Commit 1fa7f1c48 for llama.cpp
commit 1fa7f1c488cc92b8028ae780d12cff56b5113239
Author: Takamichi Ohara <tokinasin@tutanota.com>
Date: Thu Nov 5 19:02:05 2026 +0900
ggml : support `truncate=false` for YaRN (#30206)
* ggml : support `truncate=false` for YaRN
YaRN calculates correction intervals to switch frequency interpolation
methods, but the setting to not round the boundaries of these intervals
to integers (`truncate=false`) was not supported.
Models that use `truncate=false` include gpt-oss and some PLaMo-3
models; to support them correctly, the GGML operations and CPU backend
are modified to handle calculations without rounding. The default
behavior is `truncate=true`, and existing calls not explicitly
configured will maintain their previous behavior.
* support CUDA backend
* add tests
* move RoPE helpers to internal header
* format
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 3ed418140..5a5ba966b 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -2034,10 +2034,6 @@ extern "C" {
float beta_slow),
"use ggml_rope_ext_inplace instead");
- // compute correction dims for YaRN RoPE scaling
- GGML_API void ggml_rope_yarn_corr_dims(
- int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
-
// rotary position embedding backward, i.e compute dx from dy
// a - dy
GGML_API struct ggml_tensor * ggml_rope_ext_back(
@@ -2080,6 +2076,11 @@ extern "C" {
struct ggml_tensor * a,
int n_offs);
+ // control rounding of YaRN correction bounds (default: true)
+ GGML_API struct ggml_tensor * ggml_rope_set_truncate(
+ struct ggml_tensor * a,
+ bool truncate);
+
// im2col
// converts data into a format that effectively results in a convolution when combined with matrix multiplication
GGML_API struct ggml_tensor * ggml_im2col(
diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index 902d2eda6..66e011ef0 100644
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -2934,7 +2934,7 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
const float theta_scale = powf(freq_base, -2.0f / n_dims);
float corr_dims[2];
- ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+ ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, true, corr_dims);
bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
@@ -3289,7 +3289,7 @@ void ggml_cann_rope_cache_preload(ggml_backend_cann_context & ctx, ggml_tensor *
const float theta_scale = powf(freq_base, -2.0f / n_dims);
float corr_dims[2];
- ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+ ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, true, corr_dims);
bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index d3a47f839..fd2c80bd6 100644
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2547,6 +2547,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
}
case GGML_OP_ROPE:
{
+ if (!ggml_rope_get_truncate(op)) {
+ return false;
+ }
if (((const int32_t *) op->op_params)[15] != 0) {
return false; // FIXME: support ggml_rope_set_offset
}
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 007128338..55f625557 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -6310,8 +6310,8 @@ static void ggml_compute_forward_rope_flt(
//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];
const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+ const bool truncate = ggml_rope_get_truncate(dst);
memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
@@ -6356,7 +6356,7 @@ static void ggml_compute_forward_rope_flt(
const float theta_scale = powf(freq_base, -2.0f/n_dims);
float corr_dims[2];
- ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+ ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, truncate, corr_dims);
const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
const bool mrope_used = mode & GGML_ROPE_TYPE_MROPE; // ggml_rope_multi, note: also true for vision (24 & 8 == true) and for imrope
diff --git a/ggml/src/ggml-cuda/rope.cu b/ggml/src/ggml-cuda/rope.cu
index 45e813406..2604d014f 100644
--- a/ggml/src/ggml-cuda/rope.cu
+++ b/ggml/src/ggml-cuda/rope.cu
@@ -591,7 +591,6 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
//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];
