initial release
This commit is contained in:
civ
@@ -0,0 +1,32 @@
|
||||
#
|
||||
# gpt-2
|
||||
|
||||
set(TEST_TARGET gpt-2-ctx)
|
||||
add_executable(${TEST_TARGET} main-ctx.cpp)
|
||||
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
|
||||
|
||||
set(TEST_TARGET gpt-2-alloc)
|
||||
add_executable(${TEST_TARGET} main-alloc.cpp)
|
||||
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
|
||||
|
||||
set(TEST_TARGET gpt-2-backend)
|
||||
add_executable(${TEST_TARGET} main-backend.cpp)
|
||||
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
|
||||
|
||||
set(TEST_TARGET gpt-2-sched)
|
||||
add_executable(${TEST_TARGET} main-sched.cpp)
|
||||
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
|
||||
|
||||
#
|
||||
# gpt-2-quantize
|
||||
|
||||
set(TEST_TARGET gpt-2-quantize)
|
||||
add_executable(${TEST_TARGET} quantize.cpp)
|
||||
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
|
||||
|
||||
#
|
||||
# gpt-2-batched
|
||||
|
||||
set(TEST_TARGET gpt-2-batched)
|
||||
add_executable(${TEST_TARGET} main-batched.cpp)
|
||||
target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
|
||||
@@ -0,0 +1,225 @@
|
||||
# gpt-2
|
||||
|
||||
This is a C++ example running GPT-2 inference using the [ggml](https://github.com/ggerganov/ggml) library.
|
||||
|
||||
The program runs on the CPU - no video card is required.
|
||||
|
||||
The [Cerebras-GPT](https://huggingface.co/cerebras) models are also supported.
|
||||
|
||||
The example supports the following GPT-2 models:
|
||||
|
||||
| Model | Description | Disk Size |
|
||||
| --- | --- | --- |
|
||||
| 117M | Small model | 240 MB |
|
||||
| 345M | Medium model | 680 MB |
|
||||
| 774M | Large model | 1.5 GB |
|
||||
| 1558M | XL model | 3.0 GB |
|
||||
|
||||
Sample performance on MacBook M1 Pro:
|
||||
|
||||
| Model | Size | Time / Token |
|
||||
| --- | --- | --- |
|
||||
| GPT-2 | 117M | 5 ms |
|
||||
| GPT-2 | 345M | 12 ms |
|
||||
| GPT-2 | 774M | 23 ms |
|
||||
| GPT-2 | 1558M | 42 ms |
|
||||
|
||||
*TODO: add tables for Cerebras-GPT models*
|
||||
|
||||
Sample output:
|
||||
|
||||
```bash
|
||||
$ ./bin/gpt-2 -h
|
||||
usage: ./bin/gpt-2 [options]
|
||||
|
||||
options:
|
||||
-h, --help show this help message and exit
|
||||
-s SEED, --seed SEED RNG seed (default: -1)
|
||||
-t N, --threads N number of threads to use during computation (default: 8)
|
||||
-p PROMPT, --prompt PROMPT
|
||||
prompt to start generation with (default: random)
|
||||
-n N, --n_predict N number of tokens to predict (default: 200)
|
||||
--top_k N top-k sampling (default: 40)
|
||||
--top_p N top-p sampling (default: 0.9)
|
||||
--temp N temperature (default: 1.0)
|
||||
-b N, --batch_size N batch size for prompt processing (default: 8)
|
||||
-m FNAME, --model FNAME
|
||||
model path (default: models/gpt-2-117M/ggml-model.bin)
|
||||
|
||||
$ ./bin/gpt-2
|
||||
gpt2_model_load: loading model from 'models/gpt-2-117M/ggml-model.bin'
|
||||
gpt2_model_load: n_vocab = 50257
|
||||
gpt2_model_load: n_ctx = 1024
|
||||
gpt2_model_load: n_embd = 768
|
||||
gpt2_model_load: n_head = 12
|
||||
gpt2_model_load: n_layer = 12
|
||||
gpt2_model_load: f16 = 1
|
||||
gpt2_model_load: ggml ctx size = 311.12 MB
|
||||
gpt2_model_load: memory size = 72.00 MB, n_mem = 12288
|
||||
gpt2_model_load: model size = 239.08 MB
|
||||
main: number of tokens in prompt = 1
|
||||
|
||||
So this is going to be the end of the line for us.
|
||||
|
||||
If the Dolphins continue to do their business, it's possible that the team could make a bid to bring in new defensive coordinator Scott Linehan.
|
||||
|
||||
Linehan's job is a little daunting, but he's a great coach and an excellent coach. I don't believe we're going to make the playoffs.
|
||||
|
||||
We're going to have to work hard to keep our heads down and get ready to go.<|endoftext|>
|
||||
|
||||
main: mem per token = 2048612 bytes
|
||||
main: load time = 106.32 ms
|
||||
main: sample time = 7.10 ms
|
||||
main: predict time = 506.40 ms / 5.06 ms per token
|
||||
main: total time = 629.84 ms
|
||||
```
|
||||
|
||||
## Downloading and converting the original models (GPT-2)
|
||||
|
||||
You can download the original model files using the [download-model.sh](download-model.sh) Bash script. The models are
|
||||
in Tensorflow format, so in order to use them with ggml, you need to convert them to appropriate format. This is done
|
||||
via the [convert-ckpt-to-ggml.py](convert-ckpt-to-ggml.py) python script.
|
||||
|
||||
Here is the entire process for the GPT-2 117M model (download from official site + conversion):
|
||||
|
||||
```bash
|
||||
cd ggml/build
|
||||
../examples/gpt-2/download-model.sh 117M
|
||||
|
||||
Downloading model 117M ...
|
||||
models/gpt-2-117M/checkpoint 100%[=============================>] 77 --.-KB/s in 0s
|
||||
models/gpt-2-117M/encoder.json 100%[=============================>] 1018K 1.20MB/s in 0.8s
|
||||
models/gpt-2-117M/hparams.json 100%[=============================>] 90 --.-KB/s in 0s
|
||||
models/gpt-2-117M/model.ckpt.data-00000-of-00001 100%[=============================>] 474.70M 1.21MB/s in 8m 39s
|
||||
models/gpt-2-117M/model.ckpt.index 100%[=============================>] 5.09K --.-KB/s in 0s
|
||||
models/gpt-2-117M/model.ckpt.meta 100%[=============================>] 460.11K 806KB/s in 0.6s
|
||||
models/gpt-2-117M/vocab.bpe 100%[=============================>] 445.62K 799KB/s in 0.6s
|
||||
Done! Model '117M' saved in 'models/gpt-2-117M/'
|
||||
|
||||
Run the convert-ckpt-to-ggml.py script to convert the model to ggml format.
|
||||
|
||||
python /Users/john/ggml/examples/gpt-2/convert-ckpt-to-ggml.py models/gpt-2-117M/ 1
|
||||
|
||||
```
|
||||
|
||||
This conversion requires that you have python and Tensorflow installed on your computer. Still, if you want to avoid
|
||||
this, you can download the already converted ggml models as described below.
|
||||
|
||||
## Downloading and converting the original models (Cerebras-GPT)
|
||||
|
||||
Clone the respective repository from here: https://huggingface.co/cerebras
|
||||
|
||||
Use the [convert-cerebras-to-ggml.py](convert-cerebras-to-ggml.py) script to convert the model to `ggml` format:
|
||||
|
||||
```bash
|
||||
cd ggml/build
|
||||
git clone https://huggingface.co/cerebras/Cerebras-GPT-111M models/
|
||||
python ../examples/gpt-2/convert-cerebras-to-ggml.py models/Cerebras-GPT-111M/
|
||||
|
||||
```
|
||||
|
||||
## Downloading the ggml model directly (GPT-2)
|
||||
|
||||
For convenience, I will be hosting the converted ggml model files in order to make it easier to run the examples. This
|
||||
way, you can directly download a single binary file and start using it. No python or Tensorflow is required.
|
||||
|
||||
Here is how to get the 117M ggml model:
|
||||
|
||||
```bash
|
||||
cd ggml/build
|
||||
../examples/gpt-2/download-ggml-model.sh 117M
|
||||
|
||||
Downloading ggml model 117M ...
|
||||
models/gpt-2-117M/ggml-model.bin 100%[===============================>] 239.58M 8.52MB/s in 28s
|
||||
Done! Model '117M' saved in 'models/gpt-2-117M/ggml-model.bin'
|
||||
You can now use it like this:
|
||||
|
||||
$ ./bin/gpt-2 -m models/gpt-2-117M/ggml-model.bin -p "This is an example"
|
||||
|
||||
```
|
||||
|
||||
At some point, I might decide to stop hosting these models. So in that case, simply revert to the manual process above.
|
||||
|
||||
## Quantizing the models
|
||||
|
||||
You can also try to quantize the `ggml` models via 4-bit integer quantization.
|
||||
Keep in mind that for smaller models, this will render them completely useless.
|
||||
You generally want to quantize larger models.
|
||||
|
||||
```bash
|
||||
# quantize GPT-2 F16 to Q4_0 (faster but less precise)
|
||||
./bin/gpt-2-quantize models/gpt-2-1558M/ggml-model-f16.bin models/gpt-2-1558M/ggml-model-q4_0.bin 2
|
||||
./bin/gpt-2 -m models/gpt-2-1558M/ggml-model-q4_0.bin -p "This is an example"
|
||||
|
||||
# quantize Cerebras F16 to Q4_1 (slower but more precise)
|
||||
./bin/gpt-2-quantize models/Cerebras-GPT-6.7B/ggml-model-f16.bin models/Cerebras-GPT-6.7B/ggml-model-q4_1.bin 3
|
||||
./bin/gpt-2 -m models/Cerebras-GPT-6.7B/ggml-model-q4_1.bin -p "This is an example"
|
||||
|
||||
```
|
||||
|
||||
## Batched generation example
|
||||
|
||||
You can try the batched generation from a given prompt using the gpt-2-batched binary.
|
||||
|
||||
Sample output:
|
||||
|
||||
```bash
|
||||
$ gpt-2-batched -np 5 -m models/gpt-2-117M/ggml-model.bin -p "Hello my name is" -n 50
|
||||
|
||||
main: seed = 1697037431
|
||||
gpt2_model_load: loading model from 'models/gpt-2-117M/ggml-model.bin'
|
||||
gpt2_model_load: n_vocab = 50257
|
||||
gpt2_model_load: n_ctx = 1024
|
||||
gpt2_model_load: n_embd = 768
|
||||
gpt2_model_load: n_head = 12
|
||||
gpt2_model_load: n_layer = 12
|
||||
gpt2_model_load: ftype = 1
|
||||
gpt2_model_load: qntvr = 0
|
||||
gpt2_model_load: ggml tensor size = 320 bytes
|
||||
gpt2_model_load: backend buffer size = 312.72 MB
|
||||
ggml_init_cublas: found 1 CUDA devices:
|
||||
Device 0: NVIDIA GeForce GTX 1660, compute capability 7.5
|
||||
gpt2_model_load: using CPU backend
|
||||
gpt2_model_load: memory size = 72.00 MB, n_mem = 12288
|
||||
gpt2_model_load: model size = 239.08 MB
|
||||
extract_tests_from_file : No test file found.
|
||||
test_gpt_tokenizer : 0 tests failed out of 0 tests.
|
||||
main: compute buffer size: 3.26 MB
|
||||
|
||||
|
||||
main: generating 5 sequences ...
|
||||
main: prompt: 'Hello my name is'
|
||||
main: number of tokens in prompt = 4, first 8 tokens: 15496 616 1438 318
|
||||
|
||||
|
||||
sequence 0:
|
||||
|
||||
Hello my name is John. You can call me any way you want, if you want, but for my very first date, I will be on the phone with you. We're both in our early 20s, but I feel like it's all
|
||||
|
||||
sequence 1:
|
||||
|
||||
Hello my name is Robert, and I want to say that we're proud to have your company here on the world's largest platform for sharing your stories with us. This is a huge opportunity for our community. We have hundreds of people on this team and
|
||||
|
||||
sequence 2:
|
||||
|
||||
Hello my name is Jack. I'm the one who created you.
|
||||
|
||||
Jack is a boy with a big smile and a big heart. He is a handsome guy. He loves the outdoors and loves the people he meets. He wants to be a
|
||||
|
||||
sequence 3:
|
||||
|
||||
Hello my name is John. I am a Canadian citizen with a large number of family in Quebec and I am interested in studying. My aim is to take up a post in the Journal of the International Academy of Sciences of Canada which I am currently finishing.
|
||||
|
||||
sequence 4:
|
||||
|
||||
Hello my name is Dan. I am an entrepreneur. I am a great father. I am a great husband. I am a great husband. I am a great dad. And I am a great husband.
|
||||
|
||||
I love my life. I love
|
||||
|
||||
|
||||
|
||||
main: load time = 880.80 ms
|
||||
main: sample time = 91.43 ms
|
||||
main: predict time = 2518.29 ms
|
||||
main: total time = 3544.32 ms
|
||||
```
|
||||
@@ -0,0 +1,183 @@
|
||||
# Convert Cerebras models to ggml format
|
||||
#
|
||||
# ref: https://www.cerebras.net/blog/cerebras-gpt-a-family-of-open-compute-efficient-large-language-models/
|
||||
#
|
||||
|
||||
import sys
|
||||
import struct
|
||||
import json
|
||||
import torch
|
||||
import numpy as np
|
||||
import re
|
||||
|
||||
from transformers import AutoModelForCausalLM
|
||||
|
||||
# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
|
||||
def bytes_to_unicode():
|
||||
"""
|
||||
Returns list of utf-8 byte and a corresponding list of unicode strings.
|
||||
The reversible bpe codes work on unicode strings.
|
||||
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
|
||||
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
|
||||
This is a signficant percentage of your normal, say, 32K bpe vocab.
|
||||
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
|
||||
And avoids mapping to whitespace/control characters the bpe code barfs on.
