Support GLM-4/GLM-4-0414/GLM-Z1

src/llamafactory/train/kto/workflow.py

@@ -82,7 +82,13 @@ def run_kto(
         trainer.save_metrics("train", train_result.metrics)
         trainer.save_state()
         if trainer.is_world_process_zero() and finetuning_args.plot_loss:
-            plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "rewards/chosen"])
+            keys = ["loss", "rewards/chosen"]
+            if isinstance(dataset_module.get("eval_dataset"), dict):
+                keys += [f"eval_{key}_loss" for key in dataset_module["eval_dataset"].keys()]
+            else:
+                keys += ["eval_loss"]
+
+            plot_loss(training_args.output_dir, keys=keys)
 
     # Evaluation
     if training_args.do_eval:

src/llamafactory/train/pt/workflow.py

@@ -66,7 +66,13 @@ def run_pt(
         trainer.save_metrics("train", train_result.metrics)
         trainer.save_state()
         if trainer.is_world_process_zero() and finetuning_args.plot_loss:
-            plot_loss(training_args.output_dir, keys=["loss", "eval_loss"])
+            keys = ["loss"]
+            if isinstance(dataset_module.get("eval_dataset"), dict):
+                keys += [f"eval_{key}_loss" for key in dataset_module["eval_dataset"].keys()]
+            else:
+                keys += ["eval_loss"]
+
+            plot_loss(training_args.output_dir, keys=keys)
 
     # Evaluation
     if training_args.do_eval:

src/llamafactory/train/rm/workflow.py

@@ -74,7 +74,15 @@ def run_rm(
         trainer.save_metrics("train", train_result.metrics)
         trainer.save_state()
         if trainer.is_world_process_zero() and finetuning_args.plot_loss:
-            plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "eval_accuracy"])
+            keys = ["loss"]
+            if isinstance(dataset_module.get("eval_dataset"), dict):
+                keys += sum(
+                    [[f"eval_{key}_loss", f"eval_{key}_accuracy"] for key in dataset_module["eval_dataset"].keys()], []
+                )
+            else:
+                keys += ["eval_loss", "eval_accuracy"]
+
+            plot_loss(training_args.output_dir, keys=keys)
 
     # Evaluation
     if training_args.do_eval:

src/llamafactory/train/sft/workflow.py

@@ -110,7 +110,15 @@ def run_sft(
         trainer.save_metrics("train", train_result.metrics)
         trainer.save_state()
         if trainer.is_world_process_zero() and finetuning_args.plot_loss:
-            plot_loss(training_args.output_dir, keys=["loss", "eval_loss", "eval_accuracy"])
+            keys = ["loss"]
+            if isinstance(dataset_module.get("eval_dataset"), dict):
+                keys += sum(
+                    [[f"eval_{key}_loss", f"eval_{key}_accuracy"] for key in dataset_module["eval_dataset"].keys()], []
+                )
+            else:
+                keys += ["eval_loss", "eval_accuracy"]
+
+            plot_loss(training_args.output_dir, keys=keys)
 
     if training_args.predict_with_generate:
         tokenizer.padding_side = "left"  # use left-padding in generation

src/llamafactory/train/trainer_utils.py

@@ -48,6 +48,7 @@ if is_apollo_available():
 
 
 if is_ray_available():
+    import ray
     from ray.train import RunConfig, ScalingConfig
     from ray.train.torch import TorchTrainer
 
@@ -644,6 +645,9 @@ def get_ray_trainer(
     if not ray_args.use_ray:
         raise ValueError("Ray was not enabled. Please set `USE_RAY=1` to enable ray.")
 
