Port Trinity-Nano-Base (afmoe)

.github/ci/benchmark_config.json

@@ -315,6 +315,9 @@
     },
     "smolvlm_500m": {
       "config": "ported_models/llama_cpp_et/benchmarks/smolvlm_500m.json"
+    },
+    "trinity_nano_base": {
+      "config": "ported_models/llama_cpp_et/benchmarks/trinity_nano_base.json"
     }
   }
 }

ported_models/llama_cpp_et/artifacts.json

@@ -485,6 +485,24 @@
       "sha256": "d1eb8b6b23979205fdf63703ed10f788131a3f812c7b1f72e0119d5d81295150",
       "size_bytes": 108783360,
       "note": "SmolVLM 500M vision projector (SigLIP ~93M + MLP). Q8_0 quantized. Must be loaded alongside smolvlm_500m_q8_gguf."
+    },
+    "trinity_nano_base_q8_gguf": {
+      "kind": "model",
+      "framework": "llama.cpp-et",
+      "variant": "Trinity-Nano-Base-Q8_0",
+      "filename": "trinity-nano-base-Q8_0.gguf",
+      "env": "TRINITY_NANO_BASE_MODEL_PATH",
+      "source": {
+        "type": "huggingface",
+        "repo": "darthceltic85/trinity-nano-base-gguf",
+        "revision": "1745c6f9d84546f60bae6e397e918ecfedc62cd3",
+        "filename": "trinity-nano-base-Q8_0.gguf",
+        "url": "https://huggingface.co/darthceltic85/trinity-nano-base-gguf/resolve/1745c6f9d84546f60bae6e397e918ecfedc62cd3/trinity-nano-base-Q8_0.gguf"
+      },
+      "sha256": "a52296d493fbdb810008a638eda4ce7f59717ddd4282ba2750518c7cefd36872",
+      "local_cache": "local-artifacts/models/trinity-nano-base-Q8_0.gguf",
+      "board_path": "/data/models/trinity-nano-base-Q8_0.gguf",
+      "note": "Self-converted (no pre-made Q8_0 GGUF existed); hosted on Hugging Face since the 6.52 GB file exceeds GitHub's 2 GB release-asset limit. Converted cleanly with the stock, unmodified converter -- AfmoeForCausalLM was already correctly supported, no fix needed."
     }
   }
 }

ported_models/llama_cpp_et/benchmarks/trinity_nano_base.json

@@ -0,0 +1,52 @@
+{
+  "runner": "llama_server",
+  "board": true,
+  "framework": {
+    "name": "llama.cpp-et",
+    "runner": "llama_server",
+    "source_artifact": "llama_cpp_source"
+  },
+  "artifacts_file": "../artifacts.json",
+  "canonical_variant": "Trinity-Nano-Base-Q8_0",
+  "score": {
+    "metric": "tokens_per_second",
+    "label": "Decode tokens/s",
+    "higher_is_better": true
+  },
+  "llama_server": {
+    "source_artifact": "llama_cpp_source",
+    "model_artifact": "trinity_nano_base_q8_gguf",
+    "server_artifact": "llama_server",
+    "workdir_artifact": "llama_cpp_build",
+    "host": "127.0.0.1",
+    "port": 18149,
+    "device": "ET",
+    "gpu_layers": 99,
+    "ctx_size": 2048,
+    "batch_size": 256,
+    "ubatch_size": 128,
+    "parallel": 1,
+    "cache_ram_mib": 0,
+    "ready_timeout_s": 300,
+    "request_timeout_s": 420,
+    "flash_attn": false,
+    "api": "completion",
+    "prompt": "Repeat this token sequence without commentary: OK OK OK OK OK OK OK OK OK OK",
+    "max_tokens": 96,
+    "temperature": 0,
+    "ignore_eos": true,
+    "min_completion_tokens": 32,
+    "perplexity": {
+      "enabled": true,
+      "perplexity_artifact": "llama_perplexity",
+      "corpus_artifact": "wikitext2_raw_test",
+      "ctx_size": 128,
+      "batch_size": 128,
+      "ubatch_size": 128,
+      "timeout_s": 420,
+      "min_ppl": 1.0,
+      "max_ppl": 1000.0,
+      "chunks": 4
+    }
+  }
+}

ported_models/submissions/model_ports/trinity_nano_base.json

@@ -0,0 +1,14 @@
+{
+  "schema_version": 1,
+  "track": "most_models_ported",
+  "benchmark_model": "trinity_nano_base",
+  "identity_id": "afmoe",
+  "source": {
+    "repo": "arcee-ai/Trinity-Nano-Base",
+    "revision": "923a95dbba013c78c4fd6dbc6bce31eea553da7c",
+    "license": "openmdw-1.1"
+  },
+  "implementation_paths": ["ported_models/trinity_nano_base"],
+  "benchmark_config": "ported_models/llama_cpp_et/benchmarks/trinity_nano_base.json",
+  "recipe": "ported_models/trinity_nano_base/docs/RECIPE.md"
+}

