whisper.rn 0.4.0-rc.4 → 0.4.0-rc.6

package/cpp/ggml.c

@@ -1,4 +1,4 @@
-#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
+#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnings on Windows
 #define _USE_MATH_DEFINES // For M_PI on MSVC
 
 #include "ggml-impl.h"
@@ -33,7 +33,7 @@
 // we should just be careful :)
 #pragma warning(disable: 4244 4267)
 
-// disable POSIX deprecation warnigns
+// disable POSIX deprecation warnings
 // these functions are never going away, anyway
 #pragma warning(disable: 4996)
 #endif
@@ -233,24 +233,6 @@ inline static void * wsp_ggml_aligned_malloc(size_t size) {
 #define UNUSED WSP_GGML_UNUSED
 #define SWAP(x, y, T) do { T SWAP = x; x = y; y = SWAP; } while (0)
 
-//
-// tensor access macros
-//
-
-#define WSP_GGML_TENSOR_UNARY_OP_LOCALS \
-    WSP_GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
-    WSP_GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
-    WSP_GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
-    WSP_GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
-
-#define WSP_GGML_TENSOR_BINARY_OP_LOCALS \
-    WSP_GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
-    WSP_GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
-    WSP_GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
-    WSP_GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb) \
-    WSP_GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
-    WSP_GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
-
 #if defined(WSP_GGML_USE_ACCELERATE)
 #include <Accelerate/Accelerate.h>
 #if defined(WSP_GGML_USE_CLBLAST) // allow usage of CLBlast alongside Accelerate functions
@@ -455,9 +437,9 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK4_0,
         .type_size                = sizeof(block_q4_0),
         .is_quantized             = true,
-        .to_float                 = (wsp_ggml_to_float_t) dequantize_row_q4_0,
-        .from_float               = quantize_row_q4_0,
-        .from_float_reference     = (wsp_ggml_from_float_t) quantize_row_q4_0_reference,
+        .to_float                 = (wsp_ggml_to_float_t) wsp_dewsp_quantize_row_q4_0,
+        .from_float               = wsp_quantize_row_q4_0,
+        .from_float_reference     = (wsp_ggml_from_float_t) wsp_quantize_row_q4_0_reference,
         .vec_dot                  = wsp_ggml_vec_dot_q4_0_q8_0,
         .vec_dot_type             = WSP_GGML_TYPE_Q8_0,
     },
@@ -466,9 +448,9 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK4_1,
         .type_size                = sizeof(block_q4_1),
         .is_quantized             = true,
-        .to_float                 = (wsp_ggml_to_float_t) dequantize_row_q4_1,
-        .from_float               = quantize_row_q4_1,
-        .from_float_reference     = (wsp_ggml_from_float_t) quantize_row_q4_1_reference,
+        .to_float                 = (wsp_ggml_to_float_t) wsp_dewsp_quantize_row_q4_1,
+        .from_float               = wsp_quantize_row_q4_1,
+        .from_float_reference     = (wsp_ggml_from_float_t) wsp_quantize_row_q4_1_reference,
         .vec_dot                  = wsp_ggml_vec_dot_q4_1_q8_1,
         .vec_dot_type             = WSP_GGML_TYPE_Q8_1,
     },
@@ -499,9 +481,9 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK5_0,
         .type_size                = sizeof(block_q5_0),
         .is_quantized             = true,
-        .to_float                 = (wsp_ggml_to_float_t) dequantize_row_q5_0,
-        .from_float               = quantize_row_q5_0,
-        .from_float_reference     = (wsp_ggml_from_float_t) quantize_row_q5_0_reference,
+        .to_float                 = (wsp_ggml_to_float_t) wsp_dewsp_quantize_row_q5_0,
+        .from_float               = wsp_quantize_row_q5_0,
+        .from_float_reference     = (wsp_ggml_from_float_t) wsp_quantize_row_q5_0_reference,
         .vec_dot                  = wsp_ggml_vec_dot_q5_0_q8_0,
         .vec_dot_type             = WSP_GGML_TYPE_Q8_0,
     },
@@ -510,9 +492,9 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK5_1,
         .type_size                = sizeof(block_q5_1),
         .is_quantized             = true,
-        .to_float                 = (wsp_ggml_to_float_t) dequantize_row_q5_1,
-        .from_float               = quantize_row_q5_1,
-        .from_float_reference     = (wsp_ggml_from_float_t) quantize_row_q5_1_reference,
+        .to_float                 = (wsp_ggml_to_float_t) wsp_dewsp_quantize_row_q5_1,
+        .from_float               = wsp_quantize_row_q5_1,
+        .from_float_reference     = (wsp_ggml_from_float_t) wsp_quantize_row_q5_1_reference,
         .vec_dot                  = wsp_ggml_vec_dot_q5_1_q8_1,
         .vec_dot_type             = WSP_GGML_TYPE_Q8_1,
     },
@@ -521,9 +503,9 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK8_0,
         .type_size                = sizeof(block_q8_0),
         .is_quantized             = true,
-        .to_float                 = (wsp_ggml_to_float_t) dequantize_row_q8_0,
-        .from_float               = quantize_row_q8_0,
-        .from_float_reference     = (wsp_ggml_from_float_t) quantize_row_q8_0_reference,
+        .to_float                 = (wsp_ggml_to_float_t) wsp_dewsp_quantize_row_q8_0,
+        .from_float               = wsp_quantize_row_q8_0,
+        .from_float_reference     = (wsp_ggml_from_float_t) wsp_quantize_row_q8_0_reference,
         .vec_dot                  = wsp_ggml_vec_dot_q8_0_q8_0,
         .vec_dot_type             = WSP_GGML_TYPE_Q8_0,
     },
@@ -532,8 +514,8 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK8_1,
         .type_size                = sizeof(block_q8_1),
         .is_quantized             = true,
-        .from_float               = quantize_row_q8_1,
-        .from_float_reference     = (wsp_ggml_from_float_t) quantize_row_q8_1_reference,
+        .from_float               = wsp_quantize_row_q8_1,
+        .from_float_reference     = (wsp_ggml_from_float_t) wsp_quantize_row_q8_1_reference,
         .vec_dot_type             = WSP_GGML_TYPE_Q8_1,
     },
     [WSP_GGML_TYPE_Q2_K] = {
@@ -541,9 +523,9 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK_K,
         .type_size                = sizeof(block_q2_K),
         .is_quantized             = true,
-        .to_float                 = (wsp_ggml_to_float_t) dequantize_row_q2_K,
-        .from_float               = quantize_row_q2_K,
-        .from_float_reference     = (wsp_ggml_from_float_t) quantize_row_q2_K_reference,
+        .to_float                 = (wsp_ggml_to_float_t) wsp_dewsp_quantize_row_q2_K,
+        .from_float               = wsp_quantize_row_q2_K,
+        .from_float_reference     = (wsp_ggml_from_float_t) wsp_quantize_row_q2_K_reference,
         .vec_dot                  = wsp_ggml_vec_dot_q2_K_q8_K,
         .vec_dot_type             = WSP_GGML_TYPE_Q8_K,
     },
@@ -552,9 +534,9 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK_K,
         .type_size                = sizeof(block_q3_K),
         .is_quantized             = true,
-        .to_float                 = (wsp_ggml_to_float_t) dequantize_row_q3_K,
-        .from_float               = quantize_row_q3_K,
-        .from_float_reference     = (wsp_ggml_from_float_t) quantize_row_q3_K_reference,
+        .to_float                 = (wsp_ggml_to_float_t) wsp_dewsp_quantize_row_q3_K,
+        .from_float               = wsp_quantize_row_q3_K,
+        .from_float_reference     = (wsp_ggml_from_float_t) wsp_quantize_row_q3_K_reference,
         .vec_dot                  = wsp_ggml_vec_dot_q3_K_q8_K,
         .vec_dot_type             = WSP_GGML_TYPE_Q8_K,
     },
@@ -563,9 +545,9 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK_K,
         .type_size                = sizeof(block_q4_K),
         .is_quantized             = true,
-        .to_float                 = (wsp_ggml_to_float_t) dequantize_row_q4_K,
-        .from_float               = quantize_row_q4_K,
-        .from_float_reference     = (wsp_ggml_from_float_t) quantize_row_q4_K_reference,
+        .to_float                 = (wsp_ggml_to_float_t) wsp_dewsp_quantize_row_q4_K,
+        .from_float               = wsp_quantize_row_q4_K,
+        .from_float_reference     = (wsp_ggml_from_float_t) wsp_quantize_row_q4_K_reference,
         .vec_dot                  = wsp_ggml_vec_dot_q4_K_q8_K,
         .vec_dot_type             = WSP_GGML_TYPE_Q8_K,
     },
@@ -574,9 +556,9 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK_K,
         .type_size                = sizeof(block_q5_K),
         .is_quantized             = true,
-        .to_float                 = (wsp_ggml_to_float_t) dequantize_row_q5_K,
-        .from_float               = quantize_row_q5_K,
-        .from_float_reference     = (wsp_ggml_from_float_t) quantize_row_q5_K_reference,
+        .to_float                 = (wsp_ggml_to_float_t) wsp_dewsp_quantize_row_q5_K,
+        .from_float               = wsp_quantize_row_q5_K,
+        .from_float_reference     = (wsp_ggml_from_float_t) wsp_quantize_row_q5_K_reference,
         .vec_dot                  = wsp_ggml_vec_dot_q5_K_q8_K,
         .vec_dot_type             = WSP_GGML_TYPE_Q8_K,
     },
@@ -585,9 +567,9 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK_K,
         .type_size                = sizeof(block_q6_K),
         .is_quantized             = true,
-        .to_float                 = (wsp_ggml_to_float_t) dequantize_row_q6_K,
-        .from_float               = quantize_row_q6_K,
-        .from_float_reference     = (wsp_ggml_from_float_t) quantize_row_q6_K_reference,
+        .to_float                 = (wsp_ggml_to_float_t) wsp_dewsp_quantize_row_q6_K,
+        .from_float               = wsp_quantize_row_q6_K,
+        .from_float_reference     = (wsp_ggml_from_float_t) wsp_quantize_row_q6_K_reference,
         .vec_dot                  = wsp_ggml_vec_dot_q6_K_q8_K,
         .vec_dot_type             = WSP_GGML_TYPE_Q8_K,
     },
@@ -596,7 +578,7 @@ static const wsp_ggml_type_traits_t type_traits[WSP_GGML_TYPE_COUNT] = {
         .blck_size                = QK_K,
         .type_size                = sizeof(block_q8_K),
         .is_quantized             = true,
-        .from_float               = quantize_row_q8_K,
+        .from_float               = wsp_quantize_row_q8_K,
     }
 };
 
@@ -1413,7 +1395,7 @@ inline static void wsp_ggml_vec_step_f32 (const int n, float * y, const float *
 inline static void wsp_ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]);  }
 inline static void wsp_ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; }
 inline static void wsp_ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
-inline static void wsp_ggml_vec_leaky_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.1f*x[i]; }
+inline static void wsp_ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
 
 static const float GELU_COEF_A     = 0.044715f;
 static const float GELU_QUICK_COEF = -1.702f;
@@ -1613,6 +1595,7 @@ static const char * WSP_GGML_OP_NAME[WSP_GGML_OP_COUNT] = {
     "GROUP_NORM",
 
     "MUL_MAT",
+    "MUL_MAT_ID",
     "OUT_PROD",
 
     "SCALE",
@@ -1634,17 +1617,15 @@ static const char * WSP_GGML_OP_NAME[WSP_GGML_OP_COUNT] = {
     "ROPE_BACK",
     "ALIBI",
     "CLAMP",
-    "CONV_1D",
-    "CONV_1D_STAGE_0",
-    "CONV_1D_STAGE_1",
     "CONV_TRANSPOSE_1D",
-    "CONV_2D",
-    "CONV_2D_STAGE_0",
-    "CONV_2D_STAGE_1",
+    "IM2COL",
     "CONV_TRANSPOSE_2D",
     "POOL_1D",
     "POOL_2D",
     "UPSCALE",
+    "PAD",
+    "ARGSORT",
+    "LEAKY_RELU",
 
     "FLASH_ATTN",
     "FLASH_FF",
@@ -1671,7 +1652,7 @@ static const char * WSP_GGML_OP_NAME[WSP_GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS_BACK",
 };
 
-static_assert(WSP_GGML_OP_COUNT == 73, "WSP_GGML_OP_COUNT != 73");
+static_assert(WSP_GGML_OP_COUNT == 72, "WSP_GGML_OP_COUNT != 72");
 
 static const char * WSP_GGML_OP_SYMBOL[WSP_GGML_OP_COUNT] = {
     "none",
@@ -1700,6 +1681,7 @@ static const char * WSP_GGML_OP_SYMBOL[WSP_GGML_OP_COUNT] = {
     "group_norm(x)",
 
     "X*Y",
+    "X[i]*Y",
     "X*Y",
 
     "x*v",
@@ -1721,17 +1703,15 @@ static const char * WSP_GGML_OP_SYMBOL[WSP_GGML_OP_COUNT] = {
     "rope_back(x)",
     "alibi(x)",
     "clamp(x)",
-    "conv_1d(x)",
-    "conv_1d_stage_0(x)",
-    "conv_1d_stage_1(x)",
     "conv_transpose_1d(x)",
-    "conv_2d(x)",
-    "conv_2d_stage_0(x)",
-    "conv_2d_stage_1(x)",
+    "im2col(x)",
     "conv_transpose_2d(x)",
     "pool_1d(x)",
     "pool_2d(x)",
     "upscale(x)",
+    "pad(x)",
+    "argsort(x)",
+    "leaky_relu(x)",
 
     "flash_attn(x)",
     "flash_ff(x)",
@@ -1758,15 +1738,32 @@ static const char * WSP_GGML_OP_SYMBOL[WSP_GGML_OP_COUNT] = {
     "cross_entropy_loss_back(x,y)",
 };
 
-static_assert(WSP_GGML_OP_COUNT == 73, "WSP_GGML_OP_COUNT != 73");
+static_assert(WSP_GGML_OP_COUNT == 72, "WSP_GGML_OP_COUNT != 72");
 
 static_assert(WSP_GGML_OP_POOL_COUNT == 2, "WSP_GGML_OP_POOL_COUNT != 2");
 
+
+static const char * WSP_GGML_UNARY_OP_NAME[WSP_GGML_UNARY_OP_COUNT] = {
+    "ABS",
+    "SGN",
+    "NEG",
+    "STEP",
+    "TANH",
+    "ELU",
+    "RELU",
+    "GELU",
+    "GELU_QUICK",
+    "SILU",
+};
+
+static_assert(WSP_GGML_UNARY_OP_COUNT == 10, "WSP_GGML_UNARY_OP_COUNT != 10");
+
+
 static_assert(sizeof(struct wsp_ggml_object)%WSP_GGML_MEM_ALIGN == 0, "wsp_ggml_object size must be a multiple of WSP_GGML_MEM_ALIGN");
 static_assert(sizeof(struct wsp_ggml_tensor)%WSP_GGML_MEM_ALIGN == 0, "wsp_ggml_tensor size must be a multiple of WSP_GGML_MEM_ALIGN");
 
 // WARN:
-// Mis-confguration can lead to problem that's hard to reason about:
+// Mis-configuration can lead to problem that's hard to reason about:
 // * At best  it crash or talks nosense.
 // * At worst it talks slightly difference but hard to perceive.
 //
@@ -1781,18 +1778,13 @@ static void wsp_ggml_setup_op_has_task_pass(void) {
 
         p[WSP_GGML_OP_ACC                    ] = true;
         p[WSP_GGML_OP_MUL_MAT                ] = true;
+        p[WSP_GGML_OP_MUL_MAT_ID             ] = true;
         p[WSP_GGML_OP_OUT_PROD               ] = true;
         p[WSP_GGML_OP_SET                    ] = true;
         p[WSP_GGML_OP_GET_ROWS_BACK          ] = true;
         p[WSP_GGML_OP_DIAG_MASK_INF          ] = true;
         p[WSP_GGML_OP_DIAG_MASK_ZERO         ] = true;
-        p[WSP_GGML_OP_CONV_1D                ] = true;
-        p[WSP_GGML_OP_CONV_1D_STAGE_0        ] = true;
-        p[WSP_GGML_OP_CONV_1D_STAGE_1        ] = true;
         p[WSP_GGML_OP_CONV_TRANSPOSE_1D      ] = true;
-        p[WSP_GGML_OP_CONV_2D                ] = true;
-        p[WSP_GGML_OP_CONV_2D_STAGE_0        ] = true;
-        p[WSP_GGML_OP_CONV_2D_STAGE_1        ] = true;
         p[WSP_GGML_OP_CONV_TRANSPOSE_2D      ] = true;
         p[WSP_GGML_OP_FLASH_ATTN_BACK        ] = true;
         p[WSP_GGML_OP_CROSS_ENTROPY_LOSS     ] = true;
@@ -2039,6 +2031,20 @@ const char * wsp_ggml_op_symbol(enum wsp_ggml_op op) {
     return WSP_GGML_OP_SYMBOL[op];
 }
 
+const char * wsp_ggml_unary_op_name(enum wsp_ggml_unary_op op) {
+    return WSP_GGML_UNARY_OP_NAME[op];
+}
+
+const char * wsp_ggml_op_desc(const struct wsp_ggml_tensor * t) {
+    if (t->op == WSP_GGML_OP_UNARY) {
+        enum wsp_ggml_unary_op uop = wsp_ggml_get_unary_op(t);
+        return wsp_ggml_unary_op_name(uop);
+    }
+    else {
+        return wsp_ggml_op_name(t->op);
+    }
+}
+
 size_t wsp_ggml_element_size(const struct wsp_ggml_tensor * tensor) {
     return wsp_ggml_type_size(tensor->type);
 }
@@ -3170,9 +3176,7 @@ static struct wsp_ggml_tensor * wsp_ggml_add_impl(
         struct wsp_ggml_tensor * a,
         struct wsp_ggml_tensor * b,
         bool inplace) {
-    // TODO: support less-strict constraint
-    //       WSP_GGML_ASSERT(wsp_ggml_can_repeat(b, a));
-    WSP_GGML_ASSERT(wsp_ggml_can_repeat_rows(b, a));
+    WSP_GGML_ASSERT(wsp_ggml_can_repeat(b, a));
 
     bool is_node = false;
 
@@ -3387,9 +3391,7 @@ static struct wsp_ggml_tensor * wsp_ggml_mul_impl(
         struct wsp_ggml_tensor * a,
         struct wsp_ggml_tensor * b,
         bool inplace) {
-    // TODO: support less-strict constraint
-    //       WSP_GGML_ASSERT(wsp_ggml_can_repeat(b, a));
-    WSP_GGML_ASSERT(wsp_ggml_can_repeat_rows(b, a));
+    WSP_GGML_ASSERT(wsp_ggml_can_repeat(b, a));
 
     bool is_node = false;
 
@@ -3434,7 +3436,7 @@ static struct wsp_ggml_tensor * wsp_ggml_div_impl(
         struct wsp_ggml_tensor * a,
         struct wsp_ggml_tensor * b,
         bool inplace) {
-    WSP_GGML_ASSERT(wsp_ggml_are_same_shape(a, b));
+    WSP_GGML_ASSERT(wsp_ggml_can_repeat(b, a));
 
     bool is_node = false;
 
@@ -3831,12 +3833,25 @@ struct wsp_ggml_tensor * wsp_ggml_relu_inplace(
     return wsp_ggml_unary_inplace(ctx, a, WSP_GGML_UNARY_OP_RELU);
 }
 
-// wsp_ggml_leaky
+// wsp_ggml_leaky_relu
 
-struct wsp_ggml_tensor * wsp_ggml_leaky(
+struct wsp_ggml_tensor * wsp_ggml_leaky_relu(
         struct wsp_ggml_context * ctx,
-        struct wsp_ggml_tensor  * a) {
-    return wsp_ggml_unary(ctx, a, WSP_GGML_UNARY_OP_LEAKY);
+        struct wsp_ggml_tensor  * a, float negative_slope, bool inplace) {
+    bool is_node = false;
+
+    if (!inplace && (a->grad)) {
+        is_node = true;
+    }
+
+    struct wsp_ggml_tensor * result = inplace ? wsp_ggml_view_tensor(ctx, a) : wsp_ggml_dup_tensor(ctx, a);
+    wsp_ggml_set_op_params(result, &negative_slope, sizeof(negative_slope));
+
+    result->op   = WSP_GGML_OP_LEAKY_RELU;
+    result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
 }
 
 // wsp_ggml_gelu
@@ -4023,8 +4038,9 @@ static struct wsp_ggml_tensor * wsp_ggml_group_norm_impl(
 
     struct wsp_ggml_tensor * result = inplace ? wsp_ggml_view_tensor(ctx, a) : wsp_ggml_dup_tensor(ctx, a);
 
-    result->op = WSP_GGML_OP_GROUP_NORM;
     result->op_params[0] = n_groups;
+
+    result->op = WSP_GGML_OP_GROUP_NORM;
     result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = NULL; // TODO: maybe store epsilon here?
@@ -4072,6 +4088,51 @@ struct wsp_ggml_tensor * wsp_ggml_mul_mat(
     return result;
 }
 
