Merge branch 'main' into flash-optimum
Browse files- FAQS.md +3 -0
- README.md +26 -10
- configs/accelerate/default_config.yaml +0 -15
- configs/cerebras_1_3B_alpaca.yml +0 -40
- configs/galactica_1_3B.yml +0 -41
- configs/llama_13B_alpaca.yml +0 -39
- configs/llama_65B_alpaca.yml +0 -44
- configs/llama_7B_4bit.yml +0 -45
- configs/llama_7B_alpaca.yml +0 -41
- configs/quickstart.yml +0 -45
- configs/sample.yml +0 -87
- configs/vicuna_13B_4bit_reflect.yml +0 -45
- examples/cerebras/qlora.yml +60 -0
- examples/falcon/config-7b-lora.yml +1 -1
- examples/falcon/config-7b.yml +1 -1
- configs/stability_3b.yml → examples/gptj/qlora.yml +29 -28
- examples/gptq-lora-7b/README.md +1 -1
- configs/llama_7B_jeopardy.yml → examples/jeopardy-bot/config.yml +11 -14
- examples/openllama-3b/README.md +16 -0
- examples/openllama-3b/config.yml +61 -0
- examples/{lora-openllama-3b/config.yml → openllama-3b/lora.yml} +3 -3
- examples/{qlora-openllama-3b/config.yml → openllama-3b/qlora.yml} +2 -2
- configs/pythia_1_2B_alpaca.yml → examples/pythia/lora.yml +11 -15
- examples/qlora-openllama-3b/README.md +0 -6
- scripts/finetune.py +31 -10
- src/axolotl/datasets.py +4 -0
- src/axolotl/monkeypatch/llama_landmark_attn.py +56 -402
- src/axolotl/prompt_strategies/sharegpt_jokes.py +28 -0
- src/axolotl/prompt_strategies/sharegpt_simple.py +21 -0
- src/axolotl/prompters.py +19 -14
- src/axolotl/utils/data.py +9 -2
- src/axolotl/utils/models.py +14 -26
- src/axolotl/utils/trainer.py +7 -4
- src/axolotl/utils/validation.py +5 -0
- tests/test_validation.py +14 -0
FAQS.md
CHANGED
@@ -2,3 +2,6 @@
|
|
2 |
|
3 |
- Can you train StableLM with this? Yes, but only with a single GPU atm. Multi GPU support is coming soon! Just waiting on this [PR](https://github.com/huggingface/transformers/pull/22874)
|
4 |
- Will this work with Deepspeed? That's still a WIP, but setting `export ACCELERATE_USE_DEEPSPEED=true` should work in some cases
|
|
|
|
|
|
|
|
2 |
|
3 |
- Can you train StableLM with this? Yes, but only with a single GPU atm. Multi GPU support is coming soon! Just waiting on this [PR](https://github.com/huggingface/transformers/pull/22874)
|
4 |
- Will this work with Deepspeed? That's still a WIP, but setting `export ACCELERATE_USE_DEEPSPEED=true` should work in some cases
|
5 |
+
- `Error invalid argument at line 359 in file /workspace/bitsandbytes/csrc/pythonInterface.c`
|
6 |
+
`/arrow/cpp/src/arrow/filesystem/s3fs.cc:2598: arrow::fs::FinalizeS3 was not called even though S3 was initialized.`
|
7 |
+
This could lead to a segmentation fault at exit. Try reinstalling bitsandbytes and transformers from source.
|
README.md
CHANGED
@@ -16,13 +16,14 @@
|
|
16 |
|
17 |
## Axolotl supports
|
18 |
|
19 |
-
|
|
20 |
-
|
21 |
-
| llama
|
22 |
-
| Pythia
|
23 |
-
| cerebras | ✅ | ✅
|
24 |
-
| mpt
|
25 |
-
| falcon
|
|
|
26 |
|
27 |
|
28 |
## Quickstart ⚡
|
@@ -38,10 +39,10 @@ pip3 install -U git+https://github.com/huggingface/peft.git
|
|
38 |
accelerate config
|
39 |
|
40 |
# finetune lora
|
41 |
-
accelerate launch scripts/finetune.py examples/
|
42 |
|
43 |
# inference
|
44 |
-
accelerate launch scripts/finetune.py examples/
|
45 |
--inference --lora_model_dir="./lora-out"
|
46 |
```
|
47 |
|
@@ -218,6 +219,14 @@ Have dataset(s) in one of the following format (JSONL recommended):
|
|
218 |
```json
|
219 |
{"conversations": [{"role": "...", "value": "..."}]}
|
220 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
221 |
|
222 |
</details>
|
223 |
|
@@ -381,6 +390,8 @@ num_epochs: 3
|
|
381 |
warmup_steps: 100
|
382 |
learning_rate: 0.00003
|
383 |
logging_steps:
|
|
|
|
|
384 |
|
385 |
# whether to mask out or include the human's prompt from the training labels
|
386 |
train_on_inputs: false
|
@@ -497,6 +508,11 @@ Pass the appropriate flag to the train command:
|
|
497 |
```bash
|
498 |
--inference --base_model ./completed-model
|
499 |
```
|
|
|
|
|
|
|
|
|
|
|
500 |
|
501 |
### Merge LORA to base
|
502 |
|
@@ -524,7 +540,7 @@ Try set `fp16: true`
|
|
524 |
|
525 |
Try to turn off xformers.
|
526 |
|
527 |
-
## Need help?
|
528 |
|
529 |
Join our [Discord server](https://discord.gg/HhrNrHJPRb) where we can help you
|
530 |
|
|
|
16 |
|
17 |
## Axolotl supports
|
18 |
|
19 |
+
| | fp16/fp32 | lora | qlora | gptq | gptq w/ lora | gptq w/flash attn | flash attn | xformers attn |
|
20 |
+
|----------|:----------|:-----|-------|------|:-------------|-------------------|------------|---------------|
|
21 |
+
| llama | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
|
22 |
+
| Pythia | ✅ | ✅ | ✅ | ❌ | ❓ | ❌ | ❌ | ❓ |
|
23 |
+
| cerebras | ✅ | ✅ | ✅ | ❌ | ❓ | ❌ | ❌ | ✅ |
|
24 |
+
| mpt | ✅ | ❌ | ❓ | ❌ | ❓ | ❌ | ❌ | ❓ |
|
25 |
+
| falcon | ✅ | ✅ | ✅ | ❌ | ❓ | ❌ | ❌ | ✅ |
|
26 |
+
| gpt-j | ✅ | ✅ | ✅ | ❌ | ❓ | ❌ | ❓ | ✅ |
|
27 |
|
28 |
|
29 |
## Quickstart ⚡
|
|
|
39 |
accelerate config
|
40 |
|
41 |
# finetune lora
|
42 |
+
accelerate launch scripts/finetune.py examples/openllama-3b/lora.yml
|
43 |
|
44 |
# inference
|
45 |
+
accelerate launch scripts/finetune.py examples/openllama-3b/lora.yml \
|
46 |
--inference --lora_model_dir="./lora-out"
|
47 |
```
|
48 |
|
|
|
219 |
```json
|
220 |
{"conversations": [{"role": "...", "value": "..."}]}
|
221 |
```
|
222 |
+
- `sharegpt_simple.load_role`: conversations where `role` is used instead of `from`
|
223 |
+
```json
|
224 |
+
{"conversations": [{"role": "...", "value": "..."}]}
|
225 |
+
```
|
226 |
+
- `sharegpt_jokes`: creates a chat where bot is asked to tell a joke, then explain why the joke is funny
|
227 |
+
```json
|
228 |
+
{"conversations": [{"title": "...", "text": "...", "explanation": "..."}]}
|
229 |
+
```
|
230 |
|
231 |
</details>
|
232 |
|
|
|
390 |
warmup_steps: 100
|
391 |
learning_rate: 0.00003
|
392 |
logging_steps:
|
393 |
+
save_steps:
|
394 |
+
eval_steps:
|
395 |
|
396 |
# whether to mask out or include the human's prompt from the training labels
|
397 |
train_on_inputs: false
|
|
|
508 |
```bash
|
509 |
--inference --base_model ./completed-model
|
510 |
```
|
511 |
+
- Full weights finetune w/ a prompt from a text file:
|
512 |
+
```bash
|
513 |
+
cat /tmp/prompt.txt | python scripts/finetune.py configs/your_config.yml \
|
514 |
+
--base_model ./completed-model --inference --prompter=None --load_in_8bit=True
|
515 |
+
```
|
516 |
|
517 |
### Merge LORA to base
|
518 |
|
|
|
540 |
|
541 |
Try to turn off xformers.
|
542 |
|
543 |
+
## Need help? 🙋♂️
|
544 |
|
545 |
Join our [Discord server](https://discord.gg/HhrNrHJPRb) where we can help you
|
546 |
|
configs/accelerate/default_config.yaml
DELETED
@@ -1,15 +0,0 @@
|
|
1 |
-
compute_environment: LOCAL_MACHINE
|
2 |
-
distributed_type: 'NO'
|
3 |
-
downcast_bf16: 'no'
|
4 |
-
gpu_ids: all
|
5 |
-
machine_rank: 0
|
6 |
-
main_training_function: main
|
7 |
-
mixed_precision: bf16
|
8 |
-
num_machines: 1
|
9 |
-
num_processes: 1
|
10 |
-
rdzv_backend: static
|
11 |
-
same_network: true
|
12 |
-
tpu_env: []
|
13 |
-
tpu_use_cluster: false
|
14 |
-
tpu_use_sudo: false
|
15 |
-
use_cpu: false
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
configs/cerebras_1_3B_alpaca.yml
DELETED
@@ -1,40 +0,0 @@
|
|
1 |
-
base_model: cerebras/Cerebras-GPT-1.3B
|
2 |
-
model_type: AutoModelForCausalLM
|
3 |
-
tokenizer_type: AutoTokenizer
|
4 |
-
load_in_8bit: true
|
5 |
-
datasets:
|
6 |
-
- path: data/alpaca_data_gpt4.jsonl
|
7 |
-
type: alpaca
|
8 |
-
- path: data/vicuna_cleaned.jsonl
|
9 |
-
type: sharegpt
|
10 |
-
- path: data/gpt4-instruct-similarity-0.6-dataset.jsonl
|
11 |
-
type: gpteacher
|
12 |
-
- path: data/roleplay-similarity_0.6-instruct-dataset.jsonl
|
13 |
-
type: gpteacher
|
14 |
-
dataset_prepared_path: last_run_prepared
|
15 |
-
val_set_size: 0.05
|
16 |
-
adapter: lora
|
17 |
-
sequence_len: 2048
|
18 |
-
lora_r: 8
|
19 |
-
lora_alpha: 16
|
20 |
-
lora_dropout: 0.05
|
21 |
-
lora_target_modules:
|
22 |
-
- c_attn
|
23 |
-
lora_fan_in_fan_out: false
|
24 |
-
wandb_project: pythia-1.4b-lora
|
25 |
-
wandb_watch:
|
26 |
-
wandb_run_id:
|
27 |
-
wandb_log_model:
|
28 |
-
output_dir: ./lora-alpaca
|
29 |
-
gradient_accumulation_steps: 1
|
30 |
-
micro_batch_size: 4
|
31 |
-
num_epochs: 5
|
32 |
-
learning_rate: 0.0003
|
33 |
-
train_on_inputs: false
|
34 |
-
group_by_length: false
|
35 |
-
bf16: True
|
36 |
-
tf32: True
|
37 |
-
gradient_checkpointing:
|
38 |
-
early_stopping_patience:
|
39 |
-
resume_from_checkpoint:
|
40 |
-
local_rank:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
configs/galactica_1_3B.yml
DELETED
@@ -1,41 +0,0 @@
|
|
1 |
-
base_model: facebook/galactica-1.3b
|
2 |
-
model_type: AutoModelForCausalLM
|
3 |
-
tokenizer_type: AutoTokenizer
|
4 |
-
load_in_8bit: false
|
5 |
-
datasets:
|
6 |
-
- path: tatsu-lab/alpaca
|
7 |
-
type: alpaca
|
8 |
-
dataset_prepared_path: last_run_prepared
|
9 |
-
val_set_size: 0.1
|
10 |
-
adapter:
|
11 |
-
lora_model_dir:
|
12 |
-
sequence_len: 1024
|
13 |
-
max_packed_sequence_len: 1024
|
14 |
-
lora_r: 8
|
15 |
-
lora_alpha: 16
|
16 |
-
lora_dropout: 0.05
|
17 |
-
lora_target_modules:
|
18 |
-
- q_proj
|
19 |
-
- v_proj
|
20 |
-
lora_fan_in_fan_out: false
|
21 |
-
wandb_project:
|
22 |
-
wandb_watch:
|
23 |
-
wandb_run_id:
|
24 |
-
wandb_log_model:
|
25 |
-
output_dir: ./lora-llama-alpaca
|
26 |
-
gradient_accumulation_steps: 1
|
27 |
-
micro_batch_size: 16
|
28 |
-
num_epochs: 3
|
29 |
-
learning_rate: 0.00003
|
30 |
-
train_on_inputs: false
|
31 |
-
group_by_length: false
|
32 |
-
bf16: false
|
33 |
-
tf32: false
|
34 |
-
early_stopping_patience:
|
