Qualify post-v0.2 upstream changes with context ABI and FP8 CPU safety checks

Current delivery snapshot — 2026-11-04

Public core 0.17.5 / companions 0.17.3 retain v0.3.0-1. Main pins v0.3.0-2. Latest published v0.3.0-3, source 5fde7dd13544385fdacd9fe31062064496052024, retains upstream be486498eadc29dc3cd234d944c2180ed42b1415 and five carries including #41's bounded ZIP lookup. Core draft#1739 already stages pins/checksums/SwiftPM; no new wrapper/binding/public API is needed for this carry. Final integration/publication remain pending. Raw master-1047 is 984eb0686cbf5b7fc4f0063d9e0b1fb38d5c4f82, 30 commits ahead. Current evidence and gates. Bounded original-container CPU acceptance does not clear full-VAE, hardware or family limits.

Original investigation and retained acceptance

The dated snapshots below are historical. This current snapshot supersedes their pin/publication facts, not their unresolved acceptance criteria.

Triage

P2 / enhancement / owner qualification. The FP8 report is upstream-confirmed; exposure of our pinned artifact is NOT established.

Current layers — 2026-10-21

Published owner v0.2.0-4, owner c9922d960c79f7ba7381161ca6784f3b7c1e2fce, retains upstream 3f8527a46c54ecf4cb4ed6003da8e8982283c73c. The manifest bytes verify SHA-256 58f2aa7f89cdffd9f28097b47eb58750f7c24f94bf1f2fc92118a9bbe3375d4b. Public core 0.15.5, current main and SD companion 0.15.2 retain this artifact. Metal: a failed GPU command buffer still returns an image as success #19's Metal fail-loud guard is delivered, not physical-watchdog prevention.
Raw main/release master-982-0518b34 is 0518b34374761cb82aeda5ad3c2f0686fdad2cec: 53 commits beyond bundled source, five since the previous scan. Exact-head CI succeeds. No newer owner artifact or automatic pin bump.
New raw changes are runtime-quantized CLI LoRA auto-application #2162, server sigma defaults #2172, server reference-edit routing/new capability export #2167, and server recipes/loading/job timestamps #2176. These are not new image families or automatically public Dart LoRA/sigma/edit features.
The exact header delta includes batched_cfg, two context-mode pointers, preview/ref-conditioning exports and two extra convert_with_components parameters. Audit sizes/offsets/defaults/signatures and guarded wrappers, then regenerate bindings/capabilities; a pin-only update is unsafe.
New-family/derivative coverage remains raw-only: Iris-3B and Z-Image L2P use pixel-space/no-VAE recipes; FLUX.3 (#1130) needs its FLUX.2 VAE/Qwen3-VL components; MOSS (#1166) additionally needs external-tokenizer transport. Existing owner umbrella and consumer tickets remain adequate. TAEQI is a component, MoSR (#1069) an upscaler, Qwen Turbo (#1070) a schedule variant; no new family landed in today's delta.
FP8 #2139 is fixed by #2140, first master-966; conv throughput #2137 by #2141, first master-967. Pinned exposure/consumer delivery remain unproven. New #2173 reports an RDNA3 probe false-negative with identical forced/default pixels and similar timing, not proven recurrence of #2137. Metal #2149, Krea2 AMD #2154, video #1976 and ESRGAN #2163 remain open; #2166 and draft #2171 are still unmerged.
New open video reference transport #2177 and StarNodes INT8 metadata #2178 are proposals, not delivered support. BRIA Fibo2 #2179 is only a request. Long-running video/upscaler ownership wrappers remain prerequisites (#1220/#1069); software Vulkan fallback does not qualify hardware GPUs.
Current-artifact Fix legacy PyTorch checkpoint scanning above 2 GiB #23 remains confirmed: draft #25 adds regression tests with an expected failure, not a source fix. Investigate large PyTorch ZIP checkpoint load timeouts #24 remains open; #922 reports later non-reproduction, which does not reconcile the original large-ZIP timeout or prove a fix. Core #1218 retains the format qualification boundary.

Evidence and benefit

Our v0.2.0-3 manifest pins 3f8527a46c54ecf4cb4ed6003da8e8982283c73c (ggml 89c4413f). Raw main on 2026-10-15 was 120e3e875c5501c1babb50c0966122cb569d12d4 (ggml d25b121c).

Flash Attention and convolution mode updates append flash_attn_mode and conv_direct_mode pointers to sd_ctx_params_t: feat: add per-module automatic flash attention selection leejet/stable-diffusion.cpp#2132 and feat: add automatic direct convolution selection leejet/stable-diffusion.cpp#2134
Wrapped iGPU memory reclamation: fix: reclaim before accepting wrapped Vulkan iGPU memory estimates leejet/stable-diffusion.cpp#2124 (60423ab5; first raw release master-961-60423ab). Broader upstream #2022 remains open.
FP8 E4M3 CPU report #2139 now has merged repair #2140: 2891378bb17dbed110b391f9aab47d241413153d, first raw release master-966-2891378. CPU matmul falls back from FP8; pinned-artifact exposure still requires reproduction.
Vulkan auto-convolution slowdown #2137 has merged repair #2141: e47426c04460827e698b6f88ec6431a2007d5c27, first master-967-e47426c. Qualify end-to-end output and timing, not only the probe.
Historical 2026-10-16 raw head/release was 8cd731e7b4667eb8f40a153315b9a7125d09a58d / master-968-8cd731e: FLUX.3 Image Klein and Klein Base join Iris-3B and Z-Image L2P. MoSR is an upscaler, TAEQI2.1 an autoencoder, and Qwen Image 2.1 Turbo a schedule variant. None is added by unchanged published v0.2.0-3. Consumer recipes are tracked in llamadart#1130; #1069/#1070 in llamadart retain their separate scope.

