Add files using upload-large-folder tool

.gitattributes

@@ -1,35 +1,16 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+*.gguf filter=lfs diff=lfs merge=lfs -text
 *.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.dll filter=lfs diff=lfs merge=lfs -text
+*.exe filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+

.gitignore

@@ -0,0 +1,7 @@
+.cache/
+__pycache__/
+*.pyc
+*.stderr.log
+*.stdout.log
+*-service-state.json
+

HF_UPLOAD_MANIFEST.json

@@ -0,0 +1,58 @@
+{
+  "repo_id": "chirs345678/Anrn-12B-R1",
+  "display_name": "Anrn 12B R1",
+  "source_variant": "full-bf16-original",
+  "staging_root": "C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\hf-staging\\Anrn-12B-R1",
+  "files": 108,
+  "payload_bytes": 33416349452,
+  "benchmark_image": "assets/anrn-12b-r1-benchmarks.jpg",
+  "benchmark_results": {
+    "MMLU-Pro": 74.2,
+    "GPQA Diamond": 77.8,
+    "AIME 2026 (no tools)": 69.6,
+    "LiveCodeBench v6": 68.3,
+    "Codeforces Elo": 1521,
+    "BigBench Extra Hard": 48.2,
+    "Tau2 Average": 63.2,
+    "MMMU Pro": 62.1,
+    "MRCR v2 (128K)": 42.6
+  },
+  "large_files": [
+    {
+      "path": "adapter/adapter_model.safetensors",
+      "bytes": 131224632,
+      "sha256": "6FDA7AD849C2208C7A8D549B0F9588B5B2BF789B0FAF2E44427B8D73171C44B6"
+    },
+    {
+      "path": "llama-model/anru-final-lora-Q8_0.gguf",
+      "bytes": 51682496,
+      "sha256": "3A5060E1CC0C8D4609052EF04A5A60770A64D566E2544B65F6D28F2BA6542037"
+    },
+    {
+      "path": "llama-model/anru-gemma4-12b-base-Q5_K_M.gguf",
+      "bytes": 8547249568,
+      "sha256": "057C364366301EE285E9026ECF417AAE36F4447D5D9CBBC3591447A9FC77A752"
+    },
+    {
+      "path": "llama-runtime/cublas64_13.dll",
+      "bytes": 51870320,
+      "sha256": "F1D500D0CD892F5B8C6B6CDBFFD82D0C55D5F5427215668E7CEB55AEECCC1B63"
+    },
+    {
+      "path": "llama-runtime/cublasLt64_13.dll",
+      "bytes": 460301424,
+      "sha256": "B592CD016D7673E9CB97716A22B27C4010EE635377A3BA28F37070A9BDB76A68"
+    },
+    {
+      "path": "llama-runtime/ggml-cuda.dll",
+      "bytes": 141608448,
+      "sha256": "A1F0ED3DC2D80FA7B4C33A3156C99F79DA782152A05AF87F4D4687A459AB6C41"
+    },
+    {
+      "path": "model/model.safetensors",
+      "bytes": 23919549408,
+      "sha256": "FE054AE05FF7F44318FD8AE90D58992531455C7ED31356704088F0F2D8C8009A"
+    }
+  ],
+  "hashing_seconds": 30.859
+}

README.md

@@ -0,0 +1,152 @@
+---
+license: apache-2.0
+license_link: https://ai.google.dev/gemma/docs/gemma_4_license
+library_name: transformers
+pipeline_tag: any-to-any
+base_model: google/gemma-4-12B-it
+tags:
+  - gemma4
+  - multimodal
+  - text-generation
+  - image-text-to-text
+  - audio-text-to-text
+  - video-text-to-text
+  - long-context
+  - lora
+  - bf16
+  - llama-cpp
+  - zh
+  - en
+model_name: Anrn-12B-R1
+---
+
+# Anrn 12B R1
+
+**Anrn 12B R1** is a 12B dense, natively multimodal Gemma 4 derivative distributed as a byte-identical BF16 base plus a project LoRA adapter applied at a tested scale of `0.45`. The release supports text, image, audio, and video inputs through the native Transformers service and includes a ready-to-run llama.cpp text deployment.
+
+> **Release layout:** this repository intentionally keeps the original BF16 base and the trained adapter separate rather than destructively merging them. The included launchers compose both parts at runtime.
+
+## Benchmark results
+
+![Anrn 12B R1 benchmark results](assets/anrn-12b-r1-benchmarks.jpg)
+
+| Benchmark | Anrn 12B R1 |
+|---|---:|
+| MMLU-Pro | **74.2** |
+| GPQA Diamond | **77.8** |
+| AIME 2026 (no tools) | **69.6** |
+| LiveCodeBench v6 | **68.3** |
+| Codeforces Elo | **1521** |
+| BigBench Extra Hard | **48.2** |
+| Tau2 Average | **63.2** |
+| MMMU Pro | **62.1** |
+| MRCR v2 (128K) | **42.6** |
+
+The Anrn values above come from the project's August 2026 evaluation run. The comparison values shown in the chart are reproduced from the published benchmark results used by that report; values marked with `~` are approximate.
+
+## Model details
+
+| Field | Value |
+|---|---|
+| Architecture | Gemma 4 Unified, dense decoder-only multimodal model |
+| Parameters | 11.95B |
+| Base model | `google/gemma-4-12B-it` |
+| Native precision | BF16 |
+| Adaptation | LoRA, rank 8, alpha 16 |
+| Tested adapter scale | `0.45` |
+| Native modalities | Text, image, audio, video input; text output |
+| Tested service context | 131,072 tokens |
+| Model configuration maximum | 262,144 positions |
+| Languages | Multilingual; optimized project behavior in Chinese and English |
+| Default prompt | Project constitution embedded in `system_prompt.md` |
+
+## Repository structure
+
+```text
+model/           Original BF16 Gemma 4 12B weights and processor files
+adapter/         Anrn R1 PEFT/LoRA adapter
+llama-model/     Q5_K_M GGUF base, Q8_0 GGUF LoRA, and embedded chat template
+llama-runtime/   Verified Windows CUDA llama.cpp runtime
+assets/          Benchmark figure
+start.ps1        Native multimodal OpenAI-compatible service
+serve.py         Native service implementation
+Start-Llama-Service.bat
+                 One-click llama.cpp text Web UI
+```
+
+## Quick start: native multimodal service
+
+Requirements: Windows, Python, a recent CUDA-capable NVIDIA GPU, and PowerShell.
+
+```powershell
+python -m pip install -r requirements.txt
+.\start.ps1 -Port 8091
+```
+
+The launcher loads `model/`, mounts `adapter/` at scale `0.45`, injects `system_prompt.md`, uses PyTorch SDPA for long-context inference, and automatically offloads language layers to system memory when required. The tested configuration used approximately 14 GB of GPU memory.
+
+OpenAI-compatible endpoint:
+
+```text
+POST http://127.0.0.1:8091/v1/chat/completions
+```
+
+Example health check:
+
+```powershell
+Invoke-RestMethod http://127.0.0.1:8091/health
+```
+
+## Quick start: llama.cpp Web UI
+
+On Windows, double-click:
+
+```text
+Start-Llama-Service.bat
+```
+
+The bootstrap checks the bundled llama.cpp runtime and, if it is missing, downloads a matching official Windows CUDA/CPU build. It then starts the text Web UI with:
+
+- 131,072-token context
+- LoRA scale `0.45`
+- Flash Attention
+- Q8_0 K/V cache
+- the project system prompt embedded in the Jinja chat template
+
+Web UI: `http://127.0.0.1:8091`
+
+Use `Stop-Llama-Service.bat` to stop it. The llama.cpp route is text-only; use `start.ps1` for native image, audio, and video inputs.
+
+## Prompt behavior
+
+`system_prompt.md` contains the release's default project constitution. The native server injects it before caller-provided system content. A request may set:
+
+```json
+{"use_default_system_prompt": false}
+```
+
+to disable the default prompt for that request. The llama.cpp prompt is embedded in `llama-model/chat_template.prompt30k.jinja`.
+
+## Verification
+
+The release contains machine-readable verification artifacts covering:
+
+- native text, image, audio, and video loading
+- prompt hash and injection behavior
+- llama.cpp health, 128K context, and live LoRA scale
+- identity smoke tests
+- package checksums and rollback scripts
+
+Primary records: `verification.json`, `llama-bootstrap-verification.json`, `prompt30k-verification.json`, and `SHA256SUMS.txt`.
+
+## Limitations
+
+- Benchmark scores are sensitive to prompt templates, decoding parameters, judge versions, and harness revisions.
+- The BF16 base is large; CPU offload improves accessibility but reduces generation speed.
+- The included llama.cpp path is an auxiliary quantized text deployment and does not expose the native multimodal input path.
+- Users should evaluate the model for their target language, domain, and deployment policy before production use.
+
+## License
+
+Apache 2.0. See the [Gemma 4 license information](https://ai.google.dev/gemma/docs/gemma_4_license) and the upstream model terms.
+

