add qwen35 rl config and fix mix bug

examples/v1/config/rl_qwen3p5_vl_35B_grpo_mixdata.py

@@ -0,0 +1,178 @@
+import os
+from transformers import AutoTokenizer
+from xtuner.v1.config import (
+    AdamWConfig,
+    FSDPConfig,
+    LRConfig,
+)
+import json
+from xtuner.v1.ray.judger.dapo_math import DapoMathJudgerConfig
+from xtuner.v1.data_proto.rl_data import SampleParams
+from xtuner.v1.model import Qwen3_5_VLMoE35BA3Config
+from xtuner.v1.ray.base import AcceleratorResourcesConfig
+from xtuner.v1.ray.config.worker import RolloutConfig
+from xtuner.v1.ray.dataflow import DataFlowConfig, ReplayBufferConfig
+from xtuner.v1.ray.judger.controller import JudgerConfig
+from xtuner.v1.rl.base import WorkerConfig
+from xtuner.v1.rl.grpo import GRPOLossConfig
+from xtuner.v1.train.rl_trainer import RLTrainerConfig
+from xtuner.v1.datasets import RLTokenizeFnConfig, DatasetConfig, Qwen3VLTokenizeFnConfig, DataloaderConfig
+from xtuner.v1.rl.base.rollout_is import RolloutImportanceSampling
+
+work_dir = os.environ["WORK_DIR"]
+model_path = os.environ["MODEL_PATH"]
+meta_data_path = os.environ["DATA_PATH"]
+
+# basic settings
+experimental_name = "grpo_mix_data"
+total_epochs = 15
+global_batch_size = 256
+prompt_repeat_k = 8
+rollout_tp_size = 2
+rollout_ep_size = 1
+max_prompt_length = 2048
+max_response_length = 8192
+pack_max_length = 32768
+train_optimizer_steps = 8
+hf_interval = 15
+
+# 1. resources
+resources = AcceleratorResourcesConfig(
+    accelerator="GPU",
+    num_workers=8,
+    num_cpus_per_worker=12,
+    cpu_memory_per_worker=16 * 1024**3,  # 16 GB
+)
+
+# 2. rollout
+rollout_config = RolloutConfig(
+    fp32_lm_head=True,
+    env=experimental_name,
+    device=resources.accelerator,
+    model_path=model_path,
+    dtype="bfloat16",
+    tensor_parallel_size=rollout_tp_size,
+    expert_parallel_size=rollout_ep_size,
+    gpu_memory_utilization=0.8,
+    context_length = max_response_length + max_prompt_length,
+    enable_return_routed_experts=True,
+    rollout_max_batch_size_per_instance=512,
+)
+
+# sampling params
+training_sample_params = SampleParams(
+    max_tokens=max_response_length,
+    top_k=0,
+    top_p=1.0,
+    temperature=1.0,
+    min_tokens=0,
+)
+
+
+tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+
+ds_collections = json.loads(open(meta_data_path).read())
+train_dataset_cfg = []
+for name, _data in ds_collections.items():
+    tokenize_fn_cfg = Qwen3VLTokenizeFnConfig(processor_path=model_path, 
+                                            max_length=max_prompt_length, 
+                                            system_message=_data.get('system_message', None),
+                                            template_name="qwen3.5-vl-rl")
+    _data_cfg = {"dataset": DatasetConfig(name=name,
+                                          anno_path=_data['annotation'],
+                                          media_root=_data.get('media_root', ''),
+                                          sample_ratio=_data.get('sample_ratio', 1.0),
+                                          class_name='VLMJsonlDataset'),
+                "tokenize_fn": RLTokenizeFnConfig(max_length=max_prompt_length,
+                                                tokenize_fn_cfg=tokenize_fn_cfg),
+                 }
+    train_dataset_cfg.append(_data_cfg)
+
+dataloader_config = DataloaderConfig(num_workers=8,
+                                     collator="fake_collator",
+                                     pack_level="none")
+
+# 3. judger
+from xtuner.v1.utils.rl_test_utils import get_eos_token
+eos_token_id = get_eos_token(model_path)
+eos_token_str = tokenizer.convert_ids_to_tokens(eos_token_id)
+dapomath_judger_config = DapoMathJudgerConfig(
+    judger_name="dapo_math", 
+    eos_token=eos_token_str,
+    enable_overlong_buffer = True, 
+    max_response_len =max_response_length, 
+    overlong_buffer_len=4096, 
+    overlong_penalty_factor=1.0, 
+    tokenizer=tokenizer)
+judger_cfg = JudgerConfig(reward_judger_configs=[dapomath_judger_config])
+
+# 4. dataflow and evaluator
+dataflow_config = DataFlowConfig(
+    env=experimental_name,
+    prompt_repeat_k=prompt_repeat_k,
+    global_batch_size=global_batch_size,
+    sample_params=training_sample_params,
+    # max_concurrent=64,  # optional, will be determined automatically if not set
+)
+
+
+# replay buffer config: : 不需要修改
+replay_buffer_cfg = ReplayBufferConfig(
+    dataset_cfg=train_dataset_cfg, dataloader_cfg=dataloader_config, tokenizer=tokenizer
+)
+
+# 5. Train worker
+# NOTE: modify model_cfg
+model_cfg = Qwen3_5_VLMoE35BA3Config(freeze_vision=True, freeze_projector=True)
+optim_cfg = AdamWConfig(lr=1e-6, betas=(0.9, 0.999), max_grad_norm=1.0, weight_decay=0.1, foreach=False)
+loss_cfg = GRPOLossConfig(
+    policy_loss_cfg=dict(
+        cliprange_high=0.28,
+        cliprange_low=0.2,
+        loss_type="vanilla",
+        clip_ratio_c=10.0,
+        log_prob_diff_min=-20.0,
+        log_prob_diff_max=20.0,
+    ),
+    ignore_idx=-100,
+    use_kl_loss=False,
+    kl_loss_coef=0.0,
+    kl_loss_type="low_var_kl",
+    mode="chunk",
+    chunk_size=512,
+    rollout_is=RolloutImportanceSampling(
+        rollout_is_level="token",
+        rollout_is_mode="both",
+        rollout_is_threshold=(5, 0.5),
+        rollout_is_mask_threshold=(5, 0.5),
+        rollout_is_veto_threshold=(20, 0),
+    ),
+)
+lr_cfg = LRConfig(lr_type="constant", warmup_ratio=0, lr_min=1e-6)
+fsdp_cfg = FSDPConfig(torch_compile=False, cpu_offload=False, ep_size=1, fp32_lm_head=True)
+train_worker_cfg: WorkerConfig = WorkerConfig(
+    model_cfg=model_cfg,
+    load_from=model_path,
+    optim_cfg=optim_cfg,
+    loss_cfg=loss_cfg,
+    lr_cfg=lr_cfg,
+    fsdp_cfg=fsdp_cfg,
+    sp_size=1,
+    optimizer_steps=train_optimizer_steps,
+    pack_max_length=pack_max_length,
+)
+
+# 6. RL Trainer
+trainer = RLTrainerConfig(
+    load_from=model_path,
+    resources=resources,
+    rollout_config=rollout_config,
+    dataflow_config=dataflow_config,
+    judger_config=judger_cfg,
+    replay_buffer_config=replay_buffer_cfg,
+    train_worker_config=train_worker_cfg,
+    tokenizer_path=model_path,
+    work_dir=work_dir,
+    total_epochs=total_epochs,
+    hf_interval=hf_interval,
+)

