mirror of
https://github.com/ravenscroftj/turbopilot.git
synced 2024-10-01 01:06:01 -04:00
Merge pull request #43 from ravenscroftj/feature/gptneox
Feature/gptneox
This commit is contained in:
James Ravenscroft
GitHub
commit
de68ab022c
16
MODELS.md
16
MODELS.md
@ -1,10 +1,20 @@
|
||||
# Models Directory
|
||||
|
||||
## StableCode Instruct State-of-the-art for low Spec machines(Released 8th August 2023)
|
||||
|
||||
[StableCode](https://stability.ai/blog/stablecode-llm-generative-ai-coding) Instruct is a new model from [Stability.ai](https://stability.ai/) which provides reasonable autocomplete suggestions in approx 3GiB of RAM.
|
||||
|
||||
| Model Name | RAM Requirement | Direct Download | HF Project Link |
|
||||
|---------------------|-----------------|-----------------|-----------------|
|
||||
| StarCoder | ~3GiB | [:arrow_down:](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML/blob/main/stablecode-instruct-alpha-3b.ggmlv1.q4_0.bin) | [:hugs:](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML/) |
|
||||
|
||||
|
||||
|
||||
## "Coder" family models
|
||||
|
||||
WizardCoder, StarCoder and SantaCoder are current "state-of-the-art" autocomplete models
|
||||
|
||||
### SantaCoder (Best Small model)
|
||||
### SantaCoder (Small Model, Reasonable on lower spec machines - Released 13/4/2023)
|
||||
|
||||
[SantaCoder](https://huggingface.co/bigcode/santacoder) is a smaller version of the StarCoder and WizardCoder family with only 1.1 Billion parameters. The model is trained with fill-in-the-middle objective allowing it to be used to auto-complete function parameters.
|
||||
|
||||
@ -18,7 +28,7 @@ This model is primarily trained on Python, Java and Javscript.
|
||||
To run in Turbopilot set model type `-m starcoder`
|
||||
|
||||
|
||||
### WizardCoder (Best Autocomplete Performance, Compute-Hungry)
|
||||
### WizardCoder 15B Best Autocomplete Performance, Compute-Hungry (Released 15/6/2023)
|
||||
|
||||
[WizardCoder](https://github.com/nlpxucan/WizardLM/tree/main/WizardCoder) is the current SOTA auto complete model, it is an updated version of StarCoder that achieves 57.1 pass@1 on HumanEval benchmarks (essentially in 57% of cases it correctly solves a given challenge. Read more about how this metric works in the scientific paper [here](https://arxiv.org/pdf/2107.03374.pdf) ).
|
||||
|
||||
@ -32,7 +42,7 @@ Even when quantized, WizardCoder is a large model that takes up a significant am
|
||||
To run in Turbopilot set model type `-m starcoder`
|
||||
|
||||
|
||||
### StarCoder
|
||||
### StarCoder (Released 4/5/2023)
|
||||
|
||||
[StarCoder](https://huggingface.co/blog/starcoder) held the previous title of state-of-the-art coding model back in May 2023. It is still a reasonably good model by comparison but it is a similar size and has similar RAM and compute requirements to WizardCoder so you may be better off just running that. Links below provided for posterity.
|
||||
|
||||
|
||||
12
README.md
12
README.md
@ -9,10 +9,11 @@ TurboPilot is a self-hosted [copilot](https://github.com/features/copilot) clone
|
||||
|
||||
|
||||
|
||||
**✨ Now Supports [StableCode 3B Instruct](https://huggingface.co/stabilityai/stablecode-instruct-alpha-3b)** simply use [TheBloke's Quantized GGML models](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML) and set `-m stablecode`.
|
||||
|
||||
**New: Refactored + Simplified**: The source code has been improved to make it easier to extend and add new models to Turbopilot. The system now supports multiple flavours of model
|
||||
**✨ New: Refactored + Simplified**: The source code has been improved to make it easier to extend and add new models to Turbopilot. The system now supports multiple flavours of model
|
||||
|
||||
**New: Wizardcoder, Starcoder, Santacoder support** - Turbopilot now supports state of the art local code completion models which provide more programming languages and "fill in the middle" support.
|
||||
**✨ New: Wizardcoder, Starcoder, Santacoder support** - Turbopilot now supports state of the art local code completion models which provide more programming languages and "fill in the middle" support.
|
||||
|
||||
## 🤝 Contributing
|
||||
|
||||
@ -34,7 +35,7 @@ You have 2 options for getting the model
|
||||
|
||||
You can download the pre-converted, pre-quantized models from Huggingface.
