Merge to main

Dockerfile

@@ -6,20 +6,18 @@ RUN curl -O https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.s
     rm Miniconda3-latest-Linux-x86_64.sh
 ENV PATH="/opt/conda/bin:${PATH}"
 RUN conda create -n twinkle python=3.12 -y --override-channels -c conda-forge
-SHELL ["conda", "run", "-n", "twinkle", "/bin/bash", "-c"]
+ENV PATH="/opt/conda/envs/twinkle/bin:${PATH}"
 
 # Clone and install twinkle, checkout to latest v-tag
 RUN git clone https://github.com/modelscope/twinkle.git
 WORKDIR /twinkle
-RUN echo "Available v-tags:" && git tag -l 'v*' --sort=-v:refname && \
-    LATEST_TAG=$(git tag -l 'v*' --sort=-v:refname | head -n 1) && \
-    echo "Checking out: $LATEST_TAG" && \
-    git checkout "$LATEST_TAG"
+RUN echo "Available release branches:" && git branch -r -l 'origin/release/*' --sort=-v:refname && \
+    LATEST_RELEASE=$(git branch -r -l 'origin/release/*' --sort=-v:refname | head -n 1 | tr -d ' ') && \
+    echo "Checking out: $LATEST_RELEASE" && \
+    git checkout --track "$LATEST_RELEASE"
 
 RUN sh INSTALL_MEGATRON.sh
 
 RUN pip install --no-cache-dir tinker==0.14.0 "ray[serve]" transformers peft accelerate -U
 
 RUN pip install -e . --no-build-isolation
-
-CMD ["bash", "cookbook/client/server/megatron/run.sh"]

INSTALL_MEGATRON.sh

@@ -4,7 +4,7 @@
 # which always occur error
 
 set -e  # Exit immediately on error
-
+export SETUPTOOLS_USE_DISTUTILS=local
 echo "=========================================="
 echo "Starting deep learning dependencies installation..."
 echo "=========================================="
@@ -55,8 +55,8 @@ echo "Using CUDA architecture: $TORCH_CUDA_ARCH_LIST"
 
 # Install latest base packages
 echo ""
-echo "Installing peft, accelerate, transformers, modelscope, oss2..."
-pip install --upgrade peft accelerate transformers "modelscope[framework]" oss2
+echo "Installing peft, accelerate, transformers, modelscope..."
+pip install --upgrade peft accelerate transformers "modelscope[framework]"
 
 # Install latest vllm
 echo ""
@@ -71,7 +71,9 @@ echo "Site-packages path: $SITE_PACKAGES"
 
 CUDNN_PATH=$SITE_PACKAGES/nvidia/cudnn \
 CPLUS_INCLUDE_PATH=$SITE_PACKAGES/nvidia/cudnn/include \
-pip install --no-build-isolation "transformer_engine[pytorch]" megatron_core --no-cache-dir
+pip install --no-build-isolation "transformer_engine[pytorch]" --no-cache-dir
+
+pip install megatron_core mcore_bridge --no-cache-dir
 
 # Install flash-attention (force local build)
 echo ""
@@ -87,11 +89,6 @@ pip install flash-linear-attention -U
 echo ""
 echo "Installing numpy==2.2 and deep_gemm..."
 pip install numpy==2.2
-pip uninstall deep_gemm -y
-cd /tmp
-git clone --recursive https://github.com/deepseek-ai/DeepGEMM.git
-cd DeepGEMM
-pip install . --no-build-isolation
 
 # Verify installation
 echo ""
@@ -100,7 +97,7 @@ echo ""
 python -c "
 import pkg_resources
 
-packages = ['peft', 'accelerate', 'transformers', 'modelscope', 'oss2', 'vllm', 'transformer_engine', 'megatron_core', 'flash_attn', 'numpy']
+packages = ['peft', 'accelerate', 'transformers', 'modelscope', 'vllm', 'transformer_engine', 'megatron_core', 'flash_attn', 'numpy']
 
 print('Installed package versions:')
 print('-' * 40)

cookbook/client/server/megatron/run.sh

@@ -3,4 +3,4 @@ export RAY_ROTATION_BACKUP_COUNT=1
 CUDA_VISIBLE_DEVICES=0,1,2,3 ray start --head --port=6379 --num-gpus=4 --disable-usage-stats --include-dashboard=false
 CUDA_VISIBLE_DEVICES=4,5,6,7 ray start --address=127.0.0.1:6379 --num-gpus=4
 CUDA_VISIBLE_DEVICES="" ray start --address=127.0.0.1:6379 --num-gpus=0
-python server.py
+python "$(dirname "$0")/server.py"

cookbook/client/server/megatron/server.py

@@ -9,7 +9,7 @@
 import os
 
 # Enable Ray debug mode for verbose logging during development
-os.environ['TWINKLE_TRUST_REMOTE_CODE'] = '1'
+os.environ['TWINKLE_TRUST_REMOTE_CODE'] = '0'
 
