whisper.rn 0.5.0 → 0.5.1

package/cpp/ggml-metal-impl.h

@@ -1,5 +1,5 @@
-#ifndef WSP_GGML_METAL_IMPL
-#define WSP_GGML_METAL_IMPL
+#ifndef WSP_WSP_WSP_GGML_METAL_IMPL
+#define WSP_WSP_WSP_GGML_METAL_IMPL
 
 // kernel parameters for mat-vec threadgroups
 //
@@ -101,7 +101,7 @@ typedef struct {
     uint64_t nb2;
     uint64_t nb3;
     int32_t  dim;
-} wsp_ggml_metal_kargs_concat;
+} wsp_wsp_wsp_ggml_metal_kargs_concat;
 
 typedef struct {
     int32_t  ne00;
@@ -130,7 +130,7 @@ typedef struct {
     uint64_t nb3;
     uint64_t offs;
     uint64_t o1[8];
-} wsp_ggml_metal_kargs_bin;
+} wsp_wsp_wsp_ggml_metal_kargs_bin;
 
 typedef struct {
     int64_t ne0;
@@ -139,7 +139,7 @@ typedef struct {
     size_t nb02;
     size_t nb11;
     size_t nb21;
-} wsp_ggml_metal_kargs_add_id;
+} wsp_wsp_wsp_ggml_metal_kargs_add_id;
 
 typedef struct {
     int32_t  ne00;
@@ -158,7 +158,7 @@ typedef struct {
     uint64_t nb1;
     uint64_t nb2;
     uint64_t nb3;
-} wsp_ggml_metal_kargs_repeat;
+} wsp_wsp_wsp_ggml_metal_kargs_repeat;
 
 typedef struct {
     int64_t  ne00;
@@ -177,7 +177,7 @@ typedef struct {
     uint64_t nb1;
     uint64_t nb2;
     uint64_t nb3;
-} wsp_ggml_metal_kargs_cpy;
+} wsp_wsp_wsp_ggml_metal_kargs_cpy;
 
 typedef struct {
     int64_t  ne10;
@@ -192,7 +192,7 @@ typedef struct {
     uint64_t nb3;
     uint64_t offs;
     bool     inplace;
-} wsp_ggml_metal_kargs_set;
+} wsp_wsp_wsp_ggml_metal_kargs_set;
 
 typedef struct {
     int32_t  ne00;
@@ -224,7 +224,7 @@ typedef struct {
     int32_t  sect_1;
     int32_t  sect_2;
     int32_t  sect_3;
-} wsp_ggml_metal_kargs_rope;
+} wsp_wsp_wsp_ggml_metal_kargs_rope;
 
 typedef struct {
     int32_t  ne01;
@@ -255,7 +255,7 @@ typedef struct {
     float    m1;
     int32_t  n_head_log2;
     float    logit_softcap;
-} wsp_ggml_metal_kargs_flash_attn_ext;
+} wsp_wsp_wsp_ggml_metal_kargs_flash_attn_ext;
 
 typedef struct {
     int32_t  ne00;
@@ -272,7 +272,7 @@ typedef struct {
     int32_t  ne1;
     int16_t  r2;
     int16_t  r3;
-} wsp_ggml_metal_kargs_mul_mm;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mm;
 
 typedef struct {
     int32_t  ne00;
@@ -293,7 +293,7 @@ typedef struct {
     int32_t  ne1;
     int16_t  r2;
     int16_t  r3;
-} wsp_ggml_metal_kargs_mul_mv;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mv;
 
 typedef struct {
     int32_t  ne00;
@@ -317,7 +317,7 @@ typedef struct {
     int16_t  nsg;
     int16_t  nxpsg;
     int16_t  r1ptg;
-} wsp_ggml_metal_kargs_mul_mv_ext;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mv_ext;
 
 typedef struct {
     int32_t  ne10;
@@ -328,7 +328,7 @@ typedef struct {
     uint64_t nbh11;
     int32_t  ne20;  // n_expert_used
     uint64_t nb21;
-} wsp_ggml_metal_kargs_mul_mm_id_map0;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mm_id_map0;
 
 typedef struct {
     int32_t  ne20; // n_expert_used
@@ -339,7 +339,7 @@ typedef struct {
     int32_t  ne0;
     uint64_t nb1;
     uint64_t nb2;
-} wsp_ggml_metal_kargs_mul_mm_id_map1;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mm_id_map1;
 
