@novastera-oss/llamarn 0.2.7 → 0.3.0

package/cpp/llama.cpp/ggml/src/ggml.c

@@ -61,9 +61,6 @@
 #define m512i(p) (__m512i)(p)
 #endif
 
-// precomputed f32 table for f16 (256 KB) (ggml-impl.h)
-float ggml_table_f32_f16[1 << 16];
-
 #if defined(__linux__) || \
     defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \
     (defined(__APPLE__) && !TARGET_OS_TV && !TARGET_OS_WATCH)
@@ -205,19 +202,34 @@ void ggml_print_backtrace(void) {
 }
 #endif
 
+static ggml_abort_callback_t g_abort_callback = NULL;
+
+// Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
+GGML_API ggml_abort_callback_t ggml_set_abort_callback(ggml_abort_callback_t callback) {
+    ggml_abort_callback_t ret_val = g_abort_callback;
+    g_abort_callback = callback;
+    return ret_val;
+}
+
 void ggml_abort(const char * file, int line, const char * fmt, ...) {
     fflush(stdout);
 
-    fprintf(stderr, "%s:%d: ", file, line);
+    char message[2048];
+    int offset = snprintf(message, sizeof(message), "%s:%d: ", file, line);
 
     va_list args;
     va_start(args, fmt);
-    vfprintf(stderr, fmt, args);
+    vsnprintf(message + offset, sizeof(message) - offset, fmt, args);
     va_end(args);
 
-    fprintf(stderr, "\n");
+    if (g_abort_callback) {
+        g_abort_callback(message);
+    } else {
+        // default: print error and backtrace to stderr
+        fprintf(stderr, "%s\n", message);
+        ggml_print_backtrace();
+    }
 
-    ggml_print_backtrace();
     abort();
 }
 
@@ -461,6 +473,14 @@ bool ggml_guid_matches(ggml_guid_t guid_a, ggml_guid_t guid_b) {
     return memcmp(guid_a, guid_b, sizeof(ggml_guid)) == 0;
 }
 
+const char * ggml_version(void) {
+    return GGML_VERSION;
+}
+
+const char * ggml_commit(void) {
+    return GGML_COMMIT;
+}
+
 //
 // timing
 //
@@ -936,6 +956,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "TRANSPOSE",
     "GET_ROWS",
     "GET_ROWS_BACK",
+    "SET_ROWS",
     "DIAG",
     "DIAG_MASK_INF",
     "DIAG_MASK_ZERO",
@@ -947,6 +968,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CONV_TRANSPOSE_1D",
     "IM2COL",
     "IM2COL_BACK",
+    "CONV_2D",
     "CONV_2D_DW",
     "CONV_TRANSPOSE_2D",
     "POOL_1D",
@@ -955,6 +977,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "UPSCALE",
     "PAD",
     "PAD_REFLECT_1D",
+    "ROLL",
     "ARANGE",
     "TIMESTEP_EMBEDDING",
     "ARGSORT",
@@ -983,9 +1006,11 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS",
     "CROSS_ENTROPY_LOSS_BACK",
     "OPT_STEP_ADAMW",
+
+    "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+static_assert(GGML_OP_COUNT == 86, "GGML_OP_COUNT != 86");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1031,6 +1056,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "transpose(x)",
     "get_rows(x)",
     "get_rows_back(x)",
+    "set_rows(x)",
     "diag(x)",
     "diag_mask_inf(x)",
     "diag_mask_zero(x)",
@@ -1042,6 +1068,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "conv_transpose_1d(x)",
     "im2col(x)",
     "im2col_back(x)",
+    "conv_2d(x)",
     "conv_2d_dw(x)",
     "conv_transpose_2d(x)",
     "pool_1d(x)",
@@ -1050,6 +1077,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "upscale(x)",
     "pad(x)",
     "pad_reflect_1d(x)",
+    "roll(x)",
     "arange(start, stop, step)",
     "timestep_embedding(timesteps, dim, max_period)",
     "argsort(x)",
@@ -1078,9 +1106,11 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss(x,y)",
     "cross_entropy_loss_back(x,y)",
     "adamw(x)",
+
+    "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+static_assert(GGML_OP_COUNT == 86, "GGML_OP_COUNT != 86");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -1106,6 +1136,17 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
 static_assert(GGML_UNARY_OP_COUNT == 15, "GGML_UNARY_OP_COUNT != 15");
 
