glm-dsa: only rotate lid K when cache is quantized

The online Hadamard rotation (PR ggml-org#21038) exists to suppress quantization
error. On an F16 lid cache it is a no-op and costs one extra mul_mat per
layer. Remove the force-enable for GLM_DSA and DEEPSEEK2 and let the lid
cache follow the standard rule (quantized type + aligned head dim).
Guard the Hadamard mul_mats in the builder and the set_input_k_rot call
on self_k_rot_lid being non-null, since it can legitimately be null now.

Author: mmtmu

src/llama-graph.cpp

@@ -500,7 +500,10 @@ void llm_graph_input_attn_k_dsa::set_input(const llama_ubatch * ubatch) {
 
     mctx->get_lid()->set_input_k_idxs(self_k_idxs_lid, ubatch);
     mctx->get_lid()->set_input_kq_mask(self_kq_mask_lid, ubatch, cparams.causal_attn);
-    mctx->get_lid()->set_input_k_rot(self_k_rot_lid);
+
+    if (self_k_rot_lid) {
+        mctx->get_lid()->set_input_k_rot(self_k_rot_lid);
+    }
 }
 
 bool llm_graph_input_attn_k_dsa::can_reuse(const llm_graph_params & params) {

src/llama-kv-cache.cpp

@@ -281,12 +281,6 @@ llama_kv_cache::llama_kv_cache(
         !hparams.is_n_embd_k_gqa_variable() &&
         hparams.n_embd_head_k() % 64 == 0;
 
-    if ((model.arch == LLM_ARCH_GLM_DSA || model.arch == LLM_ARCH_DEEPSEEK2) &&
-        hparams.n_embd_head_k_full == hparams.indexer_head_size &&
-        hparams.indexer_head_size > 0) {
-        attn_rot_k = true;
-    }
-
     attn_rot_v =
         !attn_rot_disable &&
         ggml_is_quantized(type_v) &&

src/models/glm-dsa.cpp

@@ -91,8 +91,17 @@ llm_build_glm_dsa::llm_build_glm_dsa(const llama_model & model, const llm_graph_
             indexer_k = ggml_concat(ctx0, indexer_k_pe, indexer_k_nope, 0);
             cb(indexer_k, "indexer_k", il);
 
-            indexer_q = ggml_mul_mat(ctx0, inp_attn_dsa->self_k_rot_lid, indexer_q);
-            indexer_k = ggml_mul_mat(ctx0, inp_attn_dsa->self_k_rot_lid, indexer_k);
+            // Online Hadamard rotation (PR #21038) — only built when the lid cache is quantized.
+            // Note on direction: ggml_mul_mat(A, B) contracts on dim 0, i.e. computes Aᵀ·B,
+            // so these calls apply Hᵀ to q and k. That is safe because ggml_gen_hadamard
+            // returns a symmetric orthogonal matrix (H = Hᵀ, H² = I), so Hᵀx and Hx are the
+            // same tensor — the direction cannot mismatch with the cache's internal k-shift
+            // path. If the rotation is ever swapped for a non-symmetric orthogonal matrix,
+            // both sites must agree on H vs Hᵀ.
+            if (inp_attn_dsa->self_k_rot_lid) {
+                indexer_q = ggml_mul_mat(ctx0, inp_attn_dsa->self_k_rot_lid, indexer_q);
+                indexer_k = ggml_mul_mat(ctx0, inp_attn_dsa->self_k_rot_lid, indexer_k);
+            }
 
             const auto * mctx_lid = inp_attn_dsa->mctx->get_lid();
             const auto & k_idxs_lid = inp_attn_dsa->get_k_idxs_lid();