+// wsp_ggml_mul_mat_id
+
+struct wsp_ggml_tensor * wsp_ggml_mul_mat_id(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * const as[],
+        int                   n_as,
+        struct wsp_ggml_tensor  * ids,
+        int                   id,
+        struct wsp_ggml_tensor  * b) {
+
+    WSP_GGML_ASSERT(ids->type == WSP_GGML_TYPE_I32);
+    WSP_GGML_ASSERT(ids->ne[2] == 1 && ids->ne[3] == 1);
+    WSP_GGML_ASSERT(ids->ne[1] == b->ne[1]);
+    WSP_GGML_ASSERT(ids->ne[2] == b->ne[2] && ids->ne[3] == b->ne[3]);
+    WSP_GGML_ASSERT(n_as > 0 && n_as <= WSP_GGML_MAX_SRC - 2);
+    WSP_GGML_ASSERT(id >= 0 && id < ids->ne[0]);
+
+    bool is_node = false;
+
+    if (as[0]->grad || b->grad) {
+        is_node = true;
+    }
+
+    const int64_t ne[4] = { as[0]->ne[1], b->ne[1], b->ne[2], b->ne[3] };
+    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor(ctx, WSP_GGML_TYPE_F32, MAX(as[0]->n_dims, b->n_dims), ne);
+
+    wsp_ggml_set_op_params_i32(result, 0, id);
+    wsp_ggml_set_op_params_i32(result, 1, n_as);
+
+    result->op   = WSP_GGML_OP_MUL_MAT_ID;
+    result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = ids;
+    result->src[1] = b;
+
+    for (int i = 0; i < n_as; i++) {
+        struct wsp_ggml_tensor * a = as[i];
+        WSP_GGML_ASSERT(wsp_ggml_are_same_shape(as[0], a));
+        WSP_GGML_ASSERT(wsp_ggml_can_mul_mat(a, b));
+        WSP_GGML_ASSERT(!wsp_ggml_is_transposed(a));
+        result->src[i + 2] = a;
+    }
+
+    return result;
+}
+
 // wsp_ggml_out_prod
 
 struct wsp_ggml_tensor * wsp_ggml_out_prod(
@@ -4225,7 +4286,7 @@ struct wsp_ggml_tensor * wsp_ggml_set_2d_inplace(
         struct wsp_ggml_tensor *  b,
         size_t                nb1,
         size_t                offset) {
-    return wsp_ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false);
+    return wsp_ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, true);
 }
 
 // wsp_ggml_cpy
@@ -4689,7 +4750,9 @@ struct wsp_ggml_tensor * wsp_ggml_get_rows(
         struct wsp_ggml_context * ctx,
         struct wsp_ggml_tensor  * a,
         struct wsp_ggml_tensor  * b) {
-    WSP_GGML_ASSERT(wsp_ggml_is_matrix(a) && wsp_ggml_is_vector(b) && b->type == WSP_GGML_TYPE_I32);
+    WSP_GGML_ASSERT(a->ne[2] == b->ne[1]);
+    WSP_GGML_ASSERT(b->ne[3] == 1);
+    WSP_GGML_ASSERT(b->type == WSP_GGML_TYPE_I32);
 
     bool is_node = false;
 
@@ -4699,7 +4762,7 @@ struct wsp_ggml_tensor * wsp_ggml_get_rows(
 
     // TODO: implement non F32 return
     //struct wsp_ggml_tensor * result = wsp_ggml_new_tensor_2d(ctx, a->type, a->ne[0], b->ne[0]);
-    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor_2d(ctx, WSP_GGML_TYPE_F32, a->ne[0], b->ne[0]);
+    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor_4d(ctx, WSP_GGML_TYPE_F32, a->ne[0], b->ne[0], b->ne[1], b->ne[2]);
 
     result->op   = WSP_GGML_OP_GET_ROWS;
     result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
@@ -4842,7 +4905,17 @@ struct wsp_ggml_tensor * wsp_ggml_diag_mask_zero_inplace(
 static struct wsp_ggml_tensor * wsp_ggml_soft_max_impl(
         struct wsp_ggml_context * ctx,
         struct wsp_ggml_tensor  * a,
+        struct wsp_ggml_tensor  * mask,
+        float                 scale,
         bool                  inplace) {
+    WSP_GGML_ASSERT(wsp_ggml_is_contiguous(a));
+    if (mask) {
+        WSP_GGML_ASSERT(wsp_ggml_is_contiguous(mask));
+        WSP_GGML_ASSERT(mask->ne[2] == 1);
+        WSP_GGML_ASSERT(mask->ne[3] == 1);
+        WSP_GGML_ASSERT(wsp_ggml_can_repeat_rows(mask, a));
+    }
+
     bool is_node = false;
 
     if (a->grad) {
@@ -4851,9 +4924,13 @@ static struct wsp_ggml_tensor * wsp_ggml_soft_max_impl(
 
     struct wsp_ggml_tensor * result = inplace ? wsp_ggml_view_tensor(ctx, a) : wsp_ggml_dup_tensor(ctx, a);
 
+    float params[] = { scale };
+    wsp_ggml_set_op_params(result, params, sizeof(params));
+
     result->op   = WSP_GGML_OP_SOFT_MAX;
     result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
+    result->src[1] = mask;
 
     return result;
 }
@@ -4861,13 +4938,21 @@ static struct wsp_ggml_tensor * wsp_ggml_soft_max_impl(
 struct wsp_ggml_tensor * wsp_ggml_soft_max(
         struct wsp_ggml_context * ctx,
         struct wsp_ggml_tensor  * a) {
-    return wsp_ggml_soft_max_impl(ctx, a, false);
+    return wsp_ggml_soft_max_impl(ctx, a, NULL, 1.0f, false);
 }
 
 struct wsp_ggml_tensor * wsp_ggml_soft_max_inplace(
         struct wsp_ggml_context * ctx,
         struct wsp_ggml_tensor  * a) {
-    return wsp_ggml_soft_max_impl(ctx, a, true);
+    return wsp_ggml_soft_max_impl(ctx, a, NULL, 1.0f, true);
+}
+
+struct wsp_ggml_tensor * wsp_ggml_soft_max_ext(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        struct wsp_ggml_tensor  * mask,
+        float                 scale) {
+    return wsp_ggml_soft_max_impl(ctx, a, mask, scale, false);
 }
 
 // wsp_ggml_soft_max_back
@@ -5040,8 +5125,13 @@ struct wsp_ggml_tensor * wsp_ggml_rope_back(
         int                   n_dims,
         int                   mode,
         int                   n_ctx,
+        int                   n_orig_ctx,
         float                 freq_base,
         float                 freq_scale,
+        float                 ext_factor,
+        float                 attn_factor,
+        float                 beta_fast,
+        float                 beta_slow,
         float                 xpos_base,
         bool                  xpos_down) {
     WSP_GGML_ASSERT(wsp_ggml_is_vector(b));
@@ -5058,11 +5148,15 @@ struct wsp_ggml_tensor * wsp_ggml_rope_back(
 
     struct wsp_ggml_tensor * result = wsp_ggml_dup_tensor(ctx, a);
 
-    int32_t params[8] = { /*n_past*/ 0, n_dims, mode, n_ctx };
-    memcpy(params + 4, &freq_base,  sizeof(float));
-    memcpy(params + 5, &freq_scale, sizeof(float));
-    memcpy(params + 6, &xpos_base,  sizeof(float));
-    memcpy(params + 7, &xpos_down,  sizeof(bool));
+    int32_t params[13] = { /*n_past*/ 0, n_dims, mode, n_ctx, n_orig_ctx };
+    memcpy(params +  5, &freq_base,    sizeof(float));
+    memcpy(params +  6, &freq_scale,   sizeof(float));
+    memcpy(params +  7, &ext_factor,   sizeof(float));
+    memcpy(params +  8, &attn_factor,  sizeof(float));
+    memcpy(params +  9, &beta_fast,    sizeof(float));
+    memcpy(params + 10, &beta_slow,    sizeof(float));
+    memcpy(params + 11, &xpos_base,    sizeof(float));
+    memcpy(params + 12, &xpos_down,    sizeof(bool));
     wsp_ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = WSP_GGML_OP_ROPE_BACK;
@@ -5137,82 +5231,6 @@ static int64_t wsp_ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, in
     return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
 }
 
-// im2col: [N, IC, IL] => [N, OL, IC*K]
-// a: [OC，IC, K]
-// b: [N, IC, IL]
-// result: [N, OL, IC*K]
-static struct wsp_ggml_tensor * wsp_ggml_conv_1d_stage_0(
-    struct wsp_ggml_context * ctx,
-    struct wsp_ggml_tensor  * a,
-    struct wsp_ggml_tensor  * b,
-    int                   s0,
-    int                   p0,
-    int                   d0) {
-    WSP_GGML_ASSERT(a->ne[1] == b->ne[1]);
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        WSP_GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
-    const int64_t OL = wsp_ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
-
-    const int64_t ne[4] = {
-        a->ne[1] * a->ne[0],
-        OL,
-        b->ne[2],
-        1,
-    };
-    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor(ctx, WSP_GGML_TYPE_F16, 4, ne);
-
-    int32_t params[] = { s0, p0, d0 };
-    wsp_ggml_set_op_params(result, params, sizeof(params));
-
-    result->op = WSP_GGML_OP_CONV_1D_STAGE_0;
-    result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-    result->src[1] = b;
-
-    return result;
-}
-
-// wsp_ggml_conv_1d_stage_1
-
-// gemm: [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K]
-// a: [OC, IC, K]
-// b: [N, OL, IC * K]
-// result: [N, OC, OL]
-static struct wsp_ggml_tensor * wsp_ggml_conv_1d_stage_1(
-    struct wsp_ggml_context * ctx,
-    struct wsp_ggml_tensor  * a,
-    struct wsp_ggml_tensor  * b) {
-
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        WSP_GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
-    const int64_t ne[4] = {
-        b->ne[1],
-        a->ne[2],
-        b->ne[2],
-        1,
-    };
-    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor(ctx, WSP_GGML_TYPE_F32, 4, ne);
-
-    result->op = WSP_GGML_OP_CONV_1D_STAGE_1;
-    result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-    result->src[1] = b;
-
-    return result;
-}
-
-// wsp_ggml_conv_1d
-
 WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_1d(
         struct wsp_ggml_context * ctx,
         struct wsp_ggml_tensor  * a,
@@ -5220,43 +5238,17 @@ WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_1d(
         int                   s0,
         int                   p0,
         int                   d0) {
-    struct wsp_ggml_tensor * result = wsp_ggml_conv_1d_stage_0(ctx, a, b, s0, p0, d0);
-    result = wsp_ggml_conv_1d_stage_1(ctx, a, result);
-    return result;
-}
-
-// WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_1d(
-//         struct wsp_ggml_context * ctx,
-//         struct wsp_ggml_tensor  * a,
-//         struct wsp_ggml_tensor  * b,
-//         int                   s0,
-//         int                   p0,
-//         int                   d0) {
-//     WSP_GGML_ASSERT(wsp_ggml_is_matrix(b));
-//     WSP_GGML_ASSERT(a->ne[1] == b->ne[1]);
-//     bool is_node = false;
-
-//     if (a->grad || b->grad) {
-//         WSP_GGML_ASSERT(false); // TODO: implement backward
-//         is_node = true;
-//     }
+    struct wsp_ggml_tensor * im2col = wsp_ggml_im2col(ctx, a, b, s0, 0, p0, 0, d0, 0, false); // [N, OL, IC * K]
 
-//     const int64_t ne[4] = {
-//         wsp_ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0),
-//         a->ne[2], 1, 1,
-//     };
-//     struct wsp_ggml_tensor * result = wsp_ggml_new_tensor(ctx, WSP_GGML_TYPE_F32, 2, ne);
+    struct wsp_ggml_tensor * result =
+        wsp_ggml_mul_mat(ctx,
+                wsp_ggml_reshape_2d(ctx, im2col, im2col->ne[0], (im2col->ne[2] * im2col->ne[1])), // [N, OL, IC * K] => [N*OL, IC * K]
+                wsp_ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1]), a->ne[2]));                    // [OC，IC, K] => [OC, IC * K]
 
-//     int32_t params[] = { s0, p0, d0 };
-//     wsp_ggml_set_op_params(result, params, sizeof(params));
+    result = wsp_ggml_reshape_3d(ctx, result, im2col->ne[1], a->ne[2], im2col->ne[2]); // [N, OC, OL]
 
-//     result->op = WSP_GGML_OP_CONV_1D;
-//     result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
-//     result->src[0] = a;
-//     result->src[1] = b;
-
-//     return result;
-// }
+    return result;
+}
 
 // wsp_ggml_conv_1d_ph
 
@@ -5319,7 +5311,7 @@ WSP_GGML_API struct wsp_ggml_tensor * wsp_ggml_conv_transpose_1d(
 // a: [OC，IC, KH, KW]
 // b: [N, IC, IH, IW]
 // result: [N, OH, OW, IC*KH*KW]
-static struct wsp_ggml_tensor * wsp_ggml_conv_2d_stage_0(
+struct wsp_ggml_tensor * wsp_ggml_im2col(
     struct wsp_ggml_context * ctx,
     struct wsp_ggml_tensor  * a,
     struct wsp_ggml_tensor  * b,
@@ -5328,9 +5320,14 @@ static struct wsp_ggml_tensor * wsp_ggml_conv_2d_stage_0(
     int                  p0,
     int                  p1,
     int                  d0,
-    int                  d1) {
+    int                  d1,
+    bool                 is_2D) {
 
-    WSP_GGML_ASSERT(a->ne[2] == b->ne[2]);
+    if(is_2D) {
+        WSP_GGML_ASSERT(a->ne[2] == b->ne[2]);
+    } else {
+        WSP_GGML_ASSERT(a->ne[1] == b->ne[1]);
+    }
     bool is_node = false;
 
     if (a->grad || b->grad) {
@@ -5338,81 +5335,51 @@ static struct wsp_ggml_tensor * wsp_ggml_conv_2d_stage_0(
         is_node = true;
     }
 
-    const int64_t OH = wsp_ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1);
-    const int64_t OW = wsp_ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+    const int64_t OH = is_2D ? wsp_ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0;
+    const int64_t OW =         wsp_ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
 
     const int64_t ne[4] = {
-        a->ne[2] * a->ne[1] * a->ne[0],
+        is_2D ? (a->ne[2] * a->ne[1] * a->ne[0]) : a->ne[1] * a->ne[0],
         OW,
-        OH,
-        b->ne[3],
+        is_2D ? OH : b->ne[2],
+        is_2D ?      b->ne[3] : 1,
     };
-    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor(ctx, WSP_GGML_TYPE_F16, 4, ne);
 
-    int32_t params[] = { s0, s1, p0, p1, d0, d1 };
+    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor(ctx, WSP_GGML_TYPE_F16, 4, ne);
+    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
     wsp_ggml_set_op_params(result, params, sizeof(params));
 
-    result->op = WSP_GGML_OP_CONV_2D_STAGE_0;
-    result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-    result->src[1] = b;
-
-    return result;
-
-}
-
-// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW]
-// a: [OC, IC, KH, KW]
-// b: [N, OH, OW, IC * KH * KW]
-// result: [N, OC, OH, OW]
-static struct wsp_ggml_tensor * wsp_ggml_conv_2d_stage_1(
-    struct wsp_ggml_context * ctx,
-    struct wsp_ggml_tensor  * a,
-    struct wsp_ggml_tensor  * b) {
-
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        WSP_GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
-    const int64_t ne[4] = {
-        b->ne[1],
-        b->ne[2],
-        a->ne[3],
-        b->ne[3],
-    };
-    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor(ctx, WSP_GGML_TYPE_F32, 4, ne);
-
-    result->op = WSP_GGML_OP_CONV_2D_STAGE_1;
+    result->op = WSP_GGML_OP_IM2COL;
     result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = b;
 
     return result;
-
 }
 
 // a: [OC，IC, KH, KW]
 // b: [N, IC, IH, IW]
 // result: [N, OC, OH, OW]
 struct wsp_ggml_tensor * wsp_ggml_conv_2d(
-    struct wsp_ggml_context * ctx,
-    struct wsp_ggml_tensor  * a,
-    struct wsp_ggml_tensor  * b,
-    int                  s0,
-    int                  s1,
-    int                  p0,
-    int                  p1,
-    int                  d0,
-    int                  d1) {
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        struct wsp_ggml_tensor  * b,
+        int                  s0,
+        int                  s1,
+        int                  p0,
+        int                  p1,
+        int                  d0,
+        int                  d1) {
+    struct wsp_ggml_tensor * im2col = wsp_ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true); // [N, OH, OW, IC * KH * KW]
 
-    struct wsp_ggml_tensor * result = wsp_ggml_conv_2d_stage_0(ctx, a, b, s0, s1, p0, p1, d0, d1); // [N, OH, OW, IC * KH * KW]
-    result = wsp_ggml_conv_2d_stage_1(ctx, a, result);
+    struct wsp_ggml_tensor * result =
+        wsp_ggml_mul_mat(ctx,
+                wsp_ggml_reshape_2d(ctx, im2col, im2col->ne[0],  im2col->ne[3] * im2col->ne[2] * im2col->ne[1]), // [N, OH, OW, IC * KH * KW] => [N*OH*OW, IC * KH * KW]
+                wsp_ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1] * a->ne[2]),  a->ne[3]));                       // [OC，IC, KH, KW] => [OC, IC * KH * KW]
 
-    return result;
+    result = wsp_ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], a->ne[3], im2col->ne[3]); // [N, OC, OH, OW]
 
+    return result;
 }
 
 // wsp_ggml_conv_2d_sk_p0
@@ -5573,6 +5540,30 @@ static struct wsp_ggml_tensor * wsp_ggml_upscale_impl(
     return result;
 }
 
+struct wsp_ggml_tensor * wsp_ggml_pad(
+    struct wsp_ggml_context * ctx,
+    struct wsp_ggml_tensor  * a,
+    int p0, int p1, int p2, int p3) {
+    bool is_node = false;
+
+    if (a->grad) {
+        WSP_GGML_ASSERT(false); // TODO: implement backward
+        is_node = true;
+    }
+
+    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor_4d(ctx, a->type,
+            a->ne[0] + p0,
+            a->ne[1] + p1,
+            a->ne[2] + p2,
+            a->ne[3] + p3);
+
+    result->op = WSP_GGML_OP_PAD;
+    result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
+}
+
 struct wsp_ggml_tensor * wsp_ggml_upscale(
     struct wsp_ggml_context * ctx,
     struct wsp_ggml_tensor * a,
@@ -5580,6 +5571,43 @@ struct wsp_ggml_tensor * wsp_ggml_upscale(
     return wsp_ggml_upscale_impl(ctx, a, scale_factor);
 }
 
+// wsp_ggml_argsort
+
+struct wsp_ggml_tensor * wsp_ggml_argsort(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        enum wsp_ggml_sort_order  order) {
+    bool is_node = false;
+
+    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor(ctx, WSP_GGML_TYPE_I32, a->n_dims, a->ne);
+
+    wsp_ggml_set_op_params_i32(result, 0, (int32_t) order);
+
+    result->op   = WSP_GGML_OP_ARGSORT;
+    result->grad = is_node ? wsp_ggml_dup_tensor(ctx, result) : NULL;
+    result->src[0] = a;
+
+    return result;
+}
+
+// wsp_ggml_top_k
+
+struct wsp_ggml_tensor * wsp_ggml_top_k(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        int                   k) {
+    WSP_GGML_ASSERT(a->ne[0] >= k);
+
+    struct wsp_ggml_tensor * result = wsp_ggml_argsort(ctx, a, WSP_GGML_SORT_DESC);
+
+    result = wsp_ggml_view_4d(ctx, result,
+                k, result->ne[1], result->ne[2], result->ne[3],
+                   result->nb[1], result->nb[2], result->nb[3],
+                0);
+
+    return result;
+}
+
 // wsp_ggml_flash_attn
 
 struct wsp_ggml_tensor * wsp_ggml_flash_attn(
@@ -6472,7 +6500,7 @@ static void wsp_ggml_compute_forward_dup_f16(
                     }
                 }
             } else if (type_traits[dst->type].from_float) {
-                wsp_ggml_from_float_t const quantize_row_q = type_traits[dst->type].from_float;
+                wsp_ggml_from_float_t const wsp_quantize_row_q = type_traits[dst->type].from_float;
                 float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
 
                 size_t id = 0;
@@ -6489,7 +6517,7 @@ static void wsp_ggml_compute_forward_dup_f16(
                                 src0_f32[i00] = WSP_GGML_FP16_TO_FP32(src0_ptr[i00]);
                             }
 
