_lib() -> Option<Library> {
    unsafe {
        Library::new("libamdhip64.so.6")
            .or_else(|_| Library::new("libamdhip64.so.5"))
            .or_else(|_| Library::new("libamdhip64.so"))
            .ok()
    }
}

#[cfg(all(target_os = "windows", target_arch = "aarch64"))]
fn open_opencl_lib() -> Option<Library> {
    unsafe { Library::new("OpenCL.dll").ok() }
}

#[cfg(target_os = "windows")]
fn open_vulkan_lib() -> Option<Library> {
    unsafe { Library::new("vulkan-1.dll").ok() }
}

#[cfg(target_os = "linux")]
fn open_vulkan_lib() -> Option<Library> {
    unsafe {
        Library::new("libvulkan.so.1")
            .or_else(|_| Library::new("libvulkan.so"))
            .ok()
    }
}

// ---------------------------------------------------------------------------
// CUDA — mirrors cuda/probe.c: cuDriverGetVersion, cuInit,
// cuDeviceGetCount, then (best-effort) cuDeviceGet/cuDeviceGetAttribute
// per device. Unlike probe.c, there's no PROBE_ARCH/PROBE_VERSION
// best-arch match against llama-app's own per-SM-arch build matrix —
// llama.cpp's official GitHub releases only split CUDA builds by major
// version (12 vs. 13, from the driver's own reported version), not by SM
// architecture — so the per-device compute-capability query below is
// kept only to confirm those same calls succeed against a real device,
// not to select anything.
// ---------------------------------------------------------------------------

const CUDA_SUCCESS: i32 = 0;
const CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR: i32 = 75;
const CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR: i32 = 76;

/// CUDA's driver version int encodes `major*1000 + minor*10` (e.g. 12040
/// = "12.4", 13030 = "13.3") — this is `cuDriverGetVersion`'s own
/// documented format, the same value `nvidia-smi`'s human-readable "CUDA
/// Version: X.Y" header is derived from.
fn cuda_major_from_driver_version(driver_version: i32) -> u32 {
    (driver_version / 1000).max(0) as u32
}

/// `Some((kind, total_vram_bytes))` if at least one CUDA device is
/// present. Sums every visible device's memory (via
/// `cuDeviceTotalMem_v2`), not just device 0's, so a multi-GPU host's
/// combined VRAM weighs it correctly in aggregation (see
/// `hostgpu::memory_bytes`) — matching `detect_vulkan_inner`'s existing
/// multi-device summation below.
pub(super) fn detect_cuda() -> Option<(HostGpu, u64)> {
    let lib = open_cuda_lib()?;
    unsafe {
        let cu_driver_get_version: Symbol<unsafe extern "C" fn(*mut i32) -> i32> =
            lib.get(b"cuDriverGetVersion\0").ok()?;
        let cu_init: Symbol<unsafe extern "C" fn(u32) -> i32> = lib.get(b"cuInit\0").ok()?;
        let cu_device_get_count: Symbol<unsafe extern "C" fn(*mut i32) -> i32> =
            lib.get(b"cuDeviceGetCount\0").ok()?;

        let mut driver_version: i32 = 0;
        if cu_driver_get_version(&mut driver_version) != CUDA_SUCCESS {
            return None;
        }
        if cu_init(0) != CUDA_SUCCESS {
            return None;
        }
        let mut count: i32 = 0;
        if cu_device_get_count(&mut count) != CUDA_SUCCESS || count == 0 {
            return None;
        }

        let mut vram: u64 = 0;
        if let (Ok(cu_device_get), Ok(cu_device_get_attribute)) = (
            lib.get::<unsafe extern "C" fn(*mut i32, i32) -> i32>(b"cuDeviceGet\0"),
            lib.get::<unsafe extern "C" fn(*mut i32, i32, i32) -> i32>(b"cuDeviceGetAttribute\0"),
        ) {
            let cu_device_total_mem = lib
                .get::<unsafe extern "C" fn(*mut u64, i32) -> i32>(b"cuDeviceTotalMem_v2\0")
                .ok();
            for index in 0..count {
                let mut device: i32 = 0;
                if cu_device_get(&mut device, index) != CUDA_SUCCESS {
                    continue;
                }
                let mut major: i32 = 0;
                let mut minor: i32 = 0;
                cu_device_get_attribute(
                    &mut major,
                    CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR,
                    device,
                );
                cu_device_get_attribute(
                    &mut minor,
                    CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR,
                    device,
                );
                crate::debug_log!("CUDA device {index} compute capability: {major}.{minor}");

                if let Some(cu_device_total_mem) = &cu_device_total_mem {
                    let mut bytes: u64 = 0;
                    if cu_device_total_mem(&mut bytes, device) == CUDA_SUCCESS {
                        vram += bytes;
                    }
                }
            }
        }