New raw evidence — 2026-10-17

Current raw main/release: f0aac524d3f29140e60ba78f3cfb16d6acea87db / master-973-f0aac52. Owner, current main and published core 0.14.0 still pin v0.2.0-3 / 3f8527a46c54ecf4cb4ed6003da8e8982283c73c.

Windows UTF-8 model/CLI paths are now merged in #2144 (8c442fb0, first master-969) / #2145 (2f016e8e, first master-970). They remain absent from the owner artifact; test model and auxiliary/output paths separately.
#2148, 68285df3, adds MOSS-Image Base/SFT, Qwen3.5 encoder and external tokenizer requirements. Consumer llamadart#1166 owns distinct tokenizer/recipe work. Iris, Z-Image L2P and FLUX.3 remain raw-only too.
#2149 independently isolates Metal GROUP_NORM leaving batch items after the first unwritten in official master-968/969 libraries (CPU controls pass); batched CFG then returns white images. This is an upstream-reported candidate correctness gate, not proof our older pinned artifact is affected, and not the same mechanism as owner Metal: a failed GPU command buffer still returns an image as success #19's command-buffer failure.
#2147 refreshes ggml/removes conv1d workarounds. Latest raw CI succeeds after the earlier MUSA container failure; successful packaging is not model correctness evidence.

Additional acceptance: model-free asymmetric GROUP_NORM batch-1/batch-2 sentinels plus real-image batched/unbatched CFG controls on the exact candidate and pinned owner Metal library; fail closed or retain a supported unbatched route if affected. Cover MOSS tokenizer/FLUX.2 VAE loading (#2150 report lacks full logs) and Windows non-ASCII paths. Preserve SD#19 at P1 independent of mitigations or missing hardware.

Acceptance

Select an exact candidate; audit wrapper exports and struct sizes/offsets/default initialization. An appended pointer changes FFI layout; a pin-only update with old bindings is not acceptable.
Reproduce or explicitly bound FP8 CPU exposure on the pinned artifact and candidate. Add positive and negative controls ensuring fallback or an actionable refusal, never a process crash. Do not label current llamadart affected without evidence.
Preserve explicit on/off behavior when adopting auto modes; qualify CPU/Metal/Vulkan, mixed-device decisions, memory pressure, logs, cancellation and error-state recovery. Retain existing iOS direct-VAE refusal unless independently fixed.
Validate representative new families/components and reject unsupported weight variants; link consumer API work rather than claiming native support equals Dart support.
Produce digest/provenance-verified platform bundles, exports, Apple XCFramework minimum-OS/privacy/signing checks, then coordinate generated bindings, pins, request/capability mapping, docs and examples.
Keep existing owner Metal: a failed GPU command buffer still returns an image as success #19 and upstream video/VAE/HIP problems separate unless the exact candidate demonstrably fixes them.

Boundaries

Tracking/qualification only. This issue does not authorize publication, workflow dispatch, paid hardware runs, downstream pin changes, or weakening supported-channel policy. No PR-caused regression is alleged without a causal reproduction. Preserve immutable historical artifacts and record unavailable models/hardware explicitly; CPU/software fallback is not hardware GPU qualification.

Consumer API dependencies: llamadart#1069 (MoSR/upscaling) and llamadart#1070 (Turbo sigma configuration). Existing owner #19 is triaged separately as P1 false-success correctness; upstream-update preparation does not silently close it.

Raw advanced to master-973-f0aac52, while published owner/core 0.14.0 image pin stays v0.2.0-3. Added merged Windows UTF-8 fixes, MOSS-Image / consumer #1166, and the exact official-library Metal GROUP_NORM batch defect #2149 as a candidate gate. Pinned exposure is unproven; owner #19 remains a separate P1. Raw CI 39009185406 succeeds; the prior failure was the MUSA container job, not evidence that every backend failed. No artifact, binding or public model-support upgrade is claimed.

2026-10-18 candidate delta: raw main is d55cc0f75f9f9c71af2a7b2fd6cf3c6d10420151; latest raw release remains master-973-f0aac52, owner/core image pin remains v0.2.0-3 (3f8527a46c54). #2160 adds MOSS split-QKV/host loading (consumer #1166); #2159 adds Diffusers MiniMax-H3 LoRA mappings/alpha semantics, not a new model family or current Dart video feature. New #2154 reports Krea2 AMD-dGPU OOM around the previously tracked memory-reclamation change; follow-up implicates FA/conv-mode interactions too, so causality is not settled. Add exact pinned/candidate controls with FA/conv on/off and memory headroom before treating the iGPU repair as universally safe. No pinned exposure is established; Metal #19 remains P1 and #2149 remains separate. Raw #2150 closed by installation correction, not a proven VAE patch.

2026-10-19 delivery/candidate delta: owner v0.2.0-4 publishes #21's Metal guard but retains upstream 3f8527a4; it is not this issue's raw-source upgrade. Public core 0.15.1 still pins v0.2.0-3, while merged #1205/#1208 prepare consumer delivery. Raw master-975-95b3090 now contains MOSS split-QKV loading. Added a dated current-layers block; preserved all FP8/context-ABI, GROUP_NORM, memory and model/device gates. No new runtime qualification claim.

Delivery reconciliation: core 0.15.4 and SD companion 0.15.2 now consume v0.2.0-4; #19's owner guard is consumer-delivered, not watchdog prevention. Corrected the manifest URL. Raw source/release is unchanged; #2166 host-pinned memory work and #2171 ESRGAN controls remain unmerged. Current-artifact PyTorch defects are separately owned by #23/#24, not resolved by this upgrade umbrella.

