All issues

Search results

regression(kv): none/bf16 long-ctx decode regressed vs 0.2.5 — -22% at 32k, b_none marginal cost +25%, mlx-lm lead lost at 32k

Status: Open.

#297 In Pushkinist/rMLX;

· Pushkinist opened on Jul 23, 2026

regression(kv): *_sym fused flash-decode has 17-27x higher marginal cost than bf16 — slower than the CPU path it replaced at every context

Status: Open.

#296 In Pushkinist/rMLX;

· Pushkinist opened on Jul 23, 2026

perf(kv): k_iso*/k_rotor* restage the growing prefix host-side every decode step — caps K-only codecs at ~23-25 TPS despite the Metal kernel dispatching

Status: Open.

#295 In Pushkinist/rMLX;

· Pushkinist opened on Jul 23, 2026

bug(kv): 11 kernel-less quant codecs regress 6-34% at mid/long ctx since 0.3.0 — none/k8v8 hold exact parity

Status: Open.

#293 In Pushkinist/rMLX;

· Pushkinist opened on Jul 21, 2026

bug(kv): quantized-V flash-decode diverges — *_sym regress at every ctx, now slower at 64k than the CPU path they replaced

Status: Open.

#292 In Pushkinist/rMLX;

· Pushkinist opened on Jul 21, 2026

bug(cli): rmlx baseline --prompt-tokens tokenizes raw JSON fixture text, not message content

Status: Open.

#291 In Pushkinist/rMLX;

· Pushkinist opened on Jul 21, 2026

Track: 1-bit affine weight support (Bonsai-27B-mlx-1bit) — blocked on upstream mlx#3161

Status: Open.

#207 In Pushkinist/rMLX;

· Pushkinist opened on Jul 15, 2026

Serialize ssd_tier startup_maintenance test to kill rare OnceLock flake

Status: Open.

#117 In Pushkinist/rMLX;

· Pushkinist opened on Jun 17, 2026

feat(kv): fused quantized decode SDPA — make KV codecs live at decode (read quant store directly)

enhancementNew feature or requestNew feature or request

Status: Open.

#45 In Pushkinist/rMLX;

· Pushkinist opened on Jun 9, 2026

SSD KV tier: live reconfiguration (enable/disable + resize) on a resident model — control-plane drain+reattach

enhancementNew feature or requestNew feature or request

Status: Open.

#30 In Pushkinist/rMLX;

· Pushkinist opened on Jun 7, 2026