ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix #944

Browse Source 
Former-commit-id: 032245647848aaa4167086636b6c985268c5fee3
This commit is contained in:
hiyouga
2023-09-21 19:51:02 +08:00
parent 95c0d9ab24
commit dc68c313ee
11 changed files with 116 additions and 101 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
src/llmtuner/webui/components/eval.py
View File

@@ -28,12 +28,16 @@ def create_eval_tab(top_elems: Dict[str, "Component"], runner: "Runner") -> Dict
)
with gr.Row():
max_source_length = gr.Slider(value=512, minimum=4, maximum=4096, step=1)
max_target_length = gr.Slider(value=512, minimum=4, maximum=4096, step=1)
cutoff_len = gr.Slider(value=1024, minimum=4, maximum=8192, step=1)
max_samples = gr.Textbox(value="100000")
batch_size = gr.Slider(value=8, minimum=1, maximum=512, step=1)
predict = gr.Checkbox(value=True)
with gr.Row():
max_new_tokens = gr.Slider(10, 2048, value=128, step=1)
top_p = gr.Slider(0.01, 1, value=0.7, step=0.01)
temperature = gr.Slider(0.01, 1.5, value=0.95, step=0.01)
with gr.Row():
cmd_preview_btn = gr.Button()
start_btn = gr.Button()
@@ -55,11 +59,13 @@ def create_eval_tab(top_elems: Dict[str, "Component"], runner: "Runner") -> Dict
top_elems["system_prompt"],
dataset_dir,
dataset,
max_source_length,
max_target_length,
cutoff_len,
max_samples,
batch_size,
predict
predict,
max_new_tokens,
top_p,
temperature
]
output_components = [
@@ -78,11 +84,13 @@ def create_eval_tab(top_elems: Dict[str, "Component"], runner: "Runner") -> Dict
preview_count=preview_count,
preview_samples=preview_samples,
close_btn=close_btn,
max_source_length=max_source_length,
max_target_length=max_target_length,
cutoff_len=cutoff_len,
max_samples=max_samples,
batch_size=batch_size,
predict=predict,
max_new_tokens=max_new_tokens,
top_p=top_p,
temperature=temperature,
cmd_preview_btn=cmd_preview_btn,
start_btn=start_btn,
stop_btn=stop_btn,

21
src/llmtuner/webui/components/train.py
View File

@@ -35,11 +35,11 @@ def create_train_tab(top_elems: Dict[str, "Component"], runner: "Runner") -> Dic
)
with gr.Row():
max_source_length = gr.Slider(value=512, minimum=4, maximum=4096, step=1)
max_target_length = gr.Slider(value=512, minimum=4, maximum=4096, step=1)
cutoff_len = gr.Slider(value=1024, minimum=4, maximum=8192, step=1)
learning_rate = gr.Textbox(value="5e-5")
num_train_epochs = gr.Textbox(value="3.0")
max_samples = gr.Textbox(value="100000")
compute_type = gr.Radio(choices=["fp16", "bf16"], value="fp16")
with gr.Row():
batch_size = gr.Slider(value=4, minimum=1, maximum=512, step=1)
@@ -55,7 +55,8 @@ def create_train_tab(top_elems: Dict[str, "Component"], runner: "Runner") -> Dic
logging_steps = gr.Slider(value=5, minimum=5, maximum=1000, step=5)
save_steps = gr.Slider(value=100, minimum=10, maximum=5000, step=10)
warmup_steps = gr.Slider(value=0, minimum=0, maximum=5000, step=1)
compute_type = gr.Radio(choices=["fp16", "bf16"], value="fp16")
flash_attn = gr.Checkbox(value=False)
rope_scaling = gr.Checkbox(value=False)
with gr.Accordion(label="LoRA config", open=False) as lora_tab:
with gr.Row():
@@ -107,11 +108,11 @@ def create_train_tab(top_elems: Dict[str, "Component"], runner: "Runner") -> Dic
training_stage,
dataset_dir,
dataset,
max_source_length,
max_target_length,
cutoff_len,
learning_rate,
num_train_epochs,
max_samples,
compute_type,
batch_size,
gradient_accumulation_steps,
lr_scheduler_type,
@@ -120,7 +121,8 @@ def create_train_tab(top_elems: Dict[str, "Component"], runner: "Runner") -> Dic
logging_steps,
save_steps,
warmup_steps,
compute_type,
flash_attn,
rope_scaling,
lora_rank,
lora_dropout,
lora_target,
@@ -151,11 +153,11 @@ def create_train_tab(top_elems: Dict[str, "Component"], runner: "Runner") -> Dic
preview_count=preview_count,
preview_samples=preview_samples,
close_btn=close_btn,
max_source_length=max_source_length,
max_target_length=max_target_length,
cutoff_len=cutoff_len,
learning_rate=learning_rate,
num_train_epochs=num_train_epochs,
max_samples=max_samples,
compute_type=compute_type,
batch_size=batch_size,
gradient_accumulation_steps=gradient_accumulation_steps,
lr_scheduler_type=lr_scheduler_type,
@@ -165,7 +167,8 @@ def create_train_tab(top_elems: Dict[str, "Component"], runner: "Runner") -> Dic
logging_steps=logging_steps,
save_steps=save_steps,
warmup_steps=warmup_steps,
compute_type=compute_type,
flash_attn=flash_attn,
rope_scaling=rope_scaling,
lora_tab=lora_tab,
lora_rank=lora_rank,
lora_dropout=lora_dropout,