 
+static const char * GGML_GLU_OP_NAME[GGML_GLU_OP_COUNT] = {
+    "REGLU",
+    "GEGLU",
+    "SWIGLU",
+    "GEGLU_ERF",
+    "GEGLU_QUICK",
+};
+
+static_assert(GGML_GLU_OP_COUNT == 5, "GGML_GLU_OP_COUNT != 5");
+
+
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
 static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN");
 
@@ -1208,11 +1249,19 @@ const char * ggml_unary_op_name(enum ggml_unary_op op) {
     return GGML_UNARY_OP_NAME[op];
 }
 
+const char * ggml_glu_op_name(enum ggml_glu_op op) {
+    return GGML_GLU_OP_NAME[op];
+}
+
 const char * ggml_op_desc(const struct ggml_tensor * t) {
     if (t->op == GGML_OP_UNARY) {
         enum ggml_unary_op uop = ggml_get_unary_op(t);
         return ggml_unary_op_name(uop);
     }
+    if (t->op == GGML_OP_GLU) {
+        enum ggml_glu_op gop = ggml_get_glu_op(t);
+        return ggml_glu_op_name(gop);
+    }
     return ggml_op_name(t->op);
 }
 
@@ -1349,6 +1398,12 @@ bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor) {
         tensor->nb[2] == ggml_type_size(tensor->type);
 }
 
+bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor) {
+    return
+        tensor->ne[0] == ggml_blck_size(tensor->type) ||
+        tensor->nb[0] == ggml_type_size(tensor->type);
+}
+
 static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
@@ -1420,14 +1475,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
         // initialize time system (required on Windows)
         ggml_time_init();
 
-        for (int i = 0; i < (1 << 16); ++i) {
-            union {
-                uint16_t u16;
-                ggml_fp16_t fp16;
-            } u = {i};
-            ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
-        }
-
         is_first_call = false;
     }
 
@@ -1731,6 +1778,11 @@ enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor) {
     return (enum ggml_unary_op) ggml_get_op_params_i32(tensor, 0);
 }
 
+enum ggml_glu_op ggml_get_glu_op(const struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->op == GGML_OP_GLU);
+    return (enum ggml_glu_op) ggml_get_op_params_i32(tensor, 0);
+}
+
 const char * ggml_get_name(const struct ggml_tensor * tensor) {
     return tensor->name;
 }
@@ -2610,6 +2662,156 @@ struct ggml_tensor * ggml_exp_inplace(
     return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_EXP);
 }
 
+// ggml_glu
+
+static struct ggml_tensor * ggml_glu_impl(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        enum ggml_glu_op      op,
+        bool                  swapped) {
+    GGML_ASSERT(ggml_is_contiguous_1(a));
+
+    if (b) {
+        GGML_ASSERT(ggml_is_contiguous_1(b));
+        GGML_ASSERT(ggml_are_same_shape(a, b));
+        GGML_ASSERT(a->type == b->type);
+    }
+
+    int64_t ne[GGML_MAX_DIMS] = { a->ne[0] / 2 }; for (int i = 1; i < GGML_MAX_DIMS; i++) ne[i] = a->ne[i];
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, GGML_MAX_DIMS, b ? a->ne : ne, NULL, 0);
+
+    ggml_set_op_params_i32(result, 0, (int32_t) op);
+    ggml_set_op_params_i32(result, 1, (int32_t) swapped);
+
+    result->op     = GGML_OP_GLU;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_glu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        enum ggml_glu_op      op,
+        bool                  swapped) {
+    return ggml_glu_impl(ctx, a, NULL, op, swapped);
+}
+
+struct ggml_tensor * ggml_glu_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        enum ggml_glu_op      op) {
+    return ggml_glu_impl(ctx, a, b, op, false);
+}
+
+// ggml_reglu
+
+struct ggml_tensor * ggml_reglu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_REGLU, false);
+}
+
+struct ggml_tensor * ggml_reglu_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_REGLU, true);
+}
+
+struct ggml_tensor * ggml_reglu_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_REGLU, false);
+}
+
+// ggml_geglu
+
+struct ggml_tensor * ggml_geglu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU, false);
+}
+
+struct ggml_tensor * ggml_geglu_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU, true);
+}
+
+struct ggml_tensor * ggml_geglu_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_GEGLU, false);
+}
+
+// ggml_swiglu
+
+struct ggml_tensor * ggml_swiglu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_SWIGLU, false);
+}
+
+struct ggml_tensor * ggml_swiglu_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_SWIGLU, true);
+}
+
+struct ggml_tensor * ggml_swiglu_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_SWIGLU, false);
+}
+
+// ggml_geglu_erf
+
+struct ggml_tensor * ggml_geglu_erf(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_ERF, false);
+}
+
+struct ggml_tensor * ggml_geglu_erf_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_ERF, true);
+}
+
+struct ggml_tensor * ggml_geglu_erf_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_GEGLU_ERF, false);
+}
+
+// ggml_geglu_quick
+
+struct ggml_tensor * ggml_geglu_quick(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_QUICK, false);
+}
+
+struct ggml_tensor * ggml_geglu_quick_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_QUICK, true);
+}
+
+struct ggml_tensor * ggml_geglu_quick_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_GEGLU_QUICK, false);
+}
+
 // ggml_norm
 