|
||||
"""
|
||||
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
|
||||
cs = bs[:]
|
||||
n = 0
|
||||
for b in range(2**8):
|
||||
if b not in bs:
|
||||
bs.append(b)
|
||||
cs.append(2**8+n)
|
||||
n += 1
|
||||
cs = [chr(n) for n in cs]
|
||||
return dict(zip(bs, cs))
|
||||
|
||||
if len(sys.argv) < 2:
|
||||
print("Usage: convert-cerebras-to-ggml.py dir-model [use-f32]\n")
|
||||
sys.exit(1)
|
||||
|
||||
# output in the same directory as the model
|
||||
dir_model = sys.argv[1]
|
||||
fname_out = sys.argv[1] + "/ggml-model-f16.bin"
|
||||
|
||||
with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
|
||||
encoder = json.load(f)
|
||||
|
||||
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
|
||||
hparams = json.load(f)
|
||||
|
||||
# use 16-bit or 32-bit floats
|
||||
use_f16 = True
|
||||
if len(sys.argv) > 2:
|
||||
use_f16 = False
|
||||
fname_out = sys.argv[1] + "/ggml-model-f32.bin"
|
||||
|
||||
model = AutoModelForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
|
||||
#print (model)
|
||||
|
||||
list_vars = model.state_dict()
|
||||
#print (list_vars)
|
||||
|
||||
print(hparams)
|
||||
|
||||
fout = open(fname_out, "wb")
|
||||
|
||||
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
|
||||
fout.write(struct.pack("i", hparams["vocab_size"]))
|
||||
fout.write(struct.pack("i", hparams["n_positions"]))
|
||||
fout.write(struct.pack("i", hparams["n_embd"]))
|
||||
fout.write(struct.pack("i", hparams["n_head"]))
|
||||
fout.write(struct.pack("i", hparams["n_layer"]))
|
||||
fout.write(struct.pack("i", use_f16))
|
||||
|
||||
byte_encoder = bytes_to_unicode()
|
||||
byte_decoder = {v:k for k, v in byte_encoder.items()}
|
||||
|
||||
fout.write(struct.pack("i", len(encoder)))
|
||||
|
||||
for key in encoder:
|
||||
text = bytearray([byte_decoder[c] for c in key])
|
||||
fout.write(struct.pack("i", len(text)))
|
||||
fout.write(text)
|
||||
|
||||
for name in list_vars.keys():
|
||||
data = list_vars[name].squeeze().numpy()
|
||||
print("Processing variable: " + name + " with shape: ", data.shape)
|
||||
|
||||
# rename headers to keep compatibility
|
||||
if name == "transformer.ln_f.weight":
|
||||
name = "model/ln_f/g"
|
||||
elif name == "transformer.ln_f.bias":
|
||||
name = "model/ln_f/b"
|
||||
elif name == "transformer.wte.weight":
|
||||
name = "model/wte"
|
||||
elif name == "transformer.wpe.weight":
|
||||
name = "model/wpe"
|
||||
elif name == "lm_head.weight":
|
||||
name = "model/lm_head"
|
||||
elif re.match(r"transformer.h\.\d+\.ln_1\.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_1/g"
|
||||
elif re.match(r"transformer.h\.\d+\.ln_1\.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_1/b"
|
||||
elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_attn/w"
|
||||
elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_attn/b"
|
||||
elif re.match(r"transformer.h\.\d+\.attn\.c_proj\.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_proj/w"
|
||||
elif re.match(r"transformer.h.\d+.attn.c_proj.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_proj/b"
|
||||
elif re.match(r"transformer.h.\d+.ln_2.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_2/g"
|
||||
elif re.match(r"transformer.h.\d+.ln_2.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_2/b"
|
||||
elif re.match(r"transformer.h.\d+.mlp.c_fc.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_fc/w"
|
||||
elif re.match(r"transformer.h.\d+.mlp.c_fc.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_fc/b"
|
||||
elif re.match(r"transformer.h.\d+.mlp.c_proj.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_proj/w"
|
||||
elif re.match(r"transformer.h.\d+.mlp.c_proj.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_proj/b"
|
||||
else:
|
||||
print("Unrecognized variable name. %s", name)
|
||||
|
||||
# we don't need these
|
||||
if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
|
||||
print(" Skipping variable: " + name)
|
||||
continue
|
||||
|
||||
n_dims = len(data.shape);
|
||||
|
||||
# ftype == 0 -> float32, ftype == 1 -> float16
|
||||
ftype = 0;
|
||||
if use_f16:
|
||||
if (name == "model/wte" or name == "model/lm_head" or name[-2:] == "/g" or name[-2:] == "/w") and n_dims == 2:
|
||||
print(" Converting to float16")
|
||||
data = data.astype(np.float16)
|
||||
ftype = 1
|
||||
else:
|
||||
print(" Converting to float32")
|
||||
data = data.astype(np.float32)
|
||||
ftype = 0
|
||||
|
||||
# for efficiency - transpose the projection matrices
|
||||
# "model/h.*/attn/c_attn/w"
|
||||
# "model/h.*/attn/c_proj/w"
|
||||
# "model/h.*/mlp/c_fc/w"
|
||||
# "model/h.*/mlp/c_proj/w"
|
||||
if name[-14:] == "/attn/c_attn/w" or \
|
||||
name[-14:] == "/attn/c_proj/w" or \
|
||||
name[-11:] == "/mlp/c_fc/w" or \
|
||||
name[-13:] == "/mlp/c_proj/w":
|
||||
print(" Transposing")
|
||||
data = data.transpose()
|
||||
|
||||
# header
|
||||
str = name.encode('utf-8')
|
||||
fout.write(struct.pack("iii", n_dims, len(str), ftype))
|
||||
for i in range(n_dims):
|
||||
fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
|
||||
fout.write(str);
|
||||
|
||||
# data
|
||||
data.tofile(fout)
|
||||
|
||||
fout.close()
|
||||
|
||||
print("Done. Output file: " + fname_out)
|
||||
print("")
|
||||
@@ -0,0 +1,159 @@
|
||||
# Convert a model checkpoint to a ggml compatible file
|
||||
#
|
||||
# Load the model using TensorFlow.
|
||||
# Iterate over all variables and write them to a binary file.
|
||||
#
|
||||
# For each variable, write the following:
|
||||
# - Number of dimensions (int)
|
||||
# - Name length (int)
|
||||
# - Dimensions (int[n_dims])
|
||||
# - Name (char[name_length])
|
||||
# - Data (float[n_dims])
|
||||
#
|
||||
# By default, the bigger matrices are converted to 16-bit floats.
|
||||
# This can be disabled by adding the "use-f32" CLI argument.
|
||||
#
|
||||
# At the start of the ggml file we write the model parameters
|
||||
# and vocabulary.
|
||||
#
|
||||
|
||||
import sys
|
||||
import json
|
||||
import struct
|
||||
import numpy as np
|
||||
import tensorflow as tf
|
||||
|
||||
# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
|
||||
def bytes_to_unicode():
|
||||
"""
|
||||
Returns list of utf-8 byte and a corresponding list of unicode strings.
|
||||
The reversible bpe codes work on unicode strings.
|
||||
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
|
||||
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
|
||||
This is a signficant percentage of your normal, say, 32K bpe vocab.
|
||||
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
|
||||
And avoids mapping to whitespace/control characters the bpe code barfs on.
|
||||
"""
|
||||
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
|
||||
cs = bs[:]
|
||||
n = 0
|
||||
for b in range(2**8):
|
||||
if b not in bs:
|
||||
bs.append(b)
|
||||
cs.append(2**8+n)
|
||||
n += 1
|
||||
cs = [chr(n) for n in cs]
|
||||
return dict(zip(bs, cs))
|
||||
|
||||
# helper method to convert a numpy array to different float types
|
||||
def convert_to_ftype(data, ftype):
|
||||
# fp16
|
||||
if ftype == 1:
|
||||
return data.astype(np.float16)
|
||||
|
||||
assert False, "Invalid ftype: " + str(ftype)
|
||||
|
||||
if len(sys.argv) < 3:
|
||||
print("Usage: convert-ckpt-to-ggml.py dir-model ftype\n")
|
||||
print(" ftype == 0 -> float32")
|
||||
print(" ftype == 1 -> float16")
|
||||
sys.exit(1)
|
||||
|
||||
# output in the same directory as the model
|
||||
dir_model = sys.argv[1]
|
||||
fname_out = sys.argv[1] + "/ggml-model.bin"
|
||||
|
||||
with open(dir_model + "/encoder.json", "r", encoding="utf-8") as f:
|
||||
encoder = json.load(f)
|
||||
|
||||
with open(dir_model + "/hparams.json", "r", encoding="utf-8") as f:
|
||||
hparams = json.load(f)
|
||||
|
||||
# possible data types
|
||||
# ftype == 0 -> float32
|
||||
# ftype == 1 -> float16
|
||||
#
|
||||
# map from ftype to string
|
||||
ftype_str = ["f32", "f16"]
|
||||
|
||||
ftype = 1
|
||||
if len(sys.argv) > 2:
|
||||
ftype = int(sys.argv[2])
|
||||
if ftype < 0 or ftype > 1:
|
||||
print("Invalid ftype: " + str(ftype))
|
||||
sys.exit(1)
|
||||
fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
|
||||
|
||||
list_vars = tf.train.list_variables(dir_model)
|
||||
|
||||
fout = open(fname_out, "wb")
|
||||
|
||||
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
|
||||
fout.write(struct.pack("i", hparams["n_vocab"]))
|
||||
fout.write(struct.pack("i", hparams["n_ctx"]))
|
||||
fout.write(struct.pack("i", hparams["n_embd"]))
|
||||
fout.write(struct.pack("i", hparams["n_head"]))
|
||||
fout.write(struct.pack("i", hparams["n_layer"]))
|
||||
fout.write(struct.pack("i", ftype))
|
||||
|
||||
byte_encoder = bytes_to_unicode()
|
||||
byte_decoder = {v:k for k, v in byte_encoder.items()}
|
||||
|
||||
fout.write(struct.pack("i", len(encoder)))
|
||||
|
||||
for key in encoder:
|
||||
text = bytearray([byte_decoder[c] for c in key])
|
||||
fout.write(struct.pack("i", len(text)))
|
||||
fout.write(text)
|
||||
|
||||
for name, shape in list_vars:
|
||||
print("Processing variable: " + name + " with shape: ", shape)
|
||||
|
||||
data = tf.train.load_variable(dir_model, name).squeeze()
|
||||
n_dims = len(data.shape);
|
||||
|
||||
# for efficiency - transpose the projection matrices
|
||||
# "model/h.*/attn/c_attn/w"
|
||||
# "model/h.*/attn/c_proj/w"
|
||||
# "model/h.*/mlp/c_fc/w"
|
||||
# "model/h.*/mlp/c_proj/w"
|
||||
if name[-14:] == "/attn/c_attn/w" or \
|
||||
name[-14:] == "/attn/c_proj/w" or \
|
||||
name[-11:] == "/mlp/c_fc/w" or \
|
||||
name[-13:] == "/mlp/c_proj/w":
|
||||
print(" Transposing")
|
||||
data = data.transpose()
|
||||
|
||||
dshape = data.shape
|
||||
|
||||
ftype_cur = 0
|
||||
if ftype != 0:
|
||||
# match name:
|
||||
# "model/wte"
|
||||
# "model/h.*/attn/c_attn/w"
|
||||
# "model/h.*/attn/c_proj/w"
|
||||
# "model/h.*/mlp/c_fc/w"
|
||||
# "model/h.*/mlp/c_proj/w"
|
||||
if name == "model/wte" or name[-2:] == "/w":
|
||||
print(" Converting to " + ftype_str[ftype])
|
||||
data = convert_to_ftype(data, ftype)
|
||||
ftype_cur = ftype
|
||||
else:
|
||||
print(" Converting to float32")
|
||||
data = data.astype(np.float32)
|
||||
ftype_cur = 0
|
||||
|
||||
# header
|
||||
str = name.encode('utf-8')
|
||||
fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
|
||||
for i in range(n_dims):
|
||||
fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
|
||||
fout.write(str);
|
||||
|
||||
# data
|
||||
data.tofile(fout)
|
||||
|
||||
fout.close()
|
||||
|
||||
print("Done. Output file: " + fname_out)
|
||||
print("")
|
||||
@@ -0,0 +1,195 @@
|
||||
# Convert GPT-2 h5 transformer model to ggml format
|
||||
#
|
||||
# Load the model using GPT2Model.
|
||||
# Iterate over all variables and write them to a binary file.
|
||||
#
|
||||
# For each variable, write the following:
|
||||
# - Number of dimensions (int)
|
||||
# - Name length (int)
|
||||
# - Dimensions (int[n_dims])
|
||||
# - Name (char[name_length])
|
||||
# - Data (float[n_dims])
|
||||
#
|
||||
# By default, the bigger matrices are converted to 16-bit floats.
|
||||
# This can be disabled by adding the "use-f32" CLI argument.
|
||||
#
|
||||
# At the start of the ggml file we write the model parameters
|
||||
# and vocabulary.
|
||||
#
|
||||
|
||||
import sys
|
||||
import struct
|
||||
import json
|
||||
import numpy as np
|
||||
import re
|
||||
|
||||
from transformers import GPT2Model
|
||||
|
||||
# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
|
||||
def bytes_to_unicode():
|
||||
"""
|
||||
Returns list of utf-8 byte and a corresponding list of unicode strings.
|
||||
The reversible bpe codes work on unicode strings.
|
||||
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
|
||||
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
|
||||
This is a signficant percentage of your normal, say, 32K bpe vocab.
|
||||
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
|
||||
And avoids mapping to whitespace/control characters the bpe code barfs on.