-                            quantize_row_q(src0_f32, dst_ptr + id, ne00);
+                            wsp_quantize_row_q(src0_f32, dst_ptr + id, ne00);
                             id += rs;
                         }
                         id += rs * (ne01 - ir1);
@@ -6725,7 +6753,7 @@ static void wsp_ggml_compute_forward_dup_f32(
                     }
                 }
             } else if (type_traits[dst->type].from_float) {
-                wsp_ggml_from_float_t const quantize_row_q = type_traits[dst->type].from_float;
+                wsp_ggml_from_float_t const wsp_quantize_row_q = type_traits[dst->type].from_float;
 
                 size_t id = 0;
                 size_t rs = nb0 * (ne00 / wsp_ggml_blck_size(dst->type));
@@ -6736,7 +6764,7 @@ static void wsp_ggml_compute_forward_dup_f32(
                         id += rs * ir0;
                         for (int i01 = ir0; i01 < ir1; i01++) {
                             const float * src0_ptr = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
-                            quantize_row_q(src0_ptr, dst_ptr + id, ne00);
+                            wsp_quantize_row_q(src0_ptr, dst_ptr + id, ne00);
                             id += rs;
                         }
                         id += rs * (ne01 - ir1);
@@ -6939,7 +6967,7 @@ static void wsp_ggml_compute_forward_add_f32(
         const struct wsp_ggml_tensor * src0,
         const struct wsp_ggml_tensor * src1,
         struct wsp_ggml_tensor * dst) {
-    WSP_GGML_ASSERT(wsp_ggml_can_repeat_rows(src1, src0) && wsp_ggml_are_same_shape(src0, dst));
+    WSP_GGML_ASSERT(wsp_ggml_can_repeat(src1, src0) && wsp_ggml_are_same_shape(src0, dst));
 
     if (params->type == WSP_GGML_TASK_INIT || params->type == WSP_GGML_TASK_FINALIZE) {
         return;
@@ -6972,16 +7000,19 @@ static void wsp_ggml_compute_forward_add_f32(
             const int64_t i13 = i03 % ne13;
             const int64_t i12 = i02 % ne12;
             const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
 
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
             float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
 
+            for (int64_t r = 0; r < nr0; ++r) {
 #ifdef WSP_GGML_USE_ACCELERATE
-            vDSP_vadd(src0_ptr, 1, src1_ptr, 1, dst_ptr, 1, ne00);
+                vDSP_vadd(src0_ptr + r*ne10, 1, src1_ptr, 1, dst_ptr + r*ne10, 1, ne10);
 #else
-            wsp_ggml_vec_add_f32(ne00, dst_ptr, src0_ptr, src1_ptr);
+                wsp_ggml_vec_add_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
 #endif
+            }
         }
     } else {
         // src1 is not contiguous
@@ -6998,8 +7029,9 @@ static void wsp_ggml_compute_forward_add_f32(
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
 
-            for (int i0 = 0; i0 < ne0; i0++) {
-                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i0*nb10);
+            for (int64_t i0 = 0; i0 < ne0; ++i0) {
+                const int64_t i10 = i0 % ne10;
+                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10);
 
                 dst_ptr[i0] = src0_ptr[i0] + *src1_ptr;
             }
@@ -7158,8 +7190,8 @@ static void wsp_ggml_compute_forward_add_q_f32(
 
     const enum wsp_ggml_type type = src0->type;
     const enum wsp_ggml_type dtype = dst->type;
-    wsp_ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
-    wsp_ggml_from_float_t const quantize_row_q = type_traits[dtype].from_float;
+    wsp_ggml_to_float_t const wsp_dewsp_quantize_row_q = type_traits[type].to_float;
+    wsp_ggml_from_float_t const wsp_quantize_row_q = type_traits[dtype].from_float;
 
     // we don't support permuted src0 or src1
     WSP_GGML_ASSERT(nb00 == wsp_ggml_type_size(type));
@@ -7204,12 +7236,12 @@ static void wsp_ggml_compute_forward_add_q_f32(
         assert(ne00 % 32 == 0);
 
         // unquantize row from src0 to temp buffer
-        dequantize_row_q(src0_row, wdata, ne00);
+        wsp_dewsp_quantize_row_q(src0_row, wdata, ne00);
         // add src1
         wsp_ggml_vec_acc_f32(ne00, wdata, src1_row);
         // quantize row to dst
-        if (quantize_row_q != NULL) {
-            quantize_row_q(wdata, dst_row, ne00);
+        if (wsp_quantize_row_q != NULL) {
+            wsp_quantize_row_q(wdata, dst_row, ne00);
         } else {
             memcpy(dst_row, wdata, ne0*nb0);
         }
@@ -7435,8 +7467,8 @@ static void wsp_ggml_compute_forward_add1_q_f32(
     WSP_GGML_TENSOR_UNARY_OP_LOCALS
 
     const enum wsp_ggml_type type = src0->type;
-    wsp_ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
-    wsp_ggml_from_float_t const quantize_row_q = type_traits[type].from_float;
+    wsp_ggml_to_float_t const wsp_dewsp_quantize_row_q = type_traits[type].to_float;
+    wsp_ggml_from_float_t const wsp_quantize_row_q = type_traits[type].from_float;
 
     // we don't support permuted src0
     WSP_GGML_ASSERT(nb00 == wsp_ggml_type_size(type));
@@ -7471,11 +7503,11 @@ static void wsp_ggml_compute_forward_add1_q_f32(
         assert(ne0 % 32 == 0);
 
         // unquantize row from src0 to temp buffer
-        dequantize_row_q(src0_row, wdata, ne0);
+        wsp_dewsp_quantize_row_q(src0_row, wdata, ne0);
         // add src1
         wsp_ggml_vec_acc1_f32(ne0, wdata, v);
         // quantize row to dst
-        quantize_row_q(wdata, dst_row, ne0);
+        wsp_quantize_row_q(wdata, dst_row, ne0);
     }
 }
 
@@ -7533,7 +7565,7 @@ static void wsp_ggml_compute_forward_acc_f32(
     WSP_GGML_ASSERT(wsp_ggml_is_contiguous(dst) && wsp_ggml_is_contiguous(src0));
 
     // view src0 and dst with these strides and data offset inbytes during acc
-    // nb0 is implicitely element_size because src0 and dst are contiguous
+    // nb0 is implicitly element_size because src0 and dst are contiguous
     size_t nb1     = ((int32_t *) dst->op_params)[0];
     size_t nb2     = ((int32_t *) dst->op_params)[1];
     size_t nb3     = ((int32_t *) dst->op_params)[2];
@@ -7719,7 +7751,7 @@ static void wsp_ggml_compute_forward_mul_f32(
         const struct wsp_ggml_tensor * src0,
         const struct wsp_ggml_tensor * src1,
         struct wsp_ggml_tensor * dst) {
-    WSP_GGML_ASSERT(wsp_ggml_can_repeat_rows(src1, src0) && wsp_ggml_are_same_shape(src0, dst));
+    WSP_GGML_ASSERT(wsp_ggml_can_repeat(src1, src0) && wsp_ggml_are_same_shape(src0, dst));
 
     if (params->type == WSP_GGML_TASK_INIT || params->type == WSP_GGML_TASK_FINALIZE) {
         return;
@@ -7727,8 +7759,10 @@ static void wsp_ggml_compute_forward_mul_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
+// TODO: OpenCL kernel support broadcast
 #ifdef WSP_GGML_USE_CLBLAST
     if (src1->backend == WSP_GGML_BACKEND_GPU) {
+        WSP_GGML_ASSERT(wsp_ggml_are_same_shape(src0, src1));
         if (ith == 0) {
             wsp_ggml_cl_mul(src0, src1, dst);
         }
@@ -7742,7 +7776,6 @@ static void wsp_ggml_compute_forward_mul_f32(
 
     WSP_GGML_ASSERT( nb0 == sizeof(float));
     WSP_GGML_ASSERT(nb00 == sizeof(float));
-    WSP_GGML_ASSERT(ne00 == ne10);
 
     if (nb10 == sizeof(float)) {
         for (int64_t ir = ith; ir < nr; ir += nth) {
@@ -7754,20 +7787,21 @@ static void wsp_ggml_compute_forward_mul_f32(
             const int64_t i13 = i03 % ne13;
             const int64_t i12 = i02 % ne12;
             const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
 
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
             float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
 
+            for (int64_t r = 0 ; r < nr0; ++r) {
 #ifdef WSP_GGML_USE_ACCELERATE
-            UNUSED(wsp_ggml_vec_mul_f32);
+                UNUSED(wsp_ggml_vec_mul_f32);
 
-            vDSP_vmul( src0_ptr, 1, src1_ptr, 1, dst_ptr,  1, ne00);
+                vDSP_vmul(src0_ptr + r*ne10, 1, src1_ptr, 1, dst_ptr + r*ne10, 1, ne10);
 #else
-            wsp_ggml_vec_mul_f32(ne00, dst_ptr, src0_ptr, src1_ptr);
+                wsp_ggml_vec_mul_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
 #endif
-                // }
-            // }
+            }
         }
     } else {
         // src1 is not contiguous
@@ -7785,8 +7819,9 @@ static void wsp_ggml_compute_forward_mul_f32(
             float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
             float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
 
-            for (int64_t i0 = 0; i0 < ne00; i0++) {
-                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i0*nb10);
+            for (int64_t i0 = 0; i0 < ne00; ++i0) {
+                const int64_t i10 = i0 % ne10;
+                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10);
 
                 dst_ptr[i0] = src0_ptr[i0] * (*src1_ptr);
             }
@@ -7820,14 +7855,16 @@ static void wsp_ggml_compute_forward_div_f32(
         const struct wsp_ggml_tensor * src0,
         const struct wsp_ggml_tensor * src1,
         struct wsp_ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(wsp_ggml_are_same_shape(src0, src1) && wsp_ggml_are_same_shape(src0, dst));
+    WSP_GGML_ASSERT(wsp_ggml_can_repeat(src1, src0) && wsp_ggml_are_same_shape(src0, dst));
 
     if (params->type == WSP_GGML_TASK_INIT || params->type == WSP_GGML_TASK_FINALIZE) {
         return;
     }
 
-    const int nr  = wsp_ggml_nrows(src0);
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nr = wsp_ggml_nrows(src0);
 
     WSP_GGML_TENSOR_BINARY_OP_LOCALS
 
@@ -7835,41 +7872,50 @@ static void wsp_ggml_compute_forward_div_f32(
     WSP_GGML_ASSERT(nb00 == sizeof(float));
 
     if (nb10 == sizeof(float)) {
-        for (int ir = 0; ir < nr; ++ir) {
-            // src0, src1 and dst are same shape => same indices
-            const int i3 = ir/(ne2*ne1);
-            const int i2 = (ir - i3*ne2*ne1)/ne1;
-            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+        for (int64_t ir = ith; ir < nr; ir += nth) {
+            // src0 and dst are same shape => same indices
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
 
-#ifdef WSP_GGML_USE_ACCELERATE
-            UNUSED(wsp_ggml_vec_div_f32);
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
 
-            vDSP_vdiv(
-                    (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1,
-                    (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1,
-                    (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ), 1,
-                    ne0);
+            float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+            float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
+
+            for (int64_t r = 0; r < nr0; ++r) {
+#ifdef WSP_GGML_USE_ACCELERATE
+                UNUSED(wsp_ggml_vec_div_f32);
+
+                vDSP_vdiv(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10);
 #else
-            wsp_ggml_vec_div_f32(ne0,
-                    (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ),
-                    (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01),
-                    (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11));
+                wsp_ggml_vec_div_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
 #endif
-                // }
-            // }
+            }
         }
     } else {
         // src1 is not contiguous
-        for (int ir = 0; ir < nr; ++ir) {
-            // src0, src1 and dst are same shape => same indices
-            const int i3 = ir/(ne2*ne1);
-            const int i2 = (ir - i3*ne2*ne1)/ne1;
-            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+        for (int64_t ir = ith; ir < nr; ir += nth) {
+            // src0 and dst are same shape => same indices
+            // src1 is broadcastable across src0 and dst in i1, i2, i3
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
 
-            float * dst_ptr  = (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 );
-            float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
-            for (int i0 = 0; i0 < ne0; i0++) {
-                float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10);
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+
+            float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+
+            for (int64_t i0 = 0; i0 < ne00; ++i0) {
+                const int64_t i10 = i0 % ne10;
+                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10);
 
                 dst_ptr[i0] = src0_ptr[i0] / (*src1_ptr);
             }
@@ -8315,7 +8361,7 @@ static void wsp_ggml_compute_forward_repeat_f16(
         return;
     }
 
-    WSP_GGML_TENSOR_UNARY_OP_LOCALS;
+    WSP_GGML_TENSOR_UNARY_OP_LOCALS
 
     // guaranteed to be an integer due to the check in wsp_ggml_can_repeat
     const int nr0 = (int)(ne0/ne00);
@@ -8460,6 +8506,7 @@ static void wsp_ggml_compute_forward_concat_f32(
     WSP_GGML_ASSERT(src0->nb[0] == sizeof(float));
 
     const int ith = params->ith;
+    const int nth = params->nth;
 
     WSP_GGML_TENSOR_BINARY_OP_LOCALS
 
@@ -8469,7 +8516,7 @@ static void wsp_ggml_compute_forward_concat_f32(
     WSP_GGML_ASSERT(nb10 == sizeof(float));
 
     for (int i3 = 0; i3 < ne3; i3++) {
-        for (int i2 = ith; i2 < ne2; i2++) {
+        for (int i2 = ith; i2 < ne2; i2 += nth) {
             if (i2 < ne02) { // src0
                 for (int i1 = 0; i1 < ne1; i1++) {
                     for (int i0 = 0; i0 < ne0; i0++) {
@@ -8981,10 +9028,9 @@ static void wsp_ggml_compute_forward_silu(
             } break;
     }
 }
+// wsp_ggml_compute_forward_leaky_relu
 
-// wsp_ggml_compute_forward_leaky
-
-static void wsp_ggml_compute_forward_leaky_f32(
+static void wsp_ggml_compute_forward_leaky_relu_f32(
         const struct wsp_ggml_compute_params * params,
         const struct wsp_ggml_tensor * src0,
         struct wsp_ggml_tensor * dst) {
@@ -8998,24 +9044,27 @@ static void wsp_ggml_compute_forward_leaky_f32(
     const int n  = wsp_ggml_nrows(src0);
     const int nc = src0->ne[0];
 
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+
     assert(dst->nb[0]  == sizeof(float));
     assert(src0->nb[0] == sizeof(float));
 
     for (int i = 0; i < n; i++) {
-        wsp_ggml_vec_leaky_f32(nc,
+        wsp_ggml_vec_leaky_relu_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
-                (float *) ((char *) src0->data + i*(src0->nb[1])));
+                (float *) ((char *) src0->data + i*(src0->nb[1])), negative_slope);
     }
 }
 
-static void wsp_ggml_compute_forward_leaky(
+static void wsp_ggml_compute_forward_leaky_relu(
         const struct wsp_ggml_compute_params * params,
         const struct wsp_ggml_tensor * src0,
         struct wsp_ggml_tensor * dst) {
     switch (src0->type) {
         case WSP_GGML_TYPE_F32:
             {
-                wsp_ggml_compute_forward_leaky_f32(params, src0, dst);
+                wsp_ggml_compute_forward_leaky_relu_f32(params, src0, dst);
             } break;
         default:
             {
@@ -9504,9 +9553,14 @@ static bool wsp_ggml_compute_forward_mul_mat_use_blas(
     const int64_t ne0 = dst->ne[0];
     const int64_t ne1 = dst->ne[1];
 
+    // NOTE: with WSP_GGML_OP_MUL_MAT_ID we don't want to go through the BLAS branch because it will dequantize (to_float)
+    //       all the experts for each batch element and the processing would become incredibly slow
     // TODO: find the optimal values for these
-    if (wsp_ggml_is_contiguous(src0) &&
+    if (dst->op != WSP_GGML_OP_MUL_MAT_ID &&
+        wsp_ggml_is_contiguous(src0) &&
         wsp_ggml_is_contiguous(src1) &&
+      //src0->type == WSP_GGML_TYPE_F32 &&
+        src1->type == WSP_GGML_TYPE_F32 &&
         (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
 
         /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
@@ -9517,11 +9571,16 @@ static bool wsp_ggml_compute_forward_mul_mat_use_blas(
 }
 #endif
 
+// off1 = offset in i11 and i1
+// cne1 = ne11 and ne1
+// in a normal matrix multiplication, off1 = 0 and cne1 = ne1
+// during WSP_GGML_TASK_INIT, the full src1 is converted regardless of off1 and cne1
 static void wsp_ggml_compute_forward_mul_mat(
         const struct wsp_ggml_compute_params * params,
         const struct wsp_ggml_tensor * src0,
         const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
+              struct wsp_ggml_tensor * dst,
+              int64_t off1, int64_t cne1) {
     int64_t t0 = wsp_ggml_perf_time_us();
     UNUSED(t0);
 
@@ -9545,7 +9604,7 @@ static void wsp_ggml_compute_forward_mul_mat(
 
     // we don't support permuted src0 or src1
     WSP_GGML_ASSERT(nb00 == wsp_ggml_type_size(type));
-    WSP_GGML_ASSERT(nb10 == sizeof(float));
+    WSP_GGML_ASSERT(nb10 == wsp_ggml_type_size(src1->type));
 
     // dst cannot be transposed or permuted
     WSP_GGML_ASSERT(nb0 == sizeof(float));
@@ -9589,10 +9648,9 @@ static void wsp_ggml_compute_forward_mul_mat(
                 const int64_t i03 = i13/r3;
                 const int64_t i02 = i12/r2;
 
-                const void  * x = (char *)            src0->data + i02*nb02 + i03*nb03;
-                const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
-
-                float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
+                const void  * x = (char *)            src0->data +             i02*nb02 + i03*nb03;
+                const float * y = (float *) ((char *) src1->data + off1*nb11 + i12*nb12 + i13*nb13);
+                      float * d = (float *) ((char *)  dst->data + off1*nb1  + i12*nb2  + i13*nb3);
 
                 if (type != WSP_GGML_TYPE_F32) {
                             float * const wdata    = params->wdata;
@@ -9609,10 +9667,10 @@ static void wsp_ggml_compute_forward_mul_mat(
                 }
 
                 cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
-                        ne11, ne01, ne10,
-                        1.0f,    y, ne10,
-                                 x, ne00,
-                        0.0f,    d, ne01);
+                         cne1, ne01, ne10,
+                         1.0f,    y, ne10,
+                                  x, ne00,
+                         0.0f,    d, ne01);
             }
         }
 
@@ -9627,6 +9685,9 @@ static void wsp_ggml_compute_forward_mul_mat(
             char * wdata = params->wdata;
             const size_t row_size = ne10*wsp_ggml_type_size(vec_dot_type)/wsp_ggml_blck_size(vec_dot_type);
 
+            assert(params->wsize >= ne11*ne12*ne13*row_size);
+            assert(src1->type == WSP_GGML_TYPE_F32);
+
             for (int64_t i13 = 0; i13 < ne13; ++i13) {
                 for (int64_t i12 = 0; i12 < ne12; ++i12) {
                     for (int64_t i11 = 0; i11 < ne11; ++i11) {
@@ -9648,7 +9709,7 @@ static void wsp_ggml_compute_forward_mul_mat(
     const size_t row_size = ne10*wsp_ggml_type_size(vec_dot_type)/wsp_ggml_blck_size(vec_dot_type);
 
     const int64_t nr0 = ne01;           // src0 rows
-    const int64_t nr1 = ne11*ne12*ne13; // src1 rows
+    const int64_t nr1 = cne1*ne12*ne13; // src1 rows
 
     //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
 
@@ -9690,9 +9751,9 @@ static void wsp_ggml_compute_forward_mul_mat(
     for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
         for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
             for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
-                const int64_t i13 = (ir1/(ne12*ne11));
-                const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
-                const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
+                const int64_t i13 = (ir1/(ne12*cne1));
+                const int64_t i12 = (ir1 - i13*ne12*cne1)/cne1;
+                const int64_t i11 = (ir1 - i13*ne12*cne1 - i12*cne1) + off1;
 
                 // broadcast src0 into src1
                 const int64_t i03 = i13/r3;
@@ -9728,6 +9789,34 @@ static void wsp_ggml_compute_forward_mul_mat(
     }
 }
 
+// wsp_ggml_compute_forward_mul_mat_id
+
+static void wsp_ggml_compute_forward_mul_mat_id(
+        const struct wsp_ggml_compute_params * params,
+        const struct wsp_ggml_tensor * src0,
+        const struct wsp_ggml_tensor * src1,
+              struct wsp_ggml_tensor * dst) {
+
+    if (params->type == WSP_GGML_TASK_INIT || params->type == WSP_GGML_TASK_FINALIZE) {
+        // during WSP_GGML_TASK_INIT the entire src1 is converted to vec_dot_type
+        wsp_ggml_compute_forward_mul_mat(params, dst->src[2], src1, dst, 0, dst->ne[1]);
+        return;
+    }
+
+    const struct wsp_ggml_tensor * ids = src0;
+    const int id   = wsp_ggml_get_op_params_i32(dst, 0);
+    const int n_as = wsp_ggml_get_op_params_i32(dst, 1);
+
+    for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
+        const int32_t row_id = *(const int32_t *) ((const char *) ids->data + i01*ids->nb[1] + id*ids->nb[0]);
+
+        WSP_GGML_ASSERT(row_id >= 0 && row_id < n_as);
+
+        const struct wsp_ggml_tensor * src0_row = dst->src[row_id + 2];
+        wsp_ggml_compute_forward_mul_mat(params, src0_row, src1, dst, i01, 1);
+    }
+}
+
 // wsp_ggml_compute_forward_out_prod
 
 static void wsp_ggml_compute_forward_out_prod_f32(
@@ -9743,10 +9832,12 @@ static void wsp_ggml_compute_forward_out_prod_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
+    WSP_GGML_ASSERT(ne0  == ne00);
+    WSP_GGML_ASSERT(ne1  == ne10);
+    WSP_GGML_ASSERT(ne2  == ne02);
     WSP_GGML_ASSERT(ne02 == ne12);
-    WSP_GGML_ASSERT(ne03 == ne13);
-    WSP_GGML_ASSERT(ne2  == ne12);
     WSP_GGML_ASSERT(ne3  == ne13);
+    WSP_GGML_ASSERT(ne03 == ne13);
 
