Merge branch 'main' into hf-ui-demo

llama_lora/ui/finetune_ui.py

@@ -316,6 +316,13 @@ def do_train(
             resume_from_checkpoint = os.path.join(Global.data_dir, "lora_models", continue_from_model)
             if continue_from_checkpoint:
                 resume_from_checkpoint = os.path.join(resume_from_checkpoint, continue_from_checkpoint)
+                will_be_resume_from_checkpoint_file = os.path.join(resume_from_checkpoint, "pytorch_model.bin")
+                if not os.path.exists(will_be_resume_from_checkpoint_file):
+                    raise ValueError(f"Unable to resume from checkpoint {continue_from_model}/{continue_from_checkpoint}. Resuming is only possible from checkpoints stored locally in the data directory. Please ensure that the file '{will_be_resume_from_checkpoint_file}' exists.")
+            else:
+                will_be_resume_from_checkpoint_file = os.path.join(resume_from_checkpoint, "adapter_model.bin")
+                if not os.path.exists(will_be_resume_from_checkpoint_file):
+                    raise ValueError(f"Unable to continue from model {continue_from_model}. Continuation is only possible from models stored locally in the data directory. Please ensure that the file '{will_be_resume_from_checkpoint_file}' exists.")
 
         output_dir = os.path.join(Global.data_dir, "lora_models", model_name)
         if os.path.exists(output_dir):
@@ -858,7 +865,8 @@ def finetune_ui():
                 evaluate_data_count = gr.Slider(
                     minimum=0, maximum=1, step=1, value=0,
                     label="Evaluation Data Count",
-                    info="The number of data to be used for evaluation. This amount of data will not be used for training and will be used to assess the performance of the model during the process."
+                    info="The number of data to be used for evaluation. This specific amount of data will be randomly chosen from the training dataset for evaluating the model's performance during the process, without contributing to the actual training.",
+                    elem_id="finetune_evaluate_data_count"
                 )
 
                 with gr.Box(elem_id="finetune_continue_from_model_box"):
@@ -870,7 +878,10 @@ def finetune_ui():
                             elem_id="finetune_continue_from_model"
                         )
                         continue_from_checkpoint = gr.Dropdown(
-                            value="-", label="Checkpoint", choices=["-"])
+                            value="-",
+                            label="Resume from Checkpoint",
+                            choices=["-"],
+                            elem_id="finetune_continue_from_checkpoint")
                     with gr.Column():
                         load_params_from_model_btn = gr.Button(
                             "Load training parameters from selected model", visible=False)
@@ -911,8 +922,6 @@ def finetune_ui():
                     info="The dropout probability for LoRA, which controls the fraction of LoRA parameters that are set to zero during training. A larger lora_dropout increases the regularization effect of LoRA but also increases the risk of underfitting."
                 )
 
-                lora_target_module_choices = gr.State(value=default_lora_target_module_choices)
-
                 lora_target_modules = gr.CheckboxGroup(
                     label="LoRA Target Modules",
                     choices=default_lora_target_module_choices,
@@ -920,6 +929,7 @@ def finetune_ui():
                     info="Modules to replace with LoRA.",
                     elem_id="finetune_lora_target_modules"
                 )
+                lora_target_module_choices = gr.State(value=default_lora_target_module_choices)
                 with gr.Box(elem_id="finetune_lora_target_modules_add_box"):
                     with gr.Row():
                         lora_target_modules_add = gr.Textbox(
@@ -1136,6 +1146,14 @@ def finetune_ui():
             'Press to load a sample dataset of the current selected format into the textbox.',
         });
 
