vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

implement new YaRN algorithm

Browse source
This commit is contained in:
Cebtenzzre 2023-09-04 20:08:17 -04:00
parent 9348aa4df9
commit a30ae2095c
10 changed files with 303 additions and 269 deletions
Download patch file Download diff file Expand all files Collapse all files

32
examples/common.cpp
View file

@ -194,18 +194,30 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
break;
}
params.rope_freq_scale = std::stof(argv[i]);
} else if (arg == "--rope-ntk-factor") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.rope_ntk_factor = std::stof(argv[i]);
} else if (arg == "--rope-ext-factor") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.rope_ext_factor = std::stof(argv[i]);
} else if (arg == "--rope-attn-factor") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.rope_attn_factor = std::stof(argv[i]);
} else if (arg == "--rope-beta-fast") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.rope_beta_fast = std::stof(argv[i]);
} else if (arg == "--rope-beta-slow") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.rope_beta_slow = std::stof(argv[i]);
} else if (arg == "--memory-f32") {
params.memory_f16 = false;
} else if (arg == "--top-p") {
@ -578,8 +590,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stdout, " --cfg-scale N strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
fprintf(stdout, " --rope-freq-base N RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
fprintf(stdout, " --rope-freq-scale N RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
fprintf(stdout, " --rope-ntk-factor N RoPE NTK mix factor (default: %.1f)\n", params.rope_ntk_factor);
fprintf(stdout, " --rope-ext-factor N RoPE extrapolation mix factor (default: %.1f)\n", params.rope_ext_factor);
fprintf(stdout, " --rope-attn-factor N RoPE magnitude scaling factor (default: %.1f)\n", params.rope_attn_factor);
fprintf(stdout, " --rope-beta-fast N RoPE low correction dim (default: %.1f)\n", params.rope_beta_fast);
fprintf(stdout, " --rope-beta-slow N RoPE high correction dim (default: %.1f)\n", params.rope_beta_slow);
fprintf(stdout, " --ignore-eos ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
fprintf(stdout, " --no-penalize-nl do not penalize newline token\n");
fprintf(stdout, " --memory-f32 use f32 instead of f16 for memory key+value (default: disabled)\n");
@ -671,8 +685,10 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
lparams.embedding = params.embedding;
lparams.rope_freq_base = params.rope_freq_base;
lparams.rope_freq_scale = params.rope_freq_scale;
lparams.rope_ntk_factor = params.rope_ntk_factor;
lparams.rope_ext_factor = params.rope_ext_factor;
lparams.rope_attn_factor = params.rope_attn_factor;
lparams.rope_beta_fast = params.rope_beta_fast;
lparams.rope_beta_slow = params.rope_beta_slow;
return lparams;
}

4
examples/common.h
View file

@ -32,8 +32,10 @@ struct gpt_params {
float rms_norm_eps = LLAMA_DEFAULT_RMS_EPS; // rms norm epsilon
float rope_freq_base = 10000.0f; // RoPE base frequency
float rope_freq_scale = 1.0f; // RoPE frequency scaling factor
float rope_ntk_factor = 0.0f; // RoPE NTK mix factor
float rope_ext_factor = 0.0f; // RoPE extrapolation mix factor
float rope_attn_factor = 1.0f; // RoPE magnitude scaling factor
float rope_beta_fast = 32.0f; // RoPE low correction dim
float rope_beta_slow = 1.0f; // RoPE high correction dim
// sampling parameters
std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens

