Skip to content

preflight: the Metal v0.40.2-dynres.1 release build passes mlx-metal-0-40-2, and is deployed #486

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
glennneuber merged 3 commits into from
Oct 13, 2026
Merged

preflight: the Metal v0.40.2-dynres.1 release build passes mlx-metal-0-40-2, and is deployed #486

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9024927
preflight: mlx-metal-0-40-2 admits the v0.40.2-dynres.N point tags
glennneuber Oct 13, 2026
40b9ef8
preflight: the mlx-metal-0-40-2 test admits the point-tag stamp
glennneuber Oct 13, 2026
40df1ad
preflight: the Metal v0.40.2-dynres.1 release build passes mlx-metal-…
glennneuber Oct 13, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 5 additions & 1 deletion docs/maxusai/vision-suite/BINARIES.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,9 @@ A host can be rebuilt from it; a binary sitting only in `/tmp` cannot.
| `0.33.2-maxusai-2b95b4a5` | [`2b95b4a5`](https://github.com/MaxusAI/ollama/commit/2b95b4a5) | llama.cpp **b10630** | c793734e pin | `cd6033d6e72a430e…` | served :11435 2026-08-30 → 09-18; rollback target (BOTH halves archived); the v0.33.2 fold (#232, ADR 0033 xgrammar). Provenance for the `mlx-metal-0-33-2` profile measurement; preflight PASS 2026-08-30, goldens NOT recalibrated at this MLX pin. **Rolling back to it reinstates the pre-#3912 `fp_qmm_t` defect** on the gemma4 26b/31b vision `down_proj` (K = 4304): OCR-neutral on OCRBench, arithmetically wrong (ADR 0037) |
| `0.34.0-maxusai-8a7ba949` | [`8a7ba949`](https://github.com/MaxusAI/ollama/commit/8a7ba949) | llama.cpp **b10864** | d9add9d1 pin | `867ee5d6063d8f77…` | the v0.34.0 + v0.34.1 folds (#297, #302). The same build as `0.34.1-dynres-0-g8a7ba94` (ADR 0032, 2026-09-19 amendment): built 49 minutes before the fold tag, under the retired native stamp. Provenance for the `mlx-metal-0-34-0` profile (#306); preflight PASS 2026-09-18 (19/12); whitespace bound (#301) confirmed behaviourally on real documents across the MLX pin move. **rollback target** (both halves archived) — served `:11435` from 2026-09-18 13:17 to 2026-09-28 14:07, verified by version, a live MLX load reporting `0.32.2-61-gd9add9d`, and the dlopen path resolving the repo payload. Native gates pass: `mlxrunner` and all subpackages, and vision golden parity on 31b/26b/12b (max element delta 0.109 / 0.051 / 0.063) — the first build whose goldens were CHECKED at its MLX pin. **Promoted 2026-09-19** with the MLX #3912 kernel fix kept (ADR 0037): OCRBench v1, all 1000 items, `31b-nvfp4` 835/1000 against 0.33.2's 833/1000, exact McNemar p = 0.80 (`bench-runs/ocrbench-v1-1000-gemma4-31b-nvfp4-0340-vs-0332.json`) |
| `0.34.4-dynres-0-gb43ee8e` | [`b43ee8e3`](https://github.com/MaxusAI/ollama/commit/b43ee8e3) | llama.cpp **b11081** | 59d600b5 pin | `f343fea9ea7bc66e…` | the `v0.34.4-dynres` tag (#375), both halves built at the tag in a detached worktree. Not a profile's provenance: `mlx-metal-0-34-4` (#384) was measured on the fold's build, which has the same pins. Served `:11435` 2026-09-28 14:07 → 2026-10-01 23:45, and the rollback target for `0.35.0-dynres-0-g043f441`, which runs on this payload. Deployed with `OLLAMA_FORMAT_TWO_PASS=1` (ADR 0045) and `OLLAMA_KV_CACHE_TYPE=f16` (ADR 0043) added to the launchd environment, verified by version, the startup config, a live MLX load reporting `0.32.2-65-g59d600b`, and llama-server resolving the repo payload. Preflight PASS 23/12 on a `:11437` stage and again on production itself (`preflight/runs/preflight-mlx-metal-0344-prod-gb43ee8e.json`). Native gates pass: `go test ./mlxrunner/... -p 1 -count=1` with `OLLAMA_VISION_E2E=1`, 459 tests and 347 subtests passed, 0 failed, 1 skipped (`TestGlobalScaleSurvivesStorageBitExactly`), in the 18 of 24 packages that have tests; vision golden parity per model, each named by its manifest digest in the Apple Silicon host's store (ADR 0038): 31b-nvfp4 `637cc0ff1570` and 26b-nvfp4 `c8656f50f0a6`, both the 4-bit-tower artifacts, and 12b-nvfp4 `117d0d84cf2a`, at max sampled element delta 0.0898 / 0.0508 / 0.0625, and 26b-mlx-bf16 `fa29b1495bbb` at 0.0625. `go test ./mlx/... -p 1 -count=1`: 59 tests and 38 subtests passed, 0 failed, 1 skipped (`TestControlsDoNotStopWorker/replay_skip`), with `TestMulGatherQMMGlobalScale` passing on Metal (open item 4 of the [v0.34.4 fold record](../tasks/upstream-sync-0.34.4.md#open-items)). A think+format request on production drafts in pass one only, and its first GGUF load shows `--cache-type-k f16 --cache-type-v f16` and an f16 KV allocation. On 2026-09-29 the archived pairing, on a scratch server in production's environment, gave the fold's think-off answers byte for byte (189 of 189) and scored OCRBench v1 835/1000 against 0.34.0's 835/1000, exact McNemar p = 1.000 (`bench-runs/ocrbench-v1-1000-gemma4-31b-nvfp4-0344-vs-0340.json`). |
| `0.35.0-dynres-0-g043f441` | [`043f441a`](https://github.com/MaxusAI/ollama/commit/043f441a) | llama.cpp **b11081** | 59d600b5 pin | `7ea4fd966cc100f0…` | the `v0.35.0-dynres` tag (#427), a Go-only build at the tag in a detached worktree (`go build` with `build-macos.sh`'s flags and stamp). It runs on `0.34.4-dynres-0-gb43ee8e`'s payload, unchanged, so it has no payload archive of its own: the fold moves no native input. **current** — deployed to `:11435` 2026-10-01 23:45, binary only; the launchd plist is unchanged. Verified by version, the startup config, a live MLX load reporting `0.32.2-65-g59d600b` and a two-pass smoke. Preflight PASS 23/16 on a `:11437` stage and again on production itself (`preflight/runs/preflight-mlx-metal-0350-release-g043f441.json`, `-prod-g043f441.json`), resolved by `mlx-metal-0-34-4`, which now admits 0.35.0 (#437). Native gates pass: `go test ./mlxrunner/... -p 1 -count=1` with `OLLAMA_VISION_E2E=1`, 806 passed and 1 skipped. Rollback: the 0.34.4 binary; the payload is the same |
| `0.35.0-dynres-0-g043f441` | [`043f441a`](https://github.com/MaxusAI/ollama/commit/043f441a) | llama.cpp **b11081** | 59d600b5 pin | `7ea4fd966cc100f0…` | the `v0.35.0-dynres` tag (#427), a Go-only build at the tag in a detached worktree (`go build` with `build-macos.sh`'s flags and stamp). It runs on `0.34.4-dynres-0-gb43ee8e`'s payload, unchanged, so it has no payload archive of its own: the fold moves no native input. served `:11435` 2026-10-01 23:45 → 2026-10-11 09:29, binary only; the launchd plist is unchanged. Verified by version, the startup config, a live MLX load reporting `0.32.2-65-g59d600b` and a two-pass smoke. Preflight PASS 23/16 on a `:11437` stage and again on production itself (`preflight/runs/preflight-mlx-metal-0350-release-g043f441.json`, `-prod-g043f441.json`), resolved by `mlx-metal-0-34-4`, which now admits 0.35.0 (#437). Native gates pass: `go test ./mlxrunner/... -p 1 -count=1` with `OLLAMA_VISION_E2E=1`, 806 passed and 1 skipped. Rollback: the 0.34.4 binary; the payload is the same |
| `0.40.2-dynres-0-g5623192` | [`5623192c`](https://github.com/MaxusAI/ollama/commit/5623192c) | llama.cpp **b11351** (`631109b34`) | a59cc231 pin | `85597909da3be5ae…` | the `v0.40.2-dynres` tag (#460), both halves built at the tag in a detached worktree; payload archived as `payload-0.40.2-dynres-0-g5623192`. Served `:11435` 2026-10-11 09:29 → 2026-10-13 22:01. Preflight PASS 23/16 on `mlx-metal-0-40-2`, stage and production. **Rollback target** for `0.40.2-dynres.1-0-g3960244`: restore this binary and its payload, same plist |
| `0.40.2-dynres.1-0-g3960244` | [`39602449`](https://github.com/MaxusAI/ollama/commit/39602449) | llama.cpp **b11351** (`631109b34`) | a59cc231 pin | `023b4082a42240cf…` | the `v0.40.2-dynres.1` point tag, both halves built at the tag in a detached worktree; payload archived as `payload-0.40.2-dynres.1-0-g3960244`. Go-only on Metal: the tag's one native change (#462's 903) is `ggml-cuda/mmq.cu`; 14 of 19 libraries are byte-identical to 0.40.2's and 5 differ only by the build path in `__FILE__` strings. **current** — deployed to `:11435` 2026-10-13 22:01 by a stopped swap of both halves; the launchd plist is unchanged. Native gate `go test ./mlx/... ./mlxrunner/... -p 1` with `OLLAMA_VISION_E2E=1`: 1296 passed, 0 failed, 8 skipped. Preflight PASS 23/16 on `mlx-metal-0-40-2`, stage and production (`preflight/runs/preflight-mlx-metal-0402-dynres1-release-g3960244.json`, `-prod-g3960244.json`), plus the deploy smoke (two-pass MLX and converted GGUF, GGUF vision, System One 400) |

