Skip to content
Closed
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
53 changes: 50 additions & 3 deletions docs/arch-gaps.md
Original file line number Diff line number Diff line change
Expand Up @@ -1041,7 +1041,10 @@ takes the reviewed set: such a pack prints `. FLM ... registry: none (reviewed,
Still open, and alerting on purpose: `MiniCPMV4_7ForConditionalGeneration` (needs one converter registration, not an
architecture), `MiniCPMOcrForConditionalGeneration` (own `minicpm-ocr` library), `smolvla-NPU2` (no config at all), and
the `!! NPU FLM` families whose registry mapping is HRX-only (`hy_v3`, `dflash`, `gemma4`, `hunyuan-dense`, `gemma3`) —
that last group is engine #588 (the family must be served on the NPU route), not a registry edit.
that last group is **not a registry edit**: the family has to be served on the NPU route. It used to be recorded here as
engine #588, which is **closed** (2026-10-27) and was about *watching FLM's releases*, not about serving a family on the
route. The issue that does own this class is engine#1786 (every `FastFlowLM/*-NPU2` source serves on `--device npu`,
opened 2026-11-10). See the last section of this page for the arrivals that are sitting in it.

## Census run 2026-11-09 (`engine #1667`)

Expand Down Expand Up @@ -1072,9 +1075,11 @@ Left as they are, on purpose, under the policy above: the six `. REMOTE-CODE` cl
(`HajeriOperatorForCausalLM`, `MM200ForCausalLM`, `K2HorizonForCausalLM`, `KLMForCausalLM`,
`KimiK31SuperExpertForCausalLM`, `ThalamusReasoningQwen3_5ForConditionalGeneration`); the `? UNVERIFIABLE` gated repos,
which are retried on every run; and the `!! FLM` lines. Those keep the 2026-11-08 reasons: `MiniCPMV4_7`, `MiniCPMOcr`
and `smolvla` as above, and `!! NPU FLM` is engine #588. This run's issue window also includes the `gpt-oss`
and `smolvla` as above, and `!! NPU FLM` was recorded as engine #588 — which is **closed** (2026-10-27) and was
about watching FLM's releases, not about serving a family on the NPU route; the class is owned by engine#1786
(corrected 2026-11-11; see the section at the end of this page). This run's issue window also includes the `gpt-oss`
(`GPT-OSS-20B`, `GPT-OSS-Safeguard-20b`) and `lfm2` (`LFM2-1.2B`, `LFM2-2.6B`) Q4NX packs. Their mapping is HRX-only
too, so they belong to #588, not to a registry edit.
too, so they are the same class — a route to serve, not a registry edit.

## 2026-11-10: `DeepseekV4FlashSuperExpertForCausalLM` — reviewed, and recorded here

Expand All @@ -1087,3 +1092,45 @@ the single undocumented entry out of 37.
MLA-style geometry (`kv_lora_rank` 512, `q_lora_rank` 1536, `qk_rope_head_dim` 64, `v_head_dim` 128, 32 heads /
16 KV, IM 4096). **Not an alias of `deepseek_v41`**: nothing at the pin decodes AW-SECD weights, so mapping it to
that architecture would name a backend that cannot read the file. 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,
so the checklist was enumerated from it rather than from the release notes:

```
$ ls ~/flm-f9/xclbins | wc -l -> 54 (per-family xclbin directories)
$ ls ~/flm-f9/lib/*.so | wc -l -> 34 (per-family shared objects)
```

A word on the name: #1787 calls this "FLM 1.0.12's HRX build", and the section used to. What is counted here is a
directory of **xclbins, every one named `*-NPU2`**: they are NPU artifacts. A reader looking for GPU families will not
find them in this list.

Mapping each of the 54 by **family root** (not by model name: `Gemma3-1B` normalises to `gemma31b`, which appears nowhere,
and matching that way reported 53 of 54 as unknown) against `registry/*.json`, `docs/registry.md` and this page gave 48
with a trace and 6 xclbins over 4 roots without one. **That second number was wrong** (review of engine#2065): the root
was compared as a literal against class names and `model_type`s that do not contain it (`nanbeige4.1` against
`NanbeigeForCausalLM`, `medgemma` against `gemma3`). Each of the four, searched on `main` case-insensitively and
restated 2026-11-11:

| xclbin directory | family root | what `main` has | what is actually owed |
|---|---|---|---|
| `Translategemma-4B-Instruct-NPU2` | `translategemma` | nothing: `git grep -il translategemma` finds 0 files | everything. This is the one root with no trace. By its name it would be a `gemma3` arrival like `medgemma`; nobody has opened its config, so that is a guess and not a mapping |
| `LFM2.5-1.2B-NPU2`, `LFM2.5-1.2B-Thinking-NPU2` | `lfm2.5` | the packs' own `config.json` says `Lfm2ForCausalLM` / `model_type: lfm2`, which the registry maps, so there is no "2.5 class" to add; `models/pins.json` pins the `LFM2.5-2.6B` and `LFM2.5-8B-A1B` GGUFs; and since engine#2170 both packs **serve on the NPU route** with measured rows in [npu.md](npu.md) ("FastFlowLM's own containers") | nothing in the registry. On the route: speed (per-op, about 4 tok/s against their 48 to 83), not coverage |
| `Medgemma-4B-NPU2`, `Medgemma-1.5-4B-NPU2` | `medgemma` | a `gemma3` arrival, already routed: [npu.md](npu.md) has the row `gemma3` / `medgemma-1.5-4b-it-NPU2` / hrx / not served, and `tests/census_watch_test.py` pins that pack to `Gemma3ForCausalLM` | an NPU row. `Medgemma-4B-NPU2` is on the box; its geometry (H 2560, IM 10240) needs two kernel shapes that are not built (`GU K2560 N20480`, `D K10240 N2560`), so it is a kernel-set item, not a model-type one |
| `Nanbeige4.1-3B-NPU2` | `nanbeige4.1` | mapped: `registry/architectures.json` has `NanbeigeForCausalLM` (gguf `nanbeige`), `registry/census.json` counts it, `registry/support_pins.json` lists its converter config, and `q4nx_semantics.h` and [tokenizers.md](tokenizers.md) name `Nanbeige4.1-3B` itself (11 files) | an NPU row. Nothing on [npu.md](npu.md) says whether it serves there, and it has not been tried |

So: **one** family root with nothing (`translategemma`), and three that are mapped or served and owe a measured NPU row
or nothing at all. The count of 48 "with a trace" was spot-checked, not re-derived, and after this correction it should
be read as "at least 51".

**The rows that are still owed are un-inspected, and say so on purpose.** Nothing above describes what `translategemma`
or `nanbeige4.1` computes: an xclbin name is evidence that *their* build carries a compiled family for it, and that is
all. §1's five conditions are what converts an entry from *listed* to *settled* — imports with zero gaps and serving on
the device, a known-answer prompt under the model's own template with three identical answers, decode tok/s and TTFT
under a ticket, and the kernel set and engine commit named. The two `lfm2.5` packs are the example of an entry that
went from listed to settled that way.

The `!! NPU FLM` class above is where the owed rows belong. Its owning issue is engine#1786 (every `*-NPU2` source
serves), which is open.
Loading