【CI】add qwen3.5 cases

autotest/cluster/clusterx.py

@@ -50,6 +50,7 @@ def execute_task(self, task_config: Dict[str, Any]):
                 num_nodes=resource.get("num_nodes", 1),
                 image=resource.get("image", None),
                 no_env=resource.get("no_env", True),
+                image_pull_policy=resource.get("image_pull_policy","Always"),
             )
 
             job_schema = self.cluster.run(params)

autotest/config.yaml

@@ -277,6 +277,54 @@ case:
                     runtime_info/text_tokens: 0
             timeout: 1080
 
+    qwen3-sft-cache:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/qwen3_sft_cache.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                cpus_per_task: 80
+                envs:
+                    - QWEN3_MOE_PATH=/mnt/shared-storage-user/llmrazor-share/model/Qwen3-30B-A3B
+                    - ALPACA_PATH=/mnt/shared-storage-user/llmrazor-share/data/alpaca
+                    - CACHE_DIR=/mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/.cache
+                    - XTUNER_DETERMINISTIC=true
+            assert_info:
+                base_metric: qwen3-sft-cache/e968368a/tracker.jsonl
+                check_metrics:
+                    grad_norm: 0.000001
+                    loss/reduced_llm_loss: 0.000001
+                    lr: 0
+                    memory/max_memory_GB: 0.2
+                    runtime_info/tgs: 0.05
+                    runtime_info/text_tokens: 0
+            timeout: 10800
+
+    qwen3-5-sft-vl-moe:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/qwen3_5_30B_sft_vl.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                cpus_per_task: 80
+                envs:
+                    - MODEL_PATH=/mnt/shared-storage-user/llmrazor-share/model/Qwen3.5-35B-A3B
+                    - DATA_PATH=/mnt/shared-storage-user/llmrazor-share/data/ci_vl
+                    - MEDIA_ROOT=/mnt/shared-storage-user/llmrazor-share/data/ci_vl
+                    - XTUNER_DETERMINISTIC=true
+            assert_info:
+                base_metric: qwen3-5-sft-vl-moe/e968368a/tracker.jsonl
+                check_metrics:
+                    grad_norm: 5
+                    loss/reduced_llm_loss: 5
+                    lr: 0
+                    memory/max_memory_GB: 0.2
+                    runtime_info/tgs: 0.05
+                    runtime_info/text_tokens: 0
+            timeout: 10800
+
     qwen3-rl-lmdeploy:
         -
             type: rl

autotest/config/gptoss.py

@@ -5,8 +5,8 @@
     FSDPConfig,
     LRConfig,
 )
-from xtuner.v1.datasets import FTDPTokenizeFnConfig
 from xtuner.v1.datasets.config import DataloaderConfig, DatasetConfig
+from xtuner.v1.datasets.sft_tokenize_fn import OpenaiTokenizeFunctionConfig
 from xtuner.v1.loss.ce_loss import CELossConfig
 from xtuner.v1.model.moe.gpt_oss import GptOss21BA3P6Config
 from xtuner.v1.module.rope import RopeScalingConfig
@@ -38,7 +38,7 @@
 dataset_config = [
     {
         "dataset": DatasetConfig(name="alpaca", anno_path=ALPACA_PATH, sample_ratio=1.0),
-        "tokenize_fn": FTDPTokenizeFnConfig(max_length=16384),
+        "tokenize_fn": OpenaiTokenizeFunctionConfig(chat_template='gpt-oss', max_length=16384),
     },
 ]
 

