graph : ensure DS32 kq_mask_lid is F32

[CISC]

Overview

cont #23346
cont #23764

Additional information

Since build_attn_inp_kq_mask returns F16 mask when flash attention is enabled, pass a modified copy of cparams for kq_mask_lid.

llama.cpp/src/models/deepseek32.cpp

Lines 341 to 344 in 1f0aa2a

	// mask indexer scores
	ggml_tensor * indexer_kq_mask = inp_attn_dsa->get_kq_mask_lid();
	indexer_score = ggml_add(ctx0, indexer_score, indexer_kq_mask);
	cb(indexer_score, "indexer_score", il);

This is a bit hacky, open for better solutions. cc/ @am17an

Requirements

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

[am17an]

Does this mask need to be f32?

[CISC]

Does this mask need to be f32?

Either that or we have to cast indexer_score to F16.

[fairydreaming]

So... I checked how DeepSeek V3.2 works in master (a couple of hours too late) and ended up here. But this PR helps, ggml_cuda_op_add error is gone.