const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
const int n_offs = ((int32_t *) dst->op_params)[15];
mrope_sections sections;
@@ -637,7 +636,7 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
}
rope_corr_dims corr_dims;
- ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v);
+ ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, ggml_rope_get_truncate(dst), corr_dims.v);
// compute
if (is_neox) {
@@ -929,7 +928,7 @@ void ggml_cuda_op_rms_norm_mul_rope_fused(ggml_backend_cuda_context & ctx,
const float * freq_factors = rope->src[2] != nullptr ? (const float *) rope->src[2]->data : nullptr;
rope_corr_dims corr_dims;
- ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v);
+ ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, ggml_rope_get_truncate(rope), corr_dims.v);
const size_t ts0 = ggml_type_size(x->type);
GGML_ASSERT(x->nb[0] == ts0);
diff --git a/ggml/src/ggml-et/ggml-et.cpp b/ggml/src/ggml-et/ggml-et.cpp
index cdd947c52..466e68904 100644
--- a/ggml/src/ggml-et/ggml-et.cpp
+++ b/ggml/src/ggml-et/ggml-et.cpp
@@ -1069,7 +1069,7 @@ static bool ggml_backend_et_device_supports_op(ggml_backend_dev_t dev, const ggm
const bool zero_rot_offset = ggml_get_op_params_i32(op, 15) == 0;
supported =
- zero_view_offset && zero_rot_offset && ndims <= 512 &&
+ ggml_rope_get_truncate(op) && zero_view_offset && zero_rot_offset && ndims <= 512 &&
(is_normal || (is_neox && ndims % 16 == 0) || (is_imrope && ndims % 16 == 0 && has_sections));
} else {
supported = false;
diff --git a/ggml/src/ggml-hexagon/ggml-hexagon.cpp b/ggml/src/ggml-hexagon/ggml-hexagon.cpp
index 47e90b030..a883c585c 100644
--- a/ggml/src/ggml-hexagon/ggml-hexagon.cpp
+++ b/ggml/src/ggml-hexagon/ggml-hexagon.cpp
@@ -2352,6 +2352,7 @@ ggml_hexagon_session::~ggml_hexagon_session() noexcept(true) {
// ** backend interface
+
static const char * ggml_backend_hexagon_name(ggml_backend_t backend) {
auto sess = static_cast<ggml_hexagon_session *>(backend->context);
return sess->c_name();
diff --git a/ggml/src/ggml-hexagon/htp-kparams.cpp b/ggml/src/ggml-hexagon/htp-kparams.cpp
index 48bd93277..822d1eb08 100644
--- a/ggml/src/ggml-hexagon/htp-kparams.cpp
+++ b/ggml/src/ggml-hexagon/htp-kparams.cpp
@@ -1874,6 +1874,10 @@ static bool htp_kparams_solve_rope(
const struct ggml_tensor * src2 = op->src[2];
const struct ggml_tensor * dst = op;
+ if (!ggml_rope_get_truncate(op)) {
+ return false;
+ }
+
if (!ggml_are_same_shape(src0, dst)) {
return false;
}
diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
index 4dc55e78e..7c6a605af 100644
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@@ -107,6 +107,13 @@ static inline bool ggml_impl_is_view(const struct ggml_tensor * t) {
static inline float ggml_compute_softplus_f32(float input) {
return (input > 20.0f) ? input : logf(1 + expf(input));
}
+
+// compute correction dims for YaRN RoPE scaling
+GGML_API void ggml_rope_yarn_corr_dims(
+ int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, bool truncate, float dims[2]);
+
+GGML_API bool ggml_rope_get_truncate(const struct ggml_tensor * a);
+
//
// logging
//
diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.m
index 68a9addb4..63fbe2734 100644
--- a/ggml/src/ggml-metal/ggml-metal-device.m
+++ b/ggml/src/ggml-metal/ggml-metal-device.m
@@ -1696,7 +1696,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
return has_simdgroup_reduction && (ggml_is_contiguous_rows(op->src[0]));
case GGML_OP_ROPE:
case GGML_OP_ROPE_BACK:
- return true;
+ return ggml_rope_get_truncate(op);
case GGML_OP_IM2COL:
return ggml_is_contiguous(op->src[1]) && op->src[1]->type == GGML_TYPE_F32 && (op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32);
case GGML_OP_CONV_2D:
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index dc096dbb0..983e61cca 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -12025,6 +12025,9 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
return op->ne[3] == 1;
case GGML_OP_ROPE: {
const int mode = ((const int32_t *) op->op_params)[2];
+ if (!ggml_rope_get_truncate(op)) {
+ return false;
+ }
const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
if (is_mrope && !is_vision) {