2026-10-21: refreshed current layers to master-982-0518b34, 53 commits beyond unchanged owner v0.2.0-4. Recorded the new ref-conditioning export/conversion signature boundary, server-versus-Dart distinctions and new RDNA3 probe watch. Published core 0.15.5 remains on the same owner source. Existing FP8, GROUP_NORM, memory, family/component, wrapper-lifecycle, Apple-manifest and real-device acceptance is preserved; no model/GPU qualification or pin-only adoption is claimed.

2026-10-22 raw/adoption update — retain P2 and existing qualification gates.

Published v0.2.0-4 / owner c9922d960c79f7ba7381161ca6784f3b7c1e2fce still bundles 3f8527a46c54ecf4cb4ed6003da8e8982283c73c; public core 0.15.6 and SD companion 0.15.2 retain it. Manifest bytes verify 58f2aa7f89cdffd9f28097b47eb58750f7c24f94bf1f2fc92118a9bbe3375d4b; its 25 artifact sizes/digests match GitHub metadata and SHA256SUMS. This is metadata verification, not a new device run.

Latest raw master-987-a77d362, a77d3623180b2bc1886f3dd04c0aa04156a987ef, is 58 commits beyond bundled source/five beyond yesterday; exact-head CI passes.

Material delta	First containing raw release / downstream boundary
Fibo 2.0 Dev / Edit Turbo #2180, `2e2ba3faec6717378e74d92fc5da867a0e13078b`	master-985; #2179 is now closed-supported. One family with variants, raw-only. Recipe: transformer + FLUX.2 VAE + Qwen3-VL4B, built-in tokenizer, /16 dimensions. Existing roles may support a qualified T2I slice; editing needs reference/vision-projector transport (core#1222), and indexes/shards need core#1219. No native Fibo masks/glyph rendering/VGL expansion is delivered.
Krea2 memory-capacity #2183, `a77d362`	master-987; actual repair closes #2154: recheck fits() even when eviction candidates do not grow, without redundant releases. Test affected candidate/pin and memory-pressure/output/latency controls; not universal GPU evidence.
Server ADetailer #2164, `845646b0`	master-983; orchestration, not removal of core#1229's detector/guarded-lifecycle prerequisites.
Remaining-time #2076, `939895f0`	master-984; terminal formatting, not public Dart ETA/callback delivery.
RPC discovery #2182, `b15c8aac`	master-986; new sd_add_rpc_servers export joins the candidate audit; no public Dart RPC placement support.

Corrected watch interpretation: #2173 is closed without a patch after establishing that forced Conv3D did not exercise Qwen Image's Conv2D path; no image regression proved there. #2178 is closed unmerged/rejected: use current per-module .comfy_quant metadata/migrate legacy weights, not alleged StarNodes compatibility. Metal GROUP_NORM #2149 remains open; candidate default batched_cfg is still true and needs repair or safe exclusion. #2166 remains an unmerged MiniMax-H3 branch despite new author success; it does not repair all Wan/LTX/video paths. #1976/#2163 remain open, #2171 draft and #2177 unmerged.

All five family/derivative lanes since pin—Z-Image L2P, Iris-3B, FLUX.3, MOSS and Fibo—remain absent from the owner artifact/public qualified Dart subset. Keep components/variants distinct (TAEQI, MoSR, Qwen Turbo). Preserve 320→336 context/conversion ABI/default/export audits, guarded wrappers, patch-retirement evidence, Apple manifests/SwiftPM/digests, affected-family CPU and actual GPU/lifecycle/version-skew validation. #23's thirteen span narrowings remain in raw; #25 is expected-failure tests only; #24 remains causally unresolved. No new ticket, pin bump, retirement, publication or broad model/platform claim.

Draft #27 moves to master-982-0518b34 at head 50230f7. Metal builds run CFG unbatched through sd_dart_new_sd_ctx (leejet/stable-diffusion.cpp#2149 reproduced on the candidate ggml; guarded output matches unbatched byte for byte). Consumer ABI: sd_ctx_params_t 320→336 bytes, convert_with_components +2 args; bindings must be regenerated. No carry retired. Unqualified platforms and remaining gates are in the PR body.

2026-10-23 delivery and raw correctness delta — retain P2 qualification.

Layer	Verified state
Published owner	v0.3.0, owner `992d3f70e77137243c833f5c2d9d0f9bb4664d0d`, upstream master-982 `0518b34374761cb82aeda5ad3c2f0686fdad2cec` plus pickle_line_span.patch. Manifest SHA-256 `dc79dec2a890dd761d3643f657a6df1bb018ded67a192236195d56459cca6544`; all 25 declared artifact digests/sizes agree with release metadata, and the actual Apple ZIP digest agrees with SwiftPM. No CUDA bundle.
Consumer source / public package	Main `5795dc10` includes #1321: regenerated 336-byte context/conversion ABI, explicit mode defaults and batched_cfg=false. Public core 0.15.6/SD companion 0.15.2 still pin v0.2.0-4/`3f8527a46c54ecf4cb4ed6003da8e8982283c73c`; pending 0.16 prep #1334 is not publication.
Raw upstream	master-995-fd21ecf, `fd21ecfceef2665d2c5f2db8536e3202d864fd30`: +8 since yesterday, +13 over owner source, +66 over public pin; CI passes.

Publication advances FP8 #2140/master-966, convolution #2141/master-967, Windows Unicode paths and Iris-3B/Z-Image L2P/FLUX.3/MOSS source support into the owner artifact. It does not qualify every family/platform in Dart. Fibo (#2180/master-985) and Krea2 capacity repair (#2183/master-987) remain raw-only. MoSR and Turbo are an upscaler/schedule variant, not new base families.