|
||||
"""
|
||||
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
|
||||
cs = bs[:]
|
||||
n = 0
|
||||
for b in range(2**8):
|
||||
if b not in bs:
|
||||
bs.append(b)
|
||||
cs.append(2**8+n)
|
||||
n += 1
|
||||
cs = [chr(n) for n in cs]
|
||||
return dict(zip(bs, cs))
|
||||
|
||||
if len(sys.argv) < 2:
|
||||
print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
|
||||
sys.exit(1)
|
||||
|
||||
# output in the same directory as the model
|
||||
dir_model = sys.argv[1]
|
||||
fname_out = sys.argv[1] + "/ggml-model.bin"
|
||||
|
||||
with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
|
||||
encoder = json.load(f)
|
||||
|
||||
with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
|
||||
encoder_added = json.load(f)
|
||||
|
||||
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
|
||||
hparams = json.load(f)
|
||||
|
||||
# use 16-bit or 32-bit floats
|
||||
use_f16 = True
|
||||
if len(sys.argv) > 2:
|
||||
use_f16 = False
|
||||
fname_out = sys.argv[1] + "/ggml-model-f32.bin"
|
||||
|
||||
model = GPT2Model.from_pretrained(dir_model, low_cpu_mem_usage=True)
|
||||
#print (model)
|
||||
|
||||
list_vars = model.state_dict()
|
||||
#print (list_vars)
|
||||
|
||||
fout = open(fname_out, "wb")
|
||||
|
||||
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
|
||||
fout.write(struct.pack("i", hparams["vocab_size"]))
|
||||
fout.write(struct.pack("i", hparams["n_positions"]))
|
||||
fout.write(struct.pack("i", hparams["n_embd"]))
|
||||
fout.write(struct.pack("i", hparams["n_head"]))
|
||||
fout.write(struct.pack("i", hparams["n_layer"]))
|
||||
#fout.write(struct.pack("i", hparams["rotary_dim"]))
|
||||
fout.write(struct.pack("i", use_f16))
|
||||
|
||||
byte_encoder = bytes_to_unicode()
|
||||
byte_decoder = {v:k for k, v in byte_encoder.items()}
|
||||
|
||||
fout.write(struct.pack("i", len(encoder) + len(encoder_added)))
|
||||
|
||||
for key in encoder:
|
||||
text = bytearray([byte_decoder[c] for c in key])
|
||||
fout.write(struct.pack("i", len(text)))
|
||||
fout.write(text)
|
||||
|
||||
for key in encoder_added:
|
||||
text = bytearray([byte_decoder[c] for c in key])
|
||||
fout.write(struct.pack("i", len(text)))
|
||||
fout.write(text)
|
||||
|
||||
for name in list_vars.keys():
|
||||
data = list_vars[name].squeeze().numpy()
|
||||
print("Processing variable: " + name + " with shape: ", data.shape)
|
||||
|
||||
# we don't need these
|
||||
if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
|
||||
print(" Skipping variable: " + name)
|
||||
continue
|
||||
|
||||
n_dims = len(data.shape);
|
||||
|
||||
# ftype == 0 -> float32, ftype == 1 -> float16
|
||||
ftype = 0;
|
||||
if use_f16:
|
||||
if name[-7:] == ".weight" and n_dims == 2:
|
||||
print(" Converting to float16")
|
||||
data = data.astype(np.float16)
|
||||
ftype = 1
|
||||
else:
|
||||
print(" Converting to float32")
|
||||
data = data.astype(np.float32)
|
||||
ftype = 0
|
||||
|
||||
# for efficiency - transpose these matrices:
|
||||
# "transformer.h.*.mlp.c_proj.weight
|
||||
if name.endswith(".mlp.c_proj.weight"):
|
||||
print(" Transposing")
|
||||
data = data.transpose()
|
||||
|
||||
# rename headers to keep compatibility
|
||||
if name == "ln_f.weight":
|
||||
name = "model/ln_f/g"
|
||||
elif name == "ln_f.bias":
|
||||
name = "model/ln_f/b"
|
||||
elif name == "wte.weight":
|
||||
name = "model/wte"
|
||||
elif name == "wpe.weight":
|
||||
name = "model/wpe"
|
||||
elif re.match(r"h\.\d+\.ln_1\.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_1/g"
|
||||
elif re.match(r"h\.\d+\.ln_1\.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_1/b"
|
||||
elif re.match(r"h\.\d+\.attn\.c_attn\.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_attn/w"
|
||||
elif re.match(r"h\.\d+\.attn\.c_attn\.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_attn/b"
|
||||
elif re.match(r"h\.\d+\.attn\.c_proj\.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_proj/w"
|
||||
elif re.match(r"h.\d+.attn.c_proj.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/attn/c_proj/b"
|
||||
elif re.match(r"h.\d+.ln_2.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_2/g"
|
||||
elif re.match(r"h.\d+.ln_2.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/ln_2/b"
|
||||
elif re.match(r"h.\d+.mlp.c_fc.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_fc/w"
|
||||
elif re.match(r"h.\d+.mlp.c_fc.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_fc/b"
|
||||
elif re.match(r"h.\d+.mlp.c_proj.weight", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_proj/w"
|
||||
elif re.match(r"h.\d+.mlp.c_proj.bias", name):
|
||||
i = re.findall("\d+", name)[0]
|
||||
name = f"model/h{i}/mlp/c_proj/b"
|
||||
else:
|
||||
print("Unrecognized variable name. %s", name)
|
||||
|
||||
str = name.encode('utf-8')
|
||||
|
||||
fout.write(struct.pack("iii", n_dims, len(str), ftype))
|
||||
for i in range(n_dims):
|
||||
fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
|
||||
fout.write(str);
|
||||
|
||||
# data
|
||||
data.tofile(fout)
|
||||
|
||||
fout.close()
|
||||
|
||||
print("Done. Output file: " + fname_out)
|
||||
print("")
|
||||
Executable
+69
@@ -0,0 +1,69 @@
|
||||
#!/bin/bash
|
||||
|
||||
# This script downloads GPT-2 model files that have already been converted to ggml format.
|
||||
# This way you don't have to convert them yourself.
|
||||
#
|
||||
# If you want to download the original GPT-2 model files, use the "download-model.sh" script instead.
|
||||
|
||||
#src="https://ggml.ggerganov.com"
|
||||
#pfx="ggml-model-gpt-2"
|
||||
|
||||
src="https://huggingface.co/ggerganov/ggml"
|
||||
pfx="resolve/main/ggml-model-gpt-2"
|
||||
|
||||
ggml_path=$(dirname $(realpath $0))
|
||||
|
||||
# GPT-2 models
|
||||
models=( "117M" "345M" "774M" "1558M" )
|
||||
|
||||
# list available models
|
||||
function list_models {
|
||||
printf "\n"
|
||||
printf " Available models:"
|
||||
for model in "${models[@]}"; do
|
||||
printf " $model"
|
||||
done
|
||||
printf "\n\n"
|
||||
}
|
||||
|
||||
if [ "$#" -ne 1 ]; then
|
||||
printf "Usage: $0 <model>\n"
|
||||
list_models
|
||||
|
||||
exit 1
|
||||
fi
|
||||
|
||||
model=$1
|
||||
|
||||
if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
|
||||
printf "Invalid model: $model\n"
|
||||
list_models
|
||||
|
||||
exit 1
|
||||
fi
|
||||
|
||||
# download ggml model
|
||||
|
||||
printf "Downloading ggml model $model ...\n"
|
||||
|
||||
mkdir -p models/gpt-2-$model
|
||||
|
||||
if [ -x "$(command -v wget)" ]; then
|
||||
wget --quiet --show-progress -O models/gpt-2-$model/ggml-model.bin $src/$pfx-$model.bin
|
||||
elif [ -x "$(command -v curl)" ]; then
|
||||
curl -L --output models/gpt-2-$model/ggml-model.bin $src/$pfx-$model.bin
|
||||
else
|
||||
printf "Either wget or curl is required to download models.\n"
|
||||
exit 1
|
||||
fi
|
||||
|
||||
if [ $? -ne 0 ]; then
|
||||
printf "Failed to download ggml model $model \n"
|
||||
printf "Please try again later or download the original GPT-2 model files and convert them yourself.\n"
|
||||
exit 1
|
||||
fi
|
||||
|
||||
printf "Done! Model '$model' saved in 'models/gpt-2-$model/ggml-model.bin'\n"
|
||||
printf "You can now use it like this:\n\n"
|
||||
printf " $ ./bin/gpt-2 -m models/gpt-2-$model/ggml-model.bin -p \"This is an example\"\n"
|
||||
printf "\n"
|
||||
Executable
+48
@@ -0,0 +1,48 @@
|
||||
#!/bin/bash
|
||||
|
||||
ggml_path=$(dirname $(realpath $0))
|
||||
|
||||
# GPT-2 models
|
||||
models=( "117M" "345M" "774M" "1558M" )
|
||||
|
||||
# list available models
|
||||
function list_models {
|
||||
printf "\n"
|
||||
printf " Available models:"
|
||||
for model in "${models[@]}"; do
|
||||
printf " $model"
|
||||
done
|
||||
printf "\n\n"
|
||||
}
|
||||
|
||||
if [ "$#" -ne 1 ]; then
|
||||
printf "Usage: $0 <model>\n"
|
||||
list_models
|
||||
|
||||
exit 1
|
||||
fi
|
||||
|
||||
model=$1
|
||||
|
||||
if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
|
||||
printf "Invalid model: $model\n"
|
||||
list_models
|
||||
|
||||
exit 1
|
||||
fi
|
||||
|
||||
# download model
|
||||
|
||||
printf "Downloading model $model ...\n"
|
||||
|
||||
mkdir -p models/gpt-2-$model
|
||||
|
||||
for file in checkpoint encoder.json hparams.json model.ckpt.data-00000-of-00001 model.ckpt.index model.ckpt.meta vocab.bpe; do
|
||||
wget --quiet --show-progress -O models/gpt-2-$model/$file https://openaipublic.blob.core.windows.net/gpt-2/models/$model/$file
|
||||
done
|
||||
|
||||
printf "Done! Model '$model' saved in 'models/gpt-2-$model/'\n\n"
|
||||
printf "Run the convert-ckpt-to-ggml.py script to convert the model to ggml format.\n"
|
||||
printf "\n"
|
||||
printf " python $ggml_path/convert-ckpt-to-ggml.py models/gpt-2-$model/\n"
|
||||
printf "\n"
|
||||
@@ -0,0 +1,880 @@
|
||||
#include "ggml.h"
|
||||
#include "ggml-cpu.h"
|
||||
#include "ggml-alloc.h"
|
||||
#include "ggml-backend.h"
|
||||
|
||||
#include "common.h"
|
||||
#include "common-ggml.h"
|
||||
|
||||
#include <cassert>
|
||||
#include <cmath>
|
||||
#include <cstdio>
|
||||
#include <cstring>
|
||||
#include <fstream>
|
||||
#include <map>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
#if defined(_MSC_VER)
|
||||
#pragma warning(disable: 4244 4267) // possible loss of data
|
||||
#endif
|
||||
|
||||
// default hparams (GPT-2 117M)
|
||||
struct gpt2_hparams {
|
||||
int32_t n_vocab = 50257;
|
||||
int32_t n_ctx = 1024;
|
||||
int32_t n_embd = 768;
|
||||
int32_t n_head = 12;
|
||||
int32_t n_layer = 12;
|
||||
int32_t ftype = 1;
|
||||
float eps = 1e-5f;
|
||||
};
|
||||
|
||||
struct gpt2_layer {
|
||||
// normalization
|
||||
struct ggml_tensor * ln_1_g;
|
||||
struct ggml_tensor * ln_1_b;
|
||||
|
||||
struct ggml_tensor * ln_2_g;
|
||||
struct ggml_tensor * ln_2_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;
|
||||
|
||||
// mlp
|
||||
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 gpt2_model {
|
||||
gpt2_hparams hparams;
|
||||
|
||||
// normalization
|
||||
struct ggml_tensor * ln_f_g;
|
||||
struct ggml_tensor * ln_f_b;
|
||||
|
||||
struct ggml_tensor * wte; // token embedding
|
||||
struct ggml_tensor * wpe; // position embedding
|
||||
struct ggml_tensor * lm_head; // language model head
|
||||
|
||||
std::vector<gpt2_layer> layers;
|
||||
|
||||
// key + value memory
|
||||
struct ggml_tensor * memory_k;
|
||||
struct ggml_tensor * memory_v;
|
||||
|
||||
//
|
||||
struct ggml_context * ctx_w;
|
||||
std::map<std::string, struct ggml_tensor *> tensors;
|
||||
};
|
||||
|
||||
// load the model's weights from a file
|
||||
bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
|
||||
printf("%s: loading model from '%s'\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.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: ftype = %d\n", __func__, hparams.ftype);
|
||||
printf("%s: qntvr = %d\n", __func__, qntvr);
|
||||
|
||||
hparams.ftype %= GGML_QNT_VERSION_FACTOR;
|
||||
}
|
||||
|
||||
// load vocab
|
||||
{
|
||||
int32_t n_vocab = 0;
|
||||
fin.read((char *) &n_vocab, sizeof(n_vocab));
|
||||
|
||||
if (n_vocab != model.hparams.n_vocab) {
|
||||
fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
|
||||
__func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
|
||||
return false;
|
||||
}
|
||||
|
||||
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_w;
|
||||
|
||||
size_t ctx_size = 0;
|
||||
|
||||
{
|
||||
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_vocab = hparams.n_vocab;
|
||||
|
||||
ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
|
||||
ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
|
||||
|
||||
ctx_size += ggml_row_size(wtype, n_vocab*n_embd); // wte
|
||||
ctx_size += ggml_row_size(GGML_TYPE_F32 , n_ctx*n_embd); // wpe
|
||||
ctx_size += ggml_row_size(wtype, n_vocab*n_embd); // lm_head
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_g
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_b
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(wtype, 3*n_embd*n_embd)); // c_attn_attn_w
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 3*n_embd)); // c_attn_attn_b
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(wtype, n_embd*n_embd)); // c_attn_proj_w
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // c_attn_proj_b
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_fc_w
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd)); // c_mlp_fc_b
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_proj_w
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd)); // c_mlp_proj_b
|
||||
|
||||
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_k
|
||||
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_v
|
||||
|
||||
ctx_size += (6 + 12*n_layer)*512; // object overhead
|
||||
|
||||
printf("%s: ggml tensor size = %d bytes\n", __func__, (int) sizeof(ggml_tensor));
|
||||
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_w = ggml_init(params);
|
||||
if (!model.ctx_w) {
|
||||
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_ctx = hparams.n_ctx;
|
||||
const int n_vocab = hparams.n_vocab;
|
||||
|
||||
model.layers.resize(n_layer);