35 |
-
resume_from_checkpoint:
|
36 |
-
local_rank:
|
37 |
-
tokens:
|
38 |
-
pad_token: "[PAD]"
|
39 |
-
bos_token: "<s>"
|
40 |
-
eos_token: "</s>"
|
41 |
-
unk_token: "<unk>"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
configs/llama_13B_alpaca.yml
DELETED
@@ -1,39 +0,0 @@
|
|
1 |
-
base_model: huggyllama/llama-13b
|
2 |
-
model_type: LlamaForCausalLM
|
3 |
-
tokenizer_type: LlamaTokenizer
|
4 |
-
load_in_8bit: true
|
5 |
-
datasets:
|
6 |
-
- path: anon8231489123/ShareGPT_Vicuna_unfiltered
|
7 |
-
data_files: ShareGPT_V3_unfiltered_cleaned_split_no_imsorry.json
|
8 |
-
type: sharegpt
|
9 |
-
dataset_prepared_path: last_run_prepared
|
10 |
-
val_set_size: 0.002
|
11 |
-
adapter:
|
12 |
-
lora_model_dir:
|
13 |
-
sequence_len: 2048
|
14 |
-
lora_r: 8
|
15 |
-
lora_alpha: 16
|
16 |
-
lora_dropout: 0.05
|
17 |
-
lora_target_modules:
|
18 |
-
- q_proj
|
19 |
-
- v_proj
|
20 |
-
lora_fan_in_fan_out: false
|
21 |
-
wandb_project:
|
22 |
-
wandb_watch:
|
23 |
-
wandb_run_id:
|
24 |
-
wandb_log_model:
|
25 |
-
output_dir: ./llama-13b-sharegpt
|
26 |
-
gradient_accumulation_steps: 1
|
27 |
-
micro_batch_size: 2
|
28 |
-
warmup_steps: 1000
|
29 |
-
save_steps:
|
30 |
-
eval_steps:
|
31 |
-
num_epochs: 5
|
32 |
-
learning_rate: 0.00003
|
33 |
-
train_on_inputs: false
|
34 |
-
group_by_length: false
|
35 |
-
bf16: true
|
36 |
-
tf32: true
|
37 |
-
early_stopping_patience: 5
|
38 |
-
resume_from_checkpoint:
|
39 |
-
local_rank:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
configs/llama_65B_alpaca.yml
DELETED
@@ -1,44 +0,0 @@
|
|
1 |
-
base_model: huggyllama/llama-65b
|
2 |
-
model_type: LlamaForCausalLM
|
3 |
-
tokenizer_type: LlamaTokenizer
|
4 |
-
load_in_8bit: true
|
5 |
-
datasets:
|
6 |
-
- path: data/alpaca_data_gpt4.jsonl
|
7 |
-
type: alpaca
|
8 |
-
- path: anon8231489123/ShareGPT_Vicuna_unfiltered
|
9 |
-
data_files: ShareGPT_V3_unfiltered_cleaned_split_no_imsorry.json
|
10 |
-
type: sharegpt
|
11 |
-
- path: data/gpt4-instruct-similarity-0.6-dataset.jsonl
|
12 |
-
type: gpteacher
|
13 |
-
- path: data/roleplay-similarity_0.6-instruct-dataset.jsonl
|
14 |
-
type: gpteacher
|
15 |
-
dataset_prepared_path: last_run_prepared
|
16 |
-
val_set_size: 0.04
|
17 |
-
adapter: lora
|
18 |
-
lora_model_dir:
|
19 |
-
sequence_len: 2048
|
20 |
-
lora_r: 8
|
21 |
-
lora_alpha: 16
|
22 |
-
lora_dropout: 0.05
|
23 |
-
lora_target_modules:
|
24 |
-
- q_proj
|
25 |
-
- v_proj
|
26 |
-
lora_fan_in_fan_out: false
|
27 |
-
wandb_project: llama-65b-lora
|
28 |
-
wandb_watch:
|
29 |
-
wandb_run_id:
|
30 |
-
wandb_log_model:
|
31 |
-
output_dir: ./lora-llama-alpaca
|
32 |
-
gradient_accumulation_steps: 1
|
33 |
-
micro_batch_size: 16
|
34 |
-
warmup_steps: 1000
|
35 |
-
save_steps:
|
36 |
-
num_epochs: 5
|
37 |
-
learning_rate: 0.00003
|
38 |
-
train_on_inputs: false
|
39 |
-
group_by_length: false
|
40 |
-
bf16: true
|
41 |
-
tf32: true
|
42 |
-
early_stopping_patience:
|
43 |
-
resume_from_checkpoint:
|
44 |
-
local_rank:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
configs/llama_7B_4bit.yml
DELETED
@@ -1,45 +0,0 @@
|
|
1 |
-
base_model: decapoda-research/llama-7b-hf-int4
|
2 |
-
base_model_config: decapoda-research/llama-7b-hf
|
3 |
-
model_type: LlamaForCausalLM
|
4 |
-
tokenizer_type: LlamaTokenizer
|
5 |
-
load_in_8bit: true
|
6 |
-
datasets:
|
7 |
-
- path: tatsu-lab/alpaca # original alpaca dataset
|
8 |
-
type: alpaca
|
9 |
-
dataset_prepared_path: data/last_run_prepared
|
10 |
-
val_set_size: 0.04
|
11 |
-
adapter: lora
|
12 |
-
lora_model_dir:
|
13 |
-
sequence_len: 2048
|
14 |
-
max_packed_sequence_len: 1024
|
15 |
-
lora_r: 8
|
16 |
-
lora_alpha: 16
|
17 |
-
lora_dropout: 0.05
|
18 |
-
lora_target_modules:
|
19 |
-
- q_proj
|
20 |
-
- v_proj
|
21 |
-
# - k_proj
|
22 |
-
# - o_proj
|
23 |
-
lora_fan_in_fan_out: false
|
24 |
-
wandb_project:
|
25 |
-
wandb_watch:
|
26 |
-
wandb_run_id:
|
27 |
-
wandb_log_model:
|
28 |
-
output_dir: ./lora-test
|
29 |
-
gradient_accumulation_steps: 1
|
30 |
-
micro_batch_size: 2
|
31 |
-
num_epochs: 3
|
32 |
-
warmup_steps: 100
|
33 |
-
learning_rate: 0.00003
|
34 |
-
train_on_inputs: false
|
35 |
-
group_by_length: false
|
36 |
-
bf16: true
|
37 |
-
tf32: true
|
38 |
-
gradient_checkpointing: false
|
39 |
-
early_stopping_patience: 3
|
40 |
-
resume_from_checkpoint:
|
41 |
-
auto_resume_from_checkpoints: true
|
42 |
-
local_rank:
|
43 |
-
load_4bit: true
|
44 |
-
xformers_attention: true
|
45 |
-
flash_attention:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
configs/llama_7B_alpaca.yml
DELETED
@@ -1,41 +0,0 @@
|
|
1 |
-
base_model: huggyllama/llama-7b
|
2 |
-
model_type: LlamaForCausalLM
|
3 |
-
tokenizer_type: LlamaTokenizer
|
4 |
-
load_in_8bit: true
|
5 |
-
datasets:
|
6 |
-
- path: data/alpaca_data_gpt4.jsonl
|
7 |
-
type: alpaca
|
8 |
-
- path: data/vicuna_cleaned.jsonl
|
9 |
-
type: sharegpt
|
10 |
-
- path: data/gpt4-instruct-similarity-0.6-dataset.jsonl
|
11 |
-
type: gpteacher
|
12 |
-
- path: data/roleplay-similarity_0.6-instruct-dataset.jsonl
|
13 |
-
type: gpteacher
|
14 |
-
dataset_prepared_path: last_run_prepared
|
15 |
-
val_set_size: 0.04
|
16 |
-
adapter: lora
|
17 |
-
lora_model_dir:
|
18 |
-
sequence_len: 2048
|
19 |
-
lora_r: 8
|
20 |
-
lora_alpha: 16
|
21 |
-
lora_dropout: 0.05
|
22 |
-
lora_target_modules:
|
23 |
-
- q_proj
|
24 |
-
- v_proj
|
25 |
-
lora_fan_in_fan_out: false
|
26 |
-
wandb_project: llama-7b-lora
|
27 |
-
wandb_watch:
|
28 |
-
wandb_run_id:
|
29 |
-
wandb_log_model:
|
30 |
-
output_dir: ./lora-llama-alpaca
|
31 |
-
gradient_accumulation_steps: 1
|
32 |
-
micro_batch_size: 16
|
33 |
-
num_epochs: 5
|
34 |
-
learning_rate: 0.00003
|
35 |
-
train_on_inputs: false
|
36 |
-
group_by_length: false
|
37 |
-
bf16: true
|
38 |
-
tf32: true
|
39 |
-
early_stopping_patience:
|
40 |
-
resume_from_checkpoint:
|
41 |
-
local_rank:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
configs/quickstart.yml
DELETED
@@ -1,45 +0,0 @@
|
|
1 |
-
base_model: decapoda-research/llama-7b-hf-int4
|
2 |
-
base_model_config: decapoda-research/llama-7b-hf
|
3 |
-
model_type: LlamaForCausalLM
|
4 |
-
tokenizer_type: LlamaTokenizer
|
5 |
-
load_in_8bit: true
|
6 |
-
datasets:
|
7 |
-
- path: tatsu-lab/alpaca # original alpaca dataset
|
8 |
-
type: alpaca
|
9 |
-
dataset_prepared_path: data/last_run_prepared
|
10 |
-
val_set_size: 0.04
|
11 |
-
adapter: lora
|
12 |
-
lora_model_dir:
|
13 |
-
sequence_len: 1024
|
14 |
-
max_packed_sequence_len: 1024
|
15 |
-
lora_r: 8
|
16 |
-
lora_alpha: 16
|
17 |
-
lora_dropout: 0.05
|
18 |
-
lora_target_modules:
|
19 |
-
- q_proj
|
20 |
-
- v_proj
|
21 |
-
# - k_proj
|
22 |
-
# - o_proj
|
23 |
-
lora_fan_in_fan_out: false
|
24 |
-
wandb_project:
|
25 |
-
wandb_watch:
|
26 |
-
wandb_run_id:
|
27 |
-
wandb_log_model:
|
28 |
-
output_dir: ./lora-test
|
29 |
-
gradient_accumulation_steps: 1
|
30 |
-
micro_batch_size: 1
|
31 |
-
num_epochs: 3
|
32 |
-
warmup_steps: 100
|
33 |
-
learning_rate: 0.00003
|
34 |
-
train_on_inputs: false
|
35 |
-
group_by_length: false
|
36 |
-
bf16: true
|
37 |
-
tf32: true
|
38 |
-
gradient_checkpointing: false
|
39 |
-
early_stopping_patience: 3
|
40 |
-
resume_from_checkpoint:
|
41 |
-
auto_resume_from_checkpoints: true
|
42 |
-
local_rank:
|
43 |
-
gptq: true
|
44 |
-
xformers_attention: true
|
45 |
-
flash_attention:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
configs/sample.yml
DELETED
@@ -1,87 +0,0 @@
|
|
1 |
-
# this is the huggingface model that contains *.pt, *.safetensors, or *.bin files
|
2 |
-
# this can also be a relative path to a model on disk
|
3 |
-
base_model: decapoda-research/llama-7b-hf-int4
|
4 |
-
# you can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)
|
5 |
-
base_model_ignore_patterns:
|
6 |
-
# if the base_model repo on hf hub doesn't include configuration .json files,
|
7 |
-
# you can set that here, or leave this empty to default to base_model
|
8 |
-
base_model_config: decapoda-research/llama-7b-hf
|
9 |
-
# If you want to specify the type of model to load, AutoModelForCausalLM is a good choice too
|
10 |
-
model_type: AutoModelForCausalLM
|
11 |
-
# Corresponding tokenizer for the model AutoTokenizer is a good choice
|
12 |
-
tokenizer_type: AutoTokenizer
|
13 |
-
# whether you are training a 4-bit quantized model
|
14 |
-
load_4bit: true
|
15 |
-
# this will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer
|
16 |
-
load_in_8bit: true
|
17 |
-
# a list of one or more datasets to finetune the model with
|
18 |
-
datasets:
|
19 |
-
# this can be either a hf dataset, or relative path
|
20 |
-
- path: vicgalle/alpaca-gpt4
|
21 |
-
# The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection]
|
22 |
-
type: alpaca
|
23 |
-
# axolotl attempts to save the dataset as an arrow after packing the data together so
|
24 |
-
# subsequent training attempts load faster, relative path
|
25 |
-
dataset_prepared_path: data/last_run_prepared
|
26 |
-
# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc
|
27 |
-
val_set_size: 0.04
|
28 |
-
# if you want to use lora, leave blank to train all parameters in original model
|
29 |
-
adapter: lora
|
30 |
-
# if you already have a lora model trained that you want to load, put that here
|
31 |
-
lora_model_dir:
|
32 |
-
# the maximum length of an input to train with, this should typically be less than 2048
|
33 |
-
# as most models have a token/context limit of 2048
|
34 |
-
sequence_len: 2048
|
35 |
-
# max sequence length to concatenate training samples together up to
|
36 |
-