 typedef struct {
     int32_t  ne00;
@@ -356,7 +356,7 @@ typedef struct {
     int32_t  neh1;
     int16_t  r2;
     int16_t  r3;
-} wsp_ggml_metal_kargs_mul_mm_id;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mm_id;
 
 typedef struct {
     int32_t  nei0;
@@ -378,14 +378,14 @@ typedef struct {
     int32_t  ne0;
     int32_t  ne1;
     uint64_t nb1;
-} wsp_ggml_metal_kargs_mul_mv_id;
+} wsp_wsp_wsp_ggml_metal_kargs_mul_mv_id;
 
 typedef struct {
     int32_t  ne00;
     int32_t  ne00_4;
     uint64_t nb01;
     float    eps;
-} wsp_ggml_metal_kargs_norm;
+} wsp_wsp_wsp_ggml_metal_kargs_norm;
 
 typedef struct {
     int32_t  ne00;
@@ -400,14 +400,14 @@ typedef struct {
     uint64_t nbf1[3];
     uint64_t nbf2[3];
     uint64_t nbf3[3];
-} wsp_ggml_metal_kargs_rms_norm;
+} wsp_wsp_wsp_ggml_metal_kargs_rms_norm;
 
 typedef struct {
     int32_t  ne00;
     int32_t  ne00_4;
     uint64_t nb01;
     float    eps;
-} wsp_ggml_metal_kargs_l2_norm;
+} wsp_wsp_wsp_ggml_metal_kargs_l2_norm;
 
 typedef struct {
     int64_t  ne00;
@@ -418,7 +418,7 @@ typedef struct {
     uint64_t nb02;
     int32_t  n_groups;
     float    eps;
-} wsp_ggml_metal_kargs_group_norm;
+} wsp_wsp_wsp_ggml_metal_kargs_group_norm;
 
 typedef struct {
     int32_t  IC;
@@ -427,7 +427,7 @@ typedef struct {
     int32_t  s0;
     uint64_t nb0;
     uint64_t nb1;
-} wsp_ggml_metal_kargs_conv_transpose_1d;
+} wsp_wsp_wsp_ggml_metal_kargs_conv_transpose_1d;
 
 typedef struct {
     uint64_t  ofs0;
@@ -445,7 +445,7 @@ typedef struct {
     int32_t  KH;
     int32_t  KW;
     int32_t  KHW; // KH * KW, pre-computed on CPU to save GPU resources
-} wsp_ggml_metal_kargs_im2col;
+} wsp_wsp_wsp_ggml_metal_kargs_im2col;
 
 typedef struct{
     int32_t  ne00;
@@ -458,7 +458,7 @@ typedef struct{
     int32_t  i10;
     float    alpha;
     float    limit;
-} wsp_ggml_metal_kargs_glu;
+} wsp_wsp_wsp_ggml_metal_kargs_glu;
 
 typedef struct {
     int64_t  ne00;
@@ -485,7 +485,7 @@ typedef struct {
     uint64_t nb1;
     uint64_t nb2;
     uint64_t nb3;
-} wsp_ggml_metal_kargs_sum_rows;
+} wsp_wsp_wsp_ggml_metal_kargs_sum_rows;
 
 typedef struct {
     int32_t  ne00;
@@ -508,13 +508,13 @@ typedef struct {
     float    m0;
     float    m1;
     int32_t  n_head_log2;
-} wsp_ggml_metal_kargs_soft_max;
+} wsp_wsp_wsp_ggml_metal_kargs_soft_max;
 
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;
     int      n_past;
-} wsp_ggml_metal_kargs_diag_mask_inf;
+} wsp_wsp_wsp_ggml_metal_kargs_diag_mask_inf;
 
 typedef struct {
     int64_t  ne00;
@@ -533,7 +533,7 @@ typedef struct {
     uint64_t nb0;
     uint64_t nb1;
     uint64_t nb2;
-} wsp_ggml_metal_kargs_ssm_conv;
+} wsp_wsp_wsp_ggml_metal_kargs_ssm_conv;
 
 typedef struct {
     int64_t  d_state;
@@ -558,7 +558,7 @@ typedef struct {
     uint64_t nb51;
     uint64_t nb52;
     uint64_t nb53;
-} wsp_ggml_metal_kargs_ssm_scan;
+} wsp_wsp_wsp_ggml_metal_kargs_ssm_scan;
 
