New merge method API

[cg123]

Note

High Risk
Large refactor of merge execution, dtype promotion, and vocabulary metadata touches every weight merge path; misaligned vocab or truncation options can silently produce invalid checkpoints.

Overview
Introduces a unified merge-method execution model centered on TensorGroup / TensorBatch, with public merge_tensors and merge_state_dicts for in-memory merges (batched floating weights, strict buffer handling, dtype / out_dtype, and BatchOptions).

Built-in and custom methods are refactored around MergeMethodSpec, @merge_method (group vs batch kernels), register() / get(), and shared parameter binding (per-input maps, PerGroupValues, typed YAML resolution via evaluate_setting). Graph-facing code is adapted to method.spec metadata and optional-tensor policies instead of per-method Task classes where migrated (e.g. Arcee Fusion).

Vocabulary handling replaces WeightInfo.is_embed with vocabulary_axis across architecture JSON and auto-detection (infer_vocabulary_axes); several arch defs drop legacy tensor-space/head-split metadata. Docs expand tokenizer alignment, computation precision, and --unsafe-truncate-embeddings.

ConfigReader no longer resolves parameters per tensor name inline—it exposes parameter_sources for the planner to resolve with filters. Minor: float64 dtype in config, mergekit index version 0.1.5.

^{Reviewed by Cursor Bugbot for commit 34972dd. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want fixes drafted automatically? Bugbot Autofix can create code changes for findings. A team admin can enable Autofix in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 53757a3. Configure here.}

[cursor]

YAML dtype recasts non-floating buffers

Medium Severity

The YAML planner never passes dtype or out_dtype into ExecuteMergeMethodTask, so input casting happens in GatherTensors and output casting in SaveTensor. Both apply to every tensor, including integer or boolean buffers. That defeats copy_non_floating_buffer and can numerically merge or recast BatchNorm-style counters, contrary to the new API contract that non-floating buffers stay unchanged. The raw-PyTorch path already avoids this by casting only floating tensors at load and applying out_dtype inside the merge task after buffers are copied.

Additional Locations (2)

• mergekit/plan.py#L204-L208
• mergekit/merge_methods/task_adapter.py#L161-L164

 

^{Reviewed by Cursor Bugbot for commit 53757a3. Configure here.}