+        tippy('#finetune_evaluate_data_count', {
+          placement: 'bottom',
+          delay: [500, 0],
+          animation: 'scale-subtle',
+          content:
+            'While setting a value larger than 0, the checkpoint with the lowest loss on the evaluation data will be saved as the final trained model, thereby helping to prevent overfitting.',
+        });
+
         tippy('#finetune_save_total_limit', {
           placement: 'bottom',
           delay: [500, 0],
@@ -1165,6 +1183,24 @@ def finetune_ui():
           content:
             'The name of the new LoRA model. Must be unique.',
         });
+
+        tippy('#finetune_continue_from_model', {
+          placement: 'bottom',
+          delay: [500, 0],
+          animation: 'scale-subtle',
+          content:
+            'Select a LoRA model to train a new model on top of that model.<br /><br />💡 To use the same training parameters of a previously trained model, select it here and click the <code>Load training parameters from selected model</code> button, then un-select it.',
+          allowHTML: true,
+        });
+
+        tippy('#finetune_continue_from_checkpoint', {
+          placement: 'bottom',
+          delay: [500, 0],
+          animation: 'scale-subtle',
+          content:
+            'If a checkpoint is selected, training will resume from that specific checkpoint, bypassing any previously completed steps up to the checkpoint\\'s moment. <br /><br />💡 Use this option to resume an unfinished training session. Remember to click the <code>Load training parameters from selected model</code> button to load the training parameters of the selected model.',
+          allowHTML: true,
+        });
       }, 100);
 
       // Show/hide start and stop button base on the state.

llama_lora/ui/inference_ui.py

@@ -104,11 +104,12 @@ def do_inference(
         model = get_model(base_model_name, lora_model_name)
 
         generation_config = GenerationConfig(
-            temperature=temperature,
+            temperature=float(temperature),  # to avoid ValueError('`temperature` has to be a strictly positive float, but is 2')
             top_p=top_p,
             top_k=top_k,
             repetition_penalty=repetition_penalty,
             num_beams=num_beams,
+            do_sample=temperature > 0,  # https://github.com/huggingface/transformers/issues/22405#issuecomment-1485527953
         )
 
         def ui_generation_stopping_criteria(input_ids, score, **kwargs):
@@ -190,6 +191,23 @@ def reload_selections(current_lora_model, current_prompt_template):
             gr.Dropdown.update(choices=available_template_names_with_none, value=current_prompt_template))
 
 
+def get_warning_message_for_lora_model_and_prompt_template(lora_model, prompt_template):
+    messages = []
+
+    lora_mode_info = get_info_of_available_lora_model(lora_model)
+
+    if lora_mode_info and isinstance(lora_mode_info, dict):
+        model_base_model = lora_mode_info.get("base_model")
+        if model_base_model and model_base_model != Global.base_model_name:
+            messages.append(f"⚠️ This model was trained on top of base model `{model_base_model}`, it might not work properly with the selected base model `{Global.base_model_name}`.")
+
+        model_prompt_template = lora_mode_info.get("prompt_template")
+        if model_prompt_template and model_prompt_template != prompt_template:
+            messages.append(f"This model was trained with prompt template `{model_prompt_template}`.")
+
+    return " ".join(messages)
+
+
 def handle_prompt_template_change(prompt_template, lora_model):
     prompter = Prompter(prompt_template)
     var_names = prompter.get_variable_names()
@@ -203,37 +221,32 @@ def handle_prompt_template_change(prompt_template, lora_model):
 
     model_prompt_template_message_update = gr.Markdown.update(
         "", visible=False)
-    lora_mode_info = get_info_of_available_lora_model(lora_model)
-    if lora_mode_info and isinstance(lora_mode_info, dict):
-        model_base_model = lora_mode_info.get("base_model")
-        model_prompt_template = lora_mode_info.get("prompt_template")
-        if model_base_model and model_base_model != Global.base_model_name:
-            model_prompt_template_message_update = gr.Markdown.update(
-                f"⚠️ This model was trained on top of base model `{model_base_model}`, it might not work properly with the selected base model `{Global.base_model_name}`.", visible=True)
-        elif model_prompt_template and model_prompt_template != prompt_template:
-            model_prompt_template_message_update = gr.Markdown.update(
-                f"This model was trained with prompt template `{model_prompt_template}`.", visible=True)
+    warning_message = get_warning_message_for_lora_model_and_prompt_template(lora_model, prompt_template)
+    if warning_message:
+        model_prompt_template_message_update = gr.Markdown.update(
+            warning_message, visible=True)
 
     return [model_prompt_template_message_update] + gr_updates
 
 
 def handle_lora_model_change(lora_model, prompt_template):
     lora_mode_info = get_info_of_available_lora_model(lora_model)
-    if not lora_mode_info:
-        return gr.Markdown.update("", visible=False), prompt_template
-
-    if not isinstance(lora_mode_info, dict):
-        return gr.Markdown.update("", visible=False), prompt_template
 
