Mamba2 SSD core is unported: KERNEL-SSM-MAMBA has only the GDN arm, so every hybrid-SSM arch (Nemotron-H/3.5, Bamba, Granite-4, Zamba2, Falcon-H1, Mamba2) is blocked

[localai-bot]

Gap

KERNEL-SSM-MAMBA is INVENTORIED. Everything that has landed under it is
GDN-specific — the gated-delta-rule recurrence, its causal conv, its
conv-state gather/scatter, and the sm_120 tiling campaign. The row's own matrix
text says so: "these GDN tests do not prove generic Mamba support"
(.agents/kernel-matrix.md:157).

The Mamba2 SSD core is absent. include/vt/ops.h carries
kCausalConv1dFwd / kCausalConv1dUpdate / kCausalConv1dSpecUpdate,
kRmsNormGated, kGdnPrefill / kGdnDecode / kGdnPackedDecode,
kKdaGatedDeltaRule — and no chunked state-space scan and no selective
state update. Gated delta rule is not Mamba2: GDN carries the
(I - beta k kᵀ) delta-removal term, Mamba2's SSD is a diagonally-decayed
gated linear recurrence driven by A_log / dt / D. Different kernel,
same surrounding machinery.

Why now

nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 (published 2026-08-04,
20.1 GiB, fits one GB10) is NemotronHForCausalLM — 52 layers, 23 Mamba2 +
6 GQA attention + 23 MoE, with 64 mamba heads x head_dim 64, n_groups=8,
ssm_state_size=128, conv_kernel=4, chunk_size=128. The pinned oracle
already runs it (registry.py:179 -> nemotron_h.py; MambaMixer2 at
nemotron_h.py:373; NemotronHMoEDecoderLayer at nemotron_h.py:317;
ModelOptMixedPrecisionConfig at quantization/modelopt.py:2280), so this is
purely our gap, and the SSD core is the part that is not reachable by
composing what we already have.

This row is the shared unblocker, not a one-model detour. Every
INVENTORIED hybrid-SSM arch in .agents/model-matrix.md waits on the same
kernel: NemotronHForCausalLM (:248), Mamba2ForCausalLM (:234),
GraniteMoeHybridForCausalLM (:214), FalconH1ForCausalLM (:194),
Zamba2ForCausalLM (:280), JambaForCausalLM (:227), Lfm2ForCausalLM
(:229), Lfm2MoeForCausalLM (:230), Plamo2ForCausalLM (:264),
OlmoHybridForCausalLM (:251).

Upstream anchors (pinned oracle 555967922, vLLM 0.26.0.dev0)

• vllm/model_executor/layers/mamba/mamba_mixer2.py — MambaMixer2 (1170 lines)
• vllm/model_executor/layers/mamba/ops/ssd_combined.py
• vllm/model_executor/layers/mamba/ops/ssd_chunk_scan.py
• vllm/model_executor/layers/mamba/ops/ssd_chunk_state.py
• vllm/model_executor/layers/mamba/ops/ssd_state_passing.py
• vllm/model_executor/layers/mamba/ops/ssd_bmm.py
• vllm/model_executor/layers/mamba/ops/mamba_ssm.py (selective_state_update)
• vllm/model_executor/layers/mamba/ops/layernorm_gated.py
• vllm/model_executor/layers/mamba/mamba_utils.py

What is already ours and gets reused

MambaSpec + the hybrid KV coordinator (porting-inventory.md:78-79,109),
all three causal-conv1d arms including the spec-decode variant,
kRmsNormGated, kGdnStateGather / kGdnStateScatter, and the het-KV group
construction the gate models already exercise. The conv half of a Mamba2 layer
is done; the recurrence half is not.

Scope of this row

The kernel only: portable host reference for the SSD chunked scan and the
single-token selective state update, unit-gated against a from-first-principles
double-precision reference, then the CUDA arm, then wiring through the existing
MambaSpec state group. The Nemotron-3.5 model port (non-gated relu-squared
MoE grouped GEMM, ModelOpt MIXED_PRECISION per-module loading, the MTP head)
is a separate row on
MODEL-TEXT-nemotron-h-nemotron-hfor-causal-lm and is explicitly out of scope
here.

Spec: .agents/specs/mamba2-ssd.md (linked when the spec commit lands).

[localai-bot]

W0 landed: the spec is committed at .agents/specs/mamba2-ssd.md via #502 (d6b868c6). The row stays INVENTORIED — no lifecycle move, no measurement, no product code.

Records now agree in three places: the roadmap issue table, kernel-matrix.md (KERNEL-SSM-MAMBA checkpoint naming the SSD gap), and model-matrix.md (the NemotronH row records this as its blocker plus the three further things that arch owes). docs/FEATURES.md gains an "Inventoried but blocked" entry, because Nemotron-H is capability-blocked rather than hardware-blocked and the page should not imply we are waiting on a bigger box.

Next: W1, for a fresh implementer, RED first — portable CPU references for vt::Mamba2ChunkScan, vt::Mamba2StateUpdate and vt::RmsNormGatedGroup, gated by chunked-equals-sequential against an independent double-precision recurrence and by chunk-boundary invariance across chunk_size in {8,16,32,64,128}.