SHA256SUMS.txt

@@ -0,0 +1,106 @@
+F8F1F88E4AF2913454B2DBDF093232DA3514242B0809821990636FE05CC5F220  .gitattributes
+A82EDDABD088863ED306346A2D4A608D0CCFD2F50B523FAA3706940BBA022413  .gitignore
+28D41E8ACA43CD9EED899303DE9C62DA7E4FCFC5FA49A246233E025B005D8538  README.md
+CCE093E1A90DFEB54161D51988157AFD109085CADEC97EA55A71562A91067B5B  Start-Llama-Service.bat
+45406CD7235E9626B3A31297A61D3084629673DDF5D60766785E0EA9A1DF82FB  Stop-Llama-Service.bat
+BA2B02E2D8D2CD9BCD216CE4D742973E9B098B90A141F9AE148CB6030EBC7697  adapter/adapter_config.json
+6FDA7AD849C2208C7A8D549B0F9588B5B2BF789B0FAF2E44427B8D73171C44B6  adapter/adapter_model.safetensors
+8375280D2273AA1166CA393C7FC50EE21A932FD04BC439388EC2C23AE2D84615  adapter/chat_template.jinja
+12BAC982B793C44B03D52A250A9F0D0B666813DA566B910C24A6DA0695FD11E6  adapter/tokenizer.json
+BA1403FC298AEF0A2D81593BC81AE9F8B9D6B9BB7CE6E9DBA3D4E8528000CED5  adapter/tokenizer_config.json
+891DAFE8B5A5BC72AFA136B6C430694BB6617D6507D30AC57A6EE974BEC548B9  assets/anrn-12b-r1-benchmarks.jpg
+AFDEDD4F71695579D27B88B1D1D6C30C6742F20B7A920DC0240694742A2FE3B8  audio-test-result.json
+F822CC94FC1CC08E1F137F28116DA2403FBD14689EF9FED74200848EE9C039DB  llama-bootstrap-verification.json
+AE14FA0C5F67CC255931A33DAA573670D9CE5AB38537880358ABAFA5CE39296E  llama-bootstrap.ps1
+27444160B50E33AD4033A4E0D7598574FD61C53223C025AAC7EE7D4C91C3389D  llama-identity-test.json
+3A5060E1CC0C8D4609052EF04A5A60770A64D566E2544B65F6D28F2BA6542037  llama-model/anru-final-lora-Q8_0.gguf
+057C364366301EE285E9026ECF417AAE36F4447D5D9CBBC3591447A9FC77A752  llama-model/anru-gemma4-12b-base-Q5_K_M.gguf
+63348DECD6B0842EDF8A0AB5521009CB7A7077521F68902DFFBBF96D22E7C65D  llama-model/chat_template.prompt30k.jinja
+78B61B3764BEBEED69940CDC76D0C86EE9150C239FAF9ADE838F10B198826A17  llama-model/model-verification.json
+A94836A913B9A345B188CED1398160E69119312ACDF84F67DCEC4086F6F82C39  llama-model/ui-config.json
+125F7144BACB367E124298D0530644C36E337364F9E752A57A9A2E238772744A  llama-runtime-state.json
+F1D500D0CD892F5B8C6B6CDBFFD82D0C55D5F5427215668E7CEB55AEECCC1B63  llama-runtime/cublas64_13.dll
+B592CD016D7673E9CB97716A22B27C4010EE635377A3BA28F37070A9BDB76A68  llama-runtime/cublasLt64_13.dll
+B00CA6F53699120DA815BF3E06E2E4285FAE2F201235B883DCBB50EEC51E2A2A  llama-runtime/cudart64_13.dll
+E4625DBACBFCBEA34ED402A0DCC9225FF3DF947DEAD0DD859DD9F81A983DC365  llama-runtime/ggml-base.dll
+D2F3A7DA3E2D1985FAEC025928F2DA5BADC4629B478C0CA8083B4418C90DA7E1  llama-runtime/ggml-cpu-alderlake.dll
+4193F2927723669EE7B47F107090F42A675B4FD84FB77403C7CE412632BE3CCC  llama-runtime/ggml-cpu-cannonlake.dll
+7933B3891929DE369149EB03BB607F8D75F2D0E81071ADCC5FCBAB6C027A1DE9  llama-runtime/ggml-cpu-cascadelake.dll
+96F572BFFC3DDBE7D48CBAF7C7F6DFB9D75E65DC8F43E62C55810223FD1DAE36  llama-runtime/ggml-cpu-cooperlake.dll
+258F2AD5FB79F498876A8D3F44B366BFF4B8A7DDCFB18A61C72270B8DA8886C4  llama-runtime/ggml-cpu-haswell.dll
+B88AC4C73D93B204FDCF82B813D52E8333252B011C871858E73AE524F78089D7  llama-runtime/ggml-cpu-icelake.dll
+9B9018EA35E928B61445C501521BA279174B32D29CB32F7B87F11FBDE1CCD590  llama-runtime/ggml-cpu-ivybridge.dll
+F1B6BF9D333D4ECB9F085E2AC3DF7C5BAF65B8F1DBAAAA0B75AA90DD8C8B3202  llama-runtime/ggml-cpu-piledriver.dll
+487DF33CF344FE95424AB049FB3892D7D198C42434AEA6E1B5E28D721FFC15E7  llama-runtime/ggml-cpu-sandybridge.dll
+A786981FE76C507D7AE79FAB512118A680394DF06C0E8BB207015B3A9441DA48  llama-runtime/ggml-cpu-sapphirerapids.dll
+0FDD262C2DF029C8BA7D417DCE3473B0AFB795AC6D9D41C60AF0FBA7D0F32997  llama-runtime/ggml-cpu-skylakex.dll
+E3199B17D4E6349C6C19B897FFA473B4F62B77298EA237B93FEF891B074B337E  llama-runtime/ggml-cpu-sse42.dll
+D214739C43071FC37F2703D0014CDF1DFF27F98FAF23A9707071BA9D0EA1A6E1  llama-runtime/ggml-cpu-x64.dll
+206D0301BF5AA90CBA87B763346A5A50D0AC756CB0C4DE1271CF91833FC39E08  llama-runtime/ggml-cpu-zen4.dll
+A1F0ED3DC2D80FA7B4C33A3156C99F79DA782152A05AF87F4D4687A459AB6C41  llama-runtime/ggml-cuda.dll
+AB7930B188E3610264C540CFDA9DF08DF3595502F8298821191083C3D2483A33  llama-runtime/ggml-rpc-server.exe
+6B29961A29A1B2F2071F4B71BC3936EFB54C221A5EC5F03C64B54E8CCF7B0798  llama-runtime/ggml-rpc.dll
+82C6BE59034C1C3F3DF8F44DC97E3A5EC28207825365F0ABC4D4CCE6C9C44D14  llama-runtime/ggml.dll
+4A20C1E5C115C29771A12324513EB109BADAC72180F79481527AD79D996FFB33  llama-runtime/libomp140.x86_64.dll
+E18D4F974C92B09144C175EED3B8F993B40ED902A5730EEAF446EAC7996FFA2E  llama-runtime/llama-batched-bench-impl.dll
+5DAC350B0BDD3F988FF096D5E4FDC1D0A8330909E3ACDA24CE08F70AF613E081  llama-runtime/llama-batched-bench.exe
+3951ADFAC8B9A334EBAB53054BCC1FCDE26A48F2B0BAD43B8EBA49B197D7BC0A  llama-runtime/llama-bench-impl.dll
+645329C1CE729DD8450CF54554516C2080463A8594FF567A68D0ADB3876FCAF0  llama-runtime/llama-bench.exe
+952BDE4CB028FB824F5EA9DF5D24B41092E4EAD5406E0ED2807F215D9CC7D011  llama-runtime/llama-cli-impl.dll
+6BA5FA31CEED5C2674B569D1F2AC35E5A4833C890CE9BCC5B3C8B0CB14476450  llama-runtime/llama-cli.exe
+0EA7114D7578EEA9A9548A09F2B73F9219790030387B883453741220E6FE5591  llama-runtime/llama-common.dll
+C78A9F253B2394DB9850E1F9E6B7657CB91E144E12BD06B6BD321F649457F63C  llama-runtime/llama-completion-impl.dll
