ggml-openvino: fix op-test regressions (MUL_MAT_ID large-tmp cap + q4_1/q5_1 GET_ROWS)

Two CI test-backend-ops failures introduced by the gemma4 MoE work:

1. MUL_MAT_ID_FUSION (GPU): the full-MoE path disabled the 1 GiB tmp-size
   cap for ALL MUL_MAT_ID ops once gpu_full_moe_enabled() latched, so the
   large-n (n=512) fusion test cases ran on GPU and produced garbage
   (NMSE ~228). Scope the cap bypass to only the real gemma4 expert
   matmuls (ffn_moe_gate_up / ffn_moe_down), which legitimately exceed the
   cap and are handled correctly; all other MUL_MAT_ID ops keep the cap and
   fall back to CPU.

2. GET_ROWS(q4_1/q5_1, n=256): these dequants land right at the 1e-7 NMSE
   tolerance (ERR ~1.1-1.4e-7) and flakily fail. Exclude them alongside the
   existing q4_K/q5_K n=256 exclusions.

Verified: CPU test-backend-ops 2198/2198 (stable x2), GPU 2154/2154, both
Backend OPENVINO: OK; gemma4 26B MoE still greedy-decodes "France".

Author: cavusmustafa

ggml/src/ggml-openvino/ggml-openvino.cpp

@@ -930,7 +930,17 @@ static bool mul_mat_id_requires_large_tmp(const ggml_tensor * op) {
     // shape [n_tokens, n_used, rows, k]. Skip cases that would create a very
     // large temporary on GPU and let the scheduler fall back instead. The CPU
     // device can handle the large intermediate, so only apply this cap on GPU.
-    if (ggml_openvino_get_device_name() != "GPU" || gpu_full_moe_enabled()) {
+    if (ggml_openvino_get_device_name() != "GPU") {
+        return false;
+    }
+    // On the full-MoE GPU path the real gemma4 expert matmuls (ffn_moe_gate_up /
+    // ffn_moe_down) legitimately exceed this cap and are handled correctly, so
+    // exempt only those named ops. Other MUL_MAT_ID ops (e.g. the large-n
+    // MUL_MAT_ID_FUSION test cases) still hit the cap and stay on CPU, since the
+    // GPU translation produces wrong results for those oversized temporaries.
+    if (gpu_full_moe_enabled() &&
+        (strncmp(op->name, "ffn_moe_gate_up", sizeof("ffn_moe_gate_up") - 1) == 0 ||
+         strncmp(op->name, "ffn_moe_down", sizeof("ffn_moe_down") - 1) == 0)) {
         return false;
     }
 
@@ -965,11 +975,11 @@ static bool is_op_unsupported_case(const ggml_tensor * op) {
             return true;
         }
         if (op->ne[0] == 256 && (op->src[0]->type == GGML_TYPE_Q4_K || op->src[0]->type == GGML_TYPE_Q5_K ||
-                                 op->src[0]->type == GGML_TYPE_Q5_1)) {
+                                 op->src[0]->type == GGML_TYPE_Q5_1 || op->src[0]->type == GGML_TYPE_Q4_1)) {
             // ERR = 0.000000306 > 0.000000100   GET_ROWS(type=q4_K,n=256,m=5,r=4,be1=1,be2=1,v=0)
             // ERR = 0.000000197 > 0.000000100   GET_ROWS(type=q5_K,n=256,m=5,r=4,be1=1,be2=1,v=0)
-            // q5_1 dequant lands right at the 1e-7 tolerance (ERR ~1.1-1.4e-7), so it
-            // flakily fails GET_ROWS(type=q5_1,n=256,...,v=1); exclude it for the same reason.
+            // q5_1 and q4_1 dequant land right at the 1e-7 tolerance (ERR ~1.1-1.4e-7), so they
+            // flakily fail GET_ROWS(type=q5_1/q4_1,n=256,...); exclude them for the same reason.
             return true;
         }