 static struct ggml_tensor * ggml_norm_impl(
@@ -2867,12 +3069,14 @@ static struct ggml_tensor * ggml_scale_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         float                 s,
+        float                 b,
         bool                  inplace) {
     GGML_ASSERT(ggml_is_padded_1d(a));
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    ggml_set_op_params(result, &s, sizeof(s));
+    float params[2] = { s, b };
+    ggml_set_op_params(result, &params, sizeof(params));
 
     result->op     = GGML_OP_SCALE;
     result->src[0] = a;
@@ -2884,14 +3088,30 @@ struct ggml_tensor * ggml_scale(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         float                 s) {
-    return ggml_scale_impl(ctx, a, s, false);
+    return ggml_scale_impl(ctx, a, s, 0.0, false);
 }
 
 struct ggml_tensor * ggml_scale_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         float                 s) {
-    return ggml_scale_impl(ctx, a, s, true);
+    return ggml_scale_impl(ctx, a, s, 0.0, true);
+}
+
+struct ggml_tensor * ggml_scale_bias(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b) {
+    return ggml_scale_impl(ctx, a, s, b, false);
+}
+
+struct ggml_tensor * ggml_scale_bias_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b) {
+    return ggml_scale_impl(ctx, a, s, b, true);
 }
 
 // ggml_set
@@ -3393,6 +3613,35 @@ struct ggml_tensor * ggml_get_rows_back(
     return result;
 }
 
+// ggml_set_rows
+
+struct ggml_tensor * ggml_set_rows(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        struct ggml_tensor  * c) {
+    GGML_ASSERT(a->ne[0] == b->ne[0]);
+    GGML_ASSERT(a->ne[2] == b->ne[2]);
+    GGML_ASSERT(a->ne[3] == b->ne[3]);
+    GGML_ASSERT(b->ne[1] == c->ne[0]);
+    GGML_ASSERT(b->ne[2] % c->ne[1] == 0);
+    GGML_ASSERT(b->ne[3] % c->ne[2] == 0);
+    GGML_ASSERT(c->ne[3] == 1);
+    GGML_ASSERT(b->type == GGML_TYPE_F32);
+    GGML_ASSERT(c->type == GGML_TYPE_I64);
+
+    GGML_ASSERT(ggml_is_contiguous_rows(a));
+    GGML_ASSERT(ggml_is_contiguous_rows(b));
+
+    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
+
+    result->op     = GGML_OP_SET_ROWS;
+    result->src[0] = b;
+    result->src[1] = c;
+
+    return result;
+}
+
 // ggml_diag
 
 struct ggml_tensor * ggml_diag(
@@ -3487,9 +3736,10 @@ static struct ggml_tensor * ggml_soft_max_impl(
     if (mask) {
         GGML_ASSERT(mask->type == GGML_TYPE_F16 || mask->type == GGML_TYPE_F32);
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(ggml_is_matrix(mask));
         GGML_ASSERT(mask->ne[0] == a->ne[0]);
         GGML_ASSERT(mask->ne[1] >= a->ne[1]);
+        GGML_ASSERT(a->ne[2]%mask->ne[2] == 0);
+        GGML_ASSERT(a->ne[3]%mask->ne[3] == 0);
     }
 
     if (max_bias > 0.0f) {
@@ -4129,6 +4379,44 @@ struct ggml_tensor * ggml_conv_2d_dw_direct(
     return result;
 }
 