     // we don't support permuted src0 or src1
     WSP_GGML_ASSERT(nb00 == sizeof(float));
@@ -9757,18 +9848,25 @@ static void wsp_ggml_compute_forward_out_prod_f32(
     // WSP_GGML_ASSERT(nb1 <= nb2);
     // WSP_GGML_ASSERT(nb2 <= nb3);
 
-    WSP_GGML_ASSERT(ne0 == ne00);
-    WSP_GGML_ASSERT(ne1 == ne10);
-    WSP_GGML_ASSERT(ne2 == ne02);
-    WSP_GGML_ASSERT(ne3 == ne03);
-
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
     // TODO: #if defined(WSP_GGML_USE_CUBLAS) wsp_ggml_cuda_out_prod
-    // TODO: #if defined(WSP_GGML_USE_ACCELERATE) || defined(WSP_GGML_USE_OPENBLAS) || defined(WSP_GGML_USE_CLBLAST)
+    // TODO: #if defined(WSP_GGML_USE_CLBLAST)
+
+#if defined(WSP_GGML_USE_ACCELERATE) || defined(WSP_GGML_USE_OPENBLAS)
+    bool use_blas = wsp_ggml_is_matrix(src0) &&
+        wsp_ggml_is_matrix(src1) &&
+        wsp_ggml_is_contiguous(src0) &&
+        (wsp_ggml_is_contiguous(src1) || wsp_ggml_is_transposed(src1));
+#endif
 
     if (params->type == WSP_GGML_TASK_INIT) {
+#if defined(WSP_GGML_USE_ACCELERATE) || defined(WSP_GGML_USE_OPENBLAS) // gemm beta will zero dst
+        if (use_blas) {
+            return;
+        }
+#endif
         wsp_ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0);
         return;
     }
@@ -9777,6 +9875,50 @@ static void wsp_ggml_compute_forward_out_prod_f32(
         return;
     }
 
+#if defined(WSP_GGML_USE_ACCELERATE) || defined(WSP_GGML_USE_OPENBLAS)
+    if (use_blas) {
+        if (params->ith != 0) { // All threads other than the first do no work.
+            return;
+        }
+        // Arguments to wsp_ggml_compute_forward_out_prod (expressed as major,minor)
+        // src0: (k,n)
+        // src1: (k,m)
+        // dst:  (m,n)
+        //
+        // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
+        // Also expressed as (major,minor)
+        // a: (m,k): so src1 transposed
+        // b: (k,n): so src0
+        // c: (m,n)
+        //
+        // However, if wsp_ggml_is_transposed(src1) is true, then
+        // src1->data already contains a transposed version, so sgemm mustn't
+        // transpose it further.
+
+        int n = src0->ne[0];
+        int k = src0->ne[1];
+        int m = src1->ne[0];
+
+        int transposeA, lda;
+
+        if (!wsp_ggml_is_transposed(src1)) {
+            transposeA = CblasTrans;
+            lda = m;
+        } else {
+            transposeA = CblasNoTrans;
+            lda = k;
+        }
+
+        float * a = (float *) ((char *) src1->data);
+        float * b = (float *) ((char *) src0->data);
+        float * c = (float *) ((char *) dst->data);
+
+        cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
+
+        return;
+    }
+#endif
+
     // dst[:,:,:,:] = 0
     // for i2,i3:
     //   for i1:
@@ -9880,7 +10022,7 @@ static void wsp_ggml_compute_forward_out_prod_q_f32(
     const int nth = params->nth;
 
     const enum wsp_ggml_type type = src0->type;
-    wsp_ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
+    wsp_ggml_to_float_t const wsp_dewsp_quantize_row_q = type_traits[type].to_float;
 
     WSP_GGML_ASSERT(ne02 == ne12);
     WSP_GGML_ASSERT(ne03 == ne13);
@@ -9957,7 +10099,7 @@ static void wsp_ggml_compute_forward_out_prod_q_f32(
             float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13));
             float * d  = (float *) ((char *)  dst->data + (          i1*nb1 + i2*nb2 + i3*nb3));
 
-            dequantize_row_q(s0, wdata, ne0);
+            wsp_dewsp_quantize_row_q(s0, wdata, ne0);
             wsp_ggml_vec_mad_f32(ne0, d, wdata, *s1);
         }
     }
@@ -10084,7 +10226,7 @@ static void wsp_ggml_compute_forward_set_f32(
     WSP_GGML_ASSERT(wsp_ggml_is_contiguous(dst) && wsp_ggml_is_contiguous(src0));
 
     // view src0 and dst with these strides and data offset inbytes during set
-    // nb0 is implicitely element_size because src0 and dst are contiguous
+    // nb0 is implicitly element_size because src0 and dst are contiguous
     size_t nb1     = ((int32_t *) dst->op_params)[0];
     size_t nb2     = ((int32_t *) dst->op_params)[1];
     size_t nb3     = ((int32_t *) dst->op_params)[2];
@@ -10248,21 +10390,30 @@ static void wsp_ggml_compute_forward_get_rows_q(
         return;
     }
 
-    const int nc = src0->ne[0];
-    const int nr = wsp_ggml_nelements(src1);
-    const enum wsp_ggml_type type = src0->type;
-    wsp_ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
-
-    assert( dst->ne[0] == nc);
-    assert( dst->ne[1] == nr);
-    assert(src0->nb[0] == wsp_ggml_type_size(type));
+    WSP_GGML_TENSOR_BINARY_OP_LOCALS
 
-    for (int i = 0; i < nr; ++i) {
-        const int r = ((int32_t *) src1->data)[i];
+    const int64_t nc = ne00;
+    const int64_t nr = wsp_ggml_nelements(src1); WSP_GGML_UNUSED(nr);
 
-        dequantize_row_q(
-                (const void *) ((char *) src0->data + r*src0->nb[1]),
-                     (float *) ((char *)  dst->data + i*dst->nb[1]), nc);
+    const enum wsp_ggml_type type = src0->type;
+    wsp_ggml_to_float_t const wsp_dewsp_quantize_row_q = type_traits[type].to_float;
+
+    assert(ne0  == nc);
+    assert(ne02 == ne11);
+    assert(nb00 == wsp_ggml_type_size(type));
+    assert(wsp_ggml_nrows(dst) == nr);
+
+    // TODO: multi-thread
+    for (int64_t i12 = 0; i12 < ne12; ++i12) {
+        for (int64_t i11 = 0; i11 < ne11; ++i11) {
+            for (int64_t i10 = 0; i10 < ne10; ++i10) {
+                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+                wsp_dewsp_quantize_row_q(
+                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
+            }
+        }
     }
 }
 
@@ -10277,19 +10428,26 @@ static void wsp_ggml_compute_forward_get_rows_f16(
         return;
     }
 
-    const int nc = src0->ne[0];
-    const int nr = wsp_ggml_nelements(src1);
+    WSP_GGML_TENSOR_BINARY_OP_LOCALS
 
-    assert( dst->ne[0] == nc);
-    assert( dst->ne[1] == nr);
-    assert(src0->nb[0] == sizeof(wsp_ggml_fp16_t));
+    const int64_t nc = ne00;
+    const int64_t nr = wsp_ggml_nelements(src1); WSP_GGML_UNUSED(nr);
 
-    for (int i = 0; i < nr; ++i) {
-        const int r = ((int32_t *) src1->data)[i];
+    assert(ne0  == nc);
+    assert(ne02 == ne11);
+    assert(nb00 == sizeof(wsp_ggml_fp16_t));
+    assert(wsp_ggml_nrows(dst) == nr);
 
-        for (int j = 0; j < nc; ++j) {
-            wsp_ggml_fp16_t v = ((wsp_ggml_fp16_t *) ((char *) src0->data + r*src0->nb[1]))[j];
-            ((float *) ((char *)  dst->data + i*dst->nb[1]))[j] = WSP_GGML_FP16_TO_FP32(v);
+    // TODO: multi-thread
+    for (int64_t i12 = 0; i12 < ne12; ++i12) {
+        for (int64_t i11 = 0; i11 < ne11; ++i11) {
+            for (int64_t i10 = 0; i10 < ne10; ++i10) {
+                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+                wsp_ggml_fp16_to_fp32_row(
+                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
+            }
         }
     }
 }
@@ -10305,19 +10463,27 @@ static void wsp_ggml_compute_forward_get_rows_f32(
         return;
     }
 
-    const int nc = src0->ne[0];
-    const int nr = wsp_ggml_nelements(src1);
+    WSP_GGML_TENSOR_BINARY_OP_LOCALS
 
-    assert( dst->ne[0] == nc);
-    assert( dst->ne[1] == nr);
-    assert(src0->nb[0] == sizeof(float));
+    const int64_t nc = ne00;
+    const int64_t nr = wsp_ggml_nelements(src1); WSP_GGML_UNUSED(nr);
 
-    for (int i = 0; i < nr; ++i) {
-        const int r = ((int32_t *) src1->data)[i];
+    assert(ne0  == nc);
+    assert(ne02 == ne11);
+    assert(nb00 == sizeof(float));
+    assert(wsp_ggml_nrows(dst) == nr);
 
-        wsp_ggml_vec_cpy_f32(nc,
-                (float *) ((char *)  dst->data + i*dst->nb[1]),
-                (float *) ((char *) src0->data + r*src0->nb[1]));
+    // TODO: multi-thread
+    for (int64_t i12 = 0; i12 < ne12; ++i12) {
+        for (int64_t i11 = 0; i11 < ne11; ++i11) {
+            for (int64_t i10 = 0; i10 < ne10; ++i10) {
+                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+                wsp_ggml_vec_cpy_f32(nc,
+                        (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3),
+                        (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03));
+            }
+        }
     }
 }
 
@@ -10630,20 +10796,25 @@ static void wsp_ggml_compute_forward_diag_mask_zero(
 static void wsp_ggml_compute_forward_soft_max_f32(
         const struct wsp_ggml_compute_params * params,
         const struct wsp_ggml_tensor * src0,
-        struct wsp_ggml_tensor * dst) {
-    WSP_GGML_ASSERT(wsp_ggml_is_contiguous(src0));
-    WSP_GGML_ASSERT(wsp_ggml_is_contiguous(dst));
-    WSP_GGML_ASSERT(wsp_ggml_are_same_shape(src0, dst));
+        const struct wsp_ggml_tensor * src1,
+              struct wsp_ggml_tensor * dst) {
+    assert(wsp_ggml_is_contiguous(dst));
+    assert(wsp_ggml_are_same_shape(src0, dst));
 
     if (params->type == WSP_GGML_TASK_INIT || params->type == WSP_GGML_TASK_FINALIZE) {
         return;
     }
 
+    float scale = 1.0f;
+    memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
+
     // TODO: handle transposed/permuted matrices
 
     const int ith = params->ith;
     const int nth = params->nth;
 
+    const int64_t ne11 = src1 ? src1->ne[1] : 1;
+
     const int nc = src0->ne[0];
     const int nr = wsp_ggml_nrows(src0);
 
@@ -10654,29 +10825,40 @@ static void wsp_ggml_compute_forward_soft_max_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
+    float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith;
+
     for (int i1 = ir0; i1 < ir1; i1++) {
-        float *sp = (float *)((char *) src0->data + i1*src0->nb[1]);
-        float *dp = (float *)((char *)  dst->data +  i1*dst->nb[1]);
+        float * sp = (float *)((char *) src0->data + i1*src0->nb[1]);
+        float * dp = (float *)((char *)  dst->data +  i1*dst->nb[1]);
+
+        // broadcast the mask across rows
+        float * mp = src1 ? (float *)((char *) src1->data + (i1%ne11)*src1->nb[1]) : NULL;
+
+        wsp_ggml_vec_cpy_f32  (nc, wp, sp);
+        wsp_ggml_vec_scale_f32(nc, wp, scale);
+        if (mp) {
+            wsp_ggml_vec_acc_f32(nc, wp, mp);
+        }
 
 #ifndef NDEBUG
         for (int i = 0; i < nc; ++i) {
             //printf("p[%d] = %f\n", i, p[i]);
-            assert(!isnan(sp[i]));
+            assert(!isnan(wp[i]));
         }
 #endif
 
         float max = -INFINITY;
-        wsp_ggml_vec_max_f32(nc, &max, sp);
+        wsp_ggml_vec_max_f32(nc, &max, wp);
 
         wsp_ggml_float sum = 0.0;
 
         uint16_t scvt;
         for (int i = 0; i < nc; i++) {
-            if (sp[i] == -INFINITY) {
+            if (wp[i] == -INFINITY) {
                 dp[i] = 0.0f;
             } else {
-                // const float val = (sp[i] == -INFINITY) ? 0.0 : exp(sp[i] - max);
-                wsp_ggml_fp16_t s = WSP_GGML_FP32_TO_FP16(sp[i] - max);
+                // const float val = (wp[i] == -INFINITY) ? 0.0 : exp(wp[i] - max);
+                wsp_ggml_fp16_t s = WSP_GGML_FP32_TO_FP16(wp[i] - max);
                 memcpy(&scvt, &s, sizeof(scvt));
                 const float val = WSP_GGML_FP16_TO_FP32(wsp_ggml_table_exp_f16[scvt]);
                 sum += (wsp_ggml_float)val;
@@ -10701,11 +10883,12 @@ static void wsp_ggml_compute_forward_soft_max_f32(
 static void wsp_ggml_compute_forward_soft_max(
         const struct wsp_ggml_compute_params * params,
         const struct wsp_ggml_tensor * src0,
-        struct wsp_ggml_tensor * dst) {
+        const struct wsp_ggml_tensor * src1,
+              struct wsp_ggml_tensor * dst) {
     switch (src0->type) {
         case WSP_GGML_TYPE_F32:
             {
-                wsp_ggml_compute_forward_soft_max_f32(params, src0, dst);
+                wsp_ggml_compute_forward_soft_max_f32(params, src0, src1, dst);
             } break;
         default:
             {
@@ -11086,7 +11269,8 @@ static void wsp_ggml_compute_forward_rope_f32(
         const struct wsp_ggml_compute_params * params,
         const struct wsp_ggml_tensor * src0,
         const struct wsp_ggml_tensor * src1,
-        struct wsp_ggml_tensor * dst) {
+        struct wsp_ggml_tensor * dst,
+        const bool forward) {
     if (params->type == WSP_GGML_TASK_INIT || params->type == WSP_GGML_TASK_FINALIZE) {
         return;
     }
@@ -11145,6 +11329,11 @@ static void wsp_ggml_compute_forward_rope_f32(
     const bool is_neox = mode & 2;
     const bool is_glm  = mode & 4;
 
+    // backward process uses inverse rotation by cos and sin.
+    // cos and sin build a rotation matrix, where the inverse is the transpose.
+    // this essentially just switches the sign of sin.
+    const float sin_sign = forward ? 1.0f : -1.0f;
+
     const int32_t * pos = (const int32_t *) src1->data;
 
     for (int64_t i3 = 0; i3 < ne3; i3++) {
@@ -11161,9 +11350,9 @@ static void wsp_ggml_compute_forward_rope_f32(
                     float block_theta = MAX(p - (n_ctx - 2), 0);
                     for (int64_t i0 = 0; i0 < ne0 / 4; i0++) {
                         const float cos_theta = cosf(theta_base);
-                        const float sin_theta = sinf(theta_base);
+                        const float sin_theta = sinf(theta_base) * sin_sign;
                         const float cos_block_theta = cosf(block_theta);
-                        const float sin_block_theta = sinf(block_theta);
+                        const float sin_block_theta = sinf(block_theta) * sin_sign;
 
                         theta_base *= theta_scale;
                         block_theta *= theta_scale;
@@ -11187,6 +11376,7 @@ static void wsp_ggml_compute_forward_rope_f32(
                         rope_yarn(
                             theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta
                         );
+                        sin_theta *= sin_sign;
 
                         // zeta scaling for xPos only:
                         float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), p / xpos_base) : 1.0f;
@@ -11217,6 +11407,7 @@ static void wsp_ggml_compute_forward_rope_f32(
                                 theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor,
                                 &cos_theta, &sin_theta
                             );
+                            sin_theta *= sin_sign;
 
                             theta_base *= theta_scale;
 
@@ -11242,7 +11433,8 @@ static void wsp_ggml_compute_forward_rope_f16(
         const struct wsp_ggml_compute_params * params,
         const struct wsp_ggml_tensor * src0,
         const struct wsp_ggml_tensor * src1,
-        struct wsp_ggml_tensor * dst) {
+        struct wsp_ggml_tensor * dst,
+        const bool forward) {
     if (params->type == WSP_GGML_TASK_INIT || params->type == WSP_GGML_TASK_FINALIZE) {
         return;
     }
@@ -11294,6 +11486,11 @@ static void wsp_ggml_compute_forward_rope_f16(
     const bool is_neox = mode & 2;
     const bool is_glm  = mode & 4;
 
+    // backward process uses inverse rotation by cos and sin.
+    // cos and sin build a rotation matrix, where the inverse is the transpose.
+    // this essentially just switches the sign of sin.
+    const float sin_sign = forward ? 1.0f : -1.0f;
+
     const int32_t * pos = (const int32_t *) src1->data;
 
     for (int64_t i3 = 0; i3 < ne3; i3++) {
@@ -11310,9 +11507,9 @@ static void wsp_ggml_compute_forward_rope_f16(
                     float block_theta = MAX(p - (n_ctx - 2), 0);
                     for (int64_t i0 = 0; i0 < ne0 / 4; i0++) {
                         const float cos_theta = cosf(theta_base);
-                        const float sin_theta = sinf(theta_base);
+                        const float sin_theta = sinf(theta_base) * sin_sign;
                         const float cos_block_theta = cosf(block_theta);
-                        const float sin_block_theta = sinf(block_theta);
+                        const float sin_block_theta = sinf(block_theta) * sin_sign;
 
                         theta_base *= theta_scale;
                         block_theta *= theta_scale;
@@ -11336,6 +11533,7 @@ static void wsp_ggml_compute_forward_rope_f16(
                         rope_yarn(
                             theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta
                         );
+                        sin_theta *= sin_sign;
 
                         theta_base *= theta_scale;
 
@@ -11362,6 +11560,7 @@ static void wsp_ggml_compute_forward_rope_f16(
                                 theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor,
                                 &cos_theta, &sin_theta
                             );
+                            sin_theta *= sin_sign;
 
                             theta_base *= theta_scale;
 
@@ -11391,11 +11590,11 @@ static void wsp_ggml_compute_forward_rope(
     switch (src0->type) {
         case WSP_GGML_TYPE_F16:
             {
-                wsp_ggml_compute_forward_rope_f16(params, src0, src1, dst);
+                wsp_ggml_compute_forward_rope_f16(params, src0, src1, dst, true);
             } break;
         case WSP_GGML_TYPE_F32:
             {
-                wsp_ggml_compute_forward_rope_f32(params, src0, src1, dst);
+                wsp_ggml_compute_forward_rope_f32(params, src0, src1, dst, true);
             } break;
         default:
             {
@@ -11406,726 +11605,106 @@ static void wsp_ggml_compute_forward_rope(
 
 // wsp_ggml_compute_forward_rope_back
 
-static void wsp_ggml_compute_forward_rope_back_f32(
+static void wsp_ggml_compute_forward_rope_back(
         const struct wsp_ggml_compute_params * params,
         const struct wsp_ggml_tensor * src0,
         const struct wsp_ggml_tensor * src1,
         struct wsp_ggml_tensor * dst) {
-
-    if (params->type == WSP_GGML_TASK_INIT || params->type == WSP_GGML_TASK_FINALIZE) {
-        return;
+    switch (src0->type) {
+        case WSP_GGML_TYPE_F16:
+            {
+                wsp_ggml_compute_forward_rope_f16(params, src0, src1, dst, false);
+            } break;
+        case WSP_GGML_TYPE_F32:
+            {
+                wsp_ggml_compute_forward_rope_f32(params, src0, src1, dst, false);
+            } break;
+        default:
+            {
+                WSP_GGML_ASSERT(false);
+            } break;
     }
+}
 
-    // y = rope(x, src1)
-    // dx = rope_back(dy, src1)
-    // src0 is dy, src1 contains options
-
-    float freq_base;
-    float freq_scale;
-
-    // these two only relevant for xPos RoPE:
-    float xpos_base;
-    bool xpos_down;
-
-    //const int n_past = ((int32_t *) dst->op_params)[0];
-    const int n_dims = ((int32_t *) dst->op_params)[1];
-    const int mode   = ((int32_t *) dst->op_params)[2];
-    const int n_ctx  = ((int32_t *) dst->op_params)[3]; UNUSED(n_ctx);
-    memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
-    memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
-    memcpy(&xpos_base,  (int32_t *) dst->op_params + 6, sizeof(float));
-    memcpy(&xpos_down,  (int32_t *) dst->op_params + 7, sizeof(bool));
+// wsp_ggml_compute_forward_conv_transpose_1d
 