New raw repairs/watches:

#2181, first master-992-39bc9d0, fixes #1976 by excluding small host-visible/device-local BAR heaps, including fallback allocation. It is absent from v0.3.0. Alternative #2166 closed unmerged; do not attribute its host-accounting proposal or general video qualification to this fix.
#2185, first master-990, rejects unsupported Z-Image reference conditioning rather than crashing; it does not add editing support.
FLUX.3 #2192: author retracts coopmat diagnosis. On raw995, Vulkan VAE direct/im2col paths corrupt output; CPU VAE/tiled controls reported clean. Qwen Image 2.1 #2193 reports grid artifacts including Turbo. Exact owner exposure/causes remain unproved. Add affected-model/backend/default-mode negatives to existing core #1130/#1070 acceptance, not duplicate tickets.
#2189 (master-993) enforces MiniT2I 512²; #2190 (master-994) is server cancellation; #2191 (master-995) changes LLM image placeholders. None automatically delivers new Dart families/cancellation/reference APIs.

#23 is fixed by owner #26 and published v0.3.0, not upstream-fixed: raw995 retains all thirteen narrowing casts. #24 ZIP timeout remains causally unresolved. Metal #2149 remains open; wrapper/main safe exclusions are not an upstream repair or retirement. Preserve #19's fail-loud guard and physical-watchdog boundaries.

Owner build/release and Validate succeed. Apple privacy/dependency/linkage checks and metadata verification are not new device or hardware-GPU qualification. No new artifact/pin/workflow action is requested by this tracking update.

2026-10-24 delivery/raw ABI delta — retain P2 qualification.

Public core0.16.0 and SD companion0.16.0 now actually consume owner v0.3.0, 992d3f70e77137243c833f5c2d9d0f9bb4664d0d, upstream 0518b34374761cb82aeda5ad3c2f0686fdad2cec plus pickle_line_span.patch. Core archive SHA256 2c1bd4be3876cc3ee6f83b9ca39757eb040fd6114f3e9c5dd294bd101e0fd8e5; actual pins and companion SwiftPM v0.3.0/Apple digest verified. Main97d366da has the same SD runtime.

Raw master-1000-95ff762, 95ff76230889898e1bf40058ba272a3699b50dc6, is +18 over owner source/+5 since yesterday; CI39995698697 succeeds.

Next raw adoption is not pin-only:

#2177, first master996, appends ref_image_args to sd_vid_gen_params_t and fixes MiniMax-H3 reference resizing (#2168). Audit video layout/default/ownership; no Dart video/editing API delivered.
#2198, master997, centralizes factories/defaults for existing versions, appends FLUX3/H3_AV/MING prediction values and removes sd_ctx_supports_ref_image_conditioning. Audit exports/bindings/default prediction and family outputs. These are not 58 new model families.
#2196, master998, adds opt-in Wan2.2 grouped VAE decoding and repairs frame-stride selection. Preserve output/VRAM/cancel/teardown controls; #2197 tail balancing is still unmerged.
#2199, master1000, fixes older-GPU SageAttention build #2195; no owner CUDA bundle exists, so not a delivered artifact fix.

New watch evidence: FLUX.3 #2192 controls distinguish Q8 im2col OOM→clean tiling from direct-convolution mosaics, not a universal params-backend workaround. Qwen#2193 persists with texture-fix VAE. New dual RX570 #2200 reports system hangs vs clean single-card controls; cause and exact owner exposure remain unproved. Retain real-device/default-mode negatives; software fallback is not GPU qualification.

Historical LoRA#1849 controls support #1872 as an older repair already ancestral to owner982, not a newly missing fix. Iris/Z-Image L2P/FLUX.3/MOSS source is now in the public runtime, not blanket qualified Dart support; Fibo remains raw-only, with existing recipe/edit/shard tickets. #23 is owner-fixed/upstream-unfixed (thirteen casts still raw1000); #24/#2149 remain unresolved, and no guard/carry retirement follows. Latest owner release/Validate remain successful, with no automatic raw adoption.

2026-10-25 raw ownership ABI delta — retain P2; no automatic repin.

Owner remains v0.3.0, owner992d3f70e77137243c833f5c2d9d0f9bb4664d0d, upstream0518b34374761cb82aeda5ad3c2f0686fdad2cec plus pickle_line_span.patch. Public core0.16.2/main081690b and SD companion0.16.0 consume it; actual archives/pins verified. Fresh manifest SHA256dc79dec2a890dd761d3643f657a6df1bb018ded67a192236195d56459cca6544 agrees; Apple digest/minimums unchanged. No new owner artifact or CUDA bundle.

Latest raw master-1006-0886f11 is0886f11b6b8b4d1f6259e0f1337cae554284c37a; mainbb6a0ccfa31ef24d776d951580a51599797d8cc0 is +28 over owner/+10 since yesterday, four commits beyond that release. Release-head CI40133711576 succeeds; main CI40178569574 was queued.

Blocking raw adoption audit: #2206, first raw1004, replaces image/video/upscale/ADetailer pointer/count return contracts with owned generation-result objects, JSON metadata and dedicated destructors. The current guarded wrapper and Dart calls still use the old image-array contract. Decide explicitly whether to preserve or version the owner wrapper ABI; audit transfer/copy/free on success, partial batch, full cancellation, failure and teardown, retaining guard/error-window safety. Only then regenerate matching bindings and worker mapping. Metadata contains input paths/prompts and is not a privacy-safe or complete reproducibility envelope by default; review before any public exposure.

Other material raw changes:

LoRA#2202, raw1001: baked_scale avoids double alpha/rank scaling; FLUX.3 names#2214 closes #2212 on main only, not release1006. Existing core#1223 owns typed adapter transport/state/recipe tests; public image LoRA is absent, so no new Dart regression is established.
Z-Image ControlNet#2187, raw1003, is a Union2.1 component variant, not a base family. Tiled control encoding at1024px has reported artifacts; keep core#1224's SD1.5-first scope.
Wan tail balancing#2197, raw1002, is now merged; reported reduction is compute-buffer size, not whole-device VRAM. Reference#2207, raw1006, skips unconsumed latents; it does not deliver MOSS editing (#2203 remains open).
CPU positional caching#2211 is raw1005; MOSS#2208/#2209 are main-only, fused-SiTU#2210 open. No measured consumer speedup or new public family support. External tokenizer and API qualification still gate MOSS.