# inspired by StackLLaMA. see https://huggingface.co/blog/stackllama#supervised-fine-tuning
|
37 |
-
max_packed_sequence_len: 1024
|
38 |
-
# lora hyperparameters
|
39 |
-
lora_r: 8
|
40 |
-
lora_alpha: 16
|
41 |
-
lora_dropout: 0.05
|
42 |
-
lora_target_modules:
|
43 |
-
- q_proj
|
44 |
-
- v_proj
|
45 |
-
# - k_proj
|
46 |
-
# - o_proj
|
47 |
-
lora_fan_in_fan_out: false
|
48 |
-
# wandb configuration if your're using it
|
49 |
-
wandb_project:
|
50 |
-
wandb_watch:
|
51 |
-
wandb_run_id:
|
52 |
-
wandb_log_model:
|
53 |
-
# where to save the finsihed model to
|
54 |
-
output_dir: ./completed-model
|
55 |
-
# training hyperparameters
|
56 |
-
gradient_accumulation_steps: 1
|
57 |
-
batch_size:
|
58 |
-
micro_batch_size: 2
|
59 |
-
num_epochs: 3
|
60 |
-
warmup_steps: 100
|
61 |
-
learning_rate: 0.00003
|
62 |
-
# whether to mask out or include the human's prompt from the training labels
|
63 |
-
train_on_inputs: false
|
64 |
-
# don't use this, leads to wonky training (according to someone on the internet)
|
65 |
-
group_by_length: false
|
66 |
-
# Use CUDA bf16
|
67 |
-
bf16: true
|
68 |
-
# Use CUDA tf32
|
69 |
-
tf32: true
|
70 |
-
# does not work with current implementation of 4-bit LoRA
|
71 |
-
gradient_checkpointing: false
|
72 |
-
# stop training after this many evaluation losses have increased in a row
|
73 |
-
# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback
|
74 |
-
early_stopping_patience: 3
|
75 |
-
# specify a scheduler to use with the optimizer. only one_cycle is supported currently
|
76 |
-
lr_scheduler:
|
77 |
-
# whether to use xformers attention patch https://github.com/facebookresearch/xformers:
|
78 |
-
xformers_attention:
|
79 |
-
# whether to use flash attention patch https://github.com/HazyResearch/flash-attention:
|
80 |
-
flash_attention:
|
81 |
-
# resume from a specific checkpoint dir
|
82 |
-
resume_from_checkpoint:
|
83 |
-
# if resume_from_checkpoint isn't set and you simply want it to start where it left off
|
84 |
-
# be careful with this being turned on between different models
|
85 |
-
auto_resume_from_checkpoints: false
|
86 |
-
# don't mess with this, it's here for accelerate and torchrun
|
87 |
-
local_rank:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
configs/vicuna_13B_4bit_reflect.yml
DELETED
@@ -1,45 +0,0 @@
|
|
1 |
-
base_model: anon8231489123/vicuna-13b-GPTQ-4bit-128g
|
2 |
-
base_model_config: anon8231489123/vicuna-13b-GPTQ-4bit-128g
|
3 |
-
model_type: LlamaForCausalLM
|
4 |
-
tokenizer_type: LlamaTokenizer
|
5 |
-
load_in_8bit: false
|
6 |
-
load_4bit: true
|
7 |
-
gptq_groupsize: 128
|
8 |
-
gptq_model_v1: false
|
9 |
-
datasets:
|
10 |
-
# https://github.com/vaguenebula/AlpacaDataReflect/blob/main/alpaca_reflect_pruned.json
|
11 |
-
- path: data/alpaca_reflect_pruned.jsonl
|
12 |
-
type: reflection
|
13 |
-
dataset_prepared_path: data/last_run_prepared
|
14 |
-
val_set_size: 0.04
|
15 |
-
adapter: lora
|
16 |
-
lora_model_dir:
|
17 |
-
sequence_len: 2048
|
18 |
-
max_packed_sequence_len: 2048
|
19 |
-
lora_r: 8
|
20 |
-
lora_alpha: 16
|
21 |
-
lora_dropout: 0.05
|
22 |
-
lora_target_modules:
|
23 |
-
- q_proj
|
24 |
-
- v_proj
|
25 |
-
# - k_proj
|
26 |
-
# - o_proj
|
27 |
-
lora_fan_in_fan_out: false
|
28 |
-
wandb_project:
|
29 |
-
wandb_watch:
|
30 |
-
wandb_run_id:
|
31 |
-
wandb_log_model:
|
32 |
-
output_dir: ./lora-reflect
|
33 |
-
gradient_accumulation_steps: 1
|
34 |
-
micro_batch_size: 2
|
35 |
-
num_epochs: 3
|
36 |
-
learning_rate: 0.00003
|
37 |
-
train_on_inputs: false
|
38 |
-
group_by_length: false
|
39 |
-
bf16: true
|
40 |
-
tf32: true
|
41 |
-
gradient_checkpointing: false
|
42 |
-
early_stopping_patience: 3
|
43 |
-
resume_from_checkpoint:
|
44 |
-
local_rank:
|
45 |
-
flash_attention: true
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
examples/cerebras/qlora.yml
ADDED
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
base_model: cerebras/Cerebras-GPT-1.3B
|
2 |
+
base_model_config: cerebras/Cerebras-GPT-1.3B
|
3 |
+
load_in_8bit: false
|
4 |
+
load_in_4bit: true
|
5 |
+
strict: false
|
6 |
+
push_dataset_to_hub:
|
7 |
+
datasets:
|
8 |
+
- path: teknium/GPT4-LLM-Cleaned
|
9 |
+
type: alpaca
|
10 |
+
dataset_prepared_path: last_run_prepared
|
11 |
+
val_set_size: 0.01
|
12 |
+
adapter: qlora
|
13 |
+
lora_model_dir:
|
14 |
+
sequence_len: 2048
|
15 |
+
max_packed_sequence_len: 2048
|
16 |
+
lora_r: 16
|
17 |
+
lora_alpha: 32
|
18 |
+
lora_dropout: 0.05
|
19 |
+
lora_target_modules:
|
20 |
+
- c_fc
|
21 |
+
- c_attn
|
22 |
+
- c_proj
|
23 |
+
lora_target_linear:
|
24 |
+
lora_fan_in_fan_out:
|
25 |
+
wandb_project:
|
26 |
+
wandb_watch:
|
27 |
+
wandb_run_id:
|
28 |
+
wandb_log_model:
|
29 |
+
output_dir: ./qlora-out
|
30 |
+
batch_size: 4
|
31 |
+
micro_batch_size: 4
|
32 |
+
num_epochs: 2
|
33 |
+
optimizer: paged_adamw_8bit
|
34 |
+
torchdistx_path:
|
35 |
+
lr_scheduler: cosine
|
36 |
+
learning_rate: 0.0002
|
37 |
+
train_on_inputs: false
|
38 |
+
group_by_length: true
|
39 |
+
bf16: true
|
40 |
+
fp16: false
|
41 |
+
tf32: true
|
42 |
+
gradient_checkpointing: true
|
43 |
+
early_stopping_patience:
|
44 |
+
resume_from_checkpoint:
|
45 |
+
local_rank:
|
46 |
+
logging_steps: 1
|
47 |
+
xformers_attention: true
|
48 |
+
flash_attention:
|
49 |
+
gptq_groupsize:
|
50 |
+
gptq_model_v1:
|
51 |
+
warmup_steps: 10
|
52 |
+
eval_steps: 20
|
53 |
+
save_steps:
|
54 |
+
debug:
|
55 |
+
deepspeed:
|
56 |
+
weight_decay: 0.1
|
57 |
+
fsdp:
|
58 |
+
fsdp_config:
|
59 |
+
special_tokens:
|
60 |
+
pad_token: "<|endoftext|>"
|
examples/falcon/config-7b-lora.yml
CHANGED
@@ -23,7 +23,7 @@ lora_dropout: 0.0
|
|
23 |
lora_target_modules:
|
24 |
lora_target_linear: true
|
25 |
lora_fan_in_fan_out:
|
26 |
-
wandb_project:
|
27 |
wandb_watch:
|
28 |
wandb_run_id:
|
29 |
wandb_log_model:
|
|
|
23 |
lora_target_modules:
|
24 |
lora_target_linear: true
|
25 |
lora_fan_in_fan_out:
|
26 |
+
wandb_project:
|
27 |
wandb_watch:
|
28 |
wandb_run_id:
|
29 |
wandb_log_model:
|
examples/falcon/config-7b.yml
CHANGED
@@ -23,7 +23,7 @@ lora_dropout: 0.0
|
|
23 |
lora_target_modules:
|
24 |
lora_target_linear: true
|
25 |
lora_fan_in_fan_out:
|
26 |
-
wandb_project:
|
27 |
wandb_watch:
|
28 |
wandb_run_id:
|
29 |
wandb_log_model:
|
|
|
23 |
lora_target_modules:
|
24 |
lora_target_linear: true
|
25 |
lora_fan_in_fan_out:
|
26 |
+
wandb_project:
|
27 |
wandb_watch:
|
28 |
wandb_run_id:
|
29 |
wandb_log_model:
|
configs/stability_3b.yml → examples/gptj/qlora.yml
RENAMED
@@ -1,38 +1,42 @@
|
|
1 |
-
base_model:
|
2 |
-
base_model_config:
|
3 |
load_in_8bit: false
|
|
|
|
|
|
|
4 |
datasets:
|
5 |
-
- path:
|
6 |
type: alpaca
|
7 |
dataset_prepared_path: last_run_prepared
|
8 |
-
val_set_size: 0.
|
9 |
-
adapter:
|
10 |
lora_model_dir:
|
11 |
-
sequence_len:
|
12 |
-
max_packed_sequence_len:
|
13 |
lora_r: 8
|
14 |
-
lora_alpha:
|
15 |
lora_dropout: 0.05
|
16 |
lora_target_modules:
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
wandb_project: stable-alpaca-3b
|
21 |
wandb_watch:
|
22 |
wandb_run_id:
|
23 |
wandb_log_model:
|
24 |
-
output_dir: ./
|
25 |
-
gradient_accumulation_steps:
|
26 |
-
micro_batch_size:
|
27 |
-
num_epochs:
|
28 |
-
optimizer:
|
29 |
torchdistx_path:
|
30 |
lr_scheduler: cosine
|
31 |
-
learning_rate: 0.
|
32 |
train_on_inputs: false
|
33 |
-
group_by_length:
|
34 |
bf16: true
|
|
|
35 |
tf32: true
|
|
|
36 |
early_stopping_patience:
|
37 |
resume_from_checkpoint:
|
38 |
local_rank:
|
@@ -41,16 +45,13 @@ xformers_attention: true
|
|
41 |
flash_attention:
|
42 |
gptq_groupsize:
|
43 |
gptq_model_v1:
|
44 |
-
warmup_steps:
|
45 |
-
eval_steps:
|
46 |
-
save_steps:
|
47 |
debug:
|
48 |
deepspeed:
|
49 |
-
weight_decay: 0.
|
50 |
fsdp:
|
51 |
fsdp_config:
|
52 |
-
|
53 |
-
|
54 |
-
# bos_token: "<s>"
|
55 |
-
# eos_token: "</s>"
|
56 |
-
# unk_token: "<unk>"
|
|
|
1 |
+
base_model: EleutherAI/gpt-j-6b
|
2 |
+
base_model_config: EleutherAI/gpt-j-6b
|
3 |
load_in_8bit: false
|
4 |
+
load_in_4bit: true
|
5 |
+
strict: false
|
6 |
+
push_dataset_to_hub:
|
7 |
datasets:
|
8 |
+
- path: teknium/GPT4-LLM-Cleaned
|
9 |
type: alpaca
|
10 |
dataset_prepared_path: last_run_prepared
|
11 |
+
val_set_size: 0.01
|
12 |
+
adapter: qlora
|
13 |
lora_model_dir:
|
14 |
+
sequence_len: 2048
|
15 |
+
max_packed_sequence_len:
|
16 |
lora_r: 8
|
17 |
+
lora_alpha: 32
|
18 |
lora_dropout: 0.05
|
19 |
lora_target_modules:
|
20 |
+
lora_target_linear: true
|
21 |
+
lora_fan_in_fan_out:
|
22 |
+
wandb_project:
|
|
|
23 |
wandb_watch:
|
24 |
wandb_run_id:
|
25 |
wandb_log_model:
|
26 |
+
output_dir: ./qlora-out
|
27 |
+
gradient_accumulation_steps: 2
|
28 |
+
micro_batch_size: 2
|
29 |
+
num_epochs: 2
|
30 |
+
optimizer: paged_adamw_8bit
|
31 |
torchdistx_path:
|
32 |
lr_scheduler: cosine
|
33 |
+
learning_rate: 0.0001
|
34 |
train_on_inputs: false
|
35 |
+
group_by_length: true
|
36 |
bf16: true
|
37 |
+
fp16: false
|
38 |
tf32: true
|
39 |
+
gradient_checkpointing: true
|
40 |
early_stopping_patience:
|
41 |
resume_from_checkpoint:
|
42 |
local_rank:
|
|
|
45 |
flash_attention:
|
46 |
gptq_groupsize:
|
47 |
gptq_model_v1:
|
48 |
+
warmup_steps: 10
|
49 |
+
eval_steps: 20
|
50 |
+
save_steps:
|
51 |
debug:
|
52 |
deepspeed:
|
53 |
+
weight_decay: 0.1
|
54 |
fsdp:
|
55 |
fsdp_config:
|
56 |
+
special_tokens:
|
57 |
+
pad_token: "<|endoftext|>"
|
|
|
|
|
|
examples/gptq-lora-7b/README.md
CHANGED
@@ -3,6 +3,6 @@
|
|
3 |
This is a good place to start for beginners. This will run on an NVIDIA RTX4090 with no other changes needed.