diff --git a/ggml/src/ggml-openvino/ggml-openvino.cpp b/ggml/src/ggml-openvino/ggml-openvino.cpp
index 88634557c..a63f69bf6 100644
--- a/ggml/src/ggml-openvino/ggml-openvino.cpp
+++ b/ggml/src/ggml-openvino/ggml-openvino.cpp
@@ -1636,6 +1636,9 @@ static ggml_openvino_op_support is_op_supported_case(const ggml_tensor * op) {
break;
}
case GGML_OP_ROPE: {
+ if (!ggml_rope_get_truncate(op)) {
+ return {false, "ROPE without YaRN truncation is not supported"};
+ }
if (op->view_src != nullptr && !ggml_is_contiguous(op->src[0])) {
return {false, "ROPE on VIEW / non-contiguous input is not supported"};
}
diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
index 966bd9975..d346bb12b 100644
--- a/ggml/src/ggml-rpc/ggml-rpc.cpp
+++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
@@ -3055,7 +3055,10 @@ static ggml_backend_buffer_type_t ggml_backend_rpc_device_get_buffer_type(ggml_b
static bool ggml_backend_rpc_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
GGML_UNUSED(dev);
- GGML_UNUSED(op);
+ if ((op->op == GGML_OP_ROPE || op->op == GGML_OP_ROPE_BACK) &&
+ !ggml_rope_get_truncate(op)) {
+ return false;
+ }
//TODO: call the remote backend and cache the results
return true;
}
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index 9ee16423a..ac752f9b8 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -7569,6 +7569,7 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
}
case GGML_OP_ROPE:
case GGML_OP_ROPE_BACK:
+ return ggml_rope_get_truncate(op);
case GGML_OP_IM2COL:
case GGML_OP_IM2COL_3D:
case GGML_OP_UPSCALE:
diff --git a/ggml/src/ggml-sycl/rope.cpp b/ggml/src/ggml-sycl/rope.cpp
index b6d22559d..f2ca60349 100644
--- a/ggml/src/ggml-sycl/rope.cpp
+++ b/ggml/src/ggml-sycl/rope.cpp
@@ -547,7 +547,7 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
rope_corr_dims corr_dims;
ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast,
- beta_slow, corr_dims.v);
+ beta_slow, true, corr_dims.v);
// compute
if (is_neox) {
diff --git a/ggml/src/ggml-virtgpu/ggml-backend-device.cpp b/ggml/src/ggml-virtgpu/ggml-backend-device.cpp
index 987ce9dd1..86df4cc5f 100644
--- a/ggml/src/ggml-virtgpu/ggml-backend-device.cpp
+++ b/ggml/src/ggml-virtgpu/ggml-backend-device.cpp
@@ -30,11 +30,14 @@ static void ggml_backend_remoting_device_get_memory(ggml_backend_dev_t dev, size
}
static bool ggml_backend_remoting_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+ if ((op->op == GGML_OP_ROPE || op->op == GGML_OP_ROPE_BACK) &&
+ !ggml_rope_get_truncate(op)) {
+ return false;
+ }
#if USE_ALWAYS_TRUE_SUPPORTS_OP == 1
/* ggml-rpc cheats it like this */
/* with the current implementation of serialize_tensor, the src/view aren't properly passed */
UNUSED(dev);
- UNUSED(op);
return true;
#else
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 3a010dbd5..48281acf9 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -10915,7 +10915,7 @@ static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *
const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
float corr_dims[2];
- ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+ ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, true, corr_dims);
const float theta_scale = powf(freq_base, -2.0f/n_dims);
@@ -11413,7 +11413,7 @@ void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_
}
float corr_dims[2];
- ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+ ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, true, corr_dims);
uint32_t set_rows_stride = 0;
// Fused rope + view + set_rows passes the set_rows destination stride in set_rows_stride
@@ -14190,7 +14190,7 @@ bool ggml_vk_can_fuse_rope_set_rows(ggml_backend_vk_context * ctx, const struct
return false;
}
- return true;
+ return ggml_rope_get_truncate(rope);
}
bool ggml_vk_can_fuse_rms_norm_set_rows(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph,
@@ -14328,7 +14328,7 @@ bool ggml_vk_can_fuse_rms_norm_mul_rope(ggml_backend_vk_context * ctx, const str
return false;
}
- return true;
+ return ggml_rope_get_truncate(rope);
}
uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx) {
@@ -15877,8 +15877,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_REPEAT_BACK:
return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
case GGML_OP_ROPE:
+ if (!ggml_rope_get_truncate(op)) {
+ return false;
+ }
return ggml_is_contiguous_rows(op) && ggml_is_contiguous_rows(op->src[0]);
case GGML_OP_ROPE_BACK:
+ return ggml_rope_get_truncate(op);
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
index 6f68d4623..2e65677c1 100644
--- a/ggml/src/ggml-webgpu/ggml-webgpu.cpp
+++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
@@ -2826,7 +2826,7 @@ static webgpu_encoded_op ggml_webgpu_rope(webgpu_context & ctx,
float theta_scale = powf(freq_base, -2.0f / n_dims);
float corr_dims[2];
- ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+ ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, true, corr_dims);
std::vector<uint32_t> params = {
(uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
@@ -4696,7 +4696,7 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
supports_op = (op->type == GGML_TYPE_F32 && src0->type == GGML_TYPE_F32) && ggml_is_contiguous_rows(src0);
break;
case GGML_OP_ROPE:
- supports_op = op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16;
+ supports_op = ggml_rope_get_truncate(op) && (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
break;
case GGML_OP_GLU:
switch (ggml_get_glu_op(op)) {
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index e81148106..f8cf75b3a 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -4359,7 +4359,7 @@ static struct ggml_tensor * ggml_rope_impl(
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
- int32_t params[16] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
+ int32_t params[16] = { /*n_past*/ 0, n_dims, mode, /*no_trunc*/ 0, n_ctx_orig };
memcpy(params + 5, &freq_base, sizeof(float));
memcpy(params + 6, &freq_scale, sizeof(float));
memcpy(params + 7, &ext_factor, sizeof(float));
@@ -4532,11 +4532,15 @@ static float ggml_rope_yarn_corr_dim(int n_dims, int n_ctx_orig, float n_rot, fl
}
void ggml_rope_yarn_corr_dims(
- int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
+ int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, bool truncate, float dims[2]
) {
// start and end correction dims
- float start = floorf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_fast, freq_base));
- float end = ceilf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_slow, freq_base));
+ float start = ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_fast, freq_base);
+ float end = ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_slow, freq_base);
+ if (truncate) {
+ start = floorf(start);
+ end = ceilf(end);
+ }
dims[0] = MAX(0, start);
dims[1] = MIN(n_dims - 1, end);
}
@@ -4597,6 +4601,19 @@ struct ggml_tensor * ggml_rope_set_offset(
return a;
}
+struct ggml_tensor * ggml_rope_set_truncate(
+ struct ggml_tensor * a,
+ bool truncate) {
+ GGML_ASSERT(a->op == GGML_OP_ROPE || a->op == GGML_OP_ROPE_BACK);
+ ggml_set_op_params_i32(a, 3, !truncate);
+ return a;
+}
+
+bool ggml_rope_get_truncate(const struct ggml_tensor * a) {
+ GGML_ASSERT(a->op == GGML_OP_ROPE || a->op == GGML_OP_ROPE_BACK);
+ return ggml_get_op_params_i32(a, 3) == 0;
+}
+
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;
}
@@ -7237,7 +7254,6 @@ static void ggml_compute_backward(
//const int n_past = ((int32_t *) tensor->op_params)[0];
const int n_dims = ((const int32_t *) tensor->op_params)[1];
const int mode = ((const int32_t *) tensor->op_params)[2];
- //const int n_ctx = ((int32_t *) tensor->op_params)[3];
const int n_ctx_orig = ((const int32_t *) tensor->op_params)[4];
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
int sections[4] = {0, 0, 0, 0};
@@ -7255,6 +7271,7 @@ static void ggml_compute_backward(
mode, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow) :
ggml_rope_multi_back(ctx, grad, src1, src2, n_dims, sections,
mode, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+ ggml_rope_set_truncate(rope_back, ggml_rope_get_truncate(tensor));
ggml_add_or_set(ctx, cgraph, isrc0, rope_back);
}
GGML_ASSERT((!src2 || !src2_needs_grads) && "gradients for freq factors not implemented");