 typedef struct {
     int64_t  ne00;
@@ -569,7 +569,7 @@ typedef struct {
     uint64_t nb11;
     uint64_t nb1;
     uint64_t nb2;
-} wsp_ggml_metal_kargs_get_rows;
+} wsp_wsp_wsp_ggml_metal_kargs_get_rows;
 
 typedef struct {
     int32_t  nk0;
@@ -585,7 +585,7 @@ typedef struct {
     uint64_t nb1;
     uint64_t nb2;
     uint64_t nb3;
-} wsp_ggml_metal_kargs_set_rows;
+} wsp_wsp_wsp_ggml_metal_kargs_set_rows;
 
 typedef struct {
     int64_t  ne00;
@@ -608,7 +608,7 @@ typedef struct {
     float    sf1;
     float    sf2;
     float    sf3;
-} wsp_ggml_metal_kargs_upscale;
+} wsp_wsp_wsp_ggml_metal_kargs_upscale;
 
 typedef struct {
     int64_t  ne00;
@@ -627,7 +627,7 @@ typedef struct {
     uint64_t nb1;
     uint64_t nb2;
     uint64_t nb3;
-} wsp_ggml_metal_kargs_pad;
+} wsp_wsp_wsp_ggml_metal_kargs_pad;
 
 typedef struct {
     int64_t  ne00;
@@ -648,28 +648,28 @@ typedef struct {
     uint64_t nb3;
     int32_t  p0;
     int32_t  p1;
-} wsp_ggml_metal_kargs_pad_reflect_1d;
+} wsp_wsp_wsp_ggml_metal_kargs_pad_reflect_1d;
 
 typedef struct {
     uint64_t nb1;
     int      dim;
     int      max_period;
-} wsp_ggml_metal_kargs_timestep_embedding;
+} wsp_wsp_wsp_ggml_metal_kargs_timestep_embedding;
 
 typedef struct {
     float    slope;
-} wsp_ggml_metal_kargs_leaky_relu;
+} wsp_wsp_wsp_ggml_metal_kargs_leaky_relu;
 
 typedef struct {
     int64_t  ncols;
     int64_t  ncols_pad;
-} wsp_ggml_metal_kargs_argsort;
+} wsp_wsp_wsp_ggml_metal_kargs_argsort;
 
 typedef struct {
     int64_t  ne0;
     float    start;
     float    step;
-} wsp_ggml_metal_kargs_arange;
+} wsp_wsp_wsp_ggml_metal_kargs_arange;
 
 typedef struct {
     int32_t  k0;
@@ -683,6 +683,6 @@ typedef struct {
     int64_t  OH;
     int64_t  OW;
     int64_t  parallel_elements;
-} wsp_ggml_metal_kargs_pool_2d;
+} wsp_wsp_wsp_ggml_metal_kargs_pool_2d;
 
-#endif // WSP_GGML_METAL_IMPL
+#endif // WSP_WSP_WSP_GGML_METAL_IMPL

package/cpp/ggml-metal.h

@@ -39,18 +39,13 @@ extern "C" {
 // user-code should use only these functions
 //
 
+// TODO: remove in the future
 WSP_GGML_BACKEND_API wsp_ggml_backend_t wsp_ggml_backend_metal_init(void);
 
 WSP_GGML_BACKEND_API bool wsp_ggml_backend_is_metal(wsp_ggml_backend_t backend);
 
-WSP_GGML_DEPRECATED(
-        WSP_GGML_BACKEND_API wsp_ggml_backend_buffer_t wsp_ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
-        "obsoleted by the new device interface - https://github.com/ggml-org/llama.cpp/pull/9713");
-
 WSP_GGML_BACKEND_API void wsp_ggml_backend_metal_set_abort_callback(wsp_ggml_backend_t backend, wsp_ggml_abort_callback abort_callback, void * user_data);
 