Fibofamily remains raw-only since effective0518 pin; Iris/Z-Image L2P/FLUX.3/MOSS source already exists in owner but not blanket qualified Dart recipes. #2193's reporter now also sees grids in hosted QwenImage3.0; this weakens an sd.cpp-only diagnosis, not proof of a fix. #2192/#2200/#2149/#24 otherwise unchanged; rawmain still has thirteen narrowing casts, so retain #23's owner carry. Owner release/Validate remain healthy; no new runtime/device/model test or retirement.

2026-10-26 raw delta — retain P2 and owner v0.3.0.

Owner992d3f70/v0.3.0 still bundles upstream0518b34374761cb82aeda5ad3c2f0686fdad2cec plus pickle carry. Public core0.16.2/SD companion0.16.0 consume it; main6305b34c and frozen0.16.3 prep keep that pin. Source-only VAE tiling, seamless and conditioning-capacity APIs reuse the guarded pinned ABI; none is a raw runtime adoption.

Raw main/latest release now coincide at master-1015-e41fbcf, e41fbcf455e0fbb45e08bd535b00d6cba865e921 (+5 since yesterday/+33 over owner); CI40214476521 succeeds. Manifest dc79dec2.../Apple digest remain unchanged; owner release/Validate are still healthy.

New substantive watches, not established Dart regressions:

GGUF#2218 reports per-string metadata skip/seek overhead. Owner982 has the same safe_skip implementation; measure exact current-artifact cold/warm loading before attributing the reporter's timing to our bundle. Separate from ZIP#24.
FLUX.3#2219 reports over-budget Q8 as one graph segment on6GB RTX4050 while fittingQ4 works; owner982 also lacks graph-cut markers in flux3.hpp. Core#1130 gets fitting/oversized and segmented/unsegmented controls. No owner CUDA artifact or tested owner/backend exposure.
LTX2.5#2220 reports prompt loss around25+frames against matched Lightricks controls; cause unresolved. Core#1220 retains semantic/frame-grid/reference qualification; no public video API exists.

Server ref media#2194, first raw1011, is server parsing, not new family/Dart transport. Recipes#2217, first raw1015, adds dimension/frame/FPS constraints and changes mismatched-mask preprocessing to nearest crop-resize; keep #1221 normalization/refusal semantics explicit.

Previously main-only FLUX.3 LoRA names#2214/raw1010 and MOSS#2208/raw1008/#2209/raw1009 are now published raw, not owner/public support. Fibo masked/text-image proposals#2215/#2216 remain open, as do MOSS editing#2203 and backend-dependent SiTU#2210; existing component/edit/shard/API issues suffice. No new base family today; Fibo remains raw-only relative to owner982.

No new C-header ABI delta today: #2206/first1004 owned-result/signature/destructor migration remains mandatory before wrapper/bindings adoption. LoRA#1849 is now closed after reporter confirms raw1000 works, but that repair was already ancestral to owner982—not a new missing fix. Keep thirteen pickle casts/carry, Metal guard, ZIP and other GPU watches unchanged. No model/device/workflow action or new retirement.

2026-10-27 raw fixes and draft adaptation — retain P2; published runtime remains v0.3.0.

Published v0.3.0, owner992d3f70e77137243c833f5c2d9d0f9bb4664d0d, bundles upstream0518b34374761cb82aeda5ad3c2f0686fdad2cec plus pickle_line_span.patch. Public core0.16.4/SD companion0.16.0 and current main retain it. Manifest SHA256dc79dec2a890dd761d3643f657a6df1bb018ded67a192236195d56459cca6544 is unchanged. No newer owner artifact or CUDA bundle.

Raw latest/main master-1017-be48649, be486498eadc29dc3cd234d944c2180ed42b1415, is +2 since yesterday/+35 over bundled source; CI passes.

#2223, first1016, fixes #2218's GGUF metadata seek overhead.
#2224, first1017, fixes #2219 with FLUX.3 graph-cut markers. These are raw fixes, not published owner performance/low-VRAM guarantees; core#1130 retains exact-family/backend controls.

Draft#28, head9333172da73a719012a86f6976d2a22ba098cdf2, now adapts #2206 while preserving sd_dart_generate_image's pointer/count ABI: take_images moves pixels, clears moved pointers, then frees owned results/metadata. This supersedes the unresolved wrapper-design checkpoint, not publication. Its independent review records no P1/P2 blockers, with take_images failure-test coverage and full-cancellation batch-drop documentation still called out. Keep the draft hold until0.16.5. Matching upstream bindings, removed-export/enum/video-layout auditing and qualified bundles/manifests/checksums/SwiftPM/consumer gates remain. No new public video/upscaler/detailer/RPC call is supplied.

Safety is canonical in #29/core#1446: Metal FP8 abort remains on old/current/candidate sources, not fixed by#28. Its source-build matrix must not be described as downloaded release-artifact qualification. Core#1453 separately proposes pre-call typed refusal and retains pinned-runtime CPU behavior; draft, not shipped. Retain P1 owner priority independently of consumer mitigation.

New watches: hires#2226/#2227 effective-step off-by-one (#1228 controls), iGPU planner#2228 automatic versus explicit selection (#806; not demonstrated CPU fallback), and Qwen multi-GPU CUDA#2225. Repairs remain open/absent1017; no owner reproduction or causal regression established.

No new base family today. Fibo remains raw-only; existing MOSS/FLUX.3 recipes and LTX semantic/device qualification stay separate. Pickle/Metal guards remain necessary; ZIP#24 remains unresolved.

