【CI】add qwen3.5 mtp and vl cases

.github/workflows/e2e_test.yaml

@@ -1,5 +1,5 @@
 name: ete_test
-
+run-name: 'e2e_test-${{ github.event.inputs.run_case }}'
 permissions:
   contents: write
   pages: write

autotest/config.yaml

@@ -546,6 +546,165 @@ case:
                     lr: 0
             timeout: 10800
 
+    qwen3-5-sft-fp8-vl:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/qwen3_5_fp8_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
+                    - XTUNER_ACTIVATION_OFFLOAD=1
+                    - XTUNER_USE_FA3=1
+            assert_info:
+                base_metric: qwen3-5-sft-fp8-vl/00f7e16/tracker.jsonl
+                check_metrics:
+                    grad_norm: 10
+                    loss/local_loss: 5
+                    loss/reduced_balancing_loss: 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
+
+    qwen3-5-sft-tp2-vl:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/qwen3_5_tp2_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-tp2-vl/00f7e16/tracker.jsonl
+                check_metrics:
+                    grad_norm: 10
+                    loss/local_loss: 5
+                    loss/reduced_balancing_loss: 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
+
+    qwen3-5-sft-sp4-resume-vl:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/qwen3_5_moe_30BA3_sp4_vl.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                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
+                    - XTUNER_GC_ENABLE=1
+            assert_info:
+                base_metric: qwen3-5-sft-sp4-resume-vl/00f7e16/tracker.jsonl
+                check_metrics:
+                    grad_norm: 10
+                    loss/local_loss: 5
+                    loss/reduced_balancing_loss: 5
+                    loss/reduced_llm_loss: 10
+                    lr: 0
+                    memory/max_memory_GB: 0.2
+                    runtime_info/tgs: 0.05
+                    runtime_info/text_tokens: 0
+            timeout: 10800
+
+        -
+            type: sft
+            pre_action:
+                command: 'python ./autotest/utils/update_meta.py /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output qwen3-5-sft-sp4-resume-vl sft'
+            parameters:
+                config: autotest/config/qwen3_5_moe_30BA3_sp4_vl.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                memory_per_task: 1200
+                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
+                    - XTUNER_GC_ENABLE=1
+            assert_info:
+                base_metric: qwen3-5-sft-sp4-resume-vl/00f7e16_resume/tracker.jsonl
+                check_metrics:
+                    grad_norm: 10
+                    loss/local_loss: 5
+                    loss/reduced_balancing_loss: 5
+                    loss/reduced_llm_loss: 5
+                    lr: 0
+                    memory/max_memory_GB: 0.2
+                    runtime_info/tgs: 0.05
+                    runtime_info/text_tokens: 0
+            post_action:
+                command: 'python ./autotest/utils/resume_validation.py /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output qwen3-5-sft-sp4-resume-vl sft 10 "runtime_info/text_tokens,runtime_info/efficient_attn_ratio,loss/reduced_balancing_loss,loss/reduced_llm_loss,loss/local_loss"'
+            timeout: 10800
+
+    qwen3-5-sft-mtp:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/qwen3_5_mtp.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                envs:
+                    - MODEL_PATH=/mnt/shared-storage-user/llmrazor-share/model/Qwen3.5-35B-A3B
+                    - DATA_PATH=/mnt/shared-storage-user/llmrazor-share/data/alpaca
+                    - XTUNER_DETERMINISTIC=true
+            assert_info:
+                base_metric: qwen3-5-sft-mtp/00f7e16/tracker.jsonl
+                check_metrics:
+                    grad_norm: 0.05
+                    loss/local_loss: 0.000001
+                    loss/reduced_balancing_loss: 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-mtp-vl:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/qwen3_5_mtp_vl.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                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-mtp-vl/00f7e16/tracker.jsonl
+                check_metrics:
+                    grad_norm: 0.000001
+                    loss/local_loss: 0.000001
+                    loss/reduced_balancing_loss: 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-rl-lmdeploy:
         -
             type: rl

autotest/config/qwen3_5_fp8_vl.py

@@ -0,0 +1,67 @@
+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.float8.config import Float8Config, ScalingGranularity
+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"]
+
+float8_cfg = Float8Config(
+    scaling_granularity_gemm=ScalingGranularity.TILEWISE,
+    scaling_granularity_grouped_gemm=ScalingGranularity.TILEWISE,
+)
+
+moe_cfg = Qwen3_5_VLMoE35BA3Config(compile_cfg=False, float8_cfg=float8_cfg)
+
+optim_cfg = AdamWConfig(lr=6e-05)
+lr_cfg = LRConfig(lr_type="cosine", lr_min=1e-6)
+fsdp_cfg = FSDPConfig(
+    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_5_moe_30BA3_sp4_vl.py

@@ -0,0 +1,66 @@
+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 ResumeConfig, TrainerConfig
+
+
+MEDIA_ROOT = os.environ["MEDIA_ROOT"]
+MODEL_PATH = os.environ["MODEL_PATH"]
+DATA_PATH = os.environ["DATA_PATH"]
+
+
+moe_cfg = Qwen3_5_VLMoE35BA3Config(compile_cfg=False)
+
+optim_cfg = AdamWConfig(lr=6e-05)
+lr_cfg = LRConfig(lr_type="cosine", lr_min=1e-6)
+fsdp_cfg = FSDPConfig(
+    cpu_offload=False,
+    tp_size=2,
+)
+
+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,
+    resume_cfg=ResumeConfig(auto_resume=True),
+    checkpoint_interval=10,
+    checkpoint_maxkeep=2,
+)

autotest/config/qwen3_5_mtp.py

@@ -0,0 +1,50 @@
+import os
+
+from xtuner.v1.config import AdamWConfig, FSDPConfig, LRConfig
+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 import Qwen3_5_VLMoE35BA3Config
+from xtuner.v1.module.mtp import MTPConfig
+from xtuner.v1.train import TrainerConfig
+
+
+MODEL_PATH = os.environ["MODEL_PATH"]
+DATA_PATH = os.environ["DATA_PATH"]
+
+
+moe_cfg = Qwen3_5_VLMoE35BA3Config(compile_cfg=False)
+moe_cfg.text_config.mtp_config = MTPConfig(num_layers=1, loss_scaling_factor=0.1)
+
+optim_cfg = AdamWConfig(lr=6e-05)
+lr_cfg = LRConfig(lr_type="cosine", lr_min=1e-6)
+fsdp_cfg = FSDPConfig(
+    cpu_offload=False,
+)
+
+dataset_config = [
+    {
+        "dataset": DatasetConfig(name="alpaca", anno_path=DATA_PATH, sample_ratio=1.0),
+        "tokenize_fn": OpenaiTokenizeFunctionConfig(chat_template="qwen3", max_length=16384),
+    },
+]
+
+dataloader_config = DataloaderConfig(pack_max_length=16384)
+
+loss_cfg = CELossConfig(mode="chunk", chunk_size=1024, loss_reduction="square")
+
+trainer = TrainerConfig(
+    load_from=MODEL_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=MODEL_PATH,
+    global_batch_size=16,
+    total_epoch=1,
+    work_dir=f"{os.environ['WORK_DIR']}",
+    seed=0,
+)