-WSP_GGML_BACKEND_API wsp_ggml_backend_buffer_type_t wsp_ggml_backend_metal_buffer_type(void);
-
 // helper to check if the device supports a specific family
 // ideally, the user code should be doing these checks
 // ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf

package/cpp/ggml-quants.c

@@ -3721,6 +3721,7 @@ static void wsp_quantize_row_iq3_xxs_impl(int grid_size, const float * WSP_GGML_
             }
             float best = 0;
             float scale = max/(2*kMaxQ-1);
+            for (int k = 0; k < 8; ++k) is_on_grid[k] = true;
             for (int is = -15; is <= 15; ++is) {
                 float id = (2*kMaxQ-1+is*0.2f)/max;
                 float this_scale = 1/id;

package/cpp/ggml.c

@@ -982,7 +982,9 @@ static const char * WSP_GGML_OP_NAME[WSP_GGML_OP_COUNT] = {
     "CONV_TRANSPOSE_1D",
     "IM2COL",
     "IM2COL_BACK",
+    "IM2COL_3D",
     "CONV_2D",
+    "CONV_3D",
     "CONV_2D_DW",
     "CONV_TRANSPOSE_2D",
     "POOL_1D",
@@ -1025,7 +1027,7 @@ static const char * WSP_GGML_OP_NAME[WSP_GGML_OP_COUNT] = {
     "GLU",
 };
 
-static_assert(WSP_GGML_OP_COUNT == 88, "WSP_GGML_OP_COUNT != 88");
+static_assert(WSP_GGML_OP_COUNT == 90, "WSP_GGML_OP_COUNT != 90");
 
 static const char * WSP_GGML_OP_SYMBOL[WSP_GGML_OP_COUNT] = {
     "none",
@@ -1084,7 +1086,9 @@ static const char * WSP_GGML_OP_SYMBOL[WSP_GGML_OP_COUNT] = {
     "conv_transpose_1d(x)",
     "im2col(x)",
     "im2col_back(x)",
+    "im2col_3d(x)",
     "conv_2d(x)",
+    "conv_3d(x)",
     "conv_2d_dw(x)",
     "conv_transpose_2d(x)",
     "pool_1d(x)",
@@ -1127,7 +1131,7 @@ static const char * WSP_GGML_OP_SYMBOL[WSP_GGML_OP_COUNT] = {
     "glu(x)",
 };
 
-static_assert(WSP_GGML_OP_COUNT == 88, "WSP_GGML_OP_COUNT != 88");
+static_assert(WSP_GGML_OP_COUNT == 90, "WSP_GGML_OP_COUNT != 90");
 
 static_assert(WSP_GGML_OP_POOL_COUNT == 2, "WSP_GGML_OP_POOL_COUNT != 2");
 
@@ -3627,6 +3631,7 @@ struct wsp_ggml_tensor * wsp_ggml_get_rows(
         struct wsp_ggml_tensor  * a,
         struct wsp_ggml_tensor  * b) {
     WSP_GGML_ASSERT(a->ne[2] == b->ne[1]);
+    WSP_GGML_ASSERT(a->ne[3] == b->ne[2]);
     WSP_GGML_ASSERT(b->ne[3] == 1);
     WSP_GGML_ASSERT(b->type == WSP_GGML_TYPE_I32);
 
@@ -3680,7 +3685,7 @@ struct wsp_ggml_tensor * wsp_ggml_set_rows(
     WSP_GGML_ASSERT(b->ne[3] % c->ne[2] == 0);
     WSP_GGML_ASSERT(c->ne[3] == 1);
     WSP_GGML_ASSERT(b->type == WSP_GGML_TYPE_F32);
-    WSP_GGML_ASSERT(c->type == WSP_GGML_TYPE_I64);
+    WSP_GGML_ASSERT(c->type == WSP_GGML_TYPE_I64 || c->type == WSP_GGML_TYPE_I32);
 
     WSP_GGML_ASSERT(wsp_ggml_is_contiguous_rows(a));
     WSP_GGML_ASSERT(wsp_ggml_is_contiguous_rows(b));
@@ -3690,6 +3695,7 @@ struct wsp_ggml_tensor * wsp_ggml_set_rows(
     result->op     = WSP_GGML_OP_SET_ROWS;
     result->src[0] = b;
     result->src[1] = c;
+    result->src[2] = a; // note: order is weird due to legacy reasons (https://github.com/ggml-org/llama.cpp/pull/16063#discussion_r2385795931)
 
     return result;
 }
@@ -3930,7 +3936,7 @@ static struct wsp_ggml_tensor * wsp_ggml_rope_impl(
     memcpy(params +  8, &attn_factor,  sizeof(float));
     memcpy(params +  9, &beta_fast,    sizeof(float));
     memcpy(params + 10, &beta_slow,    sizeof(float));
-    if (mrope_used) {
+    if (mrope_used && sections) {
         memcpy(params + 11, sections,  sizeof(int32_t) * WSP_GGML_MROPE_SECTIONS);
     } else {
         memset(params + 11, 0,         sizeof(int32_t) * WSP_GGML_MROPE_SECTIONS);
@@ -4367,6 +4373,91 @@ struct wsp_ggml_tensor * wsp_ggml_conv_2d(
     return result;
 }
 