2026-10-28 owner publication and raw-only corrections — retain P2 qualification.

Newest published artifact is v0.3.0-1, owner54e97df3b8bfa52f18074bbcc6dab557b323b243, upstream be486498eadc29dc3cd234d944c2180ed42b1415 (master1017). It is a GitHub prerelease, so releases/latest still returningv0.3.0 is not evidence it is absent. Publication40556402932 passes all14 jobs. Manifest SHA2562334e0f2c87f7335d68448c54a45cbb49609c20de50715eb2b4f5212bc778834; Apple ZIP cebe0771734a657ff02420daa1190257185ced49f606e177c8e42789bd4323a2. Twelve bundles, no CUDA bundle.

#28 is merged; prior take_images failure coverage/full-cancel documentation follow-ups landed. Owned-result adaptation, GGUF seek#2223 and FLUX.3 graph cuts#2224 are now artifact-supported, not draft-only. #30 and #32 first ship here: FP8→F16 Metal loading, E5M2/fit accounting and immediate/runtime LoRA fallback repairs.

Three carries remain: fp8_metal_f16_load, lora_immediate_fallback and pickle_line_span. None was retired. LoRA fallback still does not apply TAESD adapters; generic warnings and FP8 conversion risks must not become silently supported public options.

Core#1474 already adopts the artifact on main with regenerated bindings, checksums/SwiftPM, release+symbol FP8 gate, F16 memory accounting and bounded-recipe requalification. Public core0.16.5/SD companion0.16.0 remainv0.3.0/upstream0518b343 plus pickle; they provide typed Metal FP8 refusal, not the new kernel-load route. Matching0.17 core/companions and wider device evidence remain separate.

Raw latest release is still1017, but main687915518efd69548965d24818f0d3bbbb903718 is five commits newer:

Hires#2227 fixes the extra-step calculation, and iGPU#2228 includes integrated GPUs in auto-fit enumeration. Both main-only/absent owner-1. Existing core#1228/#806 own affected step/progress/output and automatic/explicit UMA-budget controls; not a new consumer regression or CPU-fallback repair.
Raw release40553159044 failed Create release with “Resource not accessible by integration” after builds; later main40554881650 was cancelled/release skipped. Permission/workflow source edits are not successful publication evidence.

Fibo2 is newly owner-artifact-supported versus public0518, not a new family landing today. Text-to-image needs a qualified component recipe/docs; editing needs llm_vision transport absent from current core roles. Existing #20/core#1222/#1219 suffice; Kontext#1515 is not Fibo qualification. No new duplicate. ZIP#24, LTX semantics and wider hardware watches retain acceptance.

2026-10-29 raw publication recovered; retain P2 owner qualification.

Raw latest release/main now coincide at master-1024-77b72a2, 77b72a2efed78f55de26892c7e909cee001e8b83. CI40702101039 passed all14 jobs including release creation/uploads. Yesterday's permission failure is historical, not a continuing publication stall.

Published owner remains prerelease v0.3.0-1/54e97df3b8bfa52f18074bbcc6dab557b323b243 on be486498eadc29dc3cd234d944c2180ed42b1415 (raw1017), already pinned on core main. Public core0.16.5/SD companion0.16.0 retain v0.3.0/0518b343. The seven newer raw commits are absent both owner and public pins; no additional main pin bump is needed for the already-adopted artifact.

#2227's hires extra-step correction and #2228's integrated-GPU auto-fit enumeration are first published in raw1024, no longer main-only. Keep core#1228's documented old step/output semantics and #806's automatic/explicit UMA-budget controls until requalified on a future owner artifact.

#2235, first1024, is not merely an unused CUDA optimization: ComputeWorkspace now always uses the scheduler, retains tensor metadata between steps and propagates eval cancellation. CUDA adds INT8/CUTLASS/fusion/attention paths and recursive-submodule/build requirements. Require CPU/Metal/Vulkan output, workspace/memory, cancel/failure/teardown/reuse controls plus applicable CUDA build/device evidence. Its RTX4090 Qwen2.1/Krea2 sampling-only timings exclude loading/TE/VAE/save; there is no owner CUDA bundle or measured Dart speedup. No public stable-diffusion.h change was found in this delta, so no new Dart API/binding change is implied, but owner rebuild/requalification remains necessary.

Chroma#2237 is existing-family version-ID/detection refactoring, not new base-family support. Fibo's artifact-versus-Dart recipe/vision-projector gates remain. Open H3 Vulkan >4GiB buffer#2242 and CLI audio/video muxing#2233 are upstream evidence, not established consumer regressions; existing #1220/#1219 cover relevant future controls.

All three carries remain, including the thirteen upstream pickle narrowings. ZIP#24/LTX/device watches remain unchanged; no carry, workflow, model/device job or consumer pin was changed by this scan.

2026-10-30 raw candidate update — retain P2 qualification.

Published owner remains v0.3.0-1, owner54e97df3b8bfa52f18074bbcc6dab557b323b243, upstreambe486498eadc29dc3cd234d944c2180ed42b1415 (raw1017). Public core0.17.0 pins it; yesterday's public-v0.3.0 distinction is superseded.

Latest raw publication master-1028-37bc88b, 37bc88b5e97a213821949732ea66c58ae8ffc784, is four commits beyond yesterday/eleven beyond owner. CI40913538449 passes all14 jobs.

SDXL #2247, first raw1027, fixes standalone Diffusers-named UNet detection: existing family, not general directory/shard ingestion. Qualify explicit UNet+CLIP-L+CLIP-G+VAE against same-weight checkpoint and CPU/GPU controls.
H3/prefetch #2240, first raw1028, changes shared runner async compute, prefetch submission, cached outputs, buffer-cache lifecycle/capacity—not merely CUDA video throughput. Require CPU/Metal/Vulkan output, memory-pressure, cancellation/failure/reuse/teardown controls. RTX4090 sampling measurements are not Dart/end-to-end performance evidence.
GEGLU #2245, first1025, optimizes existing models. Hires #2244, first1026, simplifies #2227's calculation, not a second independently established step-count repair.

