Issues

• Without a host wired-memory ticket nothing is wired: 100 GB-class models decode 40x slow, and the docs still say the library handles it

  osaurus-ai/vmlx-swift#418 · Filter by author rcfa opened on Sep 2, 2026

• Family-gated predicates exclude every new model family by construction

  osaurus-ai/vmlx-swift#413 · Filter by author rcfa opened on Sep 2, 2026

• CI: no macOS build coverage, and the tree fails swift-format (611 files)

  osaurus-ai/vmlx-swift#400 · Filter by author rcfa opened on Sep 1, 2026
  1

• MTPRuntimeFocusedTests: 13 JANG quant-precedence expectations stale since #373, masked by #379's import break

  osaurus-ai/vmlx-swift#393 · Filter by author jjang-ai opened on Sep 1, 2026

• Disk cache telemetry: ladder probes inflate the miss counter ~10x

  osaurus-ai/vmlx-swift#389 · Filter by author jjang-ai opened on Sep 1, 2026

• MTP verify prefetch: resolve the QSA-indexer regression (opt-in on perf/mtp-cycle-host-overhead)

  osaurus-ai/vmlx-swift#384 · Filter by author jjang-ai opened on Sep 1, 2026

• Speed parity vs Python vMLX: baseline ledger & iteration tracking

  osaurus-ai/vmlx-swift#383 · Filter by author jjang-ai opened on Sep 1, 2026
  7

• Two failures on current main: native-MTP auto-launch resolves nil, and a BF16 affine test NaNs only in a full run

  osaurus-ai/vmlx-swift#362 · Filter by author rcfa opened on Aug 30, 2026

• sanitize(weights:) can silently do nothing: two protocol requirements, both with defaults

  osaurus-ai/vmlx-swift#334 · Filter by author rcfa opened on Aug 29, 2026
  1

• [PROPOSAL] Generation-visible dynamic tools with measurable prefix reuse

  osaurus-ai/vmlx-swift#213 · Filter by author tpae opened on Aug 5, 2026

• Host failures amid severe memory pressure during 64K-token prefill on Llama-3.3-70B-8bit (M5 Max, 128 GB) — observed state, and questions about memory ceilings

  osaurus-ai/vmlx-swift#138 · Filter by author mimeding opened on Jul 11, 2026

• Hybrid SSM prefix hit restores the full payload, then unsafeFullHit discards it — identical resubmission pays restore cost, skips zero tokens

  osaurus-ai/vmlx-swift#132 · Filter by author mimeding opened on Jul 11, 2026

• Fused dequant-in-attention KV kernels: unfused TurboQuant never wins (3 architectures × 8K–64K measured); plus a 32K disk-cache cliff observation

  osaurus-ai/vmlx-swift#124 · Filter by author mimeding opened on Jul 7, 2026
  2

• Native-MTP tuned speedup (1.56×) does not survive the serving path on hybrid MoE bundles — AR/MTP parity at 76–80 tok/s

  osaurus-ai/vmlx-swift#120 · Filter by author mimeding opened on Jul 7, 2026
  1

• Inconsistent env-var prefix (VMLX_ vs VMLINUX_) for the mmap safetensors knobs

  osaurus-ai/vmlx-swift#99 · Filter by author rcfa opened on Jun 30, 2026
  1

• Rewrite README for consolidated vmlx-swift package

  osaurus-ai/vmlx-swift#2 · Filter by author jjang-ai opened on May 19, 2026