        Some((
            HostGpu::Cuda {
                major: cuda_major_from_driver_version(driver_version),
            },
            vram,
        ))
    }
}

// ---------------------------------------------------------------------------
// ROCm — mirrors rocm/probe.cc: hipGetDeviceCount. probe.cc goes on to
// read hipGetDeviceProperties(...).gcnArchName; skipped here since
// llmman's Rocm variant carries no arch detail (llama.cpp's ROCm release
// asset isn't split by GPU arch either) and hipDeviceProp_t's layout
// isn't ABI-frozen the way CUDA/Vulkan's structs used elsewhere in this
// file are (it has grown fields across ROCm releases), making it unsafe
// to lay out by hand here.
// ---------------------------------------------------------------------------

const HIP_SUCCESS: i32 = 0;

/// `Some(total_vram_bytes)` if at least one HIP device is present, `None`
/// if not. Sums every device's memory rather than just device 0's — see
/// `detect_cuda`'s doc comment.
pub(super) fn detect_rocm() -> Option<u64> {
    let lib = open_hip_lib()?;
    unsafe {
        let hip_get_device_count = lib
            .get::<unsafe extern "C" fn(*mut i32) -> i32>(b"hipGetDeviceCount\0")
            .ok()?;
        let mut count: i32 = 0;
        if hip_get_device_count(&mut count) != HIP_SUCCESS || count == 0 {
            return None;
        }

        let mut vram: u64 = 0;
        if let (Ok(hip_set_device), Ok(hip_mem_get_info)) = (
            lib.get::<unsafe extern "C" fn(i32) -> i32>(b"hipSetDevice\0"),
            lib.get::<unsafe extern "C" fn(*mut u64, *mut u64) -> i32>(b"hipMemGetInfo\0"),
        ) {
            for index in 0..count {
                if hip_set_device(index) != HIP_SUCCESS {
                    continue;
                }
                let (mut free, mut total): (u64, u64) = (0, 0);
                if hip_mem_get_info(&mut free, &mut total) == HIP_SUCCESS {
                    vram += total;
                }
            }
        }
        Some(vram)
    }
}

// ---------------------------------------------------------------------------
// OpenCL — Windows ARM64 only: llama.cpp's sole OpenCL release asset is
// the win-opencl-adreno-arm64 build, so this matches Adreno GPU devices
// and nothing else. On every other target it is a no-op that lets
// detection fall through to Vulkan.
// ---------------------------------------------------------------------------

/// `Some(total_global_mem_bytes)` if at least one Adreno OpenCL GPU is
/// present, summed across devices like `detect_cuda`.
#[cfg(all(target_os = "windows", target_arch = "aarch64"))]
pub(super) fn detect_opencl() -> Option<u64> {
    const CL_SUCCESS: i32 = 0;
    const CL_DEVICE_TYPE_GPU: u64 = 1 << 2;
    const CL_DEVICE_GLOBAL_MEM_SIZE: u32 = 0x101F;
    const CL_DEVICE_NAME: u32 = 0x102B;

    let lib = open_opencl_lib()?;
    unsafe {
        let get_platform_ids: Symbol<unsafe extern "C" fn(u32, *mut *mut c_void, *mut u32) -> i32> =
            lib.get(b"clGetPlatformIDs\0").ok()?;
        let get_device_ids: Symbol<
            unsafe extern "C" fn(*mut c_void, u64, u32, *mut *mut c_void, *mut u32) -> i32,
        > = lib.get(b"clGetDeviceIDs\0").ok()?;
        let get_device_info: Symbol<
            unsafe extern "C" fn(*mut c_void, u32, usize, *mut c_void, *mut usize) -> i32,
        > = lib.get(b"clGetDeviceInfo\0").ok()?;

        let mut platform_count: u32 = 0;
        if get_platform_ids(0, std::ptr::null_mut(), &mut platform_count) != CL_SUCCESS
            || platform_count == 0
        {
            return None;
        }
        let mut platforms = vec![std::ptr::null_mut(); platform_count as usize];
        if get_platform_ids(platform_count, platforms.as_mut_ptr(), std::ptr::null_mut())
            != CL_SUCCESS
        {
            return None;
        }

