feat(model-ports): add falcon_h1_1_5b

.github/ci/benchmark_config.json

@@ -107,12 +107,16 @@
         },
         {
           "address": "0x02000000",
-          "paths": ["yolo/weights_region.bin"],
+          "paths": [
+            "yolo/weights_region.bin"
+          ],
           "required": true
         },
         {
           "address": "0x04A00000",
-          "paths": ["yolo/coco_room_000139_raw_480x640x3_uint8_rgb.bin"],
+          "paths": [
+            "yolo/coco_room_000139_raw_480x640x3_uint8_rgb.bin"
+          ],
           "required": true
         }
       ],
@@ -122,15 +126,21 @@
         "image_count": 5,
         "min_image_count": 5,
         "reference_contract": ".github/ci/reference/yolo.json",
-        "source_shape": [480, 640, 3]
+        "source_shape": [
+          480,
+          640,
+          3
+        ]
       },
       "benchmark_cases": [
         {
           "name": "coco_room_000139",
           "file_loads": [
             {
               "address": "0x04A00000",
-              "paths": ["yolo/coco_room_000139_raw_480x640x3_uint8_rgb.bin"],
+              "paths": [
+                "yolo/coco_room_000139_raw_480x640x3_uint8_rgb.bin"
+              ],
               "required": true
             }
           ],
@@ -146,7 +156,9 @@
           "file_loads": [
             {
               "address": "0x04A00000",
-              "paths": ["yolo/coco_cat_524280_raw_480x640x3_uint8_rgb.bin"],
+              "paths": [
+                "yolo/coco_cat_524280_raw_480x640x3_uint8_rgb.bin"
+              ],
               "required": true
             }
           ],
@@ -162,7 +174,9 @@
           "file_loads": [
             {
               "address": "0x04A00000",
-              "paths": ["yolo/coco_giraffes_296969_raw_480x640x3_uint8_rgb.bin"],
+              "paths": [
+                "yolo/coco_giraffes_296969_raw_480x640x3_uint8_rgb.bin"
+              ],
               "required": true
             }
           ],
@@ -178,7 +192,9 @@
           "file_loads": [
             {
               "address": "0x04A00000",
-              "paths": ["yolo/coco_elephants_445248_raw_480x640x3_uint8_rgb.bin"],
+              "paths": [
+                "yolo/coco_elephants_445248_raw_480x640x3_uint8_rgb.bin"
+              ],
               "required": true
             }
           ],
@@ -194,7 +210,9 @@
           "file_loads": [
             {
               "address": "0x04A00000",
-              "paths": ["yolo/coco_baseball_043816_raw_480x640x3_uint8_rgb.bin"],
+              "paths": [
+                "yolo/coco_baseball_043816_raw_480x640x3_uint8_rgb.bin"
+              ],
               "required": true
             }
           ],
@@ -234,17 +252,24 @@
       "file_loads": [
         {
           "address": "0x0",
-          "paths": ["zero2m.bin", "common/zero2m.bin"],
+          "paths": [
+            "zero2m.bin",
+            "common/zero2m.bin"
+          ],
           "required": true
         },
         {
           "address": "0x2000",
-          "paths": ["dncnn/dncnn20l64_input.bin"],
+          "paths": [
+            "dncnn/dncnn20l64_input.bin"
+          ],
           "required": true
         },
         {
           "address": "0x14000",
-          "paths": ["dncnn/dncnn20l64_weights.bin"],
+          "paths": [
+            "dncnn/dncnn20l64_weights.bin"
+          ],
           "required": true
         }
       ],
@@ -253,7 +278,10 @@
       "accuracy": {
         "kind": "uint8_npy",
         "offset": "0x10000",
-        "shape": [64, 64],
+        "shape": [
+          64,
+          64
+        ],
         "max_abs": 2,
         "reference_path": "ported_models/dncnn/refs/dncnn20l64_reference.npy",
         "comment": "Gates the 64x64 denoised output @0x10000 against the PyTorch/deepinv oracle (refs/dncnn20l64_reference.npy, produced by scripts/gen_dncnn_oracle.py running deepinv.models.DnCNN on the pinned weights). Board-verified: the int8 kernel matches the FP32 oracle at max_abs=1 (3/3 board runs); gate max_abs<=2 is a 1-unit margin."
@@ -315,6 +343,9 @@
     },
     "smolvlm_500m": {
       "config": "ported_models/llama_cpp_et/benchmarks/smolvlm_500m.json"
+    },
+    "falcon_h1_1_5b": {
+      "config": "ported_models/llama_cpp_et/benchmarks/falcon_h1_1_5b.json"
     }
   }
 }