Raw main GLU #2250 is e239e4b0fd3de49834cec67031a44796ca465abc, one later commit. CI40922355947 has twelve successful builds, MUSA cancelled after exceeding six hours, release skipped; annotations establish timeout, not compiler/model failure or renewed publication-permission failure.

No new base family/public-header change in this delta, so no newly required Dart API/binding change is demonstrated. Owner rebuild/qualification remains necessary before moving bounded recipe gates, checksums/manifests/SwiftPM pins and docs. All three carries remain retained; ZIP#24 is unresolved. H3 hires#2249 is open, with audio/reference paths unverified. Existing owner/core umbrellas suffice; no automatic repin or job was started.

2026-10-31 candidate delta — retain P2 qualification; no automatic repin.

Raw main0a067120b388bfefb9a87b96960cb548b97e9157 includes merged #2251. Latest published raw remains master-1028-37bc88b, two commits behind main. Owner v0.3.0-1/upstreambe486498, public0.17.0 and current main's SD pin are unchanged.

#2251 adds Vulkan INT8 GEMM/ConvRot/attention/fusion, shared single-device workspace-measurement caching and supported direct-Conv3D fallback when full im2col exceeds backend limits. The nested ggml delta includes a Vulkan profiling pending-upload synchronization repair and basic F32 SiTU CPU/CUDA primitives. No new family, container format or public Dart capability is established.

Upstream's roughly3.7–4.4x RTX4090 Vulkan sampling claims for Qwen Image2.1/Krea2 Turbo INT8-ConvRot discard warm-up and exclude load/TE/VAE/output. They are not owner, AMD/mobile or Dart end-to-end measurements.

Additional adoption controls: unchanged versus resized/changed graphs, topology/placement/fresh-load cache invalidation, bounded memory and CPU/Metal/Vulkan output; cancel/failure/dispose/reload; single versus multiple devices; supported/unsupported direct-Conv3D at backend limits; profiling on/off and applicable INT8/fusion/attention hardware. Do not infer repair of separate H3 MLP4GiB#2242.

CI41178564677 has eleven successful builds including MUSA and two ROCm jobs running at inspection; no new publication yet. Yesterday's MUSA timeout is historical, not a continuing stall. No stable-diffusion.h/export delta was found; all three owner carries and existing guarded-wrapper/recipe/device gates remain. New owner publication, manifests/checksums/SwiftPM and focused consumer requalification would precede a later adoption; no new API/binding work is demonstrated by this delta alone.

2026-11-01: raw publication progressed, with a new Vulkan memory-regression watch. Retain P2; no automatic repin.

Raw main/release master-1031-c6dfa9d, c6dfa9def12fd0cb36ac7149ba30a75792bedbf0, is14 commits beyond ownerbe486498. CI41219737394 passed all14 jobs, superseding the previous ROCm-in-progress/publication state. Published ownerv0.3.0-1, publiccore0.17.2 and current main remain unchanged.

#2252's ggml6a894937→c57b71a adds Vulkan reuse/allocation, async uploads, INT8/ConvRot and H3 fusion. Documentation places up to1GiB/device of idle cache outside --max-vram; eligible discrete devices default to separate transfer queues and NVIDIA allocation chunks can reach4GiB. RTX4090 H3 sampling measurements are not Dart/end-to-end or general hardware qualification.

New #2254 reports Krea2/Kroma Q8 OOM on RX6700XT/RADV under background VRAM pressure, narrowed by reporter A/B controls to cc937ae. The older6a89493 control succeeds; GGML_VK_REUSE_BUFFERS=0 reportedly does not help. This is not our reproduction, proven root cause or established current-owner exposure, and not the older AMD/H3/owner#34 mechanisms.

Before adoption: compare exact pin/candidate and pre/post-cc937ae under idle/pressured VRAM, cache/transfer/chunk controls, correct output/peak memory, bounded failure, cancel/reuse/dispose/reload/teardown. Do not advertise cache-disable as a workaround without reproduction. Preserve existing model/header/guarded-wrapper/recipe/Apple-manifest gates.

No new base family, format or stable-diffusion.h change landed. Open #2253 is an unmerged backend-extension/CUTLASS migration, not delivered support. Existing #20/#34 and consumer umbrellas suffice; no duplicate issue, model job or artifact/pin change was made.

2026-11-02 candidate update — retain P2 qualification; no automatic repin.

Raw master-1039-290148d, 290148d2e31ca6ef9757d567905776b5f3646db5, is 22 commits beyond unchanged owner v0.3.0-1 / be486498eadc29dc3cd234d944c2180ed42b1415. Exact-head CI41330885992 passed all 14 jobs. Public core 0.17.5 and companions 0.17.3 retain this owner artifact; no new public capability follows.

#2256, first raw1035, closes #2254 with Vulkan synchronization/cache clearing and budget-preserving capacity retries bounded to 100 ms. No confirming affected-hardware reporter rerun was found. Retain pressured-VRAM, exact-pin/candidate, output, peak-memory and bounded-failure controls; closure alone is not delivered owner repair.

#2253, first1032, is now a merged backend-extension/CUTLASS migration. #2257, first1033, chunks H3 oversized MLP buffers and closes #2242; it does not supply public video support. #2259, first1038, prefers discrete GPUs in auto-fit. #2258, first1039, adds shared F32 modulation/ConvRot fusion.

Require owner build/dependency/export checks, CPU/Metal/Vulkan output and memory/lifecycle parity, explicit-versus-auto mixed-device assignments, and existing carry/Apple/recipe/device acceptance before publication/adoption. No stable-diffusion.h, format or new-family delta was found; no new Dart API/binding addition is demonstrated.