|
||||
|
||||
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.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
||||
model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
|
||||
model.lm_head = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
||||
|
||||
// map by name
|
||||
model.tensors["model/ln_f/g"] = model.ln_f_g;
|
||||
model.tensors["model/ln_f/b"] = model.ln_f_b;
|
||||
|
||||
model.tensors["model/wte"] = model.wte;
|
||||
model.tensors["model/wpe"] = model.wpe;
|
||||
model.tensors["model/lm_head"] = model.lm_head;
|
||||
|
||||
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.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_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.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["model/h" + std::to_string(i) + "/ln_1/g"] = layer.ln_1_g;
|
||||
model.tensors["model/h" + std::to_string(i) + "/ln_1/b"] = layer.ln_1_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/ln_2/g"] = layer.ln_2_g;
|
||||
model.tensors["model/h" + std::to_string(i) + "/ln_2/b"] = layer.ln_2_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"] = layer.c_mlp_fc_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"] = layer.c_mlp_fc_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"] = layer.c_mlp_proj_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"] = 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 int n_mem = n_layer*n_ctx;
|
||||
const int n_elements = n_embd*n_mem;
|
||||
|
||||
model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
|
||||
model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 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 = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
|
||||
}
|
||||
|
||||
// load weights
|
||||
{
|
||||
size_t total_size = 0;
|
||||
|
||||
bool has_lm_head = false;
|
||||
|
||||
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) == model.tensors.end()) {
|
||||
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
|
||||
return false;
|
||||
}
|
||||
|
||||
auto tensor = model.tensors[name];
|
||||
if (ggml_nelements(tensor) != nelements) {
|
||||
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
|
||||
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 [%d, %d], expected [%d, %d]\n",
|
||||
__func__, name.c_str(), (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.c_str(), 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.c_str(), ggml_nbytes(tensor), nelements*bpe);
|
||||
return false;
|
||||
}
|
||||
|
||||
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
|
||||
|
||||
// GPT-2 models share the WTE tensor as the LM head
|
||||
if (name == "model/wte" && has_lm_head == false) {
|
||||
memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
|
||||
}
|
||||
|
||||
if (name == "model/lm_head") {
|
||||
has_lm_head = true;
|
||||
}
|
||||
|
||||
total_size += ggml_nbytes(tensor);
|
||||
}
|
||||
|
||||
printf("%s: model size = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
|
||||
}
|
||||
|
||||
fin.close();
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
// build the computation graph
|
||||
struct ggml_cgraph * gpt2_graph(
|
||||
const gpt2_model & model,
|
||||
const int n_past,
|
||||
const int n_tokens) {
|
||||
const int N = n_tokens;
|
||||
|
||||
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;
|
||||
|
||||
// since we are using ggml-alloc, this buffer only needs enough space to hold the ggml_tensor and ggml_cgraph structs, but not the tensor data
|
||||
static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
|
||||
static std::vector<uint8_t> buf(buf_size);
|
||||
|
||||
struct ggml_init_params params = {
|
||||
/*.mem_size =*/ buf_size,
|
||||
/*.mem_buffer =*/ buf.data(),
|
||||
/*.no_alloc =*/ true, // the tensors will be allocated later by ggml_gallocr_alloc_graph()
|
||||
};
|
||||
|
||||
struct ggml_context * ctx = ggml_init(params);
|
||||
|
||||
struct ggml_cgraph * gf = ggml_new_graph(ctx);
|
||||
|
||||
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
|
||||
// at this point, the tensor data is not allocated yet and cannot be set
|
||||
// we will find the tensor after the graph is allocated by its name, and set the data then
|
||||
ggml_set_name(embd, "embd");
|
||||
// setting a tensor as an input will ensure that it is allocated at the beginning of the graph
|
||||
// this is important to ensure that the input tensors are not overwritten before they are used
|
||||
ggml_set_input(embd);
|
||||
|
||||
struct ggml_tensor * position = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
|
||||
ggml_set_name(position, "position");
|
||||
ggml_set_input(position);
|
||||
|
||||
// wte + wpe
|
||||
struct ggml_tensor * inpL =
|
||||
ggml_add(ctx,
|
||||
ggml_get_rows(ctx, model.wte, embd),
|
||||
ggml_get_rows(ctx, model.wpe, position));
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
struct ggml_tensor * cur;
|
||||
|
||||
// norm
|
||||
{
|
||||
// [ 768, N]
|
||||
cur = ggml_norm(ctx, inpL, hparams.eps);
|
||||
|
||||
// cur = ln_1_g*cur + ln_1_b
|
||||
// [ 768, N]
|
||||
cur = ggml_add(ctx,
|
||||
ggml_mul(ctx,
|
||||
ggml_repeat(ctx, model.layers[il].ln_1_g, cur),
|
||||
cur),
|
||||
ggml_repeat(ctx, model.layers[il].ln_1_b, cur));
|
||||
}
|
||||
|
||||
// attn
|
||||
// [2304, 768] - model.layers[il].c_attn_attn_w
|
||||
// [2304, 1] - model.layers[il].c_attn_attn_b
|
||||
// [ 768, N] - cur (in)
|
||||
// [2304, N] - cur (out)
|
||||
//
|
||||
// cur = attn_w*cur + attn_b
|
||||
// [2304, N]
|
||||
{
|
||||
cur = ggml_mul_mat(ctx,
|
||||
model.layers[il].c_attn_attn_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx,
|
||||
ggml_repeat(ctx, model.layers[il].c_attn_attn_b, cur),
|
||||
cur);
|
||||
}
|
||||
|
||||
// self-attention
|
||||
{
|
||||
struct ggml_tensor * Qcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
|
||||
struct ggml_tensor * Kcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
|
||||
struct ggml_tensor * Vcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
|
||||
|
||||
// store key and value to memory
|
||||
if (N >= 1) {
|
||||
struct ggml_tensor * k = ggml_view_1d(ctx, 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_1d(ctx, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
|
||||
|
||||
ggml_build_forward_expand(gf, ggml_cpy(ctx, Kcur, k));
|
||||
ggml_build_forward_expand(gf, ggml_cpy(ctx, Vcur, v));
|
||||
}
|
||||
|
||||
// Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
|
||||
// [64, N, 12]
|
||||
struct ggml_tensor * Q =
|
||||
ggml_permute(ctx,
|
||||
ggml_cont_3d(ctx, Qcur, n_embd/n_head, n_head, N),
|
||||
0, 2, 1, 3);
|
||||
|
||||
// K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
|
||||
// [64, n_past + N, 12]
|
||||
struct ggml_tensor * K =
|
||||
ggml_permute(ctx,
|
||||
ggml_reshape_3d(ctx,
|
||||
ggml_view_1d(ctx, 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);
|
||||
|
||||
// GG: flash attention
|
||||
//struct ggml_tensor * V =
|
||||
// ggml_cpy(ctx0,
|
||||
// ggml_permute(ctx0,
|
||||
// ggml_reshape_3d(ctx0,
|
||||
// ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
|
||||
// n_embd/n_head, n_head, n_past + N),
|
||||
// 1, 2, 0, 3),
|
||||
// ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
|
||||
|
||||
//struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
|
||||
|
||||
// K * Q
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ = ggml_mul_mat(ctx, K, Q);
|
||||
|
||||
// KQ_scaled = KQ / sqrt(n_embd/n_head)
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ_scaled =
|
||||
ggml_scale(ctx,
|
||||
KQ,
|
||||
1.0f/sqrtf(float(n_embd)/n_head));
|
||||
|
||||
// KQ_masked = mask_past(KQ_scaled)
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx, KQ_scaled, n_past);
|
||||
|
||||
// KQ = soft_max(KQ_masked)
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx, KQ_masked);
|
||||
|
||||
// V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
|
||||
// [n_past + N, 64, 12]
|
||||
struct ggml_tensor * V_trans =
|
||||
ggml_cont_3d(ctx,
|
||||
ggml_permute(ctx,
|
||||
ggml_reshape_3d(ctx,
|
||||
ggml_view_1d(ctx, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
|
||||
n_embd/n_head, n_head, n_past + N),
|
||||
1, 2, 0, 3),
|
||||
n_past + N, n_embd/n_head, n_head);
|
||||
|
||||
// KQV = transpose(V) * KQ_soft_max
|
||||
// [64, N, 12]
|
||||
struct ggml_tensor * KQV = ggml_mul_mat(ctx, V_trans, KQ_soft_max);
|
||||
|
||||
// KQV_merged = KQV.permute(0, 2, 1, 3)
|
||||
// [64, 12, N]
|
||||
struct ggml_tensor * KQV_merged = ggml_permute(ctx, KQV, 0, 2, 1, 3);
|
||||
|
||||
// cur = KQV_merged.contiguous().view(n_embd, N)
|
||||
// [768, N]
|
||||
cur = ggml_cont_2d(ctx, KQV_merged, n_embd, N);
|
||||
}
|
||||
|
||||
// projection
|
||||
// [ 768, 768] - model.layers[il].c_attn_proj_w
|
||||
// [ 768, 1] - model.layers[il].c_attn_proj_b
|
||||
// [ 768, N] - cur (in)
|
||||
// [ 768, N] - cur (out)
|
||||
//
|
||||
// cur = proj_w*cur + proj_b
|
||||
// [768, N]
|
||||
{
|
||||
cur = ggml_mul_mat(ctx,
|
||||
model.layers[il].c_attn_proj_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx,
|
||||
ggml_repeat(ctx, model.layers[il].c_attn_proj_b, cur),
|
||||
cur);
|
||||
}
|
||||
|
||||
// add the input
|
||||
cur = ggml_add(ctx, cur, inpL);
|
||||
|
||||
struct ggml_tensor * inpFF = cur;
|
||||
|
||||
// feed-forward network
|
||||
{
|
||||
// norm
|
||||
{
|
||||
cur = ggml_norm(ctx, inpFF, hparams.eps);
|
||||
|
||||
// cur = ln_2_g*cur + ln_2_b
|
||||
// [ 768, N]
|
||||
cur = ggml_add(ctx,
|
||||
ggml_mul(ctx,
|
||||
ggml_repeat(ctx, model.layers[il].ln_2_g, cur),
|
||||
cur),
|
||||
ggml_repeat(ctx, model.layers[il].ln_2_b, cur));
|
||||
}
|
||||
|
||||
// fully connected
|
||||
// [3072, 768] - model.layers[il].c_mlp_fc_w
|
||||
// [3072, 1] - model.layers[il].c_mlp_fc_b
|
||||
// [ 768, N] - cur (in)
|
||||
// [3072, N] - cur (out)
|
||||
//
|
||||
// cur = fc_w*cur + fc_b
|
||||
// [3072, N]
|
||||
cur = ggml_mul_mat(ctx,
|
||||
model.layers[il].c_mlp_fc_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx,
|
||||
ggml_repeat(ctx, model.layers[il].c_mlp_fc_b, cur),
|
||||
cur);
|
||||
|
||||
// GELU activation
|
||||
// [3072, N]
|
||||
cur = ggml_gelu(ctx, cur);
|
||||
|
||||
// projection
|
||||
// [ 768, 3072] - model.layers[il].c_mlp_proj_w
|
||||
// [ 768, 1] - model.layers[il].c_mlp_proj_b
|
||||
// [3072, N] - cur (in)
|
||||
// [ 768, N] - cur (out)
|
||||
//
|
||||
// cur = proj_w*cur + proj_b
|
||||
// [768, N]
|
||||
cur = ggml_mul_mat(ctx,
|
||||
model.layers[il].c_mlp_proj_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx,
|
||||
ggml_repeat(ctx, model.layers[il].c_mlp_proj_b, cur),
|
||||
cur);
|
||||
}
|
||||
|
||||
// input for next layer
|
||||
inpL = ggml_add(ctx, cur, inpFF);
|
||||
}
|
||||
|
||||
// norm
|
||||
{
|
||||
// [ 768, N]
|
||||
inpL = ggml_norm(ctx, inpL, hparams.eps);
|
||||
|
||||
// inpL = ln_f_g*inpL + ln_f_b
|
||||
// [ 768, N]
|
||||
inpL = ggml_add(ctx,
|
||||
ggml_mul(ctx,
|
||||
ggml_repeat(ctx, model.ln_f_g, inpL),
|
||||
inpL),
|
||||
ggml_repeat(ctx, model.ln_f_b, inpL));
|
||||
}
|
||||
|
||||
// inpL = WTE * inpL
|
||||
// [ 768, 50257] - model.lm_head
|
||||
// [ 768, N] - inpL
|
||||
inpL = ggml_mul_mat(ctx, model.lm_head, inpL);
|
||||
ggml_set_name(inpL, "logits");
|
||||
// setting a tensor as the output will ensure that it is not overwritten by subsequent operations
|
||||
ggml_set_output(inpL);
|
||||
|
||||
// logits -> probs
|
||||
//inpL = ggml_soft_max(ctx0, inpL);
|
||||
|
||||
ggml_build_forward_expand(gf, inpL);
|
||||
|
||||
ggml_free(ctx);
|
||||
|
||||
return gf;
|
||||
}
|
||||
|
||||
// evaluate the transformer
|
||||
//
|
||||
// - model: the model
|
||||
// - allocr: ggml_gallocr to use to allocate the compute buffer
|
||||
// - 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 gpt2_eval(
|
||||
const gpt2_model & model,
|
||||
ggml_gallocr_t allocr,
|
||||
const int n_threads,
|
||||
const int n_past,
|
||||
const std::vector<gpt_vocab::id> & embd_inp,
|
||||
std::vector<float> & embd_w) {
|
||||
const int N = embd_inp.size();
|
||||
|
||||
const auto & hparams = model.hparams;
|
||||
|
||||
const int n_vocab = hparams.n_vocab;
|
||||
|
||||
struct ggml_cgraph * gf = gpt2_graph(model, n_past, embd_inp.size());
|
||||
|
||||
// allocate the graph tensors
|
||||
ggml_gallocr_alloc_graph(allocr, gf);
|
||||
|
||||
// set the graph inputs
|
||||
struct ggml_tensor * embd = ggml_graph_get_tensor(gf, "embd");
|
||||
memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
|
||||
|
||||
struct ggml_tensor * position = ggml_graph_get_tensor(gf, "position");
|
||||
for (int i = 0; i < N; ++i) {
|
||||
((int32_t *) position->data)[i] = n_past + i;
|
||||
}
|
||||
|
||||
// run the computation
|
||||
struct ggml_cplan plan = ggml_graph_plan(gf, n_threads, nullptr);
|
||||
static std::vector<uint8_t> work_buffer;
|
||||
work_buffer.resize(plan.work_size);
|
||||
plan.work_data = work_buffer.data();
|
||||
ggml_graph_compute(gf, &plan);
|
||||
|
||||
//if (n_past%100 == 0) {
|
||||
// ggml_graph_print (&gf);