ported_models/falcon_h1_1_5b/docs/RECIPE.md

@@ -0,0 +1,66 @@
+# Falcon-H1-1.5B-Instruct Porting Recipe
+
+## Overview
+
+Adds `tiiuae/Falcon-H1-1.5B-Instruct` (1.5B-parameter causal LM, TII's
+newer hybrid attention+SSM architecture) to the `llama.cpp-et` framework.
+Confirmed via local GGUF metadata inspection: `general.architecture =
+falcon-h1` -- distinct from `falcon` (already ported earlier in this
+campaign), a genuinely different architecture, not just a size variant:
+Falcon-H1 interleaves standard attention layers with Mamba-style recurrent
+layers in the same model (hybrid, not pure transformer or pure SSM).
+
+## Model Reference
+
+- **Source**: `tiiuae/Falcon-H1-1.5B-Instruct` (Hugging Face), revision
+  `80ebc50d7799a440b96c93bb6686a3924a09b0cb`
+- **License**: Falcon LLM License (`falcon-llm-license`,
+  https://falconllm.tii.ae/falcon-terms-and-conditions.html) -- a custom
+  TII license, not a standard OSI license; same family as falcon7b's
+  license in this campaign but a distinct license id
+- **GGUF source**: `tiiuae/Falcon-H1-1.5B-Instruct-GGUF` (official), file
+  `Falcon-H1-1.5B-Instruct-Q8_0.gguf`
+- **Quantization**: Q8_0,
+  `sha256=0044322ad5aaea5ccc2264f83f430f4ab2d20acc6db71a12d79b3f5acb0f69d5`
+  (verified locally against the downloaded file)
+- **Architecture**: `arch = falcon-h1` per GGUF metadata, 411 tensors.
+
+## Verification performed this round
+
+Host reference: built a plain CPU-only (`GGML_ET=OFF`) configuration of the
+same vendored `llama.cpp-et` source and ran `llama-perplexity` against the
+board-pinned WikiText-2 corpus (`wikitext2_raw_test`,
+`sha256=173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08`),
+context 128 / batch 128 / ubatch 128 / 4 chunks. The model loads and runs
+cleanly, allocating **both** a standard transformer KV cache (`llama_kv_cache`,
+256 cells / 24 layers) **and** a recurrent SSM state cache
+(`llama_memory_recurrent`, 1 cell / 24 layers) simultaneously -- direct
+confirmation this is a genuine hybrid, not mislabeled:
+
+```
+Final estimate: PPL = 11.6708 +/- 2.60807
+```
+
+This is the second SSM-adjacent model in this campaign (after
+`mamba_1_4b`) to confirm `SSM_CONV`/`SSM_SCAN` work on `ggml-cpu`, and the
+first to combine them with standard attention in one graph.
+
+## Why this port's ET-SoC1 kernel support is a real, open question
+
+Same caveat as `mamba_1_4b`: CPU-backend success does not prove the
+ET-SoC1 backend supports `SSM_CONV`/`SSM_SCAN`. This is a second, mostly
+independent data point on the same open question (different model, same
+underlying ops) -- if either mamba_1_4b or this port ever gets a real
+board run, it directly answers whether ET-SoC1 can run recurrent-state
+architectures at all.
+
+## Open items for maintainer review
+
+- Registered in `artifacts.json`, `ported_models/llama_cpp_et/benchmarks/falcon_h1_1_5b.json`,
+  and `.github/ci/benchmark_config.json` (port 18141) -- board-testable now,
+  independent of the model-ports track claim below.
+- `ported_models/submissions/model_ports/falcon_h1_1_5b.json` is the
+  model-ports track claim, pending identity approval.
+- No changes to any protected file or the vendored submodule.
+- SSM ops (`SSM_CONV`/`SSM_SCAN`) confirmed on CPU, NOT live-verified
+  against ET sysemu specifically.