|
||||
|
||||
For low RAM users (4-8 GiB), I recommend [SantaCoder](https://huggingface.co/mike-ravkine/gpt_bigcode-santacoder-GGML/resolve/main/santacoder-q4_0.bin) and for high power users (16+ GiB RAM, discrete GPU or apple silicon) I recomnmend [WizardCoder](https://huggingface.co/TheBloke/WizardCoder-15B-1.0-GGML/resolve/main/WizardCoder-15B-1.0.ggmlv3.q4_0.bin).
|
||||
For low RAM users (4-8 GiB), I recommend [StableCode](https://huggingface.co/TheBloke/stablecode-instruct-alpha-3b-GGML) and for high power users (16+ GiB RAM, discrete GPU or apple silicon) I recomnmend [WizardCoder](https://huggingface.co/TheBloke/WizardCoder-15B-1.0-GGML/resolve/main/WizardCoder-15B-1.0.ggmlv3.q4_0.bin).
|
||||
|
||||
Turbopilot still supports the first generation codegen models from `v0.0.5` and earlier builds. Although old models do need to be requantized.
|
||||
|
||||
@ -65,7 +66,7 @@ If you have a multi-core system you can control how many CPUs are used with the
|
||||
|
||||
To run the legacy codegen models. Just change the model type flag `-m` to `codegen` instead.
|
||||
|
||||
**NOTE: the latest version of GGML requires that you re-quantize your codegen models. Old models downloaded from here will no longer work. I am working on providing updated quantized codegen models**
|
||||
**NOTE: Turbopilot 0.1.0 and newer re-quantize your codegen models old models from v0.0.5 and older. I am working on providing updated quantized codegen models**
|
||||
|
||||
### 📦 Running From Docker
|
||||
|
||||
@ -91,7 +92,8 @@ As of release v0.0.5 turbocode now supports CUDA inference. In order to run the
|
||||
docker run --gpus=all --rm -it \
|
||||
-v ./models:/models \
|
||||
-e THREADS=6 \
|
||||
-e MODEL="/models/codegen-2B-multi-ggml-4bit-quant.bin" \
|
||||
-e MODEL_TYPE=starcoder \
|
||||
-e MODEL="/models/santacoder-q4_0.bin" \
|
||||
-p 18080:18080 \
|
||||
ghcr.io/ravenscroftj/turbopilot:v0.1.0-cuda11
|
||||
```
|
||||
|
||||
87
include/turbopilot/gptneox.hpp
Normal file
87
include/turbopilot/gptneox.hpp
Normal file
@ -0,0 +1,87 @@
|
||||
#ifndef __TURBOPILOT_GPTNEOX_H
|
||||
#define __TURBOPILOT_GPTNEOX_H
|
||||
|
||||
#include <turbopilot/model.hpp>
|
||||
|
||||
#include <vector>
|
||||
#include <map>
|
||||
|
||||
// default hparams (StableLM 3B)
|
||||
struct gpt_neox_hparams {
|
||||
int32_t n_vocab = 50257;
|
||||
int32_t n_ctx = 4096;
|
||||
int32_t n_embd = 4096;
|
||||
int32_t n_head = 32;
|
||||
int32_t n_layer = 16;
|
||||
int32_t n_rot = 32; // rotary_pct * (n_embd / n_head)
|
||||
int32_t par_res = 1; // 1 = true, 0 = false
|
||||
int32_t ftype = 1;
|
||||
};
|
||||
|
||||
struct gpt_neox_layer {
|
||||
// pre normalization
|
||||
struct ggml_tensor * ln_1_g;
|
||||
struct ggml_tensor * ln_1_b;
|
||||
|
||||
// attention
|
||||
struct ggml_tensor * c_attn_attn_w;
|
||||
struct ggml_tensor * c_attn_attn_b;
|
||||
|
||||
struct ggml_tensor * c_attn_proj_w;
|
||||
struct ggml_tensor * c_attn_proj_b;
|
||||
|
||||
// post normalization
|
||||
struct ggml_tensor * ln_2_g;
|
||||
struct ggml_tensor * ln_2_b;
|
||||
|
||||
// ff
|
||||
struct ggml_tensor * c_mlp_fc_w;
|
||||
struct ggml_tensor * c_mlp_fc_b;
|
||||
|
||||
struct ggml_tensor * c_mlp_proj_w;
|
||||
struct ggml_tensor * c_mlp_proj_b;
|
||||
};
|
||||
|
||||
struct gpt_neox_model {
|
||||
gpt_neox_hparams hparams;
|
||||
|
||||
// normalization
|
||||
struct ggml_tensor * ln_f_g;
|
||||
struct ggml_tensor * ln_f_b;
|
||||
|
||||
struct ggml_tensor * wte; // position embedding
|
||||
|
||||
struct ggml_tensor * lmh_g; // language model head
|
||||
//struct ggml_tensor * lmh_b; // language model bias
|
||||
|
||||
std::vector<gpt_neox_layer> layers;
|
||||
|
||||
// key + value memory
|
||||
struct ggml_tensor * memory_k;
|
||||
struct ggml_tensor * memory_v;
|
||||
|
||||
//
|
||||
struct ggml_context * ctx;
|
||||
std::map<std::string, struct ggml_tensor *> tensors;
|
||||
};
|
||||
|
||||
|
||||
class GPTNEOXModel : public TurbopilotModel {
|
||||
|
||||
public:
|
||||
GPTNEOXModel(ModelConfig config, std::mt19937 &rng) : TurbopilotModel(config, rng){
|
||||
this->model = new gpt_neox_model{};
|
||||
this->vocab = new gpt_vocab{};
