Sync with upstream master-951-f89d9b1

[oobabooga]

Syncs the fork with upstream master-951-f89d9b1 (2026-10-09). The last sync, unslothai#15 on 2026-09-21, brought the tree up to master-889-c678dfe, so this adds 62 upstream commits. Among them are fixes Studio users need:

• the Qwen-Image-2.1 VAE scaling and flow schedule;
• the Z-Image f16 overflow on CUDA;
• Metal mmap buffers;
• the new tile planner;
• a large ggml update (f583f39 -> d25b121).

Merge this with "Create a merge commit", not squash or rebase. unslothai#15 was squashed, so git never recorded it as a merge. Git still thinks the last sync was the 2026-08-07 merge, and a plain git merge re-raised every conflict unslothai#15 had already resolved. The first commit here (e739cd3) is git merge -s ours c678dfe. It records unslothai#15's sync point without changing a file: the tree is identical to master, and c678dfe..master contains only fork files. The real merge (8a48526) therefore used c678dfe as its base. A squash would throw that away again.

What needed more than a textual merge

Seven files conflicted. Beyond those, these upstream changes overlap in meaning with fork code:

• ggml update. All six patches in scripts/unsloth/ggml-patches stopped applying, so the prebuilt pipeline would have failed at its patch step. They are rebased onto d25b121 with git am -3. Each patch keeps its author and message, and applying them to a clean d25b121 reproduces the rebased tree exactly.
  • The long-sequence FlashAttention patch (0002) now carries upstream's new sparse-attention template argument. It keeps using the stock kernel's grid, so it splits the work the same way the stock kernel does.
  • The fused-op patch (0004) adds its three ops next to upstream's new SOL_ATTN and ROPE_APPLY (GGML_OP_COUNT 108).
• Tile planner (fix: correct non-circular tile placement and blending leejet/stable-diffusion.cpp#2088, refactor: unify circular and non-circular tiling leejet/stable-diffusion.cpp#2105). Upstream replaced the planner that Keep VAE tiles overlapping on the out-of-memory retry (no more seam lines), and retry the encode too unslothai/stable-diffusion.cpp#21's sd_tiling_seam_safe_tile_size worked around. The thin case still happens with the new planner: on 62 latents with 32-latent tiles, two tiles overlap by only 2 latents. The guard is reimplemented on top of sd_tiling_plan_axis and is unchanged in intent. It shrinks the tile first; it can widen it to at most 2x; every interior overlap is at least half the target and at least 2 latents.
  • The fork's batched, asynchronous tile merge and plane-parallel blend now sit on upstream's per-edge overlaps. They use the same arithmetic as upstream, element for element.
• Out-of-memory retry (fix: restrict VAE tiling retries to allocation failures leejet/stable-diffusion.cpp#2019, refactor: define VAE tile dimensions in image pixels leejet/stable-diffusion.cpp#2059). Upstream now retries the decode, on allocation failures only, halving absolute tile sizes each time. This replaces Keep VAE tiles overlapping on the out-of-memory retry (no more seam lines), and retry the encode too unslothai/stable-diffusion.cpp#21's relative-size retry. Keep VAE tiles overlapping on the out-of-memory retry (no more seam lines), and retry the encode too unslothai/stable-diffusion.cpp#21's encode retry, which upstream lacks, is ported onto the same function.
  • Because the guard can widen a retry tile, the decode and encode loops stop once a spatial retry no longer shrinks the effective tile. Without that, some sizes would retry forever: on 9 latents a 4-latent tile widens to 6, halves back to 4, and so on.
• Tile sizes are now in image pixels (refactor: define VAE tile dimensions in image pixels leejet/stable-diffusion.cpp#2059), with a 256-pixel default for every VAE. Studio passes only --vae-tiling, so nothing it sends changes meaning. MiniMax-H3 resolves its tiling through upstream's resolve_tiling_params, and SD_H3_VAE_TILE=N still means N latents. The H3 batched decode now handles circular axes the same way as VAE::tiled_compute.
• H3 VAE. Upstream's change to keep H3 VAE weights resident across chunks duplicates SD_H3_VAE_KEEP_RESIDENT, so the fork's decode loop is kept as is.
• close-organization-fork-prs.yml is a new upstream workflow that closes PRs opened from organization-owned forks. It is deleted here: that is upstream's contribution policy, not this repository's.

