feat(model-ports): add mamba_1_4b, chatglm3_6b

.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 +343,12 @@
     },
     "smolvlm_500m": {
       "config": "ported_models/llama_cpp_et/benchmarks/smolvlm_500m.json"
+    },
+    "mamba_1_4b": {
+      "config": "ported_models/llama_cpp_et/benchmarks/mamba_1_4b.json"
+    },
+    "chatglm3_6b": {
+      "config": "ported_models/llama_cpp_et/benchmarks/chatglm3_6b.json"
     }
   }
 }

ported_models/chatglm3_6b/docs/RECIPE.md

@@ -0,0 +1,55 @@
+# ChatGLM3-6B Porting Recipe
+
+## Overview
+
+Adds `zai-org/chatglm3-6b` (6B-parameter causal LM, Zhipu AI's ChatGLM3;
+this model was previously published under `THUDM/chatglm3-6b`, which now
+redirects to `zai-org`) to the `llama.cpp-et` framework. Confirmed via
+local GGUF metadata inspection: `general.architecture = chatglm` --
+distinct from `glm4` (already ported earlier in this campaign), a
+separate, older execution family in llama.cpp (ChatGLM uses a distinct
+attention-bias/RoPE convention from GLM4).
+
+## Model Reference
+
+- **Source**: `zai-org/chatglm3-6b` (Hugging Face), revision
+  `e9e0406d062cdb887444fe5bd546833920abd4ac`
+- **License**: Apache-2.0
+- **GGUF source**: `hellork/chatglm3-6b-128k-Q8_0-GGUF` (a 128k-context
+  fine-tune/extension of the base model, same architecture), file
+  `chatglm3-6b-128k-q8_0.gguf`
+- **Quantization**: Q8_0,
+  `sha256=805e3761a997486eb08f45e0f74efa5c1cc0be8475afa2b170263b490a3ba8c8`
+  (verified locally against the downloaded file)
+- **Architecture**: `arch = chatglm` per GGUF metadata, 199 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:
+
+```
+Final estimate: PPL = 15.9100 +/- 3.37450
+```
+
+## Why this port likely needs no new ET-SoC1 kernel work
+
+`LLM_ARCH_CHATGLM`'s graph builder uses standard ops already proven on the
+ET backend (RMSNorm, RoPE, MHA/GQA with a QKV bias term, GELU-based FFN) --
+no new op type. ET-SoC1 board execution itself is not something this
+session can produce.
+
+## Open items for maintainer review
+
+- Registered in `artifacts.json`, `ported_models/llama_cpp_et/benchmarks/chatglm3_6b.json`,
+  and `.github/ci/benchmark_config.json` (port 18140) -- board-testable now,
+  independent of the model-ports track claim below.
+- `ported_models/submissions/model_ports/chatglm3_6b.json` is the
+  model-ports track claim, pending identity approval.
+- No changes to any protected file or the vendored submodule.
+- GGUF used is a 128k-context fine-tune, not a base-model quant -- same
+  architecture and weights lineage, flagged for transparency.

ported_models/chatglm3_6b/oracle/perplexity_oracle.json

@@ -0,0 +1,25 @@
+{
+  "oracle_type": "perplexity_threshold",
+  "model_artifact": {
+    "repo": "hellork/chatglm3-6b-128k-Q8_0-GGUF",
+    "revision": "main",
+    "filename": "chatglm3-6b-128k-q8_0.gguf",
+    "sha256": "805e3761a997486eb08f45e0f74efa5c1cc0be8475afa2b170263b490a3ba8c8"
+  },
+  "corpus": {
+    "artifact": "wikitext2_raw_test",
+    "sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08"
+  },
+  "command": "llama-perplexity --model chatglm3-6b-128k-Q8_0.gguf -f wiki.test.raw -c 128 -b 128 -ub 128 --chunks 4",
+  "reference_run": {
+    "final_ppl": 15.91,
+    "final_ppl_stderr": 3.3745,
+    "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 [12.73, 19.09] 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/flan_t5_base/docs/RECIPE.md