36
examples/server/server.cpp
View file

@ -612,8 +612,10 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
fprintf(stdout, " -eps N, --rms-norm-eps N rms norm eps (TEMP!!! use 1e-5 for LLaMAv2) (default: %.1e)\n", params.rms_norm_eps);
fprintf(stdout, " --rope-freq-base N RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
fprintf(stdout, " --rope-freq-scale N RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
fprintf(stdout, " --rope-ntk-factor N RoPE NTK mix factor (default: %.1f)\n", params.rope_ntk_factor);
fprintf(stdout, " --rope-ext-factor N RoPE extrapolation mix factor (default: %.1f)\n", params.rope_ext_factor);
fprintf(stdout, " --rope-attn-factor N RoPE magnitude scaling factor (default: %.1f)\n", params.rope_attn_factor);
fprintf(stdout, " --rope-beta-fast N RoPE low correction dim (default: %.1f)\n", params.rope_beta_fast);
fprintf(stdout, " --rope-beta-slow N RoPE high correction dim (default: %.1f)\n", params.rope_beta_slow);
fprintf(stdout, " -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch);
fprintf(stdout, " --memory-f32 use f32 instead of f16 for memory key+value (default: disabled)\n");
fprintf(stdout, " not recommended: doubles context memory required and no measurable increase in quality\n");
@ -766,14 +768,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
}
params.rope_freq_scale = std::stof(argv[i]);
}
else if (arg == "--rope-ntk-factor")
{
if (++i >= argc) {
invalid_param = true;
break;
}
params.rope_ntk_factor = std::stof(argv[i]);
}
else if (arg == "--rope-ext-factor")
{
if (++i >= argc) {
@ -782,6 +776,30 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
}
params.rope_ext_factor = std::stof(argv[i]);
}
else if (arg == "--rope-attn-factor")
{
if (++i >= argc) {
invalid_param = true;
break;
}
params.rope_attn_factor = std::stof(argv[i]);
}
else if (arg == "--rope-beta-fast")
{
if (++i >= argc) {
invalid_param = true;
break;
}
params.rope_beta_fast = std::stof(argv[i]);
}
else if (arg == "--rope-beta-slow")
{
if (++i >= argc) {
invalid_param = true;
break;
}
params.rope_beta_slow = std::stof(argv[i]);
}
else if (arg == "--memory-f32" || arg == "--memory_f32")
{
params.memory_f16 = false;

75
ggml-cuda.cu
View file

@ -3558,34 +3558,31 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
cpy_1(cx + x_offset, cdst + dst_offset);
}
static __device__ float rope_ntkv2_ramp(const float low, const float high, const int i0) {
static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) {
const float y = (i0 / 2 - low) / min(0.001f, high - low);
return 1.0f - min(1.0f, max(0.0f, y));
}
struct rope_corr_factors {
struct rope_corr_dims {
float v[4];
};
// NTKv2 algorithm based on LlamaPartNTKScaledRotaryEmbedding.py from https://github.com/jquesnelle/scaled-rope
// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
static __device__ float rope_ntkv2(
const float theta_base,
const float theta_linear,
const float theta_ntk,
const rope_corr_factors corr_factors,
const int64_t i0,
const float ntk_factor,
const float ext_factor) {
float ramp_mix;
float theta;
static __device__ void rope_yarn(
float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale,
float * cos_theta, float * sin_theta
) {
// Get n-d rotational scaling corrected for extrapolation
float theta_interp = freq_scale * theta_extrap;
float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor;
float theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
ramp_mix = rope_ntkv2_ramp(corr_factors.v[0], corr_factors.v[1], i0) * ntk_factor;
theta = theta_linear * (1 - ramp_mix) + theta_ntk * ramp_mix;
ramp_mix = rope_ntkv2_ramp(corr_factors.v[2], corr_factors.v[3], i0) * ext_factor;
theta = theta * (1 - ramp_mix) + theta_base * ramp_mix;
return theta;
// Get n-d magnitude scaling corrected for interpolation
if (freq_scale > 1.0f)
mscale *= 1.0f + 0.1f * logf(freq_scale);
*cos_theta = cosf(theta) * mscale;
*sin_theta = sinf(theta) * mscale;
}
// rope == RoPE == rotary positional embedding
@ -3594,13 +3591,11 @@ static __global__ void rope_f32(
float * dst,
const int ncols,
const float freq_scale,
const float ntk_factor,
const float ext_factor,
const float theta_scale,
const float theta_ntk_scale,
const float p0,
const int p_delta_rows,
const rope_corr_factors corr_factors) {
const rope_corr_dims corr_dims) {
const int col = 2*(blockDim.x*blockIdx.x + threadIdx.x);
if (col >= ncols) {
@ -3612,11 +3607,9 @@ static __global__ void rope_f32(
const float p = p0 + row / p_delta_rows;
const float theta_base = p*powf(theta_scale, col/2);
const float theta_linear = freq_scale * theta_base;
const float theta_ntk = p*powf(theta_ntk_scale, col/2);
const float theta = rope_ntkv2(theta_base, theta_linear, theta_ntk, corr_factors, col, ntk_factor, ext_factor);
const float sin_theta = sinf(theta);
const float cos_theta = cosf(theta);
float cos_theta, sin_theta;
rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta);
const float x0 = x[i + 0];
const float x1 = x[i + 1];
@ -4284,20 +4277,19 @@ static void rope_f32_cuda(
const int ncols,
const int nrows,
const float freq_scale,
const float ntk_factor,
const float ext_factor,
const float theta_scale,
const float theta_ntk_scale,
const float p0,
const int p_delta_rows,
const rope_corr_factors corr_factors,
const rope_corr_dims corr_dims,
cudaStream_t stream) {
GGML_ASSERT(nrows % 2 == 0);
const dim3 block_dims(2*CUDA_ROPE_BLOCK_SIZE, 1, 1);
const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
const dim3 block_nums(num_blocks_x, nrows, 1);
rope_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, freq_scale, ntk_factor, ext_factor, theta_scale,
theta_ntk_scale, p0, p_delta_rows, corr_factors);
rope_f32<<<block_nums, block_dims, 0, stream>>>(
x, dst, ncols, freq_scale, ext_factor, theta_scale, p0, p_delta_rows, corr_dims
);
}
static void rope_glm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float p, const float block_p, const float theta_scale, cudaStream_t stream) {
@ -5000,11 +4992,13 @@ inline void ggml_cuda_op_rope(
const int n_ctx = ((int32_t *) dst->op_params)[3];
// RoPE alteration for extended context
float freq_base, freq_scale, ntk_factor, ext_factor;
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&ntk_factor, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&attn_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&beta_fast, (int32_t *) dst->op_params + 8, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 9, sizeof(float));
const float theta_scale = powf(freq_base, -2.0f/n_dims);
@ -5018,12 +5012,13 @@ inline void ggml_cuda_op_rope(
rope_glm_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, id_p, block_p, theta_scale, cudaStream_main);
} else {
const float p0 = (mode & 1) == 0 ? n_past : 0;
const float theta_ntk_scale = powf(freq_base * powf(freq_scale, (n_dims / (n_dims - 2.0f))), -2.0f/n_dims);
rope_corr_factors corr_factors;
ggml_rope_ntkv2_corr_factors(n_dims, freq_base, corr_factors.v);
rope_corr_dims corr_dims;
ggml_rope_yarn_corr_dims(n_dims, freq_base, beta_fast, beta_slow, corr_dims.v);
rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, freq_scale, ntk_factor, ext_factor, theta_scale,
theta_ntk_scale, p0, ne01, corr_factors, cudaStream_main);
rope_f32_cuda(
src0_ddf_i, dst_ddf_i, ne00, i01_diff, freq_scale, ext_factor, theta_scale, p0, ne01, corr_dims,
cudaStream_main
);
}
(void) src1;

