Compare commits

..

2 Commits

Author SHA1 Message Date
Jeff Bolz 886b97a5d6 tests: Generate unique input values for count_equal (#15487)
This avoids backend-dependent behavior for argmax that leads to intermittent failures.
2025-08-25 10:47:16 -05:00
Ihar Hrachyshka 111f8d06f0 metal: fix regression when no metal devices are present (#15531) 2025-08-25 18:27:34 +03:00
2 changed files with 40 additions and 18 deletions
+20 -18
View File
@@ -93,35 +93,37 @@ static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_dev
if (ctx->mtl_device == nil) {
ctx->mtl_device = MTLCreateSystemDefaultDevice();
ctx->has_simdgroup_reduction = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
ctx->has_simdgroup_reduction |= [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
if (ctx->mtl_device) {
ctx->has_simdgroup_reduction = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
ctx->has_simdgroup_reduction |= [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
ctx->has_simdgroup_mm = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
ctx->has_simdgroup_mm = [ctx->mtl_device supportsFamily:MTLGPUFamilyApple7];
#if defined(GGML_METAL_HAS_RESIDENCY_SETS)
ctx->has_residency_sets = getenv("GGML_METAL_NO_RESIDENCY") == nil;
ctx->has_residency_sets = getenv("GGML_METAL_NO_RESIDENCY") == nil;
#endif
ctx->has_bfloat = [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
ctx->has_bfloat |= [ctx->mtl_device supportsFamily:MTLGPUFamilyApple6];
ctx->has_bfloat = [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
ctx->has_bfloat |= [ctx->mtl_device supportsFamily:MTLGPUFamilyApple6];
#if defined(GGML_METAL_USE_BF16)
ctx->use_bfloat = ctx->has_bfloat;
ctx->use_bfloat = ctx->has_bfloat;
#else
ctx->use_bfloat = false;
ctx->use_bfloat = false;
#endif
ctx->use_fusion = getenv("GGML_METAL_FUSION_DISABLE") == nil;
ctx->use_fusion = getenv("GGML_METAL_FUSION_DISABLE") == nil;
{
const char * val = getenv("GGML_METAL_FUSION_DEBUG");
ctx->debug_fusion = val ? atoi(val) : 0;
{
const char * val = getenv("GGML_METAL_FUSION_DEBUG");
ctx->debug_fusion = val ? atoi(val) : 0;
}
memset(ctx->fuse_cnt, 0, sizeof(ctx->fuse_cnt));
ctx->max_size = ctx->mtl_device.maxBufferLength;
strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name) - 1);
}
memset(ctx->fuse_cnt, 0, sizeof(ctx->fuse_cnt));
ctx->max_size = ctx->mtl_device.maxBufferLength;
strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name) - 1);
}
ctx->mtl_device_ref_count++;
+20
View File
@@ -2209,6 +2209,26 @@ struct test_count_equal : public test_case {
double max_nmse_err() override {
return 0.0;
}
void initialize_tensors(ggml_context * ctx) override {
std::random_device rd;
std::default_random_engine rng(rd());
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
if (t->type == GGML_TYPE_F32) {
// initialize with unique values to avoid ties
for (int64_t r = 0; r < ggml_nrows(t); r++) {
std::vector<float> data(t->ne[0]);
for (int i = 0; i < t->ne[0]; i++) {
data[i] = i;
}
std::shuffle(data.begin(), data.end(), rng);
ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(float));
}
} else {
init_tensor_uniform(t);
}
}
}
};
// GGML_OP_REPEAT