autotest/config/qwen3_5_30B_sft_vl.py

@@ -0,0 +1,63 @@
+import os
+
+from xtuner.v1.config import AdamWConfig, FSDPConfig, LRConfig
+from xtuner.v1.datasets import Qwen3VLTokenizeFnConfig
+from xtuner.v1.datasets.config import DataloaderConfig, DatasetConfig
+from xtuner.v1.loss.ce_loss import CELossConfig
+from xtuner.v1.model import Qwen3_5_VLMoE35BA3Config
+from xtuner.v1.train import TrainerConfig
+
+
+MEDIA_ROOT = os.environ["MEDIA_ROOT"]
+MODEL_PATH = os.environ["MODEL_PATH"]
+DATA_PATH = os.environ["DATA_PATH"]
+
+
+moe_cfg = Qwen3_5_VLMoE35BA3Config()
+
+optim_cfg = AdamWConfig(lr=6e-05)
+lr_cfg = LRConfig(lr_type="cosine", lr_min=1e-6)
+fsdp_cfg = FSDPConfig(
+    torch_compile=True,
+    cpu_offload=False,
+)
+
+dataset_config = [
+    {
+        "dataset": DatasetConfig(
+            name="sft",
+            anno_path=DATA_PATH,
+            class_name="VLMJsonlDataset",
+            media_root=MEDIA_ROOT,
+            sample_ratio=1.0,
+        ),
+        "tokenize_fn": Qwen3VLTokenizeFnConfig(
+            processor_path=MODEL_PATH,
+            max_length=16384,
+            add_vision_id=True,
+        ),
+    },
+]
+
+dataloader_config = DataloaderConfig(
+    dataset_config_list=dataset_config,
+    pack_max_length=16384,
+    collator="qwen3_vl_sft_collator",
+)
+
+loss_cfg = CELossConfig(mode="chunk", chunk_size=1024)
+
+trainer = TrainerConfig(
+    load_from=MODEL_PATH,
+    model_cfg=moe_cfg,
+    optim_cfg=optim_cfg,
+    fsdp_cfg=fsdp_cfg,
+    dataloader_cfg=dataloader_config,
+    lr_cfg=lr_cfg,
+    loss_cfg=loss_cfg,
+    tokenizer_path=MODEL_PATH,
+    global_batch_size=16,
+    total_epoch=1,
+    work_dir=f"{os.environ['WORK_DIR']}",
+    seed=0,
+)

autotest/config/qwen3_sft_cache.py

@@ -0,0 +1,55 @@
+import os
+
+from xtuner.v1.config import AdamWConfig, FSDPConfig, LRConfig
+from xtuner.v1.datasets import FTDPTokenizeFnConfig
+from xtuner.v1.datasets.config import DataloaderConfig, DatasetConfig
+from xtuner.v1.loss.ce_loss import CELossConfig
+from xtuner.v1.model.moe.qwen3 import Qwen3MoE30BA3Config
+from xtuner.v1.train import TrainerConfig
+
+
+QWEN3_MOE_PATH = os.environ["QWEN3_MOE_PATH"]
+ALPACA_PATH = os.environ["ALPACA_PATH"]
+CACHE_DIR = os.environ["CACHE_DIR"]
+
+
+moe_cfg = Qwen3MoE30BA3Config()
+optim_cfg = AdamWConfig(lr=6e-05)
+lr_cfg = LRConfig(lr_type="cosine", lr_min=1e-6)
+fsdp_cfg = FSDPConfig(
+    torch_compile=False,
+    cpu_offload=False,
+    ep_size=moe_cfg.ep_size,
+)
+
+dataset_config = [
+    {
+        "dataset": DatasetConfig(
+            name="alpaca",
+            anno_path=ALPACA_PATH,
+            sample_ratio=1.0,
+            cache_dir=CACHE_DIR,
+        ),
+        "tokenize_fn": FTDPTokenizeFnConfig(max_length=16384),
+    },
+]
+
+dataloader_config = DataloaderConfig(pack_max_length=16384)
+
+loss_cfg = CELossConfig(mode="chunk", chunk_size=1024)
+
+trainer = TrainerConfig(
+    load_from=QWEN3_MOE_PATH,
+    model_cfg=moe_cfg,
+    optim_cfg=optim_cfg,
+    fsdp_cfg=fsdp_cfg,
+    dataset_cfg=dataset_config,
+    dataloader_cfg=dataloader_config,
+    lr_cfg=lr_cfg,
+    loss_cfg=loss_cfg,
+    tokenizer_path=QWEN3_MOE_PATH,
+    global_batch_size=16,
+    total_epoch=1,
+    work_dir=f"{os.environ['WORK_DIR']}",
+    seed=0,
+)