|
4 |
|
5 |
```shell
|
6 |
-
accelerate launch scripts/finetune.py examples/
|
7 |
|
8 |
```
|
|
|
3 |
This is a good place to start for beginners. This will run on an NVIDIA RTX4090 with no other changes needed.
|
4 |
|
5 |
```shell
|
6 |
+
accelerate launch scripts/finetune.py examples/gptq-lora-7b/config.yml
|
7 |
|
8 |
```
|
configs/llama_7B_jeopardy.yml → examples/jeopardy-bot/config.yml
RENAMED
@@ -7,30 +7,28 @@ datasets:
|
|
7 |
- path: openaccess-ai-collective/jeopardy
|
8 |
type: jeopardy
|
9 |
dataset_prepared_path: last_run_prepared
|
10 |
-
val_set_size: 0.
|
11 |
adapter:
|
12 |
lora_model_dir:
|
13 |
-
sequence_len:
|
14 |
-
max_packed_sequence_len:
|
15 |
-
lora_r:
|
16 |
-
lora_alpha:
|
17 |
-
lora_dropout:
|
18 |
lora_target_modules:
|
19 |
-
- q_proj
|
20 |
-
- v_proj
|
21 |
lora_fan_in_fan_out: false
|
22 |
-
wandb_project:
|
23 |
wandb_watch:
|
24 |
wandb_run_id:
|
25 |
wandb_log_model:
|
26 |
output_dir: ./jeopardy-bot-7b
|
27 |
-
gradient_accumulation_steps:
|
28 |
micro_batch_size: 1
|
29 |
-
num_epochs:
|
30 |
optimizer: adamw_bnb_8bit
|
31 |
torchdistx_path:
|
32 |
lr_scheduler: cosine
|
33 |
-
learning_rate: 0.
|
34 |
train_on_inputs: false
|
35 |
group_by_length: false
|
36 |
bf16: true
|
@@ -48,11 +46,10 @@ eval_steps: 110
|
|
48 |
save_steps: 660
|
49 |
debug:
|
50 |
deepspeed:
|
51 |
-
weight_decay: 0.
|
52 |
fsdp:
|
53 |
fsdp_config:
|
54 |
tokens:
|
55 |
-
pad_token: "[PAD]"
|
56 |
bos_token: "<s>"
|
57 |
eos_token: "</s>"
|
58 |
unk_token: "<unk>"
|
|
|
7 |
- path: openaccess-ai-collective/jeopardy
|
8 |
type: jeopardy
|
9 |
dataset_prepared_path: last_run_prepared
|
10 |
+
val_set_size: 0.02
|
11 |
adapter:
|
12 |
lora_model_dir:
|
13 |
+
sequence_len: 512
|
14 |
+
max_packed_sequence_len:
|
15 |
+
lora_r:
|
16 |
+
lora_alpha:
|
17 |
+
lora_dropout:
|
18 |
lora_target_modules:
|
|
|
|
|
19 |
lora_fan_in_fan_out: false
|
20 |
+
wandb_project:
|
21 |
wandb_watch:
|
22 |
wandb_run_id:
|
23 |
wandb_log_model:
|
24 |
output_dir: ./jeopardy-bot-7b
|
25 |
+
gradient_accumulation_steps: 1
|
26 |
micro_batch_size: 1
|
27 |
+
num_epochs: 3
|
28 |
optimizer: adamw_bnb_8bit
|
29 |
torchdistx_path:
|
30 |
lr_scheduler: cosine
|
31 |
+
learning_rate: 0.00003
|
32 |
train_on_inputs: false
|
33 |
group_by_length: false
|
34 |
bf16: true
|
|
|
46 |
save_steps: 660
|
47 |
debug:
|
48 |
deepspeed:
|
49 |
+
weight_decay: 0.1
|
50 |
fsdp:
|
51 |
fsdp_config:
|
52 |
tokens:
|
|
|
53 |
bos_token: "<s>"
|
54 |
eos_token: "</s>"
|
55 |
unk_token: "<unk>"
|
examples/openllama-3b/README.md
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# openllama-3b
|
2 |
+
|
3 |
+
Basic full tune
|
4 |
+
```shell
|
5 |
+
accelerate launch scripts/finetune.py examples/openllama-3b/config.yml
|
6 |
+
```
|
7 |
+
|
8 |
+
LoRA
|
9 |
+
```shell
|
10 |
+
accelerate launch scripts/finetune.py examples/openllama-3b/lora.yml
|
11 |
+
```
|
12 |
+
|
13 |
+
QLoRA
|
14 |
+
```shell
|
15 |
+
accelerate launch scripts/finetune.py examples/openllama-3b/qlora.yml
|
16 |
+
```
|
examples/openllama-3b/config.yml
ADDED
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
base_model: openlm-research/open_llama_3b
|
2 |
+
base_model_config: openlm-research/open_llama_3b
|
3 |
+
model_type: LlamaForCausalLM
|
4 |
+
tokenizer_type: LlamaTokenizer
|
5 |
+
load_in_8bit: false
|
6 |
+
load_in_4bit: false
|
7 |
+
strict: false
|
8 |
+
push_dataset_to_hub:
|
9 |
+
datasets:
|
10 |
+
- path: teknium/GPT4-LLM-Cleaned
|
11 |
+
type: alpaca
|
12 |
+
dataset_prepared_path: last_run_prepared
|
13 |
+
val_set_size: 0.02
|
14 |
+
adapter:
|
15 |
+
lora_model_dir:
|
16 |
+
sequence_len: 256
|
17 |
+
max_packed_sequence_len:
|
18 |
+
lora_r:
|
19 |
+
lora_alpha:
|
20 |
+
lora_dropout:
|
21 |
+
lora_target_modules:
|
22 |
+
lora_target_linear:
|
23 |
+
lora_fan_in_fan_out:
|
24 |
+
wandb_project:
|
25 |
+
wandb_watch:
|
26 |
+
wandb_run_id:
|
27 |
+
wandb_log_model:
|
28 |
+
output_dir: ./openllama-out
|
29 |
+
batch_size: 16
|
30 |
+
micro_batch_size: 4
|
31 |
+
num_epochs: 3
|
32 |
+
optimizer: adamw_bnb_8bit
|
33 |
+
torchdistx_path:
|
34 |
+
lr_scheduler: cosine
|
35 |
+
learning_rate: 0.0002
|
36 |
+
train_on_inputs: false
|
37 |
+
group_by_length: false
|
38 |
+
bf16: false
|
39 |
+
fp16: true
|
40 |
+
tf32: false
|
41 |
+
gradient_checkpointing: true
|
42 |
+
early_stopping_patience:
|
43 |
+
resume_from_checkpoint:
|
44 |
+
local_rank:
|
45 |
+
logging_steps: 1
|
46 |
+
xformers_attention: true
|
47 |
+
flash_attention:
|
48 |
+
gptq_groupsize:
|
49 |
+
gptq_model_v1:
|
50 |
+
warmup_steps: 10
|
51 |
+
eval_steps: 50
|
52 |
+
save_steps:
|
53 |
+
debug:
|
54 |
+
deepspeed:
|
55 |
+
weight_decay: 0.0
|
56 |
+
fsdp:
|
57 |
+
fsdp_config:
|
58 |
+
special_tokens:
|
59 |
+
bos_token: "<s>"
|
60 |
+
eos_token: "</s>"
|
61 |
+
unk_token: "<unk>"
|
examples/{lora-openllama-3b/config.yml → openllama-3b/lora.yml}
RENAMED
@@ -1,5 +1,5 @@
|
|
1 |
-
base_model: openlm-research/
|
2 |
-
base_model_config: openlm-research/
|
3 |
model_type: LlamaForCausalLM
|
4 |
tokenizer_type: LlamaTokenizer
|
5 |
load_in_8bit: true
|
@@ -49,7 +49,7 @@ early_stopping_patience:
|
|
49 |
resume_from_checkpoint:
|
50 |
local_rank:
|
51 |
logging_steps: 1
|
52 |
-
xformers_attention:
|
53 |
flash_attention:
|
54 |
gptq_groupsize:
|
55 |
gptq_model_v1:
|
|
|
1 |
+
base_model: openlm-research/open_llama_3b
|
2 |
+
base_model_config: openlm-research/open_llama_3b
|
3 |
model_type: LlamaForCausalLM
|
4 |
tokenizer_type: LlamaTokenizer
|
5 |
load_in_8bit: true
|
|
|
49 |
resume_from_checkpoint:
|
50 |
local_rank:
|
51 |
logging_steps: 1
|
52 |
+
xformers_attention: true
|
53 |
flash_attention:
|
54 |
gptq_groupsize:
|
55 |
gptq_model_v1:
|
examples/{qlora-openllama-3b/config.yml → openllama-3b/qlora.yml}
RENAMED
@@ -1,5 +1,5 @@
|
|
1 |
-
base_model: openlm-research/
|
2 |
-
base_model_config: openlm-research/
|
3 |
model_type: LlamaForCausalLM
|
4 |
tokenizer_type: LlamaTokenizer
|
5 |
load_in_8bit: false
|
|
|
1 |
+
base_model: openlm-research/open_llama_3b
|
2 |
+
base_model_config: openlm-research/open_llama_3b
|
3 |
model_type: LlamaForCausalLM
|
4 |
tokenizer_type: LlamaTokenizer
|
5 |
load_in_8bit: false
|
configs/pythia_1_2B_alpaca.yml → examples/pythia/lora.yml
RENAMED
@@ -1,36 +1,29 @@
|
|
1 |
base_model: EleutherAI/pythia-1.4b-deduped
|
2 |
-
|
3 |
-
tokenizer_type: AutoTokenizer
|
4 |
load_in_8bit: true
|
5 |
datasets:
|
6 |
-
- path:
|
7 |
type: alpaca
|
8 |
-
- path: data/vicuna_cleaned.jsonl
|
9 |
-
type: sharegpt
|
10 |
-
- path: data/gpt4-instruct-similarity-0.6-dataset.jsonl
|
11 |
-
type: gpteacher
|
12 |
-
- path: data/roleplay-similarity_0.6-instruct-dataset.jsonl
|
13 |
-
type: gpteacher
|
14 |
dataset_prepared_path: last_run_prepared
|
15 |
val_set_size: 0.05
|
16 |
adapter: lora
|
17 |
lora_model_dir:
|
18 |
-
sequence_len:
|
19 |
-
lora_r:
|
20 |
lora_alpha: 32
|
21 |
lora_dropout: 0.05
|
22 |
lora_target_modules:
|
23 |
- query_key_value
|
24 |
-
|
25 |
lora_fan_in_fan_out: true # pythia/GPTNeoX lora specific
|
26 |
-
wandb_project:
|
27 |
wandb_watch:
|
28 |
wandb_run_id:
|
29 |
wandb_log_model:
|
30 |
-
output_dir: ./lora-alpaca
|
31 |
gradient_accumulation_steps: 1
|
32 |
micro_batch_size: 4
|
33 |
-
num_epochs:
|
34 |
learning_rate: 0.00001
|
35 |
train_on_inputs: false
|
36 |
group_by_length: false
|
@@ -39,3 +32,6 @@ tf32: True
|
|
39 |
early_stopping_patience:
|
40 |
resume_from_checkpoint:
|
41 |
local_rank:
|
|
|
|
|
|
|
|
1 |
base_model: EleutherAI/pythia-1.4b-deduped
|
2 |
+
base_model_config: EleutherAI/pythia-1.4b-deduped
|
|
|
3 |
load_in_8bit: true
|
4 |
datasets:
|
5 |
+
- path: teknium/GPT4-LLM-Cleaned
|
6 |
type: alpaca
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
dataset_prepared_path: last_run_prepared
|
8 |
val_set_size: 0.05
|
9 |
adapter: lora
|
10 |
lora_model_dir:
|
11 |
+
sequence_len: 512
|
12 |
+
lora_r: 16
|
13 |
lora_alpha: 32
|
14 |
lora_dropout: 0.05
|
15 |
lora_target_modules:
|
16 |
- query_key_value
|
17 |
+
lora_target_linear:
|
18 |
lora_fan_in_fan_out: true # pythia/GPTNeoX lora specific
|
19 |
+
wandb_project:
|
20 |
wandb_watch:
|
21 |
wandb_run_id:
|
22 |
wandb_log_model:
|
23 |
+
output_dir: ./lora-alpaca-pythia
|
24 |
gradient_accumulation_steps: 1
|
25 |
micro_batch_size: 4
|
26 |
+
num_epochs: 3
|
27 |
learning_rate: 0.00001
|
28 |
train_on_inputs: false
|
29 |
group_by_length: false
|
|
|
32 |
early_stopping_patience:
|
33 |
resume_from_checkpoint:
|
34 |
local_rank:
|
35 |
+
weight_decay: 0.1
|
36 |
+
eval_steps: 20
|
37 |
+
logging_steps: 1
|
examples/qlora-openllama-3b/README.md
DELETED
@@ -1,6 +0,0 @@
|
|
1 |
-
# qlora-openllama-3b
|
2 |
-
|
3 |
-
```shell
|
4 |
-
accelerate launch scripts/finetune.py examples/qlora-openllama-3b/config.yml
|
5 |
-
|
6 |
-
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
scripts/finetune.py
CHANGED
@@ -72,7 +72,19 @@ def do_inference(cfg, model, tokenizer, prompter="AlpacaPrompter"):
|
|
72 |
if not (cfg.special_tokens and token in cfg.special_tokens):
|
73 |
tokenizer.add_special_tokens({token: symbol})
|
74 |
|
75 |
-
prompter_module =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
|
77 |
while True:
|
78 |
print("=" * 80)
|
@@ -80,10 +92,14 @@ def do_inference(cfg, model, tokenizer, prompter="AlpacaPrompter"):
|
|
80 |
instruction = get_multi_line_input()
|
81 |
if not instruction:
|
82 |
return
|
83 |
-
|
84 |
-
|
85 |
-
|
|
|
|
|
|
|
86 |
batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
|
|
|
87 |
print("=" * 40)
|
88 |
model.eval()
|
89 |
with torch.no_grad():
|
@@ -159,7 +175,7 @@ def train(
|
|
159 |
cfg_keys = cfg.keys()
|
160 |
for k, _ in kwargs.items():
|
161 |
# if not strict, allow writing to cfg even if it's not in the yml already
|
162 |
-
