glm-dsa : fix missing indexer tensors on sparse-indexer blocks

GLM-5.2 only has DSA lightning indexer tensors on a subset of blocks:
the 3 dense-lead blocks (0-2) and every 4th MoE block (6, 10, 14, ..., 78).
The remaining MoE blocks have no indexer tensors at all.

glm-dsa.cpp required them on every block, so loading failed with:
  "missing tensor 'blk.3.indexer.k_norm.weight'"

Two changes:
- glm-dsa.cpp: mark all 5 indexer tensor creates as TENSOR_NOT_REQUIRED
- deepseek32.cpp: guard the lightning indexer block with a null check on
  indexer_attn_q_b; when absent, top_k stays nullptr and build_attn_mla
  falls back to full attention for that layer

Regressed in ggml-org#23346 (2026-05-29) which wired the DSA lightning indexer
into the shared deepseek2 graph used by GLM_DSA. Before that PR the
indexer tensors were loaded but never accessed in the graph.

Assisted-by: Claude Sonnet

Author: csabakecskemeti

src/models/deepseek32.cpp

@@ -219,8 +219,8 @@ llama_model_deepseek32::graph::graph(const llama_model & model, const llm_graph_
 
             ggml_tensor * top_k = nullptr;
 
-            // lightning indexer
-            {
+            // lightning indexer - skipped when tensors are absent (e.g. GLM-5.2 only has them on certain blocks)
+            if (model.layers[il].indexer_attn_q_b) {
                 ggml_tensor * indexer_q = ggml_mul_mat(ctx0, model.layers[il].indexer_attn_q_b, qr);
                 cb(indexer_q, "indexer_q", il);
 

src/models/glm-dsa.cpp

@@ -100,12 +100,13 @@ void llama_model_glm_dsa::load_arch_tensors(llama_model_loader &) {
 
         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
 
-        // DSA indexer
-        layer.indexer_k_norm   = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "weight", i), {hparams.indexer_head_size}, flags);
-        layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "bias",   i), {hparams.indexer_head_size}, flags);
-        layer.indexer_proj     = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ,     "weight", i), {n_embd, hparams.indexer_n_head}, flags);
-        layer.indexer_attn_k   = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K,   "weight", i), {n_embd, hparams.indexer_head_size}, flags);
-        layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags);
+        // DSA indexer - only present on a subset of blocks (dense-lead blocks + every 4th MoE block),
+        // so mark all as TENSOR_NOT_REQUIRED; the graph skips the indexer path when these are null.
+        layer.indexer_k_norm   = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "weight", i), {hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
+        layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "bias",   i), {hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
+        layer.indexer_proj     = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ,     "weight", i), {n_embd, hparams.indexer_n_head}, flags | TENSOR_NOT_REQUIRED);
+        layer.indexer_attn_k   = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K,   "weight", i), {n_embd, hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
+        layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags | TENSOR_NOT_REQUIRED);
         if (i < (int) hparams.n_layer_dense_lead) {
             layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);
             layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);