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
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
use cmake::Config;
use glob::glob;
use std::path::{Path, PathBuf};
use std::process::Command;
use std::{env, fs};
macro_rules! debug_log {
($($arg:tt)*) => {
if std::env::var("BUILD_DEBUG").is_ok() {
println!("cargo:warning=[DEBUG] {}", format!($($arg)*));
}
};
}
fn get_cargo_target_dir() -> Result<std::path::PathBuf, Box<dyn std::error::Error>> {
let out_dir = std::path::PathBuf::from(std::env::var("OUT_DIR")?);
let profile = std::env::var("PROFILE")?;
let mut target_dir = None;
let mut sub_path = out_dir.as_path();
while let Some(parent) = sub_path.parent() {
if parent.ends_with(&profile) {
target_dir = Some(parent);
break;
}
sub_path = parent;
}
let target_dir = target_dir.ok_or("not found")?;
Ok(target_dir.to_path_buf())
}
/// Return a string that uniquely identifies the current state of the llama.cpp
/// submodule so we know when a re-copy is needed.
///
/// Priority:
/// 1. The commit hash from the submodule's git HEAD (most precise).
/// 2. The mtime of `CMakeLists.txt` (fallback for non-git trees).
fn llama_src_version(src: &Path) -> String {
// In a git submodule the `.git` entry is a *file* whose content is:
// gitdir: ../../.git/modules/llama-cpp-sys-4/llama.cpp
let git_file = src.join(".git");
if git_file.is_file() {
if let Ok(text) = std::fs::read_to_string(&git_file) {
if let Some(rel) = text.strip_prefix("gitdir:").map(str::trim) {
let head_path = git_file.parent().unwrap().join(rel).join("HEAD");
if let Ok(head) = std::fs::read_to_string(&head_path) {
// HEAD is either a commit hash or "ref: refs/heads/…"
let head = head.trim();
if head.starts_with("ref:") {
// Resolve the ref to the actual commit hash.
let ref_path = head
.strip_prefix("ref:")
.map(str::trim)
.unwrap_or(head);
let commit_path =
git_file.parent().unwrap().join(rel).join(ref_path);
if let Ok(hash) = std::fs::read_to_string(commit_path) {
return hash.trim().to_owned();
}
}
return head.to_owned();
}
}
}
}
// Fallback: modification time of the top-level CMakeLists.txt.
src.join("CMakeLists.txt")
.metadata()
.and_then(|m| m.modified())
.map(|t| format!("{t:?}"))
.unwrap_or_else(|_| "unknown".to_owned())
}
/// Copy a directory tree. This runs on the *host*, so cfg!(unix/windows) is correct here.
fn copy_folder(src: &Path, dst: &Path) {
std::fs::create_dir_all(dst).expect("Failed to create dst directory");
if cfg!(unix) {
std::process::Command::new("cp")
.arg("-rf")
.arg(src)
.arg(dst.parent().unwrap())
.status()
.expect("Failed to execute cp command");
}
if cfg!(windows) {
std::process::Command::new("robocopy.exe")
.arg("/e")
.arg(src)
.arg(dst)
.status()
.expect("Failed to execute robocopy command");
}
}
/// Extract library names from the build output directory.
///
/// `target` is the Rust target triple of the *cross-compilation target* so
/// that the correct file extensions are chosen even when cross-compiling.
fn extract_lib_names(out_dir: &Path, build_shared_libs: bool, target: &str) -> Vec<String> {
let lib_pattern = if target.contains("windows") {
"*.lib"
} else if target.contains("apple") {
if build_shared_libs {
"*.dylib"
} else {
"*.a"
}
} else {
if build_shared_libs {
"*.so"
} else {
"*.a"
}
};
let libs_dir = out_dir.join("lib*");
let pattern = libs_dir.join(lib_pattern);
debug_log!("Extract libs {}", pattern.display());
let mut lib_names: Vec<String> = Vec::new();
// Process the libraries based on the pattern
for entry in glob(pattern.to_str().unwrap()).unwrap() {
match entry {
Ok(path) => {
let stem = path.file_stem().unwrap();
let stem_str = stem.to_str().unwrap();
// Remove the "lib" prefix if present
let lib_name = if stem_str.starts_with("lib") {
stem_str.strip_prefix("lib").unwrap_or(stem_str)
} else {
stem_str
};
lib_names.push(lib_name.to_string());
}
Err(e) => println!("cargo:warning=error={}", e),
}
}
lib_names
}
/// Extract shared-library asset paths from the build output directory.