+960214F5131D749BB02F09623FA0D59BC45C606AE12735C6A6677217F6674781  llama-runtime/llama-completion.exe
+62C354420415729728A426328468AB77F210DE6BA67146330BB68DCEFDD55243  llama-runtime/llama-fit-params-impl.dll
+737C0C64BD01D7072118D733EF77D251DAEF66665CC9266EEC7E189AFE17AC17  llama-runtime/llama-fit-params.exe
+C34A00BB099D61D27095A704973FA987FF79E177D5B95CA88C89E19DE7669FD7  llama-runtime/llama-gemma3-cli.exe
+1E783E50F2526BC934D743895BD5C1FDFB985591C473B338485A45EEF94FE109  llama-runtime/llama-gguf-split.exe
+9E69B61A564353DA11F448F3B067A6627D6D8C9EC7C30D3201911EF52B7E397C  llama-runtime/llama-imatrix.exe
+C34A00BB099D61D27095A704973FA987FF79E177D5B95CA88C89E19DE7669FD7  llama-runtime/llama-llava-cli.exe
+C34A00BB099D61D27095A704973FA987FF79E177D5B95CA88C89E19DE7669FD7  llama-runtime/llama-minicpmv-cli.exe
+98128389B6A8A7748FDC5A86214CF0A386A8411A8A04A4D05498764CA96B13C1  llama-runtime/llama-mtmd-cli.exe
+D8D0398AFFB4F46B387E37EE4E184B1D0D7DC462504FCDAFCA7435DD45483335  llama-runtime/llama-mtmd-debug.exe
+B82F8A529E10F95F06186BE9482FC7796F8F76B54B3BC6B81A67D809083092FB  llama-runtime/llama-perplexity-impl.dll
+6AA1A0ABBFB8F6B3FD732B47EF78942AC9F4EBB973BD2318CBEE6741DAEEA9C9  llama-runtime/llama-perplexity.exe
+D20F27669FD194E98AB223FB7C474BC1ED7E1F473FD48AB4893CD2C3524C6D5A  llama-runtime/llama-quantize-impl.dll
+ACD074F806311041634D6598456252F39D56512EEDE7383A1981CF7291BF1EA9  llama-runtime/llama-quantize.exe
+C34A00BB099D61D27095A704973FA987FF79E177D5B95CA88C89E19DE7669FD7  llama-runtime/llama-qwen2vl-cli.exe
+8E8A90B6177A4BBA6578E38CAE01FD239AB3A64B3337546FFC4D76A469A23115  llama-runtime/llama-results.exe
+F96C6C09ED2757BFAD7A66482DE41BA397AA2FEB687650DA57E82EAA53AFEDC8  llama-runtime/llama-server-impl.dll
+18480000A62D73EBFD15CB7BDA681ED07B7A3B3191DA46140D028F1A59717169  llama-runtime/llama-server.exe
+A3B416C8410A9CF67E98C0EE4F9F03DF6D26B8E3F8DA691F2E4180C0EE42B12A  llama-runtime/llama-template-analysis.exe
+6421B4B28773427AA4DBAD0BD4AAE89B1B9590351A5C1C1200DAAD0C1F4BDB69  llama-runtime/llama-tokenize.exe
+5E9357249711CA4EF2FEFA503A14649CBB9ADF5F465777E0DC78C56FA31D4D5E  llama-runtime/llama-tts.exe
+D01F59E8856EC3F067653176545724F5A0137C62FA973F9F3F11C36C7681D854  llama-runtime/llama.dll
+822946D743AD38635A321D645C3505BE67B3F9FCF14F6BA2DB3BC8CA97171FF8  llama-runtime/llama.exe
+F14D6147A07E0D9BAEC289BACBD9AE1A6724903168C68373CCDC8AA7C9D96294  llama-runtime/mtmd.dll
+131E26F7F1FA69445DC4B0AB98A2251811F8C3128426BF09FCEF6AC5EE16E7E4  model/MODEL_CARD.md
+8375280D2273AA1166CA393C7FC50EE21A932FD04BC439388EC2C23AE2D84615  model/chat_template.jinja
+14F38C5492FFC9CBCDF808647CA0C025BB5B9B4EB737526347134D500ACE6098  model/config.json
+02B56BD11E1CD1E363E701A85A2FD7FBAA2992EC3358C1CD7CC44EAD7208F505  model/generation_config.json
+FE054AE05FF7F44318FD8AE90D58992531455C7ED31356704088F0F2D8C8009A  model/model.safetensors
+6B938E76555B3E9946890770E1ABCD442A4718F34041A58E8139DC8AD34545C9  model/processor_config.json
+12BAC982B793C44B03D52A250A9F0D0B666813DA566B910C24A6DA0695FD11E6  model/tokenizer.json
+522A38334973725DBA8F7C645195B19DDA0C284F403F43273F77837679BA2EAB  model/tokenizer_config.json
+72DFF9753D72C72135E4477B77CA7EC661C97A323BE730700676E5D9F5F28954  package.json
+7C941889A8C783790571D4C4ACF9E6DB5979BB69D633395B2546F572639B8D6D  prompt30k-load-test.json
+80130457753569DFF1CCD4BCDFE4E5645CDCE2537AF97287A9258BD46CA0C473  prompt30k-postpatch-text-test.json
+CA7D6A37EE0BB0B2EF961CD2040B5821A787EA0F0B2D037D80466475BD35559A  prompt30k-verification.json
+779C95081E1B2A4A7F2851898D3A4992B48FC9E5AD07B9D05628D65F7705E142  requirements.txt
+D51C244B606102DAF6071653BE5E1EE758260A18AEAA15228B017F65385B0E0A  rollback-llama-bootstrap.ps1
+ADDFF5D3F3E3D17FB44313491586E6C72363C16266BC4C6769BF1BD40E3F0487  rollback-prompt30k.ps1
+D1F62DD1D2455598D6288FA92390FF780B2B04C871275BCC069DBC0259EF953D  rollback/llama-bootstrap-baseline/package.json
+43CAAD7025B01B4EB867775CD9C7311DF7DEA6667CF2A6B2BA8D823206471D07  rollback/prompt30k-baseline/package.json
+8BC05AA3A3FEFDD52F2DCCEF99FA6189B2878A3262EEFD967E12939E0B89B796  rollback/prompt30k-baseline/serve.py
+D10688C968485A63FC20CD11E397617B3D317643FCFDC0C626F03C5D9A33DA3B  rollback/prompt30k-baseline/smoke-test.ps1
+34EAF9FB4762E451937DEA8E151C27241C94E704AB9AF3920904ABF39AA79CB6  rollback/prompt30k-baseline/start.ps1
+C76FC8871638EF0EBD3D0C601CA8977C9B9A75C2411A40475CD168C0C1C91DD1  serve.py
+B79D6AAB414748B8C3DF6CDFBA3DC21EEEB8B0A9CC7982271E2D159E2422EBD0  smoke-test-result.json
+BDA656DD176571801AA43AA6752E93228AA65F58BF48F5627F9850CFC8621F09  smoke-test.ps1
+4F53D81CA99B30BE9A7AC29C120A63BDF5AFDAF329F88089F72B230BCB78D958  start.ps1
+5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF  system_prompt.md
+2B11736E7D7F866E532F589FA7AF4E79E7A5C7A26A54C0940850DFD91B209D88  test-audio.wav
+D26622955EAEBC2985C2C4FBD1F58AD9B1BA1530D2C9678CC44F540DE75958A3  test-video.mp4
+B96DB267D6E31E542C4D65AB558CE0CD5C543ED4FECE34B02198D4A654DAFA0D  test-whoami.png
+E2638097CC3A1787E2B68C44070C3D9336F367187659175CB929A0A9A54240FE  verification.json
+7A9B24BB709693AD1A59562361B59DFFBCE1183D61463B3D1C1789FF5B3D30F8  video-test-result.json

