@fugood/llama.node 0.3.15 → 0.3.16

package/src/llama.cpp/src/llama-model.cpp

@@ -271,19 +271,32 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
         }
     }
 
-    // add extra buffer types
-    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
-    auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
-        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
-    if (ggml_backend_dev_get_extra_bufts_fn) {
-        ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
-        while (extra_bufts && *extra_bufts) {
-            buft_list.emplace_back(cpu_dev, *extra_bufts);
-            ++extra_bufts;
+    bool has_gpu_device = false;
+    for (auto * dev : devices) {
+        if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
+            has_gpu_device = true;
+            break;
         }
     }
 
+    // add extra buffer types, only if no GPU device is present
+    // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094
+    if (!has_gpu_device) {
+        auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+        auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+            ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
+        if (ggml_backend_dev_get_extra_bufts_fn) {
+            ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
+            while (extra_bufts && *extra_bufts) {
+                buft_list.emplace_back(cpu_dev, *extra_bufts);
+                ++extra_bufts;
+            }
+        }
+    } else {
+        LLAMA_LOG_WARN("%s: disabling extra buffer types (i.e. repacking) since a GPU device is available\n", __func__);
+    }
+
     // add a host buffer type
     // storing the tensors in a host buffer is useful when the processing of large batches
     // is offloaded to a GPU device, since it reduces the time spent on data transfers
@@ -2210,9 +2223,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
 
                         // optional bias tensors
-                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, 0);
-                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, 0);
-                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
@@ -2329,7 +2342,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
                         layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias",   i), { n_embd }, 0);
 
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, TENSOR_NOT_REQUIRED);
                         if (layer.wqkv == nullptr) {
                             layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
                             layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias",   i), {n_embd}, 0);
@@ -2558,7 +2571,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                     // output
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,   "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
@@ -3215,16 +3233,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         auto & layer = layers[i];
 
                         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
                         if (layer.wqkv == nullptr) {
                             layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
                             layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
                             layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
-                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
-                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
                         }
 
                         layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
@@ -3335,12 +3353,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {time_mix_extra_dim, n_embd, 5}, 0);
 
                         layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0);
-                        layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, TENSOR_NOT_REQUIRED);
                         GGML_ASSERT(!(layer.time_mix_lerp_fused == NULL && layer.time_mix_lerp_w == NULL));
 
                         layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, 0);
@@ -3370,7 +3388,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
                     output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, TENSOR_NOT_REQUIRED);
                     output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
 
                     const int time_mix_extra_dim = hparams.time_mix_extra_dim;
@@ -3396,7 +3414,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0);
                         layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, 0);
 
-                        layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, TENSOR_NOT_REQUIRED);
                         layer.time_mix_decay = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY, "weight", i), {n_embd}, 0);
                         layer.time_mix_decay_w1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W1, "weight", i), {n_embd, time_decay_extra_dim}, 0);
                         layer.time_mix_decay_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W2, "weight", i), {time_decay_extra_dim, attn_hidden_size}, 0);
@@ -3405,9 +3423,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         layer.time_mix_gate = create_tensor(tn(LLM_TENSOR_TIME_MIX_GATE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         // optional bias tensors
-                        layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, TENSOR_NOT_REQUIRED);
 
                         layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0);
 
@@ -3528,8 +3546,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_value_res_mix, n_embd}, 0);
                         }
 
-                        layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, TENSOR_NOT_REQUIRED);
 
                         try {
                             layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 6}, 0);
@@ -3546,8 +3564,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.time_mix_value = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "weight", i), {attn_hidden_size, n_embd}, 0);
                         layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0);
 
-                        layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                        layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                        layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
                         layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
@@ -6193,16 +6211,25 @@ struct llm_build_qwen2moe : public llm_graph_context {
             {
                 // compute Q and K and RoPE them
                 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
                 cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
 
                 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
                 cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
 
                 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
                 cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
 
                 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
                 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);

package/src/llama.cpp/tests/test-backend-ops.cpp

@@ -1463,11 +1463,13 @@ struct test_cpy : public test_case {
     const ggml_type type_src;
     const ggml_type type_dst;
     const std::array<int64_t, 4> ne;
-    const std::array<int64_t, 4> permute;
+    const std::array<int64_t, 4> permute_src;
+    const std::array<int64_t, 4> permute_dst;
     bool _src_use_permute;
+    bool _dst_use_permute;
 
     std::string vars() override {
-        return VARS_TO_STR4(type_src, type_dst, ne, permute);
+        return VARS_TO_STR5(type_src, type_dst, ne, permute_src, permute_dst);
     }
 
     double max_nmse_err() override {
@@ -1480,9 +1482,11 @@ struct test_cpy : public test_case {
 
     test_cpy(ggml_type type_src = GGML_TYPE_F32, ggml_type type_dst = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {10, 10, 10, 1},
-            std::array<int64_t, 4> permute = {0, 0, 0, 0})
-        : type_src(type_src), type_dst(type_dst), ne(ne), permute(permute),
-          _src_use_permute(permute[0] + permute[1] + permute[2] + permute[3] > 0) {}
+            std::array<int64_t, 4> permute_src = {0, 0, 0, 0},
+            std::array<int64_t, 4> permute_dst = {0, 0, 0, 0})
+        : type_src(type_src), type_dst(type_dst), ne(ne), permute_src(permute_src), permute_dst(permute_dst),
+          _src_use_permute(permute_src[0] + permute_src[1] + permute_src[2] + permute_src[3] > 0),
+          _dst_use_permute(permute_dst[0] + permute_dst[1] + permute_dst[2] + permute_dst[3] > 0) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
@@ -1490,13 +1494,18 @@ struct test_cpy : public test_case {
         ggml_set_name(src, "src");
 
         if (_src_use_permute) {
-            src = ggml_permute(ctx, src, permute[0], permute[1], permute[2], permute[3]);
+            src = ggml_permute(ctx, src, permute_src[0], permute_src[1], permute_src[2], permute_src[3]);
             ggml_set_name(src, "src_permuted");
         }
 
