v4 architecture realign: v4.1–v4.6 stack + trained SUT on GPU (merge gate FAILED — ship-blocking follow-up v4.7 identified)#30

Draft

FluffyAIcode wants to merge 16 commits intomainfrom

AgentMemory/v347-architecture-realign-b7fa

Commits on Apr 22, 2026

v4 architecture realign: skeleton (design + typed stubs, no runtime)

cursoragent
and
FluffyAIcode
committed
v4 impl spec: per-PR design for v4.1 through v4.5 (+ v4.6 non-scope)

cursoragent
and
FluffyAIcode
committed
v4.1: port geometry primitives + MemStore + DirectionTreeV4

cursoragent
and
FluffyAIcode
committed
v4.2: three encoders + three concrete Bundle subclasses

cursoragent
and
FluffyAIcode
committed
v4.3: KakeyaSet + KakeyaRegistry + alignment math

cursoragent
and
FluffyAIcode
committed
v4.4: BundleQueryHeads + CrossBundleAttention

cursoragent
and
FluffyAIcode
committed
v4.5: EmbBridge4 + LLMBackbone4 + MemLLM4 end-to-end + CPU smoke test

cursoragent
and
FluffyAIcode
committed
session_viability_v4: fresh-init parity harness for v4

cursoragent
and
FluffyAIcode
committed
v4.5: LLMBackbone4.load auto-moves model to cuda if available

cursoragent
and
FluffyAIcode
committed
v4.5: remove redundant prefix LN to fix bf16/fp32 mismatch on GPU

cursoragent
and
FluffyAIcode
committed
v4 fresh-init GPU SUT results (Qwen2.5-1.5B, NVIDIA H200, mt=30)

cursoragent
and
FluffyAIcode
committed
v4.6: Trainer4 + five loss terms + train_v4.py driver

cursoragent
and
FluffyAIcode
committed
session_viability_v4: add --trained-weights flag for v4.6 SUT

cursoragent
and
FluffyAIcode
committed
v4.6 fix: learnable prefix_scale in CrossBundleAttention

cursoragent
and
FluffyAIcode
committed
v4.6 fix: topic-axis loss — content-token Jaccard + diversity regularizer

cursoragent
and
FluffyAIcode
committed
v4.6 trained SUT results (H200, Qwen2.5-1.5B, 60 steps, prefix_scale + topic-diversity fixes applied)

cursoragent
and
FluffyAIcode
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

v4 architecture realign: v4.1–v4.6 stack + trained SUT on GPU (merge gate FAILED — ship-blocking follow-up v4.7 identified)#30

v4 architecture realign: v4.1–v4.6 stack + trained SUT on GPU (merge gate FAILED — ship-blocking follow-up v4.7 identified)#30
FluffyAIcode wants to merge 16 commits intomainfrom
AgentMemory/v347-architecture-realign-b7fa

Commits on Apr 22, 2026

v4 architecture realign: skeleton (design + typed stubs, no runtime)

v4 impl spec: per-PR design for v4.1 through v4.5 (+ v4.6 non-scope)

v4.1: port geometry primitives + MemStore + DirectionTreeV4

v4.2: three encoders + three concrete Bundle subclasses

v4.3: KakeyaSet + KakeyaRegistry + alignment math

v4.4: BundleQueryHeads + CrossBundleAttention

v4.5: EmbBridge4 + LLMBackbone4 + MemLLM4 end-to-end + CPU smoke test

session_viability_v4: fresh-init parity harness for v4

v4.5: LLMBackbone4.load auto-moves model to cuda if available

v4.5: remove redundant prefix LN to fix bf16/fp32 mismatch on GPU

v4 fresh-init GPU SUT results (Qwen2.5-1.5B, NVIDIA H200, mt=30)

v4.6: Trainer4 + five loss terms + train_v4.py driver

session_viability_v4: add --trained-weights flag for v4.6 SUT

v4.6 fix: learnable prefix_scale in CrossBundleAttention

v4.6 fix: topic-axis loss — content-token Jaccard + diversity regularizer

v4.6 trained SUT results (H200, Qwen2.5-1.5B, 60 steps, prefix_scale + topic-diversity fixes applied)