|
||||
}
|
||||
virtual ~GPTNEOXModel();
|
||||
bool load_model(std::string path);
|
||||
virtual std::stringstream predict(std::string prompt, int max_length, bool include_prompt);
|
||||
|
||||
private:
|
||||
gpt_neox_model *model = NULL;
|
||||
gpt_vocab *vocab = NULL;
|
||||
|
||||
|
||||
};
|
||||
|
||||
#endif // __TURBOPILOT_GPTNEOX_H
|
||||
@ -6,11 +6,13 @@ include_directories(${Boost_INCLUDE_DIRS})
|
||||
add_executable(${TURBOPILOT_TARGET}
|
||||
main.cpp
|
||||
gptj.cpp
|
||||
gptneox.cpp
|
||||
common.cpp
|
||||
server.cpp
|
||||
starcoder.cpp
|
||||
../include/turbopilot/model.hpp
|
||||
../include/turbopilot/gptj.hpp
|
||||
../include/turbopilot/gptneox.hpp
|
||||
../include/turbopilot/starcoder.hpp
|
||||
)
|
||||
|
||||
|
||||
707
src/gptneox.cpp
Normal file
707
src/gptneox.cpp
Normal file
@ -0,0 +1,707 @@
|
||||
#include <turbopilot/gptneox.hpp>
|
||||
#include <spdlog/spdlog.h>
|
||||
|
||||
#include <ggml/ggml.h>
|
||||
|
||||
#include <cinttypes>
|
||||
|
||||
#include <iostream>
|
||||
#include <fstream>
|
||||
|
||||
#if defined(_MSC_VER)
|
||||
#pragma warning(disable: 4244 4267) // possible loss of data
|
||||
#endif
|
||||
|
||||
|
||||
// feed-forward network
|
||||
ggml_tensor * gpt_neox_ff(
|
||||
const gpt_neox_layer &layer,
|
||||
ggml_context * ctx0,
|
||||
ggml_tensor * inp) {
|
||||
ggml_tensor * cur = ggml_norm(ctx0, inp);
|
||||
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_mul(ctx0,
|
||||
ggml_repeat(ctx0, layer.ln_2_g, cur),
|
||||
cur),
|
||||
ggml_repeat(ctx0, layer.ln_2_b, cur));
|
||||
|
||||
cur = ggml_mul_mat(ctx0,
|
||||
layer.c_mlp_fc_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_repeat(ctx0, layer.c_mlp_fc_b, cur),
|
||||
cur);
|
||||
|
||||
// GELU activation
|
||||
cur = ggml_gelu(ctx0, cur);
|
||||
|
||||
// projection
|
||||
// cur = proj_w*cur + proj_b
|
||||
cur = ggml_mul_mat(ctx0,
|
||||
layer.c_mlp_proj_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_repeat(ctx0, layer.c_mlp_proj_b, cur),
|
||||
cur);
|
||||
return cur;
|
||||
}
|
||||
|
||||
|
||||
// evaluate the transformer
|
||||
//
|
||||
// - model: the model
|
||||
// - n_threads: number of threads to use
|
||||
// - n_past: the context size so far
|
||||
// - embd_inp: the embeddings of the tokens in the context
|
||||
// - embd_w: the predicted logits for the next token
|
||||
//
|
||||
bool gpt_neox_eval(
|
||||
const gpt_neox_model & model,
|
||||
const int n_threads,
|
||||
const int n_past,
|
||||
const std::vector<gpt_vocab::id> & embd_inp,
|
||||
std::vector<float> & embd_w,
|
||||
size_t & mem_per_token) {
|
||||
const int N = embd_inp.size();
|
||||
|
||||
const auto & hparams = model.hparams;
|
||||
|
||||
const int n_embd = hparams.n_embd;
|
||||
const int n_layer = hparams.n_layer;
|
||||
const int n_ctx = hparams.n_ctx;
|
||||
const int n_head = hparams.n_head;
|
||||
const int n_vocab = hparams.n_vocab;
|
||||
const int n_rot = hparams.n_rot;
|
||||
|
||||
static size_t buf_size = 256u*1024*1024;
|
||||
static void * buf = malloc(buf_size);
|
||||
|
||||
// use 2 scratch buffers
|
||||
// TODO: very hacky solution - reimplement in a more elegant way
|
||||
static size_t scr0_size = 256u*1024*1024;
|
||||
static void * scr0 = malloc(scr0_size);
|
||||
|
||||
static size_t scr1_size = 256u*1024*1024;
|
||||
static void * scr1 = malloc(scr1_size);
|
||||
|
||||
if (mem_per_token > 0 && mem_per_token*N > buf_size) {
|
||||
const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
|
||||
//printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
|
||||
|
||||
// reallocate
|
||||
buf_size = buf_size_new;
|
||||
buf = realloc(buf, buf_size);
|
||||
if (buf == nullptr) {
|
||||
fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
struct ggml_init_params params = {
|
||||
/*.mem_size =*/ buf_size,
|
||||
/*.mem_buffer =*/ buf,
|
||||
/*.no_alloc =*/ false,
|
||||
};
|
||||
|
||||
struct ggml_context * ctx0 = ggml_init(params);
|
||||
struct ggml_cgraph gf = {};
|
||||
|
||||
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
|
||||
memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
|
||||
|
||||
// wte
|
||||
