2603

pipeline_sdxs.py

@@ -158,7 +158,7 @@ class SdxsPipeline(DiffusionPipeline):
                 messages = [{"role": "user", "content": [{"type": "text", "text": sys_msg + p}]}]
                 
                 # Используем Qwen-Instruct формат (apply_chat_template сам подставит system/user/assistant токены)
-                inputs = self.tokenizer2.apply_chat_template(
+                inputs = self.tokenizer.apply_chat_template(
                     messages,
                     tokenize=True,
                     add_generation_prompt=True,
@@ -166,15 +166,15 @@ class SdxsPipeline(DiffusionPipeline):
                     return_tensors="pt"
                 ).to(device)
                 
-                generated_ids = self.text_encoder2.generate(
-                    **inputs, max_new_tokens=self.text_encoder.config.max_position_embeddings, do_sample=True,temperature = 0.7
+                generated_ids = self.text_encoder.generate(
+                    **inputs, max_new_tokens=255, do_sample=True,temperature = 0.7
                 )
                 
                 # Обрезаем входные токены из ответа
                 generated_ids_trimmed = [
                     out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
                 ]
-                output_text = self.tokenizer2.batch_decode(
+                output_text = self.tokenizer.batch_decode(
                     generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
                 )
                 refined_list.append(output_text)

test.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5b1dc22c65c672008be6159c53028670acace729df6e26c3d1199a8929a07060
-size 6130032
+oid sha256:14eeb3271e1dc39b928d1006c55c727b28b1c5c320fe8d16467b7542bef7196a
+size 6720678

train.py

@@ -43,7 +43,7 @@ batch_size = local_bs * num_gpus
 print(f"GPUs: {num_gpus}, Local BS: {local_bs}, Global BS: {local_bs * num_gpus}")
 base_learning_rate = 2e-5
 min_learning_rate = 3e-6
-num_epochs = num_gpus #8 * max(1, int(num_gpus / 2))
+num_epochs = 1 #num_gpus #8 * max(1, int(num_gpus / 2))
 sample_interval_share = 20
 cfg_dropout = 0.10
 max_length = 248
@@ -149,8 +149,8 @@ if accelerator.is_main_process:
 
 # --------------------------- Загрузка моделей ---------------------------
 vae = AutoencoderKL.from_pretrained("vae", torch_dtype=dtype).to(device).eval()
-tokenizer = Qwen3_5Tokenizer.from_pretrained("tokenizer2")
-text_encoder = Qwen3_5ForConditionalGeneration.from_pretrained("text_encoder2", torch_dtype=torch.float16).to(device).eval()
+tokenizer = Qwen3_5Tokenizer.from_pretrained("tokenizer")
+text_encoder = Qwen3_5ForConditionalGeneration.from_pretrained("text_encoder", torch_dtype=torch.float16).to(device).eval()
 scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained("scheduler")
 
 def encode_texts(texts, max_length=max_length):

unet/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a1a8e5563a08991becd74bb41292f800b383fb757cafe97ce5b87b6004dfb87a
+oid sha256:6c2f0d7d80722dbfc05048a68068d7be8c4c8866e5890d63a6fcc5609316f538
 size 6318956752