All pull requests

Pull requests

• Video generation: MiniMax-H3 (VDN-H3) engine for M5 Pro

  #502·zaiyang-zhang opened Oct 19, 2026Contributor·

  ·

• Add a live engine performance dashboard at /dashboard

  #501·vjanelle opened Oct 19, 2026·

  ·

• Prepare a model for the server from local files

  #500·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Add experimental DFlash2 draft fine-tuning command

  #499·linson007 opened Oct 19, 2026Contributor·

  ·

• Empty the persistent cache only for a repeated or engine-reported failure

  #498·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Add an opt-in deterministic policy for greedy identity across batch, cache and prefill shapes (stacked on #469)

  #497·Vardaan97 opened Oct 19, 2026·

  ·

• Let concurrent MTP lanes choose their own depth and draft vocabulary (stacked on #448)

  #496·Vardaan97 opened Oct 19, 2026·

  ·

• Give auto's PLE row cache 512 MiB and the streamed experts the rest

  #495·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Write host-step events from one side and give a failure's first cause

  #494·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Let Flash-Next's down projection prepare its own table at decode

  #493·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Split K sixteen ways on the register tile's narrowest grids

  #492·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Read the float projection's weights 16 bytes a lane

  #491·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Decode one lane under 4.5 bits per weight on the 16-row tile

  #490·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Keep each block list row's mask word in a register

  #489·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Spread the sparse verify split over a one-dimensional grid

  #488·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Choose a MoE target's verify rows by expected tokens per cost

  #487·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Drop a </think> the model writes after its reasoning ended

  #486·jianc99 opened Oct 19, 2026Collaborator·

  ·

• File backend_regression's short-prompt samples by their rows again

  #485·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Keep a turn's replay point under host memory pressure

  #484·jianc99 opened Oct 19, 2026Collaborator·

  ·

• WIP: M2 / apple8 support

  #482·kenji21 opened Oct 19, 2026·

  ·

• Plan memory from one budget and turn on KV on the SSD

  #481·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Decide per request whether its KV spills, and run prefill in passes

  #480·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Read spilled K/V in the runtime and the Qwen mixers

  #479·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Spill committed KV to the SSD cache through the tiered block store

  #478·jianc99 opened Oct 19, 2026Collaborator·

  ·

• Add the hot block cache and the history stream that read spilled KV

  #477·jianc99 opened Oct 19, 2026Collaborator·

  ·