 from twinkle.server import launch_server
 

cookbook/client/server/megatron/server_config.yaml

@@ -42,7 +42,7 @@ applications:
     import_path: sampler
     args:
       model_id: "ms://Qwen/Qwen3.5-27B"  # ModelScope model identifier
-      nproc_per_node: 8               # Number of GPU processes per node
+      nproc_per_node: 4               # Number of GPU processes per node
       sampler_type: vllm              # Inference engine: 'vllm' (fast) or 'torch' (TorchSampler)
       engine_args:                    # vLLM engine-specific settings
         max_model_len: 32000           # Maximum sequence length the engine supports
@@ -84,7 +84,7 @@ applications:
       model_id: "ms://Qwen/Qwen3.5-27B" # ModelScope model identifier
       max_length: 32000                           # model max length
       max_loras: 5                                # model max loras
-      nproc_per_node: 8                           # Number of GPU processes per node
+      nproc_per_node: 4                           # Number of GPU processes per node
       device_group:
         name: model
         ranks: 4       # GPU rank indices

cookbook/client/tinker/modelscope/sample.py

@@ -45,7 +45,7 @@
     ]
 )
 
-input_feature = template.encode(trajectory, add_generation_prompt=True)
+input_feature = template.batch_encode([trajectory], add_generation_prompt=True)[0]
 
 input_ids = input_feature['input_ids'].tolist()
 

pyproject.toml

@@ -10,7 +10,7 @@ dependencies = [
   "datasets>=3.0,<4.0",
   "omegaconf>=2.3.0,<3.0.0",
   "fastapi",
-  "modelscope[framework]>=1.34.0",
+  "modelscope[framework]>=1.35.0",
   "safetensors",
   "peft>=0.11.0,<=0.19.0",
   "transformers",

src/twinkle/dataset/base.py

@@ -51,6 +51,9 @@ class Dataset(TorchDataset):
     """
 
     def __init__(self, dataset_meta: DatasetMeta, **kwargs):
+        trust_remote_code = bool(os.environ.get('TWINKLE_TRUST_REMOTE_CODE', '1'))
+        if not trust_remote_code:
+            kwargs['trust_remote_code'] = False
         dataset = self._load_dataset(dataset_meta, **kwargs)
         self.datasets = {dataset_meta.get_id(): dataset}
         self.dataset = dataset
@@ -247,6 +250,9 @@ def add_dataset(self, dataset_meta: DatasetMeta, **kwargs):
         Args:
             dataset_meta: The dataset_meta information of the loaded dataset.
         """
+        trust_remote_code = bool(os.environ.get('TWINKLE_TRUST_REMOTE_CODE', '1'))
+        if not trust_remote_code:
+            kwargs['trust_remote_code'] = False
         dataset = self._load_dataset(dataset_meta, **kwargs)
         self.datasets[dataset_meta.get_id()] = dataset
 

src/twinkle/hub/hub.py

@@ -401,7 +401,7 @@ def load_dataset(cls,
         cls.try_login(token)
         if revision is None or revision == 'main':
             revision = 'master'
-        load_kwargs = {'trust_remote_code': True}
+        load_kwargs = {'trust_remote_code': kwargs.get('trust_remote_code', True)}
         return MsDataset.load(
             dataset_id,
             subset_name=subset_name,
@@ -595,13 +595,15 @@ def load_dataset(cls,
         from datasets import load_dataset
         if revision is None or revision == 'master':
             revision = 'main'
+        trust_remote_code = kwargs.get('trust_remote_code', True)
         return load_dataset(
             dataset_id,
             name=subset_name,
             split=split,
             streaming=streaming,
             revision=revision,
             download_mode=download_mode,
+            trust_remote_code=trust_remote_code,
             num_proc=num_proc)
 
     @classmethod

src/twinkle/processor/base.py

@@ -97,7 +97,8 @@ def to_tensor(_input):
                 # so tensor ops like labels != ignore_index or .to(device) would fail without this.
                 if isinstance(value, np.ndarray):
                     value = torch.from_numpy(value)
-                elif isinstance(value, list) and isinstance(value[0], (int, float, np.number)):
+                elif (isinstance(value, list) and isinstance(value[0],
+                                                             (int, float, np.number))) or key == 'position_ids':
                     value = torch.tensor(value)
                 elif key in self.VLM_CONCAT_FIELDS:
                     if not isinstance(value[0], torch.Tensor):

src/twinkle/server/gateway/server.py

@@ -36,7 +36,7 @@ def __init__(self,
         self.http_options = http_options or {}
         self.proxy = ServiceProxy(http_options=http_options, route_prefix=self.route_prefix)
         self.supported_models = self._normalize_models(supported_models) or [
-            types.SupportedModel(model_name='Qwen/Qwen3.5-4B'),
+            types.SupportedModel(model_name='Qwen/Qwen3.5-27B'),
         ]
         self._modelscope_config_lock = asyncio.Lock()