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 89050ed92..c63b5a571 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -2799,10 +2799,12 @@ struct test_rope_set_rows : public test_case {
const std::array<int64_t, 4> ne_a;
int mode;
const int n_dims;
+ const bool yarn;
+ const bool truncate;
const int n_ctx{512};
std::string vars() override {
- return VARS_TO_STR5(type, type_idx, ne_a, mode, n_dims);
+ return VARS_TO_STR7(type, type_idx, ne_a, mode, n_dims, yarn, truncate);
}
std::string op_desc(ggml_tensor * t) override {
@@ -2816,8 +2818,8 @@ struct test_rope_set_rows : public test_case {
ggml_type type_idx,
std::array<int64_t, 4> ne_a,
int mode,
- int n_dims = 128)
- : type(type), type_idx(type_idx), ne_a(ne_a), mode(mode), n_dims(n_dims) {}
+ int n_dims = 128, bool yarn = false, bool truncate = true)
+ : type(type), type_idx(type_idx), ne_a(ne_a), mode(mode), n_dims(n_dims), yarn(yarn), truncate(truncate) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, ne_a[0], ne_a[1], ne_a[2], 1);
@@ -2834,9 +2836,12 @@ struct test_rope_set_rows : public test_case {
}
ggml_set_name(pos, "pos");
- float fs = 1.4245f;
- float ef = 0.7465f;
- float af = 1.4245f;
+ const float fs = yarn ? 1.0f/256 : 1.4245f;
+ const float ef = yarn ? 1.0f : 0.7465f;
+ const float af = yarn ? 1.0f : 1.4245f;
+ const int n_ctx_orig = yarn ? 4096 : 0;
+ const float freq_base = yarn ? 1000000.0f : 10000.0f;
+ const float beta_fast = yarn ? 32.0f : 1.0f;
ggml_tensor * freq = nullptr;
ggml_tensor * rope = nullptr;
@@ -2844,15 +2849,18 @@ struct test_rope_set_rows : public test_case {
if (is_vision) {
GGML_ASSERT(n_dims/4 > 0);
int rope_sections[4] = {n_dims/4, n_dims/4, 0, 0}; // Vision-RoPE only use first two dimension for image (x, y) coordinate
- rope = ggml_rope_multi(ctx, a, pos, freq, n_dims/2, rope_sections, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
+ rope = ggml_rope_multi(ctx, a, pos, freq, n_dims/2, rope_sections, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
} else {
GGML_ASSERT(n_dims/3 > 0);
int rope_sections[4] = {n_dims/3, n_dims/3, n_dims/3, 0};
- rope = ggml_rope_multi(ctx, a, pos, freq, n_dims, rope_sections, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
+ rope = ggml_rope_multi(ctx, a, pos, freq, n_dims, rope_sections, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
}
+ } else if (yarn) {
+ rope = ggml_rope_ext(ctx, a, pos, freq, n_dims, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
} else {
rope = ggml_rope(ctx, a, pos, n_dims, mode);
}
+ ggml_rope_set_truncate(rope, truncate);
ggml_tensor * view = ggml_view_2d(ctx, rope, ne_a[0] * ne_a[1], ne_a[2], rope->nb[2], 0);
@@ -2880,7 +2888,7 @@ struct test_rope_set_rows : public test_case {
const int num_pos_ids = (mode & GGML_ROPE_TYPE_MROPE) ? ne_a[2] * 4 : ne_a[2];
std::vector<int> data(num_pos_ids);
for (int i = 0; i < num_pos_ids; i++) {
- data[i] = rand() % n_ctx;
+ data[i] = yarn ? 4096 + i * 67 : rand() % n_ctx;
}
ggml_backend_tensor_set(t, data.data(), 0, num_pos_ids * sizeof(int));
} else {
@@ -2906,6 +2914,8 @@ struct test_rms_norm_mul_rope : public test_case {
const bool broadcast; // multiply by a 1D [ne0] weight, as model norm weights are
const ggml_type set_rows_type;
int mode;
+ const bool yarn;
+ const bool truncate;
std::string op_desc(ggml_tensor * t) override {
GGML_UNUSED(t);
@@ -2915,14 +2925,15 @@ struct test_rms_norm_mul_rope : public test_case {
bool run_whole_graph() override { return true; }
std::string vars() override {
- return VARS_TO_STR9(ne, eps, multi_add, mul, rope, set_rows, broadcast, mode, set_rows_type);
+ return VARS_TO_STR11(ne, eps, multi_add, mul, rope, set_rows, broadcast, mode, set_rows_type, yarn, truncate);
}
test_rms_norm_mul_rope(std::array<int64_t, 4> ne, float eps = 1e-6f, bool multi_add = false,
bool set_rows = false, bool broadcast = false, int mode = GGML_ROPE_TYPE_NORMAL,
- bool mul = true, bool rope = true, ggml_type set_rows_type = GGML_TYPE_F16)
+ bool mul = true, bool rope = true, ggml_type set_rows_type = GGML_TYPE_F16,
+ bool yarn = false, bool truncate = true)
: ne(ne), eps(eps), multi_add(multi_add), mul(mul), rope(rope), set_rows(set_rows), broadcast(broadcast),