Full checksums:

Expand All @@ -34,6 +36,8 @@ cd6033d6e72a430e0b96170adb3d7408c908edfdae392654faeaae46f31e9ee2 ~/.ollama/bina
867ee5d6063d8f77be6d8278c51d18f49892f00aa80c38c51630227aa9f68a57 ~/.ollama/binaries/ollama-0.34.0-maxusai-8a7ba949
f343fea9ea7bc66e7c85fd47500fb2b6e5cc45aacc2f0dad49f7adac06940d83 ~/.ollama/binaries/ollama-0.34.4-dynres-0-gb43ee8e
7ea4fd966cc100f085e7c5d1e7e68096e538fb989af39e03a7e0204116eae249 ~/.ollama/binaries/ollama-0.35.0-dynres-0-g043f441
85597909da3be5ae8f3d4a203c80959bd9c1743ce834fa3c69b1de4f343ee434 ~/.ollama/binaries/ollama-0.40.2-dynres-0-g5623192
023b4082a42240cf237f5f77e078d3bad1f99b0c573592fe46915a875d446401 ~/.ollama/binaries/ollama-0.40.2-dynres.1-0-g3960244
```

## What b10353 is the provenance for
Expand Down
21 changes: 18 additions & 3 deletions docs/maxusai/vision-suite/preflight/expectations.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -779,9 +779,11 @@ platform = "mlx-metal"
# (631109b34), with MLX-C unchanged at ebc88f10. A new profile by the ADR 0011 rule
# that created every predecessor: a payload move gets its own profile even when
# the numbers turn out identical, because "identical" is a measurement.
version_pattern = '^0\.40\.2-dynres-\d+-g[0-9a-f]{7,40}$'
# Admits the v0.40.2-dynres tag's lineage and nothing else: not -dirty, not a .N
# point tag, not the legacy 0.34.x-maxusai-<sha> shape. The fold's interim builds
version_pattern = '^0\.40\.2-dynres(\.\d+)?-\d+-g[0-9a-f]{7,40}$'
# Admits the v0.40.2-dynres tag's lineage and its .N point tags (v0.40.2-dynres.1
# moves Go only on Metal; its one native change, #462's 903, is in ggml-cuda/mmq.cu,
# which this build does not compile), and nothing else: not -dirty, not the legacy
# 0.34.x-maxusai-<sha> shape. The fold's interim builds
# stamp 0.35.0-dynres-N (mlx-metal-0-34-4's lineage, whose pins they fail); the
# Metal host validated the interim build with a campaign copy of this file whose
# only change pinned this pattern to that one stamp (#460).
Expand Down Expand Up @@ -819,6 +821,19 @@ APFS clone of production's (#460). The four arches below are MLX safetensors and
are not converted; the clone does not change what they load.
aspect_ladder and the think_format probes are carried from 0-34-4 unchanged and
run live on every preflight.

THE v0.40.2-dynres.1 RELEASE BUILD PASSES AND IS DEPLOYED, 2026-10-13:
0.40.2-dynres.1-0-g3960244, built at the point tag in a detached worktree, VERDICT
PASS, PASS=23 SKIP=16 on a :11437 stage against a cloned store
(runs/preflight-mlx-metal-0402-dynres1-release-g3960244.json), and again on
production :11435 after the stopped swap (-prod-g3960244.json). 36 of 39 rows
match the v0.40.2 release run byte for byte; the other three are the stamp, the
stage path and gemma4_unified think_format's token count (242 stage, 739 prod),
which repeats as 496/260/260 on production and 242/260/261 on the stage: the
first request after a load drafts to a different depth, then both settle at 260.
The payload is the v0.40.2 one rebuilt: 14 of 19 libraries byte-identical, and
the other 5 differ only by the build worktree's path in __FILE__ strings (same
xgrammar 82505d0d source, same MLX a59cc231 and llama.cpp 631109b34 pins).
"""

[profiles.metal-0-32-14]
Expand Down
Loading