MCPcopy Create free account
hub / github.com/Tiiny-AI/PowerInfer / test_gradient_accumulation

Function test_gradient_accumulation

smallthinker/tests/test-opt.cpp:580–686  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

578}
579
580static std::pair<int, int> test_gradient_accumulation(
581 ggml_backend_sched_t backend_sched, ggml_backend_t backend, const int32_t nbatch_physical, const enum ggml_opt_loss_type loss_type) {
582 int ntest = 0;
583 int npass = 0;
584
585 struct helper_ctx_data cd = helper_get_ctx_data(
586 backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false, /*nbatch_logical =*/ 6, nbatch_physical, loss_type);
587
588 std::vector<float> grad_history(ndata);
589 for (int64_t idata = 0; idata < ndata; ++idata) {
590 grad_history[idata] = NAN;
591 }
592
593 for (int epoch = 1; epoch <= 4; ++epoch) {
594 if (nbatch_physical == 1) {
595 for (int idata = 0; idata < ndata; ++idata) {
596 const float idataf = idata;
597 ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
598 ggml_backend_tensor_set(cd.inputs, &idataf, 0, 1*sizeof(float));
599 ggml_opt_eval(cd.opt_ctx, cd.result);
600 ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata, 0, 1*sizeof(float));
601 }
602 } else if (nbatch_physical == 2) {
603 for (int idata = 0; idata < ndata; idata += 2) {
604 const float idataf[2] = {float(idata + 0), float(idata + 1)};
605 ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
606 ggml_backend_tensor_set(cd.inputs, idataf, 0, 2*sizeof(float));
607 ggml_opt_eval(cd.opt_ctx, cd.result);
608
609 grad_history[idata + 0] = 0.0f;
610 ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata + 1, 0, 1*sizeof(float));
611 }
612 } else {
613 GGML_ASSERT(false);
614 }
615
616 {
617 GGML_ASSERT(ndata == 6);
618 constexpr double atol = 1e-6;
619 bool subtest_ok = true;
620 if (loss_type == GGML_OPT_LOSS_TYPE_SUM) {
621 if (nbatch_physical == 1) {
622 subtest_ok = subtest_ok && almost_equal(grad_history[0], 1.0, atol);
623 subtest_ok = subtest_ok && almost_equal(grad_history[2], 3.0, atol);
624 subtest_ok = subtest_ok && almost_equal(grad_history[4], 5.0, atol);
625 } else {
626 subtest_ok = subtest_ok && almost_equal(grad_history[0], 0.0, atol);
627 subtest_ok = subtest_ok && almost_equal(grad_history[2], 0.0, atol);
628 subtest_ok = subtest_ok && almost_equal(grad_history[4], 0.0, atol);
629 }
630 subtest_ok = subtest_ok && almost_equal(grad_history[1], 2.0, atol);
631 subtest_ok = subtest_ok && almost_equal(grad_history[3], 4.0, atol);
632 subtest_ok = subtest_ok && almost_equal(grad_history[5], 6.0, atol);
633 } else if (loss_type == GGML_OPT_LOSS_TYPE_MEAN) {
634 if (nbatch_physical == 1) {
635 subtest_ok = subtest_ok && almost_equal(grad_history[0], 1.0/ndata, atol);
636 subtest_ok = subtest_ok && almost_equal(grad_history[2], 3.0/ndata, atol);
637 subtest_ok = subtest_ok && almost_equal(grad_history[4], 5.0/ndata, atol);

Callers 1

test_backendFunction · 0.85

Calls 14

helper_get_ctx_dataFunction · 0.85
ggml_opt_allocFunction · 0.85
ggml_opt_evalFunction · 0.85
ggml_opt_grad_accFunction · 0.85
almost_equalFunction · 0.85
ggml_opt_result_ndataFunction · 0.85
ggml_opt_result_lossFunction · 0.85
ggml_opt_result_accuracyFunction · 0.85
ggml_opt_result_resetFunction · 0.85
helper_free_ctx_dataFunction · 0.85
ggml_backend_tensor_setFunction · 0.50

Tested by

no test coverage detected