///
/// `target` is the Rust target triple of the *cross-compilation target*.
fn extract_lib_assets(out_dir: &Path, target: &str) -> Vec<PathBuf> {
let shared_lib_pattern = if target.contains("windows") {
"*.dll"
} else if target.contains("apple") {
"*.dylib"
} else {
"*.so"
};
let shared_libs_dir = if target.contains("windows") { "bin" } else { "lib" };
let libs_dir = out_dir.join(shared_libs_dir);
let pattern = libs_dir.join(shared_lib_pattern);
debug_log!("Extract lib assets {}", pattern.display());
let mut files = Vec::new();
for entry in glob(pattern.to_str().unwrap()).unwrap() {
match entry {
Ok(path) => {
files.push(path);
}
Err(e) => eprintln!("cargo:warning=error={}", e),
}
}
files
}
/// Ask a clang binary for its library search path (macOS link helper).
///
/// `clang_binary` should be the bare name or full path of the clang binary to
/// query — e.g. `"clang"` for native builds or `"aarch64-apple-darwin-clang"`
/// for a cross-compiler.
fn macos_link_search_path(clang_binary: &str) -> Option<String> {
let output = Command::new(clang_binary)
.arg("--print-search-dirs")
.output()
.ok()?;
if !output.status.success() {
println!(
"failed to run '{clang_binary} --print-search-dirs', continuing without a link search path"
);
return None;
}
let stdout = String::from_utf8_lossy(&output.stdout);
for line in stdout.lines() {
if line.contains("libraries: =") {
let path = line.split('=').nth(1)?;
return Some(format!("{}/lib/darwin", path));
}
}
println!("failed to determine link search path, continuing without it");
None
}
/// Map a Rust target triple to the CMake `CMAKE_SYSTEM_NAME` value.
fn cmake_system_name(target: &str) -> &'static str {
if target.contains("-android") || target.contains("android-") {
"Android"
} else if target.contains("-apple-ios") {
"iOS"
} else if target.contains("-apple-") {
"Darwin"
} else if target.contains("-windows") {
"Windows"
} else if target.contains("-linux") {
"Linux"
} else {
// Generic UNIX-like fallback
"Linux"
}
}
/// Derive a MinGW cross-compiler binary name from a Rust `windows-gnu` target triple.
///
/// Rust uses `x86_64-pc-windows-gnu` / `x86_64-pc-windows-gnullvm` while the
/// MinGW toolchain conventionally uses `x86_64-w64-mingw32`. The `gnullvm`
/// variant uses Clang instead of GCC.
///
/// Returns `None` for `windows-msvc` targets — MSVC cannot cross-compile from
/// a non-Windows host and users must supply `CC`/`CXX` themselves.
fn mingw_compiler(target: &str, cxx: bool) -> Option<String> {
if !target.contains("windows-gnu") {
return None;
}
let arch = if target.contains("x86_64") {
"x86_64"
} else if target.contains("i686") || target.contains("i586") {
"i686"
} else if target.contains("aarch64") {
"aarch64"
} else {
target.split('-').next()?