ported_models/falcon_h1_1_5b/oracle/perplexity_oracle.json

@@ -0,0 +1,25 @@
+{
+  "oracle_type": "perplexity_threshold",
+  "model_artifact": {
+    "repo": "tiiuae/Falcon-H1-1.5B-Instruct-GGUF",
+    "revision": "main",
+    "filename": "Falcon-H1-1.5B-Instruct-Q8_0.gguf",
+    "sha256": "0044322ad5aaea5ccc2264f83f430f4ab2d20acc6db71a12d79b3f5acb0f69d5"
+  },
+  "corpus": {
+    "artifact": "wikitext2_raw_test",
+    "sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08"
+  },
+  "command": "llama-perplexity --model Falcon-H1-1.5B-Instruct-Q8_0.gguf -f wiki.test.raw -c 128 -b 128 -ub 128 --chunks 4",
+  "reference_run": {
+    "final_ppl": 11.6708,
+    "final_ppl_stderr": 2.60807,
+    "measured_on": "CPU (ggml-cpu backend, GGML_ET=OFF build)",
+    "measured_date": "2026-07-25"
+  },
+  "comparison_threshold": {
+    "metric": "final_ppl",
+    "max_relative_deviation": 0.2,
+    "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 [9.34, 14.00] to be considered consistent with this reference run. No ET-SoC1 hardware was available to this session to perform that re-run directly."
+  }
+}

ported_models/llama_cpp_et/artifacts.json

@@ -485,6 +485,23 @@
       "sha256": "d1eb8b6b23979205fdf63703ed10f788131a3f812c7b1f72e0119d5d81295150",
       "size_bytes": 108783360,
       "note": "SmolVLM 500M vision projector (SigLIP ~93M + MLP). Q8_0 quantized. Must be loaded alongside smolvlm_500m_q8_gguf."
+    },
+    "falcon_h1_1_5b_q8_gguf": {
+      "kind": "model",
+      "framework": "llama.cpp-et",
+      "variant": "Falcon-H1-1.5B-Instruct-Q8_0",
+      "filename": "Falcon-H1-1.5B-Instruct-Q8_0.gguf",
+      "env": "FALCON_H1_1_5B_MODEL_PATH",
+      "source": {
+        "type": "huggingface",
+        "repo": "tiiuae/Falcon-H1-1.5B-Instruct-GGUF",
+        "revision": "main",
+        "filename": "Falcon-H1-1.5B-Instruct-Q8_0.gguf",
+        "url": "https://huggingface.co/tiiuae/Falcon-H1-1.5B-Instruct-GGUF/resolve/main/Falcon-H1-1.5B-Instruct-Q8_0.gguf"
+      },
+      "sha256": "0044322ad5aaea5ccc2264f83f430f4ab2d20acc6db71a12d79b3f5acb0f69d5",
+      "local_cache": "local-artifacts/models/Falcon-H1-1.5B-Instruct-Q8_0.gguf",
+      "board_path": "/data/models/Falcon-H1-1.5B-Instruct-Q8_0.gguf"
     }
   }
 }

ported_models/llama_cpp_et/benchmarks/falcon_h1_1_5b.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": "Falcon-H1-1.5B-Instruct-Q8_0",
+  "score": {
+    "metric": "tokens_per_second",
+    "label": "Decode tokens/s",
+    "higher_is_better": true
+  },
+  "llama_server": {
+    "source_artifact": "llama_cpp_source",
+    "model_artifact": "falcon_h1_1_5b_q8_gguf",
+    "server_artifact": "llama_server",
+    "workdir_artifact": "llama_cpp_build",
+    "host": "127.0.0.1",
+    "port": 18141,
+    "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/persimmon_8b/docs/RECIPE.md

