minimax_music3: performance pack — native rope/SwiGLU lowering + opt-in CFG reuse, chunk hop, batched ensemble takes, q4_k depth by iamwavecut · Pull Request #321 · 0xShug0/audio.cpp

MiniMax Music 3 base optimizations: slim semantic LM head, CUDA graph-cache eviction on release, pipeline buffer reuse

iamwavecut

committed

2b296ad

Overlap AR decoding with per-chunk denoise/vocode on a second, lower-priority CUDA stream

iamwavecut

committed

3614e77

Flow CFG guidance-delta reuse (flow_uncond_interval) with an exactness probe (MM3_DC_VERIFY)

iamwavecut

committed

ac2da48

Native NEOX rope for the batch-1 flow graph (works around a gallocr view-aliasing corruption) + b1 determinism diagnostics (MM3_DC_B1_*), MMQ fixup-buffer zeroing

iamwavecut

committed

7f33526

M13 recipe: native NEOX rope on all batches, fused swapped-half SwiGLU, overlap auto-off on long requests, delta-cache interval=2 default

iamwavecut

committed

0e1b7d6

Ensemble decoding (ensemble_takes): K takes share one batched AR pass; K=1 byte-equivalence force probe; per-stage AR timings

iamwavecut

committed

1d065ca

GPU-resident depth sampling (byte-exact, perf-neutral options): device philox top-k sampler (MM3_DEPTH_GPU_SAMPLE) and full GPU frame with device residual-ids and hidden accumulation (MM3_DEPTH_GPU…

iamwavecut

committed

5419c99

Wave-2 default: q4_k depth decoder (the stage is weight-bandwidth-bound: -24% depth, ~-6% wall; q8_0 stays available)

iamwavecut

committed

602df48

Intro-lock ensemble fork (ensemble_prefix_frames): shared master prefix at batch 2, batched decode KV export/replication, shared flow seed under the lock

iamwavecut

committed

ede87fb

Carry-free flow probe behind MM3_FLOW_NO_CARRY=1 (listening-rejected: seams audible; kept for hybrid experiments)

iamwavecut

committed

c9f8d30

Wave-5 recipe: configurable flow chunk hop (flow_chunk_hop_frames) with consistently rederived crops/carry; defaults hop=150 + delta-cache interval=3 (flow -40%, 60s solo 74.1s, K=6 ensemble 46.9s/…

iamwavecut

committed

4bb174e

Restore upstream defaults: every optimization stays opt-in

iamwavecut

committed

e458211

Navigation Menu

minimax_music3: performance pack — native rope/SwiGLU lowering + opt-in CFG reuse, chunk hop, batched ensemble takes, q4_k depth - #321

minimax_music3: performance pack — native rope/SwiGLU lowering + opt-in CFG reuse, chunk hop, batched ensemble takes, q4_k depth#321
iamwavecut wants to merge 12 commits into
0xShug0:mainfrom
iamwavecut:pr/mm3-perf-pack

Commits on Aug 28, 2026

MiniMax Music 3 base optimizations: slim semantic LM head, CUDA graph-cache eviction on release, pipeline buffer reuse

Overlap AR decoding with per-chunk denoise/vocode on a second, lower-priority CUDA stream

Flow CFG guidance-delta reuse (flow_uncond_interval) with an exactness probe (MM3_DC_VERIFY)

Native NEOX rope for the batch-1 flow graph (works around a gallocr view-aliasing corruption) + b1 determinism diagnostics (MM3_DC_B1_*), MMQ fixup-buffer zeroing

M13 recipe: native NEOX rope on all batches, fused swapped-half SwiGLU, overlap auto-off on long requests, delta-cache interval=2 default

Ensemble decoding (ensemble_takes): K takes share one batched AR pass; K=1 byte-equivalence force probe; per-stage AR timings

GPU-resident depth sampling (byte-exact, perf-neutral options): device philox top-k sampler (MM3_DEPTH_GPU_SAMPLE) and full GPU frame with device residual-ids and hidden accumulation (MM3_DEPTH_GPU…

Wave-2 default: q4_k depth decoder (the stage is weight-bandwidth-bound: -24% depth, ~-6% wall; q8_0 stays available)

Intro-lock ensemble fork (ensemble_prefix_frames): shared master prefix at batch 2, batched decode KV export/replication, shared flow seed under the lock

Carry-free flow probe behind MM3_FLOW_NO_CARRY=1 (listening-rejected: seams audible; kept for hybrid experiments)

Wave-5 recipe: configurable flow chunk hop (flow_chunk_hop_frames) with consistently rederived crops/carry; defaults hop=150 + delta-cache interval=3 (flow -40%, 60s solo 74.1s, K=6 ensemble 46.9s/…

Restore upstream defaults: every optimization stays opt-in