llama-quant : honor --tensor-type override when it matches the global ftype

[Anai-Guo]

Summary

Fixes #22544.

When a user supplies an explicit --tensor-type "<pattern>=<type>" mapping that happens to match the requested global ftype, the user's intent (lock that tensor to that exact type) is silently dropped and the imatrix/heuristic path is allowed to override it.

Root cause

llama_tensor_get_type only set manual = true from inside the qtype != new_type branch:

for (const auto & [pattern, qtype] : qs.tensor_type_patterns) {
    if (std::regex_search(tensor_name, pattern)) {
        if (qtype != new_type) {
            LLAMA_LOG_WARN(...);
            new_type = qtype;
            manual = true;
            break;
        }
        // pattern matched but qtype == new_type:
        //   no break, no manual = true
        //   loop continues, manual stays false
    }
}

// since !manual:
new_type = llama_tensor_get_type_impl(...);  // overrides the user's choice

So invoking llama-quantize with --tensor-type "^blk\.0\.attn_qkv\.weight$=iq4_xs" and global ftype iq4_xs falls through to the heuristic, which can pick a different type (e.g. q5_K) — exactly what --tensor-type is meant to prevent.

This was introduced by 1dab5f5 (refactor type selection), reported in #22544.

Fix

Set manual = true and break unconditionally once a pattern matches. Only emit the "applying manual override" log line when the requested type actually differs from the current one:

if (std::regex_search(tensor_name, pattern)) {
    if (qtype != new_type) {
        LLAMA_LOG_WARN(...);
        new_type = qtype;
    }
    manual = true;
    break;
}

A pattern match is the user pinning that tensor — it should suppress the heuristic regardless of whether it changes the type.

Test plan

• ./llama-quantize --imatrix model.imatrix.gguf --tensor-type "^blk\.0\.attn_qkv\.weight$=iq4_xs" model.gguf out.gguf iq4_xs 8 — attn_qkv should now be reported as iq4_xs (no longer redirected to q5_K).
• Existing override case (qtype != new_type) still emits the log line and switches type.
• No regressions when --tensor-type is not supplied (tensor_type_patterns.empty() → unchanged path).

🤖 Generated with Claude Code

[ggml-gh-bot]

Hi @Anai-Guo, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• Multiple open PRs from a new contributor: We limit new contributors (those without a previously merged PR) to 1 open PR at a time. You currently have 2 open PRs.

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[Anai-Guo]

Apologies — I missed the new-contributor 1-open-PR limit and the AI-disclosure policy. Closing this in favor of my older open PR (#21838). I'll wait for that one to land before opening anything else here. Thanks for the patience.

[ddh0]

Hi @Anai-Guo, since you had to close this one, I took the liberty of submitting the fix here: #22572

Thank you! I'm sorry for introducing this mistake.