[CI]add cases about glm5.2

autotest/config-npu.yaml

@@ -26,7 +26,7 @@ default_config:
                 - HF_HUB_OFFLINE=1
 
 case: 
-    npu-qwen3-sft:
+    qwen3:
         -
             type: sft
             parameters:
@@ -53,7 +53,7 @@ case:
                     runtime_info/text_tokens: 0
             timeout: 10800
 
-    npu-qwen3-sft-ep8:
+    qwen3-ep8:
         -
             type: sft
             phase: first
@@ -107,7 +107,7 @@ case:
                     runtime_info/text_tokens: 0
             timeout: 10800
 
-    npu-qwen3-sft-tp2:
+    qwen3-tp2:
         -
             type: sft
             parameters:
@@ -133,7 +133,7 @@ case:
                     runtime_info/text_tokens: 0
             timeout: 10800
 
-    npu-qwen3-sft-recompute:
+    qwen3-recompute:
         -
             type: sft
             parameters:
@@ -161,7 +161,7 @@ case:
                     runtime_info/text_tokens: 0
             timeout: 10800
 
-    npu-qwen3-sft-16nums:
+    qwen3-16nums:
         -
             type: sft
             parameters:
@@ -184,7 +184,7 @@ case:
                     lr: 0
             timeout: 10800
 
-    npu-qwen3-sft-celoss-vs-gpu:
+    qwen3-celoss:
         -
             type: sft
             parameters:
@@ -207,7 +207,7 @@ case:
                     lr: 0.01
             timeout: 10800
 
-    npu-qwen3-rl-vllm:
+    qwen3-vllm:
         -
             type: rl
             parameters:

autotest/config.yaml

@@ -26,6 +26,268 @@ default_config:
                 - HF_HUB_OFFLINE=1
 
 case: 
+
+    glm5-2-rl-lmdeploy-grpo:
+        -
+            type: rl
+            parameters:
+                config: autotest/config/rl_glm5p2_30B_gsm8k_grpo.py
+                infer_backend: lmdeploy
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                pip_package: FLASH_MLA_DISABLE_SM100=1 pip install -v /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/flash-mla;pip install -e .[all]; pip install more-itertools pytest-xdist
+                memory_per_task: 1200
+                envs:
+                    - MODEL_PATH=/mnt/shared-storage-user/llmrazor-share/model/GLM-5.2-30B
+                    - DATA_PATH=/mnt/shared-storage-user/llmrazor-share/data/gsm8k/train-mini.jsonl
+                    - EVAL_DATA_PATH=/mnt/shared-storage-user/llmrazor-share/data/gsm8k/test.jsonl
+                    - XTUNER_USE_LMDEPLOY=1
+                    - XTUNER_DETERMINISTIC=true
+            assert_info:
+                base_metric: glm5-2-rl-lmdeploy-grpo/tracker.jsonl
+                check_metrics:
+                    -
+                       metric:  eval/accuracy
+                       threshold: 0.1
+                       method: absolute
+                       operator: <
+                    -
+                       metric: response/rewards/mean
+                       threshold: 0.3
+                       method: absolute
+                       operator: <
+                    -
+                       metric:  mismatch/mismatch_k3_kl
+                       threshold: 0.001
+                       method: absolute
+                       operator: <
+                    -
+                       metric: response/response_len/mean
+                       threshold: 0.25
+                       method: relative
+                       operator: <
+                    -
+                       metric: time/step
+                       threshold: 20
+                       method: absolute
+                       operator: <
+            timeout: 7200
+
+    glm5-2-rl-lmdeploy-dapo:
+        -
+            type: rl
+            parameters:
+                config: autotest/config/rl_glm5p2_30B_gsm8k_dapo.py
+                infer_backend: lmdeploy
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                memory_per_task: 1200
+                pip_package: FLASH_MLA_DISABLE_SM100=1 pip install -v /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/flash-mla;pip install -e .[all]; pip install more-itertools pytest-xdist
+                envs:
+                    - MODEL_PATH=/mnt/shared-storage-user/llmrazor-share/model/GLM-5.2-30B
+                    - DATA_PATH=/mnt/shared-storage-user/llmrazor-share/data/gsm8k/train-mini.jsonl
+                    - EVAL_DATA_PATH=/mnt/shared-storage-user/llmrazor-share/data/gsm8k/test.jsonl
+                    - XTUNER_USE_LMDEPLOY=1
+                    - XTUNER_DETERMINISTIC=true
+            assert_info:
+                base_metric: glm5-2-rl-lmdeploy-dapo/tracker.jsonl
+                check_metrics:
+                    -
+                       metric:  eval/accuracy
+                       threshold: 0.1
+                       method: absolute
+                       operator: <
+                    -
+                       metric: response/rewards/mean
+                       threshold: 0.3
+                       method: absolute
+                       operator: <
+                    -
+                       metric:  mismatch/mismatch_k3_kl
+                       threshold: 0.001
+                       method: absolute
+                       operator: <
+                    -
+                       metric: response/response_len/mean
+                       threshold: 0.25
+                       method: relative
+                       operator: <
+                    -
+                       metric: time/step
+                       threshold: 20
+                       method: absolute
+                       operator: <
+            timeout: 7200
+    glm5-2-sft-30B:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/glm5p2_30B.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                cpus_per_task: 120
+                memory_per_task: 1200
+                envs:
+                    - MODEL_PATH=/mnt/shared-storage-user/llmrazor-share/model/GLM-5.2-30B
+                    - ALPACA_PATH=/mnt/shared-storage-user/llmrazor-share/data/alpaca
+                    - XTUNER_GC_ENABLE=1
+                    - SWAP_OPTIMIZER=0
+                    - XTUNER_ACTIVATION_OFFLOAD=0
+                    - XTUNER_USE_CUTLASS_GROUP_GEMM=1
+                    - XTUNER_DETERMINISTIC=true
+            assert_info:
+                base_metric: glm5-2-sft-30B/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
+
+    glm5-2-sft-30B-mtp-fp8:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/glm5p2_30B_mtp_fp8.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                cpus_per_task: 120
+                memory_per_task: 1200
+                envs:
+                    - MODEL_PATH=/mnt/shared-storage-user/llmrazor-share/model/GLM-5.2-30B
+                    - ALPACA_PATH=/mnt/shared-storage-user/llmrazor-share/data/alpaca
+                    - XTUNER_DETERMINISTIC=true
+                    - XTUNER_ACTIVATION_OFFLOAD=1
+            assert_info:
+                base_metric: glm5-2-sft-30B-mtp-fp8/tracker.jsonl
+                check_metrics:
+                    grad_norm: 0.1
+                    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
+
+    glm5-2-sft-30B-sp2:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/glm5p2_30B_sp2.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                cpus_per_task: 120
+                memory_per_task: 1200
+                envs:
+                    - MODEL_PATH=/mnt/shared-storage-user/llmrazor-share/model/GLM-5.2-30B
+                    - ALPACA_PATH=/mnt/shared-storage-user/llmrazor-share/data/alpaca
+                    - XTUNER_DETERMINISTIC=true
+                    - PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+            assert_info:
+                base_metric: glm5-2-sft-30B-sp2/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
+
+    glm5-2-sft-30B-ep2-resume:
+        -
+            type: sft
+            phase: first
+            parameters:
+                config: autotest/config/glm5p2_30B_ep2_resume.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/GLM-5.2-30B
+                    - ALPACA_PATH=/mnt/shared-storage-user/llmrazor-share/data/alpaca
+                    - XTUNER_DETERMINISTIC=true
+                    - XTUNER_GC_ENABLE=1
+            assert_info:
+                base_metric: glm5-2-sft-30B-ep2-resume/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
+
+        -
+            type: sft
+            phase: resume
+            pre_action:
+                command: 'python ./autotest/utils/update_meta.py /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output glm5-2-sft-30B-ep2-resume sft'
+            parameters:
+                config: autotest/config/glm5p2_30B_ep2_resume.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                num_nodes: 2
+                cpus_per_task: 80
+                memory_per_task: 1200
+                envs:
+                    - MODEL_PATH=/mnt/shared-storage-user/llmrazor-share/model/GLM-5.2-30B
+                    - ALPACA_PATH=/mnt/shared-storage-user/llmrazor-share/data/alpaca
+                    - XTUNER_DETERMINISTIC=true
+                    - XTUNER_GC_ENABLE=1
+            assert_info:
+                base_metric: glm5-2-sft-30B-ep2-resume/tracker-resume.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
+            post_action:
+                command: 'python ./autotest/utils/resume_validation.py /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output glm5-2-sft-30B-ep2-resume sft 10 "runtime_info/text_tokens,runtime_info/efficient_attn_ratio,loss/reduced_balancing_loss,loss/reduced_llm_loss,loss/local_loss"'
+            timeout: 1500
+
+    glm5-2-sft-30B-ep8-sp2-tp2:
+        -
+            type: sft
+            parameters:
+                config: autotest/config/glm5p2_30B_ep8_sp2_tp2.py
+                output_path: /mnt/shared-storage-user/llmrazor-share/qa-llm-cicd/test_output
+            resource:
+                num_nodes: 2
+                cpus_per_task: 80
+                envs:
+                    - MODEL_PATH=/mnt/shared-storage-user/llmrazor-share/model/GLM-5.2-30B
+                    - ALPACA_PATH=/mnt/shared-storage-user/llmrazor-share/data/alpaca
+                    - XTUNER_DETERMINISTIC=true
+                    - PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+            assert_info:
+                base_metric: glm5-2-sft-30B-ep8-sp2-tp2/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: 1800
+
     # InternS2-Preview: Qwen3.5-VL-MoE-35B-A3B architecture; MODEL_PATH may need adjust on share.
     s2-preview-rl-vl-lmdeploy-grpo:
         -