|
||||
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
|
||||
//}
|
||||
|
||||
// get the graph outputs
|
||||
struct ggml_tensor * logits = ggml_graph_get_tensor(gf, "logits");
|
||||
|
||||
//embd_w.resize(n_vocab*N);
|
||||
//memcpy(embd_w.data(), ggml_get_data(logits), sizeof(float)*n_vocab*N);
|
||||
|
||||
// return result just for the last token
|
||||
embd_w.resize(n_vocab);
|
||||
memcpy(embd_w.data(), (float *) ggml_get_data(logits) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
int main(int argc, char ** argv) {
|
||||
ggml_time_init();
|
||||
|
||||
const int64_t t_main_start_us = ggml_time_us();
|
||||
|
||||
gpt_params params;
|
||||
params.model = "models/gpt-2-117M/ggml-model.bin";
|
||||
|
||||
if (gpt_params_parse(argc, argv, params) == false) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
if (params.seed < 0) {
|
||||
params.seed = time(NULL);
|
||||
}
|
||||
|
||||
printf("%s: seed = %d\n", __func__, params.seed);
|
||||
|
||||
std::mt19937 rng(params.seed);
|
||||
if (params.prompt.empty()) {
|
||||
params.prompt = gpt_random_prompt(rng);
|
||||
}
|
||||
|
||||
int64_t t_load_us = 0;
|
||||
|
||||
gpt_vocab vocab;
|
||||
gpt2_model model;
|
||||
|
||||
// load the model
|
||||
{
|
||||
const int64_t t_start_us = ggml_time_us();
|
||||
|
||||
if (!gpt2_model_load(params.model, model, vocab)) {
|
||||
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
|
||||
return 1;
|
||||
}
|
||||
|
||||
t_load_us = ggml_time_us() - t_start_us;
|
||||
|
||||
test_gpt_tokenizer(vocab, params.token_test);
|
||||
}
|
||||
|
||||
ggml_gallocr_t allocr = NULL;
|
||||
// allocate the compute buffer
|
||||
{
|
||||
allocr = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
|
||||
|
||||
// create the worst case graph for memory usage estimation
|
||||
int n_tokens = std::min(model.hparams.n_ctx, params.n_batch);
|
||||
int n_past = model.hparams.n_ctx - n_tokens;
|
||||
struct ggml_cgraph * gf = gpt2_graph(model, n_past, n_tokens);
|
||||
|
||||
// pre-allocate the compute buffer for the worst case (optional)
|
||||
ggml_gallocr_reserve(allocr, gf);
|
||||
size_t mem_size = ggml_gallocr_get_buffer_size(allocr, 0);
|
||||
fprintf(stderr, "%s: compute buffer size: %.2f MB\n", __func__, mem_size/1024.0/1024.0);
|
||||
}
|
||||
|
||||
int n_past = 0;
|
||||
|
||||
int64_t t_sample_us = 0;
|
||||
int64_t t_predict_us = 0;
|
||||
|
||||
std::vector<float> logits;
|
||||
|
||||
// tokenize the prompt
|
||||
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
|
||||
|
||||
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
|
||||
|
||||
printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
|
||||
printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size());
|
||||
for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) {
|
||||
printf("%d ", embd_inp[i]);
|
||||
}
|
||||
printf("\n\n");
|
||||
|
||||
// submit the input prompt token-by-token
|
||||
// this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
|
||||
std::vector<gpt_vocab::id> embd;
|
||||
|
||||
for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
|
||||
// predict
|
||||
if (embd.size() > 0) {
|
||||
const int64_t t_start_us = ggml_time_us();
|
||||
|
||||
if (!gpt2_eval(model, allocr, params.n_threads, n_past, embd, logits)) {
|
||||
printf("Failed to predict\n");
|
||||
return 1;
|
||||
}
|
||||
|
||||
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 = params.top_k;
|
||||
const float top_p = params.top_p;
|
||||
const float temp = params.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);
|
||||
} else {
|
||||
// if here, it means we are still processing the input prompt
|
||||
for (size_t k = i; k < embd_inp.size(); k++) {
|
||||
embd.push_back(embd_inp[k]);
|
||||
if (int32_t(embd.size()) >= params.n_batch) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
i += embd.size() - 1;
|
||||
}
|
||||
|
||||
// display text
|
||||
for (auto id : embd) {
|
||||
printf("%s", vocab.id_to_token[id].c_str());
|
||||
}
|
||||
fflush(stdout);
|
||||
|
||||
// end of text token
|
||||
if (embd.back() == 50256) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
// report timing
|
||||
{
|
||||
const int64_t t_main_end_us = ggml_time_us();
|
||||
|
||||
printf("\n\n");
|
||||
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
|
||||
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
|
||||
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
|
||||
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
|
||||
}
|
||||
|
||||
ggml_free(model.ctx_w);
|
||||
|
||||
return 0;
|
||||
}
|
||||
@@ -0,0 +1,946 @@
|
||||
#include "ggml.h"
|
||||
#include "ggml-cpu.h"
|
||||
#include "ggml-alloc.h"
|
||||
#include "ggml-backend.h"
|
||||
|
||||
#ifdef GGML_USE_CUDA
|
||||
#include "ggml-cuda.h"
|
||||
#endif
|
||||
|
||||
#ifdef GGML_USE_METAL
|
||||
#include "ggml-metal.h"
|
||||
#endif
|
||||
|
||||
#include "common.h"
|
||||
#include "common-ggml.h"
|
||||
|
||||
#include <cassert>
|
||||
#include <cmath>
|
||||
#include <cstdio>
|
||||
#include <cstring>
|
||||
#include <fstream>
|
||||
#include <map>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
#if defined(_MSC_VER)
|
||||
#pragma warning(disable: 4244 4267) // possible loss of data
|
||||
#endif
|
||||
|
||||
#define GPT2_MAX_NODES 4096
|
||||
|
||||
static void ggml_log_callback_default(ggml_log_level level, const char * text, void * user_data) {
|
||||
(void) level;
|
||||
(void) user_data;
|
||||
fputs(text, stderr);
|
||||
fflush(stderr);
|
||||
}
|
||||
|
||||
// default hparams (GPT-2 117M)
|
||||
struct gpt2_hparams {
|
||||
int32_t n_vocab = 50257;
|
||||
int32_t n_ctx = 1024;
|
||||
int32_t n_embd = 768;
|
||||
int32_t n_head = 12;
|
||||
int32_t n_layer = 12;
|
||||
int32_t ftype = 1;
|
||||
float eps = 1e-5f;
|
||||
};
|
||||
|
||||
struct gpt2_layer {
|
||||
// normalization
|
||||
struct ggml_tensor * ln_1_g;
|
||||
struct ggml_tensor * ln_1_b;
|
||||
|
||||
struct ggml_tensor * ln_2_g;
|
||||
struct ggml_tensor * ln_2_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;
|
||||
|
||||
// mlp
|
||||
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 gpt2_model {
|
||||
gpt2_hparams hparams;
|
||||
|
||||
// normalization
|
||||
struct ggml_tensor * ln_f_g;
|
||||
struct ggml_tensor * ln_f_b;
|
||||
|
||||
struct ggml_tensor * wte; // token embedding
|
||||
struct ggml_tensor * wpe; // position embedding
|
||||
struct ggml_tensor * lm_head; // language model head
|
||||
|
||||
std::vector<gpt2_layer> layers;
|
||||
|
||||
// key + value memory
|
||||
struct ggml_tensor * memory_k;
|
||||
struct ggml_tensor * memory_v;
|
||||
|
||||
//
|
||||
struct ggml_context * ctx_w;
|
||||
struct ggml_context * ctx_kv;
|
||||
|
||||
ggml_backend_t backend = NULL;
|
||||
|
||||
ggml_backend_buffer_t buffer_w;
|
||||
ggml_backend_buffer_t buffer_kv;
|
||||
|
||||
std::map<std::string, struct ggml_tensor *> tensors;
|
||||
};
|
||||
|
||||
// load the model's weights from a file
|
||||
bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab, int n_ctx, int n_gpu_layers) {
|
||||
printf("%s: loading model from '%s'\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.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: ftype = %d\n", __func__, hparams.ftype);
|
||||
printf("%s: qntvr = %d\n", __func__, qntvr);
|
||||
|
||||
hparams.ftype %= GGML_QNT_VERSION_FACTOR;
|
||||
}
|
||||
|
||||
// load vocab
|
||||
{
|
||||
int32_t n_vocab = 0;
|
||||
fin.read((char *) &n_vocab, sizeof(n_vocab));
|
||||
|
||||
if (n_vocab != model.hparams.n_vocab) {
|
||||
fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
|
||||
__func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
|
||||
return false;
|
||||
}
|
||||
|
||||
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;
|
||||
}
|
||||
|
||||
ggml_log_set(ggml_log_callback_default, nullptr);
|
||||
|
||||
auto & ctx = model.ctx_w;
|
||||
|
||||
// create the ggml context
|
||||
{
|
||||
size_t n_tensors = 2 + 6 + 12*model.hparams.n_layer;
|
||||
struct ggml_init_params params = {
|
||||
/*.mem_size =*/ ggml_tensor_overhead() * n_tensors,
|
||||
/*.mem_buffer =*/ NULL,
|
||||
/*.no_alloc =*/ true,
|
||||
};
|
||||
|
||||
ctx = ggml_init(params);
|
||||
if (!ctx) {
|
||||
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
// initialize the backend
|
||||
#ifdef GGML_USE_CUDA
|
||||
if (n_gpu_layers > 0) {
|
||||
fprintf(stderr, "%s: using CUDA backend\n", __func__);
|
||||
model.backend = ggml_backend_cuda_init(0);
|
||||
if (!model.backend) {
|
||||
fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
#ifdef GGML_USE_METAL
|
||||
if (n_gpu_layers > 0) {
|
||||
fprintf(stderr, "%s: using Metal backend\n", __func__);
|
||||
model.backend = ggml_backend_metal_init();
|
||||
if (!model.backend) {
|
||||
fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
if (!model.backend) {
|
||||
// fallback to CPU backend
|
||||
fprintf(stderr, "%s: using CPU backend\n", __func__);
|
||||
model.backend = ggml_backend_cpu_init();
|
||||
}
|
||||
|
||||
if (!model.backend) {
|
||||
fprintf(stderr, "%s: ggml_backend_cpu_init() failed\n", __func__);
|
||||
return false;
|
||||
}
|
||||
|
||||
// create the tensors for the model
|
||||
{
|
||||
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_vocab = hparams.n_vocab;
|
||||
|
||||
model.layers.resize(n_layer);
|
||||
|
||||
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.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
||||
model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
|
||||
model.lm_head = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
||||
|
||||
// map by name
|
||||
model.tensors["model/ln_f/g"] = model.ln_f_g;
|
||||
model.tensors["model/ln_f/b"] = model.ln_f_b;
|
||||
|
||||
model.tensors["model/wte"] = model.wte;
|
||||
model.tensors["model/wpe"] = model.wpe;
|
||||
model.tensors["model/lm_head"] = model.lm_head;
|
||||
|
||||
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.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_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.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["model/h" + std::to_string(i) + "/ln_1/g"] = layer.ln_1_g;
|
||||
model.tensors["model/h" + std::to_string(i) + "/ln_1/b"] = layer.ln_1_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/ln_2/g"] = layer.ln_2_g;
|
||||
model.tensors["model/h" + std::to_string(i) + "/ln_2/b"] = layer.ln_2_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"] = layer.c_mlp_fc_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"] = layer.c_mlp_fc_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"] = layer.c_mlp_proj_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"] = layer.c_mlp_proj_b;
|
||||
}
|
||||
}
|
||||
|
||||
// allocate the model tensors in a backend buffer
|
||||
model.buffer_w = ggml_backend_alloc_ctx_tensors(ctx, model.backend);
|
||||
|
||||
printf("%s: ggml tensor size = %d bytes\n", __func__, (int) sizeof(ggml_tensor));
|
||||
printf("%s: backend buffer size = %6.2f MB\n", __func__, ggml_backend_buffer_get_size(model.buffer_w)/(1024.0*1024.0));
|
||||
|
||||
// override the default training context with the user-provided
|
||||
model.hparams.n_ctx = n_ctx;
|
||||
|
||||
// key + value memory
|
||||
{
|
||||
auto * ctx = model.ctx_kv;
|
||||
|
||||
// create the ggml context
|
||||
{
|
||||
size_t n_tensors = 2;
|
||||
struct ggml_init_params params = {
|
||||
/*.mem_size =*/ ggml_tensor_overhead() * n_tensors,
|
||||
/*.mem_buffer =*/ NULL,
|
||||
/*.no_alloc =*/ true,
|
||||
};
|
||||
|
||||
ctx = ggml_init(params);
|
||||
if (!ctx) {
|
||||
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
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_mem = n_layer*n_ctx;
|
||||
const int n_elements = n_embd*n_mem;
|
||||
|
||||
// k and v here can also be GGML_TYPE_F16 to save memory and speed up the computation
|
||||
// if backend supports it
|
||||
model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
|
||||
model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
|
||||
|
||||
// allocate the KV memory in a backend buffer
|
||||
model.buffer_kv = ggml_backend_alloc_ctx_tensors(ctx, model.backend);
|
||||
|
||||
const size_t memory_size = ggml_backend_buffer_get_size(model.buffer_kv);
|
||||
printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
|
||||
}
|
||||
|
||||
// load weights
|
||||
{
|
||||
size_t total_size = 0;
|
||||
|
||||
bool has_lm_head = false;
|
||||
|
||||
std::vector<char> read_buf;
|
||||
|
||||
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) == model.tensors.end()) {
|
||||
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
|
||||
return false;
|
||||
}
|
||||
|
||||
auto tensor = model.tensors[name];
|
||||
ggml_set_name(tensor, name.c_str());
|
||||
if (ggml_nelements(tensor) != nelements) {
|
||||
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
|
||||
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 [%d, %d], expected [%d, %d]\n",
|
||||