+// ggml_conv_2d_direct
+
+struct ggml_tensor * ggml_conv_2d_direct(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,   // convolution kernel [KW, KH, IC, OC]
+        struct ggml_tensor  * b,   // input data [W, H, C, N]
+        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_ASSERT(a->ne[2] == b->ne[2]);
+    //GGML_ASSERT(a->type == b->type);
+
+    int64_t ne[4];
+    ne[0] = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+    ne[1] = ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1);
+    ne[2] = a->ne[3];
+    ne[3] = b->ne[3];
+
+    struct ggml_tensor * result = ggml_new_tensor(ctx, b->type, 4, ne);
+
+    ggml_set_op_params_i32(result, 0, s0);
+    ggml_set_op_params_i32(result, 1, s1);
+    ggml_set_op_params_i32(result, 2, p0);
+    ggml_set_op_params_i32(result, 3, p1);
+    ggml_set_op_params_i32(result, 4, d0);
+    ggml_set_op_params_i32(result, 5, d1);
+
+    result->op = GGML_OP_CONV_2D;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    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) {
@@ -4245,24 +4533,21 @@ struct ggml_tensor * ggml_pool_2d_back(
     return result;
 }
 
-// ggml_upscale
+// ggml_upscale / ggml_interpolate
 
-static struct ggml_tensor * ggml_upscale_impl(
+static struct ggml_tensor * ggml_interpolate_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
-        int                   ne2,
-        int                   ne3,
-        enum ggml_scale_mode  mode) {
-    GGML_ASSERT(a->ne[0] <= ne0);
-    GGML_ASSERT(a->ne[1] <= ne1);
-    GGML_ASSERT(a->ne[2] <= ne2);
-    GGML_ASSERT(a->ne[3] <= ne3);
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2,
+        int64_t               ne3,
+        uint32_t              mode) {
+    GGML_ASSERT((mode & 0xFF) < GGML_SCALE_MODE_COUNT);
 
     struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);
 
-    ggml_set_op_params_i32(result, 0, mode);
+    ggml_set_op_params_i32(result, 0, (int32_t)mode);
 
     result->op     = GGML_OP_UPSCALE;
     result->src[0] = a;
@@ -4275,7 +4560,8 @@ struct ggml_tensor * ggml_upscale(
         struct ggml_tensor  * a,
         int                   scale_factor,
         enum ggml_scale_mode  mode) {
-    return ggml_upscale_impl(ctx, a, a->ne[0] * scale_factor, a->ne[1] * scale_factor, a->ne[2], a->ne[3], mode);
+    GGML_ASSERT(scale_factor > 1);
+    return ggml_interpolate_impl(ctx, a, a->ne[0] * scale_factor, a->ne[1] * scale_factor, a->ne[2], a->ne[3], mode);
 }
 
 struct ggml_tensor * ggml_upscale_ext(
@@ -4286,7 +4572,18 @@ struct ggml_tensor * ggml_upscale_ext(
         int                   ne2,
         int                   ne3,
         enum ggml_scale_mode  mode) {
-    return ggml_upscale_impl(ctx, a, ne0, ne1, ne2, ne3, mode);
+    return ggml_interpolate_impl(ctx, a, ne0, ne1, ne2, ne3, mode);
+}
+
+struct ggml_tensor * ggml_interpolate(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2,
+        int64_t               ne3,
+        uint32_t              mode) {
+    return ggml_interpolate_impl(ctx, a, ne0, ne1, ne2, ne3, mode);
 }
 
 // ggml_pad
@@ -4341,6 +4638,34 @@ struct ggml_tensor * ggml_pad_reflect_1d(
     return result;
 }
 
