Function sam3_read_f32

sam3.cpp:1069–1085

Read tensor data from backend into a float buffer, handling F32, F16, and quantized types. n = number of float elements to produce.

// Read tensor data from backend into a float buffer, handling F32, F16, and
// quantized types.  n = number of float elements to produce.
static void sam3_read_f32(struct ggml_tensor* t, float* dst, int64_t n) {
    if (t->type == GGML_TYPE_F32) {
        ggml_backend_tensor_get(t, dst, 0, n * sizeof(float));
    } else if (t->type == GGML_TYPE_F16) {
        std::vector<ggml_fp16_t> tmp(n);
        ggml_backend_tensor_get(t, tmp.data(), 0, n * sizeof(ggml_fp16_t));
        ggml_fp16_to_fp32_row(tmp.data(), dst, (int)n);
    } else if (ggml_is_quantized(t->type)) {
        const size_t nbytes = ggml_nbytes(t);
        std::vector<uint8_t> raw(nbytes);
        ggml_backend_tensor_get(t, raw.data(), 0, nbytes);
        const auto * traits = ggml_get_type_traits(t->type);
        traits->to_float(raw.data(), dst, n);
    } else {
        fprintf(stderr, "%s: unsupported tensor type %d\n", __func__, (int)t->type);
    }
}

static struct ggml_tensor* sam3_conv_transpose_weight(struct ggml_context* ctx,
                                                      struct ggml_tensor* w) {