14
ggml-metal.m
View file

@ -1035,11 +1035,13 @@ void ggml_metal_graph_compute(
const int n_dims = ((int32_t *) dst->op_params)[1];
const int mode = ((int32_t *) dst->op_params)[2];
float freq_base, freq_scale, ntk_factor, ext_factor;
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&ntk_factor, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&attn_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&beta_fast, (int32_t *) dst->op_params + 8, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 9, sizeof(float));
[encoder setComputePipelineState:ctx->pipeline_rope];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@ -1065,8 +1067,10 @@ void ggml_metal_graph_compute(
[encoder setBytes:&mode length:sizeof( int) atIndex:20];
[encoder setBytes:&freq_base length:sizeof(float) atIndex:21];
[encoder setBytes:&freq_scale length:sizeof(float) atIndex:22];
[encoder setBytes:&ntk_factor length:sizeof(float) atIndex:23];
[encoder setBytes:&ext_factor length:sizeof(float) atIndex:24];
[encoder setBytes:&ext_factor length:sizeof(float) atIndex:23];
[encoder setBytes:&attn_factor length:sizeof(float) atIndex:24];
[encoder setBytes:&beta_fast length:sizeof(float) atIndex:25];
[encoder setBytes:&beta_slow length:sizeof(float) atIndex:26];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;