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
struct ggml_tensor * cur;
|
||||
|
||||
ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
|
||||
|
||||
// self-attention
|
||||
{
|
||||
{
|
||||
cur = ggml_norm(ctx0, inpL);
|
||||
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_mul(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
|
||||
cur),
|
||||
ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
|
||||
}
|
||||
|
||||
// compute QKV
|
||||
{
|
||||
cur = ggml_mul_mat(ctx0,
|
||||
model.layers[il].c_attn_attn_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
|
||||
cur);
|
||||
}
|
||||
|
||||
struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 0*sizeof(float)*n_embd/n_head));
|
||||
struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 1*sizeof(float)*n_embd/n_head));
|
||||
struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 2*sizeof(float)*n_embd/n_head));
|
||||
|
||||
// using mode = 2 for GPT-NeoX mode
|
||||
Qcur = ggml_rope_inplace(ctx0, Qcur, n_past, n_rot, 2, 0);
|
||||
Kcur = ggml_rope_inplace(ctx0, Kcur, n_past, n_rot, 2, 0);
|
||||
|
||||
// store key and value to memory
|
||||
{
|
||||
Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd, N));
|
||||
|
||||
struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
|
||||
struct ggml_tensor * v = ggml_view_2d(ctx0, model.memory_v, N, n_embd,
|
||||
( n_ctx)*ggml_element_size(model.memory_v),
|
||||
(il*n_ctx)*ggml_element_size(model.memory_v)*n_embd + n_past*ggml_element_size(model.memory_v));
|
||||
|
||||
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
|
||||
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
|
||||
}
|
||||
|
||||
// Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
|
||||
struct ggml_tensor * Q =
|
||||
ggml_permute(ctx0,
|
||||
Qcur,
|
||||
0, 2, 1, 3);
|
||||
|
||||
// K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
|
||||
struct ggml_tensor * K =
|
||||
ggml_permute(ctx0,
|
||||
ggml_reshape_3d(ctx0,
|
||||
ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
|
||||
n_embd/n_head, n_head, n_past + N),
|
||||
0, 2, 1, 3);
|
||||
|
||||
// K * Q
|
||||
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
|
||||
|
||||
// KQ_scaled = KQ / sqrt(n_embd/n_head)
|
||||
struct ggml_tensor * KQ_scaled =
|
||||
ggml_scale_inplace(ctx0,
|
||||
KQ,
|
||||
ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
|
||||
);
|
||||
|
||||
// KQ_masked = mask_past(KQ_scaled)
|
||||
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
|
||||
|
||||
// KQ = soft_max(KQ_masked)
|
||||
struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
|
||||
|
||||
// V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
|
||||
struct ggml_tensor * V =
|
||||
ggml_view_3d(ctx0, model.memory_v,
|
||||
n_past + N, n_embd/n_head, n_head,
|
||||
n_ctx*ggml_element_size(model.memory_v),
|
||||
n_ctx*ggml_element_size(model.memory_v)*n_embd/n_head,
|
||||
il*n_ctx*ggml_element_size(model.memory_v)*n_embd);
|
||||
|
||||
// KQV = transpose(V) * KQ_soft_max
|
||||
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
|
||||
|
||||
// KQV_merged = KQV.permute(0, 2, 1, 3)
|
||||
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
|
||||
|
||||
// cur = KQV_merged.contiguous().view(n_embd, N)
|
||||
cur = ggml_cpy(ctx0,
|
||||
KQV_merged,
|
||||
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
|
||||
|
||||
// projection
|
||||
{
|
||||
cur = ggml_mul_mat(ctx0,
|
||||
model.layers[il].c_attn_proj_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur), cur);
|
||||
}
|
||||
}
|
||||
|
||||
ggml_set_scratch(ctx0, { 0, scr1_size, scr1, });
|
||||
|
||||
if (hparams.par_res == 0) {
|
||||
struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpL);
|
||||
|
||||
cur = gpt_neox_ff(model.layers[il], ctx0, inpFF);
|
||||
|
||||
// input for next layer
|
||||
inpL = ggml_add(ctx0, cur, inpFF);
|
||||
} else {
|
||||
struct ggml_tensor * inpFF = cur;
|
||||
|
||||
// this is independent of the self-attention result, so it could be done in parallel to the self-attention
|
||||
// note here we pass inpL instead of cur
|
||||
cur = gpt_neox_ff(model.layers[il], ctx0, inpL);