- set_rows_type(set_rows_type), mode(mode) {}
+ set_rows_type(set_rows_type), mode(mode), yarn(yarn), truncate(truncate) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, ne[0], ne[1], ne[2], ne[3]);
@@ -2954,14 +2965,21 @@ struct test_rms_norm_mul_rope : public test_case {
if (rope) {
const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne[2] * (is_mrope ? 4 : 1));
+ const int n_ctx_orig = yarn ? 4096 : 0;
+ const float freq_base = yarn ? 1000000.0f : 10000.0f;
+ const float freq_scale = yarn ? 1.0f/256 : 1.0f;
+ const float ext_factor = yarn ? 1.0f : 0.0f;
if (is_mrope) {
const int n_dims = ne[0];
int sections[4] = { n_dims/3, n_dims/3, n_dims/3, 0 };
- a = ggml_rope_multi(ctx, a, pos, nullptr, n_dims, sections, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 32.0f, 1.0f);
+ a = ggml_rope_multi(ctx, a, pos, nullptr, n_dims, sections, mode, n_ctx_orig, freq_base, freq_scale, ext_factor, 1.0f, 32.0f, 1.0f);
+ } else if (yarn) {
+ a = ggml_rope_ext(ctx, a, pos, nullptr, ne[0], mode, n_ctx_orig, freq_base, freq_scale, ext_factor, 1.0f, 32.0f, 1.0f);
} else {
a = ggml_rope(ctx, a, pos, ne[0], mode);
}
+ ggml_rope_set_truncate(a, truncate);
}
if (set_rows) {
@@ -2986,8 +3004,8 @@ struct test_rms_norm_mul_rope : public test_case {
init_set_rows_row_ids(t, ne[2] * 2);
} else if (t->type == GGML_TYPE_I32) {
std::vector<int32_t> data(ggml_nelements(t));
- for (int32_t & value : data) {
- value = rand() % 512;
+ for (size_t i = 0; i < data.size(); i++) {
+ data[i] = yarn ? 127 + i * 67 : rand() % 512;
}
ggml_backend_tensor_set(t, data.data(), 0, ggml_nbytes(t));
} else {
@@ -6452,18 +6470,23 @@ struct test_rope : public test_case {
bool forward;
bool inplace;
int n_offs; // offset of the rotated dims window, set via ggml_rope_set_offset()
+ int n_ctx_orig;
+ float freq_base;
+ float beta_fast;
+ bool truncate;
std::string vars() override {
// forward can be inferred from the op, does not need to be printed
- return VARS_TO_STR12(type, ne_a, n_dims, mode, n_ctx, fs, ef, af, ff, v, inplace, n_offs);
+ return VARS_TO_STR16(type, ne_a, n_dims, mode, n_ctx, fs, ef, af, ff, v, inplace, n_offs, n_ctx_orig, freq_base, beta_fast, truncate);
}
test_rope(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne_a = {10, 5, 3, 1},
int n_dims = 10, int mode = GGML_ROPE_TYPE_NORMAL, int n_ctx = 512, float fs = 1.0f,
float ef = 0.0f, float af = 0.0f, bool ff = false, int v = 0, bool forward = true, bool inplace = false,
- int n_offs = 0)
- : type(type), ne_a(ne_a), n_dims(n_dims), mode(mode), n_ctx(n_ctx), fs(fs), ef(ef), af(af), ff(ff), v(v), forward(forward), inplace(inplace), n_offs(n_offs) {}
+ int n_offs = 0, int n_ctx_orig = 0, float freq_base = 10000.0f, float beta_fast = 1.0f, bool truncate = true)
+ : type(type), ne_a(ne_a), n_dims(n_dims), mode(mode), n_ctx(n_ctx), fs(fs), ef(ef), af(af), ff(ff), v(v), forward(forward), inplace(inplace), n_offs(n_offs),
+ n_ctx_orig(n_ctx_orig), freq_base(freq_base), beta_fast(beta_fast), truncate(truncate) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a;
@@ -6527,37 +6550,38 @@ struct test_rope : public test_case {
int rope_sections[4] = {n_dims/4, n_dims/4, 0, 0}; // Vision-RoPE only use first two dimension for image (x, y) coordinate
if (forward) {
if (inplace) {
- out = ggml_rope_multi_inplace(ctx, a, pos, freq, n_dims/2, rope_sections, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
+ out = ggml_rope_multi_inplace(ctx, a, pos, freq, n_dims/2, rope_sections, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
} else {
- out = ggml_rope_multi(ctx, a, pos, freq, n_dims/2, rope_sections, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
+ out = ggml_rope_multi(ctx, a, pos, freq, n_dims/2, rope_sections, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
}
} else {
- out = ggml_rope_multi_back(ctx, a, pos, freq, n_dims/2, rope_sections, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
+ out = ggml_rope_multi_back(ctx, a, pos, freq, n_dims/2, rope_sections, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
}
} else {
GGML_ASSERT(n_dims/3 > 0);