__func__, name.c_str(), (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.c_str(), 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.c_str(), ggml_nbytes(tensor), nelements*bpe);
|
||||
return false;
|
||||
}
|
||||
|
||||
if (ggml_backend_buffer_is_host(model.buffer_w)) {
|
||||
// for some backends such as CPU and Metal, the tensor data is in system memory and we can read directly into it
|
||||
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
|
||||
} else {
|
||||
// read into a temporary buffer first, then copy to device memory
|
||||
read_buf.resize(ggml_nbytes(tensor));
|
||||
fin.read(read_buf.data(), ggml_nbytes(tensor));
|
||||
ggml_backend_tensor_set(tensor, read_buf.data(), 0, ggml_nbytes(tensor));
|
||||
}
|
||||
|
||||
// GPT-2 models share the WTE tensor as the LM head
|
||||
if (name == "model/wte" && has_lm_head == false) {
|
||||
//ggml_backend_tensor_copy(tensor, model.lm_head);
|
||||
model.lm_head = tensor;
|
||||
}
|
||||
|
||||
if (name == "model/lm_head") {
|
||||
has_lm_head = true;
|
||||
}
|
||||
|
||||
total_size += ggml_nbytes(tensor);
|
||||
}
|
||||
|
||||
printf("%s: model size = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
|
||||
}
|
||||
|
||||
fin.close();
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
// build the computation graph
|
||||
struct ggml_cgraph * gpt2_graph(
|
||||
const gpt2_model & model,
|
||||
const int n_past,
|
||||
const int n_tokens) {
|
||||
const int N = n_tokens;
|
||||
|
||||
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;
|
||||
|
||||
// since we are using ggml-alloc, this buffer only needs enough space to hold the ggml_tensor and ggml_cgraph structs, but not the tensor data
|
||||
static size_t buf_size = ggml_tensor_overhead()*GPT2_MAX_NODES + ggml_graph_overhead_custom(GPT2_MAX_NODES, false);
|
||||
static std::vector<uint8_t> buf(buf_size);
|
||||
|
||||
struct ggml_init_params params = {
|
||||
/*.mem_size =*/ buf_size,
|
||||
/*.mem_buffer =*/ buf.data(),
|
||||
/*.no_alloc =*/ true, // the tensors will be allocated later by ggml_gallocr_alloc_graph()
|
||||
};
|
||||
|
||||
struct ggml_context * ctx = ggml_init(params);
|
||||
|
||||
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx, GPT2_MAX_NODES, false);
|
||||
|
||||
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
|
||||
// at this point, the tensor data is not allocated yet and cannot be set
|
||||
// we will find the tensor after the graph is allocated by its name, and set the data then
|
||||
ggml_set_name(embd, "embd");
|
||||
// setting a tensor as an input will ensure that it is allocated at the beginning of the graph
|
||||
// this is important to ensure that the input tensors are not overwritten before they are used
|
||||
ggml_set_input(embd);
|
||||
|
||||
struct ggml_tensor * position = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
|
||||
ggml_set_name(position, "position");
|
||||
ggml_set_input(position);
|
||||
|
||||
// wte + wpe
|
||||
struct ggml_tensor * inpL =
|
||||
ggml_add(ctx,
|
||||
ggml_get_rows(ctx, model.wte, embd),
|
||||
ggml_get_rows(ctx, model.wpe, position));
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
struct ggml_tensor * cur;
|
||||
|
||||
// norm
|
||||
{
|
||||
// [ 768, N]
|
||||
cur = ggml_norm(ctx, inpL, hparams.eps);
|
||||
|
||||
// cur = ln_1_g*cur + ln_1_b
|
||||
// [ 768, N]
|
||||
cur = ggml_add(ctx,
|
||||
ggml_mul(ctx,
|
||||
cur,
|
||||
model.layers[il].ln_1_g),
|
||||
model.layers[il].ln_1_b);
|
||||
}
|
||||
|
||||
// attn
|
||||
// [2304, 768] - model.layers[il].c_attn_attn_w
|
||||
// [2304, 1] - model.layers[il].c_attn_attn_b
|
||||
// [ 768, N] - cur (in)
|
||||
// [2304, N] - cur (out)
|
||||
//
|
||||
// cur = attn_w*cur + attn_b
|
||||
// [2304, N]
|
||||
{
|
||||
cur = ggml_mul_mat(ctx,
|
||||
model.layers[il].c_attn_attn_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx,
|
||||
cur,
|
||||
model.layers[il].c_attn_attn_b);
|
||||
}
|
||||
|
||||
// self-attention
|
||||
{
|
||||
struct ggml_tensor * Qcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
|
||||
struct ggml_tensor * Kcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
|
||||
struct ggml_tensor * Vcur = ggml_view_2d(ctx, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
|
||||
|
||||
// store key and value to memory
|
||||
if (N >= 1) {
|
||||
struct ggml_tensor * k = ggml_view_1d(ctx, 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_1d(ctx, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
|
||||
|
||||
ggml_build_forward_expand(gf, ggml_cpy(ctx, Kcur, k));
|
||||
ggml_build_forward_expand(gf, ggml_cpy(ctx, Vcur, v));
|
||||
}
|
||||
|
||||
// Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
|
||||
// [64, N, 12]
|
||||
struct ggml_tensor * Q =
|
||||
ggml_permute(ctx,
|
||||
ggml_cont_3d(ctx, Qcur, n_embd/n_head, n_head, N),
|
||||
0, 2, 1, 3);
|
||||
|
||||
// K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
|
||||
// [64, n_past + N, 12]
|
||||
struct ggml_tensor * K =
|
||||
ggml_permute(ctx,
|
||||
ggml_reshape_3d(ctx,
|
||||
ggml_view_1d(ctx, 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);
|
||||
|
||||
// GG: flash attention
|
||||
//struct ggml_tensor * V =
|
||||
// ggml_cpy(ctx0,
|
||||
// ggml_permute(ctx0,
|
||||
// ggml_reshape_3d(ctx0,
|
||||
// ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
|
||||
// n_embd/n_head, n_head, n_past + N),
|
||||
// 1, 2, 0, 3),
|
||||
// ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
|
||||
|
||||
//struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
|
||||
|
||||
// K * Q
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ = ggml_mul_mat(ctx, K, Q);
|
||||
|
||||
// KQ_scaled = KQ / sqrt(n_embd/n_head)
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ_scaled =
|
||||
ggml_scale(ctx,
|
||||
KQ,
|
||||
1.0f/sqrtf(float(n_embd)/n_head));
|
||||
|
||||
// KQ_masked = mask_past(KQ_scaled)
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx, KQ_scaled, n_past);
|
||||
|
||||
// KQ = soft_max(KQ_masked)
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx, KQ_masked);
|
||||
|
||||
// V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
|
||||
// [n_past + N, 64, 12]
|
||||
struct ggml_tensor * V_trans =
|
||||
ggml_cont_3d(ctx,
|
||||
ggml_permute(ctx,
|
||||
ggml_reshape_3d(ctx,
|
||||
ggml_view_1d(ctx, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
|
||||
n_embd/n_head, n_head, n_past + N),
|
||||
1, 2, 0, 3),
|
||||
n_past + N, n_embd/n_head, n_head);
|
||||
|
||||
// KQV = transpose(V) * KQ_soft_max
|
||||
// [64, N, 12]
|
||||
struct ggml_tensor * KQV = ggml_mul_mat(ctx, V_trans, KQ_soft_max);
|
||||
|
||||
// KQV_merged = KQV.permute(0, 2, 1, 3)
|
||||
// [64, 12, N]
|
||||
struct ggml_tensor * KQV_merged = ggml_permute(ctx, KQV, 0, 2, 1, 3);
|
||||
|
||||
// cur = KQV_merged.contiguous().view(n_embd, N)
|
||||
// [768, N]
|
||||
cur = ggml_cont_2d(ctx, KQV_merged, n_embd, N);
|
||||
}
|
||||
|
||||
// projection
|
||||
// [ 768, 768] - model.layers[il].c_attn_proj_w
|
||||
// [ 768, 1] - model.layers[il].c_attn_proj_b
|
||||
// [ 768, N] - cur (in)
|
||||
// [ 768, N] - cur (out)
|
||||
//
|
||||
// cur = proj_w*cur + proj_b
|
||||
// [768, N]
|
||||
{
|
||||
cur = ggml_mul_mat(ctx,
|
||||
model.layers[il].c_attn_proj_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx,
|
||||
cur,
|
||||
model.layers[il].c_attn_proj_b);
|
||||
}
|
||||
|
||||
// add the input
|
||||
cur = ggml_add(ctx, cur, inpL);
|
||||
|
||||
struct ggml_tensor * inpFF = cur;
|
||||
|
||||
// feed-forward network
|
||||
{
|
||||
// norm
|
||||
{
|
||||
cur = ggml_norm(ctx, inpFF, hparams.eps);
|
||||
|
||||
// cur = ln_2_g*cur + ln_2_b
|
||||
// [ 768, N]
|
||||
cur = ggml_add(ctx,
|
||||
ggml_mul(ctx,
|
||||
cur,
|
||||
model.layers[il].ln_2_g),
|
||||
model.layers[il].ln_2_b);
|
||||
}
|
||||
|
||||
// fully connected
|
||||
// [3072, 768] - model.layers[il].c_mlp_fc_w
|
||||
// [3072, 1] - model.layers[il].c_mlp_fc_b
|
||||
// [ 768, N] - cur (in)
|
||||
// [3072, N] - cur (out)
|
||||
//
|
||||
// cur = fc_w*cur + fc_b
|
||||
// [3072, N]
|
||||
cur = ggml_mul_mat(ctx,
|
||||
model.layers[il].c_mlp_fc_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx,
|
||||
cur,
|
||||
model.layers[il].c_mlp_fc_b);
|
||||
|
||||
// GELU activation
|
||||
// [3072, N]
|
||||
cur = ggml_gelu(ctx, cur);
|
||||
|
||||
// projection
|
||||
// [ 768, 3072] - model.layers[il].c_mlp_proj_w
|
||||
// [ 768, 1] - model.layers[il].c_mlp_proj_b
|
||||
// [3072, N] - cur (in)
|
||||
// [ 768, N] - cur (out)
|
||||
//
|
||||
// cur = proj_w*cur + proj_b
|
||||
// [768, N]
|
||||
cur = ggml_mul_mat(ctx,
|
||||
model.layers[il].c_mlp_proj_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx,
|
||||
cur,
|
||||
model.layers[il].c_mlp_proj_b);
|
||||
}
|
||||
|
||||
// input for next layer
|
||||
inpL = ggml_add(ctx, cur, inpFF);
|
||||
}
|
||||
|
||||
// norm
|
||||
{
|
||||
// [ 768, N]
|
||||
inpL = ggml_norm(ctx, inpL, hparams.eps);
|
||||
|
||||
// inpL = ln_f_g*inpL + ln_f_b
|
||||
// [ 768, N]
|
||||
inpL = ggml_add(ctx,
|
||||
ggml_mul(ctx,
|
||||
inpL,
|
||||
model.ln_f_g),
|
||||
model.ln_f_b);
|
||||
}
|
||||
|
||||
// inpL = WTE * inpL
|
||||
// [ 768, 50257] - model.lm_head
|
||||
// [ 768, N] - inpL
|
||||
inpL = ggml_mul_mat(ctx, model.lm_head, inpL);
|
||||
ggml_set_name(inpL, "logits");
|
||||
// setting a tensor as the output will ensure that it is not overwritten by subsequent operations
|
||||
ggml_set_output(inpL);
|
||||
|
||||
// logits -> probs
|
||||
//inpL = ggml_soft_max(ctx0, inpL);
|
||||
|
||||
ggml_build_forward_expand(gf, inpL);
|
||||
|
||||
ggml_free(ctx);
|
||||
|
||||
return gf;
|
||||
}
|
||||
|
||||
// evaluate the transformer
|
||||
//
|
||||
// - model: the model
|
||||
// - allocr: ggml_gallocr to use to allocate the compute buffer
|
||||
// - 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 gpt2_eval(
|
||||
const gpt2_model & model,
|
||||
ggml_gallocr_t allocr,
|
||||
const int n_threads,
|
||||
const int n_past,
|
||||
const std::vector<gpt_vocab::id> & embd_inp,
|
||||
std::vector<float> & embd_w) {
|
||||
const int N = embd_inp.size();
|
||||
|
||||
const auto & hparams = model.hparams;
|
||||
|
||||
const int n_vocab = hparams.n_vocab;
|
||||
|
||||
struct ggml_cgraph * gf = gpt2_graph(model, n_past, embd_inp.size());
|
||||
|
||||
// allocate the graph tensors
|
||||
ggml_gallocr_alloc_graph(allocr, gf);
|
||||
|
||||
// set the graph inputs
|
||||
struct ggml_tensor * embd = ggml_graph_get_tensor(gf, "embd");
|
||||
ggml_backend_tensor_set(embd, embd_inp.data(), 0, N*ggml_element_size(embd));
|
||||
|
||||
struct ggml_tensor * position = ggml_graph_get_tensor(gf, "position");
|
||||
for (int i = 0; i < N; ++i) {
|
||||
int32_t v = n_past + i;
|
||||
ggml_backend_tensor_set(position, &v, i*sizeof(int32_t), sizeof(v));
|
||||
}
|
||||
|
||||
// set backend options
|
||||
if (ggml_backend_is_cpu(model.backend)) {
|
||||
ggml_backend_cpu_set_n_threads(model.backend, n_threads);
|
||||
}
|
||||
|
||||
// run the computation
|
||||
ggml_backend_graph_compute(model.backend, gf);
|
||||
|
||||
//if (n_past%100 == 0) {
|
||||
// ggml_graph_print (&gf);
|
||||
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
|
||||
//}
|
||||
|
||||
// get the graph outputs
|
||||
struct ggml_tensor * logits = ggml_graph_get_tensor(gf, "logits");
|
||||
|
||||
//embd_w.resize(n_vocab*N);
|
||||
//ggml_backend_tensor_get(logits, embd_w.data(), 0, sizeof(float)*n_vocab*N);
|
||||
|
||||
// return result just for the last token
|
||||
embd_w.resize(n_vocab);
|
||||
ggml_backend_tensor_get(logits, embd_w.data(), (n_vocab*(N-1))*sizeof(float), sizeof(float)*n_vocab);
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
int main(int argc, char ** argv) {
|
||||
ggml_time_init();
|
||||
|
||||
const int64_t t_main_start_us = ggml_time_us();
|
||||
|
||||
gpt_params params;
|
||||
params.model = "models/gpt-2-117M/ggml-model.bin";
|
||||
|
||||
if (gpt_params_parse(argc, argv, params) == false) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
if (params.seed < 0) {
|
||||
params.seed = time(NULL);
|
||||
}
|
||||
|
||||
printf("%s: seed = %d\n", __func__, params.seed);
|
||||
|
||||
std::mt19937 rng(params.seed);
|
||||
if (params.prompt.empty()) {
|
||||
params.prompt = gpt_random_prompt(rng);
|
||||
}
|
||||
|
||||
int64_t t_load_us = 0;
|
||||
|
||||
gpt_vocab vocab;
|
||||
gpt2_model model;
|
||||
|
||||
// load the model
|
||||
{
|
||||
const int64_t t_start_us = ggml_time_us();
|
||||
|
||||
if (!gpt2_model_load(params.model, model, vocab, params.n_ctx, params.n_gpu_layers)) {
|
||||
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
|
||||
return 1;
|
||||
}
|
||||
|
||||
t_load_us = ggml_time_us() - t_start_us;
|
||||
|
||||
test_gpt_tokenizer(vocab, params.token_test);
|
||||
}
|
||||
|
||||
ggml_gallocr_t allocr = NULL;
|
||||
// allocate the compute buffer
|
||||
{
|
||||
// create a graph allocator with the backend's default buffer type
|
||||
allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));
|
||||
|
||||
// create the worst case graph for memory usage estimation
|
||||