|
||||
|
||||
// layer input + FF
|
||||
cur = ggml_add(ctx0, cur, inpFF);
|
||||
|
||||
// input for next layer
|
||||
inpL = ggml_add(ctx0, cur, inpL);
|
||||
}
|
||||
}
|
||||
|
||||
ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
|
||||
|
||||
// norm
|
||||
{
|
||||
inpL = ggml_norm(ctx0, inpL);
|
||||
|
||||
// inpL = ln_f_g*inpL + ln_f_b
|
||||
inpL = ggml_add(ctx0,
|
||||
ggml_mul(ctx0,
|
||||
ggml_repeat(ctx0, model.ln_f_g, inpL),
|
||||
inpL),
|
||||
ggml_repeat(ctx0, model.ln_f_b, inpL));
|
||||
}
|
||||
|
||||
ggml_set_scratch(ctx0, { 0, 0, nullptr, });
|
||||
|
||||
// lm_head
|
||||
{
|
||||
inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
|
||||
|
||||
//inpL = ggml_add(ctx0,
|
||||
// ggml_repeat(ctx0, model.lmh_b, inpL),
|
||||
// inpL);
|
||||
}
|
||||
|
||||
// logits -> probs
|
||||
//inpL = ggml_soft_max_inplace(ctx0, inpL);
|
||||
|
||||
// run the computation
|
||||
ggml_build_forward_expand(&gf, inpL);
|
||||
ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
|
||||
|
||||
//if (n_past%100 == 0) {
|
||||
// ggml_graph_print (&gf);
|
||||
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
|
||||
//}
|
||||
|
||||
//embd_w.resize(n_vocab*N);
|
||||
//memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
|
||||
|
||||
// return result for just the last token
|
||||
embd_w.resize(n_vocab);
|
||||
memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
|
||||
|
||||
if (mem_per_token == 0) {
|
||||
mem_per_token = ggml_used_mem(ctx0)/N;
|
||||
}
|
||||
//printf("used_mem = %zu\n", ggml_used_mem(ctx0));
|
||||
|
||||
ggml_free(ctx0);
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
|
||||
GPTNEOXModel::~GPTNEOXModel(){
|
||||
ggml_free(model->ctx);
|
||||
free(model);
|
||||
free(vocab);
|
||||
}
|
||||
|
||||
bool GPTNEOXModel::load_model(std::string fname) {
|
||||
printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
|
||||
|
||||
auto fin = std::ifstream(fname, std::ios::binary);
|
||||
if (!fin) {
|
||||
fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
|
||||
return false;
|
||||
}
|
||||
|
||||
// verify magic
|
||||
{
|
||||
uint32_t magic;
|
||||
fin.read((char *) &magic, sizeof(magic));
|
||||
if (magic != GGML_FILE_MAGIC) {
|
||||
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
// load hparams
|
||||
{
|
||||
auto & hparams = model->hparams;
|
||||
|
||||
fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
|
||||
fin.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
|
||||
fin.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
|
||||
fin.read((char *) &hparams.n_head, sizeof(hparams.n_head));
|
||||
fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
|
||||
fin.read((char *) &hparams.n_rot, sizeof(hparams.n_rot));
|
||||
fin.read((char *) &hparams.par_res, sizeof(hparams.par_res));
|
||||
fin.read((char *) &hparams.ftype, sizeof(hparams.ftype));
|
||||
|
||||
const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
|
||||
|
||||
printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
|
||||
printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx);
|
||||
printf("%s: n_embd = %d\n", __func__, hparams.n_embd);
|
||||
printf("%s: n_head = %d\n", __func__, hparams.n_head);
|
||||
printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
|
||||
printf("%s: n_rot = %d\n", __func__, hparams.n_rot);
|
||||
printf("%s: par_res = %d\n", __func__, hparams.par_res);
|
||||
printf("%s: ftype = %d\n", __func__, hparams.ftype);
|
||||
printf("%s: qntvr = %d\n", __func__, qntvr);
|
||||
|
||||
hparams.ftype %= GGML_QNT_VERSION_FACTOR;
|
||||
}
|
||||
|
||||
// load vocab
|
||||
{
|
||||
const int32_t n_vocab = model->hparams.n_vocab;
|
||||
|
||||
std::string word;
|
||||
std::vector<char> buf(128);
|
||||
|
||||
for (int i = 0; i < n_vocab; i++) {
|
||||
uint32_t len;
|
||||
fin.read((char *) &len, sizeof(len));
|
||||
|
||||
buf.resize(len);
|
||||
fin.read((char *) buf.data(), len);
|
||||
word.assign(buf.data(), len);
|
||||
|
||||
vocab->token_to_id[word] = i;
|
||||
vocab->id_to_token[i] = word;
|
||||
}
|
||||
}
|
||||
|
||||
// for the big tensors, we have the option to store the data in 16-bit floats or quantized