};
// `gnullvm` targets use LLVM/Clang; plain `gnu` targets use GCC.
let compiler = if target.contains("gnullvm") {
if cxx { "clang++" } else { "clang" }
} else {
if cxx { "g++" } else { "gcc" }
};
Some(format!("{}-w64-mingw32-{}", arch, compiler))
}
/// Map a Rust target triple to the CMake `CMAKE_SYSTEM_PROCESSOR` value.
fn cmake_system_processor(target: &str) -> String {
let arch = target.split('-').next().unwrap_or("unknown");
match arch {
"x86_64" => "x86_64".to_owned(),
"i686" | "i386" => "x86".to_owned(),
"aarch64" | "arm64" => "aarch64".to_owned(),
"armv7" | "armv7s" | "armv7k" => "armv7-a".to_owned(),
"arm" => "arm".to_owned(),
"riscv64gc" | "riscv64" => "riscv64".to_owned(),
"powerpc64le" => "ppc64le".to_owned(),
"powerpc64" => "ppc64".to_owned(),
"s390x" => "s390x".to_owned(),
"wasm32" => "wasm32".to_owned(),
other => other.to_owned(),
}
}
fn main() {
let target = env::var("TARGET").unwrap();
let host = env::var("HOST").unwrap();
let is_cross = host != target;
let out_dir = PathBuf::from(env::var("OUT_DIR").unwrap());
let target_dir = get_cargo_target_dir().unwrap();
let llama_dst = out_dir.join("llama.cpp");
let manifest_dir = env::var("CARGO_MANIFEST_DIR").expect("Failed to get CARGO_MANIFEST_DIR");
let llama_src = Path::new(&manifest_dir).join("llama.cpp");
let build_shared_libs = cfg!(feature = "cuda") || cfg!(feature = "dynamic-link");
let build_shared_libs = std::env::var("LLAMA_BUILD_SHARED_LIBS")
.map(|v| v == "1")
.unwrap_or(build_shared_libs);
let profile = env::var("LLAMA_LIB_PROFILE").unwrap_or("Release".to_string());
let static_crt = env::var("LLAMA_STATIC_CRT")
.map(|v| v == "1")
.unwrap_or(false);
debug_log!("HOST: {}", host);
debug_log!("TARGET: {}", target);
debug_log!("CROSS_COMPILING: {}", is_cross);
debug_log!("CARGO_MANIFEST_DIR: {}", manifest_dir);
debug_log!("TARGET_DIR: {}", target_dir.display());
debug_log!("OUT_DIR: {}", out_dir.display());
debug_log!("BUILD_SHARED: {}", build_shared_libs);
// ── Source copy with version tracking ────────────────────────────────────
// The copy only ran when the OUT_DIR was fresh, so updating the submodule
// (which adds/removes files like ggml-cpu/) would silently use stale data.
// We now store the current submodule HEAD in a sentinel file and re-copy
// whenever it changes.
let sentinel = out_dir.join(".llama-src-version");
let current_version = llama_src_version(&llama_src);
let stored_version = std::fs::read_to_string(&sentinel).unwrap_or_default();
let needs_copy = !llama_dst.exists() || stored_version.trim() != current_version.trim();
if needs_copy {
if llama_dst.exists() {
debug_log!("Source version changed — removing stale OUT_DIR copy");
std::fs::remove_dir_all(&llama_dst).ok();
}
debug_log!("Copy {} to {}", llama_src.display(), llama_dst.display());
copy_folder(&llama_src, &llama_dst);
std::fs::write(&sentinel, ¤t_version)
.expect("failed to write source version sentinel");
}
// Tell cargo to rerun this script when the submodule HEAD changes.
// In a git submodule, llama.cpp/.git is a file pointing at the real HEAD.
let submodule_git = llama_src.join(".git");
if submodule_git.is_file() {
// .git file contains "gitdir: ../../.git/modules/llama-cpp-sys-4/llama.cpp"
if let Ok(contents) = std::fs::read_to_string(&submodule_git) {
if let Some(gitdir) = contents.strip_prefix("gitdir:").map(|s| s.trim()) {
let head = submodule_git.parent().unwrap().join(gitdir).join("HEAD");
if head.exists() {
println!("cargo:rerun-if-changed={}", head.display());
}
}
}
}
// Speed up build
// TODO: Audit that the environment access only happens in single-threaded code.
unsafe {
env::set_var(
"CMAKE_BUILD_PARALLEL_LEVEL",
std::thread::available_parallelism()
.unwrap()
.get()
.to_string(),
)
};
// Point CC/CXX at the MPI wrappers when building with MPI on macOS.