+    if ray_args.ray_init_kwargs is not None:
+        ray.init(**ray_args.ray_init_kwargs)
+
     trainer = TorchTrainer(
         training_function,
         train_loop_config=train_loop_config,

src/llamafactory/webui/components/export.py

@@ -111,7 +111,7 @@ def create_export_tab(engine: "Engine") -> dict[str, "Component"]:
     with gr.Row():
         export_size = gr.Slider(minimum=1, maximum=100, value=5, step=1)
         export_quantization_bit = gr.Dropdown(choices=["none"] + GPTQ_BITS, value="none")
-        export_quantization_dataset = gr.Textbox(value="data/c4_demo.json")
+        export_quantization_dataset = gr.Textbox(value="data/c4_demo.jsonl")
         export_device = gr.Radio(choices=["cpu", "auto"], value="cpu")
         export_legacy_format = gr.Checkbox()
 

src/llamafactory/webui/components/top.py

@@ -42,7 +42,7 @@ def create_top() -> dict[str, "Component"]:
 
     with gr.Row():
         quantization_bit = gr.Dropdown(choices=["none", "8", "4"], value="none", allow_custom_value=True)
-        quantization_method = gr.Dropdown(choices=["bitsandbytes", "hqq", "eetq"], value="bitsandbytes")
+        quantization_method = gr.Dropdown(choices=["bnb", "hqq", "eetq"], value="bnb")
         template = gr.Dropdown(choices=list(TEMPLATES.keys()), value="default")
         rope_scaling = gr.Dropdown(choices=["none", "linear", "dynamic", "yarn", "llama3"], value="none")
         booster = gr.Dropdown(choices=["auto", "flashattn2", "unsloth", "liger_kernel"], value="auto")

src/llamafactory/webui/runner.py

@@ -368,6 +368,7 @@ class Runner:
             if args.get("deepspeed", None) is not None:
                 env["FORCE_TORCHRUN"] = "1"
 
+            # NOTE: DO NOT USE shell=True to avoid security risk
             self.trainer = Popen(["llamafactory-cli", "train", save_cmd(args)], env=env)
             yield from self.monitor()
 

tests/data/processor/test_feedback.py

@@ -25,10 +25,10 @@ from llamafactory.train.test_utils import load_dataset_module
 
 DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 
 TRAIN_ARGS = {
-    "model_name_or_path": TINY_LLAMA,
+    "model_name_or_path": TINY_LLAMA3,
     "stage": "kto",
     "do_train": True,
     "finetuning_type": "full",
@@ -45,7 +45,7 @@ TRAIN_ARGS = {
 @pytest.mark.parametrize("num_samples", [16])
 def test_feedback_data(num_samples: int):
     train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
-    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
     original_data = load_dataset(DEMO_DATA, name="kto_en_demo", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)
     for index in indexes:

tests/data/processor/test_pairwise.py

@@ -25,10 +25,10 @@ from llamafactory.train.test_utils import load_dataset_module
 
 DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 
 TRAIN_ARGS = {
-    "model_name_or_path": TINY_LLAMA,
+    "model_name_or_path": TINY_LLAMA3,
     "stage": "rm",
     "do_train": True,
     "finetuning_type": "full",
@@ -54,7 +54,7 @@ def _convert_sharegpt_to_openai(messages: list[dict[str, str]]) -> list[dict[str
 @pytest.mark.parametrize("num_samples", [16])
 def test_pairwise_data(num_samples: int):
     train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
-    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
     original_data = load_dataset(DEMO_DATA, name="dpo_en_demo", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)
     for index in indexes:

tests/data/processor/test_supervised.py

@@ -25,12 +25,12 @@ from llamafactory.train.test_utils import load_dataset_module
 
 DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 
 TINY_DATA = os.getenv("TINY_DATA", "llamafactory/tiny-supervised-dataset")
 
