robbiemu/paper2-dia2-coreml-output-heads

Dia2 Core ML output heads

Three downloadable Core ML model packages: the action head, codebook-0 head and depformer-stage-14 head.

Model identity and compatibility

• Upstream model: nari-labs/Dia2-2B. The study checkpoint is pinned to 7abae125471a73b0fc6b9d413cb15f4ae1e771d8.
• Derivation: Core ML conversion of three output projections used with an MPS FP32 generator and guarded host selection.
• Language scope: English, matching the upstream model and the retained generation request.
• Interface: The action and codebook-0 heads take hidden states with shape [2,1,2048]; the stage-14 depformer head takes [2,1,1024]. They return logits. The generator, recurrent caches, sampler, native fallback and Mimi decoder are separate dependencies.
• Validation scope: The retained composition agrees with the strict CUDA FP32 reference for one 32-frame request, repeated twice. It does not match the CUDA BF16 target and is not a complete generator checkpoint.
• Data role: The released projections are converted model components, not models pretrained on the private evidence dataset.

The base-model link records component lineage. Conversion, graph adjustment and fitting are described separately above. Core ML input/output dtypes do not by themselves specify internal compute precision.

The documented Python Core ML loading path requires a compatible macOS Core ML runtime. Recorded Core ML Tools versions: 9.0. Exact specification versions, conversion dates, source-framework versions and I/O dtypes are preserved in coreml_specs.json; current cross-version runtime compatibility has not been retested.

release_metadata.json preserves this scope as machine-readable data. The private evidence dataset is linked for experiment records and is not labeled as a training dataset.

File and tensor details

This release uses Core ML packages. It contains no Safetensors weights, so the Hub's automatic Safetensors panel does not apply.

Hugging Face's Safetensors metadata parser reads tensor headers from Safetensors files; README metadata alone cannot supply that panel.

Format	Packages or files	Stored size
Core ML package	3	16.03 MiB

Browse the per-file details, or inspect tensor names, shapes and storage dtypes. These sizes include separate alternatives; they are not a combined model parameter count.

Package	Stored size	Input → output dtypes	Specification
models/action-head.mlpackage	10,574 B	FLOAT32 → FLOAT32	parseable
models/codebook-0-head.mlpackage	8.01 MiB	FLOAT32 → FLOAT32	parseable
models/depformer-14-head.mlpackage	8.01 MiB	FLOAT32 → FLOAT32	parseable

Core ML input/output dtypes describe the interface, not internal compute or weight precision.

Model files

Model / configuration	Files	Role
action-head	models/action-head.mlpackage	Action head
codebook-0-head	models/codebook-0-head.mlpackage	Codebook-0 head
depformer-14-head	models/depformer-14-head.mlpackage	Depformer-stage-14 head

Download and load

from huggingface_hub import snapshot_download
from pathlib import Path
import coremltools as ct

root = Path(snapshot_download("robbiemu/paper2-dia2-coreml-output-heads"))
package = root / "models/action-head.mlpackage"
spec = ct.utils.load_spec(str(package))
print(spec.description.input)
model = ct.models.MLModel(str(package), compute_units=ct.ComputeUnit.CPU_ONLY)
# Supply the named arrays and shapes listed in coreml_specs.json:
# output = model.predict(inputs)

The example selects CPU execution explicitly. Use the study’s declared compute policy when reproducing a comparison. coreml_specs.json lists exact input/output names, shapes and conversion metadata. manifest.json records file sizes, SHA-256 hashes and provenance paths. python restore.py --verify checks all model files.

Scope and result

These heads accompany an MPS FP32 generator and guarded host selection with native fallback. The retained composition matches a strict CUDA FP32 reference for one 32-frame request, repeated twice. The complete upstream generator and Mimi codec are separate dependencies. The CUDA BF16 target was not matched.

These model files were published from retained study artifacts. Their integrity and Core ML interfaces were checked; publication does not constitute a new inference or deployment validation. Experimental evidence, diagnostic probes, transfer intermediates and optimization history are held in separate private repositories.

Provenance and license

Upstream notices/model cards are retained under licenses/; the applicable model license is in LICENSE. The private evidence dataset records historical source revisions, execution policies and results.