-    WSP_GGML_TENSOR_UNARY_OP_LOCALS
+static void wsp_ggml_compute_forward_conv_transpose_1d_f16_f32(
+        const struct wsp_ggml_compute_params * params,
+        const struct wsp_ggml_tensor * src0,
+        const struct wsp_ggml_tensor * src1,
+              struct wsp_ggml_tensor * dst) {
+    WSP_GGML_ASSERT(src0->type == WSP_GGML_TYPE_F16);
+    WSP_GGML_ASSERT(src1->type == WSP_GGML_TYPE_F32);
+    WSP_GGML_ASSERT( dst->type == WSP_GGML_TYPE_F32);
 
-    //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
-    //printf("n_past = %d, ne2 = %d\n", n_past, ne2);
+    int64_t t0 = wsp_ggml_perf_time_us();
+    UNUSED(t0);
 
-    assert(nb0 == sizeof(float));
+    WSP_GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int nr = wsp_ggml_nrows(dst);
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
+    const int nk = ne00*ne01*ne02;
 
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
+    WSP_GGML_ASSERT(nb00 == sizeof(wsp_ggml_fp16_t));
+    WSP_GGML_ASSERT(nb10 == sizeof(float));
 
-    // row index used to determine which thread to use
-    int ir = 0;
+    if (params->type == WSP_GGML_TASK_INIT) {
+        memset(params->wdata, 0, params->wsize);
 
-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+        // permute kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout)
+        {
+            wsp_ggml_fp16_t * const wdata = (wsp_ggml_fp16_t *) params->wdata + 0;
 
-    const bool is_neox = mode & 2;
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
+                    const wsp_ggml_fp16_t * const src = (wsp_ggml_fp16_t *)((char *) src0->data + i02*nb02 + i01*nb01);
+                    wsp_ggml_fp16_t * dst_data = wdata + i01*ne00*ne02;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        dst_data[i00*ne02 + i02] = src[i00];
+                    }
+                }
+            }
+        }
 
-    const int32_t * pos = (const int32_t *) src1->data;
+        // permute source data (src1) from (L x Cin) to (Cin x L)
+        {
+            wsp_ggml_fp16_t * const wdata = (wsp_ggml_fp16_t *) params->wdata + nk;
+            wsp_ggml_fp16_t * dst_data = wdata;
 
-    for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = 0; i2 < ne2; i2++) {
-            const int64_t p = pos[i2];
-            for (int64_t i1 = 0; i1 < ne1; i1++) {
-                if (ir++ < ir0) continue;
-                if (ir   > ir1) break;
+            for (int64_t i11 = 0; i11 < ne11; i11++) {
+                const float * const src = (float *)((char *) src1->data + i11*nb11);
+                for (int64_t i10 = 0; i10 < ne10; i10++) {
+                    dst_data[i10*ne11 + i11] = WSP_GGML_FP32_TO_FP16(src[i10]);
+                }
+            }
+        }
 
-                float theta_base = freq_scale * (float)p;
+        // need to zero dst since we are accumulating into it
+        memset(dst->data, 0, wsp_ggml_nbytes(dst));
 
-                if (!is_neox) {
-                    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
-                        const float cos_theta = cosf(theta_base);
-                        const float sin_theta = sinf(theta_base);
+        return;
+    }
 
-                        // zeta scaling for xPos only:
-                        float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), p / xpos_base) : 1.0f;
-                        if (xpos_down) zeta = 1.0f / zeta;
+    if (params->type == WSP_GGML_TASK_FINALIZE) {
+        return;
+    }
 
-                        theta_base *= theta_scale;
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
 
-                        const float * const dy  = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                              float *       dx  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+    // total rows in dst
+    const int nr = ne1;
 
-                        const float dy0 = dy[0];
-                        const float dy1 = dy[1];
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
 
-                        dx[0] =   dy0*cos_theta*zeta + dy1*sin_theta*zeta;
-                        dx[1] = - dy0*sin_theta*zeta + dy1*cos_theta*zeta;
-                    }
-                } else {
-                    for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
-                        for (int64_t ic = 0; ic < n_dims; ic += 2) {
-                            const float cos_theta = cosf(theta_base);
-                            const float sin_theta = sinf(theta_base);
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
 
-                            theta_base *= theta_scale;
-
-                            const int64_t i0 = ib*n_dims + ic/2;
-
-                            const float * const dy  = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                                  float *       dx  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                            const float dy0 = dy[0];
-                            const float dy1 = dy[n_dims/2];
-
-                            dx[0]        =   dy0*cos_theta + dy1*sin_theta;
-                            dx[n_dims/2] = - dy0*sin_theta + dy1*cos_theta;
-                        }
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void wsp_ggml_compute_forward_rope_back_f16(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-        struct wsp_ggml_tensor * dst) {
-
-    if (params->type == WSP_GGML_TASK_INIT || params->type == WSP_GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    // y = rope(x, src1)
-    // dx = rope_back(dy, src1)
-    // src0 is dy, src1 contains options
-
-    //const int n_past = ((int32_t *) dst->op_params)[0];
-    const int n_dims = ((int32_t *) dst->op_params)[1];
-    const int mode   = ((int32_t *) dst->op_params)[2];
-
-    WSP_GGML_TENSOR_UNARY_OP_LOCALS
-
-    //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
-    //printf("n_past = %d, ne2 = %d\n", n_past, ne2);
-
-    assert(nb0 == sizeof(wsp_ggml_fp16_t));
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int nr = wsp_ggml_nrows(dst);
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    // row index used to determine which thread to use
-    int ir = 0;
-
-    const float theta_scale = powf(10000.0, -2.0f/n_dims);
-
-    const bool is_neox = mode & 2;
-
-    const int32_t * pos = (const int32_t *) src1->data;
-
-    for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = 0; i2 < ne2; i2++) {
-            const int64_t p = pos[i2];
-            for (int64_t i1 = 0; i1 < ne1; i1++) {
-                if (ir++ < ir0) continue;
-                if (ir   > ir1) break;
-
-                float theta_base = (float)p;
-
-                if (!is_neox) {
-                    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
-                        const float cos_theta = cosf(theta_base);
-                        const float sin_theta = sinf(theta_base);
-
-                        theta_base *= theta_scale;
-
-                        const wsp_ggml_fp16_t * const dy  = (wsp_ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                              wsp_ggml_fp16_t *       dx  = (wsp_ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                        const float dy0 = WSP_GGML_FP16_TO_FP32(dy[0]);
-                        const float dy1 = WSP_GGML_FP16_TO_FP32(dy[1]);
-
-                        dx[0] = WSP_GGML_FP32_TO_FP16( dy0*cos_theta + dy1*sin_theta);
-                        dx[1] = WSP_GGML_FP32_TO_FP16(-dy0*sin_theta + dy1*cos_theta);
-                    }
-                } else {
-                    for (int64_t ib = 0; ib < ne0/n_dims; ++ib) {
-                        for (int64_t ic = 0; ic < n_dims; ic += 2) {
-                            const float cos_theta = cosf(theta_base);
-                            const float sin_theta = sinf(theta_base);
-
-                            theta_base *= theta_scale;
-
-                            const int64_t i0 = ib*n_dims + ic/2;
-
-                            const wsp_ggml_fp16_t * const dy  = (wsp_ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                                  wsp_ggml_fp16_t *       dx  = (wsp_ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                            const float dy0 = WSP_GGML_FP16_TO_FP32(dy[0]);
-                            const float dy1 = WSP_GGML_FP16_TO_FP32(dy[n_dims/2]);
-
-                            dx[0]        = WSP_GGML_FP32_TO_FP16( dy0*cos_theta + dy1*sin_theta);
-                            dx[n_dims/2] = WSP_GGML_FP32_TO_FP16(-dy0*sin_theta + dy1*cos_theta);
-                        }
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void wsp_ggml_compute_forward_rope_back(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-        struct wsp_ggml_tensor * dst) {
-    switch (src0->type) {
-        case WSP_GGML_TYPE_F16:
-            {
-                wsp_ggml_compute_forward_rope_back_f16(params, src0, src1, dst);
-            } break;
-        case WSP_GGML_TYPE_F32:
-            {
-                wsp_ggml_compute_forward_rope_back_f32(params, src0, src1, dst);
-            } break;
-        default:
-            {
-                WSP_GGML_ASSERT(false);
-            } break;
-    }
-}
-
-// wsp_ggml_compute_forward_conv_1d
-
-static void wsp_ggml_compute_forward_conv_1d_f16_f32(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    WSP_GGML_ASSERT(src0->type == WSP_GGML_TYPE_F16);
-    WSP_GGML_ASSERT(src1->type == WSP_GGML_TYPE_F32);
-    WSP_GGML_ASSERT( dst->type == WSP_GGML_TYPE_F32);
-
-    int64_t t0 = wsp_ggml_perf_time_us();
-    UNUSED(t0);
-
-    WSP_GGML_TENSOR_BINARY_OP_LOCALS
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int nk = ne00;
-
-    // size of the convolution row - the kernel size unrolled across all input channels
-    const int ew0 = nk*ne01;
-
-    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t p0 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t d0 = ((const int32_t*)(dst->op_params))[2];
-
-    WSP_GGML_ASSERT(nb00 == sizeof(wsp_ggml_fp16_t));
-    WSP_GGML_ASSERT(nb10 == sizeof(float));
-
-    if (params->type == WSP_GGML_TASK_INIT) {
-        memset(params->wdata, 0, params->wsize);
-
-        wsp_ggml_fp16_t * const wdata = (wsp_ggml_fp16_t *) params->wdata + 0;
-
-        for (int64_t i11 = 0; i11 < ne11; i11++) {
-            const float * const src = (float *)((char *) src1->data + i11*nb11);
-            wsp_ggml_fp16_t * dst_data = wdata;
-
-            for (int64_t i0 = 0; i0 < ne0; i0++) {
-                for (int64_t ik = 0; ik < nk; ik++) {
-                    const int idx0 = i0*s0 + ik*d0 - p0;
-
-                    if(!(idx0 < 0 || idx0 >= ne10)) {
-                        dst_data[i0*ew0 + i11*nk + ik] = WSP_GGML_FP32_TO_FP16(src[idx0]);
-                    }
-                }
-            }
-        }
-
-        return;
-    }
-
-    if (params->type == WSP_GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    // total rows in dst
-    const int nr = ne2;
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    wsp_ggml_fp16_t * const wdata = (wsp_ggml_fp16_t *) params->wdata + 0;
-
-    for (int i2 = 0; i2 < ne2; i2++) {
-        for (int i1 = ir0; i1 < ir1; i1++) {
-            float * dst_data = (float *)((char *) dst->data + i2*nb2 + i1*nb1);
-
-            for (int i0 = 0; i0 < ne0; i0++) {
-                wsp_ggml_vec_dot_f16(ew0, dst_data + i0,
-                        (wsp_ggml_fp16_t *) ((char *) src0->data + i1*nb02),
-                        (wsp_ggml_fp16_t *)                wdata + i2*nb2 + i0*ew0);
-            }
-        }
-    }
-}
-
-static void wsp_ggml_compute_forward_conv_1d_f32(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    WSP_GGML_ASSERT(src0->type == WSP_GGML_TYPE_F32);
-    WSP_GGML_ASSERT(src1->type == WSP_GGML_TYPE_F32);
-    WSP_GGML_ASSERT( dst->type == WSP_GGML_TYPE_F32);
-
-    int64_t t0 = wsp_ggml_perf_time_us();
-    UNUSED(t0);
-
-    WSP_GGML_TENSOR_BINARY_OP_LOCALS
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int nk = ne00;
-
-    const int ew0 = nk*ne01;
-
-    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t p0 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t d0 = ((const int32_t*)(dst->op_params))[2];
-
-    WSP_GGML_ASSERT(nb00 == sizeof(float));
-    WSP_GGML_ASSERT(nb10 == sizeof(float));
-
-    if (params->type == WSP_GGML_TASK_INIT) {
-        memset(params->wdata, 0, params->wsize);
-
-        float * const wdata = (float *) params->wdata + 0;
-
-        for (int64_t i11 = 0; i11 < ne11; i11++) {
-            const float * const src = (float *)((char *) src1->data + i11*nb11);
-            float * dst_data = wdata;
-
-            for (int64_t i0 = 0; i0 < ne0; i0++) {
-                for (int64_t ik = 0; ik < nk; ik++) {
-                    const int idx0 = i0*s0 + ik*d0 - p0;
-
-                    if(!(idx0 < 0 || idx0 >= ne10)) {
-                        dst_data[i0*ew0 + i11*nk + ik] = src[idx0];
-                    }
-                }
-            }
-        }
-
-        return;
-    }
-
-    if (params->type == WSP_GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    // total rows in dst
-    const int nr = ne02;
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    float * const wdata = (float *) params->wdata + 0;
-
-    for (int i2 = 0; i2 < ne2; i2++) {
-        for (int i1 = ir0; i1 < ir1; i1++) {
-            float * dst_data = (float *)((char *) dst->data + i2*nb2 + i1*nb1);
-
-            for (int i0 = 0; i0 < ne0; i0++) {
-                wsp_ggml_vec_dot_f32(ew0, dst_data + i0,
-                        (float *) ((char *) src0->data + i1*nb02),
-                        (float *)                wdata + i2*nb2 + i0*ew0);
-            }
-        }
-    }
-}
-
-// TODO: reuse wsp_ggml_mul_mat or implement wsp_ggml_im2col and remove stage_0 and stage_1
-static void gemm_f16_out_f32(int64_t m, int64_t n, int64_t k,
-                             wsp_ggml_fp16_t * A,
-                             wsp_ggml_fp16_t * B,
-                             float * C,
-                             const int ith, const int nth) {
-    // does not seem to make a difference
-    int64_t m0, m1, n0, n1;
-    // patches per thread
-    if (m > n) {
-        n0 = 0;
-        n1 = n;
-
-        // total patches in dst
-        const int np = m;
-
-        // patches per thread
-        const int dp = (np + nth - 1)/nth;
-
-        // patch range for this thread
-        m0 = dp*ith;
-        m1 = MIN(m0 + dp, np);
-    } else {
-        m0 = 0;
-        m1 = m;
-
-        // total patches in dst
-        const int np = n;
-
-        // patches per thread
-        const int dp = (np + nth - 1)/nth;
-
-        // patch range for this thread
-        n0 = dp*ith;
-        n1 = MIN(n0 + dp, np);
-    }
-
-    // block-tiling attempt
-    int64_t blck_n = 16;
-    int64_t blck_m = 16;
-
-    // int64_t CACHE_SIZE = 2 * 1024 * 1024; // 2MB
-    // int64_t blck_size = CACHE_SIZE / (sizeof(float) + 2 * sizeof(wsp_ggml_fp16_t) * K);
-    // if (blck_size > 0) {
-    //     blck_0 = 4;
-    //     blck_1 = blck_size / blck_0;
-    //     if (blck_1 < 0) {
-    //         blck_1 = 1;
-    //     }
-    //     // blck_0 = (int64_t)sqrt(blck_size);
-    //     // blck_1 = blck_0;
-    // }
-    // // printf("%zd %zd %zd %zd\n", blck_size, K, blck_0, blck_1);
-
-    for (int j = n0; j < n1; j+=blck_n) {
-        for (int i = m0; i < m1; i+=blck_m) {
-            // printf("i j k => %d %d %d\n", i, j, K);
-            for (int ii = i; ii < i + blck_m && ii < m1; ii++) {
-                for (int jj = j; jj < j + blck_n && jj < n1; jj++) {
-                    wsp_ggml_vec_dot_f16(k,
-                                    C + ii*n + jj,
-                                    A + ii * k,
-                                    B + jj * k);
-                }
-            }
-        }
-    }
-}
-
-// src0: kernel [OC, IC, K]
-// src1: signal [N, IC, IL]
-// dst:  result [N, OL, IC*K]
-static void wsp_ggml_compute_forward_conv_1d_stage_0_f32(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    WSP_GGML_ASSERT(src0->type == WSP_GGML_TYPE_F16);
-    WSP_GGML_ASSERT(src1->type == WSP_GGML_TYPE_F32);
-    WSP_GGML_ASSERT( dst->type == WSP_GGML_TYPE_F16);
-
-    int64_t t0 = wsp_ggml_perf_time_us();
-    UNUSED(t0);
-
-    WSP_GGML_TENSOR_BINARY_OP_LOCALS;
-
-    const int64_t N  = ne12;
-    const int64_t IC = ne11;
-    const int64_t IL = ne10;
-
-    const int64_t K = ne00;
-
-    const int64_t OL = ne1;
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t p0 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t d0 = ((const int32_t*)(dst->op_params))[2];
-
-    WSP_GGML_ASSERT(nb00 == sizeof(wsp_ggml_fp16_t));
-    WSP_GGML_ASSERT(nb10 == sizeof(float));
-
-    if (params->type == WSP_GGML_TASK_INIT) {
-        memset(dst->data, 0, wsp_ggml_nbytes(dst));
-        return;
-    }
-
-    if (params->type == WSP_GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    // im2col: [N, IC, IL] => [N, OL, IC*K]
-    {
-        wsp_ggml_fp16_t * const wdata = (wsp_ggml_fp16_t *) dst->data;
-
-        for (int64_t in = 0; in < N; in++) {
-            for (int64_t iol = 0; iol < OL; iol++) {
-                for (int64_t iic = ith; iic < IC; iic+=nth) {
-
-                    // micro kernel
-                    wsp_ggml_fp16_t * dst_data = wdata + (in*OL + iol)*(IC*K); // [IC, K]
-                    const float * const src_data = (float *)((char *) src1->data + in*nb12 + iic*nb11); // [IL]
-
-                    for (int64_t ik = 0; ik < K; ik++) {
-                        const int64_t iil = iol*s0 + ik*d0 - p0;
-
-                        if (!(iil < 0 || iil >= IL)) {
-                            dst_data[iic*K + ik] = WSP_GGML_FP32_TO_FP16(src_data[iil]);
-                        }
-                    }
-                }
-            }
-        }
-    }
-}
-
-// gemm: [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K]
-// src0: [OC, IC, K]
-// src1: [N, OL, IC * K]
-// result: [N, OC, OL]
-static void wsp_ggml_compute_forward_conv_1d_stage_1_f16(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    WSP_GGML_ASSERT(src0->type == WSP_GGML_TYPE_F16);
-    WSP_GGML_ASSERT(src1->type == WSP_GGML_TYPE_F16);
-    WSP_GGML_ASSERT( dst->type == WSP_GGML_TYPE_F32);
-
-    int64_t t0 = wsp_ggml_perf_time_us();
-    UNUSED(t0);
-
-    if (params->type == WSP_GGML_TASK_INIT) {
-        return;
-    }
-
-    if (params->type == WSP_GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    WSP_GGML_TENSOR_BINARY_OP_LOCALS;
-
-    WSP_GGML_ASSERT(nb00 == sizeof(wsp_ggml_fp16_t));
-    WSP_GGML_ASSERT(nb10 == sizeof(wsp_ggml_fp16_t));
-    WSP_GGML_ASSERT(nb0  == sizeof(float));
-
-    const int N = ne12;
-    const int OL = ne11;
-
-    const int OC = ne02;
-    const int IC = ne01;
-    const int K  = ne00;
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    int64_t m = OC;
-    int64_t n = OL;
-    int64_t k = IC * K;
-
-    // [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K]
-    for (int i = 0; i < N; i++) {
-        wsp_ggml_fp16_t * A = (wsp_ggml_fp16_t *)src0->data; // [m, k]
-        wsp_ggml_fp16_t * B = (wsp_ggml_fp16_t *)src1->data + i * m * k; // [n, k]
-        float * C = (float *)dst->data + i * m * n; // [m, n]
-
-        gemm_f16_out_f32(m, n, k, A, B, C, ith, nth);
-    }
-}
-
-static void wsp_ggml_compute_forward_conv_1d(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    switch(src0->type) {
-        case WSP_GGML_TYPE_F16:
-            {
-                wsp_ggml_compute_forward_conv_1d_f16_f32(params, src0, src1, dst);
-            } break;
-        case WSP_GGML_TYPE_F32:
-            {
-                wsp_ggml_compute_forward_conv_1d_f32(params, src0, src1, dst);
-            } break;
-        default:
-            {
-                WSP_GGML_ASSERT(false);
-            } break;
-    }
-}
-
-static void wsp_ggml_compute_forward_conv_1d_stage_0(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    switch(src0->type) {
-        case WSP_GGML_TYPE_F16:
-            {
-                wsp_ggml_compute_forward_conv_1d_stage_0_f32(params, src0, src1, dst);
-            } break;
-        default:
-            {
-                WSP_GGML_ASSERT(false);
-            } break;
-    }
-}
-
-static void wsp_ggml_compute_forward_conv_1d_stage_1(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    switch(src0->type) {
-        case WSP_GGML_TYPE_F16:
-            {
-                wsp_ggml_compute_forward_conv_1d_stage_1_f16(params, src0, src1, dst);
-            } break;
-        default:
-            {
-                WSP_GGML_ASSERT(false);
-            } break;
-    }
-}
-
-// wsp_ggml_compute_forward_conv_transpose_1d
-
-static void wsp_ggml_compute_forward_conv_transpose_1d_f16_f32(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    WSP_GGML_ASSERT(src0->type == WSP_GGML_TYPE_F16);
-    WSP_GGML_ASSERT(src1->type == WSP_GGML_TYPE_F32);
-    WSP_GGML_ASSERT( dst->type == WSP_GGML_TYPE_F32);
-
-    int64_t t0 = wsp_ggml_perf_time_us();
-    UNUSED(t0);
-
-    WSP_GGML_TENSOR_BINARY_OP_LOCALS
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int nk = ne00*ne01*ne02;
-
-    WSP_GGML_ASSERT(nb00 == sizeof(wsp_ggml_fp16_t));
-    WSP_GGML_ASSERT(nb10 == sizeof(float));
-
-    if (params->type == WSP_GGML_TASK_INIT) {
-        memset(params->wdata, 0, params->wsize);
-
-        // permute kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout)
-        {
-            wsp_ggml_fp16_t * const wdata = (wsp_ggml_fp16_t *) params->wdata + 0;
-
-            for (int64_t i02 = 0; i02 < ne02; i02++) {
-                for (int64_t i01 = 0; i01 < ne01; i01++) {
-                    const wsp_ggml_fp16_t * const src = (wsp_ggml_fp16_t *)((char *) src0->data + i02*nb02 + i01*nb01);
-                    wsp_ggml_fp16_t * dst_data = wdata + i01*ne00*ne02;
-                    for (int64_t i00 = 0; i00 < ne00; i00++) {
-                        dst_data[i00*ne02 + i02] = src[i00];
-                    }
-                }
-            }
-        }
-
-        // permute source data (src1) from (L x Cin) to (Cin x L)
-        {
-            wsp_ggml_fp16_t * const wdata = (wsp_ggml_fp16_t *) params->wdata + nk;
-            wsp_ggml_fp16_t * dst_data = wdata;
-
-            for (int64_t i11 = 0; i11 < ne11; i11++) {
-                const float * const src = (float *)((char *) src1->data + i11*nb11);
-                for (int64_t i10 = 0; i10 < ne10; i10++) {
-                    dst_data[i10*ne11 + i11] = WSP_GGML_FP32_TO_FP16(src[i10]);
-                }
-            }
-        }
-
-        // need to zero dst since we are accumulating into it
-        memset(dst->data, 0, wsp_ggml_nbytes(dst));
-
-        return;
-    }
-
-    if (params->type == WSP_GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-
-    // total rows in dst
-    const int nr = ne1;
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    wsp_ggml_fp16_t * const wdata     = (wsp_ggml_fp16_t *) params->wdata + 0;
-    wsp_ggml_fp16_t * const wdata_src = wdata + nk;
+    wsp_ggml_fp16_t * const wdata     = (wsp_ggml_fp16_t *) params->wdata + 0;
+    wsp_ggml_fp16_t * const wdata_src = wdata + nk;
 
     for (int i1 = ir0; i1 < ir1; i1++) {
         float * dst_data = (float *)((char *) dst->data + i1*nb1);
@@ -12258,12 +11837,10 @@ static void wsp_ggml_compute_forward_conv_transpose_1d(
     }
 }
 
