Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 23 additions & 0 deletions docs/arch-gaps.md
Original file line number Diff line number Diff line change
Expand Up @@ -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,
Expand Down
5 changes: 5 additions & 0 deletions registry/significant.json
Original file line number Diff line number Diff line change
Expand Up @@ -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"
}
}
}
12 changes: 12 additions & 0 deletions tests/census_watch_test.py
Original file line number Diff line number Diff line change
Expand Up @@ -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.
Expand Down
Loading