@@ -0,0 +1,59 @@
+# Persimmon-8B -- Negative Result
+
+## Summary
+
+Attempted to add `adept/persimmon-8b-base` (8B-parameter causal LM, Adept's
+Persimmon architecture) to the `llama.cpp-et` framework. **This repo's
+vendored `llama.cpp-et` submodule has no `persimmon` architecture support
+compiled in at all -- the model fails to load with `unknown model
+architecture: 'persimmon'`.** No model-ports claim is filed.
+
+## Model Reference
+
+- **Source**: `adept/persimmon-8b-base` (Hugging Face), revision
+  `94dc4e0bb7eeb26ec521eb3f78c36c91f6fe866b`
+- **License**: Apache-2.0
+- **GGUF source**: `maddes8cht/adept-persimmon-8b-base-gguf`, file
+  `adept-persimmon-8b-base-Q8_0.gguf`
+- **Quantization**: Q8_0, 9.30 GiB,
+  `sha256=25eb5e4901cd45ae03c9a447bb49c7716a71a26562d1a5a3b291de232c27b0c4`
+  (verified locally against the downloaded file)
+- GGUF metadata (read successfully -- the failure is architecture support,
+  not file corruption): `general.architecture = persimmon`, 580 tensors,
+  36 layers, 4096 embedding, 64 attention heads (no GQA -- head_count ==
+  head_count_kv), 16384-token context, RoPE freq_base 25000.
+
+## Root cause: confirmed, not speculative
+
+Built a plain CPU-only (`GGML_ET=OFF`) configuration of the exact vendored
+`llama.cpp-et` source in this repo and tried to load the file with
+`llama-perplexity`. The GGUF header parses completely (llama.cpp's own
+metadata dump above proves the file itself is well-formed), but model
+construction fails immediately after:
+
+```
+llama_model_load: error loading model: error loading model architecture:
+unknown model architecture: 'persimmon'
+llama_model_load_from_file_impl: failed to load model
+```
+
+Upstream llama.cpp has historically had Persimmon support (it was one of
+the earlier non-Llama architectures added), but this repo's vendored fork
+either predates that support or has since dropped it. Either way, the
+architecture registry `llama_model_load()` dispatches on simply does not
+contain an entry for `persimmon` in this exact vendored commit.
+
+## Why no claim is filed
+
+The failure is unambiguous and reproducible -- not a tooling gap like
+`flan_t5_base` (where the model loads but the CLI can't drive it) and not
+an exotic-quant-type collision like `bitnet_2b` (where the file itself
+gets rejected). This is the simplest and most direct kind of negative
+result: the architecture is not implemented in this vendored source at
+all.
+
+## What would need to change
+
+Persimmon support would need to be added to (or restored in) this repo's
+vendored `llama.cpp-et` fork -- a submodule content change, out of scope
+for a model-port PR.

ported_models/submissions/model_ports/falcon_h1_1_5b.json

@@ -0,0 +1,16 @@
+{
+  "schema_version": 1,
+  "track": "most_models_ported",
+  "benchmark_model": "falcon_h1_1_5b",
+  "identity_id": "falcon-h1",
+  "source": {
+    "repo": "tiiuae/Falcon-H1-1.5B-Instruct",
+    "revision": "80ebc50d7799a440b96c93bb6686a3924a09b0cb",
+    "license": "falcon-llm-license"
+  },
+  "implementation_paths": [
+    "ported_models/falcon_h1_1_5b"
+  ],
+  "benchmark_config": "ported_models/llama_cpp_et/benchmarks/falcon_h1_1_5b.json",
+  "recipe": "ported_models/falcon_h1_1_5b/docs/RECIPE.md"
+}