@@ -0,0 +1,67 @@
+# FLAN-T5-base -- Inconclusive Result (No Claim Filed)
+
+## Summary
+
+Attempted to add `google/flan-t5-base` (250M-parameter encoder-decoder
+causal LM, an instruction-tuned T5) to the `llama.cpp-et` framework.
+**The model loads and its compute graph builds successfully, but neither of
+the two standard tools in this vendored `llama.cpp-et` submodule
+(`llama-perplexity`, `llama-cli`) can actually drive it to produce output.**
+No model-ports claim is filed -- this recipe documents the finding rather
+than a working port.
+
+## Model Reference
+
+- **Source**: `google/flan-t5-base` (Hugging Face), revision
+  `7bcac572ce56db69c1ea7c8af255c5d7c9672fc2`
+- **License**: Apache-2.0
+- **GGUF source**: `Felladrin/gguf-flan-t5-base`, file
+  `flan-t5-base.Q8_0.gguf`
+- **Quantization**: Q8_0,
+  `sha256=73eb18e2341b28f40e9596c3740ef46941075a57c79af1ebffab177fd66cbe50`
+  (verified locally against the downloaded file)
+- **Architecture**: `arch = t5` per GGUF metadata, 282 tensors.
+
+## What actually happens (both confirmed live, not speculative)
+
+The model loads cleanly and reserves a compute graph (`llama_kv_cache`
+allocated, 704 graph nodes) -- so far identical to every working port in
+this campaign. Then:
+
+1. **`llama-perplexity`** crashes with
+   `GGML_ASSERT(!llama_vocab_get_add_eos(vocab)) failed`. T5's tokenizer
+   always appends an EOS token (a fixed property of its seq2seq
+   vocabulary) -- the perplexity tool's raw-continuation PPL method
+   explicitly asserts the OPPOSITE for every model it evaluates, since
+   that method assumes a plain causal-LM tokenizer. This is a tool/model
+   mismatch, not a model defect.
+2. **`llama-cli`** crashes with
+   `GGML_ASSERT(!cross->seq_ids_enc.empty() && "llama_encode must be called
+   first") failed`. T5 is encoder-decoder: its cross-attention layers need
+   the encoder pass (`llama_encode()`) run once before any decoder step
+   (`llama_decode()`). `llama-cli` goes straight to `llama_decode()` --
+   it never calls `llama_encode()` at all. The underlying library
+   (`libllama`) clearly supports the encode/decode split (the assert
+   exists specifically to catch exactly this ordering mistake), but
+   neither CLI frontend in this vendored build orchestrates it.
+
+## Why no claim is filed
+
+Both failures happen in the harness/tooling, not the model itself -- the
+GGUF is valid and the graph builds. But this campaign's standard is
+"prove the model actually runs and produces output," and neither available
+tool can demonstrate that for an encoder-decoder model. Filing a claim on
+"it loads" alone, with the same evidence bar as this campaign's actually-
+completed ports, would overstate what was actually shown. This is a
+different category from `bitnet_2b` (which fails at model load) --
+here the model format is fine, but this repo's tooling has no
+encoder-decoder driver.
+
+## What would need to change
+
+A custom driver that calls `llama_encode()` once against the input, then
+loops `llama_decode()` for the output tokens (this is exactly what
+llama.cpp's upstream examples do for T5/text2text tasks in some example
+programs, but that driver is not present in the ported subset of tools
+built in this repo's vendored submodule). Out of scope for a model-port
+recipe -- would be new tooling work, not board wiring.

ported_models/llama_cpp_et/artifacts.json

@@ -485,6 +485,40 @@
       "sha256": "d1eb8b6b23979205fdf63703ed10f788131a3f812c7b1f72e0119d5d81295150",
       "size_bytes": 108783360,
       "note": "SmolVLM 500M vision projector (SigLIP ~93M + MLP). Q8_0 quantized. Must be loaded alongside smolvlm_500m_q8_gguf."
+    },
+    "mamba_1_4b_q6k_gguf": {
+      "kind": "model",
+      "framework": "llama.cpp-et",
+      "variant": "mamba-1.4b-hf-Q6_K",
+      "filename": "mamba-1.4b-hf-Q6_K.gguf",
+      "env": "MAMBA_1_4B_MODEL_PATH",
+      "source": {
+        "type": "huggingface",
+        "repo": "RichardErkhov/state-spaces_-_mamba-1.4b-hf-gguf",
+        "revision": "main",
+        "filename": "mamba-1.4b-hf.Q6_K.gguf",
+        "url": "https://huggingface.co/RichardErkhov/state-spaces_-_mamba-1.4b-hf-gguf/resolve/main/mamba-1.4b-hf.Q6_K.gguf"
+      },
+      "sha256": "a5dde81ce41c34213bf26bcaf2ba26cd1c6fde20fd20b5af6a4a76a5272478bc",
+      "local_cache": "local-artifacts/models/mamba-1.4b-hf-Q6_K.gguf",
+      "board_path": "/data/models/mamba-1.4b-hf-Q6_K.gguf"
+    },
+    "chatglm3_6b_q8_gguf": {
+      "kind": "model",
+      "framework": "llama.cpp-et",
+      "variant": "chatglm3-6b-128k-Q8_0",
+      "filename": "chatglm3-6b-128k-Q8_0.gguf",
+      "env": "CHATGLM3_6B_MODEL_PATH",
+      "source": {
+        "type": "huggingface",
+        "repo": "hellork/chatglm3-6b-128k-Q8_0-GGUF",
+        "revision": "main",
+        "filename": "chatglm3-6b-128k-q8_0.gguf",
+        "url": "https://huggingface.co/hellork/chatglm3-6b-128k-Q8_0-GGUF/resolve/main/chatglm3-6b-128k-q8_0.gguf"
+      },
+      "sha256": "805e3761a997486eb08f45e0f74efa5c1cc0be8475afa2b170263b490a3ba8c8",
+      "local_cache": "local-artifacts/models/chatglm3-6b-128k-Q8_0.gguf",
+      "board_path": "/data/models/chatglm3-6b-128k-Q8_0.gguf"
     }
   }
 }

ported_models/llama_cpp_et/benchmarks/chatglm3_6b.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": "chatglm3-6b-128k-Q8_0",
+  "score": {
+    "metric": "tokens_per_second",
+    "label": "Decode tokens/s",
+    "higher_is_better": true
+  },
+  "llama_server": {
+    "source_artifact": "llama_cpp_source",
+    "model_artifact": "chatglm3_6b_q8_gguf",
+    "server_artifact": "llama_server",
+    "workdir_artifact": "llama_cpp_build",
+    "host": "127.0.0.1",
+    "port": 18140,
+    "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
+    }
+  }
+}