docs: the MLX drafting retention reproduced and bounded; gemma4 resize/#28954 record; parked upstream draft by glennneuber · Pull Request #311 · MaxusAI/ollama

task: gemma4's bicubic resize entered with the 0.33.1 fold (#27594), unflagged; not in this fold's delta

glennneuber

and

claude

committed

45aa955

task: gemma4 veto closed (repeat on b10760 identical); llama.cpp #28954 — fork immune to the abort, split-decode cost noted; register row

glennneuber

and

claude

committed

9cd1825

task: the drafting retention reproduced (image + stop + speculation), exclusions by instrumented runs, device vs MLX accounting

glennneuber

and

claude

committed

333a6a4

task: two components — A (MLX-tracked, speculation + image + stop) and B (CUDA-internal, shape variety); the knob removes A, not B

glennneuber

and

claude

committed

7b566c6

docs: parked upstream draft — MLX speculation + image + stop retention (ollama), CUDA-internal shape-variety growth (mlx)

glennneuber

and

claude

committed

ecbe9ce

register: the drafting knob as memory mitigation, with what retires it

glennneuber

and

claude

committed

85cb506

task: component B plateaus (bounded, ~4 GiB CUDA-internal at steady state); A is the only unbounded growth

glennneuber

and

claude

committed

8bd5812

task: gemma4 image chunk vs. generation batch, measured — no contract cost, 1.5–1.9× prefill in one piece, 31b 9 px tier moves

glennneuber

and

claude

committed

c15a294

task: #313's pinned budget ladder on discrete CUDA — no cliff at 1120, geometry identical to the ROCm host

glennneuber

and

claude

committed

b7a05ba

Navigation Menu

Repository navigation

docs: the MLX drafting retention reproduced and bounded; gemma4 resize/#28954 record; parked upstream draft - #311

docs: the MLX drafting retention reproduced and bounded; gemma4 resize/#28954 record; parked upstream draft#311
glennneuber merged 9 commits into
mainfrom
docs/leak-repro-0341

Commits on Sep 18, 2026

task: gemma4's bicubic resize entered with the 0.33.1 fold (#27594), unflagged; not in this fold's delta

task: gemma4 veto closed (repeat on b10760 identical); llama.cpp #28954 — fork immune to the abort, split-decode cost noted; register row

task: the drafting retention reproduced (image + stop + speculation), exclusions by instrumented runs, device vs MLX accounting

task: two components — A (MLX-tracked, speculation + image + stop) and B (CUDA-internal, shape variety); the knob removes A, not B

docs: parked upstream draft — MLX speculation + image + stop retention (ollama), CUDA-internal shape-variety growth (mlx)

register: the drafting knob as memory mitigation, with what retires it

task: component B plateaus (bounded, ~4 GiB CUDA-internal at steady state); A is the only unbounded growth

task: gemma4 image chunk vs. generation batch, measured — no contract cost, 1.5–1.9× prefill in one piece, 31b 9 px tier moves

task: #313's pinned budget ladder on discrete CUDA — no cliff at 1120, geometry identical to the ROCm host