if k in cfg_keys or cfg.strict
|
163 |
# handle booleans
|
164 |
if isinstance(cfg[k], bool):
|
165 |
cfg[k] = bool(kwargs[k])
|
@@ -199,8 +215,8 @@ def train(
|
|
199 |
logging.info(f"loading tokenizer... {tokenizer_config}")
|
200 |
tokenizer = load_tokenizer(tokenizer_config, cfg.tokenizer_type, cfg)
|
201 |
|
202 |
-
if
|
203 |
-
["
|
204 |
): # don't need to load dataset for these
|
205 |
if not cfg.pretraining_dataset:
|
206 |
train_dataset, eval_dataset = load_prepare_datasets(
|
@@ -239,7 +255,6 @@ def train(
|
|
239 |
tokenizer,
|
240 |
cfg,
|
241 |
adapter=cfg.adapter,
|
242 |
-
inference=("inference" in kwargs),
|
243 |
)
|
244 |
|
245 |
if "merge_lora" in kwargs and cfg.adapter is not None:
|
@@ -252,9 +267,15 @@ def train(
|
|
252 |
model.save_pretrained(str(Path(cfg.output_dir) / "merged"))
|
253 |
return
|
254 |
|
255 |
-
if
|
256 |
logging.info("calling do_inference function")
|
257 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
258 |
return
|
259 |
|
260 |
if "shard" in kwargs:
|
|
|
72 |
if not (cfg.special_tokens and token in cfg.special_tokens):
|
73 |
tokenizer.add_special_tokens({token: symbol})
|
74 |
|
75 |
+
prompter_module = None
|
76 |
+
if prompter:
|
77 |
+
prompter_module = getattr(
|
78 |
+
importlib.import_module("axolotl.prompters"), prompter
|
79 |
+
)
|
80 |
+
|
81 |
+
if cfg.landmark_attention:
|
82 |
+
from axolotl.monkeypatch.llama_landmark_attn import set_model_mem_id
|
83 |
+
|
84 |
+
set_model_mem_id(model, tokenizer)
|
85 |
+
model.set_mem_cache_args(
|
86 |
+
max_seq_len=255, mem_freq=50, top_k=5, max_cache_size=None
|
87 |
+
)
|
88 |
|
89 |
while True:
|
90 |
print("=" * 80)
|
|
|
92 |
instruction = get_multi_line_input()
|
93 |
if not instruction:
|
94 |
return
|
95 |
+
if prompter_module:
|
96 |
+
prompt: str = next(
|
97 |
+
prompter_module().build_prompt(instruction=instruction.strip("\n"))
|
98 |
+
)
|
99 |
+
else:
|
100 |
+
prompt = instruction.strip()
|
101 |
batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
|
102 |
+
|
103 |
print("=" * 40)
|
104 |
model.eval()
|
105 |
with torch.no_grad():
|
|
|
175 |
cfg_keys = cfg.keys()
|
176 |
for k, _ in kwargs.items():
|
177 |
# if not strict, allow writing to cfg even if it's not in the yml already
|
178 |
+
if k in cfg_keys or not cfg.strict:
|
179 |
# handle booleans
|
180 |
if isinstance(cfg[k], bool):
|
181 |
cfg[k] = bool(kwargs[k])
|
|
|
215 |
logging.info(f"loading tokenizer... {tokenizer_config}")
|
216 |
tokenizer = load_tokenizer(tokenizer_config, cfg.tokenizer_type, cfg)
|
217 |
|
218 |
+
if (
|
219 |
+
check_not_in(["shard", "merge_lora"], kwargs) and not cfg.inference
|
220 |
): # don't need to load dataset for these
|
221 |
if not cfg.pretraining_dataset:
|
222 |
train_dataset, eval_dataset = load_prepare_datasets(
|
|
|
255 |
tokenizer,
|
256 |
cfg,
|
257 |
adapter=cfg.adapter,
|
|
|
258 |
)
|
259 |
|
260 |
if "merge_lora" in kwargs and cfg.adapter is not None:
|
|
|
267 |
model.save_pretrained(str(Path(cfg.output_dir) / "merged"))
|
268 |
return
|
269 |
|
270 |
+
if cfg.inference:
|
271 |
logging.info("calling do_inference function")
|
272 |
+
inf_kwargs: Dict[str, Any] = {}
|
273 |
+
if "prompter" in kwargs:
|
274 |
+
if kwargs["prompter"] == "None":
|
275 |
+
inf_kwargs["prompter"] = None
|
276 |
+
else:
|
277 |
+
inf_kwargs["prompter"] = kwargs["prompter"]
|
278 |
+
do_inference(cfg, model, tokenizer, **inf_kwargs)
|
279 |
return
|
280 |
|
281 |
if "shard" in kwargs:
|
src/axolotl/datasets.py
CHANGED
@@ -33,12 +33,16 @@ class TokenizedPromptDataset(IterableDataset):
|
|
33 |
|
34 |
def __iter__(self):
|
35 |
iterator = iter(self.dataset)
|
|
|
36 |
# Loop through the entire dataset
|
37 |
for example in iterator:
|
38 |
try:
|
39 |
yield self.prompt_tokenizer.tokenize_prompt(example)
|
|
|
40 |
except InvalidDataException:
|
41 |
pass
|
|
|
|
|
42 |
|
43 |
|
44 |
# TODO this isn't the best since it can't interleave datasets
|
|
|
33 |
|
34 |
def __iter__(self):
|
35 |
iterator = iter(self.dataset)
|
36 |
+
count = 0
|
37 |
# Loop through the entire dataset
|
38 |
for example in iterator:
|
39 |
try:
|
40 |
yield self.prompt_tokenizer.tokenize_prompt(example)
|
41 |
+
count += 1
|
42 |
except InvalidDataException:
|
43 |
pass
|
44 |
+
if count == 0:
|
45 |
+
raise RuntimeError("Expected at least one datapoint in dataset.")
|
46 |
|
47 |
|
48 |
# TODO this isn't the best since it can't interleave datasets
|
src/axolotl/monkeypatch/llama_landmark_attn.py
CHANGED
@@ -28,15 +28,24 @@ from typing import List, Optional, Tuple, Union
|
|
28 |
import torch
|
29 |
import torch.utils.checkpoint
|
30 |
from torch import nn
|
31 |
-
from torch.nn import
|
32 |
-
from transformers
|
33 |
from transformers.modeling_outputs import (
|
34 |
BaseModelOutputWithPast,
|
35 |
CausalLMOutputWithPast,
|
36 |
-
SequenceClassifierOutputWithPast,
|
37 |
)
|
38 |
-
from transformers.modeling_utils import PreTrainedModel
|
39 |
from transformers.models.llama.configuration_llama import LlamaConfig
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
from transformers.utils import (
|
41 |
add_start_docstrings,
|
42 |
add_start_docstrings_to_model_forward,
|
@@ -51,131 +60,6 @@ _CONFIG_FOR_DOC = "LlamaConfig"
|
|
51 |
MEM_TOKEN = "<landmark>" # nosec
|
52 |
|
53 |
|
54 |
-
# Copied from transformers.models.bart.modeling_bart._make_causal_mask
|
55 |
-
def _make_causal_mask(
|
56 |
-
input_ids_shape: torch.Size,
|
57 |
-
dtype: torch.dtype,
|
58 |
-
device: torch.device,
|
59 |
-
past_key_values_length: int = 0,
|
60 |
-
):
|
61 |
-
"""
|
62 |
-
Make causal mask used for bi-directional self-attention.
|
63 |
-
"""
|
64 |
-
bsz, tgt_len = input_ids_shape
|
65 |
-
mask = torch.full(
|
66 |
-
(tgt_len, tgt_len),
|
67 |
-
torch.tensor(torch.finfo(dtype).min, device=device),
|
68 |
-
device=device,
|
69 |
-
)
|
70 |
-
mask_cond = torch.arange(mask.size(-1), device=device)
|
71 |
-
mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
|
72 |
-
mask = mask.to(dtype)
|
73 |
-
|
74 |
-
if past_key_values_length > 0:
|
75 |
-
mask = torch.cat(
|
76 |
-
[
|
77 |
-
torch.zeros(
|
78 |
-
tgt_len, past_key_values_length, dtype=dtype, device=device
|
79 |
-
),
|
80 |
-
mask,
|
81 |
-
],
|
82 |
-
dim=-1,
|
83 |
-
)
|
84 |
-
return mask[None, None, :, :].expand(
|
85 |
-
bsz, 1, tgt_len, tgt_len + past_key_values_length
|
86 |
-
)
|
87 |
-
|
88 |
-
|
89 |
-
# Copied from transformers.models.bart.modeling_bart._expand_mask
|
90 |
-
def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
|
91 |
-
"""
|
92 |
-
Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
|
93 |
-
"""
|
94 |
-
bsz, src_len = mask.size()
|
95 |
-
tgt_len = tgt_len if tgt_len is not None else src_len
|
96 |
-
|
97 |
-
expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
|
98 |
-
|
99 |
-
inverted_mask = 1.0 - expanded_mask
|
100 |
-
|
101 |
-
return inverted_mask.masked_fill(
|
102 |
-
inverted_mask.to(torch.bool), torch.finfo(dtype).min
|
103 |
-
)
|
104 |
-
|
105 |
-
|
106 |
-
class LlamaRMSNorm(nn.Module):
|
107 |
-
def __init__(self, hidden_size, eps=1e-6):
|
108 |
-
"""
|
109 |
-
LlamaRMSNorm is equivalent to T5LayerNorm
|
110 |
-
"""
|
111 |
-
super().__init__()
|
112 |
-
self.weight = nn.Parameter(torch.ones(hidden_size))
|
113 |
-
self.variance_epsilon = eps
|
114 |
-
|
115 |
-
def forward(self, hidden_states):
|
116 |
-
variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
|
117 |
-
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
118 |
-
|
119 |
-
# convert into half-precision if necessary
|
120 |
-
if self.weight.dtype in [torch.float16, torch.bfloat16]:
|
121 |
-
hidden_states = hidden_states.to(self.weight.dtype)
|
122 |
-
|
123 |
-
return self.weight * hidden_states
|
124 |
-
|
125 |
-
|
126 |
-
class LlamaRotaryEmbedding(torch.nn.Module):
|
127 |
-
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
128 |
-
super().__init__()
|
129 |
-
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
|
130 |
-
self.register_buffer("inv_freq", inv_freq)
|
131 |
-
|
132 |
-
# Build here to make `torch.jit.trace` work.
|
133 |
-
self.max_seq_len_cached = max_position_embeddings
|
134 |
-
t = torch.arange(
|
135 |
-
self.max_seq_len_cached,
|
136 |
-
device=self.inv_freq.device,
|
137 |
-
dtype=self.inv_freq.dtype,
|
138 |
-
)
|
139 |
-
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
140 |
-
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
141 |
-
emb = torch.cat((freqs, freqs), dim=-1)
|
142 |
-
self.register_buffer(
|
143 |
-
"cos_cached", emb.cos()[None, None, :, :], persistent=False
|
144 |
-
)
|
145 |
-
self.register_buffer(
|
146 |
-
"sin_cached", emb.sin()[None, None, :, :], persistent=False
|
147 |
-
)
|
148 |
-
|
149 |
-
def forward(self, x, seq_len=None):
|
150 |
-
# x: [bs, num_attention_heads, seq_len, head_size]
|
151 |
-
# This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
|
152 |
-
if seq_len > self.max_seq_len_cached:
|
153 |
-
self.max_seq_len_cached = seq_len
|
154 |
-
t = torch.arange(
|
155 |
-
self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype
|
156 |
-
)
|
157 |
-
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
158 |
-
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
159 |
-
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
|
160 |
-
self.register_buffer(
|
161 |
-
"cos_cached", emb.cos()[None, None, :, :], persistent=False
|
162 |
-
)
|
163 |
-
self.register_buffer(
|
164 |
-
"sin_cached", emb.sin()[None, None, :, :], persistent=False
|
165 |
-
)
|
166 |
-
return (
|
167 |
-
self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
|
168 |
-
self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
|
169 |
-
)
|
170 |
-
|
171 |
-
|
172 |
-
def rotate_half(x):
|
173 |
-
"""Rotates half the hidden dims of the input."""