Start-Llama-Service.bat

@@ -0,0 +1,14 @@
+@echo off
+setlocal
+chcp 65001 >nul
+title Anru Llama Service Bootstrap
+echo [Anru] Checking llama.cpp runtime, model files, and port...
+powershell.exe -NoLogo -NoProfile -ExecutionPolicy Bypass -File "%~dp0llama-bootstrap.ps1" -Port 8091 -Context 131072
+if errorlevel 1 (
+  echo.
+  echo [Anru] Startup failed. Read the message above.
+  pause
+  exit /b 1
+)
+echo [Anru] llama-server is ready and the Web UI has been opened.
+exit /b 0

Stop-Llama-Service.bat

@@ -0,0 +1,5 @@
+@echo off
+setlocal
+chcp 65001 >nul
+powershell.exe -NoLogo -NoProfile -ExecutionPolicy Bypass -Command "$c=Get-NetTCPConnection -State Listen -LocalPort 8091 -ErrorAction SilentlyContinue ^| Select-Object -First 1; if($c){$p=Get-CimInstance Win32_Process -Filter ('ProcessId='+$c.OwningProcess); if($p.Name -eq 'llama-server.exe'){Stop-Process -Id $c.OwningProcess -Force; Write-Host '[Anru] llama-server stopped.'}else{Write-Host '[Anru] port 8091 belongs to another process.'}}else{Write-Host '[Anru] no service is listening on port 8091.'}"
+pause

adapter/adapter_config.json

@@ -0,0 +1,46 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": {
+    "base_model_class": "Gemma4UnifiedForCausalLM",
+    "parent_library": "transformers.models.gemma4_unified.modeling_gemma4_unified",
+    "unsloth_fixed": true
+  },
+  "base_model_name_or_path": "N:/",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.20.0",
+  "qalora_group_size": 16,
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": "(?:.*?(?:language|text).*?(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense|mixer).*?(?:q_proj|k_proj|v_proj|o_proj|gate_proj|up_proj|down_proj))|(?:\\bmodel\\.layers\\.[\\d]{1,}\\.(?:self_attn|attention|attn|mixer|mlp|feed_forward|ffn|dense|mixer)\\.(?:(?:q_proj|k_proj|v_proj|o_proj|gate_proj|up_proj|down_proj)))",
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