        let mut found: Option<u64> = None;
        for platform in platforms {
            let mut device_count: u32 = 0;
            if get_device_ids(
                platform,
                CL_DEVICE_TYPE_GPU,
                0,
                std::ptr::null_mut(),
                &mut device_count,
            ) != CL_SUCCESS
                || device_count == 0
            {
                continue;
            }
            let mut devices = vec![std::ptr::null_mut(); device_count as usize];
            if get_device_ids(
                platform,
                CL_DEVICE_TYPE_GPU,
                device_count,
                devices.as_mut_ptr(),
                std::ptr::null_mut(),
            ) != CL_SUCCESS
            {
                continue;
            }
            for device in devices {
                let mut size: usize = 0;
                if get_device_info(device, CL_DEVICE_NAME, 0, std::ptr::null_mut(), &mut size)
                    != CL_SUCCESS
                    || size == 0
                {
                    continue;
                }
                let mut name = vec![0u8; size];
                if get_device_info(
                    device,
                    CL_DEVICE_NAME,
                    name.len(),
                    name.as_mut_ptr().cast(),
                    std::ptr::null_mut(),
                ) != CL_SUCCESS
                {
                    continue;
                }
                if !String::from_utf8_lossy(&name)
                    .to_ascii_lowercase()
                    .contains("adreno")
                {
                    continue;
                }
                let mut mem: u64 = 0;
                if get_device_info(
                    device,
                    CL_DEVICE_GLOBAL_MEM_SIZE,
                    std::mem::size_of::<u64>(),
                    (&mut mem as *mut u64).cast(),
                    std::ptr::null_mut(),
                ) != CL_SUCCESS
                {
                    mem = 0;
                }
                found = Some(found.unwrap_or(0) + mem);
            }
        }
        found
    }
}

#[cfg(not(all(target_os = "windows", target_arch = "aarch64")))]
pub(super) fn detect_opencl() -> Option<u64> {
    None
}

// ---------------------------------------------------------------------------
// Vulkan — mirrors vulkan/probe.c: vkCreateInstance,
// vkEnumeratePhysicalDevices, vkGetPhysicalDeviceProperties (skip
// VK_PHYSICAL_DEVICE_TYPE_CPU always, and _INTEGRATED_GPU unless
// LLMMAN_IGPU_ENABLE is set — llmman's own addition, not in probe.c —
// and require apiVersion >= 1.2),
// vkGetPhysicalDeviceQueueFamilyProperties (require a queue family with
// both COMPUTE and TRANSFER bits set).
// ---------------------------------------------------------------------------

const VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO: i32 = 1;
const VK_SUCCESS: i32 = 0;
const VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU: i32 = 1;
const VK_PHYSICAL_DEVICE_TYPE_CPU: i32 = 4;
const VK_QUEUE_COMPUTE_BIT: u32 = 0x2;
const VK_QUEUE_TRANSFER_BIT: u32 = 0x4;

fn vk_make_api_version(variant: u32, major: u32, minor: u32, patch: u32) -> u32 {
    (variant << 29) | (major << 22) | (minor << 12) | patch
}

type VkInstance = *mut c_void;
type VkPhysicalDevice = *mut c_void;

#[repr(C)]
struct VkInstanceCreateInfo {
    s_type: i32,
    p_next: *const c_void,
    flags: u32,
    p_application_info: *const c_void,
    enabled_layer_count: u32,
    pp_enabled_layer_names: *const *const c_char,
    enabled_extension_count: u32,
    pp_enabled_extension_names: *const *const c_char,
}

#[repr(C)]
#[derive(Clone, Copy, Default)]
struct VkExtent3D {
    width: u32,
    height: u32,
    depth: u32,
}

#[repr(C)]
#[derive(Clone, Copy, Default)]
struct VkQueueFamilyProperties {
    queue_flags: u32,
    queue_count: u32,
    timestamp_valid_bits: u32,
    min_image_transfer_granularity: VkExtent3D,
}

