llama: share the nextn tensor flags between models

[ServeurpersoCom]

Overview

Follow-up of the TODO in glm5-next: move the trunk-only and MTP-only detection that each model copied into a nextn_flags helper of llama_model_base. It probes the first trunk layer and the first NextN layer, and adds TENSOR_SKIP when MTP is not loaded. qwen4exp probes hc_attn_norm since it has no attn_norm.

deepseek4, nemotron-h, qwen35, qwen35moe, qwen3next and qwen4exp now also accept a trunk-only file, like the other models.

Additional information

Follow-up #29928

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES

[CISC]

Huh, GCC15 didn't like that:
https://github.com/ggml-org/llama.cpp/actions/runs/37614260360/job/112768687955

Edit: Guess you'll have to return {} directly.

[ServeurpersoCom]

Huh, GCC15 didn't like that: https://github.com/ggml-org/llama.cpp/actions/runs/37614260360/job/112768687955

Edit: Guess you'll have to return {} directly.

Ha, Thanks! I flagged this exact one on another PR and still walked right into it. Fixed: the lambdas in step35 and the qwen35 family captured the structured bindings, so the models now read the flags into plain variables, and the helper returns them directly.

[CISC]

Huh, GCC15 didn't like that: https://github.com/ggml-org/llama.cpp/actions/runs/37614260360/job/112768687955
Edit: Guess you'll have to return {} directly.

Ha, Thanks! I flagged this exact one on another PR and still walked right into it. Fixed: the lambdas in step35 and the qwen35 family captured the structured bindings, so the models now read the flags into plain variables, and the helper returns them directly.

Pretty sure you can still do const auto [trunk_flags, mtp_flags] = nextn_flags(ml);.

[ServeurpersoCom]

Pretty sure you can still do const auto [trunk_flags, mtp_flags] = nextn_flags(ml);.

The binding itself is what GCC 15 rejects once a lambda uses it: in step35 and the qwen35 family mtp_flags is read inside a [&] lambda, and the log points at that capture, not at the return. I kept plain variables in all 16 models so they stay uniform.

[CISC]

Pretty sure you can still do const auto [trunk_flags, mtp_flags] = nextn_flags(ml);.

The binding itself is what GCC 15 rejects once a lambda uses it: in step35 and the qwen35 family mtp_flags is read inside a [&] lambda, and the log points at that capture, not at the return. I kept plain variables in all 16 models so they stay uniform.

Ahhh, I see.

[CISC]

Looks like llama-install will be hogging the runners for the next 24 hours. :P

[ggerganov]

I think it's OK to merge - this is unlikely to affect the AMD-based workflows.