-        ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, src->ne);
+        ggml_tensor * dst = ggml_new_tensor(ctx, type_dst, 4, src->ne);
         ggml_set_name(dst, "dst");
 
+        if (_dst_use_permute) {
+            dst = ggml_permute(ctx, dst, permute_dst[0], permute_dst[1], permute_dst[2], permute_dst[3]);
+            ggml_set_name(dst, "dst_permuted");
+        }
+
         ggml_tensor * out = ggml_cpy(ctx, src, dst);
         ggml_set_name(out, "out");
 
@@ -1964,9 +1973,10 @@ struct test_mul_mat : public test_case {
     const std::array<int64_t, 2> bs;  // dims 3 and 4
     const std::array<int64_t, 2> nr;  // repeat in dims 3 and 4
     const std::array<int64_t, 4> per; // permutation of dimensions
+    const bool v; // whether a is a non-contiguous view
 
     std::string vars() override {
-        return VARS_TO_STR8(type_a, type_b, m, n, k, bs, nr, per);
+        return VARS_TO_STR9(type_a, type_b, m, n, k, bs, nr, per, v);
     }
 
     double max_nmse_err() override {
@@ -1986,8 +1996,9 @@ struct test_mul_mat : public test_case {
             int64_t m = 32, int64_t n = 32, int64_t k = 32,
             std::array<int64_t, 2> bs = {10, 10},
             std::array<int64_t, 2> nr = {2, 2},
-            std::array<int64_t, 4> per = {0, 1, 2, 3})
-        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per) {}
+            std::array<int64_t, 4> per = {0, 1, 2, 3},
+            bool v = false)
+        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), v(v) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         // C^T = A * B^T: (k, m) * (k, n) => (m, n)
@@ -1997,6 +2008,7 @@ struct test_mul_mat : public test_case {
         const int npermuted = (per[0] != 0) + (per[1] != 1) + (per[2] != 2) + (per[3] != 3);
         if (npermuted > 0) {
             GGML_ASSERT(npermuted == 2);
+            GGML_ASSERT(!v); // not handled
             GGML_ASSERT(!ggml_is_quantized(type_a) || per[0] == 0);
             GGML_ASSERT(!ggml_is_quantized(type_b) || per[0] == 0);
 
@@ -2020,7 +2032,13 @@ struct test_mul_mat : public test_case {
             ggml_set_name(a, "a_permuted");
             ggml_set_name(b, "b_permuted");
         } else {
-            a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
+
+            if (v) {
+                a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0], bs[1]);
+                a = ggml_view_4d(ctx, a, k, m, bs[0], bs[1], a->nb[1], a->nb[2], a->nb[3], 0);
+            } else {
+                a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
+            }
             b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
             if (!ggml_is_quantized(type_a)) {
                 if (bs[1] == 1 && nr[1] == 1) {
@@ -3995,14 +4013,25 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_set(GGML_TYPE_I32, GGML_TYPE_I32, {6, 5, 4, 3}, dim));
     }
 
-    for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_F32}) {
+    // same-type copy
+    for (ggml_type type : all_types) {
+        const auto nk = ggml_blck_size(type);
+
+        for (int k = 1; k < 4; ++k) {
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}));
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}, {0, 2, 1, 3}));
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}, {0, 3, 1, 2}, {0, 2, 1, 3}));
+        }
+    }
+
+    for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_F32}) {
         for (ggml_type type_dst : all_types) {
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 4, 4, 4}));
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
         }
     }
-    for (ggml_type type_dst : {GGML_TYPE_F32}) {
-        for (ggml_type type_src : all_types) {
+    for (ggml_type type_src : all_types) {
+        for (ggml_type type_dst : {GGML_TYPE_F32}) {
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 4, 4, 4}));
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
         }
@@ -4176,6 +4205,17 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 45, 128, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45,  64, { 8,  1}, {4, 1}));
 
+    for (auto bs : {1,2,4,8}) {
+        for (auto nr : {1,4}) {
+            for (uint32_t m = 0; m < 2; ++m) {
+                for (uint32_t k = 0; k < 2; ++k) {
+                    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056 + m, 1, 128 + k,  {bs,  1}, {nr, 1}, {0, 2, 1, 3}));
+                    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128 + m,  1, 1056 + k, {bs,  1}, {nr, 1}, {0, 1, 2, 3}, true));
+                }
+            }
+        }
+    }
+
     // sycl backend will limit task global_range < MAX_INT
     // test case for f16-type-convert-to-fp32 kernel with large k under fp32 compute dtype (occurs in stable-diffusion)
     // however this case needs to alloc more memory which may fail in some devices (Intel Arc770, etc.)
@@ -4444,6 +4484,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32000, 512, 1, 1}));
 
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 16416, 1, 128, {8,  1}, {4, 1}, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 1, 16416, {8,  1}, {4, 1}, {0, 1, 2, 3}, true));
+
     for (int bs : {1, 2, 3, 4, 5, 8, 512}) {
         for (ggml_type type_a : all_types) {
             for (ggml_type type_b : {GGML_TYPE_F32}) {