|
||||
// in order to save memory and also to speed up the computation
|
||||
ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model->hparams.ftype));
|
||||
if (wtype == GGML_TYPE_COUNT) {
|
||||
fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
|
||||
__func__, fname.c_str(), model->hparams.ftype);
|
||||
return false;
|
||||
}
|
||||
|
||||
auto & ctx = model->ctx;
|
||||
|
||||
size_t ctx_size = 0;
|
||||
|
||||
{
|
||||
const auto & hparams = model->hparams;
|
||||
|
||||
const size_t n_embd = hparams.n_embd;
|
||||
const size_t n_layer = hparams.n_layer;
|
||||
const size_t n_ctx = hparams.n_ctx;
|
||||
const size_t n_vocab = hparams.n_vocab;
|
||||
|
||||
ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
|
||||
ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
|
||||
|
||||
ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
|
||||
|
||||
ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // lmh_g
|
||||
//ctx_size += n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
|
||||
|
||||
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
|
||||
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
|
||||
|
||||
ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_attn_w
|
||||
ctx_size += n_layer*( 3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
|
||||
|
||||
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
|
||||
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_proj_b
|
||||
|
||||
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
|
||||
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
|
||||
|
||||
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_fc_w
|
||||
ctx_size += n_layer*( 4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
|
||||
|
||||
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_proj_w
|
||||
ctx_size += n_layer*( n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
|
||||
|
||||
ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
|
||||
ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
|
||||
|
||||
ctx_size += (6 + 16*n_layer)*1024; // object overhead
|
||||
|
||||
printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
|
||||
}
|
||||
|
||||
// create the ggml context
|
||||
{
|
||||
struct ggml_init_params params = {
|
||||
/*.mem_size =*/ ctx_size,
|
||||
/*.mem_buffer =*/ NULL,
|
||||
/*.no_alloc =*/ false,
|
||||
};
|
||||
|
||||
model->ctx = ggml_init(params);
|
||||
if (!model->ctx) {
|
||||
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
// prepare memory for the weights
|
||||
{
|
||||
const auto & hparams = model->hparams;
|
||||
|
||||
const int n_embd = hparams.n_embd;
|
||||
const int n_layer = hparams.n_layer;
|
||||
const int n_vocab = hparams.n_vocab;
|
||||
|
||||
model->layers.resize(n_layer);
|
||||
|
||||
model->wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
||||
|
||||
model->ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
model->ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
|
||||
model->lmh_g = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
||||
//model->lmh_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
|
||||
|
||||
// map by name
|
||||
model->tensors["gpt_neox.embed_in.weight"] = model->wte;
|
||||
|
||||
model->tensors["gpt_neox.final_layer_norm.weight"] = model->ln_f_g;
|
||||
model->tensors["gpt_neox.final_layer_norm.bias"] = model->ln_f_b;
|
||||
|
||||
model->tensors["embed_out.weight"] = model->lmh_g;
|
||||
//model->tensors["lm_head.bias"] = model->lmh_b;
|
||||
|
||||
for (int i = 0; i < n_layer; ++i) {
|
||||
auto & layer = model->layers[i];
|
||||
|
||||
layer.ln_1_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
|
||||
layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 3*n_embd);
|
||||
layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
|
||||
|
||||
layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
|
||||
layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
|
||||
layer.ln_2_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
layer.ln_2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
|
||||
layer.c_mlp_fc_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 4*n_embd);