-    model_prompt_template = lora_mode_info.get("prompt_template")
-    if not model_prompt_template:
-        return gr.Markdown.update("", visible=False), prompt_template
+    if lora_mode_info and isinstance(lora_mode_info, dict):
+        model_prompt_template = lora_mode_info.get("prompt_template")
+        if model_prompt_template:
+            available_template_names = get_available_template_names()
+            if model_prompt_template in available_template_names:
+                prompt_template = model_prompt_template
 
-    available_template_names = get_available_template_names()
-    if model_prompt_template in available_template_names:
-        return gr.Markdown.update("", visible=False), model_prompt_template
+    model_prompt_template_message_update = gr.Markdown.update(
+        "", visible=False)
+    warning_message = get_warning_message_for_lora_model_and_prompt_template(lora_model, prompt_template)
+    if warning_message:
+        model_prompt_template_message_update = gr.Markdown.update(
+            warning_message, visible=True)
 
-    return gr.Markdown.update(f"Trained with prompt template `{model_prompt_template}`", visible=True), prompt_template
+    return model_prompt_template_message_update, prompt_template
 
 
 def update_prompt_preview(prompt_template,
@@ -313,7 +326,7 @@ def inference_ui():
                 # with gr.Column():
                 with gr.Accordion("Options", open=True, elem_id="inference_options_accordion"):
                     temperature = gr.Slider(
-                        minimum=0, maximum=1, value=0.1, step=0.01,
+                        minimum=0, maximum=2, value=0, step=0.01,
                         label="Temperature",
                         elem_id="inference_temperature"
                     )
@@ -525,7 +538,8 @@ def inference_ui():
             delay: [500, 0],
             animation: 'scale-subtle',
             content:
-              'Controls randomness: Lowering results in less random completions. Higher values (e.g., 1.0) make the model generate more diverse and random outputs. As the temperature approaches zero, the model will become deterministic and repetitive.',
+              '<strong>Controls randomness</strong>: Higher values (e.g., <code>1.0</code>) make the model generate more diverse and random outputs. As the temperature approaches zero, the model will become deterministic and repetitive.<br /><i>Setting a value larger then <code>0</code> will enable sampling.</i>',
+            allowHTML: true,
           });
 
           tippy('#inference_top_p', {
@@ -533,7 +547,8 @@ def inference_ui():
             delay: [500, 0],
             animation: 'scale-subtle',
             content:
-              'Controls diversity via nucleus sampling: only the tokens whose cumulative probability exceeds "top_p" are considered. 0.5 means half of all likelihood-weighted options are considered.',
+              'Controls diversity via nucleus sampling: only the tokens whose cumulative probability exceeds <code>top_p</code> are considered. <code>0.5</code> means half of all likelihood-weighted options are considered.<br />Will only take effect if Temperature is set to > 0.',
+            allowHTML: true,
           });
 
           tippy('#inference_top_k', {
@@ -541,7 +556,8 @@ def inference_ui():
             delay: [500, 0],
             animation: 'scale-subtle',
             content:
-              'Controls diversity of the generated text by only considering the "top_k" tokens with the highest probabilities. This method can lead to more focused and coherent outputs by reducing the impact of low probability tokens.',
+              'Controls diversity of the generated text by only considering the <code>top_k</code> tokens with the highest probabilities. This method can lead to more focused and coherent outputs by reducing the impact of low probability tokens.<br />Will only take effect if Temperature is set to > 0.',
+            allowHTML: true,
           });
 
