Function run_sampled_generation

ds4_cli.c:421–543

    const char *think_open = "<think>";
    const char *think_close = "</think>";
    size_t total = p->pending_len + len;
    char *buf = malloc(total ? total : 1);
    if (!buf) return;
    if (p->pending_len) memcpy(buf, p->pending, p->pending_len);
    if (len) memcpy(buf + p->pending_len, text, len);
    p->pending_len = 0;

    size_t i = 0;
    while (i < total) {
        const char *cur = buf + i;
        const size_t rem = total - i;
        if (bytes_has_prefix(cur, rem, think_open)) {
            p->in_think = true;
            i += strlen(think_open);
            continue;
        }
        if (bytes_has_prefix(cur, rem, think_close)) {
            p->in_think = false;
            token_printer_reset_color(p);
            if (!p->last_output_newline) {
                fputc('\n', p->fp);
                p->last_output_newline = true;
            }
            i += strlen(think_close);
            continue;
        }
        if (!finish && cur[0] == '<' &&
            (bytes_is_partial_prefix(cur, rem, think_open) ||
             bytes_is_partial_prefix(cur, rem, think_close)))
        {
            if (rem < sizeof(p->pending)) {
                memcpy(p->pending, cur, rem);
                p->pending_len = rem;
            }
            break;
        }
        token_printer_write_char(p, cur[0]);
        i++;
    }

    free(buf);
}

static void token_printer_finish(token_printer *p) {
    if (p->format_thinking) {
        token_printer_process(p, NULL, 0, true);
        token_printer_reset_color(p);
    }
    fflush(p->fp);
}

static void generation_done(void *ud) {
    token_printer *p = ud;
    token_printer_finish(p);
    if (!p->last_output_newline) {
        fputc('\n', p->fp);
        p->last_output_newline = true;
    }