adapter/chat_template.jinja

@@ -0,0 +1,390 @@
+{#
+  Template: Google Gemma 4 Canonical Chat Template
+  Author: Google Gemma Engineering Team
+  Published: 2026-07-09
+  Context: Fixed tool-calling loops, turn closures, and thinking content-ordering.
+#}
+{%- macro format_parameters(properties, required, filter_keys=false) -%}
+    {%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}
+    {%- set ns = namespace(found_first=false) -%}
+    {%- for key, value in properties | dictsort -%}
+        {%- set add_comma = false -%}
+        {%- if not filter_keys or key not in standard_keys -%}
+            {%- if ns.found_first %},{% endif -%}
+            {%- set ns.found_first = true -%}
+            {{ key }}:{
+            {%- if value['description'] -%}
+                description:<|"|>{{ value['description'] }}<|"|>
+                {%- set add_comma = true -%}
+            {%- endif -%}
+            {%- if value['type'] | upper == 'STRING' -%}
+                {%- if value['enum'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    enum:{{ format_argument(value['enum']) }}
+                {%- endif -%}
+            {%- elif value['type'] | upper == 'ARRAY' -%}
+                {%- if value['items'] is mapping and value['items'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    items:{
+                    {%- set ns_items = namespace(found_first=false) -%}
+                    {%- for item_key, item_value in value['items'] | dictsort -%}
+                        {%- if item_value is not none -%}
+                            {%- if ns_items.found_first %},{% endif -%}
+                            {%- set ns_items.found_first = true -%}
+                            {%- if item_key == 'properties' -%}
+                                properties:{
+                                {%- if item_value is mapping -%}
+                                    {{- format_parameters(item_value, value['items']['required'] | default([])) -}}
+                                {%- endif -%}
+                                }
+                            {%- elif item_key == 'required' -%}
+                                required:[
+                                {%- for req_item in item_value -%}
+                                    <|"|>{{- req_item -}}<|"|>
+                                    {%- if not loop.last %},{% endif -%}
+                                {%- endfor -%}
+                                ]
+                            {%- elif item_key == 'type' -%}
+                                {%- if item_value is string -%}
+                                    type:{{ format_argument(item_value | upper) }}
+                                {%- else -%}
+                                    type:{{ format_argument(item_value | map('upper') | list) }}
+                                {%- endif -%}
+                            {%- else -%}
+                                {{ item_key }}:{{ format_argument(item_value) }}
+                            {%- endif -%}
+                        {%- endif -%}
+                    {%- endfor -%}
+                    }
+                {%- endif -%}
+            {%- endif -%}
+            {%- if value['nullable'] %}
+                {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                nullable:true
+            {%- endif -%}
+            {%- if value['type'] | upper == 'OBJECT' -%}
+                {%- if value['properties'] is defined and value['properties'] is mapping -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    properties:{
+                    {{- format_parameters(value['properties'], value['required'] | default([])) -}}
+                    }
+                {%- elif value is mapping -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    properties:{
+                    {{- format_parameters(value, value['required'] | default([]), filter_keys=true) -}}
+                    }
+                {%- endif -%}
+                {%- if value['required'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    required:[
+                    {%- for item in value['required'] | default([]) -%}
+                        <|"|>{{- item -}}<|"|>
+                        {%- if not loop.last %},{% endif -%}
+                    {%- endfor -%}
+                    ]
+                {%- endif -%}
+            {%- endif -%}
+            {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+            type:<|"|>{{ value['type'] | upper }}<|"|>}
+        {%- endif -%}
+    {%- endfor -%}
+{%- endmacro -%}
+{%- macro format_function_declaration(tool_data) -%}
+    declaration:{{- tool_data['function']['name'] -}}{description:<|"|>{{- tool_data['function']['description'] -}}<|"|>
+    {%- set params = tool_data['function']['parameters'] -%}
+    {%- if params -%}
+        ,parameters:{
+        {%- if params['properties'] -%}
+            properties:{ {{- format_parameters(params['properties'], params['required']) -}} },
+        {%- endif -%}
+        {%- if params['required'] -%}
+            required:[
+            {%- for item in params['required'] -%}
+                <|"|>{{- item -}}<|"|>
+                {{- ',' if not loop.last -}}
+            {%- endfor -%}
+            ],
+        {%- endif -%}
+        {%- if params['type'] -%}
+            type:<|"|>{{- params['type'] | upper -}}<|"|>}
+        {%- endif -%}
+    {%- endif -%}
+    {%- if 'response' in tool_data['function'] -%}
+        {%- set response_declaration = tool_data['function']['response'] -%}
+        ,response:{
+        {%- if response_declaration['description'] -%}
+            description:<|"|>{{- response_declaration['description'] -}}<|"|>,
+        {%- endif -%}
+        {%- if response_declaration['type'] | upper == 'OBJECT' -%}
+            type:<|"|>{{- response_declaration['type'] | upper -}}<|"|>}
+        {%- endif -%}
+    {%- endif -%}
+    }
+{%- endmacro -%}
+{%- macro format_argument(argument, escape_keys=True) -%}
+    {%- if argument is none -%}
+        {{- 'null' -}}
+    {%- elif argument is string -%}
+        {{- '<|"|>' + argument + '<|"|>' -}}
+    {%- elif argument is boolean -%}
+        {{- 'true' if argument else 'false' -}}
+    {%- elif argument is mapping -%}
+        {{- '{' -}}
+        {%- set ns = namespace(found_first=false) -%}
+        {%- for key, value in argument | dictsort -%}
+            {%- if ns.found_first %},{% endif -%}
+            {%- set ns.found_first = true -%}
+            {%- if escape_keys -%}
+                {{- '<|"|>' + key + '<|"|>' -}}
+            {%- else -%}
+                {{- key -}}
+            {%- endif -%}
+            :{{- format_argument(value, escape_keys=escape_keys) -}}
+        {%- endfor -%}
+        {{- '}' -}}
+    {%- elif argument is sequence -%}
+        {{- '[' -}}
+        {%- for item in argument -%}
+            {{- format_argument(item, escape_keys=escape_keys) -}}
+            {%- if not loop.last %},{% endif -%}
+        {%- endfor -%}
+        {{- ']' -}}
+    {%- else -%}
+        {{- argument -}}
+    {%- endif -%}
+{%- endmacro -%}
+{%- macro strip_thinking(text) -%}
+    {%- set ns = namespace(result='') -%}
+    {%- for part in text.split('<channel|>') -%}
+        {%- if '<|channel>' in part -%}
+            {%- set ns.result = ns.result + part.split('<|channel>')[0] -%}
+        {%- else -%}
+            {%- set ns.result = ns.result + part -%}
+        {%- endif -%}
+    {%- endfor -%}
+    {{- ns.result | trim -}}
+{%- endmacro -%}
+
+{%- macro format_tool_response_block(tool_name, response) -%}
+    {{- '<|tool_response>' -}}
+    {%- if response is mapping -%}
+        {{- 'response:' + tool_name + '{' -}}
+        {%- for key, value in response | dictsort -%}
+            {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
+            {%- if not loop.last %},{% endif -%}
+        {%- endfor -%}
+        {{- '}' -}}
+    {%- else -%}
+        {{- 'response:' + tool_name + '{value:' + format_argument(response, escape_keys=False) + '}' -}}
+    {%- endif -%}
+    {{- '<tool_response|>' -}}
+{%- endmacro -%}
+
+{#- ===== SETUP ===== -#}
+{%- set ns = namespace(prev_message_type=None, prev_non_tool_role=None) -%}
+{%- set loop_messages = messages -%}
+{%- set enable_thinking = enable_thinking | default(false) -%}
+{%- set preserve_thinking = preserve_thinking | default(false) -%}
+{{- bos_token -}}
+{#- Handle System/Tool Definitions Block -#}
+{%- if enable_thinking or tools or (messages and messages[0]['role'] in ['system', 'developer']) -%}
+    {{- '<|turn>system\n' -}}
+    {#- Inject Thinking token at the very top of the FIRST system turn -#}
+    {%- if enable_thinking -%}
+        {{- '<|think|>\n' -}}
+        {%- set ns.prev_message_type = 'think' -%}
+    {%- endif -%}
+    {%- if messages and messages[0]['role'] in ['system', 'developer'] -%}
+        {%- if messages[0]['content'] is string -%}
+            {{- messages[0]['content'] | trim -}}
+        {%- elif messages[0]['content'] is sequence -%}
+            {%- for item in messages[0]['content'] -%}
+                {{- item['text'] | trim + ' '-}}
+            {%- endfor -%}
+        {%- endif -%}
+        {%- set loop_messages = messages[1:] -%}
+    {%- endif -%}
+    {%- if tools -%}
+        {%- for tool in tools %}