|
174 |
-
x1 = x[..., : x.shape[-1] // 2]
|
175 |
-
x2 = x[..., x.shape[-1] // 2 :]
|
176 |
-
return torch.cat((-x2, x1), dim=-1)
|
177 |
-
|
178 |
-
|
179 |
def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
|
180 |
# The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
|
181 |
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
|
@@ -190,24 +74,11 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
|
|
190 |
return q_embed, k_embed
|
191 |
|
192 |
|
193 |
-
class LlamaMLP(nn.Module):
|
194 |
-
def __init__(
|
195 |
-
self,
|
196 |
-
hidden_size: int,
|
197 |
-
intermediate_size: int,
|
198 |
-
hidden_act: str,
|
199 |
-
):
|
200 |
-
super().__init__()
|
201 |
-
self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
|
202 |
-
self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
|
203 |
-
self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
|
204 |
-
self.act_fn = ACT2FN[hidden_act]
|
205 |
-
|
206 |
-
def forward(self, x):
|
207 |
-
return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
208 |
-
|
209 |
-
|
210 |
class LandmarkGroupedSoftmaxFunction(torch.autograd.Function):
|
|
|
|
|
|
|
|
|
211 |
# Note that forward, setup_context, and backward are @staticmethods
|
212 |
@staticmethod
|
213 |
def forward(ctx, x, dim, mem_cnt, resp_mem_idx):
|
@@ -682,16 +553,14 @@ class LlamaAttention(nn.Module):
|
|
682 |
# upcast attention to fp32
|
683 |
if is_mem is None:
|
684 |
raise ValueError("Don't use this without landmarks")
|
685 |
-
|
686 |
-
|
687 |
-
|
688 |
-
|
689 |
-
|
690 |
-
|
691 |
-
|
692 |
-
|
693 |
-
last_section_mask=last_section_mask,
|
694 |
-
).to(query_states.dtype)
|
695 |
if attn_prefix is not None:
|
696 |
attn_prefix, attn_weights = torch.split(
|
697 |
attn_weights,
|
@@ -722,6 +591,10 @@ class LlamaAttention(nn.Module):
|
|
722 |
|
723 |
|
724 |
class LlamaDecoderLayer(nn.Module):
|
|
|
|
|
|
|
|
|
725 |
def __init__(self, config: LlamaConfig):
|
726 |
super().__init__()
|
727 |
self.hidden_size = config.hidden_size
|
@@ -802,114 +675,6 @@ class LlamaDecoderLayer(nn.Module):
|
|
802 |
return outputs
|
803 |
|
804 |
|
805 |
-
LLAMA_START_DOCSTRING = r"""
|
806 |
-
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
807 |
-
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
808 |
-
etc.)
|
809 |
-
|
810 |
-
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
811 |
-
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
812 |
-
and behavior.
|
813 |
-
|
814 |
-
Parameters:
|
815 |
-
config ([`LlamaConfig`]):
|
816 |
-
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
817 |
-
load the weights associated with the model, only the configuration. Check out the
|
818 |
-
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
819 |
-
"""
|
820 |
-
|
821 |
-
|
822 |
-
@add_start_docstrings(
|
823 |
-
"The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
|
824 |
-
LLAMA_START_DOCSTRING,
|
825 |
-
)
|
826 |
-
class LlamaPreTrainedModel(PreTrainedModel):
|
827 |
-
config_class = LlamaConfig
|
828 |
-
base_model_prefix = "model"
|
829 |
-
supports_gradient_checkpointing = True
|
830 |
-
_no_split_modules = ["LlamaDecoderLayer"]
|
831 |
-
_keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
|
832 |
-
|
833 |
-
def _init_weights(self, module):
|
834 |
-
std = self.config.initializer_range
|
835 |
-
if isinstance(module, nn.Linear):
|
836 |
-
module.weight.data.normal_(mean=0.0, std=std)
|
837 |
-
if module.bias is not None:
|
838 |
-
module.bias.data.zero_()
|
839 |
-
elif isinstance(module, nn.Embedding):
|
840 |
-
module.weight.data.normal_(mean=0.0, std=std)
|
841 |
-
if module.padding_idx is not None:
|
842 |
-
module.weight.data[module.padding_idx].zero_()
|
843 |
-
|
844 |
-
def _set_gradient_checkpointing(self, module, value=False):
|
845 |
-
if isinstance(module, LlamaModel):
|
846 |
-
module.gradient_checkpointing = value
|
847 |
-
|
848 |
-
|
849 |
-
LLAMA_INPUTS_DOCSTRING = r"""
|
850 |
-
Args:
|
851 |
-
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
852 |
-
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
853 |
-
it.
|
854 |
-
|
855 |
-
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
856 |
-
[`PreTrainedTokenizer.__call__`] for details.
|
857 |
-
|
858 |
-
[What are input IDs?](../glossary#input-ids)
|
859 |
-
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
860 |
-
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
861 |
-
|
862 |
-
- 1 for tokens that are **not masked**,
|
863 |
-
- 0 for tokens that are **masked**.
|
864 |
-
|
865 |
-
[What are attention masks?](../glossary#attention-mask)
|
866 |
-
|
867 |
-
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
868 |
-
[`PreTrainedTokenizer.__call__`] for details.
|
869 |
-
|
870 |
-
If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
|
871 |
-
`past_key_values`).
|
872 |
-
|
873 |
-
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
874 |
-
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
875 |
-
information on the default strategy.
|
876 |
-
|
877 |
-
- 1 indicates the head is **not masked**,
|
878 |
-
- 0 indicates the head is **masked**.
|
879 |
-
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
880 |
-
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
881 |
-
config.n_positions - 1]`.
|
882 |
-
|
883 |
-
[What are position IDs?](../glossary#position-ids)
|
884 |
-
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
885 |
-
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
|
886 |
-
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
|
887 |
-
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
|
888 |
-
|
889 |
-
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
890 |
-
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
|
891 |
-
|
892 |
-
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
|
893 |
-
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
|
894 |
-
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
|
895 |
-
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
896 |
-
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
897 |
-
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
898 |
-
model's internal embedding lookup matrix.
|
899 |
-
use_cache (`bool`, *optional*):
|
900 |
-
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
901 |
-
`past_key_values`).
|
902 |
-
output_attentions (`bool`, *optional*):
|
903 |
-
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
904 |
-
tensors for more detail.
|
905 |
-
output_hidden_states (`bool`, *optional*):
|
906 |
-
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
907 |
-
more detail.
|
908 |
-
return_dict (`bool`, *optional*):
|
909 |
-
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
910 |
-
"""
|
911 |
-
|
912 |
-
|
913 |
@add_start_docstrings(
|
914 |
"The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
|
915 |
LLAMA_START_DOCSTRING,
|
@@ -1178,6 +943,10 @@ class LlamaModel(LlamaPreTrainedModel):
|
|
1178 |
|
1179 |
|
1180 |
class LlamaForCausalLM(LlamaPreTrainedModel):
|
|
|
|
|
|
|
|
|
1181 |
def __init__(self, config):
|
1182 |
super().__init__(config)
|
1183 |
self.model = LlamaModel(config)
|
@@ -1448,148 +1217,33 @@ class LlamaForCausalLM(LlamaPreTrainedModel):
|
|
1448 |
return reordered_past
|
1449 |
|
1450 |
|
1451 |
-
@add_start_docstrings(
|
1452 |
-
"""
|
1453 |
-
The LLaMa Model transformer with a sequence classification head on top (linear layer).
|
1454 |
-
|
1455 |
-
[`LlamaForSequenceClassification`] uses the last token in order to do the classification, as other causal models
|
1456 |
-
(e.g. GPT-2) do.
|
1457 |
-
|
1458 |
-
Since it does classification on the last token, it requires to know the position of the last token. If a
|
1459 |
-
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
|
1460 |
-
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
|
1461 |
-
padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
|
1462 |
-
each row of the batch).
|
1463 |
-
""",
|
1464 |
-
LLAMA_START_DOCSTRING,
|
1465 |
-
)
|
1466 |
-
class LlamaForSequenceClassification(LlamaPreTrainedModel):
|
1467 |
-
_keys_to_ignore_on_load_missing = [r"lm_head.weight"]
|
1468 |
-
|
1469 |
-
def __init__(self, config):
|
1470 |
-
super().__init__(config)
|
1471 |
-
self.num_labels = config.num_labels
|
1472 |
-
self.model = LlamaModel(config)
|
1473 |
-
self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
|
1474 |
-
|
1475 |
-
# Initialize weights and apply final processing
|
1476 |
-
self.post_init()
|
1477 |
-
|
1478 |
-
def get_input_embeddings(self):
|
1479 |
-
return self.model.embed_tokens
|
1480 |
-
|
1481 |
-
def set_input_embeddings(self, value):
|
1482 |
-
self.model.embed_tokens = value
|
1483 |
-
|
1484 |
-
@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
|
1485 |
-
def forward(
|
1486 |
-
self,
|
1487 |
-
input_ids: torch.LongTensor = None,
|
1488 |
-
attention_mask: Optional[torch.Tensor] = None,
|
1489 |
-
position_ids: Optional[torch.LongTensor] = None,
|
1490 |
-
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
1491 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1492 |
-
labels: Optional[torch.LongTensor] = None,
|
1493 |
-
use_cache: Optional[bool] = None,
|
1494 |
-
output_attentions: Optional[bool] = None,
|
1495 |
-
output_hidden_states: Optional[bool] = None,
|
1496 |
-
return_dict: Optional[bool] = None,
|
1497 |
-
) -> Union[Tuple, SequenceClassifierOutputWithPast]:
|
1498 |
-
r"""
|
1499 |
-
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
1500 |
-
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
1501 |
-
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
1502 |
-
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
1503 |
-
"""
|
1504 |
-
return_dict = (
|
1505 |
-
return_dict if return_dict is not None else self.config.use_return_dict
|
1506 |
-
)
|
1507 |
-
|
1508 |
-
transformer_outputs = self.model(
|
1509 |
-
input_ids,
|
1510 |
-
attention_mask=attention_mask,
|
1511 |
-
position_ids=position_ids,
|
1512 |
-
past_key_values=past_key_values,
|
1513 |
-
inputs_embeds=inputs_embeds,
|
1514 |
-
use_cache=use_cache,
|
1515 |
-
output_attentions=output_attentions,
|
1516 |
-
output_hidden_states=output_hidden_states,
|
1517 |
-
return_dict=return_dict,
|
1518 |
-
)
|
1519 |
-
hidden_states = transformer_outputs[0]
|
1520 |
-
logits = self.score(hidden_states)
|
1521 |
-
|
1522 |
-
if input_ids is not None:
|
1523 |
-
batch_size = input_ids.shape[0]
|
1524 |
-
else:
|
1525 |
-
batch_size = inputs_embeds.shape[0]
|
1526 |
-
|
1527 |
-
if self.config.pad_token_id is None and batch_size != 1:
|
1528 |
-
raise ValueError(
|
1529 |
-
"Cannot handle batch sizes > 1 if no padding token is defined."
