Skip to content

Latest commit

 

History

History
24 lines (21 loc) · 3.19 KB

File metadata and controls

24 lines (21 loc) · 3.19 KB

MLX Workshop Beta Support Matrix

This matrix distinguishes qualified release paths from experimental tools and fail-closed capability checks. “Implemented” does not mean every model family is compatible or that an experimental candidate is safe to promote.

Path Beta evidence Classification User-visible rule
Dense Llama uniform MXFP4/MXFP8/affine conversion Inspect, plan review, explicit confirmation, immutable output, reload, exact-parent verification, and three hashed gates Supported when inspection and resource review pass The original model is never modified
Run lifecycle Streaming logs, cancellation, interruption, relaunch recovery, safe resume, and recovered cancellation Supported Every mutating run requires confirmation
Compare and local release staging Parent/candidate fingerprints, gate evidence, honest optional metrics, and immutable reference-only release records Supported for qualified uniform candidates Missing speed, KL, or score data is shown as “Not measured”
Calibration-driven mixed precision Real parent-relative logits/KL measurements, Llama layer-group adapter, Pareto assignments, exact saved assignment, and reload verification Experimental and unqualified Analyze and materialize are available; the result cannot be promoted without full qualification
Refusal-direction behavior editing Reviewed six-step allowlisted executor, separated dataset groups, provenance hashes, adapter checks, logs, and held-out gate manifest Experimental; executable only for eligible quantized Qwen3.5 models Critical/tool/template parity is not yet measured, so promotion fails closed
Vision smoke test Capability gate plus a local image-to-text generation through the bundled mlx-vlm runtime Capability check for models that advertise vision weights Text-only models block before launching inference; this is not vision grafting
MTPLX compatibility Read-only local mtplx inspect --require-mtp --json Capability check Never starts, stops, or reconfigures an MTPLX daemon
Host snapshot Chip, memory, disk, runtime versions, power/thermal context, and sanitized MLX-related workload names Supported and read-only Active workloads are reported, never stopped
Unknown tensor semantics, native FP8 inputs, unknown architectures, or missing adapters Explicit blocker with remediation guidance Supported fail-closed outcome No conversion is attempted
Generated tiny dense Llama fixture Real MXFP4 conversion, reload/generation, parent immutability, mixed-precision measurement, qualification, and a second full run through the bundled Release-app runtime without manifest drift Required CI/reference evidence; no quality claim Tiny and deterministic, not representative of model quality

Beta v1 does not claim universal model compatibility, automatic vision grafting, automatic MTP packaging, canonical-model replacement, behavior-edit promotion, or model quality. Public source builds and ad-hoc local packages are ready; a binary that friends can open without Gatekeeper workarounds still requires Developer ID signing and Apple notarization as recorded in ../BETA_REPORT.md.