不同模型保存为gguf格式的方法不同?

003加载自己的数据集微调.py

@@ -84,6 +84,8 @@ trainer = SFTTrainer(
 trainer.train()
 
 
+
+
 messages = [
     {"role" : "user", "content" : "请介绍一下昊天"}
 ]
@@ -110,3 +112,7 @@ model.save_pretrained_gguf(
     # quantization_method="q4_k_m",  # 或 "q8_0"  # 量化模式--默认 q8_0, 可选f16, "q4_k_m", "q8_0", "q5_k_m",
     maximum_memory_usage=0.7  # 限制使用 GPU 显存为总容量的 50%
 )
+
+# print("*"*150, " Modefile")
+# print(tokenizer._ollama_modelfile)
+# print("*"*150)

003微调deepseek.py

@@ -1,8 +1,8 @@
 from datasets import load_dataset
 from unsloth import FastLanguageModel
 import torch
-# import os
-# os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'
+import os
+os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'
 
 # 加载 jsonl 文件
 dataset = load_dataset("json", data_files="dataset/test_dataset.jsonl", split="train")
@@ -30,8 +30,9 @@ quant_cfg = BitsAndBytesConfig(
 )
 
 model, tokenizer = FastLanguageModel.from_pretrained(
-    "deepseek-ai/DeepSeek-V2-Lite",
-    max_seq_length = 1024,
+    # "deepseek-ai/DeepSeek-V2-Lite",
+    "unsloth/Qwen3-14B-unsloth-bnb-4bit",
+    max_seq_length = 2048,
     load_in_4bit=False,
     load_in_8bit=True,
     quantization_config=quant_cfg,
@@ -44,10 +45,10 @@ model, tokenizer = FastLanguageModel.from_pretrained(
 
 model = FastLanguageModel.get_peft_model(
     model,
-    r = 8,           # Choose any number > 0! Suggested 8, 16, 32, 64, 128
+    r = 16,           # Choose any number > 0! Suggested 8, 16, 32, 64, 128
     target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                       "gate_proj", "up_proj", "down_proj",],
-    lora_alpha = 8,  # Best to choose alpha = rank or rank*2
+    lora_alpha = 16,  # Best to choose alpha = rank or rank*2
     lora_dropout = 0, # Supports any, but = 0 is optimized
     bias = "none",    # Supports any, but = "none" is optimized
     # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
@@ -73,7 +74,7 @@ trainer = SFTTrainer(
         gradient_accumulation_steps=8,
         warmup_steps = 5,
         # num_train_epochs = 1, # Set this for 1 full training run.
-        max_steps = 30,
+        max_steps = 5,
         learning_rate = 2e-4, # Reduce to 2e-5 for long training runs
         logging_steps = 1,
         optim = "adamw_8bit",
@@ -108,10 +109,26 @@ trainer.train()
 
 # model.cpu()
 
+# model.save_pretrained_gguf(
+#     "DeepSeek-V2-Lite", 
+#     # tokenizer = tokenizer, 
+#     # quantization_method="q4_k_m",  # 或 "q8_0"  # 量化模式--默认 q8_0, 可选f16, "q4_k_m", "q8_0", "q5_k_m",
+#     # quantization_type="q4_k_m"
+#     # maximum_memory_usage=0.7  # 限制使用 GPU 显存为总容量的 50%
+# )
+
+
+# print("*"*150, " Modefile")
+# print(tokenizer._ollama_modelfile)
+# print("*"*150)
+
+# model.save_pretrained_merged("merged_model", maximum_memory_usage=0.7)
+# tokenizer.save_pretrained("merged_model")
+
+
 model.save_pretrained_gguf(
-    "DeepSeek-V2-Lite", 
+    "Qwen3-14B-unsloth-bnb-4bit", 
     tokenizer, 
     # quantization_method="q4_k_m",  # 或 "q8_0"  # 量化模式--默认 q8_0, 可选f16, "q4_k_m", "q8_0", "q5_k_m",
-    # quantization_type="q4_k_m"
-    # maximum_memory_usage=0.7  # 限制使用 GPU 显存为总容量的 50%
-)
+    maximum_memory_usage=0.7  # 限制使用 GPU 显存为总容量的 50%
+)

Modelfile

@@ -1,4 +1,4 @@
-FROM ./unsloth.Q8_0.gguf
+FROM .Qwen3-8B/unsloth.Q4_K_M.gguf
 
 TEMPLATE """{{ if .System }}<|im_start|>system
 {{ .System }}<|im_end|>
@@ -6,3 +6,8 @@ TEMPLATE """{{ if .System }}<|im_start|>system
 {{ .Prompt }}<|im_end|>
 {{ end }}<|im_start|>assistant
 """
+PARAMETER stop "<|im_start|>"
+PARAMETER stop "<|im_end|>"
+PARAMETER temperature 0.7
+PARAMETER top_p 0.9
+PARAMETER top_k 40

unsloth_compiled_cache/UnslothAlignPropTrainer.py

@@ -130,7 +130,7 @@ class UnslothAlignPropConfig(AlignPropConfig):
     )
     def __init__(
         self,
-        exp_name = '003微调deepseek',
+        exp_name = '003加载自己的数据集微调',
         run_name = '',
         seed = 3407,
         log_with = None,

unsloth_compiled_cache/UnslothDDPOTrainer.py

@@ -146,7 +146,7 @@ class UnslothDDPOConfig(DDPOConfig):
     )
     def __init__(
         self,
-        exp_name = '003微调deepseek',
+        exp_name = '003加载自己的数据集微调',
         run_name = '',
         seed = 3407,
         log_with = None,