// Check the *target* OS, not the host, so that cross-compilation from a
// macOS host to a non-Apple target does not accidentally set these.
if cfg!(feature = "mpi") && target.contains("apple") {
// TODO: Audit that the environment access only happens in single-threaded code.
unsafe { env::set_var("CC", "/opt/homebrew/bin/mpicc") };
// TODO: Audit that the environment access only happens in single-threaded code.
unsafe { env::set_var("CXX", "/opt/homebrew/bin/mpicxx") };
}
// ── Bindgen ──────────────────────────────────────────────────────────────
let mut builder = bindgen::Builder::default()
.header("wrapper.h")
.generate_comments(true)
// https://github.com/rust-lang/rust-bindgen/issues/1834
// "fatal error: 'string' file not found" on macOS
.clang_arg("-xc++")
.clang_arg("-std=c++17")
// When cross-compiling, tell libclang/bindgen the target triple so
// that layout, pointer sizes, and type widths are computed for the
// *target* architecture rather than the host.
.clang_arg(format!("--target={}", target))
// .raw_line("#![feature(unsafe_extern_blocks)]") // https://github.com/rust-lang/rust/issues/123743
.clang_arg(format!("-I{}", llama_dst.join("include").display()))
.clang_arg(format!("-I{}", llama_dst.join("ggml/include").display()))
.clang_arg(format!("-I{}", llama_dst.join("src").display()))
.clang_arg(format!("-I{}", llama_dst.join("common").display()))
.parse_callbacks(Box::new(bindgen::CargoCallbacks::new()))
.derive_partialeq(true)
// Do not derive PartialEq on types that contain function-pointer fields.
// Deriving PartialEq on those triggers the
// `unpredictable_function_pointer_comparisons` lint on newer rustc
// because function addresses are not stable across codegen units.
// macOS FILE internals (function pointers _close/_read/_seek/_write)
.no_partialeq("__sFILE")
.no_partialeq("ggml_cplan")
.no_partialeq("ggml_type_traits")
.no_partialeq("ggml_type_traits_cpu")
.no_partialeq("ggml_context")
.no_partialeq("ggml_opt_params")
.no_partialeq("llama_model_params")
.no_partialeq("llama_context_params")
.no_partialeq("llama_sampler_i")
.no_partialeq("llama_opt_params")
.allowlist_function("ggml_.*")
.allowlist_type("ggml_.*")
.allowlist_function("llama_.*")
.allowlist_function("llama_lora_.*")
.allowlist_type("llama_.*")
.allowlist_function("common_token_to_piece")
.allowlist_function("common_tokenize")
// .allowlist_item("common_.*")
// .allowlist_function("common_tokenize")
// .allowlist_function("common_detokenize")
// .allowlist_type("common_.*")
// .allowlist_item("common_params")
// .allowlist_item("common_sampler_type")
// .allowlist_item("common_sampler_params")
.allowlist_item("LLAMA_.*")
// .opaque_type("common_lora_adapter_info")
.opaque_type("llama_grammar")
.opaque_type("llama_grammar_parser")
.opaque_type("llama_sampler_chain")
// .opaque_type("llama_context_deleter")
// .blocklist_type("llama_model_deleter")
.opaque_type("std::.*");
// Add RPC support if feature is enabled
if cfg!(feature = "rpc") {
builder = builder
.clang_arg("-DRPC_SUPPORT")
.allowlist_function("ggml_backend_rpc_.*")
.allowlist_type("ggml_backend_rpc_.*");
}
// Add mtmd (multimodal) support if feature is enabled
if cfg!(feature = "mtmd") {
builder = builder
.clang_arg("-DMTMD_SUPPORT")
.clang_arg(format!("-I{}", llama_dst.join("tools/mtmd").display()))
.allowlist_function("mtmd_.*")
.allowlist_type("mtmd_.*")
.allowlist_item("MTMD_.*")
.no_partialeq("mtmd_context_params");
}
let bindings = builder
// .layout_tests(false)
// .derive_default(true)