70
ggml-metal.metal
View file

@ -597,53 +597,41 @@ kernel void kernel_alibi_f32(
}
}
static float rope_ntkv2_ramp(const float low, const float high, const int i0) {
static float rope_yarn_ramp(const float low, const float high, const int i0) {
const float y = (i0 / 2 - low) / min(0.001f, high - low);
return 1.0f - min(1.0f, max(0.0f, y));
}
// NTKv2 algorithm based on LlamaPartNTKScaledRotaryEmbedding.py from https://github.com/jquesnelle/scaled-rope
// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
static float rope_ntkv2(
const float theta_base,
const float theta_linear,
const float theta_ntk,
const float corr_factors[4],
const int64_t i0,
const float ntk_factor,
const float ext_factor) {
float ramp_mix;
float theta;
static void rope_yarn(
float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
float * cos_theta, float * sin_theta
) {
// Get n-d rotational scaling corrected for extrapolation
float theta_interp = freq_scale * theta_extrap;
float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
float theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
ramp_mix = rope_ntkv2_ramp(corr_factors[0], corr_factors[1], i0) * ntk_factor;
theta = theta_linear * (1 - ramp_mix) + theta_ntk * ramp_mix;
ramp_mix = rope_ntkv2_ramp(corr_factors[2], corr_factors[3], i0) * ext_factor;
theta = theta * (1 - ramp_mix) + theta_base * ramp_mix;
return theta;
// Get n-d magnitude scaling corrected for interpolation
if (freq_scale > 1.0f)
mscale *= 1.0f + 0.1f * logf(freq_scale);
*cos_theta = cosf(theta) * mscale;
*sin_theta = sinf(theta) * mscale;
}
// Interpolation constants found experimentally for LLaMA (might not be totally optimal though)
// Do not change unless there is a good reason for doing so!
constant float BETA_0 = 1.75f;
constant float BETA_1 = 1.25f;
constant float GAMMA_0 = 16.0f;
constant float GAMMA_1 = 2.0f;
constant float max_pos_emb = 2048;
// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
static float rope_ntkv2_corr_factor(const int n_dims, const float n_rot, const float base) {
static float rope_yarn_corr_factor(const int n_dims, const float n_rot, const float base) {
return n_dims * log(max_pos_emb / (n_rot * 2 * M_PI_F)) / (2 * log(base));
}
static void rope_ntkv2_corr_factors(int n_dims, const float freq_base, float factors[4]) {
// start and end correction factors
factors[0] = max(0.0f, floor(rope_ntkv2_corr_factor(n_dims, BETA_0, freq_base)));
factors[1] = min(n_dims - 1.0f, ceil(rope_ntkv2_corr_factor(n_dims, BETA_1, freq_base)));
factors[2] = max(0.0f, floor(rope_ntkv2_corr_factor(n_dims, GAMMA_0, freq_base)));
factors[3] = min(n_dims - 1.0f, ceil(rope_ntkv2_corr_factor(n_dims, GAMMA_1, freq_base)));
static void rope_yarn_corr_dims(int n_dims, const float freq_base, float beta_fast, float beta_slow, float dims[2]) {
// start and end correction dims
dims[0] = max(0.0f, floor(rope_yarn_corr_factor(n_dims, beta_fast, freq_base)));
dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, beta_slow, freq_base)));
}
kernel void kernel_rope(
@ -670,33 +658,29 @@ kernel void kernel_rope(
constant int & mode,
constant float & freq_base,
constant float & freq_scale,
constant float & ntk_factor,
constant float & ext_factor,
constant float & attn_factor,
constant float & beta_fast,
constant float & beta_slow,
uint3 tpig[[thread_position_in_grid]]) {
const int64_t i3 = tpig[2];
const int64_t i2 = tpig[1];
const int64_t i1 = tpig[0];
const float theta_scale = pow(freq_base, -2.0f/n_dims);
const float theta_ntk_scale = pow(freq_base * pow(freq_scale, (n_dims / (n_dims - 2.0f))), -2.0f/n_dims);
float corr_factors[4];
rope_ntkv2_corr_factors(n_dims, freq_base, corr_factors);
float corr_dims[2];
rope_yarn_corr_dims(n_dims, freq_base, beta_fast, beta_slow, corr_dims);
float theta_base = (mode & 1) == 0 ? n_past + i2 : i2;
float theta_ntk = theta_base;
const bool is_neox = mode & 2;
if (!is_neox) {
for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
const float theta_linear = freq_scale * theta_base;
const float theta = rope_ntkv2(theta_base, theta_linear, theta_ntk, corr_factors,
i0, ntk_factor, ext_factor);
const float cos_theta = cos(theta);
const float sin_theta = sin(theta);
float cos_theta, sin_theta;
rope_yarn(theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
theta_base *= theta_scale;
theta_ntk *= theta_ntk_scale;
device const float * const src = (device float *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
device float * dst_data = (device float *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);

