UPSTREAM PR #17898: model : Qwen3-Next-80B-A3B has 48 layers

[loci-dev]

Mirrored from ggml-org/llama.cpp#17898

Qwen3-Next-80B-A3B has 48 layers instead of 80, as pointed out by model README and a comment in original PR.

This change should be purely cosmetic, fixes "?B" model names shown by llama-bench, etc.

[loci-review]

Explore the complete analysis inside the Version Insights

Pull Request #506 Technical Review

PR Summary

Title: UPSTREAM PR #17898: model : Qwen3-Next-80B-A3B has 48 layers
Changes: Corrects layer count for Qwen3-Next-80B-A3B model from 80 to 48 layers and adds missing type name string mapping.

Code Changes Analysis

Modified File: src/llama-model.cpp

Change 1 - llm_type_name() function (line 123):

• Added case statement: case LLM_TYPE_80B_A3B: return "80B.A3B";
• Purpose: Provides string representation for the 80B_A3B model type enum
• Impact: Enables proper model name display in llama-bench and other tools

Change 2 - llama_model::load_hparams() function (line 2261):

• Modified layer count check: case 80: → case 48:
• Purpose: Corrects model architecture detection for Qwen3-Next-80B-A3B
• Impact: Ensures correct model type assignment during model loading based on actual layer count

Performance Impact Assessment

Function: llm_type_name()

• Base response time: 57 ns
• Current response time: 62 ns
• Absolute change: +5 ns
• Analysis: The addition of one case statement in the switch block adds minimal overhead. The 5 ns increase represents a single additional comparison in the switch dispatch logic.

Function: llama_model::load_hparams()

• This function is part of model loading, not inference path
• Changes affect model initialization only, executed once per model load
• No impact on per-token inference performance

Inference Performance:
No functions in the inference path (llama_decode, llama_encode, llama_tokenize) were modified. The changes are isolated to model metadata handling and initialization logic. Tokens per second remains unaffected.

Power Consumption:

• Binary: build.bin.libllama.so
• Change: +0.018% (+36 nJ)
• Analysis: Negligible increase consistent with one additional switch case

The changes are cosmetic corrections to model metadata with no measurable impact on inference performance or throughput.