E3 - qwen3_5_moe, lfm2, nemotron_h
Remaining non-dense generate families: MoE+linear-attn (qwen3_5_moe) and hybrid conv/SSM (lfm2, nemotron_h).
Spec: doc/plans/engine-inference-core.md § E3
Part of: #56
Depends on: E1 (#104 ) for qwen3_5_moe MoE block; hybrid cache is E3-local (#117 )
Tracking notes folded into sub-issues
Sub-issues
Suggested order
E3: Hybrid cache (SSM/conv) + unified hybrid support gate #117 hybrid cache + gate unification
E3: qwen3_5_moe (linear attention + MoE FFN) #118 / E3: lfm2 (short-conv + attention hybrid) #119 / E3: nemotron_h (mamba-style SSM + attention/MLP) #120 family forwards (parallelizable after E3: Hybrid cache (SSM/conv) + unified hybrid support gate #117 ; E3: qwen3_5_moe (linear attention + MoE FFN) #118 also needs E1)
E3: Parity gates (qwen3_5_moe, lfm2, nemotron_h) #121 parity as each family lands
Acceptance
Each family generates on a canonical small checkpoint (skip-if-absent)
Hybrid cache correctly hosts SSM/conv entries alongside KV entries
Pure-dense rejects no longer use the Stage D triple-gate once hybrid is first-class
E3 - qwen3_5_moe, lfm2, nemotron_h
Remaining non-dense generate families: MoE+linear-attn (qwen3_5_moe) and hybrid conv/SSM (lfm2, nemotron_h).
Spec:
doc/plans/engine-inference-core.md§ E3Part of: #56
Depends on: E1 (#104) for qwen3_5_moe MoE block; hybrid cache is E3-local (#117)
Tracking notes folded into sub-issues
weights_expected_names()per new family (pattern:qwen3_layer_matmuls/qwen3_layer_normsinweights.c) - owned by each family issueengine_model_check_supportedwhen linear attention lands (PR feat(engine): stage D6 - qwen3_5 dense full-attn deltas #99 finding 6) - owned by E3: Hybrid cache (SSM/conv) + unified hybrid support gate #117Sub-issues
MODEL_QWEN3_5_MOE)Suggested order
Acceptance