-// wsp_ggml_compute_forward_conv_2d
-
 // src0: kernel [OC, IC, KH, KW]
 // src1: image [N, IC, IH, IW]
 // dst:  result [N, OH, OW, IC*KH*KW]
-static void wsp_ggml_compute_forward_conv_2d_stage_0_f32(
+static void wsp_ggml_compute_forward_im2col_f16(
         const struct wsp_ggml_compute_params * params,
         const struct wsp_ggml_tensor * src0,
         const struct wsp_ggml_tensor * src1,
@@ -12277,218 +11854,35 @@ static void wsp_ggml_compute_forward_conv_2d_stage_0_f32(
 
     WSP_GGML_TENSOR_BINARY_OP_LOCALS;
 
-    const int64_t N = ne13;
-    const int64_t IC = ne12;
-    const int64_t IH = ne11;
-    const int64_t IW = ne10;
-
-    // const int64_t OC = ne03;
-    // const int64_t IC = ne02;
-    const int64_t KH = ne01;
-    const int64_t KW = ne00;
-
-    const int64_t OH = ne2;
-    const int64_t OW = ne1;
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
-    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
-    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
-    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
-
-    WSP_GGML_ASSERT(nb00 == sizeof(wsp_ggml_fp16_t));
-    WSP_GGML_ASSERT(nb10 == sizeof(float));
-
-    if (params->type == WSP_GGML_TASK_INIT) {
-        memset(dst->data, 0, wsp_ggml_nbytes(dst));
-        return;
-    }
-
-    if (params->type == WSP_GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
-    {
-        wsp_ggml_fp16_t * const wdata = (wsp_ggml_fp16_t *) dst->data;
-
-        for (int64_t in = 0; in < N; in++) {
-            for (int64_t ioh = 0; ioh < OH; ioh++) {
-                for (int64_t iow = 0; iow < OW; iow++) {
-                    for (int64_t iic = ith; iic < IC; iic+=nth) {
-
-                        // micro kernel
-                        wsp_ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW]
-                        const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW]
-
-                        for (int64_t ikh = 0; ikh < KH; ikh++) {
-                            for (int64_t ikw = 0; ikw < KW; ikw++) {
-                                const int64_t iiw = iow*s0 + ikw*d0 - p0;
-                                const int64_t iih = ioh*s1 + ikh*d1 - p1;
-
-                                if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) {
-                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = WSP_GGML_FP32_TO_FP16(src_data[iih*IW + iiw]);
-                                }
-                            }
-                        }
-                    }
-                }
-            }
-        }
-    }
-}
-
-// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW]
-// src0: [OC, IC, KH, KW]
-// src1: [N, OH, OW, IC * KH * KW]
-// result: [N, OC, OH, OW]
-static void wsp_ggml_compute_forward_conv_2d_stage_1_f16(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    WSP_GGML_ASSERT(src0->type == WSP_GGML_TYPE_F16);
-    WSP_GGML_ASSERT(src1->type == WSP_GGML_TYPE_F16);
-    WSP_GGML_ASSERT( dst->type == WSP_GGML_TYPE_F32);
-
-    int64_t t0 = wsp_ggml_perf_time_us();
-    UNUSED(t0);
-
-    if (params->type == WSP_GGML_TASK_INIT) {
-        return;
-    }
-
-    if (params->type == WSP_GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    WSP_GGML_TENSOR_BINARY_OP_LOCALS;
-
-    WSP_GGML_ASSERT(nb00 == sizeof(wsp_ggml_fp16_t));
-    WSP_GGML_ASSERT(nb10 == sizeof(wsp_ggml_fp16_t));
-    WSP_GGML_ASSERT(nb0  == sizeof(float));
-
-    const int N = ne13;
-    const int OH = ne12;
-    const int OW = ne11;
-
-    const int OC = ne03;
-    const int IC = ne02;
-    const int KH = ne01;
-    const int KW = ne00;
+    const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t *)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t *)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t *)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t *)(dst->op_params))[5];
+    const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1;
 
     const int ith = params->ith;
     const int nth = params->nth;
 
-    int64_t m = OC;
-    int64_t n = OH * OW;
-    int64_t k = IC * KH * KW;
-
-    // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW]
-    for (int i = 0; i < N; i++) {
-        wsp_ggml_fp16_t * A = (wsp_ggml_fp16_t *)src0->data; // [m, k]
-        wsp_ggml_fp16_t * B = (wsp_ggml_fp16_t *)src1->data + i * m * k; // [n, k]
-        float * C = (float *)dst->data + i * m * n; // [m, n]
-
-        gemm_f16_out_f32(m, n, k, A, B, C, ith, nth);
-    }
-}
-
-static void wsp_ggml_compute_forward_conv_2d_f16_f32(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    WSP_GGML_ASSERT(src0->type == WSP_GGML_TYPE_F16);
-    WSP_GGML_ASSERT(src1->type == WSP_GGML_TYPE_F32);
-    WSP_GGML_ASSERT( dst->type == WSP_GGML_TYPE_F32);
-
-    int64_t t0 = wsp_ggml_perf_time_us();
-    UNUSED(t0);
-
-    WSP_GGML_TENSOR_BINARY_OP_LOCALS
-
-    // src1: image [N, IC, IH, IW]
-    // src0: kernel [OC, IC, KH, KW]
-    // dst:  result [N, OC, OH, OW]
-    // ne12: IC
-    // ne0: OW
-    // ne1: OH
-    // nk0: KW
-    // nk1: KH
-    // ne13: N
-
-    const int N = ne13;
-    const int IC = ne12;
-    const int IH = ne11;
-    const int IW = ne10;
-
-    const int OC = ne03;
-    // const int IC = ne02;
-    const int KH = ne01;
-    const int KW = ne00;
-
-    const int OH = ne1;
-    const int OW = ne0;
-
-    const int ith = params->ith;
-    const int nth = params->nth;
+    const int64_t N  = is_2D ? ne13 : ne12;
+    const int64_t IC = is_2D ? ne12 : ne11;
+    const int64_t IH = is_2D ? ne11 : 1;
+    const int64_t IW = ne10;
 
-    // const int nk0 = ne00;
-    // const int nk1 = ne01;
+    const int64_t KH = is_2D ? ne01 : 1;
+    const int64_t KW = ne00;
 
-    // size of the convolution row - the kernel size unrolled across all channels
-    // const int ew0 = nk0*nk1*ne02;
-    // ew0: IC*KH*KW
+    const int64_t OH = is_2D ? ne2 : 1;
+    const int64_t OW = ne1;
 
-    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
-    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
-    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
-    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
+    int ofs0 = is_2D ? nb13 : nb12;
+    int ofs1 = is_2D ? nb12 : nb11;
 
     WSP_GGML_ASSERT(nb00 == sizeof(wsp_ggml_fp16_t));
     WSP_GGML_ASSERT(nb10 == sizeof(float));
 
     if (params->type == WSP_GGML_TASK_INIT) {
-        memset(params->wdata, 0, params->wsize);
-
-        // prepare source data (src1)
-        // im2col: [N, IC, IH, IW] => [N*OH*OW, IC*KH*KW]
-
-        {
-            wsp_ggml_fp16_t * const wdata = (wsp_ggml_fp16_t *) params->wdata + 0;
-
-            for (int in = 0; in < N; in++) {
-                for (int iic = 0; iic < IC; iic++) {
-                    for (int ioh = 0; ioh < OH; ioh++) {
-                        for (int iow = 0; iow < OW; iow++) {
-
-                            // micro kernel
-                            wsp_ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW]
-                            const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW]
-
-                            for (int ikh = 0; ikh < KH; ikh++) {
-                                for (int ikw = 0; ikw < KW; ikw++) {
-                                    const int iiw = iow*s0 + ikw*d0 - p0;
-                                    const int iih = ioh*s1 + ikh*d1 - p1;
-
-                                    if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) {
-                                        dst_data[iic*(KH*KW) + ikh*KW + ikw] = WSP_GGML_FP32_TO_FP16(src_data[iih*IW + iiw]);
-                                    }
-                                }
-                            }
-                        }
-                    }
-                }
-            }
-        }
-
         return;
     }
 
@@ -12496,69 +11890,39 @@ static void wsp_ggml_compute_forward_conv_2d_f16_f32(
         return;
     }
 
-    wsp_ggml_fp16_t * const wdata = (wsp_ggml_fp16_t *) params->wdata + 0;
-    // wdata: [N*OH*OW, IC*KH*KW]
-    // dst: result [N, OC, OH, OW]
-    // src0: kernel [OC, IC, KH, KW]
-
-    int64_t m = OC;
-    int64_t n = OH * OW;
-    int64_t k = IC * KH * KW;
-
-    // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW]
-    for (int i = 0; i < N; i++) {
-        wsp_ggml_fp16_t * A = (wsp_ggml_fp16_t *)src0->data; // [m, k]
-        wsp_ggml_fp16_t * B = (wsp_ggml_fp16_t *)wdata + i * m * k; // [n, k]
-        float * C = (float *)dst->data + i * m * n; // [m * k]
-
-        gemm_f16_out_f32(m, n, k, A, B, C, ith, nth);
-    }
-}
-
-static void wsp_ggml_compute_forward_conv_2d(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    switch (src0->type) {
-        case WSP_GGML_TYPE_F16:
-            {
-                wsp_ggml_compute_forward_conv_2d_f16_f32(params, src0, src1, dst);
-            } break;
-        case WSP_GGML_TYPE_F32:
-            {
-                //wsp_ggml_compute_forward_conv_2d_f32(params, src0, src1, dst);
-                WSP_GGML_ASSERT(false);
-            } break;
-        default:
-            {
-                WSP_GGML_ASSERT(false);
-            } break;
-    }
-}
-
-static void wsp_ggml_compute_forward_conv_2d_stage_0(
-        const struct wsp_ggml_compute_params * params,
-        const struct wsp_ggml_tensor * src0,
-        const struct wsp_ggml_tensor * src1,
-              struct wsp_ggml_tensor * dst) {
-    switch (src0->type) {
-        case WSP_GGML_TYPE_F16:
-            {
-                wsp_ggml_compute_forward_conv_2d_stage_0_f32(params, src0, src1, dst);
-            } break;
-        case WSP_GGML_TYPE_F32:
-            {
-                WSP_GGML_ASSERT(false);
-            } break;
-        default:
-            {
-                WSP_GGML_ASSERT(false);
-            } break;
+    // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
+    {
+        wsp_ggml_fp16_t * const wdata = (wsp_ggml_fp16_t *) dst->data;
+
+        for (int64_t in = 0; in < N; in++) {
+            for (int64_t ioh = 0; ioh < OH; ioh++) { // 1
+                for (int64_t iow = 0; iow < OW; iow++) {
+                    for (int64_t iic = ith; iic < IC; iic += nth) {
+
+                        // micro kernel
+                        wsp_ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW]
+                        const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW]
+
+                        for (int64_t ikh = 0; ikh < KH; ikh++) {  // 1
+                            for (int64_t ikw = 0; ikw < KW; ikw++) {
+                                const int64_t iiw = iow*s0 + ikw*d0 - p0;
+                                const int64_t iih = ioh*s1 + ikh*d1 - p1;
+
+                                if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0;
+                                } else {
+                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = WSP_GGML_FP32_TO_FP16(src_data[iih*IW + iiw]);
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+        }
     }
 }
 
-static void wsp_ggml_compute_forward_conv_2d_stage_1(
+static void wsp_ggml_compute_forward_im2col(
         const struct wsp_ggml_compute_params * params,
         const struct wsp_ggml_tensor * src0,
         const struct wsp_ggml_tensor * src1,
@@ -12566,7 +11930,7 @@ static void wsp_ggml_compute_forward_conv_2d_stage_1(
     switch (src0->type) {
         case WSP_GGML_TYPE_F16:
             {
-                wsp_ggml_compute_forward_conv_2d_stage_1_f16(params, src0, src1, dst);
+                wsp_ggml_compute_forward_im2col_f16(params, src0, src1, dst);
             } break;
         case WSP_GGML_TYPE_F32:
             {
@@ -12839,6 +12203,7 @@ static void wsp_ggml_compute_forward_upscale_f32(
     WSP_GGML_ASSERT(src0->nb[0] == sizeof(float));
 
     const int ith = params->ith;
+    const int nth = params->nth;
 
     WSP_GGML_TENSOR_UNARY_OP_LOCALS
 
@@ -12846,16 +12211,17 @@ static void wsp_ggml_compute_forward_upscale_f32(
 
     // TODO: optimize
 
-    for (int i03 = 0; i03 < ne03; i03++) {
-        for (int i02 = ith; i02 < ne02; i02++) {
-            for (int m = 0; m < dst->ne[1]; m++) {
-                int i01 = m / scale_factor;
-                for (int n = 0; n < dst->ne[0]; n++) {
-                    int i00 = n / scale_factor;
-
-                    const float * x = (float *)((char *) src0->data + i00 * nb00 +i01 * nb01 + i02 * nb02 + i03 * nb03);
+    for (int64_t i3 = 0; i3 < ne3; i3++) {
+        const int64_t i03 = i3;
+        for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+            const int64_t i02 = i2;
+            for (int64_t i1 = 0; i1 < ne1; i1++) {
+                const int64_t i01 = i1 / scale_factor;
+                for (int64_t i0 = 0; i0 < ne0; i0++) {
+                    const int64_t i00 = i0 / scale_factor;
 
-                    float * y = (float *)((char *) dst->data + n * dst->nb[0] + m * dst->nb[1] + i02 * dst->nb[2] + i03 * dst->nb[3]);
+                    const float * x = (float *)((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                          float * y = (float *)((char *)  dst->data +  i0*nb0  +  i1*nb1  +  i2*nb2  +  i3*nb3);
 
                     *y = *x;
                 }
@@ -12880,6 +12246,125 @@ static void wsp_ggml_compute_forward_upscale(
     }
 }
 
+// wsp_ggml_compute_forward_pad
+
+static void wsp_ggml_compute_forward_pad_f32(
+    const struct wsp_ggml_compute_params * params,
+    const struct wsp_ggml_tensor * src0,
+          struct wsp_ggml_tensor * dst) {
+
+    if (params->type == WSP_GGML_TASK_INIT || params->type == WSP_GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    WSP_GGML_ASSERT(src0->nb[0] == sizeof(float));
+    WSP_GGML_ASSERT( dst->nb[0] == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    WSP_GGML_TENSOR_UNARY_OP_LOCALS
+
+    float * dst_ptr = (float *) dst->data;
+
+    // TODO: optimize
+
+    for (int64_t i2 = 0; i2 < ne2; ++i2) {
+        for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
+            for (int64_t i0 = 0; i0 < ne0; ++i0) {
+                for (int64_t i3 = 0; i3 < ne3; ++i3) {
+                    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+
+                    const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+
+                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                        dst_ptr[dst_idx] = *src_ptr;
+                    } else {
+                        dst_ptr[dst_idx] = 0;
+                    }
+                }
+            }
+        }
+    }
+}
+
+static void wsp_ggml_compute_forward_pad(
+    const struct wsp_ggml_compute_params * params,
+    const struct wsp_ggml_tensor * src0,
+    struct wsp_ggml_tensor * dst) {
+    switch (src0->type) {
+        case WSP_GGML_TYPE_F32:
+            {
+                wsp_ggml_compute_forward_pad_f32(params, src0, dst);
+            } break;
+        default:
+            {
+                WSP_GGML_ASSERT(false);
+            } break;
+    }
+}
+
+// wsp_ggml_compute_forward_argsort
+
+static void wsp_ggml_compute_forward_argsort_f32(
+    const struct wsp_ggml_compute_params * params,
+    const struct wsp_ggml_tensor * src0,
+    struct wsp_ggml_tensor * dst) {
+
+    if (params->type == WSP_GGML_TASK_INIT || params->type == WSP_GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    WSP_GGML_TENSOR_UNARY_OP_LOCALS
+
+    WSP_GGML_ASSERT(nb0 == sizeof(float));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nr = wsp_ggml_nrows(src0);
+
+    enum wsp_ggml_sort_order order = (enum wsp_ggml_sort_order) wsp_ggml_get_op_params_i32(dst, 0);
+
+    for (int64_t i = ith; i < nr; i += nth) {
+        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
+        const float * src_data = (float *)((char *) src0->data + i*nb01);
+
+        for (int64_t j = 0; j < ne0; j++) {
+            dst_data[j] = j;
+        }
+
+        // C doesn't have a functional sort, so we do a bubble sort instead
+        for (int64_t j = 0; j < ne0; j++) {
+            for (int64_t k = j + 1; k < ne0; k++) {
+                if ((order == WSP_GGML_SORT_ASC && src_data[dst_data[j]] > src_data[dst_data[k]]) ||
+                    (order == WSP_GGML_SORT_DESC && src_data[dst_data[j]] < src_data[dst_data[k]])) {
+                    int32_t tmp = dst_data[j];
+                    dst_data[j] = dst_data[k];
+                    dst_data[k] = tmp;
+                }
+            }
+        }
+    }
+}
+
+static void wsp_ggml_compute_forward_argsort(
+    const struct wsp_ggml_compute_params * params,
+    const struct wsp_ggml_tensor * src0,
+    struct wsp_ggml_tensor * dst) {
+
+    switch (src0->type) {
+        case WSP_GGML_TYPE_F32:
+            {
+                wsp_ggml_compute_forward_argsort_f32(params, src0, dst);
+            } break;
+        default:
+            {
+                WSP_GGML_ASSERT(false);
+            } break;
+    }
+}
+
 // wsp_ggml_compute_forward_flash_attn
 
