tts : add OuteTTS support (llama/10784)

* server : add "tokens" output

ggml-ci

* server : output embeddings for all tokens when pooling = none

ggml-ci

* server : be explicit about the pooling type in the tests

ggml-ci

* server : do not normalize embeddings when there is no pooling

ggml-ci

* llama : add OuteTTS support (wip)

* wip

* extract features

* first conv

* group norm

* resnet conv

* resnet

* attn

* pos net

* layer norm

* convnext

* head

* hann window

* fix n_embd + remove llama.cpp hacks

* compute hann window

* fft

* spectrum processing

* clean-up

* tts : receive input text and generate codes

* clip : fix new conv name

* tts : minor fix

* tts : add header + minor fixes

ggml-ci

* tts : add matchematical constant

ggml-ci

* tts : fix sampling + cut initial noise

* tts : fixes

* tts : update default samplers

ggml-ci

* tts : text pre-processing

* tts : outetts-voc -> wavtokenizer-dec

* tts : remove hardcoded constants

ggml-ci

* tts : fix tensor shapes

* llama : refactor wavtokenizer tensors

ggml-ci

* cont

ggml-ci

* cont [no ci]

* llama : update WavTokenizer to non-causal attn

* llama : handle no-vocab detokenization

* tts : add Python example for OuteTTS (wip)

* tts : extend python example to generate spectrogram

ggml-ci

* server : fix rebase artifacts

* tts : enable "return_tokens" in Python example

ggml-ci

* tts : minor fixes

* common : support HF download for vocoder

ggml/include/ggml.h

@@ -1564,17 +1564,6 @@ extern "C" {
         int                   d1, // dilation dimension 1
         bool                  is_2D);
 
-    GGML_API struct ggml_tensor * ggml_conv_depthwise_2d(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,  // convolution kernel
-            struct ggml_tensor  * b,  // data
-            int                  s0,  // stride dimension 0
-            int                  s1,  // stride dimension 1
-            int                  p0,  // padding dimension 0
-            int                  p1,  // padding dimension 1
-            int                  d0,  // dilation dimension 0
-            int                  d1); // dilation dimension 1
-
     GGML_API struct ggml_tensor * ggml_conv_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,   // convolution kernel
@@ -1592,6 +1581,23 @@ extern "C" {
             int                   s,  // stride
             int                   d); // dilation
 
+    // depthwise
+    // TODO: this is very likely wrong for some cases! - needs more testing
+    GGML_API struct ggml_tensor * ggml_conv_1d_dw(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
+
+    GGML_API struct ggml_tensor * ggml_conv_1d_dw_ph(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   d0); // dilation
+
     GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,   // convolution kernel
@@ -1611,7 +1617,6 @@ extern "C" {
             int                   d0,  // dilation dimension 0
             int                   d1); // dilation dimension 1
 
-
     // kernel size is a->ne[0] x a->ne[1]
     // stride is equal to kernel size
     // padding is zero
@@ -1638,6 +1643,18 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // depthwise
+    GGML_API struct ggml_tensor * ggml_conv_2d_dw(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // convolution kernel
+            struct ggml_tensor  * b,  // data
+            int                  s0,  // stride dimension 0
+            int                  s1,  // stride dimension 1
+            int                  p0,  // padding dimension 0
+            int                  p1,  // padding dimension 1
+            int                  d0,  // dilation dimension 0
+            int                  d1); // dilation dimension 1
+
     GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,

ggml/src/ggml.c

@@ -3760,13 +3760,84 @@ struct ggml_tensor * ggml_clamp(
     return result;
 }
 
-// ggml_conv_1d
-
 static int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
     return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
 }
 
