SKaiNET-transformers → engine 0.51: full engine-loader migration, MAPPED default, BitNet fold-in by michalharakal · Pull Request #353 · SKaiNET-developers/SKaiNET-transformers

build(#338): pin engine 0.51.0-SNAPSHOT + mavenLocal dev lane + bring-up shims

michalharakal

and

claude

committed

6b67ce4

feat(#338): CLI installs the 0.51 kernel tiers; linearProject seam flip deferred with evidence

michalharakal

and

claude

committed

41f35eb

feat(apertus)!: onto the engine loader — WeightForm replaces QuantPolicy (#339, fixes #100)

michalharakal

and

claude

committed

7f20943

feat(decoder)!: llama/qwen/smollm2/voxtral onto the engine loader (#340) + qwen2 attention-bias support

michalharakal

and

claude

committed

d2cd546

feat(gemma)!: gemma onto the engine loader + delete the packing layer (#341)

michalharakal

and

claude

committed

ffec546

feat(memory)!: MAPPED residency by default + memory plan + Arena retirement (#342)

michalharakal

committed

88b228a

feat(bitnet): BitNet b1.58 architecture — DSL extension + network def (#151)

michalharakal

and

claude

committed

71c7c36

feat(bitnet): the packed I2_S load path — BitNetPackedGgufLoader (#152)

michalharakal

and

claude

committed

4f31b4e

feat(bitnet): two-stage lm_head decode with an exactness guarantee (#152)

michalharakal

and

claude

committed

d437bd4

feat(bitnet): the real model speaks — 1-bit BitNet-2B4T usable via the CLI

michalharakal

and

claude

committed

3440cb6

feat(bitnet): align the fold-in to the 0.51 engine — WeightForm lanes + MAPPED request (#336, #337)

michalharakal

committed

85bfc0c

refactor(quant)!: QuantPolicy to zero (#342)

michalharakal

committed

c1dfdbe

feat(memory): per-step ForwardScope in the DIRECT decode loop (#343)

michalharakal

committed

cd46bd4

refactor(families): shared decoder metadata parser + delete the deprecated accelerator seam (#346)

michalharakal

committed

c7a3d28

docs: engine-adoption arc epilog — WeightForm/MAPPED docs, --explain-load, CHANGELOG (#344)

michalharakal

committed

69c8a6f

build: pin skainet 0.51.0 (tagged release)

michalharakal

committed

d064839

fix(kllama): install the 0.51 kernel packs in KLlamaJava/kllama-cli (#354)

michalharakal

committed

fda0d7e

test(kllama): pin the Llama3ChatTemplate + KLlamaJava combination EdgeTranslator relies on (#354)

michalharakal

committed

e7aafdf

Navigation Menu

SKaiNET-transformers → engine 0.51: full engine-loader migration, MAPPED default, BitNet fold-in - #353

SKaiNET-transformers → engine 0.51: full engine-loader migration, MAPPED default, BitNet fold-in#353
michalharakal merged 18 commits into
developfrom
feat/0.51-migration

Commits on Aug 29, 2026

build(#338): pin engine 0.51.0-SNAPSHOT + mavenLocal dev lane + bring-up shims

feat(#338): CLI installs the 0.51 kernel tiers; linearProject seam flip deferred with evidence

feat(apertus)!: onto the engine loader — WeightForm replaces QuantPolicy (#339, fixes #100)

feat(decoder)!: llama/qwen/smollm2/voxtral onto the engine loader (#340) + qwen2 attention-bias support

feat(gemma)!: gemma onto the engine loader + delete the packing layer (#341)

feat(memory)!: MAPPED residency by default + memory plan + Arena retirement (#342)

feat(bitnet): BitNet b1.58 architecture — DSL extension + network def (#151)

feat(bitnet): the packed I2_S load path — BitNetPackedGgufLoader (#152)

feat(bitnet): two-stage lm_head decode with an exactness guarantee (#152)

feat(bitnet): the real model speaks — 1-bit BitNet-2B4T usable via the CLI

feat(bitnet): align the fold-in to the 0.51 engine — WeightForm lanes + MAPPED request (#336, #337)

refactor(quant)!: QuantPolicy to zero (#342)

feat(memory): per-step ForwardScope in the DIRECT decode loop (#343)

refactor(families): shared decoder metadata parser + delete the deprecated accelerator seam (#346)

docs: engine-adoption arc epilog — WeightForm/MAPPED docs, --explain-load, CHANGELOG (#344)

build: pin skainet 0.51.0 (tagged release)

fix(kllama): install the 0.51 kernel packs in KLlamaJava/kllama-cli (#354)

test(kllama): pin the Llama3ChatTemplate + KLlamaJava combination EdgeTranslator relies on (#354)