UPSTREAM PR #20653: Models: Add control vector functions to qwen3.5 and qwen-next models

[loci-dev]

Note

Source pull request: ggml-org/llama.cpp#20653

This PR adds control vector support to Qwen3.5, Qwen3.5-MoE and Qwen3-next (Qwen3-coder-next).
It was missing before but was present for Qwen3-vl and so on.
Control vector PR for reference: ggml-org/llama.cpp#5970

The PR resolves ggml-org/llama.cpp#20541

AI disclaimer: I had Claude point me at the missing parts and help me understand the correct position for the two lines added in each file.
I tested the changes myself and made sure it doesn't break things where I could. The local CI pipeline shows no problem with the new code.

[loci-review]

Overview

This PR adds control vector support to Qwen3.5 and Qwen3Next models with negligible performance impact. Out of 120,579 total functions analyzed, only 4 were modified (0 new, 0 removed). The single function exceeding the 2% threshold shows +342ns (+0.225%) response time increase during graph construction—a one-time initialization cost that does not affect per-token inference performance.

Power Consumption Changes:

• build.bin.libllama.so: -0.08%
• All other binaries (build.bin.llama-cvector-generator, build.bin.llama-tts, build.bin.libmtmd.so, build.bin.llama-bench, build.bin.llama-quantize, build.bin.llama-qwen2vl-cli, build.bin.llama-tokenize, build.bin.llama-gemma3-cli, build.bin.llama-gguf-split, build.bin.llama-llava-cli, build.bin.llama-minicpmv-cli, build.bin.libggml.so, build.bin.libggml-base.so, build.bin.libggml-cpu.so): 0.00%

Function Analysis

llm_build_qwen3next (build.bin.libllama.so)

• Response time: 151,935ns → 152,278ns (+342ns, +0.225%)
• Throughput time: 470ns → 483ns (+12ns, +2.6%)
• Source changes: Added 3 lines in src/models/qwen3next.cpp to call build_cvec(cur, il) after FFN layer construction, enabling control vector transformations per layer
• Impact assessment: Non-critical path (graph construction during initialization, not inference loop). The 342ns overhead is distributed across all model layers and occurs once per session, not per token. CFG analysis shows the new build_cvec call adds ~383ns per layer with minor improvements in existing calls (-39ns, -18ns) partially offsetting the overhead.

Other modified functions showed changes below the 2% threshold.

🔎 Full breakdown: Loci Inspector
💬 Questions? Tag @loci-dev