 TRAIN_ARGS = {
-    "model_name_or_path": TINY_LLAMA,
+    "model_name_or_path": TINY_LLAMA3,
     "stage": "sft",
     "do_train": True,
     "finetuning_type": "full",
@@ -45,7 +45,7 @@ TRAIN_ARGS = {
 @pytest.mark.parametrize("num_samples", [16])
 def test_supervised_single_turn(num_samples: int):
     train_dataset = load_dataset_module(dataset_dir="ONLINE", dataset=TINY_DATA, **TRAIN_ARGS)["train_dataset"]
-    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
     original_data = load_dataset(TINY_DATA, split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)
     for index in indexes:
@@ -66,7 +66,7 @@ def test_supervised_multi_turn(num_samples: int):
     train_dataset = load_dataset_module(dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", **TRAIN_ARGS)[
         "train_dataset"
     ]
-    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
     original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)
     for index in indexes:
@@ -79,7 +79,7 @@ def test_supervised_train_on_prompt(num_samples: int):
     train_dataset = load_dataset_module(
         dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", train_on_prompt=True, **TRAIN_ARGS
     )["train_dataset"]
-    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
     original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)
     for index in indexes:
@@ -93,7 +93,7 @@ def test_supervised_mask_history(num_samples: int):
     train_dataset = load_dataset_module(
         dataset_dir="REMOTE:" + DEMO_DATA, dataset="system_chat", mask_history=True, **TRAIN_ARGS
     )["train_dataset"]
-    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
     original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)
     for index in indexes:

tests/data/processor/test_unsupervised.py

@@ -24,12 +24,12 @@ from llamafactory.train.test_utils import load_dataset_module
 
 DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 
 TINY_DATA = os.getenv("TINY_DATA", "llamafactory/tiny-supervised-dataset")
 
 TRAIN_ARGS = {
-    "model_name_or_path": TINY_LLAMA,
+    "model_name_or_path": TINY_LLAMA3,
     "stage": "ppo",
     "do_train": True,
     "finetuning_type": "full",
@@ -48,7 +48,7 @@ TRAIN_ARGS = {
 @pytest.mark.parametrize("num_samples", [16])
 def test_unsupervised_data(num_samples: int):
     train_dataset = load_dataset_module(**TRAIN_ARGS)["train_dataset"]
-    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
     original_data = load_dataset(DEMO_DATA, name="system_chat", split="train")
     indexes = random.choices(range(len(original_data)), k=num_samples)
     for index in indexes:

tests/data/test_collator.py

@@ -24,11 +24,11 @@ from llamafactory.hparams import get_infer_args
 from llamafactory.model import load_tokenizer
 
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 
 
 def test_base_collator():
-    model_args, data_args, *_ = get_infer_args({"model_name_or_path": TINY_LLAMA, "template": "default"})
+    model_args, data_args, *_ = get_infer_args({"model_name_or_path": TINY_LLAMA3, "template": "default"})
     tokenizer_module = load_tokenizer(model_args)
     template = get_template_and_fix_tokenizer(tokenizer_module["tokenizer"], data_args)
     data_collator = MultiModalDataCollatorForSeq2Seq(

tests/data/test_loader.py

@@ -19,12 +19,12 @@ from llamafactory.train.test_utils import load_dataset_module
 
 DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 
 TINY_DATA = os.getenv("TINY_DATA", "llamafactory/tiny-supervised-dataset")
 
 TRAIN_ARGS = {
-    "model_name_or_path": TINY_LLAMA,
+    "model_name_or_path": TINY_LLAMA3,
     "stage": "sft",
     "do_train": True,
     "finetuning_type": "full",

tests/data/test_mm_plugin.py

@@ -20,7 +20,6 @@ import torch
 from PIL import Image
 
 from llamafactory.data.mm_plugin import get_mm_plugin
-from llamafactory.extras.packages import is_transformers_version_greater_than
 from llamafactory.hparams import get_infer_args
 from llamafactory.model import load_tokenizer
 
@@ -35,7 +34,8 @@ if TYPE_CHECKING:
 
 HF_TOKEN = os.getenv("HF_TOKEN")
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA4 = os.getenv("TINY_LLAMA4", "llamafactory/tiny-random-Llama-4")
 
