model : refactor QKV into common build_qkv and create_tensor_qkv helpers

[JoursBleu]

Overview

Currently llama.cpp supports 112 model files in src/models/.

We modified the 85 applicable model files. Our changes abstract the duplicated
Q/K/V tensors' loading and graph-building code into two reusable helpers,
following the create_tensor_gate_up_exps pattern (#19139).

• create_tensor_qkv (llama-model.cpp): tries fused wqkv/bqkv first (TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL), falls back to separate wq/wk/wv. Supports adding biases.

• build_qkv (llama-graph.h/cpp): returns {Qcur, Kcur, Vcur} as 3D tensors. Fused case: single fused qkv matmul + ggml_view_3d split. Separate case: 3 separate matmuls + ggml_reshape_3d.

Test: test-llama-archs — all OK, 0 FAIL. Zero diff on llama-arch.cpp.

The remaining 27 models are not modified for the following reasons:

Reason	Count	Models
Non-attention (SSM/linear/RNN)	10	mamba, mamba-base, rwkv6, rwkv6-base, rwkv6qwen2, rwkv7, rwkv7-base, arwkv7, delta-net-base, wavtokenizer-dec
MLA attention	4	deepseek2, minicpm3, minimax-m2, plm
Graph directly uses layer.wqkv (non-standard layout)	3	cogvlm, openelm, plamo2
Q+gate joint projection	4	qwen35, qwen35moe, qwen3next, plamo3
n_embd_head_k != n_embd_head_v	2	step35-iswa, mimo2-iswa
No fused wqkv_enc	1	t5-enc
Other special architectures	3	olmo2, olmoe, kimi-linear

Additional information

Basing on the discussion in #20628 (@am17an, @ngxson). The plan is:

1. This PR: This PR does not modify any logic, it simply extracts the redundant code into
   the two functions above, and adds handling for the fused qkv case.
2. Future PR: add --fuse-qkv to convert_hf_to_gguf.py.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - used as a translation tool for translating the PR description

[JoursBleu]

hi @CISC,

• Removed the has_bias flag.
• Bias tensors are now always created with TENSOR_NOT_REQUIRED.
• Fixed the incomplete conversions and typos mentioned above.

[JoursBleu]

@CISC Done：

• Remove unnecessary comments/restore comments that should be retained.
• JAIS2 restores manually created bias tensors.

[JoursBleu]

@CISC Done:

• Remove the Vcur reshape in afmoe.cpp

[CISC]

OP is inaccurate, there's nothing special about these:

• nemotron-h: just add build_qkv in llm_build_nemotron_h::build_attention_layer
• granite-hybrid: just add build_qkv in lm_build_granite_hybrid::build_attention_layer
• olmo/mpt/dbrx: use build_qkv, add clamping
• gemma3n-iswa: just do build_qkv
• t5-dec/t5-enc: do build_qkv on normal self-attention
• bert: use build_qkv
• lfm2: do build_qkv in build_attn_block

[CISC]

I meant move the clamping to build_qkv.

[JoursBleu]

@CISC Done:

• Extended build_qkv to bert, mpt, dbrx, olmo, lfm2, nemotron-h, granite-hybrid, gemma3n-iswa, t5-dec, t5-enc;
• Clamping handled internally in build_qkv using hparams.f_clamp_kqv.

[JoursBleu]

@ngxson @am17an @ggerganov This PR is ready. Could you take a look when you have time?

[am17an]

Good job!