|
1530 |
-
)
|
1531 |
-
if self.config.pad_token_id is None:
|
1532 |
-
sequence_lengths = -1
|
1533 |
-
else:
|
1534 |
-
if input_ids is not None:
|
1535 |
-
sequence_lengths = (
|
1536 |
-
torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1
|
1537 |
-
).to(logits.device)
|
1538 |
-
else:
|
1539 |
-
sequence_lengths = -1
|
1540 |
-
|
1541 |
-
pooled_logits = logits[
|
1542 |
-
torch.arange(batch_size, device=logits.device), sequence_lengths
|
1543 |
-
]
|
1544 |
-
|
1545 |
-
loss = None
|
1546 |
-
if labels is not None:
|
1547 |
-
labels = labels.to(logits.device)
|
1548 |
-
if self.config.problem_type is None:
|
1549 |
-
if self.num_labels == 1:
|
1550 |
-
self.config.problem_type = "regression"
|
1551 |
-
elif self.num_labels > 1 and (
|
1552 |
-
labels.dtype == torch.long or labels.dtype == torch.int
|
1553 |
-
):
|
1554 |
-
self.config.problem_type = "single_label_classification"
|
1555 |
-
else:
|
1556 |
-
self.config.problem_type = "multi_label_classification"
|
1557 |
-
|
1558 |
-
if self.config.problem_type == "regression":
|
1559 |
-
loss_fct = MSELoss()
|
1560 |
-
if self.num_labels == 1:
|
1561 |
-
loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
|
1562 |
-
else:
|
1563 |
-
loss = loss_fct(pooled_logits, labels)
|
1564 |
-
elif self.config.problem_type == "single_label_classification":
|
1565 |
-
loss_fct = CrossEntropyLoss()
|
1566 |
-
loss = loss_fct(
|
1567 |
-
pooled_logits.view(-1, self.num_labels), labels.view(-1)
|
1568 |
-
)
|
1569 |
-
elif self.config.problem_type == "multi_label_classification":
|
1570 |
-
loss_fct = BCEWithLogitsLoss()
|
1571 |
-
loss = loss_fct(pooled_logits, labels)
|
1572 |
-
if not return_dict:
|
1573 |
-
output = (pooled_logits,) + transformer_outputs[1:]
|
1574 |
-
return ((loss,) + output) if loss is not None else output
|
1575 |
-
|
1576 |
-
return SequenceClassifierOutputWithPast(
|
1577 |
-
loss=loss,
|
1578 |
-
logits=pooled_logits,
|
1579 |
-
past_key_values=transformer_outputs.past_key_values,
|
1580 |
-
hidden_states=transformer_outputs.hidden_states,
|
1581 |
-
attentions=transformer_outputs.attentions,
|
1582 |
-
)
|
1583 |
-
|
1584 |
-
|
1585 |
def add_mem_tokens(example, mem_freq, mem_id):
|
1586 |
-
|
1587 |
ret = []
|
1588 |
prev_idx = 0
|
1589 |
-
for t_idx in range(mem_freq, len(
|
1590 |
-
ret.extend(
|
1591 |
ret.append(mem_id)
|
1592 |
prev_idx = t_idx
|
1593 |
-
ret.extend(
|
1594 |
# drop attention_mask
|
1595 |
return {"input_ids": ret}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
import torch
|
29 |
import torch.utils.checkpoint
|
30 |
from torch import nn
|
31 |
+
from torch.nn import CrossEntropyLoss
|
32 |
+
from transformers import LlamaTokenizer
|
33 |
from transformers.modeling_outputs import (
|
34 |
BaseModelOutputWithPast,
|
35 |
CausalLMOutputWithPast,
|
|
|
36 |
)
|
|
|
37 |
from transformers.models.llama.configuration_llama import LlamaConfig
|
38 |
+
from transformers.models.llama.modeling_llama import (
|
39 |
+
LLAMA_INPUTS_DOCSTRING,
|
40 |
+
LLAMA_START_DOCSTRING,
|
41 |
+
LlamaMLP,
|
42 |
+
LlamaPreTrainedModel,
|
43 |
+
LlamaRMSNorm,
|
44 |
+
LlamaRotaryEmbedding,
|
45 |
+
_expand_mask,
|
46 |
+
_make_causal_mask,
|
47 |
+
rotate_half,
|
48 |
+
)
|
49 |
from transformers.utils import (
|
50 |
add_start_docstrings,
|
51 |
add_start_docstrings_to_model_forward,
|
|
|
60 |
MEM_TOKEN = "<landmark>" # nosec
|
61 |
|
62 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
|
64 |
# The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
|
65 |
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
|
|
|
74 |
return q_embed, k_embed
|
75 |
|
76 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
class LandmarkGroupedSoftmaxFunction(torch.autograd.Function):
|
78 |
+
"""
|
79 |
+
Landmark grouped softmax function.
|
80 |
+
"""
|
81 |
+
|
82 |
# Note that forward, setup_context, and backward are @staticmethods
|
83 |
@staticmethod
|
84 |
def forward(ctx, x, dim, mem_cnt, resp_mem_idx):
|
|
|
553 |
# upcast attention to fp32
|
554 |
if is_mem is None:
|
555 |
raise ValueError("Don't use this without landmarks")
|
556 |
+
|
557 |
+
attn_weights = landmark_grouped_softmax(
|
558 |
+
attn_weights,
|
559 |
+
dim=-1,
|
560 |
+
is_mem=is_mem.expand(-1, self.num_heads, -1, -1),
|
561 |
+
last_section_mask=last_section_mask,
|
562 |
+
).to(query_states.dtype)
|
563 |
+
|
|
|
|
|
564 |
if attn_prefix is not None:
|
565 |
attn_prefix, attn_weights = torch.split(
|
566 |
attn_weights,
|
|
|
591 |
|
592 |
|
593 |
class LlamaDecoderLayer(nn.Module):
|
594 |
+
"""
|
595 |
+
Llama Decoder layer
|
596 |
+
"""
|
597 |
+
|
598 |
def __init__(self, config: LlamaConfig):
|
599 |
super().__init__()
|
600 |
self.hidden_size = config.hidden_size
|
|
|
675 |
return outputs
|
676 |
|
677 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
678 |
@add_start_docstrings(
|
679 |
"The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
|
680 |
LLAMA_START_DOCSTRING,
|
|
|
943 |
|
944 |
|
945 |
class LlamaForCausalLM(LlamaPreTrainedModel):
|
946 |
+
"""
|
947 |
+
Llama model with a causal language modeling head.
|
948 |
+
"""
|
949 |
+
|
950 |
def __init__(self, config):
|
951 |
super().__init__(config)
|
952 |
self.model = LlamaModel(config)
|
|
|
1217 |
return reordered_past
|
1218 |
|
1219 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1220 |
def add_mem_tokens(example, mem_freq, mem_id):
|
1221 |
+
ids = example["input_ids"]
|
1222 |
ret = []
|
1223 |
prev_idx = 0
|
1224 |
+
for t_idx in range(mem_freq, len(ids), mem_freq):
|
1225 |
+
ret.extend(ids[prev_idx:t_idx])
|
1226 |
ret.append(mem_id)
|
1227 |
prev_idx = t_idx
|
1228 |
+
ret.extend(ids[prev_idx:])
|
1229 |
# drop attention_mask
|
1230 |
return {"input_ids": ret}
|
1231 |
+
|
1232 |
+
|
1233 |
+
def patch_llama_with_landmark_attn():
|
1234 |
+
import transformers
|
1235 |
+
|
1236 |
+
transformers.models.llama.modeling_llama.LlamaForCausalLM = LlamaForCausalLM
|
1237 |
+
transformers.models.llama.modeling_llama.LlamaModel = LlamaModel
|
1238 |
+
transformers.models.llama.modeling_llama.LlamaAttention = LlamaAttention
|
1239 |
+
transformers.models.llama.modeling_llama.LlamaDecoderLayer = LlamaDecoderLayer
|
1240 |
+
transformers.models.llama.modeling_llama.apply_rotary_pos_emb = apply_rotary_pos_emb
|
1241 |
+
|
1242 |
+
|
1243 |
+
def set_model_mem_id(model: LlamaForCausalLM, tokenizer: LlamaTokenizer):
|
1244 |
+
mem_id = tokenizer.convert_tokens_to_ids(MEM_TOKEN)
|
1245 |
+
model.set_mem_id(mem_id)
|
1246 |
+
|
1247 |
+
|
1248 |
+
def get_mem_id(tokenizer: LlamaTokenizer):
|
1249 |
+
return tokenizer.convert_tokens_to_ids(MEM_TOKEN)
|
src/axolotl/prompt_strategies/sharegpt_jokes.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""Module for Jokes prompts using sharegpt style """
|
2 |
+
from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
|
3 |
+
from axolotl.prompters import PromptStyle, ShareGPTPrompter
|
4 |
+
|
5 |
+
|
6 |
+
def load(tokenizer, cfg):
|
7 |
+
return SimpleJokesShareGPTPromptTokenizingStrategy(
|
8 |
+
ShareGPTPrompter(PromptStyle.CHAT.value),
|
9 |
+
tokenizer,
|
10 |
+
cfg.train_on_inputs,
|
11 |
+
cfg.sequence_len,
|
12 |
+
)
|
13 |
+
|
14 |
+
|
15 |
+
class SimpleJokesShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
|
16 |
+
"""
|
17 |
+
Tokenization strategy for asking bot to tell a joke and then explain why its funny
|
18 |
+
"""
|
19 |
+
|
20 |
+
# title, text, explanation
|
21 |
+
def get_conversation_thread(self, prompt):
|
22 |
+
title = "" if not prompt["title"] else prompt["title"] + " "
|
23 |
+
return [
|
24 |
+
{"from": "human", "value": "Tell me a joke."},
|
25 |
+
{"from": "gpt", "value": title + prompt["text"]},
|
26 |
+
{"from": "human", "value": "Why is that joke funny?"},
|
27 |
+
{"from": "gpt", "value": prompt["explanation"]},
|
28 |
+
]
|
src/axolotl/prompt_strategies/sharegpt_simple.py
CHANGED
@@ -13,6 +13,15 @@ def load(tokenizer, cfg):
|
|
13 |
)
|
14 |
|
15 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
def load_guanaco(tokenizer, cfg):
|
17 |
return GuanacoShareGPTPromptTokenizingStrategy(
|
18 |
ShareGPTPrompter(PromptStyle.CHAT.value),
|
@@ -31,6 +40,18 @@ class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
|
|
31 |
return prompt["conversations"]
|
32 |
|
33 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
class GuanacoShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
|
35 |
"""
|
36 |
sharegpt strategy that remaps oasst data to sharegpt format
|
|
|
13 |
)
|
14 |
|
15 |
|
16 |
+
def load_role(tokenizer, cfg):
|
17 |
+
return SimpleRoleShareGPTPromptTokenizingStrategy(
|
18 |
+
ShareGPTPrompter(PromptStyle.CHAT.value),
|
19 |
+
tokenizer,
|
20 |
+
cfg.train_on_inputs,
|
21 |
+
cfg.sequence_len,
|
22 |
+
)
|
23 |
+
|
24 |
+
|
25 |
def load_guanaco(tokenizer, cfg):
|
26 |
return GuanacoShareGPTPromptTokenizingStrategy(
|
27 |
ShareGPTPrompter(PromptStyle.CHAT.value),
|
|
|
40 |
return prompt["conversations"]
|
41 |
|
42 |
|
43 |
+
class SimpleRoleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
|
44 |
+
"""
|
45 |
+
basic sharegpt strategy to grab conversations from the sample row, but uses role instead of from
|
46 |
+
"""
|
47 |
+
|
48 |
+
def get_conversation_thread(self, prompt):
|
49 |
+
conversations = prompt["conversations"]
|
50 |
+
# remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
|
51 |
+
turns = [{"from": t["role"], "value": t["value"]} for t in conversations]
|
52 |
+
return turns
|
53 |
+
|
54 |
+
|
55 |
class GuanacoShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
|
56 |
"""
|
57 |
sharegpt strategy that remaps oasst data to sharegpt format
|
src/axolotl/prompters.py
CHANGED
@@ -261,28 +261,33 @@ class Conversation:
|
|
261 |
self.messages.append([role, message])
|
262 |
|
263 |
|
264 |
-
conv_vicuna_v1_1 = Conversation(
|
265 |
-
system="A chat between a curious user and an artificial intelligence assistant. "
|
266 |
-
"The assistant gives helpful, detailed, and polite answers to the user's questions.",
|
267 |
-
roles=["USER", "ASSISTANT"],
|
268 |
-
messages=[],
|
269 |
-
offset=0,
|
270 |
-
sep_style=SeparatorStyle.TWO,
|
271 |
-
sep=" ",
|
272 |
-
sep2=" ",
|
273 |
-
)
|
274 |
-
|
275 |
-
|
276 |
class ShareGPTPrompter: # pylint: disable=too-few-public-methods
|
277 |
"""
|
278 |
A prompter that generates prompts for the ShareGPT
|
279 |
"""
|
280 |
|
281 |
-
def __init__(self, prompt_style=None):
|
282 |
if prompt_style != PromptStyle.CHAT.value:
|
283 |
raise ValueError(
|
284 |
f"unsupported prompt_style for ShareGPTPrompter({prompt_style})"
|
285 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
286 |
|
287 |
# def match_prompt_style(self):
|
288 |
# if self.prompt_style == PromptStyle.chat.value:
|
@@ -300,7 +305,7 @@ class ShareGPTPrompter: # pylint: disable=too-few-public-methods
|
|
300 |
# also happens on the data splitting leaving empty conversations
|
301 |
raise IndexError
|
302 |
|
303 |
-
conv =
|
304 |
roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
|
305 |
|
306 |
try:
|
|
|
261 |
self.messages.append([role, message])
|
262 |
|
263 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
264 |
class ShareGPTPrompter: # pylint: disable=too-few-public-methods
|
265 |
"""
|
266 |
A prompter that generates prompts for the ShareGPT
|
267 |
"""
|
268 |
|
269 |
+
def __init__(self, prompt_style=None, system_prompt: Optional[str] = None):
|
270 |
if prompt_style != PromptStyle.CHAT.value:
|
271 |
raise ValueError(
|
272 |
f"unsupported prompt_style for ShareGPTPrompter({prompt_style})"
|
273 |
)
|
274 |
+
system: str = (
|
275 |
+
system_prompt
|
276 |
+
if system_prompt
|
277 |
+
else (
|
278 |
+
"A chat between a curious user and an artificial intelligence assistant. "
|
279 |
+
"The assistant gives helpful, detailed, and polite answers to the user's questions."
