Misc. bug: new kv cell seq implementation does not handle "seq_id = -1" specified in the API

[l3utterfly]

Name and Version

The API in llama.h specifies:

// Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
    // seq_id < 0 : match any sequence
    // p0 < 0     : [0,  p1]
    // p1 < 0     : [p0, inf)
    LLAMA_API bool llama_kv_self_seq_rm(
            struct llama_context * ctx,
                    llama_seq_id   seq_id,
                       llama_pos   p0,
                       llama_pos   p1);

This doesn't seem to work anymore with the new kv cell implementation. In particular in the llama-kv-cells.h, there are assertions for seq_id >= 0, e.g.:

// check if the cell contains seq_id
    bool seq_has(uint32_t i, llama_seq_id seq_id) const {
        assert(i < pos.size());
        assert(seq_id >= 0);

        return seq[i].test(seq_id);
    }

Are we planning to keep the "any sequence" feature in the api?

Note, the seq = -1 input is still being used in main, when loading session caches, so it currently does not work. After loading session, it does this:

// remove any "future" tokens that we might have inherited from the previous session
        llama_kv_self_seq_rm(ctx, -1, n_matching_session_tokens, -1);

Operating systems

No response

Which llama.cpp modules do you know to be affected?

No response

Command line

Problem description & steps to reproduce

1. build main project
2. run, save session, then load session

First Bad Commit

No response

Relevant log output

bitset test argument out of range

[ggerganov]

Yes, this is an oversight. Let me push a fix.

[ggerganov]

@l3utterfly Please check if #13985 resolves the issue.

[l3utterfly]

@ggerganov Yes, this fixes the issue. Thank you!