+// ggml_roll
+
+struct ggml_tensor * ggml_roll(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   shift0,
+        int                   shift1,
+        int                   shift2,
+        int                   shift3) {
+    GGML_ASSERT(a->nb[0] == ggml_type_size(a->type));
+    GGML_ASSERT(abs(shift0) < a->ne[0]);
+    GGML_ASSERT(abs(shift1) < a->ne[1]);
+    GGML_ASSERT(abs(shift2) < a->ne[2]);
+    GGML_ASSERT(abs(shift3) < a->ne[3]);
+
+    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
+
+    ggml_set_op_params_i32(result, 0, shift0);
+    ggml_set_op_params_i32(result, 1, shift1);
+    ggml_set_op_params_i32(result, 2, shift2);
+    ggml_set_op_params_i32(result, 3, shift3);
+
+    result->op     = GGML_OP_ROLL;
+    result->src[0] = a;
+
+    return result;
+}
+
 // ggml_arange
 
 struct ggml_tensor * ggml_arange(
@@ -4435,13 +4760,17 @@ struct ggml_tensor * ggml_flash_attn_ext(
     GGML_ASSERT(ggml_can_mul_mat(k, q));
     // TODO: check if vT can be multiplied by (k*qT)
 
+    GGML_ASSERT(q->ne[3] == k->ne[3]);
+    GGML_ASSERT(q->ne[3] == v->ne[3]);
+
     if (mask) {
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(mask->ne[2] == 1);
-        GGML_ASSERT(mask->ne[3] == 1);
         GGML_ASSERT(mask->ne[1] >= GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) &&
                 "the Flash-Attention kernel requires the mask to be padded to GGML_KQ_MASK_PAD and at least n_queries big");
         //GGML_ASSERT(ggml_can_repeat_rows(mask, qk));
+
+        GGML_ASSERT(q->ne[2] % mask->ne[2] == 0);
+        GGML_ASSERT(q->ne[3] % mask->ne[3] == 0);
     }
 
     if (max_bias > 0.0f) {
@@ -4569,7 +4898,6 @@ struct ggml_tensor * ggml_ssm_conv(
     const int64_t n_s     = sx->ne[2];
 
     // TODO: maybe support other strides than 1?
-    // FIXME: this is always true?
     GGML_ASSERT(sx->ne[0] == d_conv - 1 + n_t);
     GGML_ASSERT(sx->ne[1] == d_inner);
     GGML_ASSERT(n_t >= 0);
@@ -4592,36 +4920,49 @@ struct ggml_tensor * ggml_ssm_scan(
         struct ggml_tensor  * dt,
         struct ggml_tensor  * A,
         struct ggml_tensor  * B,
-        struct ggml_tensor  * C) {
+        struct ggml_tensor  * C,
+        struct ggml_tensor  * ids) {
     GGML_ASSERT(ggml_is_contiguous(s));
-    GGML_ASSERT(ggml_is_contiguous(x));
     GGML_ASSERT(ggml_is_contiguous(dt));
     GGML_ASSERT(ggml_is_contiguous(A));
-    GGML_ASSERT(ggml_is_matrix(A));
-    GGML_ASSERT(ggml_is_3d(B));
-    GGML_ASSERT(ggml_is_3d(s));
+    GGML_ASSERT(x->nb[0] == ggml_type_size(x->type));
     GGML_ASSERT(B->nb[0] == ggml_type_size(B->type));
     GGML_ASSERT(C->nb[0] == ggml_type_size(C->type));
-    GGML_ASSERT(ggml_are_same_shape(x, dt));
+    GGML_ASSERT(x->nb[1] == x->ne[0]*x->nb[0]);
+    GGML_ASSERT(B->nb[1] == B->ne[0]*B->nb[0]);
+    GGML_ASSERT(C->nb[1] == C->ne[0]*C->nb[0]);
     GGML_ASSERT(ggml_are_same_shape(B, C));
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
 