Behaviour that changes with upstream

• Z-Image output changes on CUDA, ROCm, Vulkan and Metal. Every comparison against pure upstream f89d9b1 that this PR ran is pixel-identical (Z-Image 1024 untiled/tiled and 1008 tiled on CUDA, SD1.5 and Z-Image on Metal), so the change comes from upstream.
• First OOM retry tile: a decode or encode that runs out of memory now first retries with 256-pixel tiles (6x6 at 1008), not Keep VAE tiles overlapping on the out-of-memory retry (no more seam lines), and retry the encode too unslothai/stable-diffusion.cpp#21's half-latent tiles (3x3).
• H3 overlap with --vae-tiling: the H3 video VAE now uses --vae-tile-overlap (default 0.5) instead of a forced 0.25. Studio passes --vae-tiling only under its low-VRAM offload policies.
• Encode tile multiplier removed: image VAE encode no longer doubles the tile size.

Verification

Arms are master 4a98feb vs this branch, and pure upstream f89d9b1 where noted.

CUDA, local (RTX 6000 Ada sm89, RTX 3090 sm86, cuDNN on)

• test-backend-ops, full suite: 16597/16597 on both GPUs.
• H3, 960x544x124, 4 steps:
  • Default mode: master 98.6 / 99.3 s, this branch 98.9 / 95.2 s, upstream 144.5 / 143.5 s. Max mode: master 73.0 s, this branch 72.5 s.
  • Frames are 109 dB from master (77 of 124 identical) and audio is bit-identical. Repeat runs are identical.
• Z-Image:
  • Matches upstream exactly at 1024 untiled/tiled and 1008 tiled.
  • At 496 tiled this branch uses 3x3 tiles where upstream uses 2x2. Mean difference from the untiled image across the 240-256 px tile boundary: 2.5-2.7 here vs 4.3-4.6 upstream (columns) and 0.6 vs 1.4-1.8 (rows).
• VRAM caps on the 3090:
  • Decode retry works (5 GiB free).
  • Encode retry works (img2img, 3.2 GiB free) where pure upstream fails with "failed to encode init image".
• The retry-loop stop rule, tested by linking the real sd_tiling_seam_safe_tile_size into an emulation of the loop: all 2382 cases (dims 4-400, overlap 0.5/0.25/0.1, scale 8/16) finish within 6 attempts.

AMD Strix Halo gfx1151 (DevLab runners)

• Builds on HIP, Linux Vulkan and Windows Vulkan. All op tests pass: FLASH_ATTN_EXT, CPY, MUL_MAT, RMS_NORM, ROPE, the fork's fused ops, ADD, MUL, CONV_2D_DW.

• H3 at 640x384x56:

  Backend	Master	This branch	Output vs master
  HIP	280.8 / 286.9 s	284.3 / 284.1 s	differs; see CPU check below
  Linux Vulkan	142.6 / 146.0 s	137.1 / 134.4 s	video 106.8 dB
  Windows Vulkan (interleaved)	303.0 / 272.1 s	273.3 / 246.2 s	video 106.8 dB, audio bit-identical

• CPU reference check (2 steps, each arm vs its own CPU-backend render):

  • The CPU renders of master and this branch match: audio bit-identical, video 106.8 dB.
  • HIP: this branch is slightly closer to the CPU than master (video 30.96 vs 30.19 dB, audio 5.50 vs 5.30 dB).
  • Linux Vulkan, three seeds: video equal, audio 6.79 / 7.61 / 9.00 dB vs master's 7.16 / 7.31 / 9.61. The sign varies by seed, the GPU-vs-CPU distance itself is only 7-10 dB, and Windows Vulkan audio is bit-identical. I read this as noise.

• HIP OOM retries:

  • img2img encode at 3.2 GiB free: retries with 256-pixel tiles.
  • 1536 untiled decode at 10 and 12 GiB free: retries and finishes (11x11 tiles).

