Repository navigation
Conversation
Install Transformers and PEFT in the native base, use operator-managed uv, and document ordinary uv/npm setup with explicit accelerator choices. Preserve reviewed specialist AMD contracts and probe compilation on the intended device. Correct runtime activation progress, startup/reconnect behavior and resident Auto resource accounting. Preserve image recipes through native encoding, guidance, precision, ControlNet and named ordered LoRA contracts, and bundle the migrated ordinary-template client. Add output comparison, memory capture and portable AMD preparation tools, with the Mellon comparison and remaining image-workflow validation plan. Report missing offline upscalers with their real selection and preparation action. Validation: 4857 backend tests and 11105 subtests passed; 26 tests skipped. Ruff E9/F, shell parsing, native dependency checks and tiny eager/SDPA repeated generation with named PEFT adapters passed. Earlier full-weight comparisons retain their original source/hardware scopes; Windows and dedicated AMD checks remain separate. Private media/logs and operator state are excluded.
Windows checkouts can give pinned Diffusers Python sources CRLF endings. Reuse the existing source normalization for operation inventory receipts and the missed Shap-E signature check, preserving exact JSON bytes and artifact pins. Regression cases still detect code and JSON byte drift. Bind synthetic runtime tests to explicit host/build/device fixtures. Exercise both Windows uv and Python launcher branches, retain native stderr and exit-code checks, and sample the allocator's reported PID through Windows venv redirectors. Validation: full backend gate, 4863 passed and 11108 subtests passed; 28 skipped. Ruff E9/F and source inventory reproduction passed. The preceding Windows/macOS CI failures are retained; fresh platform CI will verify these fixes.
Declare reviewed guidance meanings and opt-in Qwen component bundle inputs with exact managed ownership, role and cache-residency validation. Preserve original masked Qwen recipes through native stages, including source preprocessing, VAE sampling, mask compositing and runtime policy ownership. Handle absent runtime hints when persisting generated outputs. Ship the paired frontend build and guidance, component authoring, Mellon comparison and three-machine validation documentation.
Regenerate the upstream coverage inventory from the final mirrored client bundle and reviewed upstream sources, then refresh its four dependent ledgers. Record the four promoted Qwen native execution selections without changing historical qualification or coverage counts. Validation: all nine existing generator check/validate gates pass; full backend pytest after final mirroring passes (5002 tests, 11186 subtests, 28 skips). All 105 production assets still match the frontend, and 67 protected files are unchanged.
Add a rigorous Windows validation plan covering all ten Hugging Face feedback items, independent native/guided setup, cold and repeated Auto generation, real LoRA and refresh/stop/retry journeys, and scoped compiler/runtime qualification. Link it from the three-machine guide. Correct API and troubleshooting prose that still described mandatory base dependencies as optional. Plan commands, public schemas and links independently reviewed; targeted documentation formatting and diff checks pass. Runtime code, dependencies and production assets are unchanged from the full backend test gate.
Mirror the validated client build from MoDiff-client d889c1d1e817b5d646f3da16fa54938b749dc5fa. Backend validation: 5031 passed, 28 skipped, 11226 subtests passed. Client full check and committed bundle identity checks passed.
A fresh optional legacy graph library now returns an empty array instead of a 404. Preserve real filesystem failures and add missing-versus-empty and permission regressions. Mirror the reviewed client 7461b8fa2a60fc4823603e2486847ce3be7f2df7 and refresh only publication hashes through the existing five ledger generators. Qualification and operator state are unchanged. Validation: 5033 backend tests and 11226 subtests passed, 28 hardware/optional tests skipped; Ruff E9/F and CPU package compatibility passed. Client npm run check and focused browser regressions passed. Full browser suite and live AMD mutation retests are tracked separately.
Fresh model owners apply the existing ROCm vision helper before placement and publication. Shared components keep their source ownership and require compatible attention settings. Mirror client d52a8ea6a0463c9519df29b3fa631289e2b1aba8, including captured task subtitles and clearer image-source choices. Refresh only the five existing source-ledger hash bindings. Validation: 5052 backend tests and 11226 subtests passed, 28 skipped; 2405 client Node tests, 34 Python tests and 250 browser cases passed. Ruff, package checks, preflight and tiny native image/named LoRA smoke passed. Full model output and hardware qualification are recorded separately.
…racts Bind package-owned model loaders to the same reviewed workflow as their stages. Refresh the 17 affected registered Block definitions and the matching client bundle without expanding model admission or qualification. Keep Windows cache fixtures consistent with native path identity and record the original-weight 100B LLaDA native SDK smoke with its measured limits. Validation: backend 5,291 passed, 31 skipped and 11,331 subtests; compatible frontend full check and unit suite passed. Final source and bundled files matched the reviewed pair and preserved user files. Windows GPU and remaining image quality qualification are separate follow-ups.
Preserve shared single-owner reuse and explicit additive memory budgets. Use the existing release/recheck schedule for sequential unmeasured primary owners; reject unresolved overlapping owners without changing their graph. Validation: 304 focused tests and 445 subtests passed. Full native CPU suite: 5284 passed, 43 skipped and 11331 subtests passed; the two external CLI source-import failures passed their exact retry with the checkout import path configured.
Keep validated backend-owned JSON text unchanged for tokenizer input. Exact starter defaults now match the original publisher artifact, and invalid JSON still fails before prompt authoring. Reviewed frontend templates can inherit this default through generic authored-value provenance. Validation: 65 focused tests and 1821 subtests passed in the isolated candidate; root exact-variant gate 30 passed; Ruff E9/F passed. Full CPU gate passed 5291 tests and 11339 subtests with 43 explicit skips; one generated web/template fingerprint failure is retained for coordinated bundle-ledger regeneration before publication.
Preserve exhausted planning/driver pools instead of replacing them with a process-local estimate. Unknown values retain the existing CUDA/XPU/MPS fallback and shared-pool rules; capacity and sysfs priority are unchanged. Validation: private regression reproduced 8 failures before the fix and 20 passes after it, including the actual XPU hardware-probe-to-next-owner path. Root Auto/planner/hardware gate passed 225 tests and 450 subtests; Ruff E9/F passed.
…ance Keep compatible connected group/global loaders, but release native model-CPU and resident owners sequentially or reject unsafe overlap before loading. Preserve borrowed placement and disk ownership, select current compatible owners before deduplicating shared-weight credit, and retain known-zero free device memory. Direct required-library diagnostics to the matching accelerator base repair. Update the feedback and Auto handoff with retained evidence, Gallery limitations and the local validation scope. Validation: required Ruff gate; clean native CPU uv setup, dependency consistency and preflight; real tiny eager generation/named-LoRA and model-free service smokes; full CPU suite (5,399 passed, 43 skipped, 11,414 subtests). No new production GPU qualification.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Scope And Human Review
none:Checks
uvx --from ruff==0.12.7 ruff check . --select E9,Fuv pip check --python .venv/bin/python(use.venv/Scripts/python.exeon Windows)./.venv/bin/python -m modiff.preflight --json --check-port 8088 --fail-on-error./.venv/bin/python -m pytest -qbash -n run.shwhen a POSIX shell is availableValidation Results
not claimed:Compatibility, Security, And Proof