|
280 |
+
)
|
281 |
+
)
|
282 |
+
self._conversation = Conversation(
|
283 |
+
system=system,
|
284 |
+
roles=["USER", "ASSISTANT"],
|
285 |
+
messages=[],
|
286 |
+
offset=0,
|
287 |
+
sep_style=SeparatorStyle.TWO,
|
288 |
+
sep=" ",
|
289 |
+
sep2=" ",
|
290 |
+
)
|
291 |
|
292 |
# def match_prompt_style(self):
|
293 |
# if self.prompt_style == PromptStyle.chat.value:
|
|
|
305 |
# also happens on the data splitting leaving empty conversations
|
306 |
raise IndexError
|
307 |
|
308 |
+
conv = self._conversation.copy()
|
309 |
roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
|
310 |
|
311 |
try:
|
src/axolotl/utils/data.py
CHANGED
@@ -240,8 +240,15 @@ def load_tokenized_prepared_datasets(
|
|
240 |
ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
|
241 |
datasets.append(ds_wrapper)
|
242 |
else:
|
243 |
-
|
244 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
245 |
logging.info("tokenizing, merging, and shuffling master dataset")
|
246 |
|
247 |
samples: List[int] = []
|
|
|
240 |
ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
|
241 |
datasets.append(ds_wrapper)
|
242 |
else:
|
243 |
+
suffix = ""
|
244 |
+
if ":load_" in d.type:
|
245 |
+
suffix = f" Did you mean {d.type.replace(':load_', '.load_')}?"
|
246 |
+
logging.error(
|
247 |
+
f"unhandled prompt tokenization strategy: {d.type}. {suffix}"
|
248 |
+
)
|
249 |
+
raise ValueError(
|
250 |
+
f"unhandled prompt tokenization strategy: {d.type} {suffix}"
|
251 |
+
)
|
252 |
logging.info("tokenizing, merging, and shuffling master dataset")
|
253 |
|
254 |
samples: List[int] = []
|
src/axolotl/utils/models.py
CHANGED
@@ -20,15 +20,6 @@ from transformers import (
|
|
20 |
LlamaConfig,
|
21 |
)
|
22 |
|
23 |
-
try:
|
24 |
-
from transformers import ( # pylint: disable=unused-import # noqa: F401
|
25 |
-
LlamaForCausalLM,
|
26 |
-
)
|
27 |
-
except ImportError:
|
28 |
-
logging.warning(
|
29 |
-
"This version of transformers does not support Llama. Consider upgrading."
|
30 |
-
)
|
31 |
-
|
32 |
from axolotl.prompt_tokenizers import LLAMA_DEFAULT_PAD_TOKEN
|
33 |
|
34 |
if TYPE_CHECKING:
|
@@ -78,15 +69,9 @@ def load_tokenizer(
|
|
78 |
|
79 |
|
80 |
def load_model(
|
81 |
-
base_model,
|
82 |
-
base_model_config,
|
83 |
-
model_type,
|
84 |
-
tokenizer,
|
85 |
-
cfg,
|
86 |
-
adapter="lora",
|
87 |
-
inference=False,
|
88 |
):
|
89 |
-
# type: (str, str, str, AutoTokenizer, DictDefault, Optional[str]
|
90 |
"""
|
91 |
Load a model from a base model and a model type.
|
92 |
"""
|
@@ -98,7 +83,7 @@ def load_model(
|
|
98 |
)
|
99 |
|
100 |
if cfg.is_llama_derived_model and cfg.flash_attention:
|
101 |
-
if cfg.device not in ["mps", "cpu"] and inference
|
102 |
from axolotl.flash_attn import replace_llama_attn_with_flash_attn
|
103 |
|
104 |
logging.info("patching with flash attention")
|
@@ -118,14 +103,15 @@ def load_model(
|
|
118 |
logging.info("patching with sdp attention")
|
119 |
hijack_llama_sdp_attention()
|
120 |
elif cfg.is_llama_derived_model and cfg.landmark_attention:
|
121 |
-
from axolotl.monkeypatch.llama_landmark_attn import (
|
122 |
MEM_TOKEN,
|
123 |
-
|
124 |
)
|
125 |
|
126 |
logging.info("patching with landmark attention")
|
|
|
127 |
|
128 |
-
#
|
129 |
tokenizer.add_special_tokens({"additional_special_tokens": [MEM_TOKEN]})
|
130 |
|
131 |
if cfg.is_llama_derived_model and cfg.xpos_rope:
|
@@ -210,7 +196,9 @@ def load_model(
|
|
210 |
else True,
|
211 |
)
|
212 |
load_in_8bit = False
|
213 |
-
elif cfg.is_llama_derived_model
|
|
|
|
|
214 |
config = LlamaConfig.from_pretrained(base_model_config)
|
215 |
model = LlamaForCausalLM.from_pretrained(
|
216 |
base_model,
|
@@ -314,7 +302,9 @@ def load_model(
|
|
314 |
or (cfg.adapter == "qlora" and cfg.load_in_4bit)
|
315 |
):
|
316 |
logging.info("converting PEFT model w/ prepare_model_for_kbit_training")
|
317 |
-
model = prepare_model_for_kbit_training(
|
|
|
|
|
318 |
|
319 |
model, lora_config = load_adapter(model, cfg, adapter)
|
320 |
|
@@ -387,7 +377,6 @@ def load_llama_adapter(model, cfg):
|
|
387 |
model = PeftModel.from_pretrained(
|
388 |
model,
|
389 |
cfg.lora_model_dir,
|
390 |
-
device_map=cfg.device_map,
|
391 |
torch_dtype=torch.float16,
|
392 |
)
|
393 |
else:
|
@@ -449,8 +438,7 @@ def load_lora(model, cfg):
|
|
449 |
model = PeftModel.from_pretrained(
|
450 |
model,
|
451 |
cfg.lora_model_dir,
|
452 |
-
|
453 |
-
# torch_dtype=torch.float16,
|
454 |
)
|
455 |
else:
|
456 |
model = get_peft_model(model, lora_config)
|
|
|
20 |
LlamaConfig,
|
21 |
)
|
22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
from axolotl.prompt_tokenizers import LLAMA_DEFAULT_PAD_TOKEN
|
24 |
|
25 |
if TYPE_CHECKING:
|
|
|
69 |
|
70 |
|
71 |
def load_model(
|
72 |
+
base_model, base_model_config, model_type, tokenizer, cfg, adapter="lora"
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
):
|
74 |
+
# type: (str, str, str, AutoTokenizer, DictDefault, Optional[str]) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
|
75 |
"""
|
76 |
Load a model from a base model and a model type.
|
77 |
"""
|
|
|
83 |
)
|
84 |
|
85 |
if cfg.is_llama_derived_model and cfg.flash_attention:
|
86 |
+
if cfg.device not in ["mps", "cpu"] and not cfg.inference:
|
87 |
from axolotl.flash_attn import replace_llama_attn_with_flash_attn
|
88 |
|
89 |
logging.info("patching with flash attention")
|
|
|
103 |
logging.info("patching with sdp attention")
|
104 |
hijack_llama_sdp_attention()
|
105 |
elif cfg.is_llama_derived_model and cfg.landmark_attention:
|
106 |
+
from axolotl.monkeypatch.llama_landmark_attn import (
|
107 |
MEM_TOKEN,
|
108 |
+
patch_llama_with_landmark_attn,
|
109 |
)
|
110 |
|
111 |
logging.info("patching with landmark attention")
|
112 |
+
patch_llama_with_landmark_attn()
|
113 |
|
114 |
+
# Note: This might overwrite previous additional_special_tokens
|
115 |
tokenizer.add_special_tokens({"additional_special_tokens": [MEM_TOKEN]})
|
116 |
|
117 |
if cfg.is_llama_derived_model and cfg.xpos_rope:
|
|
|
196 |
else True,
|
197 |
)
|
198 |
load_in_8bit = False
|
199 |
+
elif cfg.is_llama_derived_model:
|
200 |
+
from transformers import LlamaForCausalLM
|
201 |
+
|
202 |
config = LlamaConfig.from_pretrained(base_model_config)
|
203 |
model = LlamaForCausalLM.from_pretrained(
|
204 |
base_model,
|
|
|
302 |
or (cfg.adapter == "qlora" and cfg.load_in_4bit)
|
303 |
):
|
304 |
logging.info("converting PEFT model w/ prepare_model_for_kbit_training")
|
305 |
+
model = prepare_model_for_kbit_training(
|
306 |
+
model, use_gradient_checkpointing=cfg.gradient_checkpointing
|
307 |
+
)
|
308 |
|
309 |
model, lora_config = load_adapter(model, cfg, adapter)
|
310 |
|
|
|
377 |
model = PeftModel.from_pretrained(
|
378 |
model,
|
379 |
cfg.lora_model_dir,
|
|
|
380 |
torch_dtype=torch.float16,
|
381 |
)
|
382 |
else:
|
|
|
438 |
model = PeftModel.from_pretrained(
|
439 |
model,
|
440 |
cfg.lora_model_dir,
|
441 |
+
is_trainable=not cfg.inference,
|
|
|
442 |
)
|
443 |
else:
|
444 |
model = get_peft_model(model, lora_config)
|
src/axolotl/utils/trainer.py
CHANGED
@@ -245,16 +245,19 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
|
|
245 |
if cfg.is_llama_derived_model and cfg.landmark_attention:
|
246 |
from functools import partial
|
247 |
|
248 |
-
from axolotl.monkeypatch.llama_landmark_attn import
|
|
|
|
|
|
|
|
|
249 |
|
250 |
-
|
251 |
-
model.set_mem_id(mem_id)
|
252 |
|
253 |
logging.info("Adding landmark attention tokens to dataset")
|
254 |
|
255 |
for dataset in [train_dataset, eval_dataset]:
|
256 |
dataset = dataset.map(
|
257 |
-
partial(add_mem_tokens, mem_freq=50, mem_id=
|
258 |
batched=False,
|
259 |
num_proc=32,
|
260 |
)
|
|
|
245 |
if cfg.is_llama_derived_model and cfg.landmark_attention:
|
246 |
from functools import partial
|
247 |
|
248 |
+
from axolotl.monkeypatch.llama_landmark_attn import (
|
249 |
+
add_mem_tokens,
|
250 |
+
get_mem_id,
|
251 |
+
set_model_mem_id,
|
252 |
+
)
|
253 |
|
254 |
+
set_model_mem_id(model, tokenizer)
|
|
|
255 |
|
256 |
logging.info("Adding landmark attention tokens to dataset")
|
257 |
|
258 |
for dataset in [train_dataset, eval_dataset]:
|
259 |
dataset = dataset.map(
|
260 |
+
partial(add_mem_tokens, mem_freq=50, mem_id=get_mem_id(tokenizer)),
|
261 |
batched=False,
|
262 |
num_proc=32,
|
263 |
)
|
src/axolotl/utils/validation.py
CHANGED
@@ -59,6 +59,11 @@ def validate_config(cfg):
|
|
59 |
if (cfg.base_model and "falcon" in cfg.base_model.lower()) and cfg.fsdp:
|
60 |
raise ValueError("FSDP is not supported for falcon models")
|
61 |
|
|
|
|
|
|
|
|
|
|
|
62 |
if cfg.flash_optimum is True:
|
63 |
if cfg.adapter:
|
64 |
logging.warning(
|
|
|
59 |
if (cfg.base_model and "falcon" in cfg.base_model.lower()) and cfg.fsdp:
|
60 |
raise ValueError("FSDP is not supported for falcon models")
|
61 |
|
62 |
+
if (
|
63 |
+
cfg.base_model and "mpt" in cfg.base_model.lower()
|
64 |
+
) and cfg.gradient_checkpointing:
|
65 |
+
raise ValueError("gradient_checkpointing is not supported for MPT models")
|
66 |
+
|
67 |
if cfg.flash_optimum is True:
|
68 |
if cfg.adapter:
|
69 |
logging.warning(
|
tests/test_validation.py
CHANGED
@@ -199,6 +199,20 @@ class ValidationTest(unittest.TestCase):
|
|
199 |
|
200 |
validate_config(cfg)
|
201 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
202 |
def test_flash_optimum(self):
|
203 |
cfg = DictDefault(
|
204 |
{
|
|
|
199 |
|
200 |
validate_config(cfg)
|
201 |
|
202 |
+
def test_mpt_gradient_checkpointing(self):
|
203 |
+
regex_exp = r".*gradient_checkpointing is not supported for MPT models*"
|
204 |
+
|
205 |
+
# Check for lower-case
|
206 |
+
cfg = DictDefault(
|
207 |
+
{
|
208 |
+
"base_model": "mosaicml/mpt-7b",
|
209 |
+
"gradient_checkpointing": True,
|
210 |
+
}
|
211 |
+
)
|
212 |
+
|
213 |
+
with pytest.raises(ValueError, match=regex_exp):
|
214 |
+
validate_config(cfg)
|
215 |
+
|
216 |
def test_flash_optimum(self):
|
217 |
cfg = DictDefault(
|
218 |
{
|