Audit fixes: MLX chat template, Ollama export preflight, migration intent preservation, smol-135m warning by espadonne · Pull Request #17 · tenseleyFlow/DocumentLanguageModel

Fix MLX backend silently producing base output for PEFT adapters

mfwolffe

committed

Fail loud when mlx-lm leaves the model un-wrapped after load

mfwolffe

committed

Apply chat template in MlxBackend.generate

mfwolffe

committed

Fall back to chat_template.jinja sibling in preflight

mfwolffe

committed

Allow GGUF reserved-slot vocab gap; relax != to >

mfwolffe

committed

Drop invalid PARAMETER draft_model from Modelfile; document via comment

mfwolffe

committed

Warn at train time when base has a registered capability_warning

mfwolffe

committed

Surface why tokenized cache is empty in 'dlm cache show'

mfwolffe

committed

Annotate doctor's MLX line with scope (prompt-only, default backend)

mfwolffe

committed

Preserve user-explicit frontmatter fields across migration

mfwolffe

committed

Apply ruff format

mfwolffe

committed

Cover collect_dict_field_paths branches

mfwolffe

committed

Cover both branches of doctor's MLX-availability render line

mfwolffe

committed

Cover MLX backend's chat-template helper

mfwolffe

committed

Cover OSError branch when reading chat_template.jinja sibling

mfwolffe

committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Audit fixes: MLX chat template, Ollama export preflight, migration intent preservation, smol-135m warning#17

Audit fixes: MLX chat template, Ollama export preflight, migration intent preservation, smol-135m warning#17
mfwolffe merged 15 commits into
trunkfrom
audit-fixes

Commits on May 2, 2026

Fix MLX backend silently producing base output for PEFT adapters

Fail loud when mlx-lm leaves the model un-wrapped after load

Apply chat template in MlxBackend.generate

Fall back to chat_template.jinja sibling in preflight

Allow GGUF reserved-slot vocab gap; relax != to >

Drop invalid PARAMETER draft_model from Modelfile; document via comment

Warn at train time when base has a registered capability_warning

Surface why tokenized cache is empty in 'dlm cache show'

Annotate doctor's MLX line with scope (prompt-only, default backend)

Preserve user-explicit frontmatter fields across migration

Apply ruff format

Cover collect_dict_field_paths branches

Cover both branches of doctor's MLX-availability render line

Cover MLX backend's chat-template helper

Cover OSError branch when reading chat_template.jinja sibling