ported_models/trinity_nano_base/docs/RECIPE.md

@@ -0,0 +1,76 @@
+# Trinity-Nano-Base Porting Recipe
+
+## Overview
+
+Adds `arcee-ai/Trinity-Nano-Base` (fine-grained Mixture-of-Experts causal
+LM, 56 layers, 128 experts, 8 active per token + 1 shared expert, 131K
+context via sliding-window + full attention layer mix, 6.12B total
+params) to the `llama_cpp_et` benchmark suite. This introduces the
+**Afmoe** execution family to the board — Arcee AI's Trinity model
+line.
+
+## Model Reference
+
+- **Source**: `arcee-ai/Trinity-Nano-Base` (Hugging Face), revision
+  `923a95dbba013c78c4fd6dbc6bce31eea553da7c`
+- **License**: OpenMDW License Agreement, version 1.1 (`license_name:
+  openmdw-1.1`) — a genuinely permissive open license (no restriction
+  on use/modification/redistribution beyond retaining copyright
+  notices), confirmed by reading the actual `LICENSE` file in the
+  source repo rather than assumed.
+- **Architecture**: `arch = afmoe` (`AfmoeForCausalLM`), 56 layers,
+  embedding length 1024, 128 experts (8 active + 1 shared),
+  vocab_size 200192.
+
+## Conversion
+
+No fix was needed. `AfmoeForCausalLM` is already a correctly-registered
+architecture class in this repo's `convert_hf_to_gguf.py`
+(`AfmoeModel(LlamaModel)`), so this converted cleanly on the first
+attempt via the stock, unmodified converter: `convert_hf_to_gguf.py
+--outtype q8_0`. Produced a 1057-tensor, 6.52 GB file,
+`sha256=a52296d493fbdb810008a638eda4ce7f59717ddd4282ba2750518c7cefd36872`.
+
+## Hosting
+
+This GGUF (6.52 GB) exceeds GitHub's 2 GB release-asset limit, so it is
+hosted on Hugging Face: `darthceltic85/trinity-nano-base-gguf`, file
+`trinity-nano-base-Q8_0.gguf`.
+
+## Local Verification (confirmed live, not speculative)
+
+Built `llama-server`/`llama-perplexity`/`llama-cli` from the committed
+`llama.cpp-et` submodule (CPU backend — see the `jamba_tiny_dev` recipe
+for the verification-tier note on why CPU, not full ET sysemu) and ran
+real inference:
+
+- Model loads cleanly: `arch = afmoe`, 128 experts / 8 used, sliding
+  window + full attention mix (`is_swa_any = 1`), fused Gated Delta Net
+  (autoregressive + chunked) enabled, clean 4223-node compute graph, 1
+  split.
+- Real perplexity run against WikiText-2 raw (4 chunks, ctx=128,
+  batch=128): **PPL = 10.5908 +/- 1.92727** — a good result, well
+  within this campaign's normal range, confirming genuine coherent
+  output.
+
+## Committed deterministic oracle (added per maintainer review)
+
+`ported_models/trinity_nano_base/oracle/perplexity_oracle.json` commits
+the exact reproduction command, pinned corpus/artifact hashes, per-chunk
+and final PPL from this session's CPU reference run, and an explicit
+±20% comparison threshold for independently verifying a future
+full-offload ET-SoC1 run against this reference.
+
+## Instructions for Reproduction
+
+```bash
+python3 -c "from huggingface_hub import snapshot_download; print(snapshot_download('arcee-ai/Trinity-Nano-Base'))"
+# from the llama.cpp-et submodule root:
+python3 convert_hf_to_gguf.py <snapshot-dir> --outfile trinity-nano-base-Q8_0.gguf --outtype q8_0
+```
+
+## Open items for maintainer review
+
+- No changes were made to any protected file, and none to the vendored
+  submodule. No standalone wrapper was needed for this port — the
+  converter already handles this architecture correctly.

ported_models/trinity_nano_base/oracle/perplexity_oracle.json

@@ -0,0 +1,26 @@
+{
+  "oracle_type": "perplexity_threshold",
+  "model_artifact": {
+    "repo": "darthceltic85/trinity-nano-base-gguf",
+    "revision": "1745c6f9d84546f60bae6e397e918ecfedc62cd3",
+    "filename": "trinity-nano-base-Q8_0.gguf",
+    "sha256": "a52296d493fbdb810008a638eda4ce7f59717ddd4282ba2750518c7cefd36872"
+  },
+  "corpus": {
+    "artifact": "wikitext2_raw_test",
+    "sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08"
+  },
+  "command": "llama-perplexity --model trinity-nano-base-Q8_0.gguf -f wiki.test.raw -c 128 -b 128 -ub 128 --chunks 4",
+  "reference_run": {
+    "per_chunk_ppl": [8.2549, 10.5577, 10.8174, 10.5908],
+    "final_ppl": 10.5908,
+    "final_ppl_stderr": 1.92727,
+    "measured_on": "CPU (ggml-cpu backend, GGML_ET=OFF build)",
+    "measured_date": "2026-07-26"
+  },
+  "comparison_threshold": {
+    "metric": "final_ppl",
+    "max_relative_deviation": 0.20,
+    "note": "Matches this repo's own leaderboard-gate policy (PPL must stay within 20% of best-seen value). A full-offload ET-SoC1 re-run against this exact command/corpus/artifact should land at final_ppl within [8.47, 12.71] to be considered consistent with this reference run. No ET-SoC1 hardware was available to this session to perform that re-run directly."
+  }
+}