diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h
index 724ac84f1d..74b886b054 100644
--- a/ggml/src/ggml-cpu/arch-fallback.h
+++ b/ggml/src/ggml-cpu/arch-fallback.h
@@ -72,6 +72,9 @@
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
+// quants.c
+#define ggml_vec_dot_tbq3_0_q8_K_generic ggml_vec_dot_tbq3_0_q8_K
+#define ggml_vec_dot_tbq4_0_q8_K_generic ggml_vec_dot_tbq4_0_q8_K
 // repack.cpp
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
@@ -207,6 +210,8 @@
 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
 #elif defined(__riscv)
 // quants.c
+#define ggml_vec_dot_tbq3_0_q8_K_generic ggml_vec_dot_tbq3_0_q8_K
+#define ggml_vec_dot_tbq4_0_q8_K_generic ggml_vec_dot_tbq4_0_q8_K
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x1_generic ggml_quantize_mat_q8_0_4x1
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 6efdcb36b3..8933acf800 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -11,6 +11,7 @@
 #include <algorithm>
 #include <cfloat>
 #include <cmath>
+#include <type_traits>
 #include <vector>
 
 // ggml_compute_forward_dup
@@ -529,8 +530,6 @@ static void ggml_compute_forward_dup_from_q(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    std::vector<float> tmp(qk);
-
     for (int64_t ir = ir0; ir < ir1; ++ir) {
 
         uint32_t i = ir * qk;
@@ -547,11 +546,19 @@ static void ggml_compute_forward_dup_from_q(
         const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
         const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
 
-        dequantize_row_q(
-                (const void *) ((char *) src0->data + x_offset),
-                tmp.data(), qk);
+        if constexpr (std::is_same_v<dst_t, float>) {
+            dequantize_row_q(
+                    (const void *) ((char *) src0->data + x_offset),
+                    (float *) ((char *) dst->data + dst_offset), qk);
+        } else {
+            std::vector<float> tmp(qk);
 
-        ggml_dup_from_float_row(tmp.data(), (dst_t *) ((char *) dst->data + dst_offset), qk);
+            dequantize_row_q(
+                    (const void *) ((char *) src0->data + x_offset),
+                    tmp.data(), qk);
+
+            ggml_dup_from_float_row(tmp.data(), (dst_t *) ((char *) dst->data + dst_offset), qk);
+        }
     }
 }
 
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index f6ce2817a8..c14fb98a5d 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -2944,10 +2944,15 @@ llama_context * llama_init_from_model(
 
     if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO && ggml_is_quantized(params.type_k)) {
         const uint32_t blck_size = ggml_blck_size(params.type_k);
+        const bool is_tbq_k = params.type_k == GGML_TYPE_TBQ3_0 || params.type_k == GGML_TYPE_TBQ4_0;
+
         for (uint32_t il = 0; il < model->hparams.n_layer; ++il) {
-            if (model->hparams.n_embd_head_k(il) % blck_size != 0) {
-                LLAMA_LOG_ERROR("%s: K cache type %s with block size %u does not divide n_embd_head_k=%u\n",
-                    __func__, ggml_type_name(params.type_k), blck_size, model->hparams.n_embd_head_k(il));
+            const uint32_t n_embd_k = is_tbq_k ? model->hparams.n_embd_k_gqa(il) : model->hparams.n_embd_head_k(il);
+
+            if (n_embd_k % blck_size != 0) {
+                LLAMA_LOG_ERROR("%s: K cache type %s with block size %u does not divide %s=%u\n",
+                    __func__, ggml_type_name(params.type_k), blck_size,
+                    is_tbq_k ? "n_embd_k_gqa" : "n_embd_head_k", n_embd_k);
                 return nullptr;
             }
         }
@@ -2955,10 +2960,15 @@ llama_context * llama_init_from_model(
 
     if (params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO && ggml_is_quantized(params.type_v)) {
         const uint32_t blck_size = ggml_blck_size(params.type_v);
+        const bool is_tbq_v = params.type_v == GGML_TYPE_TBQ3_0 || params.type_v == GGML_TYPE_TBQ4_0;
+
         for (uint32_t il = 0; il < model->hparams.n_layer; ++il) {
-            if (model->hparams.n_embd_head_v(il) % blck_size != 0) {
-                LLAMA_LOG_ERROR("%s: V cache type %s with block size %u does not divide n_embd_head_v=%u\n",
-                    __func__, ggml_type_name(params.type_v), blck_size, model->hparams.n_embd_head_v(il));
+            const uint32_t n_embd_v = is_tbq_v ? model->hparams.n_embd_v_gqa(il) : model->hparams.n_embd_head_v(il);
+
+            if (n_embd_v % blck_size != 0) {
+                LLAMA_LOG_ERROR("%s: V cache type %s with block size %u does not divide %s=%u\n",
+                    __func__, ggml_type_name(params.type_v), blck_size,
+                    is_tbq_v ? "n_embd_v_gqa" : "n_embd_head_v", n_embd_v);
                 return nullptr;
             }
         }
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index f8ce1cf314..75e2ae22d2 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -7804,15 +7804,29 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         }
     }
     for (ggml_type type_src : all_types) {
-        for (ggml_type type_dst : {GGML_TYPE_F32}) {
+        if (!ggml_is_quantized(type_src)) {
+            continue;
+        }
+        test_cases.emplace_back(new test_cpy(type_src, GGML_TYPE_F32, {256, 4, 4, 4}));
+        test_cases.emplace_back(new test_cpy(type_src, GGML_TYPE_F32, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
+    }
+    for (ggml_type type : turboq_types) {
+        test_cases.emplace_back(new test_cpy(type, GGML_TYPE_F32, {256, 4, 4, 4}));
+        test_cases.emplace_back(new test_cpy(type, GGML_TYPE_F32, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
+    }
+    for (ggml_type type_src : all_types) {
+        if (!ggml_is_quantized(type_src)) {
+            continue;
+        }
+        for (ggml_type type_dst : {GGML_TYPE_F16, GGML_TYPE_BF16}) {
             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_src : all_types) {
+    for (ggml_type type : turboq_types) {
         for (ggml_type type_dst : {GGML_TYPE_F16, GGML_TYPE_BF16}) {
-            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
+            test_cases.emplace_back(new test_cpy(type, type_dst, {256, 4, 4, 4}));
+            test_cases.emplace_back(new test_cpy(type, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
         }
     }
     for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_F32}) {