 MM_MESSAGES = [
     {"role": "user", "content": "<image>What is in this image?"},
@@ -130,13 +130,13 @@ def _check_plugin(
 
 
 def test_base_plugin():
-    tokenizer_module = _load_tokenizer_module(model_name_or_path=TINY_LLAMA)
+    tokenizer_module = _load_tokenizer_module(model_name_or_path=TINY_LLAMA3)
     base_plugin = get_mm_plugin(name="base")
     check_inputs = {"plugin": base_plugin, **tokenizer_module}
     _check_plugin(**check_inputs)
 
 
-@pytest.mark.skipif(not HF_TOKEN or not is_transformers_version_greater_than("4.50.0"), reason="Gated model.")
+@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
 def test_gemma3_plugin():
     image_seqlen = 256
     tokenizer_module = _load_tokenizer_module(model_name_or_path="google/gemma-3-4b-it")
@@ -157,6 +157,27 @@ def test_gemma3_plugin():
     _check_plugin(**check_inputs)
 
 
+@pytest.mark.xfail(reason="Unknown error.")
+def test_llama4_plugin():
+    tokenizer_module = _load_tokenizer_module(model_name_or_path=TINY_LLAMA4)
+    processor = tokenizer_module["processor"]
+    llama4_plugin = get_mm_plugin(name="llama4", image_token="<|image|>")
+    check_inputs = {"plugin": llama4_plugin, **tokenizer_module}
+    mm_inputs = _get_mm_inputs(tokenizer_module["processor"])
+    image_height, image_width = mm_inputs["pixel_values"][0].shape[-2:]
+    num_patches_per_chunk = int(
+        (image_height // processor.patch_size) * (image_width // processor.patch_size) // processor.downsample_ratio
+    )
+    aspect_ratios = mm_inputs.pop("aspect_ratios")
+    tokens_for_this_image = processor._prompt_split_image(aspect_ratios[0], num_patches_per_chunk)
+    check_inputs["expected_mm_messages"] = [
+        {key: value.replace("<image>", tokens_for_this_image) for key, value in message.items()}
+        for message in MM_MESSAGES
+    ]
+    check_inputs["expected_mm_inputs"] = mm_inputs
+    _check_plugin(**check_inputs)
+
+
 def test_llava_plugin():
     image_seqlen = 576
     tokenizer_module = _load_tokenizer_module(model_name_or_path="llava-hf/llava-1.5-7b-hf")

tests/data/test_template.py

@@ -29,7 +29,8 @@ if TYPE_CHECKING:
 
 HF_TOKEN = os.getenv("HF_TOKEN")
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA4 = os.getenv("TINY_LLAMA4", "llamafactory/tiny-random-Llama-4")
 
 MESSAGES = [
     {"role": "user", "content": "How are you"},
@@ -75,7 +76,7 @@ def _check_template(model_id: str, template_name: str, prompt_str: str, answer_s
 
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_encode_oneturn(use_fast: bool):
-    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA, use_fast=use_fast)
+    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, use_fast=use_fast)
     template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="llama3"))
     prompt_ids, answer_ids = template.encode_oneturn(tokenizer, MESSAGES)
     prompt_str = (
@@ -90,7 +91,7 @@ def test_encode_oneturn(use_fast: bool):
 
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_encode_multiturn(use_fast: bool):
-    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA, use_fast=use_fast)
+    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, use_fast=use_fast)
     template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="llama3"))
     encoded_pairs = template.encode_multiturn(tokenizer, MESSAGES)
     prompt_str_1 = (
@@ -111,8 +112,8 @@ def test_encode_multiturn(use_fast: bool):
 