-GGML_API struct ggml_tensor * ggml_conv_1d(
+// im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
+// a: [OC，IC, KH, KW]
+// b: [N, IC, IH, IW]
+// result: [N, OH, OW, IC*KH*KW]
+struct ggml_tensor * ggml_im2col(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int                   s0,
+        int                   s1,
+        int                   p0,
+        int                   p1,
+        int                   d0,
+        int                   d1,
+        bool                  is_2D,
+        enum ggml_type        dst_type) {
+    if (is_2D) {
+        GGML_ASSERT(a->ne[2] == b->ne[2]);
+    } else {
+        //GGML_ASSERT(b->ne[1] % a->ne[1] == 0);
+        GGML_ASSERT(b->ne[1] == a->ne[1]);
+        GGML_ASSERT(b->ne[3] == 1);
+    }
+
+    const int64_t OH = is_2D ? ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0;
+    const int64_t OW =         ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+
+    GGML_ASSERT((!is_2D || OH > 0) && "b too small compared to a");
+    GGML_ASSERT((OW > 0)           && "b too small compared to a");
+
+    const int64_t ne[4] = {
+        is_2D ? (a->ne[2] * a->ne[1] * a->ne[0]) : a->ne[1] * a->ne[0],
+        OW,
+        is_2D ? OH : b->ne[2],
+        is_2D ?      b->ne[3] : 1,
+    };
+
+    struct ggml_tensor * result = ggml_new_tensor(ctx, dst_type, 4, ne);
+    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = GGML_OP_IM2COL;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_im2col_back(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int64_t             * ne,
+        int                   s0,
+        int                   s1,
+        int                   p0,
+        int                   p1,
+        int                   d0,
+        int                   d1,
+        bool                  is_2D) {
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = GGML_OP_IM2COL_BACK;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+// ggml_conv_1d
+
+struct ggml_tensor * ggml_conv_1d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
@@ -3796,137 +3867,75 @@ struct ggml_tensor* ggml_conv_1d_ph(
     return ggml_conv_1d(ctx, a, b, s, a->ne[0] / 2, d);
 }
 
-// ggml_conv_transpose_1d
-
-static int64_t ggml_calc_conv_transpose_1d_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
-    return (ins - 1) * s - 2 * p + d * (ks - 1) + 1;
-}
+// ggml_conv_1d_dw
 
-GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
+struct ggml_tensor * ggml_conv_1d_dw(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
         int                   s0,
         int                   p0,
         int                   d0) {
-    GGML_ASSERT(ggml_is_matrix(b));
-    GGML_ASSERT(a->ne[2] == b->ne[1]);
-    GGML_ASSERT(a->ne[3] == 1);
+    struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], 1, a->ne[1], a->ne[2]);
+    struct ggml_tensor * new_b = ggml_reshape_4d(ctx, b, b->ne[0], 1, b->ne[1], b->ne[2]);
 
-    GGML_ASSERT(p0 == 0);
-    GGML_ASSERT(d0 == 1);
+    struct ggml_tensor * im2col = ggml_im2col(ctx, new_a, new_b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16);
 
-    const int64_t ne[4] = {
-        ggml_calc_conv_transpose_1d_output_size(b->ne[0], a->ne[0], s0, 0 /*p0*/, 1 /*d0*/),
-        a->ne[1], b->ne[2], 1,
-    };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+    struct ggml_tensor * result = ggml_mul_mat(ctx, im2col, a);
 
-    int32_t params[] = { s0, p0, d0 };
-    ggml_set_op_params(result, params, sizeof(params));
-
-    result->op     = GGML_OP_CONV_TRANSPOSE_1D;
-    result->src[0] = a;
-    result->src[1] = b;
+    result = ggml_reshape_3d(ctx, result, b->ne[0], b->ne[1], 1);
 
     return result;
 }
 
-// ggml_conv_depthwise
+// ggml_conv_1d_dw_ph
 
-struct ggml_tensor * ggml_conv_depthwise_2d(
+struct ggml_tensor * ggml_conv_1d_dw_ph(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
         int                   s0,
-        int                   s1,
-        int                   p0,
-        int                   p1,
-        int                   d0,
-        int                   d1) {
-    struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]);
-    struct ggml_tensor * im2col = ggml_im2col(ctx, new_a,
-                                        ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]),
-                                        s0, s1, p0, p1, d0, d1, true, GGML_TYPE_F16); // [N * IC, OH, OW, KH * KW]
-    struct ggml_tensor * new_b = ggml_reshape_4d(ctx, im2col, im2col->ne[0], im2col->ne[2] * im2col->ne[1], b->ne[2], b->ne[3]); // [N * IC, OH, OW, KH * KW] => [N, IC, OH * OW, KH * KW]
+        int                   d0) {
+    return ggml_conv_1d_dw(ctx, a, b, s0, a->ne[0] / 2, d0);
+}
 
-    new_a = ggml_reshape_4d(ctx, new_a, (new_a->ne[0] * new_a->ne[1]), new_a->ne[2],  new_a->ne[3], 1);                       // [OC，1, KH, KW] => [1, OC, 1, KH * KW]
-    struct ggml_tensor * result = ggml_mul_mat(ctx, new_a, new_b);
-    result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], b->ne[2], b->ne[3]); // [N, OC, OH, OW]
+// ggml_conv_transpose_1d
 