144
ggml.c
View file

@ -6712,8 +6712,10 @@ static struct ggml_tensor * ggml_rope_impl(
int n_ctx,
float freq_base,
float freq_scale,
float ntk_factor,
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow,
bool inplace) {
GGML_ASSERT(n_past >= 0);
bool is_node = false;
@ -6724,11 +6726,13 @@ 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[8] = { n_past, n_dims, mode, n_ctx };
int32_t params[10] = { n_past, n_dims, mode, n_ctx };
memcpy(params + 4, &freq_base, sizeof(float));
memcpy(params + 5, &freq_scale, sizeof(float));
memcpy(params + 6, &ntk_factor, sizeof(float));
memcpy(params + 7, &ext_factor, sizeof(float));
memcpy(params + 6, &ext_factor, sizeof(float));
memcpy(params + 7, &attn_factor, sizeof(float));
memcpy(params + 8, &beta_fast, sizeof(float));
memcpy(params + 9, &beta_slow, sizeof(float));
ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_ROPE;
@ -6745,7 +6749,7 @@ struct ggml_tensor * ggml_rope(
int n_dims,
int mode,
int n_ctx) {
return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, 0.0f, false);
return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, false);
}
struct ggml_tensor * ggml_rope_inplace(
@ -6755,7 +6759,7 @@ struct ggml_tensor * ggml_rope_inplace(
int n_dims,
int mode,
int n_ctx) {
return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, 0.0f, true);
return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, true);
}
struct ggml_tensor * ggml_rope_custom(
@ -6767,9 +6771,13 @@ struct ggml_tensor * ggml_rope_custom(
int n_ctx,
float freq_base,
float freq_scale,
float ntk_factor,
float ext_factor) {
return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, ntk_factor, ext_factor, false);
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow) {
return ggml_rope_impl(
ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, false
);
}
struct ggml_tensor * ggml_rope_custom_inplace(
@ -6781,9 +6789,13 @@ struct ggml_tensor * ggml_rope_custom_inplace(
int n_ctx,
float freq_base,
float freq_scale,
float ntk_factor,
float ext_factor) {
return ggml_rope_impl(ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, ntk_factor, ext_factor, true);
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow) {
return ggml_rope_impl(
ctx, a, n_past, n_dims, mode, n_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, true
);
}
// ggml_rope_back
@ -12012,52 +12024,40 @@ static void ggml_compute_forward_clamp(
// ggml_compute_forward_rope
static inline float rope_ntkv2_ramp(const float low, const float high, const int i0) {
static inline float rope_yarn_ramp(const float low, const float high, const int i0) {
const float y = (i0 / 2 - low) / MIN(0.001f, high - low);
return 1 - MIN(1, MAX(0, y));
}
// NTKv2 algorithm based on LlamaPartNTKScaledRotaryEmbedding.py from https://github.com/jquesnelle/scaled-rope
// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
static float rope_ntkv2(
const float theta_base,
const float theta_linear,
const float theta_ntk,
const float corr_factors[4],
const int64_t i0,
const float ntk_factor,
const float ext_factor) {
float ramp_mix;
float theta;
static void rope_yarn(
float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
float * cos_theta, float * sin_theta
) {
// Get n-d rotational scaling corrected for extrapolation
float theta_interp = freq_scale * theta_extrap;
float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
float theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
ramp_mix = rope_ntkv2_ramp(corr_factors[0], corr_factors[1], i0) * ntk_factor;
theta = theta_linear * (1 - ramp_mix) + theta_ntk * ramp_mix;
ramp_mix = rope_ntkv2_ramp(corr_factors[2], corr_factors[3], i0) * ext_factor;
theta = theta * (1 - ramp_mix) + theta_base * ramp_mix;
return theta;
// Get n-d magnitude scaling corrected for interpolation
if (freq_scale > 1.0f)
mscale *= 1.0f + 0.1f * logf(freq_scale);
*cos_theta = cosf(theta) * mscale;
*sin_theta = sinf(theta) * mscale;
}
// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
static float ggml_rope_ntkv2_corr_factor(const int n_dims, const float n_rot, const float base) {
// `corr_dim(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
static float ggml_rope_yarn_corr_dim(const int n_dims, const float n_rot, const float base) {
static const float max_pos_emb = 2048;
return n_dims * logf(max_pos_emb / (n_rot * 2 * (float)M_PI)) / (2 * logf(base));
}
void ggml_rope_ntkv2_corr_factors(int n_dims, const float freq_base, float factors[4]) {
// Interpolation constants found experimentally for LLaMA (might not be totally optimal though)
// Do not change unless there is a good reason for doing so!
static const float BETA_0 = 1.75f;
static const float BETA_1 = 1.25f;
static const float GAMMA_0 = 16.0f;
static const float GAMMA_1 = 2.0f;
// start and end correction factors