+// a: [OC*IC, KD, KH, KW]
+// b: [N*IC, ID, IH, IW]
+// result: [N*OD, OH, OW, IC * KD * KH * KW]
+struct wsp_ggml_tensor * wsp_ggml_im2col_3d(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        struct wsp_ggml_tensor  * b,
+        int64_t               IC,
+        int                   s0, // stride width
+        int                   s1, // stride height
+        int                   s2, // stride depth
+        int                   p0, // padding width
+        int                   p1, // padding height
+        int                   p2, // padding depth
+        int                   d0, // dilation width
+        int                   d1, // dilation height
+        int                   d2, // dilation depth
+        enum wsp_ggml_type        dst_type) {
+    const int64_t N = b->ne[3] / IC;
+    const int64_t ID = b->ne[2];
+    const int64_t IH = b->ne[1];
+    const int64_t IW = b->ne[0];
+
+    const int64_t OC = a->ne[3] / IC;
+    UNUSED(OC);
+    const int64_t KD = a->ne[2];
+    const int64_t KH = a->ne[1];
+    const int64_t KW = a->ne[0];
+    const int64_t OD = wsp_ggml_calc_conv_output_size(ID, KD, s2, p2, d2);
+    const int64_t OH = wsp_ggml_calc_conv_output_size(IH, KH, s1, p1, d1);
+    const int64_t OW = wsp_ggml_calc_conv_output_size(IW, KW, s0, p0, d0);
+
+    WSP_GGML_ASSERT((OD > 0)  && "b too small compared to a");
+    WSP_GGML_ASSERT((OH > 0)  && "b too small compared to a");
+    WSP_GGML_ASSERT((OW > 0)  && "b too small compared to a");
+
+
+    const int64_t ne[4] = {KW*KH*KD*IC, OW, OH, OD*N};
+
+    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor(ctx, dst_type, 4, ne);
+    int32_t params[] = { s0, s1, s2, p0, p1, p2, d0, d1, d2, (int32_t)IC};
+    wsp_ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = WSP_GGML_OP_IM2COL_3D;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+// a: [OC*IC, KD, KH, KW]
+// b: [N*IC, ID, IH, IW]
+// result: [N*OC, OD, OH, OW]
+struct wsp_ggml_tensor * wsp_ggml_conv_3d(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        struct wsp_ggml_tensor  * b,
+        int64_t               IC,
+        int                   s0, // stride width
+        int                   s1, // stride height
+        int                   s2, // stride depth
+        int                   p0, // padding width
+        int                   p1, // padding height
+        int                   p2, // padding depth
+        int                   d0, // dilation width
+        int                   d1, // dilation height
+        int                   d2  // dilation depth
+        ) {
+    struct wsp_ggml_tensor * im2col = wsp_ggml_im2col_3d(ctx, a, b, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2, a->type); // [N*OD, OH, OW, IC * KD * KH * KW]
+
+    int64_t OC = a->ne[3] / IC;
+    int64_t N = b->ne[3] / IC;
+    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*OD, OH, OW, IC * KD * KH * KW] => [N*OD*OH*OW, IC * KD * KH * KW]
+                wsp_ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1] * a->ne[2] * IC), OC));                          // [OC*IC, KD, KH, KW] => [OC, IC * KD * KH * KW]
+
+    int64_t OD = im2col->ne[3] / N;
+    result = wsp_ggml_reshape_4d(ctx, result, im2col->ne[1]*im2col->ne[2], OD, N, OC); // [OC, N*OD*OH*OW] => [OC, N, OD, OH*OW]
+    result = wsp_ggml_cont(ctx, wsp_ggml_permute(ctx, result, 0, 1, 3, 2)); // [N, OC, OD, OH*OW]
+    result = wsp_ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], OD, OC * N); // [N*OC, OD, OH, OW]
+
+    return result;
+}
+
 // wsp_ggml_conv_2d_sk_p0
 
 struct wsp_ggml_tensor * wsp_ggml_conv_2d_sk_p0(
@@ -4488,6 +4579,56 @@ struct wsp_ggml_tensor * wsp_ggml_conv_2d_direct(
     return result;
 }
 