// .enable_cxx_namespaces()
.use_core()
.prepend_enum_name(false)
.generate()
.expect("Failed to generate bindings");
// Write the generated bindings to an output file
let bindings_path = out_dir.join("bindings.rs");
bindings
.write_to_file(bindings_path.clone())
.expect("Failed to write bindings");
// temporary fix for https://github.com/rust-lang/rust/issues/123743 in
// cargo +nightly build
let contents = std::fs::read_to_string(bindings_path.clone()).unwrap();
let contents = contents.replace("unsafe extern \"C\" {", " extern \"C\" {");
fs::write(bindings_path, contents).unwrap();
println!("cargo:rerun-if-changed=wrapper.h");
println!("cargo:rerun-if-changed=./sherpa-onnx");
debug_log!("Bindings Created");
// ── CMake build ──────────────────────────────────────────────────────────
let mut config = Config::new(&llama_dst);
// Would require extra source files to pointlessly
// be included in what's uploaded to and downloaded from
// crates.io, so deactivating these instead
config.define("LLAMA_BUILD_TESTS", "OFF");
config.define("LLAMA_BUILD_EXAMPLES", "OFF");
config.define("LLAMA_BUILD_SERVER", "OFF");
// Build tools (including the mtmd library) only when the mtmd feature is
// requested. Common is also required because the CMakeLists gate for
// tools is `if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS)`.
if cfg!(feature = "mtmd") {
config.define("LLAMA_BUILD_TOOLS", "ON");
config.define("LLAMA_BUILD_COMMON", "ON");
} else {
config.define("LLAMA_BUILD_TOOLS", "OFF");
}
config.define(
"BUILD_SHARED_LIBS",
if build_shared_libs { "ON" } else { "OFF" },
);
// ── Cross-compilation CMake configuration ────────────────────────────────
// When building for a different target than the host, tell CMake the
// target system so that it does not auto-detect the host as the target.
// Android is handled separately below via its NDK toolchain file.
if is_cross && !target.contains("android") {
let system_name = cmake_system_name(&target);
let system_processor = cmake_system_processor(&target);
debug_log!("Cross-compiling: CMAKE_SYSTEM_NAME={system_name} CMAKE_SYSTEM_PROCESSOR={system_processor}");
config.define("CMAKE_SYSTEM_NAME", system_name);
config.define("CMAKE_SYSTEM_PROCESSOR", &system_processor);
// CMake only sets CMAKE_CROSSCOMPILING=TRUE automatically when
// CMAKE_SYSTEM_NAME differs from the host OS name. For same-OS
// cross-arch builds (e.g. x86_64-linux → aarch64-linux) the OS
// names are identical, so CMAKE_CROSSCOMPILING stays FALSE and
// ggml's guard (`if (CMAKE_CROSSCOMPILING)` in ggml/CMakeLists.txt)
// never fires — leaving GGML_NATIVE_DEFAULT=ON and causing
// `-march=native` (tuned for the build host) to be baked into the
// target binary, which crashes with SIGILL on the target.
// Force the flag explicitly so ggml always sees it.
config.define("CMAKE_CROSSCOMPILING", "TRUE");
if target.contains("apple") {
// ── Apple cross-arch (e.g. x86_64-apple-darwin → aarch64-apple-darwin) ──
//
// Apple's Clang is already a universal cross-compiler; switching
// to a different compiler binary is neither needed nor possible
// (there is no `aarch64-apple-darwin-gcc` in Xcode). The right
// CMake knob for same-SDK Apple cross-arch builds is
// CMAKE_OSX_ARCHITECTURES, which makes Clang add the `-arch`
// flag automatically.
let osx_arch = if target.contains("aarch64") || target.contains("arm64") {
"arm64"
} else if target.contains("x86_64") {
"x86_64"
} else if target.contains("i686") {
"i386"
} else {
// Fallback: strip the vendor/OS suffix and use the raw arch.