int rope_sections[4] = {n_dims/3, n_dims/3, n_dims/3, 0};
if (forward) {
if (inplace) {
- out = ggml_rope_multi_inplace(ctx, a, pos, freq, n_dims, rope_sections, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
+ out = ggml_rope_multi_inplace(ctx, a, pos, freq, n_dims, rope_sections, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
} else {
- out = ggml_rope_multi(ctx, a, pos, freq, n_dims, rope_sections, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
+ out = ggml_rope_multi(ctx, a, pos, freq, n_dims, rope_sections, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
}
} else {
- out = ggml_rope_multi_back(ctx, a, pos, freq, n_dims, rope_sections, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
+ out = ggml_rope_multi_back(ctx, a, pos, freq, n_dims, rope_sections, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
}
}
} else {
if (forward) {
if (inplace) {
- out = ggml_rope_ext_inplace(ctx, a, pos, freq, n_dims, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
+ out = ggml_rope_ext_inplace(ctx, a, pos, freq, n_dims, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
} else {
- out = ggml_rope_ext(ctx, a, pos, freq, n_dims, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
+ out = ggml_rope_ext(ctx, a, pos, freq, n_dims, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
}
} else {
- out = ggml_rope_ext_back(ctx, a, pos, freq, n_dims, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
+ out = ggml_rope_ext_back(ctx, a, pos, freq, n_dims, mode, n_ctx_orig, freq_base, fs, ef, af, beta_fast, 1.0f);
}
}
+ ggml_rope_set_truncate(out, truncate);
if (n_offs != 0) {
out = ggml_rope_set_offset(out, n_offs);
}
@@ -10512,6 +10536,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
}
}
+ for (int mode : {GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX}) {
+ for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
+ for (bool truncate : {false, true}) {
+ test_cases.emplace_back(new test_rope_set_rows(type, GGML_TYPE_I64, {128, 2, 3, 2}, mode, 128, true, truncate));
+ }
+ }
+ }
+
for (ggml_type type_input : {GGML_TYPE_F32}) {
for (ggml_op_pool pool_type : {GGML_OP_POOL_AVG, GGML_OP_POOL_MAX}) {
for (int k0 : {1, 3}) {
@@ -11302,6 +11334,18 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
// ne[2] > 65535 exceeds the CUDA gridDim.y limit (#27901)
test_cases.emplace_back(new test_rms_norm_mul_rope({4, 1, 65536, 1}, 1e-6f, false, false, false, GGML_ROPE_TYPE_NORMAL));
test_cases.emplace_back(new test_rms_norm_mul_rope({4, 1, 65536, 1}, 1e-6f, false, true, false, GGML_ROPE_TYPE_NEOX));
+
+ for (int mode : {GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX}) {
+ for (int64_t n : {128, 1024}) {
+ for (bool truncate : {false, true}) {
+ test_cases.emplace_back(new test_rms_norm_mul_rope({n, 2, 3, 1}, 1e-6f, false, false, true, mode, true, true, GGML_TYPE_F16, true, truncate));
+ for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
+ test_cases.emplace_back(new test_rms_norm_mul_rope({n, 2, 3, 1}, 1e-6f, false, true, true, mode, true, true, type, true, truncate));
+ }
+ }
+ }
+ }
+
for (int64_t d_conv : {3, 4, 7, 9}) {
for (int64_t d_inner: {1024, 1536, 2048}) {
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {d_conv, d_inner, 1, 1}, {d_conv, d_inner, 1, 1}));
@@ -12148,6 +12192,17 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_rope(type, {128, 32, 2, 1}, 32, GGML_ROPE_TYPE_NEOX, 512, 1.4245f, 0.7465f, 1.4245f, false, 0, true, true, 32));
}
+ for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
+ for (int mode : {GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX, GGML_ROPE_TYPE_MROPE, GGML_ROPE_TYPE_IMROPE, GGML_ROPE_TYPE_VISION}) {
+ for (bool truncate : {false, true}) {
+ for (bool fw : {false, true}) {
+ test_cases.emplace_back(new test_rope(type, {128, 2, 3, 1}, 128, mode, 8192, 1.0f/256, 1.0f, 1.0f, false, 0, fw, false, 0, 4096, 1000000.0f, 32.0f, truncate));
+ }
+ test_cases.emplace_back(new test_rope(type, {128, 2, 3, 1}, 128, mode, 8192, 1.0f/256, 1.0f, 1.0f, true, 1, true, true, 0, 4096, 1000000.0f, 32.0f, truncate));
+ }
+ }
+ }
+
// Real-model RoPE: F32 forward, packed Q, 512-token prefill.