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_jinja_template(use_fast: bool):
-    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA, use_fast=use_fast)
-    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA, use_fast=use_fast)
+    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, use_fast=use_fast)
+    ref_tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, use_fast=use_fast)
     template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="llama3"))
     tokenizer.chat_template = template._get_jinja_template(tokenizer)  # llama3 template no replace
     assert tokenizer.chat_template != ref_tokenizer.chat_template
@@ -120,7 +121,7 @@ def test_jinja_template(use_fast: bool):
 
 
 def test_ollama_modelfile():
-    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
     template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="llama3"))
     assert template.get_ollama_modelfile(tokenizer) == (
         "# ollama modelfile auto-generated by llamafactory\n\n"
@@ -137,7 +138,7 @@ def test_ollama_modelfile():
 
 
 def test_get_stop_token_ids():
-    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA)
+    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3)
     template = get_template_and_fix_tokenizer(tokenizer, DataArguments(template="llama3"))
     assert set(template.get_stop_token_ids(tokenizer)) == {128008, 128009}
 
@@ -152,7 +153,7 @@ def test_gemma_template(use_fast: bool):
         "<start_of_turn>model\n"
     )
     answer_str = "很高兴认识你！<end_of_turn>\n"
-    _check_template("google/gemma-2-9b-it", "gemma", prompt_str, answer_str, use_fast)
+    _check_template("google/gemma-3-4b-it", "gemma", prompt_str, answer_str, use_fast)
 
 
 @pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
@@ -168,7 +169,20 @@ def test_llama3_template(use_fast: bool):
     _check_template("meta-llama/Meta-Llama-3-8B-Instruct", "llama3", prompt_str, answer_str, use_fast)
 
 
-@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
+@pytest.mark.parametrize(
+    "use_fast", [True, pytest.param(False, marks=pytest.mark.xfail(reason="Llama 4 has no slow tokenizer."))]
+)
+def test_llama4_template(use_fast: bool):
+    prompt_str = (
+        "<|begin_of_text|><|header_start|>user<|header_end|>\n\nHow are you<|eot|>"
+        "<|header_start|>assistant<|header_end|>\n\nI am fine!<|eot|>"
+        "<|header_start|>user<|header_end|>\n\n你好<|eot|>"
+        "<|header_start|>assistant<|header_end|>\n\n"
+    )
+    answer_str = "很高兴认识你！<|eot|>"
+    _check_template(TINY_LLAMA4, "llama4", prompt_str, answer_str, use_fast)
+
+
 @pytest.mark.parametrize(
     "use_fast", [True, pytest.param(False, marks=pytest.mark.xfail(reason="Phi-4 slow tokenizer is broken."))]
 )
@@ -183,35 +197,21 @@ def test_phi4_template(use_fast: bool):
     _check_template("microsoft/phi-4", "phi4", prompt_str, answer_str, use_fast)
 
 
-@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")  # TODO: why it is gated?
 @pytest.mark.parametrize("use_fast", [True, False])
 def test_qwen_template(use_fast: bool):
     prompt_str = (
-        "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
-        "<|im_start|>user\nHow are you<|im_end|>\n"
-        "<|im_start|>assistant\nI am fine!<|im_end|>\n"
-        "<|im_start|>user\n你好<|im_end|>\n"
-        "<|im_start|>assistant\n"
-    )
-    answer_str = "很高兴认识你！<|im_end|>\n"
-    _check_template("Qwen/Qwen2-7B-Instruct", "qwen", prompt_str, answer_str, use_fast)
-
-
-@pytest.mark.parametrize("use_fast", [True, False])
-@pytest.mark.xfail(reason="Yi tokenizer is broken.")
-def test_yi_template(use_fast: bool):
-    prompt_str = (
+        "<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n"
         "<|im_start|>user\nHow are you<|im_end|>\n"
         "<|im_start|>assistant\nI am fine!<|im_end|>\n"
         "<|im_start|>user\n你好<|im_end|>\n"
         "<|im_start|>assistant\n"
     )
     answer_str = "很高兴认识你！<|im_end|>\n"
-    _check_template("01-ai/Yi-1.5-6B-Chat", "yi", prompt_str, answer_str, use_fast)
+    _check_template("Qwen/Qwen2.5-7B-Instruct", "qwen", prompt_str, answer_str, use_fast)
 