|
||||
layer.c_mlp_fc_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
|
||||
|
||||
layer.c_mlp_proj_w = ggml_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
|
||||
layer.c_mlp_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
|
||||
|
||||
// map by name
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.weight"] = layer.ln_1_g;
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.bias"] = layer.ln_1_b;
|
||||
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.weight"] = layer.c_attn_attn_w;
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.bias"] = layer.c_attn_attn_b;
|
||||
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.weight"] = layer.c_attn_proj_w;
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.bias"] = layer.c_attn_proj_b;
|
||||
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.weight"] = layer.ln_2_g;
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.bias"] = layer.ln_2_b;
|
||||
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.weight"] = layer.c_mlp_fc_w;
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.bias"] = layer.c_mlp_fc_b;
|
||||
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.weight"] = layer.c_mlp_proj_w;
|
||||
model->tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.bias"] = layer.c_mlp_proj_b;
|
||||
}
|
||||
}
|
||||
|
||||
// key + value memory
|
||||
{
|
||||
const auto & hparams = model->hparams;
|
||||
|
||||
const int n_embd = hparams.n_embd;
|
||||
const int n_layer = hparams.n_layer;
|
||||
const int n_ctx = hparams.n_ctx;
|
||||
|
||||
const int64_t n_mem = n_layer*n_ctx;
|
||||
const int64_t n_elements = n_embd*n_mem;
|
||||
|
||||
model->memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
|
||||
model->memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
|
||||
|
||||
const size_t memory_size = ggml_nbytes(model->memory_k) + ggml_nbytes(model->memory_v);
|
||||
|
||||
printf("%s: memory_size = %8.2f MB, n_mem = %" PRId64 "\n", __func__, memory_size/1024.0/1024.0, n_mem);
|
||||
}
|
||||
|
||||
// load weights
|
||||
{
|
||||
int n_tensors = 0;
|
||||
size_t total_size = 0;
|
||||
|
||||
printf("%s: ", __func__);
|
||||
|
||||
while (true) {
|
||||
int32_t n_dims;
|
||||
int32_t length;
|
||||
int32_t ttype;
|
||||
|
||||
fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
|
||||
fin.read(reinterpret_cast<char *>(&length), sizeof(length));
|
||||
fin.read(reinterpret_cast<char *>(&ttype), sizeof(ttype));
|
||||
|
||||
if (fin.eof()) {
|
||||
break;
|
||||
}
|
||||
|
||||
int32_t nelements = 1;
|
||||
int32_t ne[2] = { 1, 1 };
|
||||
for (int i = 0; i < n_dims; ++i) {
|
||||
fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
|
||||
nelements *= ne[i];
|
||||
}
|
||||
|
||||
std::string name(length, 0);
|
||||
fin.read(&name[0], length);
|
||||
|
||||
if (model->tensors.find(name.data()) == model->tensors.end()) {
|
||||
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
|
||||
return false;
|
||||
}
|
||||
|
||||
auto tensor = model->tensors[name.data()];
|
||||
if (ggml_nelements(tensor) != nelements) {
|
||||
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
|
||||
return false;
|
||||
}
|
||||
|
||||
if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
|
||||
fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%5d, %5d], expected [%5d, %5d]\n",
|
||||
__func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
|
||||
return false;
|
||||
}
|
||||
|
||||
// for debugging
|
||||
if (0) {
|
||||
printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
|
||||
}
|
||||
|
||||
const size_t bpe = ggml_type_size(ggml_type(ttype));
|
||||
|
||||
if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
|
||||
fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
|
||||
__func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
|
||||
return false;
|
||||
}
|
||||
|
||||
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
|
||||
|
||||
total_size += ggml_nbytes(tensor);
|
||||
if (++n_tensors % 8 == 0) {
|
||||
printf(".");
|
||||
fflush(stdout);
|
||||
}
|
||||
}
|
||||
|
||||
printf(" done\n");
|
||||
|
||||
printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