44
src/llmtuner/webui/locales.py
View File

@@ -147,26 +147,16 @@ LOCALES = {
"value": "关闭"
}
},
"max_source_length": {
"cutoff_len": {
"en": {
"label": "Max source length",
"info": "Max tokens in source sequence."
"label": "Cutoff length",
"info": "Max tokens in input sequence."
},
"zh": {
"label": "输入序列最大长度",
"label": "截断长度",
"info": "输入序列分词后的最大长度。"
}
},
"max_target_length": {
"en": {
"label": "Max target length",
"info": "Max tokens in target sequence."
},
"zh": {
"label": "输出序列最大长度",
"info": "输出序列分词后的最大长度。"
}
},
"learning_rate": {
"en": {
"label": "Learning rate",
@@ -197,6 +187,16 @@ LOCALES = {
"info": "每个数据集最多使用的样本数。"
}
},
"compute_type": {
"en": {
"label": "Compute type",
"info": "Whether to use fp16 or bf16 mixed precision training."
},
"zh": {
"label": "计算类型",
"info": "是否启用 FP16 或 BF16 混合精度训练。"
}
},
"batch_size": {
"en": {
"label": "Batch size",
@@ -277,14 +277,20 @@ LOCALES = {
"info": "学习率预热采用的步数。"
}
},
"compute_type": {
"flash_attn": {
"en": {
"label": "Compute type",
"info": "Whether to use fp16 or bf16 mixed precision training."
"label": "Use FlashAttention-2"
},
"zh": {
"label": "计算类型",
"info": "是否启用 FP16 或 BF16 混合精度训练。"
"label": "使用 FlashAttention-2"
}
},
"rope_scaling": {
"en": {
"label": "Use RoPE scaling"
},
"zh": {
"label": "使用 RoPE 插值"
}
},
"lora_tab": {

26
src/llmtuner/webui/runner.py
View File

@@ -73,11 +73,11 @@ class Runner:
training_stage: str,
dataset_dir: str,
dataset: List[str],
max_source_length: int,
max_target_length: int,
cutoff_len: int,
learning_rate: str,
num_train_epochs: str,
max_samples: str,
compute_type: str,
batch_size: int,
gradient_accumulation_steps: int,
lr_scheduler_type: str,
@@ -86,7 +86,8 @@ class Runner:
logging_steps: int,
save_steps: int,
warmup_steps: int,
compute_type: str,
flash_attn: bool,
rope_scaling: bool,
lora_rank: int,
lora_dropout: float,
lora_target: str,
@@ -120,8 +121,7 @@ class Runner:
system_prompt=system_prompt,
dataset_dir=dataset_dir,
dataset=",".join(dataset),
max_source_length=max_source_length,
max_target_length=max_target_length,
cutoff_len=cutoff_len,
learning_rate=float(learning_rate),
num_train_epochs=float(num_train_epochs),
max_samples=int(max_samples),
@@ -132,6 +132,8 @@ class Runner:
logging_steps=logging_steps,
save_steps=save_steps,
warmup_steps=warmup_steps,
flash_attn=flash_attn,
rope_scaling="linear" if rope_scaling else None,
lora_rank=lora_rank,
lora_dropout=lora_dropout,
lora_target=lora_target or DEFAULT_MODULE.get(model_name.split("-")[0], "q_proj,v_proj"),
@@ -168,11 +170,13 @@ class Runner:
system_prompt: str,
dataset_dir: str,
dataset: List[str],
max_source_length: int,
max_target_length: int,
cutoff_len: int,
max_samples: str,
batch_size: int,
predict: bool
predict: bool,
max_new_tokens: int,
top_p: float,
temperature: float
) -> Tuple[str, str, List[str], str, Dict[str, Any]]:
if checkpoints:
checkpoint_dir = ",".join(
@@ -200,10 +204,12 @@ class Runner:
system_prompt=system_prompt,
dataset_dir=dataset_dir,
dataset=",".join(dataset),
max_source_length=max_source_length,
max_target_length=max_target_length,
cutoff_len=cutoff_len,
max_samples=int(max_samples),
per_device_eval_batch_size=batch_size,
max_new_tokens=max_new_tokens,
top_p=top_p,
temperature=temperature,
output_dir=output_dir
)