/// A generously oversized, 8-byte-aligned stand-in for
/// `VkPhysicalDeviceProperties` (~824 bytes on a 64-bit build, dominated
/// by the large `VkPhysicalDeviceLimits` sub-struct). The Vulkan spec
/// freezes this struct's layout for ABI compatibility, so it's safe for
/// `vkGetPhysicalDeviceProperties` to write its real, full contents into
/// any buffer at least this big and this aligned — this only ever reads
/// back the two leading fields it actually needs (`apiVersion` at offset
/// 0, `deviceType` at offset 16), exactly what `vulkan/probe.c` reads off
/// its own fully-typed `VkPhysicalDeviceProperties`.
#[repr(C, align(8))]
struct RawPhysicalDeviceProperties([u8; 1024]);

const VK_MAX_MEMORY_HEAPS: usize = 16;
const VK_MEMORY_HEAP_DEVICE_LOCAL_BIT: u32 = 0x1;

#[repr(C)]
#[derive(Clone, Copy, Default)]
struct VkMemoryHeap {
    size: u64,
    flags: u32,
    _pad: u32,
}

/// `VkPhysicalDeviceMemoryProperties`, minus its 32-entry `memoryTypes`
/// array (unused here) — safe to overwrite as a stand-in the same way
/// [`RawPhysicalDeviceProperties`] is: `vkGetPhysicalDeviceMemoryProperties`
/// only ever reads/writes the struct's fixed, spec-frozen layout, and
/// this reads back just `memoryHeapCount`/`memoryHeaps`, past a
/// same-sized raw padding block standing in for `memoryTypes`.
#[repr(C)]
struct RawPhysicalDeviceMemoryProperties {
    memory_type_count: u32,
    _memory_types_padding: [u8; 32 * 8],
    memory_heap_count: u32,
    memory_heaps: [VkMemoryHeap; VK_MAX_MEMORY_HEAPS],
}

pub(super) fn detect_vulkan() -> Option<u64> {
    let lib = open_vulkan_lib()?;
    unsafe { detect_vulkan_inner(&lib).unwrap_or(None) }
}

unsafe fn detect_vulkan_inner(lib: &Library) -> Option<Option<u64>> {
    let vk_create_instance: Symbol<
        unsafe extern "C" fn(*const VkInstanceCreateInfo, *const c_void, *mut VkInstance) -> i32,
    > = lib.get(b"vkCreateInstance\0").ok()?;
    let vk_destroy_instance: Symbol<unsafe extern "C" fn(VkInstance, *const c_void)> =
        lib.get(b"vkDestroyInstance\0").ok()?;
    let vk_enumerate_physical_devices: Symbol<
        unsafe extern "C" fn(VkInstance, *mut u32, *mut VkPhysicalDevice) -> i32,
    > = lib.get(b"vkEnumeratePhysicalDevices\0").ok()?;
    let vk_get_physical_device_properties: Symbol<unsafe extern "C" fn(VkPhysicalDevice, *mut u8)> =
        lib.get(b"vkGetPhysicalDeviceProperties\0").ok()?;
    let vk_get_physical_device_queue_family_properties: Symbol<
        unsafe extern "C" fn(VkPhysicalDevice, *mut u32, *mut VkQueueFamilyProperties),
    > = lib
        .get(b"vkGetPhysicalDeviceQueueFamilyProperties\0")
        .ok()?;
    let vk_get_physical_device_memory_properties: Symbol<
        unsafe extern "C" fn(VkPhysicalDevice, *mut RawPhysicalDeviceMemoryProperties),
    > = lib.get(b"vkGetPhysicalDeviceMemoryProperties\0").ok()?;

    let create_info = VkInstanceCreateInfo {
        s_type: VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
        p_next: std::ptr::null(),
        flags: 0,
        p_application_info: std::ptr::null(),
        enabled_layer_count: 0,
        pp_enabled_layer_names: std::ptr::null(),
        enabled_extension_count: 0,
        pp_enabled_extension_names: std::ptr::null(),
    };
    let mut instance: VkInstance = std::ptr::null_mut();
    if vk_create_instance(&create_info, std::ptr::null(), &mut instance) != VK_SUCCESS {
        return Some(None);
    }

    let mut count: u32 = 0;
    let found = if vk_enumerate_physical_devices(instance, &mut count, std::ptr::null_mut())
        != VK_SUCCESS
        || count == 0
    {
        None
    } else {
        let mut devices: Vec<VkPhysicalDevice> = vec![std::ptr::null_mut(); count as usize];
        if vk_enumerate_physical_devices(instance, &mut count, devices.as_mut_ptr()) != VK_SUCCESS {
            None
        } else {
            let mut found: Option<u64> = None;
            for &device in &devices {
                let mut props_buf = RawPhysicalDeviceProperties([0u8; 1024]);
                vk_get_physical_device_properties(device, props_buf.0.as_mut_ptr());
                let api_version = u32::from_ne_bytes(props_buf.0[0..4].try_into().unwrap());
                let device_type = i32::from_ne_bytes(props_buf.0[16..20].try_into().unwrap());

                if device_type == VK_PHYSICAL_DEVICE_TYPE_CPU {
                    continue;
                }
                if device_type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU && !igpu_enabled() {
                    continue;
                }
                if api_version < vk_make_api_version(0, 1, 2, 0) {
                    continue;
                }

                let mut qcount: u32 = 0;
                vk_get_physical_device_queue_family_properties(
                    device,
                    &mut qcount,
                    std::ptr::null_mut(),
                );
                if qcount == 0 {
                    continue;
                }
                let mut queues = vec![VkQueueFamilyProperties::default(); qcount as usize];
                vk_get_physical_device_queue_family_properties(
                    device,
                    &mut qcount,
                    queues.as_mut_ptr(),
                );

                let has_compute = queues.iter().any(|q| {
                    q.queue_flags & VK_QUEUE_COMPUTE_BIT != 0
                        && q.queue_flags & VK_QUEUE_TRANSFER_BIT != 0
                });
                if !has_compute {
                    continue;
                }

                let mut mem_props = RawPhysicalDeviceMemoryProperties {
                    memory_type_count: 0,
                    _memory_types_padding: [0u8; 32 * 8],
                    memory_heap_count: 0,
                    memory_heaps: [VkMemoryHeap::default(); VK_MAX_MEMORY_HEAPS],
                };
                vk_get_physical_device_memory_properties(device, &mut mem_props);
                let vram: u64 = mem_props.memory_heaps
                    [..(mem_props.memory_heap_count as usize).min(VK_MAX_MEMORY_HEAPS)]
                    .iter()
                    .filter(|h| h.flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT != 0)
                    .map(|h| h.size)
                    .sum();
                found = Some(found.unwrap_or(0) + vram);
            }
            found
        }
    };

    vk_destroy_instance(instance, std::ptr::null());
    Some(found)
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn cuda_major_from_driver_version_matches_documented_encoding() {
        assert_eq!(cuda_major_from_driver_version(12040), 12); // "12.4"
        assert_eq!(cuda_major_from_driver_version(13030), 13); // "13.3"
        assert_eq!(cuda_major_from_driver_version(0), 0);
    }

    #[test]
    fn vk_make_api_version_matches_the_vulkan_header_macro() {
        // VK_API_VERSION_1_2 per vulkan_core.h: VK_MAKE_API_VERSION(0, 1, 2, 0)
        assert_eq!(vk_make_api_version(0, 1, 2, 0), (1 << 22) | (2 << 12));
        assert_eq!(vk_make_api_version(0, 1, 0, 0), 1 << 22);
    }

    /// Exercises the real dynamic-loaded CUDA/HIP/Vulkan probes against
    /// whatever hardware/drivers are actually present on the machine
    /// running the test — not run by a plain `cargo test` since its
    /// result depends entirely on the host, but useful to eyeball
    /// directly. Run explicitly with `cargo test --bin llmman -- --ignored
    /// --nocapture detect_reports_this_hosts_real_hardware`.
    #[test]
    #[ignore = "result depends on this host's actual GPU/driver setup"]
    fn detect_reports_this_hosts_real_hardware() {
        // `detect()` itself is deliberately not exercised here:
        // `detect_gpu_api_isolated` re-execs `std::env::current_exe()`,
        // which under `cargo test` is this test harness binary, not the
        // real `llmman` CLI binary — its generated `main()` doesn't check
        // for PROBE_SUBPROCESS_ARG at all, so `detect()` only ever
        // reports HostGpu::None when run this way (a `cargo test`-only
        // artifact, not something a real caller of the compiled `llmman`
        // binary hits). To verify the real re-exec path end to end
        // instead, run this directly against a real build:
        //
        //   cargo build --bin llmman
        //   ./target/debug/llmman __hostgpu-probe
        //
        // — confirmed directly this way while writing this module.
        println!("detect_cuda() -> {:?}", detect_cuda());
        println!("detect_rocm() -> {:?}", detect_rocm());
        println!("detect_vulkan() -> {:?}", detect_vulkan());
        println!(
            "detect_with_vram() -> {:?}",
            super::super::detect_with_vram()
        );
    }
}