vulkan: Fix crash when FP16 mul_mat accumulation is not supported (#16796)

* Experimenting crash fix

* added assert for aborting and fixed comment

* changed to check if a pipeline is empty or not

* Moved function in class definition

* replaced with is_empty

* Modified is_empty to check only unaligned pipelines

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -677,7 +677,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id_map0(ggml_metal_
     char name[256];
 
     snprintf(base, 256, "kernel_mul_mm_id_map0_ne20_%d", ne20);
-    snprintf(name, 256, "%s", base);
+    snprintf(name, 256, "%s_ne02=%d", base, ne02);
 
     ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
     if (res) {

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -145,8 +145,13 @@ static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline);
 struct vk_matmul_pipeline_struct {
     vk_pipeline l, m, s;
     vk_pipeline a_l, a_m, a_s;
+    // Returns true when all unaligned pipelines are null.
+    // We only check for unaligned variants since one of the unaligned pipelines must exist
+    // while aligned pipelines are optional
+    bool is_empty() const {
+        return l == nullptr && m == nullptr && s == nullptr;
+    }
 };
-
 typedef std::shared_ptr<vk_matmul_pipeline_struct> vk_matmul_pipeline;
 
 struct vk_matmul_pipeline2 {
@@ -5079,7 +5084,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     if (src1_type == GGML_TYPE_Q8_1) {
         vk_matmul_pipeline pipelines = ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f32acc;
 
-        if (pipelines->s == nullptr && pipelines->m == nullptr && pipelines->l == nullptr) {
+        if (pipelines->is_empty()) {
             return nullptr;
         }
 
@@ -5228,7 +5233,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
     if (src1_type == GGML_TYPE_Q8_1) {
         vk_matmul_pipeline pipelines = ctx->device->pipeline_dequant_mul_mat_mat_id_q8_1[src0_type].f32acc;
 
-        if (pipelines->s == nullptr && pipelines->m == nullptr && pipelines->l == nullptr) {
+        if (pipelines->is_empty()) {
             return nullptr;
         }
 
@@ -5263,16 +5268,17 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
             return nullptr;
     }
 
+    vk_matmul_pipeline2& mmp = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type];
     // XXX TODO 'prec' is not actually allowed in mul_mat_id.
     bool prefer_fp16acc = ctx->device->fp16 /*&& prec == GGML_PREC_DEFAULT*/;
-    bool support_fp16acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc != nullptr;
-    bool support_fp32acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc != nullptr;
+    bool support_fp16acc = !mmp.f16acc->is_empty();
+    bool support_fp32acc = !mmp.f32acc->is_empty();
 
     if (support_fp16acc && (prefer_fp16acc || !support_fp32acc)) {
-        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc;
+        return mmp.f16acc;
     } else {
         GGML_ASSERT(support_fp32acc);
-        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc;
+        return mmp.f32acc;
     }
 }
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp

@@ -82,9 +82,13 @@ layout (constant_id = 10) const uint WARP = 32;
 
 #include "mul_mmq_shmem_types.glsl"
 
+#ifdef MUL_MAT_ID
+#define BK_STEP 1
+#else
 #ifndef BK_STEP
 #define BK_STEP 4
 #endif
+#endif
 
 // Shared memory cache
 shared block_a_cache buf_a[BM * BK_STEP];

ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl

@@ -27,7 +27,7 @@ struct block_a_cache {
 #elif defined(DATA_A_Q8_0)
 #define QUANT_R_MMQ 1
 // AMD likes 4, Intel likes 1 and Nvidia likes 2
-#define BK_STEP 1
+// #define BK_STEP 1
 struct block_a_cache {
     int32_t qs[32/4];
     FLOAT_TYPE dm;

tests/test-backend-ops.cpp

@@ -6880,6 +6880,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 1, 1, false, 8, 16, 1));
     test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, false, 32, 32, 32, 3));
 
+    // gpt-oss issue with Vulkan mmq_id
+    test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_MXFP4, GGML_TYPE_F32, 32, 2, false, 2880, 32, 2880));
+
     for (ggml_type type_a : base_types) {
         for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
             for (int n_mats : {4, 8}) {