Guide for the weight lifecycle — downloading trained weights from Tinker, merging LoRA adapters into HuggingFace models, and publishing to HuggingFace Hub. Use when the user asks about exporting, downloading, merging, or publishing trained model weights.
中文简介正在准备中。当前显示原文说明。
安装前可以查看交给 AI 代理的指令内容。
The tinker_cookbook.weights subpackage provides a standard pipeline for trained weight management: download → build → publish.
Read these for details:
tinker_cookbook/weights/__init__.py — API overview and workflow exampletinker_cookbook/weights/_download.py — Download implementationtinker_cookbook/weights/_export.py — LoRA merge implementationtinker_cookbook/weights/_publish.py — HuggingFace Hub publishdocs/download-weights.mdx — Download guidedocs/publish-weights.mdx — Publishing guidedocs/save-load.mdx — Checkpointing (save_weights_for_sampler vs save_state)from tinker_cookbook import weights
# Step 1: Download adapter from Tinker
adapter_dir = weights.download(
tinker_path="tinker://run-id/sampler_weights/final",
output_dir="./adapter",
)
# Step 2: Merge LoRA adapter into base model
weights.build_hf_model(
base_model="Qwen/Qwen3.5-35B-A3B",
adapter_path=adapter_dir,
output_path="./model",
dtype="bfloat16", # or "float16", "float32"
)
# Step 3: Publish to HuggingFace Hub
url = weights.publish_to_hf_hub(
model_path="./model",
repo_id="user/my-finetuned-model",
private=True,
)
weights.download()Downloads and extracts a checkpoint archive from Tinker.
adapter_dir = weights.download(
tinker_path="tinker://run-id/sampler_weights/final", # Tinker checkpoint path
output_dir="./adapter", # Local directory to extract to
base_url=None, # Optional custom Tinker API URL
)
# Returns: path to extracted directory
weights.build_hf_model()Merges a LoRA adapter into a base model, producing a full HuggingFace model.
weights.build_hf_model(
base_model="Qwen/Qwen3-8B", # HF model name or local path
adapter_path="./adapter", # Directory with adapter_model.safetensors
output_path="./model", # Where to save merged model
dtype="bfloat16", # Weight dtype
trust_remote_code=None, # Override HF_TRUST_REMOTE_CODE
)
weights.publish_to_hf_hub()Pushes a local model directory to HuggingFace Hub.
url = weights.publish_to_hf_hub(
model_path="./model", # Local model directory
repo_id="user/my-finetuned-model", # HF repo ID
private=True, # Private repo
token=None, # HF token (uses HF_TOKEN env var if None)
)
# Returns: URL to published repo
weights.build_lora_adapter() (not yet implemented)Convert Tinker LoRA adapter to standard format for vLLM/SGLang. Currently raises NotImplementedError — use build_hf_model() instead.
During training, there are two types of checkpoints:
save_state() — Full state (weights + optimizer). Used for resuming training.save_weights_for_sampler() — Weights only. Used for sampling and export.The weights.download() function works with sampler weights (save_weights_for_sampler checkpoints).
download() expects a tinker:// path from save_weights_for_sampler, not save_statebuild_hf_model() requires the base model to be downloadable from HuggingFaceHF_TOKEN environment variable for private models and publishingdtype="bfloat16" is recommended for most models还没有评价。欢迎分享你的使用感受。
无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。
Guide for checkpointing — saving, loading, and resuming training with CheckpointRecord. Use when the user asks about saving weights, resuming training, checkpoint management, or the checkpoint lifecycle.
中文简介正在准备中。当前显示原文说明。
Guide for testing conventions and CI pipelines — unit tests, integration smoke tests, pytest markers, and GitHub Actions workflows. Use when the user asks about testing, CI, running tests, or adding tests for a recipe.
中文简介正在准备中。当前显示原文说明。
Guide for using completers — TokenCompleter and MessageCompleter for text generation during RL rollouts and evaluation. Use when the user asks about generating text, completing messages, or using completers in RL environments.
中文简介正在准备中。当前显示原文说明。
Guide for contributing to the tinker-cookbook repo — development setup, code style, type checking, PR process, and design conventions. Use when the user asks about how to contribute, set up the dev environment, code style, or project conventions.
中文简介正在准备中。当前显示原文说明。
Guide for dataset construction — SupervisedDatasetBuilder, RLDatasetBuilder, ChatDatasetBuilder, and custom dataset creation from JSONL, HuggingFace, or conversation data. Use when the user asks about datasets, data loading, data preparation, or custom data formats.
中文简介正在准备中。当前显示原文说明。
Set up and run knowledge distillation (on-policy, off-policy, or multi-teacher) from a teacher model to a student model using the Tinker API. Use when the user wants to distill knowledge, compress models, or train a student from a teacher.
中文简介正在准备中。当前显示原文说明。