target.split('-').next().unwrap_or("arm64")
};
config.define("CMAKE_OSX_ARCHITECTURES", osx_arch);
debug_log!("Apple cross-arch: CMAKE_OSX_ARCHITECTURES={osx_arch}");
// Propagate an explicit SDK path when the caller provides one.
if let Ok(sdk) = env::var("CMAKE_OSX_SYSROOT") {
config.define("CMAKE_OSX_SYSROOT", &sdk);
}
// Honour an explicit compiler override (e.g. osxcross), but do
// NOT guess a compiler name: the system Clang is always correct
// for same-SDK cross-arch and osxcross users set CC themselves.
if let Ok(cc) = env::var("CC") {
config.define("CMAKE_C_COMPILER", &cc);
}
if let Ok(cxx) = env::var("CXX") {
config.define("CMAKE_CXX_COMPILER", &cxx);
}
} else {
// ── Non-Apple cross-compilation ───────────────────────────────────────
//
// Honour CC / CXX set by the caller (e.g. cargo cross, zig cc, …).
// If they are not set:
// • Windows GNU targets → derive the MinGW triple name
// (e.g. x86_64-pc-windows-gnu → x86_64-w64-mingw32-gcc)
// • Windows MSVC targets → no safe default; MSVC cannot
// cross-compile from a non-Windows host, so the user must
// supply CC/CXX (e.g. clang-cl via a sysroot).
// • Everything else → {target-triple}-gcc / g++
if let Ok(cc) = env::var("CC") {
config.define("CMAKE_C_COMPILER", &cc);
} else if let Some(cc) = mingw_compiler(&target, false) {
config.define("CMAKE_C_COMPILER", &cc);
} else if !target.contains("windows-msvc") {
config.define("CMAKE_C_COMPILER", format!("{}-gcc", target));
}
if let Ok(cxx) = env::var("CXX") {
config.define("CMAKE_CXX_COMPILER", &cxx);
} else if let Some(cxx) = mingw_compiler(&target, true) {
config.define("CMAKE_CXX_COMPILER", &cxx);
} else if !target.contains("windows-msvc") {
config.define("CMAKE_CXX_COMPILER", format!("{}-g++", target));
}
// Propagate a sysroot when provided (e.g. via --sysroot or
// CARGO_TARGET_<TRIPLE>_RUSTFLAGS / CMAKE_SYSROOT env var).
if let Ok(sysroot) = env::var("CMAKE_SYSROOT") {
config.define("CMAKE_SYSROOT", &sysroot);
}
}
}
// ── GGML_NATIVE ──────────────────────────────────────────────────────────
// GGML_NATIVE=ON tells ggml to detect and use the *build host's* CPU
// features (e.g. -march=native, check_cxx_source_runs for ARM NEON/SVE,
// FindSIMD.cmake for MSVC). That is wrong for cross-compilation: the
// probed features belong to the build host, not the target, so the
// resulting binary would crash with SIGILL on a different microarch.
//
// Override the cmake default explicitly so a stale CMakeCache.txt can
// never re-enable it after the user switches from a native to a cross
// build in the same OUT_DIR.
if is_cross {
// Belt-and-suspenders: even though CMAKE_CROSSCOMPILING=TRUE above
// already causes ggml to default GGML_NATIVE to OFF, we pin it here
// too so the cmake crate's cache-skip path cannot resurrect a
// previously cached ON value.
config.define("GGML_NATIVE", "OFF");
} else if cfg!(feature = "native") {
// The `native` Cargo feature explicitly opts in to host-CPU
// optimisation for non-cross builds.
config.define("GGML_NATIVE", "ON");
} else {
// Default native builds to OFF so that the resulting library is
// portable across machines of the same architecture (matching the
// behaviour users expect from a Rust crate).
config.define("GGML_NATIVE", "OFF");
}
// Disable OpenMP on 32-bit ARM Windows (compiler support is absent).
