# Mugi "stable" MLX stack for Settings → Revert to stable.
# These versions are validated to resolve together (pip can install in one pass).
#
# The local model server is mlx-vlm's built-in OpenAI-compatible server
# (`python -m mlx_vlm.server`) — mlx-openai-server has been retired. mlx 0.31.2
# made GPU streams thread-local (mlx#3281/#3405), which broke mlx-openai-server's
# worker-thread handler; mlx-vlm 0.6.2's continuous-batching server is stable on
# 0.31.2 (validated: text-only + vision + LoRA + native delta.tool_calls + concurrency).
#
# transformers is PINNED (not floating) because mlx-embeddings loads in-process from
# THIS venv; a transformers major bump can break knowledge/session embeddings at runtime.
#
# Bump only after on-device testing (vision + text + tool calls). The floating
# requirements-mlx.txt may track newer releases; this file is the escape hatch.
# Deeper rollback to the retired mlx-openai-server stack: requirements-mlx-openai-legacy.txt.
mlx==0.31.2
mlx-metal==0.31.2
mlx-lm[train]==0.31.3
mlx-vlm==0.6.2
transformers==5.10.2
# Knowledge/session semantic search (separate from chat mlx-lm); required for vector index.
mlx-embeddings>=0.0.5,<0.2
