convert : enable expert group selection for all models with it

[CISC]

Enable expert group selection for all models with it (requires reconversion or metadata editing).

Specifically, these models (maybe more):

• DeepSeekV2.5 and onwards (DeepSeekV2.5 only used a single token instead of the sum of two, but the rest is the same)
• Dots1
• Glm4Moe

[CISC]

@jeffbolznv @am17an It would be great to have fusion for this as well. :)

llama.cpp/src/llama-graph.cpp

Lines 953 to 976 in 84bf3c6

	// select top n_group_used expert groups
	// https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/e815299b0bcbac849fa540c768ef21845365c9eb/modeling_deepseek.py#L440-L457
	if (hparams.n_expert_groups > 1 && n_tokens > 0) {
	const int64_t n_exp_per_group = n_expert / hparams.n_expert_groups;
	// organize experts into n_expert_groups
	ggml_tensor * selection_groups = ggml_reshape_3d(ctx0, selection_probs, n_exp_per_group, hparams.n_expert_groups, n_tokens); // [n_exp_per_group, n_expert_groups, n_tokens]
	ggml_tensor * group_scores = ggml_top_k(ctx0, selection_groups, 2); // [2, n_expert_groups, n_tokens]
	group_scores = ggml_get_rows(ctx0, ggml_reshape_4d(ctx0, selection_groups, 1, selection_groups->ne[0], selection_groups->ne[1], selection_groups->ne[2]), group_scores); // [1, 2, n_expert_groups, n_tokens]
	// get top n_group_used expert groups
	group_scores = ggml_sum_rows(ctx0, ggml_reshape_3d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2], group_scores->ne[3])); // [1, n_expert_groups, n_tokens]
	group_scores = ggml_reshape_2d(ctx0, group_scores, group_scores->ne[1], group_scores->ne[2]); // [n_expert_groups, n_tokens]
	ggml_tensor * expert_groups = ggml_top_k(ctx0, group_scores, hparams.n_group_used); // [n_group_used, n_tokens]
	cb(expert_groups, "ffn_moe_group_topk", il);
	// mask out the other groups
	selection_probs = ggml_get_rows(ctx0, selection_groups, expert_groups); // [n_exp_per_group, n_group_used, n_tokens]
	selection_probs = ggml_set_rows(ctx0, ggml_scale_bias(ctx0, selection_groups, 0.0f, -INFINITY), selection_probs, expert_groups); // [n_exp_per_group, n_expert_groups, n_tokens]
	selection_probs = ggml_reshape_2d(ctx0, selection_probs, n_expert, n_tokens); // [n_expert, n_tokens]
	cb(selection_probs, "ffn_moe_probs_masked", il);
	}

[jeffbolznv]

@jeffbolznv @am17an It would be great to have fusion for this as well. :)

Agreed. Let's catch up with the clamping change and get these common code utilities in, then I agree we should do this.