// Use the TARGET env var, not cfg!(), so the check works when
// cross-compiling from a non-Windows host.
if target.contains("windows") && target.starts_with("arm") && !target.starts_with("aarch64") {
config.define("GGML_OPENMP", "OFF");
}
// static_crt (MSVC /MT vs /MD) is meaningless for MinGW; only set it for
// MSVC targets to avoid confusing CMake on windows-gnu cross builds.
if target.contains("windows-msvc") {
config.static_crt(static_crt);
}
if target.contains("android") && target.contains("aarch64") {
// build flags for android taken from this doc
// https://github.com/ggerganov/llama.cpp/blob/master/docs/android.md
let android_ndk = env::var("ANDROID_NDK")
.expect("Please install Android NDK and ensure that ANDROID_NDK env variable is set");
config.define(
"CMAKE_TOOLCHAIN_FILE",
format!("{android_ndk}/build/cmake/android.toolchain.cmake"),
);
config.define("ANDROID_ABI", "arm64-v8a");
config.define("ANDROID_PLATFORM", "android-28");
config.define("CMAKE_SYSTEM_PROCESSOR", "arm64");
config.define("CMAKE_C_FLAGS", "-march=armv8.7a");
config.define("CMAKE_CXX_FLAGS", "-march=armv8.7a");
config.define("GGML_OPENMP", "OFF");
config.define("GGML_LLAMAFILE", "OFF");
}
if cfg!(feature = "vulkan") {
config.define("GGML_VULKAN", "ON");
if target.contains("windows") {
let vulkan_path = env::var("VULKAN_SDK")
.expect("Please install Vulkan SDK and ensure that VULKAN_SDK env variable is set");
let vulkan_lib_path = Path::new(&vulkan_path).join("Lib");
println!("cargo:rustc-link-search={}", vulkan_lib_path.display());
println!("cargo:rustc-link-lib=vulkan-1");
}
if target.contains("linux") {
println!("cargo:rustc-link-lib=vulkan");
}
}
if cfg!(feature = "cuda") {
config.define("GGML_CUDA", "ON");
}
if cfg!(feature = "openmp") {
config.define("GGML_OPENMP", "ON");
} else {
config.define("GGML_OPENMP", "OFF");
}
if cfg!(feature = "mpi") {
config.define("LLAMA_MPI", "ON");
}
if cfg!(feature = "rpc") {
config.define("GGML_RPC", "ON");
}
// General
config
.profile(&profile)
.very_verbose(std::env::var("CMAKE_VERBOSE").is_ok()) // Not verbose by default
.always_configure(false);
// The cmake crate skips re-configuration when CMakeCache.txt already exists
// (always_configure = false). If a previous run left a CMakeCache.txt but
// never wrote the actual build-system files (Makefile / build.ninja), the
// subsequent `cmake --build` call fails with "No such file or directory".
// Detect that broken state and remove CMakeCache.txt so cmake is forced to
// configure from scratch.
{
let cmake_build_dir = out_dir.join("build");
let cache = cmake_build_dir.join("CMakeCache.txt");
if cache.exists() {
let has_makefile = cmake_build_dir.join("Makefile").exists();
let has_ninja = cmake_build_dir.join("build.ninja").exists();
if !has_makefile && !has_ninja {
debug_log!(
"CMakeCache.txt exists but no Makefile/build.ninja found — \
removing cache to force reconfiguration"
);
std::fs::remove_file(&cache)
.expect("failed to remove stale CMakeCache.txt");
}
}
}
let build_dir = config.build();
// ── Link search paths ────────────────────────────────────────────────────
println!("cargo:rustc-link-search={}", out_dir.join("lib").display());
println!(
"cargo:rustc-link-search={}",
out_dir.join("lib64").display()
);
println!("cargo:rustc-link-search={}", build_dir.display());
// ── Link libraries ───────────────────────────────────────────────────────
let llama_libs_kind = if build_shared_libs { "dylib" } else { "static" };
let llama_libs = extract_lib_names(&out_dir, build_shared_libs, &target);
assert_ne!(llama_libs.len(), 0);
for lib in llama_libs {
debug_log!(
"LINK {}",
format!("cargo:rustc-link-lib={}={}", llama_libs_kind, lib)
);
println!(
"{}",
format!("cargo:rustc-link-lib={}={}", llama_libs_kind, lib)
);
}
// OpenMP: link gomp when the cmake build enabled it (GGML_OPENMP_ENABLED=ON).
