diff --git a/docs/arch-gaps.md b/docs/arch-gaps.md index eecaf3988..bd50f00ef 100644 --- a/docs/arch-gaps.md +++ b/docs/arch-gaps.md @@ -1127,6 +1127,29 @@ watcher already classes it as remote code) and needs no entry. (`attn_res_mixer`, `attn_res_mlp`, `res_out`), and the GQA layers have a fused `qg_proj` and a value-residual `v_mix`. No pinned graph has those tensors. Reported, not alerted (`"watch": "report"`). +## 2026-11-13: `Skylar2ForCausalLM` — reviewed, not an alias (engine#2367) + +The one `!! UNCOVERED` line of the 2026-11-13 watcher run. It was read from the repo's own `config.json` and the +safetensors header (`model.safetensors`, tensor names and shapes), not from the class name. The run's two +`REMOTE-CODE` lines (`ParenthesesForCausalLM`, `Spark2_5ForCausalLM`) are informational, and the `UNVERIFIABLE` lines +are retried by the watcher, so neither needs an entry. + +* `Skylar2ForCausalLM` (`model_type` `skylar2`, 1 model, `Skyl4r-Ai/Skylar2-1B-Instruct`) is "Skylar 2". It is a + from-scratch 1B model that transformers loads only after `import skylar` from its own pip package: the repo has no + `auto_map` and no modelling code. It has 36 layers, hidden 1536 and vocab 64000, in a 3:1 hybrid: + * 27 KDA linear-attention layers, with `A_log`, `dt_bias`, per-q/k/v short convs (`q_conv`, `k_conv`, `v_conv`) + and low-rank gates (`f_a`/`f_b`, `g_a`/`g_b`); + * 9 GQA layers (12 heads / 4 KV, `qk_norm`), each with a per-head output gate (`W_gate`, 12 x 1536). + + Every norm is a gated RMSNorm with a rank-16 `gate_up`/`gate_down`. Every attention and FFN sublayer has an + attention-residual mixer (`res_attn`/`res_ffn`, a `query` and a `key_weight`), and so does the output + (`res_final`). The FFN uses a `situ_glu` activation (`situ_beta1` 4, `situ_beta2` 25). + + **It is not an alias of `KimiLinearForCausalLM`**: that class has a different KDA gate, no gated norms and no + residual mixers. Nor is it Kimi K3, which llama.cpp #327 adds: K3 shares the situ activation and attention residuals, + but has MLA attention, a full-rank KDA gate and a latent MoE where Skylar 2 has GQA, a low-rank gate and a dense + FFN. No pinned graph has these tensors. Reported, not alerted (`"watch": "report"`). + ## FLM 1.0.12's build on the box: one family root with nothing, three that are already somewhere (engine#1787) F9 of `TNPU-ALL` §9 asks for a row for *every family FastFlowLM's build serves and we do not*. The build is on the box, diff --git a/registry/significant.json b/registry/significant.json index 05f78cc4b..9c1a46a63 100644 --- a/registry/significant.json +++ b/registry/significant.json @@ -187,6 +187,11 @@ "model_type": "oleander", "family": "Oleander-150M, a from-scratch training export (first seen: CyanMonkey/Test-Subject-3; its own config says step 2000 of 28820, 6.94 %). Hybrid 28 layers in a gdn2,gdn2,gdn2,gqa pattern, hidden 576, vocab 32770, F32. Ships no modelling code and no auto_map. Not an alias of the Qwen3.5 / Qwen3-Next linear-attention families despite the same layer_types: gdn2 layers carry per-q/k/v conv1d and low-rank f/g gate projections, every sublayer has attention-residual mixers (attn_res_*), and GQA layers a fused qg_proj and a value-residual v_mix; no pinned graph has those tensors.", "watch": "report" + }, + "Skylar2ForCausalLM": { + "model_type": "skylar2", + "family": "Skylar 2 (first seen: Skyl4r-Ai/Skylar2-1B-Instruct, the 2026-11-13 census), a from-scratch 1B hybrid that loads through its own `skylar` pip package (no auto_map, no modelling code in the repo). 36 layers, hidden 1536, vocab 64000: 27 KDA linear-attention layers (A_log, dt_bias, per-q/k/v short conv, low-rank f_a/f_b and g_a/g_b gates) and 9 GQA layers (12 heads / 4 KV, qk_norm, a per-head output gate W_gate) in a 3:1 pattern; every norm is a gated RMSNorm with rank-16 gate_up/gate_down, every sublayer has an attention-residual mixer (res_attn/res_ffn query + key_weight), and the FFN uses a situ_glu activation. Not an alias of KimiLinearForCausalLM (different KDA gate, norms and residual mixers) nor of Kimi K3 in llama.cpp #327, which shares situ and attention residuals but uses MLA attention, a full-rank KDA gate and a latent MoE. No pinned graph has these tensors.", + "watch": "report" } } } diff --git a/tests/census_watch_test.py b/tests/census_watch_test.py index 5a56cbccc..37773e893 100644 --- a/tests/census_watch_test.py +++ b/tests/census_watch_test.py @@ -964,6 +964,18 @@ def _fake_urlopen(req, timeout=0): check("engine#1922 entry records its model_type", sig.get(_DS, {}).get("model_type") == "deepseek_v4_flash_superexpert") +# engine#2367: the 2026-11-13 run's one UNCOVERED class, read from its config and safetensors header. Skylar 2 is a +# KDA + GQA hybrid with gated norms and attention-residual mixers that no pinned graph has: declined, report-only. +_SK = "Skylar2ForCausalLM" +check("engine#2367 Skylar2ForCausalLM is reviewed report-only and documented", + sig.get(_SK, {}).get("watch") == "report" and sig[_SK].get("model_type") == "skylar2" + and _SK not in archs and _SK in gap_docs) +rc, res = rc_of([model("Skyl4r-Ai/Skylar2-1B-Instruct", _SK)], archs, sig) +check("engine#2367 the run's upload -> exit 0, listed as reviewed, no alert", + rc == 0 and res["reviewed_report"] == [_SK] and res["significant"] == []) +_, out = text_of([model("Skyl4r-Ai/Skylar2-1B-Instruct", _SK)], archs, sig) +check("engine#2367 printed as REVIEWED, not UNCOVERED", ". REVIEWED " + _SK in out and "UNCOVERED" not in out) + # 27. engine#2171: an answered gap stays answered. CI keeps no state, so every scheduled run sees the whole # FastFlowLM org as new; before registry/flm_answers.json each family docs/npu.md had already answered # printed `!! NPU FLM` / `!! NEW FLM` again and re-filed the census issue the day after it closed.