     {
         const int64_t d_state      = s->ne[0];
-        const int64_t d_inner      = s->ne[1];
-        const int64_t n_seq_tokens = x->ne[1];
-        const int64_t n_seqs       = x->ne[2];
-
-        GGML_ASSERT(s->ne[2] == n_seqs);
-        GGML_ASSERT(x->ne[0] == d_inner);
-        GGML_ASSERT(A->ne[0] == d_state);
-        GGML_ASSERT(A->ne[1] == d_inner);
+        const int64_t head_dim     = x->ne[0];
+        const int64_t n_head       = x->ne[1];
+        const int64_t n_seq_tokens = x->ne[2];
+        const int64_t n_seqs       = x->ne[3];
+
+        GGML_ASSERT(dt->ne[0] == n_head);
+        GGML_ASSERT(dt->ne[1] == n_seq_tokens);
+        GGML_ASSERT(dt->ne[2] == n_seqs);
+        GGML_ASSERT(ggml_is_3d(dt));
+        GGML_ASSERT(s->ne[1] == head_dim);
+        GGML_ASSERT(s->ne[2] == n_head);
         GGML_ASSERT(B->ne[0] == d_state);
-        GGML_ASSERT(B->ne[1] == n_seq_tokens);
-        GGML_ASSERT(B->ne[2] == n_seqs);
+        GGML_ASSERT(B->ne[2] == n_seq_tokens);
+        GGML_ASSERT(B->ne[3] == n_seqs);
+        GGML_ASSERT(ids->ne[0] == n_seqs);
+        GGML_ASSERT(ggml_is_vector(ids));
+        GGML_ASSERT(A->ne[1] == n_head);
+        GGML_ASSERT(ggml_is_matrix(A));
+
+        if (A->ne[0] != 1) {
+            // Mamba-1 has more granular decay factors
+            GGML_ASSERT(A->ne[0] == d_state);
+        }
     }
 
     // concatenated y + ssm_states
-    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + ggml_nelements(s));
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + s->ne[0]*s->ne[1]*s->ne[2]*ids->ne[0]);
 
     result->op   = GGML_OP_SSM_SCAN;
     result->src[0] = s;
@@ -4630,6 +4971,7 @@ struct ggml_tensor * ggml_ssm_scan(
     result->src[3] = A;
     result->src[4] = B;
     result->src[5] = C;
+    result->src[6] = ids;
 
     return result;
 }
@@ -5453,7 +5795,7 @@ static void ggml_compute_backward(
         } break;
         case GGML_OP_MEAN: {
             if (src0_needs_grads) {
-                ggml_add1_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, 1.0f/src0->ne[0], false));
+                ggml_add1_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, 1.0f/src0->ne[0], 0.0, false));
             }
         } break;
         case GGML_OP_REPEAT: {
@@ -5530,7 +5872,7 @@ static void ggml_compute_backward(
             if (src0_needs_grads) {
                 float s;
                 memcpy(&s, tensor->op_params, sizeof(float));
-                ggml_add_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, s, false));
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, s, 0.0, false));
             }
         } break;
         case GGML_OP_SET: {
@@ -5770,13 +6112,28 @@ static void ggml_compute_backward(
             }
             GGML_ASSERT(!src1_needs_grads && "backward pass for labels not implemented");
         } break;
+        case GGML_OP_GLU: {
+            switch (ggml_get_glu_op(tensor)) {
+                case GGML_GLU_OP_SWIGLU: {
+                    if (src0_needs_grads) {
+                        GGML_ASSERT(src1 && "backward pass only implemented for split swiglu");
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_silu_back(ctx, ggml_mul(ctx, grad, src1), src0));
+                    }
+                    if (src1_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc1, ggml_mul(ctx, ggml_silu(ctx, src0), grad));
+                    }
+                } break;
+                default: {
+                    GGML_ABORT("unsupported glu op for backward pass: %s", ggml_glu_op_name(ggml_get_glu_op(tensor)));
+                } //break;
+            }
+        } break;
         case GGML_OP_NONE: {
             // noop
         } break;
         case GGML_OP_COUNT:
         default: {
-            fprintf(stderr, "%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
-            GGML_ABORT("fatal error");
+            GGML_ABORT("%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
         } //break;
     }
 
@@ -5785,19 +6142,32 @@ static void ggml_compute_backward(
     GGML_ASSERT(!src2_needs_grads || ggml_are_same_shape(src2, cgraph->grads[isrc2]));
 }
 