+// wsp_ggml_conv_3d_direct
+
+struct wsp_ggml_tensor * wsp_ggml_conv_3d_direct(
+        struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor  * a,
+        struct wsp_ggml_tensor  * b,
+        int                   s0,
+        int                   s1,
+        int                   s2,
+        int                   p0,
+        int                   p1,
+        int                   p2,
+        int                   d0,
+        int                   d1,
+        int                   d2,
+        int                   c,
+        int                   n,
+        int                   oc) {
+
+    WSP_GGML_ASSERT(a->ne[3] == (int64_t) c * oc);
+    WSP_GGML_ASSERT(b->ne[3] == (int64_t) c * n);
+
+    int64_t ne[4];
+    ne[0] = wsp_ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+    ne[1] = wsp_ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1);
+    ne[2] = wsp_ggml_calc_conv_output_size(b->ne[2], a->ne[2], s2, p2, d2);
+    ne[3] = (int64_t) oc * n;
+
+    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor(ctx, WSP_GGML_TYPE_F32, 4, ne);
+
+    wsp_ggml_set_op_params_i32(result, 0,  s0);
+    wsp_ggml_set_op_params_i32(result, 1,  s1);
+    wsp_ggml_set_op_params_i32(result, 2,  s2);
+    wsp_ggml_set_op_params_i32(result, 3,  p0);
+    wsp_ggml_set_op_params_i32(result, 4,  p1);
+    wsp_ggml_set_op_params_i32(result, 5,  p2);
+    wsp_ggml_set_op_params_i32(result, 6,  d0);
+    wsp_ggml_set_op_params_i32(result, 7,  d1);
+    wsp_ggml_set_op_params_i32(result, 8,  d2);
+    wsp_ggml_set_op_params_i32(result, 9,  c);
+    wsp_ggml_set_op_params_i32(result, 10, n);
+    wsp_ggml_set_op_params_i32(result, 11, oc);
+
+    result->op = WSP_GGML_OP_CONV_3D;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
 // wsp_ggml_conv_transpose_2d_p0
 
 static int64_t wsp_ggml_calc_conv_transpose_output_size(int64_t ins, int64_t ks, int s, int p) {
@@ -4666,11 +4807,36 @@ struct wsp_ggml_tensor * wsp_ggml_pad(
         int                   p1,
         int                   p2,
         int                   p3) {
+    return wsp_ggml_pad_ext(ctx, a, 0, p0, 0, p1, 0, p2, 0, p3);
+}
+
+struct wsp_ggml_tensor * wsp_ggml_pad_ext(
+            struct wsp_ggml_context * ctx,
+            struct wsp_ggml_tensor  * a,
+            int                  lp0,
+            int                  rp0,
+            int                  lp1,
+            int                  rp1,
+            int                  lp2,
+            int                  rp2,
+            int                  lp3,
+            int                  rp3
+            ) {
     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);
+            a->ne[0] + lp0 + rp0,
+            a->ne[1] + lp1 + rp1,
+            a->ne[2] + lp2 + rp2,
+            a->ne[3] + lp3 + rp3);
+
+    wsp_ggml_set_op_params_i32(result, 0, lp0);
+    wsp_ggml_set_op_params_i32(result, 1, rp0);
+    wsp_ggml_set_op_params_i32(result, 2, lp1);
+    wsp_ggml_set_op_params_i32(result, 3, rp1);
+    wsp_ggml_set_op_params_i32(result, 4, lp2);
+    wsp_ggml_set_op_params_i32(result, 5, rp2);
+    wsp_ggml_set_op_params_i32(result, 6, lp3);
+    wsp_ggml_set_op_params_i32(result, 7, rp3);
+
 
     result->op     = WSP_GGML_OP_PAD;
     result->src[0] = a;
@@ -4766,12 +4932,8 @@ struct wsp_ggml_tensor * wsp_ggml_timestep_embedding(
         struct wsp_ggml_tensor  * timesteps,
         int                   dim,
         int                   max_period) {
-    int actual_dim = dim;
-    if (dim % 2 != 0) {
-        actual_dim = dim + 1;
-    }
 
-    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor_2d(ctx, WSP_GGML_TYPE_F32, actual_dim, timesteps->ne[0]);
+    struct wsp_ggml_tensor * result = wsp_ggml_new_tensor_2d(ctx, WSP_GGML_TYPE_F32, dim, timesteps->ne[0]);
 
     wsp_ggml_set_op_params_i32(result, 0, dim);
     wsp_ggml_set_op_params_i32(result, 1, max_period);