 static void wsp_ggml_compute_forward_flash_attn_f32(
@@ -14026,10 +13511,6 @@ static void wsp_ggml_compute_forward_unary(
             {
                 wsp_ggml_compute_forward_silu(params, src0, dst);
             } break;
-        case WSP_GGML_UNARY_OP_LEAKY:
-            {
-                wsp_ggml_compute_forward_leaky(params, src0, dst);
-            } break;
         default:
             {
                 WSP_GGML_ASSERT(false);
@@ -14701,7 +14182,11 @@ static void wsp_ggml_compute_forward(struct wsp_ggml_compute_params * params, st
             } break;
         case WSP_GGML_OP_MUL_MAT:
             {
-                wsp_ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor);
+                wsp_ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor, 0, tensor->ne[1]);
+            } break;
+        case WSP_GGML_OP_MUL_MAT_ID:
+            {
+                wsp_ggml_compute_forward_mul_mat_id(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case WSP_GGML_OP_OUT_PROD:
             {
@@ -14761,7 +14246,7 @@ static void wsp_ggml_compute_forward(struct wsp_ggml_compute_params * params, st
             } break;
         case WSP_GGML_OP_SOFT_MAX:
             {
-                wsp_ggml_compute_forward_soft_max(params, tensor->src[0], tensor);
+                wsp_ggml_compute_forward_soft_max(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case WSP_GGML_OP_SOFT_MAX_BACK:
             {
@@ -14783,33 +14268,13 @@ static void wsp_ggml_compute_forward(struct wsp_ggml_compute_params * params, st
             {
                 wsp_ggml_compute_forward_clamp(params, tensor->src[0], tensor);
             } break;
-        case WSP_GGML_OP_CONV_1D:
-            {
-                wsp_ggml_compute_forward_conv_1d(params, tensor->src[0], tensor->src[1], tensor);
-            } break;
-        case WSP_GGML_OP_CONV_1D_STAGE_0:
-            {
-                wsp_ggml_compute_forward_conv_1d_stage_0(params, tensor->src[0], tensor->src[1], tensor);
-            } break;
-        case WSP_GGML_OP_CONV_1D_STAGE_1:
-            {
-                wsp_ggml_compute_forward_conv_1d_stage_1(params, tensor->src[0], tensor->src[1], tensor);
-            } break;
         case WSP_GGML_OP_CONV_TRANSPOSE_1D:
             {
                 wsp_ggml_compute_forward_conv_transpose_1d(params, tensor->src[0], tensor->src[1], tensor);
             } break;
-        case WSP_GGML_OP_CONV_2D:
-            {
-                wsp_ggml_compute_forward_conv_2d(params, tensor->src[0], tensor->src[1], tensor);
-            } break;
-        case WSP_GGML_OP_CONV_2D_STAGE_0:
-            {
-                wsp_ggml_compute_forward_conv_2d_stage_0(params, tensor->src[0], tensor->src[1], tensor);
-            } break;
-        case WSP_GGML_OP_CONV_2D_STAGE_1:
+        case WSP_GGML_OP_IM2COL:
             {
-                wsp_ggml_compute_forward_conv_2d_stage_1(params, tensor->src[0], tensor->src[1], tensor);
+                wsp_ggml_compute_forward_im2col(params, tensor->src[0], tensor->src[1], tensor);
             } break;
         case WSP_GGML_OP_CONV_TRANSPOSE_2D:
             {
@@ -14827,6 +14292,18 @@ static void wsp_ggml_compute_forward(struct wsp_ggml_compute_params * params, st
             {
                 wsp_ggml_compute_forward_upscale(params, tensor->src[0], tensor);
             } break;
+        case WSP_GGML_OP_PAD:
+            {
+                wsp_ggml_compute_forward_pad(params, tensor->src[0], tensor);
+            } break;
+        case WSP_GGML_OP_ARGSORT:
+            {
+                wsp_ggml_compute_forward_argsort(params, tensor->src[0], tensor);
+            } break;
+        case WSP_GGML_OP_LEAKY_RELU:
+            {
+                wsp_ggml_compute_forward_leaky_relu(params, tensor->src[0], tensor);
+            } break;
         case WSP_GGML_OP_FLASH_ATTN:
             {
                 const int32_t t = wsp_ggml_get_op_params_i32(tensor, 0);
@@ -15151,7 +14628,7 @@ void wsp_ggml_build_backward_gradient_checkpointing(
             // insert new tensors recomputing src, reusing already made replacements,
             // remember replacements: remember new tensors with mapping from corresponding gf nodes
             // recurse for input tensors,
-            // unless (i.e. terminating when) input tensors are replacments (like checkpoints)
+            // unless (i.e. terminating when) input tensors are replacements (like checkpoints)
             node->src[k] = wsp_ggml_recompute_graph_node(ctx, gf, replacements, node->src[k]);
         }
         // insert rewritten backward node with replacements made into resulting backward graph gb
@@ -15477,6 +14954,10 @@ static void wsp_ggml_compute_backward(struct wsp_ggml_context * ctx, struct wsp_
                                 zero_table);
                 }
             } break;
+        case WSP_GGML_OP_MUL_MAT_ID:
+            {
+                WSP_GGML_ASSERT(false); // TODO: not implemented
+            } break;
         case WSP_GGML_OP_OUT_PROD:
             {
                 WSP_GGML_ASSERT(false); // TODO: not implemented
@@ -15708,17 +15189,20 @@ static void wsp_ggml_compute_backward(struct wsp_ggml_context * ctx, struct wsp_
                 // necessary for llama
                 if (src0->grad) {
                     //const int n_past = ((int32_t *) tensor->op_params)[0];
-                    const int n_dims = ((int32_t *) tensor->op_params)[1];
-                    const int mode   = ((int32_t *) tensor->op_params)[2];
-                    const int n_ctx  = ((int32_t *) tensor->op_params)[3];
-                    float freq_base;
-                    float freq_scale;
-                    float xpos_base;
-                    bool  xpos_down;
-                    memcpy(&freq_base,  (int32_t *) tensor->op_params + 4, sizeof(float));
-                    memcpy(&freq_scale, (int32_t *) tensor->op_params + 5, sizeof(float));
-                    memcpy(&xpos_base,  (int32_t *) tensor->op_params + 6, sizeof(float));
-                    memcpy(&xpos_down,  (int32_t *) tensor->op_params + 7, sizeof(bool));
+                    const int n_dims     = ((int32_t *) tensor->op_params)[1];
+                    const int mode       = ((int32_t *) tensor->op_params)[2];
+                    const int n_ctx      = ((int32_t *) tensor->op_params)[3];
+                    const int n_orig_ctx = ((int32_t *) tensor->op_params)[4];
+                    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, xpos_base, xpos_down;
+
+                    memcpy(&freq_base,   (int32_t *) tensor->op_params +  5, sizeof(float));
+                    memcpy(&freq_scale,  (int32_t *) tensor->op_params +  6, sizeof(float));
+                    memcpy(&ext_factor,  (int32_t *) tensor->op_params +  7, sizeof(float));
+                    memcpy(&attn_factor, (int32_t *) tensor->op_params +  8, sizeof(float));
+                    memcpy(&beta_fast,   (int32_t *) tensor->op_params +  9, sizeof(float));
+                    memcpy(&beta_slow,   (int32_t *) tensor->op_params + 10, sizeof(float));
+                    memcpy(&xpos_base,   (int32_t *) tensor->op_params + 11, sizeof(float));
+                    memcpy(&xpos_down,   (int32_t *) tensor->op_params + 12, sizeof(bool));
 
                     src0->grad = wsp_ggml_add_or_set(ctx,
                             src0->grad,
@@ -15728,8 +15212,13 @@ static void wsp_ggml_compute_backward(struct wsp_ggml_context * ctx, struct wsp_
                                 n_dims,
                                 mode,
                                 n_ctx,
+                                n_orig_ctx,
                                 freq_base,
                                 freq_scale,
+                                ext_factor,
+                                attn_factor,
+                                beta_fast,
+                                beta_slow,
                                 xpos_base,
                                 xpos_down),
                             zero_table);
@@ -15739,17 +15228,20 @@ static void wsp_ggml_compute_backward(struct wsp_ggml_context * ctx, struct wsp_
             {
                 if (src0->grad) {
                     //const int n_past = ((int32_t *) tensor->op_params)[0];
-                    const int n_dims = ((int32_t *) tensor->op_params)[1];
-                    const int mode   = ((int32_t *) tensor->op_params)[2];
-                    const int n_ctx  = ((int32_t *) tensor->op_params)[3];
-                    float freq_base;
-                    float freq_scale;
-                    float xpos_base;
-                    bool  xpos_down;
-                    memcpy(&freq_base,  (int32_t *) tensor->op_params + 4, sizeof(float));
-                    memcpy(&freq_scale, (int32_t *) tensor->op_params + 5, sizeof(float));
-                    memcpy(&xpos_base,  (int32_t *) tensor->op_params + 6, sizeof(float));
-                    memcpy(&xpos_down,  (int32_t *) tensor->op_params + 7, sizeof(bool));
+                    const int n_dims     = ((int32_t *) tensor->op_params)[1];
+                    const int mode       = ((int32_t *) tensor->op_params)[2];
+                    const int n_ctx      = ((int32_t *) tensor->op_params)[3];
+                    const int n_orig_ctx = ((int32_t *) tensor->op_params)[4];
+                    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, xpos_base, xpos_down;
+
+                    memcpy(&freq_base,   (int32_t *) tensor->op_params +  5, sizeof(float));
+                    memcpy(&freq_scale,  (int32_t *) tensor->op_params +  6, sizeof(float));
+                    memcpy(&ext_factor,  (int32_t *) tensor->op_params +  7, sizeof(float));
+                    memcpy(&attn_factor, (int32_t *) tensor->op_params +  8, sizeof(float));
+                    memcpy(&beta_fast,   (int32_t *) tensor->op_params +  9, sizeof(float));
+                    memcpy(&beta_slow,   (int32_t *) tensor->op_params + 10, sizeof(float));
+                    memcpy(&xpos_base,   (int32_t *) tensor->op_params + 11, sizeof(float));
+                    memcpy(&xpos_down,   (int32_t *) tensor->op_params + 12, sizeof(bool));
 
                     src0->grad = wsp_ggml_add_or_set(ctx,
                             src0->grad,
@@ -15758,14 +15250,14 @@ static void wsp_ggml_compute_backward(struct wsp_ggml_context * ctx, struct wsp_
                                 src1,
                                 n_dims,
                                 mode,
-                                0,
                                 n_ctx,
+                                n_orig_ctx,
                                 freq_base,
                                 freq_scale,
-                                0.0f,
-                                1.0f,
-                                0.0f,
-                                0.0f,
+                                ext_factor,
+                                attn_factor,
+                                beta_fast,
+                                beta_slow,
                                 xpos_base,
                                 xpos_down,
                                 false),
@@ -15780,47 +15272,39 @@ static void wsp_ggml_compute_backward(struct wsp_ggml_context * ctx, struct wsp_
             {
                 WSP_GGML_ASSERT(false); // TODO: not implemented
             } break;
-        case WSP_GGML_OP_CONV_1D:
-            {
-                WSP_GGML_ASSERT(false); // TODO: not implemented
-            } break;
-        case WSP_GGML_OP_CONV_1D_STAGE_0:
-            {
-                WSP_GGML_ASSERT(false); // TODO: not implemented
-            } break;
-        case WSP_GGML_OP_CONV_1D_STAGE_1:
+        case WSP_GGML_OP_CONV_TRANSPOSE_1D:
             {
                 WSP_GGML_ASSERT(false); // TODO: not implemented
             } break;
-        case WSP_GGML_OP_CONV_TRANSPOSE_1D:
+        case WSP_GGML_OP_IM2COL:
             {
                 WSP_GGML_ASSERT(false); // TODO: not implemented
             } break;
-        case WSP_GGML_OP_CONV_2D:
+        case WSP_GGML_OP_CONV_TRANSPOSE_2D:
             {
                 WSP_GGML_ASSERT(false); // TODO: not implemented
             } break;
-        case WSP_GGML_OP_CONV_2D_STAGE_0:
+        case WSP_GGML_OP_POOL_1D:
             {
                 WSP_GGML_ASSERT(false); // TODO: not implemented
             } break;
-        case WSP_GGML_OP_CONV_2D_STAGE_1:
+        case WSP_GGML_OP_POOL_2D:
             {
                 WSP_GGML_ASSERT(false); // TODO: not implemented
             } break;
-        case WSP_GGML_OP_CONV_TRANSPOSE_2D:
+        case WSP_GGML_OP_UPSCALE:
             {
                 WSP_GGML_ASSERT(false); // TODO: not implemented
             } break;
-        case WSP_GGML_OP_POOL_1D:
+        case WSP_GGML_OP_PAD:
             {
                 WSP_GGML_ASSERT(false); // TODO: not implemented
             } break;
-        case WSP_GGML_OP_POOL_2D:
+        case WSP_GGML_OP_ARGSORT:
             {
                 WSP_GGML_ASSERT(false); // TODO: not implemented
             } break;
-        case WSP_GGML_OP_UPSCALE:
+        case WSP_GGML_OP_LEAKY_RELU:
             {
                 WSP_GGML_ASSERT(false); // TODO: not implemented
             } break;
@@ -16184,12 +15668,8 @@ struct wsp_ggml_cgraph * wsp_ggml_new_graph(struct wsp_ggml_context * ctx) {
     return wsp_ggml_new_graph_custom(ctx, WSP_GGML_DEFAULT_GRAPH_SIZE, false);
 }
 
-struct wsp_ggml_cgraph * wsp_ggml_graph_view(struct wsp_ggml_context * ctx, struct wsp_ggml_cgraph * cgraph0, int i0, int i1) {
-    const size_t obj_size = sizeof(struct wsp_ggml_cgraph);
-    struct wsp_ggml_object * obj = wsp_ggml_new_object(ctx, WSP_GGML_OBJECT_GRAPH, obj_size);
-    struct wsp_ggml_cgraph * cgraph = (struct wsp_ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs);
-
-    *cgraph = (struct wsp_ggml_cgraph) {
+struct wsp_ggml_cgraph wsp_ggml_graph_view(struct wsp_ggml_cgraph * cgraph0, int i0, int i1) {
+    struct wsp_ggml_cgraph cgraph = {
         /*.size         =*/ 0,
         /*.n_nodes      =*/ i1 - i0,
         /*.n_leafs      =*/ 0,
@@ -16424,7 +15904,6 @@ static int wsp_ggml_get_n_tasks(struct wsp_ggml_tensor * node, int n_threads) {
                 n_tasks = n_threads;
             } break;
         case WSP_GGML_OP_SUB:
-        case WSP_GGML_OP_DIV:
         case WSP_GGML_OP_SQR:
         case WSP_GGML_OP_SQRT:
         case WSP_GGML_OP_LOG:
@@ -16434,6 +15913,7 @@ static int wsp_ggml_get_n_tasks(struct wsp_ggml_tensor * node, int n_threads) {
         case WSP_GGML_OP_ARGMAX:
         case WSP_GGML_OP_REPEAT:
         case WSP_GGML_OP_REPEAT_BACK:
+        case WSP_GGML_OP_LEAKY_RELU:
             {
                 n_tasks = 1;
             } break;
@@ -16446,7 +15926,6 @@ static int wsp_ggml_get_n_tasks(struct wsp_ggml_tensor * node, int n_threads) {
                 case WSP_GGML_UNARY_OP_TANH:
                 case WSP_GGML_UNARY_OP_ELU:
                 case WSP_GGML_UNARY_OP_RELU:
-                case WSP_GGML_UNARY_OP_LEAKY:
                     {
                         n_tasks = 1;
                     } break;
@@ -16457,10 +15936,13 @@ static int wsp_ggml_get_n_tasks(struct wsp_ggml_tensor * node, int n_threads) {
                     {
                         n_tasks = n_threads;
                     } break;
+                default:
+                    WSP_GGML_ASSERT(false);
             }
             break;
         case WSP_GGML_OP_SILU_BACK:
         case WSP_GGML_OP_MUL:
+        case WSP_GGML_OP_DIV:
         case WSP_GGML_OP_NORM:
         case WSP_GGML_OP_RMS_NORM:
         case WSP_GGML_OP_RMS_NORM_BACK:
@@ -16498,6 +15980,11 @@ static int wsp_ggml_get_n_tasks(struct wsp_ggml_tensor * node, int n_threads) {
                 }
 #endif
             } break;
+        case WSP_GGML_OP_MUL_MAT_ID:
+            {
+                // FIXME: blas
+                n_tasks = n_threads;
+            } break;
         case WSP_GGML_OP_OUT_PROD:
             {
                 n_tasks = n_threads;
@@ -16517,7 +16004,6 @@ static int wsp_ggml_get_n_tasks(struct wsp_ggml_tensor * node, int n_threads) {
             } break;
         case WSP_GGML_OP_DIAG_MASK_ZERO:
         case WSP_GGML_OP_DIAG_MASK_INF:
-        case WSP_GGML_OP_SOFT_MAX:
         case WSP_GGML_OP_SOFT_MAX_BACK:
         case WSP_GGML_OP_ROPE:
         case WSP_GGML_OP_ROPE_BACK:
@@ -16533,31 +16019,15 @@ static int wsp_ggml_get_n_tasks(struct wsp_ggml_tensor * node, int n_threads) {
             {
                 n_tasks = 1; //TODO
             } break;
-        case WSP_GGML_OP_CONV_1D:
-            {
-                n_tasks = n_threads;
-            } break;
-        case WSP_GGML_OP_CONV_1D_STAGE_0:
-            {
-                n_tasks = n_threads;
-            } break;
-        case WSP_GGML_OP_CONV_1D_STAGE_1:
+        case WSP_GGML_OP_SOFT_MAX:
             {
-                n_tasks = n_threads;
+                n_tasks = MIN(MIN(4, n_threads), wsp_ggml_nrows(node->src[0]));
             } break;
         case WSP_GGML_OP_CONV_TRANSPOSE_1D:
             {
                 n_tasks = n_threads;
             } break;
-        case WSP_GGML_OP_CONV_2D:
-            {
-                n_tasks = n_threads;
-            } break;
-        case WSP_GGML_OP_CONV_2D_STAGE_0:
-            {
-                n_tasks = n_threads;
-            } break;
-        case WSP_GGML_OP_CONV_2D_STAGE_1:
+        case WSP_GGML_OP_IM2COL:
             {
                 n_tasks = n_threads;
             } break;
@@ -16574,6 +16044,14 @@ static int wsp_ggml_get_n_tasks(struct wsp_ggml_tensor * node, int n_threads) {
             {
                 n_tasks = n_threads;
             } break;
+        case WSP_GGML_OP_PAD:
+            {
+                n_tasks = n_threads;
+            } break;
+        case WSP_GGML_OP_ARGSORT:
+            {
+                n_tasks = n_threads;
+            } break;
         case WSP_GGML_OP_FLASH_ATTN:
             {
                 n_tasks = n_threads;
@@ -16642,6 +16120,12 @@ static int wsp_ggml_get_n_tasks(struct wsp_ggml_tensor * node, int n_threads) {
             } break;
         default:
             {
+                fprintf(stderr, "%s: op not implemented: ", __func__);
+                if (node->op < WSP_GGML_OP_COUNT) {
+                    fprintf(stderr, "%s\n", wsp_ggml_op_name(node->op));
+                } else {
+                    fprintf(stderr, "%d\n", node->op);
+                }
                 WSP_GGML_ASSERT(false);
             } break;
     }
@@ -16782,18 +16266,16 @@ struct wsp_ggml_cplan wsp_ggml_graph_plan(struct wsp_ggml_cgraph * cgraph, int n
 
     // thread scheduling for the different operations + work buffer size estimation
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        int n_tasks = 1;
-
         struct wsp_ggml_tensor * node = cgraph->nodes[i];
 