|
||||
}
|
||||
|
||||
fin.close();
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
std::stringstream GPTNEOXModel::predict(std::string prompt, int max_length, bool include_prompt) {
|
||||
|
||||
std::stringstream result;
|
||||
// tokenize the prompt
|
||||
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize((*vocab), prompt);
|
||||
|
||||
auto END_TOKEN_ID = vocab->token_to_id["<|endoftext|>"];
|
||||
|
||||
int n_past = 0;
|
||||
|
||||
int64_t t_sample_us = 0;
|
||||
int64_t t_predict_us = 0;
|
||||
|
||||
int n_predict = std::min(max_length, model->hparams.n_ctx - (int) embd_inp.size());
|
||||
|
||||
spdlog::debug("{}: number of tokens in prompt = {}", __func__, embd_inp.size());
|
||||
|
||||
std::vector<gpt_vocab::id> embd;
|
||||
|
||||
// determine the required inference memory per token:
|
||||
size_t mem_per_token = 0;
|
||||
|
||||
std::vector<float> logits;
|
||||
|
||||
gpt_neox_eval((*model), config.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
|
||||
|
||||
for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) {
|
||||
// predict
|
||||
if (embd.size() > 0) {
|
||||
const int64_t t_start_us = ggml_time_us();
|
||||
|
||||
if (!gpt_neox_eval((*model), config.n_threads, n_past, embd, logits, mem_per_token)) {
|
||||
throw std::runtime_error("Failed to predict");
|
||||
}
|
||||
|
||||
t_predict_us += ggml_time_us() - t_start_us;
|
||||
}
|
||||
|
||||
n_past += embd.size();
|
||||
embd.clear();
|
||||
|
||||
if (i >= embd_inp.size()) {
|
||||
// sample next token
|
||||
const int top_k = config.top_k;
|
||||
const float top_p = config.top_p;
|
||||
const float temp = config.temp;
|
||||
|
||||
const int n_vocab = model->hparams.n_vocab;
|
||||
|
||||
gpt_vocab::id id = 0;
|
||||
|
||||
{
|
||||
const int64_t t_start_sample_us = ggml_time_us();
|
||||
|
||||
id = gpt_sample_top_k_top_p((*vocab), logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
|
||||
|
||||
t_sample_us += ggml_time_us() - t_start_sample_us;
|
||||
}
|
||||
|
||||
// add it to the context
|
||||
embd.push_back(id);
|
||||
|
||||
// do not actually add endoftext char to the end string
|
||||
if(id != END_TOKEN_ID){
|
||||
result << vocab->id_to_token[id].c_str();
|
||||
}
|
||||
|
||||
} else {
|
||||
// if here, it means we are still processing the input prompt
|
||||
for (int k = i; k < embd_inp.size(); k++) {
|
||||
embd.push_back(embd_inp[k]);
|
||||
|
||||
if(include_prompt){
|
||||
result << vocab->id_to_token[embd_inp[k]].c_str();
|
||||
}
|
||||
|
||||
if (embd.size() > config.n_batch) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
i += embd.size() - 1;
|
||||
}
|
||||
|
||||
|
||||
// end of text token
|
||||
//if (embd.back() == 50256) {
|
||||
if(embd.back() == END_TOKEN_ID){
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
return result;
|
||||
}
|
||||
@ -11,6 +11,7 @@
|
||||
#include "turbopilot/model.hpp"
|
||||
#include "turbopilot/starcoder.hpp"
|
||||
#include "turbopilot/gptj.hpp"
|
||||
#include "turbopilot/gptneox.hpp"
|
||||
#include "turbopilot/server.hpp"
|
||||
|
||||
int main(int argc, char **argv)
|
||||
@ -23,7 +24,7 @@ int main(int argc, char **argv)
|
||||
.required();
|
||||
|
||||
program.add_argument("-m", "--model-type")
|
||||
.help("The type of model to load. Can be codegen,starcoder,wizardcoder")
|
||||
.help("The type of model to load. Can be codegen,starcoder,wizardcoder,stablecode")
|
||||
.default_value("codegen");
|
||||
|
||||
program.add_argument("-t", "--threads")
|
||||
@ -76,6 +77,9 @@ int main(int argc, char **argv)
|
||||
}else if(model_type.compare("starcoder") == 0 || model_type.compare("wizardcoder") == 0){
|
||||
spdlog::info("Initializing Starcoder/Wizardcoder type model for '{}' model type", model_type);
|
||||
model = new StarcoderModel(config, rng);
|
||||
}else if(model_type.compare("stablecode") == 0){
|
||||
spdlog::info("Initializing StableLM type model for '{}' model type", model_type);
|
||||
model = new GPTNEOXModel(config, rng);
|
||||
}else{
|
||||
spdlog::error("Invalid model type: {}", model_type);
|
||||
}
|
||||
|
||||
Loading…
Reference in New Issue
Block a user