factors[0] = MAX(0, floorf(ggml_rope_ntkv2_corr_factor(n_dims, BETA_0, freq_base)));
factors[1] = MIN(n_dims - 1, ceilf(ggml_rope_ntkv2_corr_factor(n_dims, BETA_1, freq_base)));
factors[2] = MAX(0, floorf(ggml_rope_ntkv2_corr_factor(n_dims, GAMMA_0, freq_base)));
factors[3] = MIN(n_dims - 1, ceilf(ggml_rope_ntkv2_corr_factor(n_dims, GAMMA_1, freq_base)));
void ggml_rope_yarn_corr_dims(int n_dims, const float freq_base, float beta_fast, float beta_slow, float dims[2]) {
// start and end correction dims
dims[0] = MAX(0, floorf(ggml_rope_yarn_corr_dim(n_dims, beta_fast, freq_base)));
dims[1] = MIN(n_dims - 1, ceilf(ggml_rope_yarn_corr_dim(n_dims, beta_slow, freq_base)));
}
static void ggml_compute_forward_rope_f32(
@ -12069,10 +12069,7 @@ static void ggml_compute_forward_rope_f32(
return;
}
float freq_base;
float freq_scale;
float ntk_factor;
float ext_factor;
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
const int n_past = ((int32_t *) dst->op_params)[0];
const int n_dims = ((int32_t *) dst->op_params)[1];
@ -12080,8 +12077,10 @@ static void ggml_compute_forward_rope_f32(
const int n_ctx = ((int32_t *) dst->op_params)[3];
memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&ntk_factor, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&attn_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&beta_fast, (int32_t *) dst->op_params + 8, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 9, sizeof(float));
assert(n_past >= 0);
@ -12111,9 +12110,8 @@ static void ggml_compute_forward_rope_f32(
int ir = 0;
const float theta_scale = powf(freq_base, -2.0f/n_dims);
const float theta_ntk_scale = powf(freq_base * powf(freq_scale, (n_dims / (n_dims - 2.0f))), -2.0f/n_dims);
float corr_factors[4];
ggml_rope_ntkv2_corr_factors(n_dims, freq_base, corr_factors);
float corr_dims[2];
ggml_rope_yarn_corr_dims(n_dims, freq_base, beta_fast, beta_slow, corr_dims);
const bool is_neox = mode & 2;
const bool is_glm = mode & 4;
@ -12126,7 +12124,6 @@ static void ggml_compute_forward_rope_f32(
if (ir > ir1) break;
float theta_base = (float)p;
float theta_ntk = theta_base;
if (is_glm) {
theta_base = MIN(p, n_ctx - 2);
@ -12155,14 +12152,12 @@ static void ggml_compute_forward_rope_f32(
}
} else if (!is_neox) {
for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
const float theta_linear = freq_scale * theta_base;
const float theta = rope_ntkv2(theta_base, theta_linear, theta_ntk, corr_factors,
i0, ntk_factor, ext_factor);
const float cos_theta = cosf(theta);
const float sin_theta = sinf(theta);
float cos_theta, sin_theta;
rope_yarn(
theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta
);
theta_base *= theta_scale;
theta_ntk *= theta_ntk_scale;
const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
@ -12211,10 +12206,7 @@ static void ggml_compute_forward_rope_f16(
return;
}
float freq_base;
float freq_scale;
float ntk_factor;
float ext_factor;
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
const int n_past = ((int32_t *) dst->op_params)[0];
const int n_dims = ((int32_t *) dst->op_params)[1];
@ -12222,8 +12214,10 @@ static void ggml_compute_forward_rope_f16(
const int n_ctx = ((int32_t *) dst->op_params)[3];
memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&ntk_factor, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&attn_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&beta_fast, (int32_t *) dst->op_params + 8, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 9, sizeof(float));
assert(n_past >= 0);
@ -12253,9 +12247,8 @@ static void ggml_compute_forward_rope_f16(
int ir = 0;
const float theta_scale = powf(freq_base, -2.0f/n_dims);
const float theta_ntk_scale = powf(freq_base * powf(freq_scale, (n_dims / (n_dims - 2.0f))), -2.0f/n_dims);
float corr_factors[4];
ggml_rope_ntkv2_corr_factors(n_dims, freq_base, corr_factors);
float corr_dims[2];
ggml_rope_yarn_corr_dims(n_dims, freq_base, beta_fast, beta_slow, corr_dims);
const bool is_neox = mode & 2;
const bool is_glm = mode & 4;
@ -12268,7 +12261,6 @@ static void ggml_compute_forward_rope_f16(
if (ir > ir1) break;
float theta_base = (float)p;
float theta_ntk = theta_base;
if (is_glm) {
theta_base = MIN(p, n_ctx - 2);
@ -12297,14 +12289,12 @@ static void ggml_compute_forward_rope_f16(
}
} if (!is_neox) {
for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
const float theta_linear = freq_scale * theta_base;
const float theta = rope_ntkv2(theta_base, theta_linear, theta_ntk, corr_factors,
i0, ntk_factor, ext_factor);
const float cos_theta = cosf(theta);
const float sin_theta = sinf(theta);
float cos_theta, sin_theta;
rope_yarn(
theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta
);
theta_base *= theta_scale;
theta_ntk *= theta_ntk_scale;
const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);

