3 Commits

Author	SHA1	Message	Date
cmoyates	f73aabf5af	chore: default bf16+fused decode on, add benchmark results to plan ... load_model() now defaults to dtype=bf16, fused_decode=True. Both are free (zero parity regression, bit-exact fused path). Backbone/sigmoid stay fp32. Plan updated with benchmark results: tiled+GC = 12x peak memory reduction at 2048x2048 (27.6GB → 2.3GB), all acceptance criteria met. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>	2026-03-08 21:27:56 -02:30
cmoyates	b3b7df271d	feat: MLX memory optimizations — bf16, fused decode, deterministic GC ... Phase 1: selective bf16 mixed precision (backbone fp32, decoders bf16, sigmoid always fp32). dtype param on GreenFormer, all outputs guaranteed fp32. Phase 2: FusedDecoderPair batches 3 upsample ops by concatenating alpha+fg projections along channel axis. Bit-exact parity with unfused path. Phase 3: deterministic GC pipeline in tile loop (del + gc.collect + mx.clear_cache) and engine.py intermediate tensor cleanup. 7 new tests: bf16 dtype/range/nan, fused-vs-unfused parity, fp32 default unchanged. 80 passed, 0 failures. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>	2026-03-08 21:21:01 -02:30
cmoyates	830e88f050	docs: add brainstorm + plan for MLX memory optimizations ... bf16 mixed precision, batched decoder upsample, deterministic GC pipeline. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>	2026-03-08 21:12:59 -02:30