-static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
+static size_t ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
     // check if already visited
-    if (ggml_hash_insert(&cgraph->visited_hash_set, node) == GGML_HASHSET_ALREADY_EXISTS) {
-        return;
+    size_t node_hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
+    GGML_ASSERT(node_hash_pos != GGML_HASHSET_FULL);
+    if (!ggml_bitset_get(cgraph->visited_hash_set.used, node_hash_pos)) {
+        // This is the first time we see this node in the current graph.
+        cgraph->visited_hash_set.keys[node_hash_pos] = node;
+        ggml_bitset_set(cgraph->visited_hash_set.used, node_hash_pos);
+        cgraph->use_counts[node_hash_pos] = 0;
+    } else {
+        // already visited
+        return node_hash_pos;
     }
 
     for (int i = 0; i < GGML_MAX_SRC; ++i) {
         const int k =
             (cgraph->order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? i :
             (cgraph->order == GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT) ? (GGML_MAX_SRC-1-i) :
-            /* unknown order, just fall back to using i*/ i;
-        if (node->src[k]) {
-            ggml_visit_parents(cgraph, node->src[k]);
+            /* unknown order, just fall back to using i */ i;
+
+        struct ggml_tensor * src = node->src[k];
+        if (src) {
+            size_t src_hash_pos = ggml_visit_parents(cgraph, src);
+
+            // Update the use count for this operand.
+            cgraph->use_counts[src_hash_pos]++;
         }
     }
 
@@ -5821,6 +6191,8 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
         cgraph->nodes[cgraph->n_nodes] = node;
         cgraph->n_nodes++;
     }
+
+    return node_hash_pos;
 }
 
 static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand) {
@@ -5958,6 +6330,7 @@ static size_t ggml_graph_nbytes(size_t size, bool grads) {
     incr_ptr_aligned(&p, sizeof(struct ggml_cgraph), 1);
     incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // nodes
     incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // leafs
+    incr_ptr_aligned(&p, hash_size * sizeof(int32_t), sizeof(int32_t)); // use_counts
     incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // hash keys
     if (grads) {
         incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grads
@@ -5987,11 +6360,12 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
 
     void * p = cgraph + 1;
 
-    struct ggml_tensor ** nodes_ptr     =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** leafs_ptr     =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** hash_keys_ptr =         incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** grads_ptr     = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
-    struct ggml_tensor ** grad_accs_ptr = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+    struct ggml_tensor ** nodes_ptr      =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** leafs_ptr      =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    int32_t             * use_counts_ptr =         incr_ptr_aligned(&p, hash_size * sizeof(int32_t), sizeof(int32_t));
+    struct ggml_tensor ** hash_keys_ptr  =         incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** grads_ptr      = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+    struct ggml_tensor ** grad_accs_ptr  = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
 
     ggml_bitset_t * hash_used = incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
 
@@ -6006,6 +6380,7 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
         /*.grads        =*/ grads_ptr,
         /*.grad_accs    =*/ grad_accs_ptr,
         /*.leafs        =*/ leafs_ptr,
+        /*.use_counts   =*/ use_counts_ptr,
         /*.hash_table   =*/ { hash_size, hash_used, hash_keys_ptr },
         /*.order        =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT,
     };
@@ -6032,7 +6407,8 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1)
         /*.grads            =*/ NULL, // gradients would need visited_hash_set
         /*.grad_accs        =*/ NULL,
         /*.leafs            =*/ NULL,
-        /*.visited_hash_set =*/ { 0, NULL, NULL },
+        /*.use_counts       =*/ cgraph0->use_counts,
+        /*.visited_hash_set =*/ cgraph0->visited_hash_set,
         /*.order            =*/ cgraph0->order,
     };
 
@@ -6059,7 +6435,8 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
     for (size_t i = 0; i < src->visited_hash_set.size; ++i) {
         // copy all hashset keys (tensors) that are in use
         if (ggml_bitset_get(src->visited_hash_set.used, i)) {
-            ggml_hash_insert(&dst->visited_hash_set, src->visited_hash_set.keys[i]);
+            size_t new_hash_pos = ggml_hash_insert(&dst->visited_hash_set, src->visited_hash_set.keys[i]);
+            dst->use_counts[new_hash_pos] = src->use_counts[i];
         }
     }