-    return result;
+static int64_t ggml_calc_conv_transpose_1d_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
+    return (ins - 1) * s - 2 * p + d * (ks - 1) + 1;
 }
-// ggml_conv_2d
 
-// im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
-// a: [OC，IC, KH, KW]
-// b: [N, IC, IH, IW]
-// result: [N, OH, OW, IC*KH*KW]
-struct ggml_tensor * ggml_im2col(
+GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
         int                   s0,
-        int                   s1,
         int                   p0,
-        int                   p1,
-        int                   d0,
-        int                   d1,
-        bool                  is_2D,
-        enum ggml_type        dst_type) {
-    if(is_2D) {
-        GGML_ASSERT(a->ne[2] == b->ne[2]);
-    } else {
-        GGML_ASSERT(a->ne[1] == b->ne[1]);
-        GGML_ASSERT(b->ne[3] == 1);
-    }
-
-    const int64_t OH = is_2D ? ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0;
-    const int64_t OW =         ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+        int                   d0) {
+    GGML_ASSERT(ggml_is_matrix(b));
+    GGML_ASSERT(a->ne[2] == b->ne[1]);
+    GGML_ASSERT(a->ne[3] == 1);
 
-    GGML_ASSERT((!is_2D || OH > 0) && "b too small compared to a");
-    GGML_ASSERT((OW > 0)           && "b too small compared to a");
+    GGML_ASSERT(p0 == 0);
+    GGML_ASSERT(d0 == 1);
 
     const int64_t ne[4] = {
-        is_2D ? (a->ne[2] * a->ne[1] * a->ne[0]) : a->ne[1] * a->ne[0],
-        OW,
-        is_2D ? OH : b->ne[2],
-        is_2D ?      b->ne[3] : 1,
+        ggml_calc_conv_transpose_1d_output_size(b->ne[0], a->ne[0], s0, 0 /*p0*/, 1 /*d0*/),
+        a->ne[1], b->ne[2], 1,
     };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, dst_type, 4, ne);
-    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
+    int32_t params[] = { s0, p0, d0 };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op     = GGML_OP_IM2COL;
+    result->op     = GGML_OP_CONV_TRANSPOSE_1D;
     result->src[0] = a;
     result->src[1] = b;
 
     return result;
 }
 
-struct ggml_tensor * ggml_im2col_back(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        struct ggml_tensor  * b,
-        int64_t             * ne,
-        int                   s0,
-        int                   s1,
-        int                   p0,
-        int                   p1,
-        int                   d0,
-        int                   d1,
-        bool                  is_2D) {
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
-    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
-    ggml_set_op_params(result, params, sizeof(params));
-
-    result->op     = GGML_OP_IM2COL_BACK;
-    result->src[0] = a;
-    result->src[1] = b;
-
-    return result;
-}
+// ggml_conv_2d
 
 // a: [OC，IC, KH, KW]
 // b: [N, IC, IH, IW]
@@ -3973,6 +3982,31 @@ struct ggml_tensor * ggml_conv_2d_s1_ph(
     return ggml_conv_2d(ctx, a, b, 1, 1, a->ne[0] / 2, a->ne[1] / 2, 1, 1);
 }
 
+// ggml_conv_2d_dw
+
+struct ggml_tensor * ggml_conv_2d_dw(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int                   s0,
+        int                   s1,
+        int                   p0,
+        int                   p1,
+        int                   d0,
+        int                   d1) {
+    struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]);
+    struct ggml_tensor * im2col = ggml_im2col(ctx, new_a,
+                                        ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]),
+                                        s0, s1, p0, p1, d0, d1, true, GGML_TYPE_F16); // [N * IC, OH, OW, KH * KW]
+    struct ggml_tensor * new_b = ggml_reshape_4d(ctx, im2col, im2col->ne[0], im2col->ne[2] * im2col->ne[1], b->ne[2], b->ne[3]); // [N * IC, OH, OW, KH * KW] => [N, IC, OH * OW, KH * KW]
+
+    new_a = ggml_reshape_4d(ctx, new_a, (new_a->ne[0] * new_a->ne[1]), new_a->ne[2],  new_a->ne[3], 1);                       // [OC，1, KH, KW] => [1, OC, 1, KH * KW]
+    struct ggml_tensor * result = ggml_mul_mat(ctx, new_a, new_b);
+    result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], b->ne[2], b->ne[3]); // [N, OC, OH, OW]
+
+    return result;
+}
+
 // ggml_conv_transpose_2d_p0
 
 static int64_t ggml_calc_conv_transpose_output_size(int64_t ins, int64_t ks, int s, int p) {