sycl: fuse mul_mat(gate) + mul_mat(up) + GLU for q4_K dense FFN (#26779)

Measured on Arc Pro B70 (Battlemage, Level Zero), llama-bench -r 20, two
interleaved rounds, tg128:

    qwen2.5-3B-Instruct Q4_K_M    154.18 -> 158.53 t/s   +2.8%
    gemma-2-2b-it Q4_K_M          162.45 -> 165.62 t/s   +2.0%

llama-batched-bench on qwen2.5-3B, S_TG by batch size:

      B=1   142.72 -> 147.57 t/s    +3.4%
      B=2   243.72 -> 268.26 t/s   +10.1%
      B=4   359.58 -> 398.02 t/s   +10.7%
      B=8   449.75 -> 505.63 t/s   +12.4%

ggml/src/ggml-sycl/fusion.cpp

@@ -2,6 +2,61 @@
 
 #include <algorithm>
 
+// mul_mat(gate) + mul_mat(up) + GLU: graph shape and tensor properties only. Backend state
+// (weight layout, split buffers, DMMV) is checked by ggml_sycl_mul_mat_glu_mmvq_fused().
+static bool ggml_sycl_should_fuse_mul_mat_glu(const ggml_tensor * gate, const ggml_tensor * up,
+                                              const ggml_tensor * glu) {
+    // the fused epilogue implements these two; the rest fall back to the standalone GLU kernels
+    const ggml_glu_op glu_op = ggml_get_glu_op(glu);
+    if (glu_op != GGML_GLU_OP_SWIGLU && glu_op != GGML_GLU_OP_GEGLU) {
+        return false;
+    }
+
+    // the kernel always treats src[0] as the activated operand and src[1] as the multiplier
+    if (ggml_get_op_params_i32(glu, 1) /* swapped */) {
+        return false;
+    }
+
+    const ggml_tensor * wu  = up->src[0];
+    const ggml_tensor * wg  = gate->src[0];
+    const ggml_tensor * act = up->src[1];
+
+    // one set of block offsets and one quantized activation must serve both weights
+    if (wu->type != wg->type || !ggml_are_same_shape(wu, wg) || !ggml_are_same_stride(wu, wg)) {
+        return false;
+    }
+    if (act != gate->src[1]) {
+        return false;
+    }
+
+    // only q4_K has a fused reorder GEMV so far, and it walks whole super-blocks
+    if (wu->type != GGML_TYPE_Q4_K || wu->ne[0] % QK_K != 0) {
+        return false;
+    }
+
+    // one 2D reorder-layout matrix in, a plain column stride out: no broadcast or padding
+    if (!ggml_is_contiguous(wu) || !ggml_is_contiguous(wg) || !ggml_is_contiguous(act) ||
+        !ggml_is_contiguous(glu)) {
+        return false;
+    }
+    if (act->type != GGML_TYPE_F32 || glu->type != GGML_TYPE_F32) {
+        return false;
+    }
+    if (act->ne[2] != 1 || act->ne[3] != 1 || wu->ne[2] != 1 || wu->ne[3] != 1) {
+        return false;
+    }
+    // the kernel writes rows [0, wu->ne[1]) of each glu column, strided by glu->ne[0]
+    if (glu->ne[0] != wu->ne[1] || glu->ne[1] != act->ne[1]) {
+        return false;
+    }
+    // mat-vec only: one column per decoded token, up to the batch the reorder kernels cover
+    if (act->ne[1] > MMVQ_MAX_BATCH_SIZE) {
+        return false;
+    }
+
+    return true;
+}
+
 bool ggml_sycl_can_fuse(const ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops,
                         std::initializer_list<enum ggml_unary_op> unary_ops) {
 #ifndef NDEBUG
@@ -13,6 +68,28 @@ bool ggml_sycl_can_fuse(const ggml_cgraph * cgraph, int node_idx, std::initializ
         return false;
     }
 
+    // gate and up are siblings, not a chain, so ggml_can_fuse cannot express this: use the
+    // subgraph form with the GLU as the only materialised output.
+    if (ops.size() == 3 && ops.begin()[0] == GGML_OP_MUL_MAT && ops.begin()[1] == GGML_OP_MUL_MAT &&
+        ops.begin()[2] == GGML_OP_GLU) {
+        if (!ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) {
+            return false;
+        }
+
+        const ggml_tensor * glu  = cgraph->nodes[node_idx + 2];
+        const ggml_tensor * gate = glu->src[0];
+        const ggml_tensor * up   = glu->src[1];
+
+        // don't assume which of the two mat-muls is the gate; infer it from the GLU's operands
+        const bool ok = (gate == cgraph->nodes[node_idx] && up == cgraph->nodes[node_idx + 1]) ||
+                        (gate == cgraph->nodes[node_idx + 1] && up == cgraph->nodes[node_idx]);
+        if (!ok) {
+            return false;
+        }
+
+        return ggml_sycl_should_fuse_mul_mat_glu(gate, up, glu);
+    }
+
     if (!ggml_can_fuse(cgraph, node_idx, ops)) {
         return false;
     }