           tippy('#inference_beams', {

llama_lora/ui/main_page.py

@@ -134,6 +134,10 @@ def main_page_custom_css():
         border: 1px solid var(--border-color-primary);
         border-radius: 4px;
         box-shadow: 0 2px 20px rgba(5,5,5,.08);
+        /* box-shadow: var(--shadow-drop-lg); */
+    }
+    body.dark .tippy-box {
+        box-shadow: 0 0 8px rgba(160,160,160,0.12);
     }
     .tippy-arrow {
         color: var(--block-background-fill);
@@ -144,6 +148,45 @@ def main_page_custom_css():
         font-weight: 100;
     }
 
+    .tippy-arrow::before {
+        z-index: 1;
+    }
+    .tippy-arrow::after {
+        content: "";
+        position: absolute;
+        z-index: -1;
+        border-color: transparent;
+        border-style: solid;
+    }
+    .tippy-box[data-placement^=top]> .tippy-arrow::after {
+        bottom: -9px;
+        left: -1px;
+        border-width: 9px 9px 0;
+        border-top-color: var(--border-color-primary);
+        transform-origin: center top;
+    }
+    .tippy-box[data-placement^=bottom]> .tippy-arrow::after {
+        top: -9px;
+        left: -1px;
+        border-width: 0 9px 9px;
+        border-bottom-color: var(--border-color-primary);
+        transform-origin: center bottom;
+    }
+    .tippy-box[data-placement^=left]> .tippy-arrow:after {
+        border-width: 9px 0 9px 9px;
+        border-left-color: var(--border-color-primary);
+        top: -1px;
+        right: -9px;
+        transform-origin: center left;
+    }
+    .tippy-box[data-placement^=right]> .tippy-arrow::after {
+        top: -1px;
+        left: -9px;
+        border-width: 9px 9px 9px 0;
+        border-right-color: var(--border-color-primary);
+        transform-origin: center right;
+    }
+
     /*
     .codemirror-wrapper .cm-editor .cm-gutters {
         background-color: var(--background-fill-secondary);
@@ -248,12 +291,18 @@ def main_page_custom_css():
     #inference_lora_model_prompt_template_message:not(.hidden) + #inference_lora_model {
         padding-bottom: 28px;
     }
+    #inference_lora_model_group {
+        flex-direction: column-reverse;
+        border-width: var(--block-border-width);
+        border-color: var(--block-border-color);
+    }
+    #inference_lora_model_group #inference_lora_model {
+        border: 0;
+    }
     #inference_lora_model_group > #inference_lora_model_prompt_template_message {
-        position: absolute;
-        bottom: 8px;
-        left: 20px;
-        z-index: 61;
-        width: 999px;
+        padding: var(--block-padding) !important;
+        margin-top: -50px !important;
+        margin-left: 4px !important;
         font-size: 12px;
         opacity: 0.7;
     }
@@ -608,7 +657,7 @@ def main_page_custom_css():
     }
 
     @media screen and (max-width: 392px) {
-        #inference_lora_model, #finetune_template {
+        #inference_lora_model, #inference_lora_model_group, #finetune_template {
             border-bottom-left-radius: 0;
             border-bottom-right-radius: 0;
         }

lora_models/unhelpful-ai-v01/finetune_params.json

@@ -1,8 +1,9 @@
 {
-  "num_train_epochs": 16,
+  "num_train_epochs": 8,
   "learning_rate": 0.0003,
   "cutoff_len": 512,
-  "lora_r": 12,
+  "val_set_size": 0,
+  "lora_r": 16,
   "lora_alpha": 32,
   "lora_dropout": 0.05,
   "lora_target_modules": [
@@ -11,9 +12,5 @@
     "k_proj",
     "o_proj"
   ],
-  "train_on_inputs": false,
-  "group_by_length": false,
-  "save_steps": 500,
-  "save_total_limit": 5,
-  "logging_steps": 10
+  "train_on_inputs": false
 }