mirror of
https://github.com/ml-explore/mlx-examples.git
synced 2025-06-27 03:05:20 +08:00
fix args, update README, remove extra files
This commit is contained in:
Awni Hannun
parent
a8d4149147
commit
1613e608a9
@ -1,24 +1,48 @@
|
||||
# Phi-2
|
||||
|
||||
Phi-2 is a 2.7B parameter model released by Microsoft and trained on a mixture of GPT-4 outputs and clean web-text.
|
||||
Its performance theoretically rivals much, much stronger models.
|
||||
Phi-2 is a 2.7B parameter model released by Microsoft[^1] and trained on a mixture
|
||||
of GPT-4 outputs and clean web-text. Its performance rivals
|
||||
much, much stronger models.
|
||||
|
||||
## Downloading and Converting Weights
|
||||
## Setup
|
||||
|
||||
To download and convert the model:
|
||||
Download and convert the model:
|
||||
|
||||
```sh
|
||||
python phi2/convert.py
|
||||
python convert.py
|
||||
```
|
||||
|
||||
That will fill in `weights/phi-2.npz`.
|
||||
which will make a file `weights.npz`.
|
||||
|
||||
## Running the Model
|
||||
## Generate
|
||||
|
||||
🚧 (Not yet done) To run the model:
|
||||
To generate text with the default prompt:
|
||||
|
||||
```sh
|
||||
python phi2/generate.py
|
||||
python model.py
|
||||
```
|
||||
|
||||
Layer-by-layer forward pass outputs are currently shown in the outputs.txt files.
|
||||
Should give the output:
|
||||
|
||||
```
|
||||
Answer: Mathematics is like a lighthouse that guides us through the darkness of
|
||||
uncertainty. Just as a lighthouse emits a steady beam of light, mathematics
|
||||
provides us with a clear path to navigate through complex problems. It
|
||||
illuminates our understanding and helps us make sense of the world around us.
|
||||
|
||||
Exercise 2:
|
||||
Compare and contrast the role of logic in mathematics and the role of a compass
|
||||
in navigation.
|
||||
|
||||
Answer: Logic in mathematics is like a compass in navigation. It helps
|
||||
```
|
||||
|
||||
To use your own prompt:
|
||||
|
||||
```sh
|
||||
python model.py --prompt <your prompt here> --max_tokens <max_token>
|
||||
```
|
||||
|
||||
[^1]: For more details on the model see the [blog post](
|
||||
https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/)
|
||||
and the [Hugging Face repo](https://huggingface.co/microsoft/phi-2)
|
||||
|
||||
@ -1,23 +0,0 @@
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
model = AutoModelForCausalLM.from_pretrained(
|
||||
"microsoft/phi-2", torch_dtype="auto", trust_remote_code=True
|
||||
)
|
||||
tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-2", trust_remote_code=True)
|
||||
|
||||
inputs = tokenizer(
|
||||
'''def print_prime(n):
|
||||
"""
|
||||
Print all primes between 1 and n
|
||||
"""''',
|
||||
return_tensors="pt",
|
||||
return_attention_mask=False,
|
||||
)
|
||||
|
||||
print(model(**inputs))
|
||||
|
||||
# outputs = model.generate(**inputs, max_length=200)
|
||||
# text = tokenizer.batch_decode(outputs)[0]
|
||||
# print(text)
|
||||
@ -203,11 +203,14 @@ if __name__ == "__main__":
|
||||
|
||||
prompt = mx.array(prompt)
|
||||
|
||||
print("[INFO] Generating with Phi-2...", flush=True)
|
||||
print(args.prompt, end="", flush=True)
|
||||
|
||||
tokens = []
|
||||
for token, _ in zip(generate(prompt, model), range(args.max_tokens)):
|
||||
tokens.append(token)
|
||||
|
||||
if (len(tokens) % args.tokens_per_eval) == 0:
|
||||
if (len(tokens) % 10) == 0:
|
||||
mx.eval(tokens)
|
||||
s = tokenizer.decode([t.item() for t in tokens])
|
||||
print(s, end="", flush=True)
|
||||
|
||||
@ -1,63 +0,0 @@
|
||||
(HF) Output of Embeddings
|
||||
|
||||
tensor([[[-0.0353, 0.0045, 0.0208, ..., -0.0117, 0.0041, 0.0075],
|
||||
[-0.0172, 0.0236, -0.0051, ..., 0.0141, 0.0115, 0.0058],
|
||||
[-0.0148, 0.0043, -0.0252, ..., 0.0179, 0.0025, -0.0008],
|
||||
...,
|
||||
[ 0.0003, 0.0051, 0.0002, ..., 0.0043, 0.0075, 0.0049],
|
||||
[-0.0110, 0.0472, 0.0030, ..., 0.0098, -0.0075, 0.0146],
|
||||
[-0.0085, -0.0219, -0.0016, ..., -0.0059, 0.0109, -0.0016]]],
|
||||
device='cuda:0', dtype=torch.float16, grad_fn=<EmbeddingBackward0>)
|
||||
|
||||
(MLX) Output of Embeddings
|
||||
|
||||
array([[[-0.0352783, 0.00445175, 0.020813, ..., -0.0117188, 0.00411606, 0.00748444],
|
||||
[-0.0171509, 0.0236053, -0.00508881, ..., 0.0141144, 0.0115204, 0.00582504],
|
||||
[-0.0147858, 0.00426102, -0.0252075, ..., 0.0179443, 0.0024662, -0.00076437],
|
||||
...,
|
||||
[0.000337124, 0.00508499, 0.000193119, ..., 0.00427628, 0.00753403, 0.00492477],
|
||||
[-0.0110092, 0.0472107, 0.00295448, ..., 0.00982666, -0.00747681, 0.0145721],
|
||||
[-0.00852203, -0.0218964, -0.00161839, ..., -0.00592422, 0.0108643, -0.00162697]]], dtype=float16)
|
||||
|
||||
(HF) Output of First Attention Layer
|
||||
|
||||
tensor([[[-0.2000, 0.4849, 0.9863, ..., -0.2209, 0.1355, 0.3469],
|
||||
[ 0.4922, -0.3865, 0.8428, ..., 0.5894, -0.0069, -0.5278],
|
||||
[ 0.0902, 0.1028, 0.6826, ..., 0.1394, -0.8145, -0.1880],
|
||||
...,
|
||||
[ 0.2380, 0.0555, -0.3005, ..., 0.0372, -0.0895, 0.0255],
|
||||
[ 0.2512, 0.1949, 0.3401, ..., 0.3625, -0.3103, -0.1064],
|
||||
[-0.0905, 0.0665, 0.5210, ..., -0.0767, -0.2460, -0.1449]]],
|
||||
device='cuda:0', dtype=torch.float16, grad_fn=<AddBackward0>)
|
||||
torch.Size([1, 23, 2560])
|
||||
|
||||
(MLX) Output of First Attention Layer
|
||||
|
||||
array([[[-0.199973, 0.485224, 0.987237, ..., -0.220847, 0.13511, 0.346074],
|
||||
[0.44883, -0.271683, 0.877478, ..., 0.653217, -0.0929724, -0.711176],
|
||||
[-0.233398, 5.7824e-05, 0.435001, ..., 0.0504494, -0.623998, -0.438785],
|
||||
...,
|
||||
[0.123587, -0.237459, -0.447518, ..., 0.0653363, -0.0767153, -0.341505],
|
||||
[0.187798, 0.331209, 0.0827338, ..., 0.529453, -0.582141, -0.165316],
|
||||
[-0.413614, 0.134572, 0.685769, ..., 0.0796088, 0.0217719, -0.118885]]], dtype=float32)
|
||||
[1, 23, 2560]
|
||||
|
||||
(HF) Overall Output of Inputs:
|
||||
|
||||
tensor([[[ 6.4688, 5.1016, 1.9658, ..., -2.9043, -2.9043, -2.9043],
|
||||
[ 5.2188, 6.4414, 5.1914, ..., -0.1852, -0.1862, -0.1866],
|
||||
[ 4.3516, 5.3281, 5.9922, ..., -0.3689, -0.3699, -0.3696],
|
||||
...,
|
||||
[10.4141, 11.7031, 12.5859, ..., 0.7778, 0.7769, 0.7754],
|
||||
[10.7188, 11.7891, 13.3125, ..., 1.6123, 1.6113, 1.6104],
|
||||
[10.8047, 12.0234, 12.4375, ..., 0.2321, 0.2314, 0.2317]]],
|
||||
|
||||
(MLX) Overall Output of Inputs:
|
||||
|
||||
array([[[6.46632, 5.10102, 1.96306, ..., -2.90427, -2.90341, -2.90392],
|
||||
[4.5092, 5.90938, 4.98036, ..., -0.411165, -0.412062, -0.412547],
|
||||
[4.34246, 5.7794, 6.13245, ..., -0.40106, -0.402052, -0.401838],
|
||||
...,
|
||||
[6.61827, 10.4022, 12.1672, ..., 0.602787, 0.602138, 0.600666],
|
||||
[7.96546, 12.9569, 14.7947, ..., -0.347764, -0.348587, -0.34937],
|
||||
[8.22272, 10.6631, 11.5968, ..., -1.12037, -1.12025, -1.12152]]], dtype=float32)
|
||||
Loading…
Reference in New Issue
Block a user