models : deduplicate delta-net graphs for Qwen family

[ggerganov]

cont #19375

• Add llm_build_delta_net_base for common delta net builds. Currently used only by qwen3next
• Rename llm_graph_context_mamba -> llm_build_mamba_base
• Includes clean-up

Next PRs:

• Use llm_build_delta_net_base for qwen35 and qwen35moe (after release)
• Adapt and use llm_build_delta_net_base for kimi_linear (cc @ymcki)

[pwilkin]

@ggerganov out of curiosity - what's the reason for moving the identity / mask constructions from once-per-graph to once-per-layer?

[ggerganov]

@ggerganov out of curiosity - what's the reason for moving the identity / mask constructions from once-per-graph to once-per-layer?

Mainly wanted to get rid of the ggml_new_tensor() calls. In this case, it does not cause any issues, but it can become problematic if the same tensor created with ggml_tensor_new() would be used for different ops that would be performed on different backends (for example due to offloading or missing support). As I mentioned, it's always better to find alternatives that don't involve creating new tensors.

Btw, creating a mask and then multiplying does not have advantage compare to just masking the data (with ggml_tri()), so it seems simpler too.

[ymcki]

Is this the new home for the unified delta_net effort? My job is to modify build_delta_net_chunking to use "if (is_kda) {} else {}" for kda and gdn?

[ggerganov]

Yes, we'll merge this PR soon and then the new llm_build_delta_net_base should be adapted to work with both GDA and KDA. Ideally, the KDA would have to add as few extra ops as possible - likely utilizing broadcasts when possible. The KDA support should be a standalone change, without any additional optimizations in order to keep things simple.

[ymcki]

@ggreganov, I am done with adding kimi linear to delta-net-base.cpp.

https://github.com/ymcki/llama.cpp/tree/dn

The original delta net path is the same except that I added o = ggml_cont(ctx0, o); after ggml_permute at the end.

This is to avoid the assertion error:
/home/user/gguf/gg/llama.cpp/ggml/src/ggml.c:3573: GGML_ASSERT(ggml_is_contiguous(a)) failed
raised by
ggml_tensor * attn_out_final = ggml_reshape_3d(ctx0, output, head_dim, n_head, n_seq_tokens * n_seqs);

Do I open another PR? Or you will do that yourself?

[ggerganov]

Please open a PR. I will attempt to simplify the implementation - hoping we can unify this and avoid branching on the g dimensions. Not sure if possible though - we'll see.