 
-def test_parse_template():
-    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA, token=HF_TOKEN)
+def test_parse_llama3_template():
+    tokenizer = AutoTokenizer.from_pretrained(TINY_LLAMA3, token=HF_TOKEN)
     template = parse_template(tokenizer)
     assert template.format_user.slots == [
         "<|start_header_id|>user<|end_header_id|>\n\n{{content}}<|eot_id|>"
@@ -223,12 +223,11 @@ def test_parse_template():
     assert template.default_system == ""
 
 
-@pytest.mark.skipif(not HF_TOKEN, reason="Gated model.")
 def test_parse_qwen_template():
-    tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-7B-Instruct", token=HF_TOKEN)
+    tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct", token=HF_TOKEN)
     template = parse_template(tokenizer)
     assert template.format_user.slots == ["<|im_start|>user\n{{content}}<|im_end|>\n<|im_start|>assistant\n"]
     assert template.format_assistant.slots == ["{{content}}<|im_end|>\n"]
     assert template.format_system.slots == ["<|im_start|>system\n{{content}}<|im_end|>\n"]
     assert template.format_prefix.slots == []
-    assert template.default_system == "You are a helpful assistant."
+    assert template.default_system == "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."

tests/e2e/test_chat.py

@@ -17,10 +17,10 @@ import os
 from llamafactory.chat import ChatModel
 
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 
 INFER_ARGS = {
-    "model_name_or_path": TINY_LLAMA,
+    "model_name_or_path": TINY_LLAMA3,
     "finetuning_type": "lora",
     "template": "llama3",
     "infer_dtype": "float16",

tests/e2e/test_train.py

@@ -21,12 +21,12 @@ from llamafactory.train.tuner import export_model, run_exp
 
 DEMO_DATA = os.getenv("DEMO_DATA", "llamafactory/demo_data")
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 
 TINY_LLAMA_ADAPTER = os.getenv("TINY_LLAMA_ADAPTER", "llamafactory/tiny-random-Llama-3-lora")
 
 TRAIN_ARGS = {
-    "model_name_or_path": TINY_LLAMA,
+    "model_name_or_path": TINY_LLAMA3,
     "do_train": True,
     "finetuning_type": "lora",
     "dataset_dir": "REMOTE:" + DEMO_DATA,
@@ -35,10 +35,11 @@ TRAIN_ARGS = {
     "overwrite_output_dir": True,
     "per_device_train_batch_size": 1,
     "max_steps": 1,
+    "report_to": "none",
 }
 
 INFER_ARGS = {
-    "model_name_or_path": TINY_LLAMA,
+    "model_name_or_path": TINY_LLAMA3,
     "adapter_name_or_path": TINY_LLAMA_ADAPTER,
     "finetuning_type": "lora",
     "template": "llama3",

tests/model/model_utils/test_attention.py

@@ -21,10 +21,10 @@ from llamafactory.extras.packages import is_transformers_version_greater_than
 from llamafactory.train.test_utils import load_infer_model
 
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 
 INFER_ARGS = {
-    "model_name_or_path": TINY_LLAMA,
+    "model_name_or_path": TINY_LLAMA3,
     "template": "llama3",
 }
 

tests/model/model_utils/test_checkpointing.py

@@ -21,10 +21,10 @@ from llamafactory.extras.misc import get_current_device
 from llamafactory.train.test_utils import load_train_model
 
 
-TINY_LLAMA = os.getenv("TINY_LLAMA", "llamafactory/tiny-random-Llama-3")
+TINY_LLAMA3 = os.getenv("TINY_LLAMA3", "llamafactory/tiny-random-Llama-3")
 
 TRAIN_ARGS = {
-    "model_name_or_path": TINY_LLAMA,
+    "model_name_or_path": TINY_LLAMA3,
     "stage": "sft",
     "do_train": True,
     "finetuning_type": "lora",