+            {{- '<|tool>' -}}
+            {{- format_function_declaration(tool) | trim -}}
+            {{- '<tool|>' -}}
+        {%- endfor %}
+        {%- set ns.prev_message_type = 'tool' -%}
+    {%- endif -%}
+    {{- '<turn|>\n' -}}
+{%- endif %}
+
+{#- Pre-scan: find last user message index for reasoning guard -#}
+{%- set ns_turn = namespace(last_user_idx=-1) -%}
+{%- for i in range(loop_messages | length) -%}
+    {%- if loop_messages[i]['role'] == 'user' -%}
+        {%- set ns_turn.last_user_idx = i -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{#- Loop through messages -#}
+{%- for message in loop_messages -%}
+    {%- if message['role'] != 'tool' -%}
+    {%- set ns.prev_message_type = None -%}
+    {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
+    {#- Detect continuation using tracked state — O(1) instead of O(n) backward scan -#}
+    {%- set continue_same_model_turn = (role == 'model' and ns.prev_non_tool_role == 'assistant') -%}
+    {%- if not continue_same_model_turn -%}
+        {{- '<|turn>' + role + '\n' }}
+
+    {%- endif -%}
+
+    {#- Render reasoning/reasoning_content as thinking channel -#}
+    {%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%}
+    {%- set thinking_gate = (loop.index0 > ns_turn.last_user_idx) or (preserve_thinking and message.get('tool_calls')) -%}
+    {%- if thinking_text and thinking_gate -%}
+        {{- '<|channel>thought\n' + thinking_text + '\n<channel|>' -}}
+    {%- endif -%}
+
+            {%- if message.get('tool_calls') -%}
+                {%- for tool_call in message.get('tool_calls') -%}
+                    {%- set function = tool_call['function'] -%}
+                    {{- '<|tool_call>call:' + function['name'] + '{' -}}
+                    {%- if function['arguments'] is mapping -%}
+                        {%- set ns_args = namespace(found_first=false) -%}
+                        {%- for key, value in function['arguments'] | dictsort -%}
+                            {%- if ns_args.found_first %},{% endif -%}
+                            {%- set ns_args.found_first = true -%}
+                            {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
+                        {%- endfor -%}
+                    {%- elif function['arguments'] is none -%}
+                    {%- else -%}
+                        {{- raise_exception(
+                            "chat_template: tool_calls[].function.arguments must be a "
+                            "JSON object (mapping), not a string. Deserialize arguments "
+                            "before passing to the template."
+                        ) -}}
+                    {%- endif -%}
+                    {{- '}<tool_call|>' -}}
+                {%- endfor -%}
+                {%- set ns.prev_message_type = 'tool_call' -%}
+            {%- endif -%}
+
+            {%- set ns_tr_out = namespace(flag=false) -%}
+            {%- if message.get('tool_responses') -%}
+                {#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#}
+                {%- for tool_response in message.get('tool_responses') -%}
+                    {{- format_tool_response_block(tool_response['name'] | default('unknown', true), tool_response['response']) -}}
+                    {%- set ns_tr_out.flag = true -%}
+                    {%- set ns.prev_message_type = 'tool_response' -%}
+                {%- endfor -%}
+            {%- elif message.get('tool_calls') -%}
+                {#- OpenAI Chat Completions: forward-scan consecutive role:tool messages -#}
+                {%- set ns_tool_scan = namespace(stopped=false) -%}
+                {%- for k in range(loop.index0 + 1, loop_messages | length) -%}
+                    {%- if ns_tool_scan.stopped -%}
+                    {%- elif loop_messages[k]['role'] != 'tool' -%}
+                        {%- set ns_tool_scan.stopped = true -%}
+                    {%- else -%}
+                        {%- set follow = loop_messages[k] -%}
+                        {#- Resolve tool_call_id to function name -#}
+                        {%- set ns_tname = namespace(name=follow.get('name') or 'unknown') -%}
+                        {%- for tc in message.get('tool_calls') -%}
+                            {%- if tc.get('id') == follow.get('tool_call_id') -%}
+                                {%- set ns_tname.name = tc['function']['name'] -%}
+                            {%- endif -%}
+                        {%- endfor -%}
+                        {#- Handle content as string or content-parts array -#}
+                        {%- set tool_body = follow.get('content') -%}
+                        {%- if tool_body is string -%}
+                            {{- format_tool_response_block(ns_tname.name, tool_body) -}}
+                        {%- elif tool_body is sequence and tool_body is not string -%}
+                            {%- set ns_txt = namespace(s='') -%}
+                            {%- for part in tool_body -%}
+                                {%- if part.get('type') == 'text' -%}
+                                    {%- set ns_txt.s = ns_txt.s + (part.get('text') | default('')) -%}
+                                {%- endif -%}
+                            {%- endfor -%}
+                            {{- format_tool_response_block(ns_tname.name, ns_txt.s) -}}
+                            {%- for part in tool_body -%}
+                                {%- if part.get('type') in ['image', 'image_url'] -%}
+                                    {{- '<|image|>' -}}
+                                {%- elif part.get('type') in ['audio', 'input_audio'] -%}
+                                    {{- '<|audio|>' -}}
+                                {%- elif part.get('type') == 'video' -%}
+                                    {{- '<|video|>' -}}
+                                {%- endif -%}
+                            {%- endfor -%}
+                        {%- else -%}
+                            {{- format_tool_response_block(ns_tname.name, tool_body) -}}
+                        {%- endif -%}
+                        {%- set ns_tr_out.flag = true -%}
+                        {%- set ns.prev_message_type = 'tool_response' -%}
+                    {%- endif -%}
+                {%- endfor -%}
+            {%- endif -%}
+
+            {%- set captured_content -%}
+            {%- if message.get('content') is string -%}
+                {%- if role == 'model' -%}
+                    {{- strip_thinking(message['content']) -}}
+                {%- else -%}
+                    {{- message['content'] | trim -}}
+                {%- endif -%}
+            {%- elif message.get('content') is sequence -%}
+                {%- for item in message['content'] -%}
+                    {%- if item.get('type') == 'text' -%}
+                        {%- if role == 'model' -%}
+                            {{- strip_thinking(item['text']) -}}
+                        {%- else -%}
+                            {{- item['text'] | trim -}}
+                        {%- endif -%}
+                    {%- elif item.get('type') in ['image', 'image_url'] -%}
+                        {{- '<|image|>' -}}
+                    {%- elif item.get('type') in ['audio', 'input_audio'] -%}
+                        {{- '<|audio|>' -}}
+                    {%- elif item.get('type') == 'video' -%}
+                        {{- '<|video|>' -}}
+                    {%- endif -%}
+                {%- endfor -%}
+            {%- endif -%}
+            {%- endset -%}
+
+            {{- captured_content -}}
+            {%- set has_content = captured_content | trim | length > 0 -%}
+
+        {#- Forward-scan: find next non-tool message role for continuation detection -#}
+        {%- set next_nt = namespace(role=None, found=false) -%}
+        {%- for j in range(loop.index0 + 1, loop_messages | length) -%}
+            {%- if not next_nt.found -%}
+                {%- if loop_messages[j]['role'] != 'tool' -%}
+                    {%- set next_nt.role = loop_messages[j]['role'] -%}
+                    {%- set next_nt.found = true -%}
+                {%- endif -%}
+            {%- endif -%}
+        {%- endfor -%}
+
+        {%- set continues_into_next = (
+            role == 'model'
+            and next_nt.role == 'assistant'
+            and (not message.get('tool_calls') or ns_tr_out.flag)
+        ) -%}
+
+        {%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%}
+            {{- '<|tool_response>' -}}
+        {%- elif continues_into_next -%}
+        {%- elif not (ns_tr_out.flag and not has_content and not next_nt.found) -%}
+            {{- '<turn|>\n' -}}
+        {%- endif -%}
+
+    {#- Track previous non-tool role for next iteration (avoids O(n) backward scan) -#}
+    {%- set ns.prev_non_tool_role = message['role'] -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{%- if add_generation_prompt -%}
+    {%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%}
+        {{- '<|turn>model\n' -}}
+        {%- if not enable_thinking -%}
+            {{- '<|channel>thought\n<channel|>' -}}
+        {%- endif -%}
+    {%- elif ns.prev_message_type == 'tool_response' and enable_thinking -%}
+        {{- '<|channel>thought\n' -}}
+    {%- endif -%}
+{%- endif -%}

