Release 0.49.0 — the SKEEP-003 architecture, complete - #1185
Merged
Merged
Conversation
The deliberate 0.40.1 → 0.49.0 jump (~100 merged PRs): one storage model, resolver-owned decisions, scope-recycled eager execution, and a compile lane that carries what the runtime decides. This is the release SKaiNET-transformers and the IREE conformity pipeline build on. - CHANGELOG: full 0.49.0 notes with an explicit Breaking-changes migration map (legacy loader axes → WeightForm; dead placement machinery → AllocationResolver; skainet.tensor_encodings → skainet.tensor_layouts). - README: What's New in 0.49.0, BOM coordinate, contributors incl. @AjithGoveas (Iris provider). - Versions: gradle.properties, docs/antora.yml skainet_version, tutorial coordinates; kernel-support matrix and operator reference regenerated at 0.49.0. - Docs audit (62 pages): every reference presenting removed API as current is fixed (WeightOrientation/staging/quantPolicy prose, SKEEP-002 snippet, tensor_encodings seam); overtaken claims updated (op outputs + carriage landed, CI api-compatibility leg + branch protection, native FFM shipped, device placement resolved, ternary 'in progress' → landed, eager-execution kernel matrix gains the exact-FP32 ternary and BITNET_PLANES rows); SKEEP-003 status → Implemented and SKEEP-003a added to the skeep index; the dangling docs/design/memory prose repaired; a leaking literal block and a wrong cross-module xref fixed; the fake StableHloOptimizer(debugMode) snippet replaced with the real HloGenerator target API. Known non-blockers, left deliberately: how-to/java-cli-app.adoc and tutorials/kllama-getting-started.adoc are orphaned from nav (content review needed before linking). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
📖 Documentation Preview The documentation has been built successfully for this PR. Generated Files:
Artifacts:
This comment will be updated automatically when the PR is updated. |
|
📖 Documentation Preview The documentation has been built successfully for this PR. Generated Files:
Artifacts:
This comment will be updated automatically when the PR is updated. |
…path On the JVM, SKEEP-004 activations live in MemorySegments, and the common views->KernelDispatch path a packed ternary weight takes (#1136) cannot read them: element access over SegmentStorage is forbidden by design, so the first linearProject over a BitNetB158TensorData weight in a real decode loop threw the SKEEP-004 guard. matmulWeightTransposed on the JVM tier now bulk-copies such an activation to the heap once per call before handing the common path its views — decode-step activations are k floats, not weights, so the copy is noise next to the matmul. Two producers are handled: TensorData that is itself segment-backed (MemorySegmentBackedData), and TensorData whose view presents a SegmentStorage. The copy deliberately constructs DenseFloatArrayTensorData directly — this tier's dataFactory adopts float arrays back INTO segments, which would round-trip the problem. Found and verified running the real microsoft/bitnet-b1.58-2B-4T I2_S GGUF end-to-end in SKaiNET-transformers: with this fix the model loads packed and generates coherent text ("The capital of France is Paris. Paris is known for its beautiful architecture, art, and cuisine...") through the vendored ternary kernels — which also confirms the GROUP_128 default of the I2_S import (#1140's open caveat) against the official BitNet.cpp file. Refs #1136 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…k ships with the release
|
📖 Documentation Preview The documentation has been built successfully for this PR. Generated Files:
Artifacts:
This comment will be updated automatically when the PR is updated. |
…ivation-ternary fix(cpu): off-heap activations reach the ternary dispatch path (#1136)
|
📖 Documentation Preview The documentation has been built successfully for this PR. Generated Files:
Artifacts:
This comment will be updated automatically when the PR is updated. |
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
1 participant
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.
Release 0.49.0 — the deliberate jump from 0.40.1 (~100 merged PRs): the SKEEP-003 memory & storage architecture complete, from accepted proposal to shipped system. This is the release SKaiNET-transformers and the IREE conformity pipeline build on.
In this PR (release prep only — no code changes):
WeightForm;@Place/@Weights/StorageSpec/oldMemoryPlanner→AllocationResolver;skainet.tensor_encodings→ structuralskainet.tensor_layouts.gradle.properties(VERSION_NAME=0.49.0), Antoraskainet_version, tutorial coordinates; regenerated kernel-support matrix and operator reference (the latter had been stale at 0.27.0 since June).WeightOrientation/staging/quantPolicyprose, SKEEP-002's loader snippet, thetensor_encodingsseam claim, the danglingdocs/design/memoryprose from docs: describe the memory model as built, and fold the loose files into Antora #1106); overtaken claims updated (op outputs + carriage landed; CI'sapi-compatibilityleg + branch protection; native FFM shipped, not "planned"; device placement resolved; ternary "in progress" → landed; eager-execution kernel matrix gains the exact-FP32 ternary andBITNET_PLANESrows); SKEEP-003 → Implemented, SKEEP-003a added to the skeep index; one leaking listing block and one wrong cross-module xref repaired; the nonexistentStableHloOptimizer(debugMode)/dumpGraphviztutorial snippet replaced with the realHloGenerator.generate(target = …)surface.Left deliberately (follow-ups, not blockers):
how-to/java-cli-app.adocandtutorials/kllama-getting-started.adocare orphaned from nav — content review needed before linking them.After merge: publish 0.49.0, then downstream: bump SKaiNET-transformers onto it, unpark the
model-footprintCLI in SKaiNET-research (bumpskainetVersion, drop the mavenLocal preamble), and draft the IREE-validation SKEEP against the conformity repo (#1148).Full pr-gate green.
🤖 Generated with Claude Code