// This can happen even without the "openmp" feature because cmake's FindOpenMP
// is invoked unconditionally on some platforms (e.g. ARM) when the
// ggml-cpu CMakeLists includes OpenMP support at the variant level.
let cmake_cache_path = out_dir.join("build").join("CMakeCache.txt");
let openmp_enabled_in_cmake = std::fs::read_to_string(&cmake_cache_path)
.map(|contents| contents.contains("GGML_OPENMP_ENABLED:INTERNAL=ON"))
.unwrap_or(false);
if cfg!(feature = "openmp") || openmp_enabled_in_cmake {
if target.contains("gnu") || target.contains("musl") {
println!("cargo:rustc-link-lib=gomp");
}
}
// msvcrtd is the MSVC debug CRT — it does not exist in MinGW toolchains.
if cfg!(debug_assertions) && target.contains("windows-msvc") {
println!("cargo:rustc-link-lib=dylib=msvcrtd");
}
// macOS frameworks and libc++
if target.contains("apple") {
println!("cargo:rustc-link-lib=framework=Foundation");
println!("cargo:rustc-link-lib=framework=Metal");
println!("cargo:rustc-link-lib=framework=MetalKit");
println!("cargo:rustc-link-lib=framework=Accelerate");
println!("cargo:rustc-link-lib=c++");
}
// Linux libstdc++
if target.contains("linux") {
println!("cargo:rustc-link-lib=dylib=stdc++");
}
// On (older) macOS / Apple targets we may need to link against the clang
// runtime, which is hidden in a non-default path.
// More details at https://github.com/alexcrichton/curl-rust/issues/279.
if target.contains("apple") {
// For same-SDK Apple cross-arch builds (e.g. x86_64-apple-darwin →
// aarch64-apple-darwin) the host's plain `clang` is still the right
// binary to ask: both arches share the same Xcode SDK and therefore
// the same library search directories.
//
// For osxcross (Linux → macOS) the user sets CC, so we honour that;
// we do NOT guess a `{target}-clang` name because it is not a stable
// convention and the SDK paths it would report are likely wrong anyway.
let clang_bin = env::var("CC").unwrap_or_else(|_| "clang".to_owned());
if let Some(path) = macos_link_search_path(&clang_bin) {
println!("cargo:rustc-link-lib=clang_rt.osx");
println!("cargo:rustc-link-search={}", path);
}
}
// ── Copy shared-library assets to the Cargo target directory ─────────────
if build_shared_libs {
let libs_assets = extract_lib_assets(&out_dir, &target);
for asset in libs_assets {
let asset_clone = asset.clone();
let filename = asset_clone.file_name().unwrap();
let filename = filename.to_str().unwrap();
let dst = target_dir.join(filename);
debug_log!("HARD LINK {} TO {}", asset.display(), dst.display());
if !dst.exists() {
std::fs::hard_link(asset.clone(), dst).unwrap();
}
// Copy DLLs to examples as well
if target_dir.join("examples").exists() {
let dst = target_dir.join("examples").join(filename);
debug_log!("HARD LINK {} TO {}", asset.display(), dst.display());
if !dst.exists() {
std::fs::hard_link(asset.clone(), dst).unwrap();
}
}
// Copy DLLs to target/profile/deps as well for tests
let dst = target_dir.join("deps").join(filename);
debug_log!("HARD LINK {} TO {}", asset.display(), dst.display());
if !dst.exists() {
std::fs::hard_link(asset.clone(), dst).unwrap();
}
}
}
}