CUDA/HIP: Fix kernel slection for mmvq mmid kernel to align host selection with devic...

The conditions cc == GGML_CUDA_CC_VOLTA || cc >= GGML_CUDA_CC_ADA_LOVELACE and cc >= GGML_CUDA_CC_TURING match all non-nvidia devices. This causes us to attempt to launch the kernel for batch sizes with larger configurations than our launch bounds on HIP devices. This pr fixes the conditionals in get_mmvq_mmid_max_batch.

Fixes #21191

diff --git a/src/ggml-cuda/mmvq.cu b/src/ggml-cuda/mmvq.cu
index 8d80d1dd9a78413afd9c9d510f095e874100036b..07b10167bc462813c513421d562612d0dd559974 100644 (file)
--- a/src/ggml-cuda/mmvq.cu
+++ b/src/ggml-cuda/mmvq.cu
@@ -235,30 +235,33 @@ static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna4(ggml_type
 // Host function: returns the max batch size for the current arch+type at runtime.
 int get_mmvq_mmid_max_batch(ggml_type type, int cc) {
     // NVIDIA: Volta, Ada Lovelace, and Blackwell always use MMVQ for MUL_MAT_ID.
-    if (cc == GGML_CUDA_CC_VOLTA || cc >= GGML_CUDA_CC_ADA_LOVELACE) {
-        return MMVQ_MAX_BATCH_SIZE;
-    }
-    if (cc >= GGML_CUDA_CC_TURING) {
-        return get_mmvq_mmid_max_batch_turing_plus(type);
-    }
     if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+        if (cc == GGML_CUDA_CC_VOLTA || cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+            return MMVQ_MAX_BATCH_SIZE;
+        }
+        if (cc >= GGML_CUDA_CC_TURING) {
+            return get_mmvq_mmid_max_batch_turing_plus(type);
+        }
         return get_mmvq_mmid_max_batch_pascal_older(type);
     }
+
     // AMD
-    if (GGML_CUDA_CC_IS_RDNA4(cc)) {
-        return get_mmvq_mmid_max_batch_rdna4(type);
-    }
-    if (GGML_CUDA_CC_IS_RDNA3(cc)) {
-        return get_mmvq_mmid_max_batch_rdna3(type);
-    }
-    if (GGML_CUDA_CC_IS_RDNA1(cc) || GGML_CUDA_CC_IS_RDNA2(cc)) {
-        return get_mmvq_mmid_max_batch_rdna1_rdna2(type);
-    }
-    if (GGML_CUDA_CC_IS_CDNA(cc)) {
-        return get_mmvq_mmid_max_batch_cdna(type);
-    }
-    if (GGML_CUDA_CC_IS_GCN(cc)) {
-        return get_mmvq_mmid_max_batch_gcn(type);
+    if (GGML_CUDA_CC_IS_AMD(cc)) {
+        if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+            return get_mmvq_mmid_max_batch_rdna4(type);
+        }
+        if (GGML_CUDA_CC_IS_RDNA3(cc)) {
+            return get_mmvq_mmid_max_batch_rdna3(type);
+        }
+        if (GGML_CUDA_CC_IS_RDNA1(cc) || GGML_CUDA_CC_IS_RDNA2(cc)) {
+            return get_mmvq_mmid_max_batch_rdna1_rdna2(type);
+        }
+        if (GGML_CUDA_CC_IS_CDNA(cc)) {
+            return get_mmvq_mmid_max_batch_cdna(type);
+        }
+        if (GGML_CUDA_CC_IS_GCN(cc)) {
+            return get_mmvq_mmid_max_batch_gcn(type);
+        }
     }
     return MMVQ_MAX_BATCH_SIZE;
 }