UPSTREAM PR #18105: model: fix LFM2 missing tensors

[loci-dev]

Mirrored from ggml-org/llama.cpp#18105

Overlooked a change from ggml-org/llama.cpp#18051

cc @tdakhran for visibility: LFM2 reports missing tensor without this fix

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary: PR #594

Pull Request: LFM2 Missing Tensors Fix
Change Scope: Single file modification (src/llama-model.cpp), 2 lines changed
Nature: Tensor naming correction for LFM2 model architecture

---

Analysis Overview

This PR addresses a missing tensor issue in the LFM2 model by correcting the tensor name used during model loading. The change modifies line 6239 in llama_model::load_tensors(), replacing LLM_TENSOR_OUTPUT_NORM with LLM_TENSOR_OUTPUT_NORM_LFM2 for the output normalization layer.

Performance Impact

Power Consumption Analysis:
All 16 analyzed binaries show zero or negligible power consumption change:

• libllama.so: -0.0001% (-0.26 nJ)
• llama-tts: +0.0004% (+0.95 nJ)
• llama-cvector-generator: +0.0002% (+0.62 nJ)
• All other binaries: 0.0% change

Function-Level Metrics:
No measurable performance deltas detected across any functions. The summary report returned no data for Response Time or Throughput Time changes, indicating the versions are functionally identical from a performance perspective.

Inference Impact:
No impact on tokens per second. The change affects only model loading logic for LFM2 architecture, specifically the tensor name lookup during weight initialization. Functions responsible for inference (llama_decode, llama_encode, llama_tokenize) show no performance variation.

Technical Assessment

The modification is a correctness fix rather than a performance optimization. It ensures the LFM2 model loader references the architecture-specific output normalization tensor name, preventing "missing tensor" errors during model initialization. This change has no runtime performance implications as it only affects the one-time model loading phase, not the inference execution path.