test_cases.emplace_back(new test_rope(GGML_TYPE_F32, {256, 8, 512, 1}, 64, GGML_ROPE_TYPE_IMROPE, 512, 1.0f, 0.0f, 1.0f, false, 0, true)); // qwen3.5 0.8B
test_cases.emplace_back(new test_rope(GGML_TYPE_F32, {256, 16, 512, 1}, 64, GGML_ROPE_TYPE_IMROPE, 512, 1.0f, 0.0f, 1.0f, false, 0, true)); // qwen3.5 4B
diff --git a/tests/test-rope.cpp b/tests/test-rope.cpp
index 801e4cd82..9b0ef69a4 100644
--- a/tests/test-rope.cpp
+++ b/tests/test-rope.cpp
@@ -124,6 +124,83 @@ static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph *
ggml_graph_compute(graph, &plan);
}
+static void test_yarn(ggml_type type, int mode, bool truncate, bool forward, bool inplace) {
+ const int n_dims = 128;
+ const int n_offs = 8;
+ const int ne0 = n_dims + 2*n_offs;
+ const int32_t positions[] = {0, 1, 4096, 8192, -67};
+ const int n_pos = sizeof(positions)/sizeof(positions[0]);
+ const double base = 1000000.0;
+ const double factor = 256.0;
+ const double pi = std::acos(-1.0);
+ double low = n_dims/2 * std::log(4096.0/(32.0*2*pi))/std::log(base);
+ double high = n_dims/2 * std::log(4096.0/( 1.0*2*pi))/std::log(base);
+ if (truncate) {
+ low = std::floor(low);
+ high = std::ceil(high);
+ }
+ const double mscale = 1.0 + 0.1*std::log(factor);
+
+ ggml_init_params params = {1024*1024, nullptr, false};
+ ggml_context * ctx = ggml_init(params);
+ ggml_tensor * x = ggml_new_tensor_3d(ctx, type, ne0, 2, n_pos);
+ ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_pos);
+ std::vector<float> input(ggml_nelements(x));
+ for (size_t i = 0; i < input.size(); ++i) {
+ input[i] = (int(i % 31) - 15)/16.0f;
+ if (type == GGML_TYPE_F32) {
+ ((float *) x->data)[i] = input[i];
+ } else {
+ ((ggml_fp16_t *) x->data)[i] = ggml_fp32_to_fp16(input[i]);
+ }
+ }
+ for (int i = 0; i < n_pos; ++i) {
+ ((int32_t *) pos->data)[i] = positions[i];
+ }
+ ggml_tensor * out;
+ if (!forward) {
+ out = ggml_rope_ext_back(ctx, x, pos, nullptr, n_dims, mode, 4096, base, 1.0/factor, 1.0f, 1.0f, 32.0f, 1.0f);
+ } else if (inplace) {
+ out = ggml_rope_ext_inplace(ctx, x, pos, nullptr, n_dims, mode, 4096, base, 1.0/factor, 1.0f, 1.0f, 32.0f, 1.0f);
+ } else {
+ out = ggml_rope_ext(ctx, x, pos, nullptr, n_dims, mode, 4096, base, 1.0/factor, 1.0f, 1.0f, 32.0f, 1.0f);
+ }
+ ggml_rope_set_truncate(out, truncate);
+ ggml_rope_set_offset(out, n_offs);
+ ggml_cgraph * graph = ggml_new_graph(ctx);
+ ggml_build_forward_expand(graph, out);
+ std::vector<uint8_t> work;
+ ggml_graph_compute_helper(work, graph, 2);
+
+ std::vector<float> expected = input;
+ for (int p = 0; p < n_pos; ++p) {
+ for (int h = 0; h < 2; ++h) {
+ const int row = (p*2 + h)*ne0 + n_offs;
+ for (int i = 0; i < n_dims/2; ++i) {
+ const double ramp = 1.0 - MIN(1.0, MAX(0.0, (i - low)/(high - low)));
+ const double freq = std::pow(base, -2.0*i/n_dims);
+ const double theta = positions[p]*freq*((1.0 - ramp)/factor + ramp);
+ const double c = std::cos(theta)*mscale;
+ const double s = std::sin(theta)*mscale*(forward ? 1.0 : -1.0);
+ const int i0 = row + (mode == GGML_ROPE_TYPE_NORMAL ? 2*i : i);
+ const int i1 = i0 + (mode == GGML_ROPE_TYPE_NORMAL ? 1 : n_dims/2);
+ expected[i0] = input[i0]*c - input[i1]*s;
+ expected[i1] = input[i0]*s + input[i1]*c;
+ }
+ }
+ }
+ float max_error = 0.0f;
+ for (size_t i = 0; i < expected.size(); ++i) {
+ const float actual = type == GGML_TYPE_F32 ? ((float *) out->data)[i] : ggml_fp16_to_fp32(((ggml_fp16_t *) out->data)[i]);
+ GGML_ASSERT(std::isfinite(actual));
+ max_error = MAX(max_error, std::fabs(actual - expected[i]));
+ }
+ printf("yarn type=%s mode=%d truncate=%d forward=%d inplace=%d max_error=%g\n",
+ ggml_type_name(type), mode, truncate, forward, inplace, max_error);
+ GGML_ASSERT(max_error < (type == GGML_TYPE_F32 ? 0.003f : 0.005f));
+ ggml_free(ctx);
+}
+
int main(int /*argc*/, const char ** /*argv*/) {
struct ggml_init_params params = {
/* .mem_size = */ 128*1024*1024,
@@ -259,5 +336,15 @@ int main(int /*argc*/, const char ** /*argv*/) {
ggml_free(ctx0);
+ for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
+ for (int mode : {GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX}) {
+ for (bool truncate : {false, true}) {
+ test_yarn(type, mode, truncate, true, false);
+ test_yarn(type, mode, truncate, true, true);
+ test_yarn(type, mode, truncate, false, false);
+ }
+ }
+ }
+
return 0;
}