16
ggml.h
View file

@ -1195,8 +1195,10 @@ extern "C" {
int n_ctx,
float freq_base,
float freq_scale,
float ntk_factor,
float ext_factor);
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow);
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
@ -1208,11 +1210,13 @@ extern "C" {
int n_ctx,
float freq_base,
float freq_scale,
float ntk_factor,
float ext_factor);
float ext_factor,
float attn_factor,
float beta_fast,
float beta_slow);
// compute correction factors for NTKv2 RoPE scaling
void ggml_rope_ntkv2_corr_factors(int n_dims, const float freq_base, float factors[4]);
// compute correction dims for YaRN RoPE scaling
void ggml_rope_yarn_corr_dims(int n_dims, const float freq_base, float beta_fast, float beta_slow, float dims[2]);
// rotary position embedding backward, i.e compute dx from dy
// a - dy

55
llama.cpp
View file

@ -196,8 +196,10 @@ struct llama_hparams {
float rope_freq_base = 10000.0f;
float rope_freq_scale = 1.0f;
float rope_ntk_factor = 0.0f;
float rope_ext_factor = 0.0f;
float rope_attn_factor = 1.0f;
float rope_beta_fast = 0.0f;
float rope_beta_slow = 0.0f;
enum llama_ftype ftype = LLAMA_FTYPE_MOSTLY_F16;
@ -900,8 +902,10 @@ struct llama_context_params llama_context_default_params() {
/*.tensor_split =*/ nullptr,
/*.rope_freq_base =*/ 10000.0f,
/*.rope_freq_scale =*/ 1.0f,
/*.rope_ntk_factor =*/ 0.0f,
/*.rope_ext_factor =*/ 0.0f,
/*.rope_attn_factor =*/ 1.0f,
/*.rope_beta_fast =*/ 32.0f,
/*.rope_beta_slow =*/ 1.0f,
/*.progress_callback =*/ nullptr,
/*.progress_callback_user_data =*/ nullptr,
/*.low_vram =*/ false,
@ -1036,8 +1040,10 @@ static void llama_model_load_internal(
const bool mul_mat_q,
float rope_freq_base,
float rope_freq_scale,
float rope_ntk_factor,
float rope_ext_factor,
float rope_attn_factor,
float rope_beta_fast,
float rope_beta_slow,
bool low_vram,
ggml_type memory_type,
bool use_mmap,
@ -1089,8 +1095,10 @@ static void llama_model_load_internal(
hparams.rope_freq_base = rope_freq_base;
hparams.rope_freq_scale = rope_freq_scale;
hparams.rope_ntk_factor = rope_ntk_factor;
hparams.rope_ext_factor = rope_ext_factor;
hparams.rope_attn_factor = rope_attn_factor;
hparams.rope_beta_fast = rope_beta_fast;
hparams.rope_beta_slow = rope_beta_slow;
}
// ref: https://github.com/facebookresearch/llama/blob/6c7fe276574e78057f917549435a2554000a876d/llama/model.py#L194-L199
@ -1114,8 +1122,10 @@ static void llama_model_load_internal(
fprintf(stderr, "%s: n_ff = %u\n", __func__, n_ff);
fprintf(stderr, "%s: freq_base = %.1f\n", __func__, hparams.rope_freq_base);
fprintf(stderr, "%s: freq_scale = %g\n", __func__, hparams.rope_freq_scale);
fprintf(stderr, "%s: ntk_factor = %g\n", __func__, hparams.rope_ntk_factor);
fprintf(stderr, "%s: ext_factor = %g\n", __func__, hparams.rope_ext_factor);
fprintf(stderr, "%s: attn_factor = %g\n", __func__, hparams.rope_attn_factor);
fprintf(stderr, "%s: beta_fast = %g\n", __func__, hparams.rope_beta_fast);
fprintf(stderr, "%s: beta_slow = %g\n", __func__, hparams.rope_beta_slow);
fprintf(stderr, "%s: ftype = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype));
fprintf(stderr, "%s: model size = %s\n", __func__, llama_model_type_name(model.type));
}
@ -1384,8 +1394,10 @@ static bool llama_model_load(
const bool mul_mat_q,
float rope_freq_base,
float rope_freq_scale,
float rope_ntk_factor,
float rope_ext_factor,
float rope_attn_factor,
float rope_beta_fast,
float rope_beta_slow,
bool low_vram,
ggml_type memory_type,
bool use_mmap,
@ -1394,10 +1406,11 @@ static bool llama_model_load(
llama_progress_callback progress_callback,
void *progress_callback_user_data) {
try {
llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gqa, rms_norm_eps, n_gpu_layers, main_gpu,
tensor_split, mul_mat_q, rope_freq_base, rope_freq_scale, rope_ntk_factor,
rope_ext_factor, low_vram, memory_type, use_mmap, use_mlock, vocab_only,
progress_callback, progress_callback_user_data);
llama_model_load_internal(
fname, model, vocab, n_ctx, n_batch, n_gqa, rms_norm_eps, n_gpu_layers, main_gpu, tensor_split, mul_mat_q,
rope_freq_base, rope_freq_scale, rope_ext_factor, rope_attn_factor, rope_beta_fast, rope_beta_slow,
low_vram, memory_type, use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data
);
return true;
} catch (const std::exception & err) {
fprintf(stderr, "error loading model: %s\n", err.what());
@ -1435,8 +1448,10 @@ static struct ggml_cgraph * llama_build_graph(
const float freq_base = hparams.rope_freq_base;
const float freq_scale = hparams.rope_freq_scale;
const float ntk_factor = hparams.rope_ntk_factor;
const float ext_factor = hparams.rope_ext_factor;
const float attn_factor = hparams.rope_attn_factor;
const float beta_fast = hparams.rope_beta_fast;
const float beta_slow = hparams.rope_beta_slow;
const float rms_norm_eps = hparams.f_rms_norm_eps;
const int n_gpu_layers = model.n_gpu_layers;
@ -1568,13 +1583,15 @@ static struct ggml_cgraph * llama_build_graph(
struct ggml_tensor * Kcur = ggml_rope_custom_inplace(
ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, N), n_past, n_embd_head, 0, 0, freq_base,
freq_scale, ntk_factor, ext_factor);
freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
offload_func_kq(Kcur);
ggml_set_name(Kcur, "Kcur");
struct ggml_tensor * Qcur = ggml_rope_custom_inplace(
ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, N), n_past, n_embd_head, 0, 0, freq_base,
freq_scale, ntk_factor, ext_factor);
freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
offload_func_kq(Qcur);
ggml_set_name(Qcur, "Qcur");
@ -3216,11 +3233,13 @@ struct llama_model * llama_load_model_from_file(
ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
if (!llama_model_load(path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gqa,
params.rms_norm_eps, params.n_gpu_layers, params.main_gpu, params.tensor_split, params.mul_mat_q,
params.rope_freq_base, params.rope_freq_scale, params.rope_ntk_factor, params.rope_ext_factor,
params.low_vram, memory_type, params.use_mmap, params.use_mlock, params.vocab_only,
params.progress_callback, params.progress_callback_user_data)) {
if (!llama_model_load(
path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gqa, params.rms_norm_eps,
params.n_gpu_layers, params.main_gpu, params.tensor_split, params.mul_mat_q, params.rope_freq_base,
params.rope_freq_scale, params.rope_ext_factor, params.rope_attn_factor, params.rope_beta_fast,
params.rope_beta_slow, params.low_vram, memory_type, params.use_mmap, params.use_mlock, params.vocab_only,
params.progress_callback, params.progress_callback_user_data
)) {
delete model;
fprintf(stderr, "%s: failed to load model\n", __func__);
return nullptr;

4
llama.h
View file

@ -100,8 +100,10 @@ extern "C" {
// ref: https://github.com/ggerganov/llama.cpp/pull/2054
float rope_freq_base; // RoPE base frequency
float rope_freq_scale; // RoPE frequency scaling factor
float rope_ntk_factor; // RoPE NTK mix factor
float rope_ext_factor; // RoPE extrapolation mix factor
float rope_attn_factor; // RoPE magnitude scaling factor
float rope_beta_fast; // RoPE low correction dim
float rope_beta_slow; // RoPE high correction dim
// called with a progress value between 0 and 1, pass NULL to disable
llama_progress_callback progress_callback;