int n_tokens = std::min(model.hparams.n_ctx, params.n_batch);
|
||||
int n_past = model.hparams.n_ctx - n_tokens;
|
||||
struct ggml_cgraph * gf = gpt2_graph(model, n_past, n_tokens);
|
||||
|
||||
// pre-allocate the compute buffer for the worst case (optional)
|
||||
ggml_gallocr_reserve(allocr, gf);
|
||||
size_t mem_size = ggml_gallocr_get_buffer_size(allocr, 0);
|
||||
fprintf(stderr, "%s: compute buffer size: %.2f MB\n", __func__, mem_size/1024.0/1024.0);
|
||||
}
|
||||
|
||||
int n_past = 0;
|
||||
|
||||
int64_t t_sample_us = 0;
|
||||
int64_t t_predict_us = 0;
|
||||
|
||||
std::vector<float> logits;
|
||||
|
||||
// tokenize the prompt
|
||||
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
|
||||
|
||||
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
|
||||
|
||||
printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
|
||||
printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size());
|
||||
for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) {
|
||||
printf("%d ", embd_inp[i]);
|
||||
}
|
||||
printf("\n\n");
|
||||
|
||||
// submit the input prompt token-by-token
|
||||
// this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
|
||||
std::vector<gpt_vocab::id> embd;
|
||||
|
||||
for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
|
||||
// predict
|
||||
if (embd.size() > 0) {
|
||||
const int64_t t_start_us = ggml_time_us();
|
||||
|
||||
if (!gpt2_eval(model, allocr, params.n_threads, n_past, embd, logits)) {
|
||||
printf("Failed to predict\n");
|
||||
return 1;
|
||||
}
|
||||
|
||||
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 = params.top_k;
|
||||
const float top_p = params.top_p;
|
||||
const float temp = params.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);
|
||||
} else {
|
||||
// if here, it means we are still processing the input prompt
|
||||
for (size_t k = i; k < embd_inp.size(); k++) {
|
||||
embd.push_back(embd_inp[k]);
|
||||
if (int32_t(embd.size()) >= params.n_batch) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
i += embd.size() - 1;
|
||||
}
|
||||
|
||||
// display text
|
||||
for (auto id : embd) {
|
||||
printf("%s", vocab.id_to_token[id].c_str());
|
||||
}
|
||||
fflush(stdout);
|
||||
|
||||
// end of text token
|
||||
if (!params.ignore_eos && embd.back() == 50256) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
// report timing
|
||||
{
|
||||
const int64_t t_main_end_us = ggml_time_us();
|
||||
|
||||
printf("\n\n");
|
||||
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
|
||||
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
|
||||
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
|
||||
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
|
||||
}
|
||||
|
||||
ggml_free(model.ctx_w);
|
||||
|
||||
ggml_gallocr_free(allocr);
|
||||
ggml_backend_buffer_free(model.buffer_w);
|
||||
ggml_backend_buffer_free(model.buffer_kv);
|
||||
ggml_backend_free(model.backend);
|
||||
|
||||
return 0;
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,840 @@
|
||||
#include "ggml.h"
|
||||
#include "ggml-cpu.h"
|
||||
#include "common.h"
|
||||
#include "common-ggml.h"
|
||||
|
||||
#include <cassert>
|
||||
#include <cmath>
|
||||
#include <cstdio>
|
||||
#include <cstring>
|
||||
#include <fstream>
|
||||
#include <map>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
#if defined(_MSC_VER)
|
||||
#pragma warning(disable: 4244 4267) // possible loss of data
|
||||
#endif
|
||||
|
||||
// default hparams (GPT-2 117M)
|
||||
struct gpt2_hparams {
|
||||
int32_t n_vocab = 50257;
|
||||
int32_t n_ctx = 1024;
|
||||
int32_t n_embd = 768;
|
||||
int32_t n_head = 12;
|
||||
int32_t n_layer = 12;
|
||||
int32_t ftype = 1;
|
||||
float eps = 1e-5f;
|
||||
};
|
||||
|
||||
struct gpt2_layer {
|
||||
// normalization
|
||||
struct ggml_tensor * ln_1_g;
|
||||
struct ggml_tensor * ln_1_b;
|
||||
|
||||
struct ggml_tensor * ln_2_g;
|
||||
struct ggml_tensor * ln_2_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;
|
||||
|
||||
// mlp
|
||||
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 gpt2_model {
|
||||
gpt2_hparams hparams;
|
||||
|
||||
// normalization
|
||||
struct ggml_tensor * ln_f_g;
|
||||
struct ggml_tensor * ln_f_b;
|
||||
|
||||
struct ggml_tensor * wte; // token embedding
|
||||
struct ggml_tensor * wpe; // position embedding
|
||||
struct ggml_tensor * lm_head; // language model head
|
||||
|
||||
std::vector<gpt2_layer> layers;
|
||||
|
||||
// key + value memory
|
||||
struct ggml_tensor * memory_k;
|
||||
struct ggml_tensor * memory_v;
|
||||
|
||||
//
|
||||
struct ggml_context * ctx_w;
|
||||
std::map<std::string, struct ggml_tensor *> tensors;
|
||||
};
|
||||
|
||||
// load the model's weights from a file
|
||||
bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
|
||||
printf("%s: loading model from '%s'\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.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: ftype = %d\n", __func__, hparams.ftype);
|
||||
printf("%s: qntvr = %d\n", __func__, qntvr);
|
||||
|
||||
hparams.ftype %= GGML_QNT_VERSION_FACTOR;
|
||||
}
|
||||
|
||||
// load vocab
|
||||
{
|
||||
int32_t n_vocab = 0;
|
||||
fin.read((char *) &n_vocab, sizeof(n_vocab));
|
||||
|
||||
if (n_vocab != model.hparams.n_vocab) {
|
||||
fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
|
||||
__func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
|
||||
return false;
|
||||
}
|
||||
|
||||
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_w;
|
||||
|
||||
size_t ctx_size = 0;
|
||||
|
||||
{
|
||||
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_vocab = hparams.n_vocab;
|
||||
|
||||
ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
|
||||
ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
|
||||
|
||||
ctx_size += ggml_row_size(wtype, n_vocab*n_embd); // wte
|
||||
ctx_size += ggml_row_size(GGML_TYPE_F32, n_ctx*n_embd); // wpe
|
||||
ctx_size += ggml_row_size(wtype, n_vocab*n_embd); // lm_head
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_g
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_b
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(wtype, 3*n_embd*n_embd)); // c_attn_attn_w
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 3*n_embd)); // c_attn_attn_b
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(wtype, n_embd*n_embd)); // c_attn_proj_w
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // c_attn_proj_b
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_fc_w
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd)); // c_mlp_fc_b
|
||||
|
||||
ctx_size += n_layer*(ggml_row_size(wtype, 4*n_embd*n_embd)); // c_mlp_proj_w
|
||||
ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd)); // c_mlp_proj_b
|
||||
|
||||
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_k
|
||||
ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_v
|
||||
|
||||
ctx_size += (6 + 12*n_layer)*512; // object overhead
|
||||
|
||||
printf("%s: ggml tensor size = %d bytes\n", __func__, (int) sizeof(ggml_tensor));
|
||||
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_w = ggml_init(params);
|
||||
if (!model.ctx_w) {
|
||||
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_ctx = hparams.n_ctx;
|
||||
const int n_vocab = hparams.n_vocab;
|
||||
|
||||
model.layers.resize(n_layer);
|
||||
|
||||
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.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
||||
model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
|
||||
model.lm_head = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
|
||||
|
||||
// map by name
|
||||
model.tensors["model/ln_f/g"] = model.ln_f_g;
|
||||
model.tensors["model/ln_f/b"] = model.ln_f_b;
|
||||
|
||||
model.tensors["model/wte"] = model.wte;
|
||||
model.tensors["model/wpe"] = model.wpe;
|
||||
model.tensors["model/lm_head"] = model.lm_head;
|
||||
|
||||
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.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_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.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["model/h" + std::to_string(i) + "/ln_1/g"] = layer.ln_1_g;
|
||||
model.tensors["model/h" + std::to_string(i) + "/ln_1/b"] = layer.ln_1_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/ln_2/g"] = layer.ln_2_g;
|
||||
model.tensors["model/h" + std::to_string(i) + "/ln_2/b"] = layer.ln_2_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"] = layer.c_mlp_fc_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"] = layer.c_mlp_fc_b;
|
||||
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"] = layer.c_mlp_proj_w;
|
||||
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"] = 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 int n_mem = n_layer*n_ctx;
|
||||
const int n_elements = n_embd*n_mem;
|
||||
|
||||
model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
|
||||
model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 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 = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
|
||||
}
|
||||
|
||||
// load weights
|
||||
{
|
||||
size_t total_size = 0;
|
||||
|
||||
bool has_lm_head = false;
|
||||
|
||||
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) == model.tensors.end()) {
|
||||
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
|
||||
return false;
|
||||
}
|
||||
|
||||
auto tensor = model.tensors[name];
|
||||
if (ggml_nelements(tensor) != nelements) {
|
||||
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
|
||||
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 [%d, %d], expected [%d, %d]\n",
|
||||
__func__, name.c_str(), (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.c_str(), 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.c_str(), ggml_nbytes(tensor), nelements*bpe);
|
||||
return false;
|
||||
}
|
||||
|
||||
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
|
||||
|
||||
// GPT-2 models share the WTE tensor as the LM head
|
||||
if (name == "model/wte" && has_lm_head == false) {
|
||||
memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
|
||||
}
|
||||
|
||||
if (name == "model/lm_head") {
|
||||
has_lm_head = true;
|
||||
}
|
||||
|
||||
total_size += ggml_nbytes(tensor);
|
||||
}
|
||||
|
||||
printf("%s: model size = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
|
||||
}
|
||||
|
||||
fin.close();
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
// 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 gpt2_eval(
|
||||
const gpt2_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;
|
||||
|
||||
static size_t buf_size = 256u*1024*1024;
|
||||
static void * buf = malloc(buf_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 = ggml_new_graph(ctx0);
|
||||
|
||||
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
|
||||
memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
|
||||
|
||||
struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
|
||||
for (int i = 0; i < N; ++i) {
|
||||
((int32_t *) position->data)[i] = n_past + i;
|
||||
}
|
||||
|
||||
// wte + wpe
|
||||
struct ggml_tensor * inpL =
|
||||
ggml_add(ctx0,
|
||||
ggml_get_rows(ctx0, model.wte, embd),
|
||||
ggml_get_rows(ctx0, model.wpe, position));
|
||||
|
||||
for (int il = 0; il < n_layer; ++il) {
|
||||
struct ggml_tensor * cur;
|
||||
|
||||
// norm
|
||||
{
|
||||
// [ 768, N]
|
||||
cur = ggml_norm(ctx0, inpL, hparams.eps);
|
||||
|
||||
// cur = ln_1_g*cur + ln_1_b
|
||||
// [ 768, N]
|
||||
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));
|
||||
}
|
||||
|
||||
// attn
|
||||
// [2304, 768] - model.layers[il].c_attn_attn_w
|
||||
// [2304, 1] - model.layers[il].c_attn_attn_b
|
||||
// [ 768, N] - cur (in)
|
||||
// [2304, N] - cur (out)
|
||||
//
|
||||
// cur = attn_w*cur + attn_b
|
||||
// [2304, N]
|
||||
{
|
||||
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);
|
||||
}
|
||||
|
||||
// self-attention
|
||||
{
|
||||
struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
|
||||
struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
|
||||
struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
|
||||
|
||||
// store key and value to memory
|
||||
if (N >= 1) {
|
||||
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_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
|
||||
|
||||
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)
|
||||
// [64, N, 12]
|
||||
struct ggml_tensor * Q =
|
||||
ggml_permute(ctx0,
|
||||
ggml_cpy(ctx0,
|
||||
Qcur,
|
||||
ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
|
||||
0, 2, 1, 3);
|
||||
|
||||
// K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
|
||||
// [64, n_past + N, 12]
|
||||
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);
|
||||
|
||||
// GG: flash attention
|
||||
//struct ggml_tensor * V =
|
||||
// ggml_cpy(ctx0,
|
||||
// ggml_permute(ctx0,
|
||||
// ggml_reshape_3d(ctx0,
|
||||
// ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
|
||||
// n_embd/n_head, n_head, n_past + N),
|
||||
// 1, 2, 0, 3),
|
||||
// ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
|
||||
|
||||
//struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
|
||||
|
||||
// K * Q
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
|
||||
|
||||
// KQ_scaled = KQ / sqrt(n_embd/n_head)
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, 1.0f/sqrt(float(n_embd)/n_head));
|
||||
|
||||
// KQ_masked = mask_past(KQ_scaled)
|
||||
// [n_past + N, N, 12]
|
||||
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
|
||||
|
||||
// KQ = soft_max(KQ_masked)
|
||||
// [n_past + N, N, 12]
|
||||
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()
|
||||
// [n_past + N, 64, 12]
|
||||
struct ggml_tensor * V_trans =
|
||||
ggml_cpy(ctx0,
|
||||
ggml_permute(ctx0,
|
||||
ggml_reshape_3d(ctx0,
|
||||
ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
|
||||
n_embd/n_head, n_head, n_past + N),
|
||||
1, 2, 0, 3),
|
||||
ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
|
||||
|
||||
// KQV = transpose(V) * KQ_soft_max
|
||||
// [64, N, 12]
|
||||
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
|
||||
|
||||
// KQV_merged = KQV.permute(0, 2, 1, 3)
|
||||
// [64, 12, N]
|
||||
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
|
||||
|
||||
// cur = KQV_merged.contiguous().view(n_embd, N)
|
||||
// [768, N]
|
||||
cur = ggml_cpy(ctx0,
|
||||
KQV_merged,
|
||||
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
|
||||
}
|
||||
|
||||
// projection
|
||||
// [ 768, 768] - model.layers[il].c_attn_proj_w
|
||||
// [ 768, 1] - model.layers[il].c_attn_proj_b
|
||||
// [ 768, N] - cur (in)
|
||||
// [ 768, N] - cur (out)
|
||||
//
|
||||
// cur = proj_w*cur + proj_b
|
||||
// [768, N]
|
||||
{
|
||||
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);
|
||||
}
|
||||
|
||||
// add the input
|
||||
cur = ggml_add(ctx0, cur, inpL);
|
||||
|
||||
struct ggml_tensor * inpFF = cur;
|
||||
|
||||
// feed-forward network
|
||||
{
|
||||
// norm
|
||||
{
|
||||
cur = ggml_norm(ctx0, inpFF, hparams.eps);
|
||||
|
||||
// cur = ln_2_g*cur + ln_2_b
|
||||
// [ 768, N]
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_mul(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
|
||||
cur),
|
||||
ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
|
||||
}
|
||||
|
||||
// fully connected
|
||||
// [3072, 768] - model.layers[il].c_mlp_fc_w
|
||||
// [3072, 1] - model.layers[il].c_mlp_fc_b
|
||||
// [ 768, N] - cur (in)
|
||||
// [3072, N] - cur (out)
|
||||
//
|
||||
// cur = fc_w*cur + fc_b
|
||||
// [3072, N]
|
||||
cur = ggml_mul_mat(ctx0,
|
||||
model.layers[il].c_mlp_fc_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
|
||||
cur);
|
||||
|
||||
// GELU activation
|
||||
// [3072, N]
|
||||
cur = ggml_gelu(ctx0, cur);
|
||||
|
||||
// projection
|
||||
// [ 768, 3072] - model.layers[il].c_mlp_proj_w
|
||||
// [ 768, 1] - model.layers[il].c_mlp_proj_b
|
||||
// [3072, N] - cur (in)
|
||||
// [ 768, N] - cur (out)
|
||||
//
|
||||
// cur = proj_w*cur + proj_b
|
||||
// [768, N]
|
||||
cur = ggml_mul_mat(ctx0,
|
||||
model.layers[il].c_mlp_proj_w,
|
||||
cur);
|
||||
|
||||
cur = ggml_add(ctx0,
|
||||
ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
|
||||
cur);
|
||||
}
|
||||
|
||||
// input for next layer
|
||||
inpL = ggml_add(ctx0, cur, inpFF);
|
||||
}
|
||||
|
||||
// norm
|
||||
{
|
||||
// [ 768, N]
|
||||
inpL = ggml_norm(ctx0, inpL, hparams.eps);
|
||||
|
||||
// inpL = ln_f_g*inpL + ln_f_b
|
||||
// [ 768, N]
|
||||
inpL = ggml_add(ctx0,
|
||||
ggml_mul(ctx0,
|
||||
ggml_repeat(ctx0, model.ln_f_g, inpL),
|
||||
inpL),
|
||||
ggml_repeat(ctx0, model.ln_f_b, inpL));
|
||||
}
|
||||
|
||||
// inpL = WTE * inpL
|
||||
// [ 768, 50257] - model.lm_head
|
||||
// [ 768, N] - inpL
|
||||
inpL = ggml_mul_mat(ctx0, model.lm_head, 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 just for 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;
|
||||
}
|
||||
|
||||
int main(int argc, char ** argv) {
|
||||
ggml_time_init();
|
||||
|
||||
const int64_t t_main_start_us = ggml_time_us();
|
||||
|
||||
gpt_params params;
|
||||
params.model = "models/gpt-2-117M/ggml-model.bin";
|
||||
|
||||
if (gpt_params_parse(argc, argv, params) == false) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
if (params.seed < 0) {
|
||||
params.seed = time(NULL);
|
||||
}
|
||||
|
||||
printf("%s: seed = %d\n", __func__, params.seed);
|
||||
|
||||
std::mt19937 rng(params.seed);
|
||||
if (params.prompt.empty()) {
|
||||
params.prompt = gpt_random_prompt(rng);
|
||||
}
|
||||
|
||||
int64_t t_load_us = 0;
|
||||
|
||||
gpt_vocab vocab;
|
||||
gpt2_model model;
|
||||
|
||||
// load the model
|
||||
{
|
||||
const int64_t t_start_us = ggml_time_us();
|
||||
|
||||
if (!gpt2_model_load(params.model, model, vocab)) {
|
||||
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
|
||||
return 1;
|
||||
}
|
||||
|
||||
t_load_us = ggml_time_us() - t_start_us;
|
||||
|
||||
test_gpt_tokenizer(vocab, params.token_test);
|
||||
}
|
||||
|
||||
int n_past = 0;
|
||||
|
||||
int64_t t_sample_us = 0;
|
||||
int64_t t_predict_us = 0;
|
||||
|
||||
std::vector<float> logits;
|
||||
|
||||
// tokenize the prompt
|
||||
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
|
||||
|
||||
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
|
||||
|
||||
printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
|
||||
printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size());
|
||||
for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) {
|
||||
printf("%d ", embd_inp[i]);
|
||||
}
|
||||
printf("\n\n");
|
||||
|
||||
// submit the input prompt token-by-token
|
||||
// this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
|
||||
std::vector<gpt_vocab::id> embd;
|
||||
|
||||
// determine the required inference memory per token:
|
||||
size_t mem_per_token = 0;
|
||||
gpt2_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
|
||||
|
||||
for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
|
||||
// predict
|
||||
if (embd.size() > 0) {
|
||||
const int64_t t_start_us = ggml_time_us();
|
||||
|
||||
if (!gpt2_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
|
||||
printf("Failed to predict\n");
|
||||
return 1;
|
||||
}
|
||||
|
||||
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 = params.top_k;
|
||||
const float top_p = params.top_p;
|
||||
const float temp = params.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);
|
||||
} else {
|
||||
// if here, it means we are still processing the input prompt
|
||||
for (size_t k = i; k < embd_inp.size(); k++) {
|
||||
embd.push_back(embd_inp[k]);
|
||||
if (int32_t(embd.size()) >= params.n_batch) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
i += embd.size() - 1;
|
||||
}
|
||||
|
||||
// display text
|
||||
for (auto id : embd) {
|
||||
printf("%s", vocab.id_to_token[id].c_str());
|
||||
}
|
||||
fflush(stdout);
|
||||
|
||||
// end of text token
|
||||
if (embd.back() == 50256) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
// report timing
|
||||
{
|
||||
const int64_t t_main_end_us = ggml_time_us();
|
||||
|
||||
printf("\n\n");
|
||||
printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
|
||||
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
|
||||
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
|
||||
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
|
||||
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
|
||||
}
|
||||
|
||||
ggml_free(model.ctx_w);
|
||||
|
||||
return 0;
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@@ -0,0 +1,184 @@
|
||||
#include "ggml.h"
|
||||
|
||||
#include "common.h"
|
||||
#include "common-ggml.h"
|
||||
|
||||
#include <cassert>
|
||||
#include <cmath>
|
||||
#include <cstdio>
|
||||
#include <cstring>
|
||||
#include <fstream>
|
||||
#include <map>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
#include <regex>
|
||||
|
||||
// default hparams (GPT-2 117M)
|
||||
struct gpt2_hparams {
|
||||
int32_t n_vocab = 50257;
|
||||
int32_t n_ctx = 1024;
|
||||
int32_t n_embd = 768;
|
||||
int32_t n_head = 12;
|
||||
int32_t n_layer = 12;
|
||||
int32_t ftype = 1;
|
||||
};
|
||||
|
||||
// quantize a model
|
||||
bool gpt2_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
|
||||
gpt_vocab vocab;
|
||||
|
||||
printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
|
||||
|
||||
auto finp = std::ifstream(fname_inp, std::ios::binary);
|
||||
if (!finp) {
|
||||
fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
|
||||
return false;
|
||||
}
|
||||
|
||||
auto fout = std::ofstream(fname_out, std::ios::binary);
|
||||
if (!fout) {
|
||||
fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
|
||||
return false;
|
||||
}
|
||||
|
||||
// verify magic
|
||||
{
|
||||
uint32_t magic;
|
||||
finp.read((char *) &magic, sizeof(magic));
|
||||
if (magic != GGML_FILE_MAGIC) {
|
||||
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
|
||||
return false;
|
||||
}
|
||||
|
||||
fout.write((char *) &magic, sizeof(magic));
|
||||
}
|
||||
|
||||
gpt2_hparams hparams;
|
||||
|
||||
// load hparams
|
||||
{
|
||||
finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
|
||||
finp.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
|
||||
finp.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
|
||||
finp.read((char *) &hparams.n_head, sizeof(hparams.n_head));
|
||||
finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
|
||||
finp.read((char *) &hparams.ftype, sizeof(hparams.ftype));
|
||||
|
||||
const int32_t qntvr_src = hparams.ftype / GGML_QNT_VERSION_FACTOR;
|
||||
const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
|
||||
|
||||
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: ftype (src) = %d\n", __func__, hparams.ftype);
|
||||
printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
|
||||
printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
|
||||
printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
|
||||
|
||||
fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
|
||||
fout.write((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
|
||||
fout.write((char *) &hparams.n_embd, sizeof(hparams.n_embd));
|
||||
fout.write((char *) &hparams.n_head, sizeof(hparams.n_head));
|
||||
fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
|
||||
fout.write((char *) &ftype_dst, sizeof(ftype_dst));
|
||||
}
|
||||
|
||||
// load vocab
|
||||
{
|
||||
int32_t n_vocab = 0;
|
||||
finp.read ((char *) &n_vocab, sizeof(n_vocab));
|
||||
fout.write((char *) &n_vocab, sizeof(n_vocab));
|
||||
|
||||
if (n_vocab != hparams.n_vocab) {
|
||||
fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
|
||||
__func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
|
||||
return false;
|
||||
}
|
||||
|
||||
std::string word;
|
||||
for (int i = 0; i < n_vocab; i++) {
|
||||
uint32_t len;
|
||||
finp.read ((char *) &len, sizeof(len));
|
||||
fout.write((char *) &len, sizeof(len));
|
||||
|
||||
word.resize(len);
|
||||
finp.read ((char *) word.data(), len);
|
||||
fout.write((char *) word.data(), len);
|
||||
|
||||
vocab.token_to_id[word] = i;
|
||||
vocab.id_to_token[i] = word;
|
||||
}
|
||||
}
|
||||
|
||||
// regexes of tensor names to be quantized
|
||||
const std::vector<std::string> to_quant = {
|
||||
"model/wte",
|
||||
"model/lm_head",
|
||||
"model/h.*/attn/c_attn/w",
|
||||
"model/h.*/attn/c_proj/w",
|
||||
"model/h.*/mlp/c_fc/w",
|
||||
"model/h.*/mlp/c_proj/w",
|
||||
};
|
||||
|
||||
if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) {
|
||||
fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
|
||||
return false;
|
||||
}
|
||||
|
||||
finp.close();
|
||||
fout.close();
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
// usage:
|
||||
// ./gpt-2-quantize models/gpt-2-117M/ggml-model.bin models/gpt-2-117M/ggml-model-quant.bin type
|
||||
//
|
||||
int main(int argc, char ** argv) {
|
||||
if (argc != 4) {
|
||||
fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
|
||||
ggml_print_ftypes(stderr);
|
||||
return 1;
|
||||
}
|
||||
|
||||
// needed to initialize f16 tables
|
||||
{
|
||||
struct ggml_init_params params = { 0, NULL, false };
|
||||
struct ggml_context * ctx = ggml_init(params);
|
||||
ggml_free(ctx);
|
||||
}
|
||||
|
||||
const std::string fname_inp = argv[1];
|
||||
const std::string fname_out = argv[2];
|
||||
|
||||
const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
|
||||
|
||||
const int64_t t_main_start_us = ggml_time_us();
|
||||
|
||||
int64_t t_quantize_us = 0;
|
||||
|
||||
// load the model
|
||||
{
|
||||
const int64_t t_start_us = ggml_time_us();
|
||||
|
||||
if (!gpt2_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
|
||||
fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
|
||||
return 1;
|
||||
}
|
||||
|
||||
t_quantize_us = ggml_time_us() - t_start_us;
|
||||
}
|
||||
|
||||
// report timing
|
||||
{
|
||||
const int64_t t_main_end_us = ggml_time_us();
|
||||
|
||||
printf("\n");
|
||||
printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
|
||||
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
Reference in New Issue
Block a user