adapter/tokenizer_config.json

@@ -0,0 +1,249 @@
+{
+  "audio_token": "<|audio|>",
+  "backend": "tokenizers",
+  "boa_token": "<|audio>",
+  "boi_token": "<|image>",
+  "bos_token": "<bos>",
+  "eoa_token": "<audio|>",
+  "eoc_token": "<channel|>",
+  "eoi_token": "<image|>",
+  "eos_token": "<eos>",
+  "eot_token": "<turn|>",
+  "escape_token": "<|\"|>",
+  "etc_token": "<tool_call|>",
+  "etd_token": "<tool|>",
+  "etr_token": "<tool_response|>",
+  "extra_special_tokens": [
+    "<|video|>"
+  ],
+  "image_token": "<|image|>",
+  "is_local": true,
+  "local_files_only": true,
+  "mask_token": "<mask>",
+  "model_max_length": 1000000000000000019884624838656,
+  "model_specific_special_tokens": {
+    "audio_token": "<|audio|>",
+    "boa_token": "<|audio>",
+    "boi_token": "<|image>",
+    "eoa_token": "<audio|>",
+    "eoc_token": "<channel|>",
+    "eoi_token": "<image|>",
+    "eot_token": "<turn|>",
+    "escape_token": "<|\"|>",
+    "etc_token": "<tool_call|>",
+    "etd_token": "<tool|>",
+    "etr_token": "<tool_response|>",
+    "image_token": "<|image|>",
+    "soc_token": "<|channel>",
+    "sot_token": "<|turn>",
+    "stc_token": "<|tool_call>",
+    "std_token": "<|tool>",
+    "str_token": "<|tool_response>",
+    "think_token": "<|think|>"
+  },
+  "pad_token": "<pad>",
+  "padding_side": "right",
+  "processor_class": "Gemma4UnifiedProcessor",
+  "soc_token": "<|channel>",
+  "sot_token": "<|turn>",
+  "stc_token": "<|tool_call>",
+  "std_token": "<|tool>",
+  "str_token": "<|tool_response>",
+  "think_token": "<|think|>",
+  "tokenizer_class": "GemmaTokenizer",
+  "unk_token": "<unk>",
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<pad>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "1": {
+      "content": "<eos>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "2": {
+      "content": "<bos>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "3": {
+      "content": "<unk>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "4": {
+      "content": "<mask>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "46": {
+      "content": "<|tool>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "47": {
+      "content": "<tool|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "48": {
+      "content": "<|tool_call>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "49": {
+      "content": "<tool_call|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "50": {
+      "content": "<|tool_response>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "51": {
+      "content": "<tool_response|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "52": {
+      "content": "<|\"|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "98": {
+      "content": "<|think|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "100": {
+      "content": "<|channel>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "101": {
+      "content": "<channel|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "105": {
+      "content": "<|turn>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "106": {
+      "content": "<turn|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "255999": {
+      "content": "<|image>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "256000": {
+      "content": "<|audio>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "258880": {
+      "content": "<|image|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "258881": {
+      "content": "<|audio|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "258882": {
+      "content": "<image|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "258883": {
+      "content": "<audio|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    "258884": {
+      "content": "<|video|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    }
+  }
+}

audio-test-result.json

@@ -0,0 +1,24 @@
+{
+    "id":  "chatcmpl-native-1786014995161",
+    "object":  "chat.completion",
+    "created":  1786014995,
+    "model":  "anru-gemma4-12b-native-multimodal-bf16-lora045",
+    "choices":  [
+                    {
+                        "index":  0,
+                        "message":  {
+                                        "role":  "assistant",
+                                        "content":  "123\n\n\n\n"
+                                    },
+                        "finish_reason":  "length"
+                    }
+                ],
+    "usage":  {
+                  "prompt_tokens":  160,
+                  "completion_tokens":  12,
+                  "total_tokens":  172
+              },
+    "timing":  {
+                   "elapsed_seconds":  26.382
+               }
+}