Open H3 hires#2249 now has CPU end-to-end evidence, but its new ROCm segfault report remains unresolved. PR CI41378538257 is action_required with zero jobs, not a compile failure or stalled release. #34's fit-off abort and the older iGPU planner issue remain distinct.

The current v0.3.0-1 manifest digest and 25 artifact size/digest records match release metadata; that is not new payload/device qualification. Three carries remain. No workflow, model/device job or owner artifact was changed.

2026-11-03: new owner artifact is published and source-adopted; newer raw remains separate. Retain P2 qualification.

Owner v0.3.0-2, source9d228522749a75af13a52cd5f285acd7547119a9, still bundles be486498eadc29dc3cd234d944c2180ed42b1415. Manifest SHA256 d4ac5e8a070ac72f250b0c7582246853119d02fa6b8c672983618554b5f15e5c and 25 artifact size/digest records match release metadata. Core#1729 merged67700531 and main adopts the matching wrapper/Apple artifacts; public core0.17.5/companions0.17.3 retain -1. No further main pin implementation is needed for -2.

The new guarded intrinsic-tiny constructor/probe fails closed rather than using an unsafe old constructor. Unsafe-budget refusal is also delivered. The recorded physical iPhone16Pro/iOS18.3 SDXS basic profile is bounded evidence, not full-VAE watchdog repair, broader GPU support or new model-family support; #1047 retains the wider limits. Ubuntu26 software-Vulkan #37 remains unresolved despite the 24.04 producer pin.

Raw master-1043, 3be23788fc0c9f35a2c24b1211caae11d4a54dbc, is 26 commits beyond the retained upstream. CI41480582683 passed all14 jobs. #2262 fixes multi-CUDA node assignment; #2263 converts integer LoRA scalars. Both first appear in raw1043 and are absent owner -2; no confirming affected-reporter rerun was found.

New Chroma scheduler/grid#2267 needs exact pin/candidate/reference-schedule controls. Shared sequence-length source with the pin does not prove affected owner pixels or causality. Preserve output/memory/lifecycle, build/export/carry, Apple and hardware gates before future raw adoption.

Correct the prior H3 watch: #2249's reporter now confirms success after branch corrections. It remains unmerged, not broad ROCm or public-video qualification. #24's fresh non-reproductions are not causal fixes. No new raw family/format/header API landed; the four owner carries and all scoped consumer limits remain. No model/device execution or pin was changed by this scan.

2026-11-04 delivery/raw delta — retain P2 qualification; no automatic raw repin.

Latest v0.3.0-3, owner5fde7dd13544385fdacd9fe31062064496052024, retains upstream be486498eadc29dc3cd234d944c2180ed42b1415 and adds #41's bounded PyTorch ZIP lookup carry. Manifest SHA25630d1eca28e4158dccf957d6f702833de8ac3b928e8a284765f5965b3f5b86869 and all25 manifest size/digest records match release metadata. No wrapper/header/export change; no new Dart API/binding work. Main remains -2, public core0.17.5/companion0.17.3 remain -1; draft core#1739 already stages -3/pins/SwiftPM. #24 / core#1703 retain bounded original-container evidence and final adoption gates.

Raw master-1047-984eb06, 984eb0686cbf5b7fc4f0063d9e0b1fb38d5c4f82, is30 commits beyond bundled upstream. New MOSS SiTU#2210/QKV#2232, Z-Image BF16 fusion#2265 and MOSS modulation/RoPE#2269 first appear in raw1044/1045/1046/1047. They add no family, container or stable-diffusion.h delta. Upstream RTX4090 sampling measurements are raw-only; the owner publishes no CUDA bundle. CI41689852788 passes all14 jobs; -3 prepare/publication replay succeeds, with intentional skipped build/release stages rather than stalled automation.

New H3 VAE latency#2268 logs place the7.6GB compute buffer on CPU. Its309-second result is not a pure hardware-Vulkan benchmark; placement/budget suggestions are not a verified fix. Existing video owner core#1220 suffices. Existing #37 software-Vulkan drift, full-VAE watchdog, Chroma grid and unmerged H3-hires limits remain separate. Retain five carries and exact-output/memory/cancel/reuse/teardown/Apple/hardware gates before newer raw adoption.

Navigation Menu

Repository navigation

Qualify post-v0.2 upstream changes with context ABI and FP8 CPU safety checks #20

Current delivery snapshot — 2026-11-04

Original investigation and retained acceptance

Triage

Current layers — 2026-10-21

Evidence and benefit

New raw evidence — 2026-10-17

Acceptance

Boundaries

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Qualify post-v0.2 upstream changes with context ABI and FP8 CPU safety checks #20

Description

Current delivery snapshot — 2026-11-04

Original investigation and retained acceptance

Triage

Current layers — 2026-10-21

Evidence and benefit

New raw evidence — 2026-10-17

Acceptance

Boundaries

Activity

leehack commented on Oct 15, 2026

leehack commented on Oct 17, 2026

leehack commented on Oct 18, 2026

leehack commented on Oct 19, 2026

leehack commented on Oct 20, 2026

leehack commented on Oct 21, 2026

leehack commented on Oct 22, 2026

leehack commented on Oct 22, 2026

leehack commented on Oct 23, 2026

leehack commented on Oct 24, 2026

leehack commented on Oct 25, 2026

leehack commented on Oct 26, 2026

leehack commented on Oct 27, 2026

leehack commented on Oct 28, 2026

leehack commented on Oct 29, 2026

leehack commented on Oct 30, 2026

leehack commented on Oct 31, 2026

leehack commented on Nov 1, 2026

leehack commented on Nov 2, 2026

leehack commented on Nov 3, 2026

leehack commented on Nov 4, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions