Rename project to InfuseLoop and update code and project links
Browse files- README.md +8 -8
- config.json +2 -2
README.md
CHANGED
|
@@ -5,39 +5,39 @@ license: apache-2.0
|
|
| 5 |
library_name: pytorch
|
| 6 |
pipeline_tag: text-generation
|
| 7 |
tags:
|
| 8 |
-
-
|
| 9 |
- looped-transformer
|
| 10 |
- causal-lm
|
| 11 |
- research
|
| 12 |
---
|
| 13 |
|
| 14 |
-
#
|
| 15 |
|
| 16 |
A 596,139,465-parameter base language model with 4 parallel recurrent loops and causal cross-trajectory Attention Residuals. Trained on 6,000,082,944 FineWeb-Edu tokens (45,777 updates).
|
| 17 |
|
| 18 |
-
**Code:** [thuEcstasy/
|
| 19 |
|
| 20 |
## Usage
|
| 21 |
|
| 22 |
Install the custom reference implementation:
|
| 23 |
|
| 24 |
```bash
|
| 25 |
-
pip install git+https://github.com/thuEcstasy/
|
| 26 |
```
|
| 27 |
|
| 28 |
```python
|
| 29 |
from transformers import AutoTokenizer
|
| 30 |
-
from
|
| 31 |
-
repo = "thuEcstasy1/
|
| 32 |
tokenizer = AutoTokenizer.from_pretrained(repo)
|
| 33 |
-
model =
|
| 34 |
prompt = "The scientist repeated the experiment and found that"
|
| 35 |
x = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).input_ids.cuda()
|
| 36 |
y = model.generate(x, max_new_tokens=32, eos_token_id=tokenizer.eos_token_id)
|
| 37 |
print(tokenizer.decode(y[0, x.shape[1]:], skip_special_tokens=True))
|
| 38 |
```
|
| 39 |
|
| 40 |
-
This is a base text model, **not an instruction-tuned assistant**. Use natural prefixes without a chat template. The custom `
|
| 41 |
|
| 42 |
The public API is a readable eager PyTorch reference. It does not reproduce the optimized CUDA Graph / FA4 / Triton latency measurements in the paper.
|
| 43 |
|
|
|
|
| 5 |
library_name: pytorch
|
| 6 |
pipeline_tag: text-generation
|
| 7 |
tags:
|
| 8 |
+
- infuseloop
|
| 9 |
- looped-transformer
|
| 10 |
- causal-lm
|
| 11 |
- research
|
| 12 |
---
|
| 13 |
|
| 14 |
+
# InfuseLoop-4
|
| 15 |
|
| 16 |
A 596,139,465-parameter base language model with 4 parallel recurrent loops and causal cross-trajectory Attention Residuals. Trained on 6,000,082,944 FineWeb-Edu tokens (45,777 updates).
|
| 17 |
|
| 18 |
+
**Project:** [InfuseLoop](https://infuseloop-demo.regal-vole-5301.chatgpt.site) 路 **Code:** [thuEcstasy/InfuseLoop](https://github.com/thuEcstasy/InfuseLoop).
|
| 19 |
|
| 20 |
## Usage
|
| 21 |
|
| 22 |
Install the custom reference implementation:
|
| 23 |
|
| 24 |
```bash
|
| 25 |
+
pip install git+https://github.com/thuEcstasy/InfuseLoop.git
|
| 26 |
```
|
| 27 |
|
| 28 |
```python
|
| 29 |
from transformers import AutoTokenizer
|
| 30 |
+
from infuseloop import InfuseLoop
|
| 31 |
+
repo = "thuEcstasy1/InfuseLoop-4"
|
| 32 |
tokenizer = AutoTokenizer.from_pretrained(repo)
|
| 33 |
+
model = InfuseLoop.from_pretrained(repo, device="cuda")
|
| 34 |
prompt = "The scientist repeated the experiment and found that"
|
| 35 |
x = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).input_ids.cuda()
|
| 36 |
y = model.generate(x, max_new_tokens=32, eos_token_id=tokenizer.eos_token_id)
|
| 37 |
print(tokenizer.decode(y[0, x.shape[1]:], skip_special_tokens=True))
|
| 38 |
```
|
| 39 |
|
| 40 |
+
This is a base text model, **not an instruction-tuned assistant**. Use natural prefixes without a chat template. The custom `InfuseLoop` class is required; this repository does not register `AutoModelForCausalLM` or promise compatibility with generic Transformers generation pipelines.
|
| 41 |
|
| 42 |
The public API is a readable eager PyTorch reference. It does not reproduce the optimized CUDA Graph / FA4 / Triton latency measurements in the paper.
|
| 43 |
|
config.json
CHANGED
|
@@ -1,7 +1,7 @@
|
|
| 1 |
{
|
| 2 |
-
"model_type": "
|
| 3 |
"architectures": [
|
| 4 |
-
"
|
| 5 |
],
|
| 6 |
"checkpoint_format": "foresightloop-safetensors-v1",
|
| 7 |
"backbone_config": {
|
|
|
|
| 1 |
{
|
| 2 |
+
"model_type": "infuseloop",
|
| 3 |
"architectures": [
|
| 4 |
+
"InfuseLoop"
|
| 5 |
],
|
| 6 |
"checkpoint_format": "foresightloop-safetensors-v1",
|
| 7 |
"backbone_config": {
|