with save

hft.py

@@ -1,7 +1,7 @@
 import os
 import torch
 import torch.nn as nn
-from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer, GenerationMixin
+from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer
 from datasets import Dataset
 from PIL import Image
 import re
@@ -107,13 +107,15 @@ def tokenize_function(examples):
     return tokenized
 
 def custom_collate_fn(batch):
-    input_ids = torch.stack([torch.tensor(b["input_ids"]) for b in batch]).cpu()
+    input_ids = torch.stack([torch.tensor(b["input_ids"]) for b in batch])
-    attention_mask = torch.stack([torch.tensor(b["attention_mask"]) for b in batch]).cpu()
+    attention_mask = torch.stack([torch.tensor(b["attention_mask"]) for b in batch])
-    labels = torch.stack([torch.tensor(b["labels"]) for b in batch]).cpu()
+    labels = torch.stack([torch.tensor(b["labels"]) for b in batch])
-    source_idx = torch.tensor([b.get("source_idx", -1) for b in batch], dtype=torch.long).cpu()
+    
+    source_idx = torch.tensor([b.get("source_idx", -1) for b in batch], dtype=torch.long)
+    #print("source_idx shape:", source_idx.shape)  # Debugowanie
     return {"input_ids": input_ids, "attention_mask": attention_mask, "labels": labels, "source_idx": source_idx}
 
-class CustomModel(nn.Module, GenerationMixin):
+class CustomModel(nn.Module):
     def __init__(self, model_name, config):
         super().__init__()
         self.base_model = AutoModelForCausalLM.from_pretrained(model_name, config=config)
@@ -122,11 +124,11 @@ class CustomModel(nn.Module, GenerationMixin):
             embedding_dim=config.hidden_size,
             padding_idx=-1
         )
-        self.config = config
-        self.device = next(self.base_model.parameters()).device
         
     def forward(self, input_ids=None, attention_mask=None, labels=None, source_idx=None, **kwargs):
         if source_idx is not None:
+            #print("Max source_idx:", torch.max(source_idx))
+            #print("Num embeddings:", self.source_embedding.num_embeddings)
             source_idx = torch.clamp(source_idx, 0, self.source_embedding.num_embeddings - 1)
             source_embeds = self.source_embedding(source_idx).unsqueeze(1).expand(-1, input_ids.size(1), -1)
             hidden_states = self.base_model.get_input_embeddings()(input_ids) + source_embeds
@@ -136,16 +138,8 @@ class CustomModel(nn.Module, GenerationMixin):
         
         return outputs
 
-    def prepare_inputs_for_generation(self, input_ids, **kwargs):
-        return self.base_model.prepare_inputs_for_generation(input_ids, **kwargs)
-
-    def _reorder_cache(self, past, beam_idx):
-        return self.base_model._reorder_cache(past, beam_idx)
-
 class CustomTrainer(Trainer):
     def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
-        device = next(model.parameters()).device
-        inputs = {k: v.to(device) for k, v in inputs.items()}
         labels = inputs.pop("labels")
         source_idx = inputs.pop("source_idx", None)
         outputs = model(**inputs, labels=labels, source_idx=source_idx)
@@ -166,9 +160,9 @@ tokenized_dataset = dataset.map(tokenize_function, batched=True, batch_size=8)
 
 # Inicjalizacja modelu
 config = AutoModelForCausalLM.from_pretrained(model_name).config
+#print("Vocabulary size:", config.vocab_size)
 model = CustomModel(model_name, config)
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model.to("cpu")  # Zmienione na CPU dla debugowania
-model = model.to(device)
 
 # Konfiguracja treningu
 training_args = TrainingArguments(
@@ -177,13 +171,13 @@ training_args = TrainingArguments(
     per_device_train_batch_size=2,
     gradient_accumulation_steps=4,
     learning_rate=2e-5,
-    fp16=torch.cuda.is_available(),
+    fp16=False,  # Wyłączone dla CPU
     logging_steps=1,
     logging_dir="./logs",
     save_strategy="steps",
     save_steps=1000,
     logging_strategy="no",
-    report_to="none"
+    report_to="none",
 )
 
 # Trening
@@ -197,10 +191,9 @@ trainer.train()
 
 # Funkcja generująca odpowiedź
 def generate_answer(question, model, tokenizer, source_mapper, max_length=200):
-    device = next(model.parameters()).device
+    inputs = tokenizer(question, return_tensors="pt", truncation=True, max_length=512)
-    inputs = tokenizer(question, return_tensors="pt", truncation=True, max_length=512).to(device)
     
-    outputs = model.generate(
+    outputs = model.base_model.generate(
         **inputs,
         max_length=max_length,
         num_return_sequences=1,
@@ -212,12 +205,26 @@ def generate_answer(question, model, tokenizer, source_mapper, max_length=200):
     
     # Pobierz źródło z ostatniego tokena
     last_token_id = outputs.sequences[0][-1].item()
-    source_idx = model.source_embedding.weight.shape[0] - 1  # Tymczasowe rozwiązanie
+    source_idx = model.source_embeddi
-    source = source_mapper.get_source(source_idx)
     
-    return f"{answer}\n\nŹródło: {source if source else 'Opracowanie własne'}"
+# Utwórz katalog do zapisu modelu
+save_directory = "./trained_model/ably.do/hse"
+os.makedirs(save_directory, exist_ok=True)
 
-# Przykład użycia
+# 1. Zapisz wagę modelu
-question = "Ile dni urlopu przysługuje pracownikowi?"
+torch.save(model.state_dict(), os.path.join(save_directory, "hse-nano-mistral.bin"))
-answer = generate_answer(question, model, tokenizer, source_mapper)
+
-print(answer)
+# 2. Zapisz tokenizer
+tokenizer.save_pretrained(save_directory)
+
+# 3. Zapisz mapowanie źródeł
+source_mapper_data = {
+    "source_to_idx": dict(source_mapper.source_to_idx),
+    "idx_to_source": source_mapper.idx_to_source
+}
+
+with open(os.path.join(save_directory, "source_mapper.json"), 'w') as f:
+    json.dump(source_mapper_data, f)
+
+# 4. Zapisz konfigurację modelu (opcjonalnie, ale zalecane)
+model.base_model.config.save_pretrained(save_directory)