Metal (M1, 8 GB)

• Op tests pass.
• SD1.5 512 untiled, 512 tiled and 504 tiled match pure upstream pixel for pixel. Timings are equal across master, this branch and upstream (sampling about 42.7 s).
• Z-Image matches upstream untiled and at 768 tiled. At 496 tiled this branch uses 3x3 tiles where upstream uses 2x2. Mean difference from the untiled image across the 236-260 px tile boundary: 4.39 vs 5.53 (columns) and 1.58 vs 4.76 (rows).

Prebuilt pipeline legs, replayed from this branch's unsloth-sd-prebuilt.yml on GitHub-hosted runners (signing skipped)

• Every leg builds: macOS arm64 (including the load gate), macOS x86_64, Linux x86_64, Linux aarch64, Linux Vulkan, Windows CPU, Windows Vulkan, and Linux CUDA 12.8.
• The CUDA leg covers 8 architectures with cuDNN on, passes the NEEDED gate, and produces a 1192.9 MiB zip.
• That CUDA zip renders H3 on the Ada with cuDNN 9.27 loaded, and renders Z-Image on both GPUs.

After merging

The prebuilt workflow names a release after the highest master-NNN-sha tag reachable in this repository. Upstream's tags after master-813 were never copied here, so releases would still be called master-813-.... Pushing upstream's tag fixes the name:

git fetch https://github.com/leejet/stable-diffusion.cpp tag master-951-f89d9b1
git push origin master-951-f89d9b1

Not covered

• CUDA runtime on sm80, sm90 and sm120 (it compiles in the leg).
• Discrete AMD GPUs.
• One HIP case got worse: Z-Image 1536 with only about 9 GiB free on the Strix Halo iGPU, where master finished and this branch fails in the DiT. Pure upstream f89d9b1 fails the same way at 9 and 8 GiB free. The new ggml reports the iGPU's real free memory, where the old one reported system memory (about 59 GB), so the DiT now switches to segments and runs out partway. This comes from upstream, not from the merge.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-10T01:55:51.700703Z	d0d47fc	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[oobabooga]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8e412971c2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Version public parameter structs before extending them

Preserve an ABI-safe way to identify the caller's struct size before reading this field. When an application compiled against the previous public header loads the updated shared library, its sd_ctx_params_t ends at sage_attn, but StableDiffusionGGML::init() now unconditionally reads conditioning_cache_size past the caller's allocation; the appended image_preprocess fields similarly affect existing image/video callers. This can interpret unrelated memory as configuration or a rules pointer and crash, so retain a legacy entry point/layout or add explicit size/version negotiation.

AGENTS.md reference: AGENTS.md:L69-L75

Useful? React with 👍 / 👎.

[oobabooga]

Not changing this. The new fields are upstream's (leejet#2028, leejet#2034, leejet#2059); this PR only merges them. The fork has no shared-library consumer to break: the prebuilt bundles ship statically linked sd-cli and sd-server (no libstable-diffusion), and Studio runs those executables, so nothing loads a newer library against an older header. Adding a size/version scheme to upstream's structs here would also diverge from upstream's API on every later sync.

[chatgpt-codex-connector]

Keep the existing conversion entry point ABI

Add a new LoRA-aware symbol or retain a wrapper with the old convert_with_components signature instead of adding arguments to this exported C function. An existing dynamically linked caller still resolves the same unmangled symbol but supplies only the former 11 arguments, so the implementation reads indeterminate values for loras and lora_count; a non-null garbage pointer or positive garbage count can then be dereferenced while logging or applying LoRAs.

AGENTS.md reference: AGENTS.md:L69-L75

Useful? React with 👍 / 👎.

[oobabooga]

Not changing this. The two LoRA arguments come from upstream (leejet#2079, merging LoRA weights on conversion); this PR only merges them. The fork distributes statically linked sd-cli and sd-server only, so no existing binary resolves convert_with_components from an upgraded shared library. A fork-only wrapper and renamed entry point would diverge from upstream's API on every later sync.