vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 17:14:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

GradientAI Auto ROPE Base calculation (#910)

Browse source 
* GradientAI Auto ROPE Base calculation

https://gradient.ai/blog/scaling-rotational-embeddings-for-long-context-language-models
has a formula that better fits the ideal rope scaling. 

Tested with Lllama3, checked calculation is correct for llama2. Retains logic for not scaling rope if under trained CTX.

* add in solar scaling logic

Solar based models require the context values to be multiplied by 8. This is (i'm guessing) because the positions as based on a 32k context, but sliding window of 4k.

* Update model_adapter.h

adding in tensor count to identify solar models based on tensor count of 435.

* Update model_adapter.cpp

add in n_tensor count for solar identification

* refactor and cleanup GradientAI rope scaling

---------

Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
This commit is contained in:
askmyteapot 2024-06-13 20:12:00 +10:00 • committed by GitHub
parent 49e4c3fd7b
commit 1e72b65c38
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 39 additions and 22 deletions
Download patch file Download diff file Expand all files Collapse all files

44
gpttype_adapter.cpp
View file

@ -7,6 +7,7 @@
//No dynamic memory allocation! Setup structs with FIXED (known) shapes and sizes for ALL output fields
//Python will ALWAYS provide the memory, we just write to it.
#include <cmath>
#include <time.h>
#include <mutex>
#include "model_adapter.h"
@ -787,6 +788,19 @@ static int GetBatchSize(int desiredBlasBatchSize,FileFormat in_file_format)
return desiredBlasBatchSize;
}
//this function applies automatic scaling to rope freq base when the desired context exceeds trained context
static float CalcGradientAIRopeFreqBase(float original_rope_base, int n_ctx_train, int n_ctx_desired, bool is_solar)
{
if(n_ctx_desired <= n_ctx_train || n_ctx_desired <= 2048)
{
return original_rope_base;
}
float ctx_multiplier = (is_solar?8.0f:1.0f);
float chi_ctx_train_value = (n_ctx_train * ctx_multiplier) / 6.28318;
float chi_ctx_value = (n_ctx_desired * ctx_multiplier) / 6.28318;
return powf(original_rope_base, logf(chi_ctx_value) / logf(chi_ctx_train_value));
}
ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in_file_format, FileFormatExtraMeta in_file_format_meta)
{
ggml_time_init();
@ -835,28 +849,16 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
}
else
{
rope_freq_scale = 1.0f;
if (kcpp_params->n_ctx <= 2048) //normie mode
//Set freq base for all, including non GGUF. If we are using GGUF, this will be overwritten with more accurate values later.
rope_freq_base = CalcGradientAIRopeFreqBase(10000.0f,2048,kcpp_params->n_ctx,false);
if(file_format==FileFormat::GGUF_GENERIC)
{
rope_freq_base = 10000.0f;
printf("Using automatic RoPE scaling. If the model has customized RoPE settings, they will be used directly instead!\n");
}
else
{
//approximate NTK aware ctx
auto effectivenctx = kcpp_params->n_ctx;
if((file_format == FileFormat::GGUF_GENERIC) && file_format_meta.n_ctx_train > 2048)
{
float factor = file_format_meta.n_ctx_train/2048;
effectivenctx = effectivenctx/factor;
}
float magic_multiplier = 8.0f;
float base_multiplier = effectivenctx*magic_multiplier;
float base_raw = 10000.0f;
rope_freq_base = (effectivenctx <= 2048 ? base_raw : base_multiplier);
printf("Using Automatic RoPE scaling, Pre-GGUF (scale:%.3f, base:%.1f).\n",rope_freq_scale, rope_freq_base);
}
printf("Using automatic RoPE scaling. If the model has customized RoPE settings, they will be used directly instead!\n");
}
gptj_ctx_v3.hparams.rope_freq_scale = neox_ctx_v3.hparams.rope_freq_scale = rope_freq_scale;
gptj_ctx_v3.hparams.rope_freq_base = neox_ctx_v3.hparams.rope_freq_base = rope_freq_base;
@ -1085,7 +1087,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
}
else
{
//if the model modifes rope in any way, use the model values. Otherwise, use our automatic ones
//if the model modifes rope in any way, or uses yarn, use the model values. Otherwise, use our automatic ones
//special exception for llama, which uses auto scale
if((llamamodel->hparams.rope_freq_base_train!=10000.0f && llamamodel->hparams.rope_freq_base_train!=500000.0f) ||
llamamodel->hparams.rope_freq_scale_train!=1.0f ||
@ -1095,8 +1097,8 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
}
else
{
float multiplier_rope_base = llamamodel->hparams.rope_freq_base_train/10000.0f;
rope_freq_base *= multiplier_rope_base;
//Calculate rope_freq_base using the gradientAI formula, solar requires ctx *8 for correct scaling
rope_freq_base = CalcGradientAIRopeFreqBase(llamamodel->hparams.rope_freq_base_train, file_format_meta.n_ctx_train, kcpp_params->n_ctx, file_format_meta.model_architecture==GGUFArch::ARCH_SOLAR);
llama_ctx_params.rope_freq_base = rope_freq_base;
llama_ctx_params.rope_freq_scale = rope_freq_scale;
printf("Automatic RoPE Scaling: Using (scale:%.3f, base:%.1f).\n", rope_freq_scale, rope_freq_base);
@ -2467,4 +2469,4 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
concat_output_mtx.unlock();
output.text = concat_output_reader_copy_res.c_str();
return output;
}
}