【CI】add cases about ep_size

.github/workflows/ete_test_gpu.yaml

@@ -32,6 +32,14 @@ on:
           Resume phases within one case stay sequential.
         type: string
         default: '4'
+      CI_E2E_DEBUG:
+        required: false
+        description: >-
+          CI_E2E_DEBUG:if true, start "sleep inf &" after install and before training so the
+          cluster job stays alive for interactive debug (including after OOM).
+          Default false.
+        type: boolean
+        default: false
   schedule:
     - cron:  '00 16 * * 0-5'
 env:
@@ -69,6 +77,7 @@ jobs:
         env:
           GITHUB_RUN_ID: ${{ github.run_id }}
           CI_REPORTS_RAW_URL_BASE: ${{ env.CI_REPORTS_RAW_URL_BASE }}
+          CI_E2E_DEBUG: ${{ github.event.inputs.CI_E2E_DEBUG || 'false' }}
         run: |
           source /mnt/shared-storage-user/opencompass-shared/qa-llm-cicd/miniconda3/bin/activate
           conda activate clusterx 

.github/workflows/ete_test_npu.yaml

@@ -24,6 +24,14 @@ on:
           (e.g. "npu-qwen3-sft,npu-qwen3-rl-vllm").
         type: string
         default: 'all'
+      CI_E2E_DEBUG:
+        required: false
+        description: >-
+          CI_E2E_DEBUG:if true, start "sleep inf &" after install and before training so the
+          cluster job stays alive for interactive debug (including after OOM).
+          Default false.
+        type: boolean
+        default: false
   schedule:
     - cron:  '00 16 * * 0-5'
 env:
@@ -59,6 +67,7 @@ jobs:
         env:
           GITHUB_RUN_ID: ${{ github.run_id }}
           CI_REPORTS_RAW_URL_BASE: ${{ env.CI_REPORTS_RAW_URL_BASE }}
+          CI_E2E_DEBUG: ${{ github.event.inputs.CI_E2E_DEBUG || 'false' }}
         run: |
           source /mnt/hwfile/vc-intern-delivery/qa-llm-cicd/miniconda3/bin/activate
           source activate npuci

autotest/config.yaml

@@ -824,6 +824,75 @@ case:
                     runtime_info/text_tokens: 0
             timeout: 1500
 
+    qwen3-5-sft-ep2-mtp4-vl:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/qwen3_5_ep2_mtp4_vl.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                num_nodes: 2
+                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
+                    - TORCH_ALLOW_TF32_CUBLAS_OVERRIDE=0
+            assert_info:
+                base_metric: qwen3-5-sft-ep2-mtp4-vl/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: 1500
+
+    qwen3-5-rl-vl-lmdeploy-mtp-ep:
+        -
+            type: rl
+            parameters:
+                config: autotest/config/rl_qwen3p5_vl_35B_mtp_ep.py
+                infer_backend: lmdeploy
+                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/rl_vl_dapomath_geo3k_train.jsonl
+                    - EVAL_DATA_PATH=/mnt/shared-storage-user/llmrazor-share/data/rl_vl_dapomath_geo3k_test.jsonl
+                    - XTUNER_USE_LMDEPLOY=1
+                    - XTUNER_DETERMINISTIC=true
+                    - LMDEPLOY_FP32_MAMBA_SSM_DTYPE=1
+            assert_info:
+                base_metric: qwen3-5-rl-vl-lmdeploy-mtp-ep/tracker.jsonl
+                check_metrics:
+                    -
+                       metric: response/rewards/mean
+                       threshold: 0.3
+                       method: absolute
+                       operator: <
+                    -
+                       metric:  mismatch/mismatch_k3_kl
+                       threshold: 0.0005
+                       method: absolute
+                       operator: <=
+                    -
+                       metric: response/response_len/mean
+                       threshold: 0.25
+                       method: relative
+                       operator: <
+                    -
+                       metric: time/step
+                       threshold: 0.15
+                       method: relative
+                       operator: <
+                       aggregate: 80
+            timeout: 10800
+
     qwen3-rl-lmdeploy:
         -
             type: rl

autotest/config/qwen3_5_ep2_mtp4_vl.py

@@ -0,0 +1,95 @@
+import os
+
+from xtuner.v1.config import FSDPConfig, LRConfig, MuonConfig
+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.module.mtp import MTPConfig
+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"]
+WORK_DIR = os.environ["WORK_DIR"]
+
+sample_max_length = 16 * 1024
+pack_max_length = 64 * 1024
+ep_size = 2
+
+moe_cfg = Qwen3_5_VLMoE35BA3Config(only_llm_forward=False, compile_cfg=True)
+moe_cfg.text_config.ep_size = ep_size
+moe_cfg.text_config.dispatcher = "deepep"
+moe_cfg.text_config.mtp_config = MTPConfig(
+    num_layers=4,
+    share_weights=True,
+    loss_scaling_factor=1.0,
+)
+
+optim_cfg = MuonConfig(lr=2e-5, weight_decay=0.05)
+lr_cfg = LRConfig(lr_type="cosine", warmup_ratio=0.1, lr_min=1e-6)
+fsdp_cfg = FSDPConfig(
+    recompute_ratio=1.0,
+    torch_compile=True,
+    cpu_offload=False,
+    ep_size=ep_size,
+    checkpoint_preserve_rng_state=False,
+)
+
+dataset_config = [
+    {
+        "dataset": DatasetConfig(
+            name="multimodal",
+            anno_path=DATA_PATH,
+            class_name="VLMJsonlDataset",
+            media_root=MEDIA_ROOT,
+            sample_ratio=1.0,
+            cache_dir=os.path.join(WORK_DIR, "jsonl_cache"),
+            cache_tag=f"qwen3p5_vl_{sample_max_length}",
+        ),
+        "tokenize_fn": Qwen3VLTokenizeFnConfig(
+            processor_path=MODEL_PATH,
+            chat_template="qwen3.5-vl",
+            llm_pack_weight=-3.2,
+            visual_pack_weight=5.0,
+            max_length=sample_max_length,
+            rand_video_max_frames=24,
+            max_pixels=16384 * 32 * 32,
+        ),
+    },
+]
+
+dataloader_config = DataloaderConfig(
+    dataset_config_list=dataset_config,
+    pack_level="soft",
+    pack_max_length=pack_max_length,
+    pack_to_max_length=True,
+    pack_chunk_size=10000,
+    pack_workers=4,
+    global_pack=True,
+    group_by_length=True,
+    collator="qwen3_vl_sft_collator",
+    pack_extra_buffer_size=20,
+    num_workers=1,
+)
+
+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,
+    dataloader_cfg=dataloader_config,
+    lr_cfg=lr_cfg,
+    loss_cfg=loss_cfg,
+    tokenizer_path=MODEL_PATH,
+    resume_cfg=ResumeConfig(auto_resume=True),
+    global_batch_size=16,
+    total_step=20,
+    sp_size=4,
+    work_dir=WORK_DIR,
+    seed=0,
+)
+