+        const int n_tasks = wsp_ggml_get_n_tasks(node, n_threads);
+
         size_t cur = 0;
 
         switch (node->op) {
             case WSP_GGML_OP_CPY:
             case WSP_GGML_OP_DUP:
                 {
-                    n_tasks = n_threads;
-
                     if (wsp_ggml_is_quantized(node->type)) {
                         cur = wsp_ggml_type_size(WSP_GGML_TYPE_F32) * node->ne[0] * n_tasks;
                     }
@@ -16801,16 +16283,12 @@ struct wsp_ggml_cplan wsp_ggml_graph_plan(struct wsp_ggml_cgraph * cgraph, int n
             case WSP_GGML_OP_ADD:
             case WSP_GGML_OP_ADD1:
                 {
-                    n_tasks = n_threads;
-
                     if (wsp_ggml_is_quantized(node->src[0]->type)) {
                         cur = wsp_ggml_type_size(WSP_GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
                     }
                 } break;
             case WSP_GGML_OP_ACC:
                 {
-                    n_tasks = n_threads;
-
                     if (wsp_ggml_is_quantized(node->src[0]->type)) {
                         cur = wsp_ggml_type_size(WSP_GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks;
                     }
@@ -16836,45 +16314,32 @@ struct wsp_ggml_cplan wsp_ggml_graph_plan(struct wsp_ggml_cgraph * cgraph, int n
                         cur = wsp_ggml_type_size(vec_dot_type)*wsp_ggml_nelements(node->src[1])/wsp_ggml_blck_size(vec_dot_type);
                     }
                 } break;
+            case WSP_GGML_OP_MUL_MAT_ID:
+                {
+                    const struct wsp_ggml_tensor * a = node->src[2];
+                    const struct wsp_ggml_tensor * b = node->src[1];
+                    const enum wsp_ggml_type vec_dot_type = type_traits[a->type].vec_dot_type;
+#if defined(WSP_GGML_USE_ACCELERATE) || defined(WSP_GGML_USE_OPENBLAS)
+                    if (wsp_ggml_compute_forward_mul_mat_use_blas(a, b, node)) {
+                        if (a->type != WSP_GGML_TYPE_F32) {
+                            // here we need memory just for single 2D matrix from src0
+                            cur = wsp_ggml_type_size(WSP_GGML_TYPE_F32)*(a->ne[0]*a->ne[1]);
+                        }
+                    } else
+#endif
+                    if (b->type != vec_dot_type) {
+                        cur = wsp_ggml_type_size(vec_dot_type)*wsp_ggml_nelements(b)/wsp_ggml_blck_size(vec_dot_type);
+                    }
+                } break;
             case WSP_GGML_OP_OUT_PROD:
                 {
-                    n_tasks = n_threads;
-
                     if (wsp_ggml_is_quantized(node->src[0]->type)) {
                         cur = wsp_ggml_type_size(WSP_GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
                     }
                 } break;
-            case WSP_GGML_OP_CONV_1D:
+            case WSP_GGML_OP_SOFT_MAX:
                 {
-                    WSP_GGML_ASSERT(node->src[0]->ne[3] == 1);
-                    WSP_GGML_ASSERT(node->src[1]->ne[2] == 1);
-                    WSP_GGML_ASSERT(node->src[1]->ne[3] == 1);
-
-                    const int64_t ne00 = node->src[0]->ne[0];
-                    const int64_t ne01 = node->src[0]->ne[1];
-                    const int64_t ne02 = node->src[0]->ne[2];
-
-                    const int64_t ne10 = node->src[1]->ne[0];
-                    const int64_t ne11 = node->src[1]->ne[1];
-
-                    const int64_t ne0 = node->ne[0];
-                    const int64_t ne1 = node->ne[1];
-                    const int64_t nk  = ne00;
-                    const int64_t ew0 = nk * ne01;
-
-                    UNUSED(ne02);
-                    UNUSED(ne10);
-                    UNUSED(ne11);
-
-                    if (node->src[0]->type == WSP_GGML_TYPE_F16 &&
-                        node->src[1]->type == WSP_GGML_TYPE_F32) {
-                        cur = sizeof(wsp_ggml_fp16_t)*(ne0*ne1*ew0);
-                    } else if (node->src[0]->type == WSP_GGML_TYPE_F32 &&
-                               node->src[1]->type == WSP_GGML_TYPE_F32) {
-                        cur = sizeof(float)*(ne0*ne1*ew0);
-                    } else {
-                        WSP_GGML_ASSERT(false);
-                    }
+                    cur = wsp_ggml_type_size(WSP_GGML_TYPE_F32) * node->ne[0] * n_tasks;
                 } break;
             case WSP_GGML_OP_CONV_TRANSPOSE_1D:
                 {
@@ -16901,38 +16366,6 @@ struct wsp_ggml_cplan wsp_ggml_graph_plan(struct wsp_ggml_cgraph * cgraph, int n
                         WSP_GGML_ASSERT(false);
                     }
                 } break;
-            case WSP_GGML_OP_CONV_2D:
-                {
-                    const int64_t ne00 = node->src[0]->ne[0]; // W
-                    const int64_t ne01 = node->src[0]->ne[1]; // H
-                    const int64_t ne02 = node->src[0]->ne[2]; // C
-                    const int64_t ne03 = node->src[0]->ne[3]; // N
-
-                    const int64_t ne10 = node->src[1]->ne[0]; // W
-                    const int64_t ne11 = node->src[1]->ne[1]; // H
-                    const int64_t ne12 = node->src[1]->ne[2]; // C
-
-                    const int64_t ne0 = node->ne[0];
-                    const int64_t ne1 = node->ne[1];
-                    const int64_t ne2 = node->ne[2];
-                    const int64_t ne3 = node->ne[3];
-                    const int64_t nk = ne00*ne01;
-                    const int64_t ew0 = nk * ne02;
-
-                    UNUSED(ne03);
-                    UNUSED(ne2);
-
-                    if (node->src[0]->type == WSP_GGML_TYPE_F16 &&
-                        node->src[1]->type == WSP_GGML_TYPE_F32) {
-                        // im2col: [N*OH*OW, IC*KH*KW]
-                        cur = sizeof(wsp_ggml_fp16_t)*(ne3*ne0*ne1*ew0);
-                    } else if (node->src[0]->type == WSP_GGML_TYPE_F32 &&
-                               node->src[1]->type == WSP_GGML_TYPE_F32) {
-                        cur = sizeof(float)*      (ne10*ne11*ne12);
-                    } else {
-                        WSP_GGML_ASSERT(false);
-                    }
-                } break;
             case WSP_GGML_OP_CONV_TRANSPOSE_2D:
                 {
                     const int64_t ne00 = node->src[0]->ne[0]; // W
@@ -16949,8 +16382,6 @@ struct wsp_ggml_cplan wsp_ggml_graph_plan(struct wsp_ggml_cgraph * cgraph, int n
                 } break;
             case WSP_GGML_OP_FLASH_ATTN:
                 {
-                    n_tasks = n_threads;
-
                     const int64_t ne11 = wsp_ggml_up(node->src[1]->ne[1], WSP_GGML_SOFT_MAX_UNROLL);
 
                     if (node->src[1]->type == WSP_GGML_TYPE_F32) {
@@ -16963,8 +16394,6 @@ struct wsp_ggml_cplan wsp_ggml_graph_plan(struct wsp_ggml_cgraph * cgraph, int n
                 } break;
             case WSP_GGML_OP_FLASH_FF:
                 {
-                    n_tasks = n_threads;
-
                     if (node->src[1]->type == WSP_GGML_TYPE_F32) {
                         cur  = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1)
                         cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2
@@ -16975,8 +16404,6 @@ struct wsp_ggml_cplan wsp_ggml_graph_plan(struct wsp_ggml_cgraph * cgraph, int n
                 } break;
             case WSP_GGML_OP_FLASH_ATTN_BACK:
                 {
-                    n_tasks = n_threads;
-
                     const int64_t    D = node->src[0]->ne[0];
                     const int64_t ne11 = wsp_ggml_up(node->src[1]->ne[1], WSP_GGML_SOFT_MAX_UNROLL);
                     const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in wsp_ggml_compute_forward_flash_attn_back
@@ -16991,8 +16418,6 @@ struct wsp_ggml_cplan wsp_ggml_graph_plan(struct wsp_ggml_cgraph * cgraph, int n
 
             case WSP_GGML_OP_CROSS_ENTROPY_LOSS:
                 {
-                    n_tasks = n_threads;
-
                     cur = wsp_ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
                 } break;
             case WSP_GGML_OP_COUNT:
@@ -18719,14 +18144,14 @@ enum wsp_ggml_opt_result wsp_ggml_opt_resume_g(
 
 ////////////////////////////////////////////////////////////////////////////////
 
-size_t wsp_ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist) {
+size_t wsp_ggml_wsp_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist) {
     assert(k % QK4_0 == 0);
     const int nb = k / QK4_0;
 
     for (int b = 0; b < n; b += k) {
         block_q4_0 * restrict y = (block_q4_0 *) dst + b/QK4_0;
 
-        quantize_row_q4_0_reference(src + b, y, k);
+        wsp_quantize_row_q4_0_reference(src + b, y, k);
 
         for (int i = 0; i < nb; i++) {
             for (int j = 0; j < QK4_0; j += 2) {
@@ -18742,14 +18167,14 @@ size_t wsp_ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64
     return (n/QK4_0*sizeof(block_q4_0));
 }
 
-size_t wsp_ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist) {
+size_t wsp_ggml_wsp_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist) {
     assert(k % QK4_1 == 0);
     const int nb = k / QK4_1;
 
     for (int b = 0; b < n; b += k) {
         block_q4_1 * restrict y = (block_q4_1 *) dst + b/QK4_1;
 
-        quantize_row_q4_1_reference(src + b, y, k);
+        wsp_quantize_row_q4_1_reference(src + b, y, k);
 
         for (int i = 0; i < nb; i++) {
             for (int j = 0; j < QK4_1; j += 2) {
@@ -18765,22 +18190,22 @@ size_t wsp_ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64
     return (n/QK4_1*sizeof(block_q4_1));
 }
 
-size_t wsp_ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist) {
+size_t wsp_ggml_wsp_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist) {
     assert(k % QK5_0 == 0);
     const int nb = k / QK5_0;
 
     for (int b = 0; b < n; b += k) {
         block_q5_0 * restrict y = (block_q5_0 *)dst + b/QK5_0;
 
-        quantize_row_q5_0_reference(src + b, y, k);
+        wsp_quantize_row_q5_0_reference(src + b, y, k);
 
         for (int i = 0; i < nb; i++) {
             uint32_t qh;
             memcpy(&qh, &y[i].qh, sizeof(qh));
 
             for (int j = 0; j < QK5_0; j += 2) {
-                const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-                const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12));
+                const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4;
+                const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12));
 
                 // cast to 16 bins
                 const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2;
@@ -18795,22 +18220,22 @@ size_t wsp_ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64
     return (n/QK5_0*sizeof(block_q5_0));
 }
 
-size_t wsp_ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist) {
+size_t wsp_ggml_wsp_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist) {
     assert(k % QK5_1 == 0);
     const int nb = k / QK5_1;
 
     for (int b = 0; b < n; b += k) {
         block_q5_1 * restrict y = (block_q5_1 *)dst + b/QK5_1;
 
-        quantize_row_q5_1_reference(src + b, y, k);
+        wsp_quantize_row_q5_1_reference(src + b, y, k);
 
         for (int i = 0; i < nb; i++) {
             uint32_t qh;
             memcpy(&qh, &y[i].qh, sizeof(qh));
 
             for (int j = 0; j < QK5_1; j += 2) {
-                const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-                const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12));
+                const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4;
+                const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12));
 
                 // cast to 16 bins
                 const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2;
@@ -18825,14 +18250,14 @@ size_t wsp_ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64
     return (n/QK5_1*sizeof(block_q5_1));
 }
 
-size_t wsp_ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist) {
+size_t wsp_ggml_wsp_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist) {
     assert(k % QK8_0 == 0);
     const int nb = k / QK8_0;
 
     for (int b = 0; b < n; b += k) {
         block_q8_0 * restrict y = (block_q8_0 *)dst + b/QK8_0;
 
-        quantize_row_q8_0_reference(src + b, y, k);
+        wsp_quantize_row_q8_0_reference(src + b, y, k);
 
         for (int i = 0; i < nb; i++) {
             for (int j = 0; j < QK8_0; ++j) {
@@ -18846,68 +18271,68 @@ size_t wsp_ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64
     return (n/QK8_0*sizeof(block_q8_0));
 }
 
-size_t wsp_ggml_quantize_chunk(enum wsp_ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist) {
+size_t wsp_ggml_wsp_quantize_chunk(enum wsp_ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist) {
     size_t result = 0;
     switch (type) {
         case WSP_GGML_TYPE_Q4_0:
             {
                 WSP_GGML_ASSERT(start % QK4_0 == 0);
                 block_q4_0 * block = (block_q4_0*)dst + start / QK4_0;
-                result = wsp_ggml_quantize_q4_0(src + start, block, n, n, hist);
+                result = wsp_ggml_wsp_quantize_q4_0(src + start, block, n, n, hist);
             } break;
         case WSP_GGML_TYPE_Q4_1:
             {
                 WSP_GGML_ASSERT(start % QK4_1 == 0);
                 block_q4_1 * block = (block_q4_1*)dst + start / QK4_1;
-                result = wsp_ggml_quantize_q4_1(src + start, block, n, n, hist);
+                result = wsp_ggml_wsp_quantize_q4_1(src + start, block, n, n, hist);
             } break;
         case WSP_GGML_TYPE_Q5_0:
             {
                 WSP_GGML_ASSERT(start % QK5_0 == 0);
                 block_q5_0 * block = (block_q5_0*)dst + start / QK5_0;
-                result = wsp_ggml_quantize_q5_0(src + start, block, n, n, hist);
+                result = wsp_ggml_wsp_quantize_q5_0(src + start, block, n, n, hist);
             } break;
         case WSP_GGML_TYPE_Q5_1:
             {
                 WSP_GGML_ASSERT(start % QK5_1 == 0);
                 block_q5_1 * block = (block_q5_1*)dst + start / QK5_1;
-                result = wsp_ggml_quantize_q5_1(src + start, block, n, n, hist);
+                result = wsp_ggml_wsp_quantize_q5_1(src + start, block, n, n, hist);
             } break;
         case WSP_GGML_TYPE_Q8_0:
             {
                 WSP_GGML_ASSERT(start % QK8_0 == 0);
                 block_q8_0 * block = (block_q8_0*)dst + start / QK8_0;
-                result = wsp_ggml_quantize_q8_0(src + start, block, n, n, hist);
+                result = wsp_ggml_wsp_quantize_q8_0(src + start, block, n, n, hist);
             } break;
         case WSP_GGML_TYPE_Q2_K:
             {
                 WSP_GGML_ASSERT(start % QK_K == 0);
                 block_q2_K * block = (block_q2_K*)dst + start / QK_K;
-                result = wsp_ggml_quantize_q2_K(src + start, block, n, n, hist);
+                result = wsp_ggml_wsp_quantize_q2_K(src + start, block, n, n, hist);
             } break;
         case WSP_GGML_TYPE_Q3_K:
             {
                 WSP_GGML_ASSERT(start % QK_K == 0);
                 block_q3_K * block = (block_q3_K*)dst + start / QK_K;
-                result = wsp_ggml_quantize_q3_K(src + start, block, n, n, hist);
+                result = wsp_ggml_wsp_quantize_q3_K(src + start, block, n, n, hist);
             } break;
         case WSP_GGML_TYPE_Q4_K:
             {
                 WSP_GGML_ASSERT(start % QK_K == 0);
                 block_q4_K * block = (block_q4_K*)dst + start / QK_K;
-                result = wsp_ggml_quantize_q4_K(src + start, block, n, n, hist);
+                result = wsp_ggml_wsp_quantize_q4_K(src + start, block, n, n, hist);
             } break;
         case WSP_GGML_TYPE_Q5_K:
             {
                 WSP_GGML_ASSERT(start % QK_K == 0);
                 block_q5_K * block = (block_q5_K*)dst + start / QK_K;
-                result = wsp_ggml_quantize_q5_K(src + start, block, n, n, hist);
+                result = wsp_ggml_wsp_quantize_q5_K(src + start, block, n, n, hist);
             } break;
         case WSP_GGML_TYPE_Q6_K:
             {
                 WSP_GGML_ASSERT(start % QK_K == 0);
                 block_q6_K * block = (block_q6_K*)dst + start / QK_K;
-                result = wsp_ggml_quantize_q6_K(src + start, block, n, n, hist);
+                result = wsp_ggml_wsp_quantize_q6_K(src + start, block, n, n, hist);
             } break;
         case WSP_GGML_TYPE_F16:
             {
@@ -19000,6 +18425,7 @@ struct wsp_gguf_kv {
 
 struct wsp_gguf_header {
     char magic[4];
+
     uint32_t version;
     uint64_t n_tensors; // GGUFv2
     uint64_t n_kv;      // GGUFv2
@@ -19089,7 +18515,7 @@ struct wsp_gguf_context * wsp_gguf_init_from_file(const char * fname, struct wsp
 
         for (uint32_t i = 0; i < sizeof(magic); i++) {
             if (magic[i] != WSP_GGUF_MAGIC[i]) {
-                fprintf(stderr, "%s: invalid magic characters %s.\n", __func__, magic);
+                fprintf(stderr, "%s: invalid magic characters '%c%c%c%c'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
                 fclose(file);
                 return NULL;
             }
@@ -19104,7 +18530,6 @@ struct wsp_gguf_context * wsp_gguf_init_from_file(const char * fname, struct wsp
     {
         strncpy(ctx->header.magic, magic, 4);
 
-
         ctx->kv    = NULL;
         ctx->infos = NULL;
         ctx->data  = NULL;
@@ -19132,7 +18557,7 @@ struct wsp_gguf_context * wsp_gguf_init_from_file(const char * fname, struct wsp
     {
         ctx->kv = malloc(ctx->header.n_kv * sizeof(struct wsp_gguf_kv));
 
-        for (uint32_t i = 0; i < ctx->header.n_kv; ++i) {
+        for (uint64_t i = 0; i < ctx->header.n_kv; ++i) {
             struct wsp_gguf_kv * kv = &ctx->kv[i];
 
             //fprintf(stderr, "%s: reading kv %d\n", __func__, i);
@@ -19179,7 +18604,7 @@ struct wsp_gguf_context * wsp_gguf_init_from_file(const char * fname, struct wsp
                             case WSP_GGUF_TYPE_STRING:
                                 {
                                     kv->value.arr.data = malloc(kv->value.arr.n * sizeof(struct wsp_gguf_str));
-                                    for (uint32_t j = 0; j < kv->value.arr.n; ++j) {
+                                    for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
                                         ok = ok && wsp_gguf_fread_str(file, &((struct wsp_gguf_str *) kv->value.arr.data)[j], &offset);
                                     }
                                 } break;
@@ -19207,7 +18632,7 @@ struct wsp_gguf_context * wsp_gguf_init_from_file(const char * fname, struct wsp
     {
         ctx->infos = malloc(ctx->header.n_tensors * sizeof(struct wsp_gguf_tensor_info));
 
-        for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) {
+        for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
             struct wsp_gguf_tensor_info * info = &ctx->infos[i];
 
             for (int j = 0; j < WSP_GGML_MAX_DIMS; ++j) {
@@ -19254,7 +18679,7 @@ struct wsp_gguf_context * wsp_gguf_init_from_file(const char * fname, struct wsp
     // compute the total size of the data section, taking into account the alignment
     {
         ctx->size = 0;
-        for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) {
+        for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
             struct wsp_gguf_tensor_info * info = &ctx->infos[i];
 
             const int64_t ne =
@@ -19323,7 +18748,7 @@ struct wsp_gguf_context * wsp_gguf_init_from_file(const char * fname, struct wsp
         wsp_ggml_set_no_alloc(ctx_data, true);
 
         // create the tensors
-        for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) {
+        for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
             const int64_t ne[WSP_GGML_MAX_DIMS] = {
                 ctx->infos[i].ne[0],
                 ctx->infos[i].ne[1],
@@ -19458,24 +18883,29 @@ int wsp_gguf_find_key(const struct wsp_gguf_context * ctx, const char * key) {
 }
 
 const char * wsp_gguf_get_key(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     return ctx->kv[key_id].key.data;
 }
 
 enum wsp_gguf_type wsp_gguf_get_kv_type(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     return ctx->kv[key_id].type;
 }
 
 enum wsp_gguf_type wsp_gguf_get_arr_type(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_ARRAY);
     return ctx->kv[key_id].value.arr.type;
 }
 
 const void * wsp_gguf_get_arr_data(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_ARRAY);
     return ctx->kv[key_id].value.arr.data;
 }
 
 const char * wsp_gguf_get_arr_str(const struct wsp_gguf_context * ctx, int key_id, int i) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_ARRAY);
     struct wsp_gguf_kv * kv = &ctx->kv[key_id];
     struct wsp_gguf_str * str = &((struct wsp_gguf_str *) kv->value.arr.data)[i];
@@ -19483,70 +18913,90 @@ const char * wsp_gguf_get_arr_str(const struct wsp_gguf_context * ctx, int key_i
 }
 
 int wsp_gguf_get_arr_n(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_ARRAY);
     return ctx->kv[key_id].value.arr.n;
 }
 
 uint8_t wsp_gguf_get_val_u8(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_UINT8);
     return ctx->kv[key_id].value.uint8;
 }
 
 int8_t wsp_gguf_get_val_i8(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_INT8);
     return ctx->kv[key_id].value.int8;
 }
 
 uint16_t wsp_gguf_get_val_u16(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_UINT16);
     return ctx->kv[key_id].value.uint16;
 }
 
 int16_t wsp_gguf_get_val_i16(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_INT16);
     return ctx->kv[key_id].value.int16;
 }
 
 uint32_t wsp_gguf_get_val_u32(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_UINT32);
     return ctx->kv[key_id].value.uint32;
 }
 
 int32_t wsp_gguf_get_val_i32(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_INT32);
     return ctx->kv[key_id].value.int32;
 }
 
 float wsp_gguf_get_val_f32(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_FLOAT32);
     return ctx->kv[key_id].value.float32;
 }
 
 uint64_t wsp_gguf_get_val_u64(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_UINT64);
     return ctx->kv[key_id].value.uint64;
 }
 
 int64_t wsp_gguf_get_val_i64(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_INT64);
     return ctx->kv[key_id].value.int64;
 }
 
 double wsp_gguf_get_val_f64(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_FLOAT64);
     return ctx->kv[key_id].value.float64;
 }
 
 bool wsp_gguf_get_val_bool(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_BOOL);
     return ctx->kv[key_id].value.bool_;
 }
 
 const char * wsp_gguf_get_val_str(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
     WSP_GGML_ASSERT(ctx->kv[key_id].type == WSP_GGUF_TYPE_STRING);
     return ctx->kv[key_id].value.str.data;
 }
 
+const void * wsp_gguf_get_val_data(const struct wsp_gguf_context * ctx, int key_id) {
+    WSP_GGML_ASSERT(key_id >= 0 && key_id < wsp_gguf_get_n_kv(ctx));
+    WSP_GGML_ASSERT(ctx->kv[key_id].type != WSP_GGUF_TYPE_ARRAY);
+    WSP_GGML_ASSERT(ctx->kv[key_id].type != WSP_GGUF_TYPE_STRING);
+    return &ctx->kv[key_id].value;
+}
+
 int wsp_gguf_get_n_tensors(const struct wsp_gguf_context * ctx) {
     return ctx->header.n_tensors;
 }