autotest/config/glm5p2_30B.py

@@ -0,0 +1,63 @@
+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 get_model_config_from_hf
+from xtuner.v1.train import TrainerConfig
+
+
+MODEL_PATH = os.environ["MODEL_PATH"]
+ALPACA_PATH = os.environ["ALPACA_PATH"]
+
+ep_size = 4
+
+moe_cfg = get_model_config_from_hf(MODEL_PATH)
+moe_cfg.dispatcher = "deepep"
+moe_cfg.ep_size = ep_size
+moe_cfg.compile_cfg = False
+if hasattr(moe_cfg.attention, "sparse_mla_backend"):
+    moe_cfg.attention.sparse_mla_backend = "tilelang"
+
+optim_cfg = AdamWConfig(lr=6e-05)
+lr_cfg = LRConfig(lr_type="cosine", lr_min=1e-6)
+fsdp_cfg = FSDPConfig(
+    cpu_offload=False,
+    ep_size=ep_size,
+)
+
+dataset_config = [
+    {
+        "dataset": DatasetConfig(name="alpaca", anno_path=ALPACA_PATH, sample_ratio=1.0),
+        "tokenize_fn": OpenaiTokenizeFunctionConfig(chat_template="glm5.2", max_length=4096),
+    },
+]
+
+dataloader_config = DataloaderConfig(pack_max_length=16384)
+
+loss_cfg = CELossConfig(mode="chunk", chunk_size=1024)
+moe_cfg.lm_loss_cfg = loss_cfg
+
+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,
+    strict_load=True,
+    global_batch_size=8,
+    intra_layer_micro_batch=1,
+    sp_size=1,
+    total_epoch=1,
+    work_dir=f"{os.environ['WORK_DIR']}",
+    seed=0,
+)