xtuner/v1/data_proto/templates/__init__.py

@@ -46,6 +46,22 @@
         image_context_token="<|image_pad|>",
         video_context_token="<|video_pad|>",
     ),
+    "qwen3.5-vl-rl": HybridChatTemplate(
+        system="<|im_start|>system\n{system}<|im_end|>\n",
+        tool_prompt="\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\n"
+        "You are provided with function signatures within <tools></tools> XML tags:\n<tools>{tool_text}"
+        "\n</tools>\n\nFor each function call, return a json object with function name and arguments within "
+        """<tool_call></tool_call> XML tags:\n<tool_call>\n{{"name": <function-name>, """
+        """"arguments": <args-json-object>}}\n</tool_call>""",  # TODO: fix tool call
+        tool_extractor="<|im_start|>user\n<tool_response>\n{tool_extractor}\n</tool_response><|im_end|>\n<|im_start|>assistant\n",
+        user="<|im_start|>user\n{user}<|im_end|>\n<|im_start|>assistant\n<think>\n",  # only add <think>\n to the end
+        stop_words=["<|im_end|>", "<|endoftext|>"],
+        assistant="{assistant}<|im_end|>",
+        image_start_token="<|vision_start|>",
+        image_end_token="<|vision_end|>",
+        image_context_token="<|image_pad|>",
+        video_context_token="<|video_pad|>",
+    ),
     "llama3": HybridChatTemplate(
         system="<|start_header_id|>system<|end_header_id|>\n\n{system}<|eot_id|>",
         user=(

xtuner/v1/datasets/mllm_tokenize_fn/qwen3_vl_tokenize_fn.py

@@ -233,6 +233,7 @@ def __init__(
         hash: str | None = None,
         add_eos_token: bool = True,  # for mllm pretrain
         add_bos_token: bool = False,  # for mllm pretrain
+        template_name: str = "qwen3-vl",
     ):
         self.oss_loader = None
         self.debug = debug
@@ -294,7 +295,7 @@ def __init__(
             f"rand_video_max_frames: {self.rand_video_max_frames}"
         )
 
-        self.chat_template = CHAT_TEMPLATE_MAP["qwen3-vl"]
+        self.chat_template = copy.deepcopy(CHAT_TEMPLATE_MAP[template_name])
         if system_message is not None:
             self.chat_template.default_system = system_message
 
@@ -904,13 +905,16 @@ class Qwen3VLTokenizeFnConfig(BaseMLLMTokenizeFnConfig):
     # it's helpful to add labels to the images and videos for better reference.
     add_vision_id: bool = True
 
+    template_name: str = "qwen3_vl"
+
     def build(
         self, tokenizer, tokenizer_hash: str | None = None, anno_name: str = "", **kwargs
     ) -> Qwen3VLTokenizeFunction:
         return Qwen3VLTokenizeFunction(
             tokenizer,
             self.processor_path,
             anno_name,
+            template_name=self.template_name,
             min_pixels=self.min_pixels,
             max_pixels=self.max_pixels,
             oss_loader_cfg=self.oss_loader_cfg,

xtuner/v1/ray/dataflow/replay_buffer.py

@@ -278,7 +278,8 @@ def sample(self, env: str, prompt_repeat_k: int) -> List[RLDataFlowItem]:
         multimodal_train_info = data.pop("multimodal_train_info", {})
         if "pixel_values" in multimodal_train_info:
             multimodal_train_info["pixel_values"] = ray.put(multimodal_train_info["pixel_values"])
-            data["multimodal_train_info"] = multimodal_train_info
+        # If it is a mixture of pure text and image data, there will be position_id but no pixel_values
+        data["multimodal_train_info"] = multimodal_train_info
 
         for data_item in group_data_item:
             data_item.uid = RLUIDItem(