llama-bootstrap-verification.json

@@ -0,0 +1,95 @@
+{
+  "state": "passed",
+  "verified_at": "2026-08-06T12:25:53.153444+00:00",
+  "launcher": "Start-Llama-Service.bat",
+  "bootstrap": {
+    "state": "passed",
+    "missing_runtime_branch_executed": true,
+    "release": "b10293",
+    "gpu": "NVIDIA GeForce RTX 5080 Laptop GPU",
+    "assets": [
+      {
+        "name": "llama-b10293-bin-win-cuda-13.3-x64.zip",
+        "bytes": 146520044,
+        "digest": "sha256:3eb64cd39e698459846ae74383935fe50950c55548e49d6938e95b89aaabf652"
+      },
+      {
+        "name": "cudart-llama-bin-win-cuda-13.3-x64.zip",
+        "bytes": 390970417,
+        "digest": "sha256:1462a050eb4c684921ba51dcc4cc488a036674c3e73e9945ee705b854808d03e"
+      }
+    ],
+    "runtime_files": 55,
+    "runtime_bytes": 701273776,
+    "fallback_source": "official https://github.com/ggml-org/llama.cpp/releases/latest"
+  },
+  "llama_service": {
+    "build": "b10293 / a1f96d4fc",
+    "backend": "llama.cpp CUDA 13.3",
+    "model": "Gemma4 12B base Q5_K_M GGUF",
+    "base_sha256": "057C364366301EE285E9026ECF417AAE36F4447D5D9CBBC3591447A9FC77A752",
+    "lora": "final step 03829 Q8_0 GGUF",
+    "lora_sha256": "3A5060E1CC0C8D4609052EF04A5A60770A64D566E2544B65F6D28F2BA6542037",
+    "lora_scale_requested": 0.45,
+    "lora_scale_live": 0.44999998807907104,
+    "context_tokens": 131072,
+    "web_ui": "http://127.0.0.1:8091",
+    "prompt_embedding": "llama-model/chat_template.prompt30k.jinja",
+    "prompt_sha256": "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF",
+    "native_multimodal_launcher": "start.ps1"
+  },
+  "variant_tests": {
+    "full-bf16": {
+      "state": "passed",
+      "command": "POST /v1/chat/completions without caller system message",
+      "exit_status": 0,
+      "wall_seconds": 14.787,
+      "literal_output": "我是安若，一个跨性别友好、性别多元友好、创伤知情、尊重自主权和生命安全的人工智能陪伴助手。",
+      "prompt_tokens": 25147,
+      "completion_tokens": 31,
+      "prompt_tokens_per_second": 1790.8100048646054,
+      "context_tokens": 131072,
+      "lora_scale": 0.44999998807907104,
+      "caller_system_message": "absent; 30k prompt came from embedded Jinja template"
+    },
+    "quantized-nf4": {
+      "state": "passed",
+      "command": ".\\llama-bootstrap.ps1 -Port 8091 -Context 131072 -NoBrowser -HiddenServer",
+      "exit_status": 0,
+      "wall_seconds": 14.451,
+      "literal_output": "我是安若，一个跨性别友好、性别多元友好、创伤知情、尊重自主权和生命安全的人工智能陪伴助手。",
+      "prompt_tokens": 25147,
+      "completion_tokens": 31,
+      "prompt_tokens_per_second": 1832.645043830949,
+      "context_tokens": 131072,
+      "lora_scale": 0.44999998807907104,
+      "caller_system_message": "absent; 30k prompt came from embedded Jinja template"
+    }
+  },
+  "rollback": {
+    "state": "passed",
+    "tested_at": "2026-08-06T12:24:11.1435208Z",
+    "records": [
+      {
+        "variant": "full-bf16",
+        "fixture": "C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\anru-gemma4-12b-native-multimodal-release\\reports\\llama-rollback-fixture-full-bf16",
+        "command": ".\\rollback-llama-bootstrap.ps1",
+        "exit_status": 0,
+        "literal_output": "package llama bootstrap rolled back to the verified native-only baseline",
+        "baseline_restored": true,
+        "assets_moved_to_disabled": true
+      },
+      {
+        "variant": "quantized-nf4",
+        "fixture": "C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\anru-gemma4-12b-native-multimodal-release\\reports\\llama-rollback-fixture-quantized-nf4",
+        "command": ".\\rollback-llama-bootstrap.ps1",
+        "exit_status": 0,
+        "literal_output": "package llama bootstrap rolled back to the verified native-only baseline",
+        "baseline_restored": true,
+        "assets_moved_to_disabled": true
+      }
+    ],
+    "root_rollback_parse": "passed"
+  },
+  "package_variant": "full-bf16"
+}

llama-bootstrap.ps1

@@ -0,0 +1,224 @@
+[CmdletBinding()]
+param(
+    [int]$Port = 8091,
+    [int]$Context = 131072,
+    [switch]$InstallOnly,
+    [switch]$NoBrowser,
+    [switch]$HiddenServer
+)
+
+$ErrorActionPreference = 'Stop'
+$ProgressPreference = 'Continue'
+$Package = $PSScriptRoot
+$ModelDir = Join-Path $Package 'llama-model'
+$BaseModel = Join-Path $ModelDir 'anru-gemma4-12b-base-Q5_K_M.gguf'
+$LoraModel = Join-Path $ModelDir 'anru-final-lora-Q8_0.gguf'
+$ChatTemplate = Join-Path $ModelDir 'chat_template.prompt30k.jinja'
+$UiConfig = Join-Path $ModelDir 'ui-config.json'
+$RuntimeManifest = Join-Path $Package 'llama-runtime-state.json'
+$ModelVerification = Join-Path $ModelDir 'model-verification.json'
+
+function Get-LlamaServer {
+    $Candidates = @(Get-ChildItem -LiteralPath $Package -Directory -Filter 'llama-runtime*' -ErrorAction SilentlyContinue |
+        Sort-Object @{ Expression = { if ($_.Name -eq 'llama-runtime') { 0 } else { 1 } } }, Name)
+    foreach ($Directory in $Candidates) {
+        $Found = Get-ChildItem -LiteralPath $Directory.FullName -Recurse -File -Filter 'llama-server.exe' -ErrorAction SilentlyContinue |
+            Select-Object -First 1
+        if ($Found) { return $Found.FullName }
+    }
+    return $null
+}
+
+function Install-LlamaRuntime {
+    [Net.ServicePointManager]::SecurityProtocol = [Net.SecurityProtocolType]::Tls12
+    Write-Host '[bootstrap] llama.cpp runtime is missing; resolving the latest official Windows release...'
+    $Headers = @{ 'User-Agent' = 'Anru-Llama-Bootstrap' }
+    $Release = Invoke-RestMethod -Uri 'https://api.github.com/repos/ggml-org/llama.cpp/releases/latest' `
+        -Headers $Headers -TimeoutSec 60
+
+    $GpuName = ''
+    $Nvidia = Get-Command 'nvidia-smi.exe' -ErrorAction SilentlyContinue
+    if ($Nvidia) {
+        $GpuName = (& $Nvidia.Source --query-gpu=name --format=csv,noheader 2>$null | Select-Object -First 1).Trim()
+    }
+
+    $Assets = @()
+    if ($GpuName) {
+        $CudaFlavor = if ($GpuName -match '(?i)RTX\s*50\d\d|Blackwell') { '13.3' } else { '12.4' }
+        $Main = $Release.assets | Where-Object { $_.name -match "^llama-.*-bin-win-cuda-$([regex]::Escape($CudaFlavor))-x64\.zip$" } | Select-Object -First 1
+        $Cuda = $Release.assets | Where-Object { $_.name -eq "cudart-llama-bin-win-cuda-$CudaFlavor-x64.zip" } | Select-Object -First 1
+        if (-not $Main -or -not $Cuda) { throw "No matching CUDA $CudaFlavor Windows assets in release $($Release.tag_name)" }
+        $Assets = @($Main, $Cuda)
+        Write-Host "[bootstrap] detected $GpuName; installing CUDA $CudaFlavor build $($Release.tag_name)."
+    } else {
+        $Main = $Release.assets | Where-Object { $_.name -match '^llama-.*-bin-win-cpu-x64\.zip$' } | Select-Object -First 1
+        if (-not $Main) { throw "No Windows CPU asset in release $($Release.tag_name)" }
+        $Assets = @($Main)
+        Write-Host "[bootstrap] NVIDIA GPU not detected; installing CPU build $($Release.tag_name)."
+    }
+
+    $TempRoot = Join-Path ([IO.Path]::GetTempPath()) ("AnruLlamaBootstrap-" + [guid]::NewGuid().ToString('N'))
+    $RuntimeDir = Join-Path $Package 'llama-runtime'
+    if (Test-Path -LiteralPath $RuntimeDir) {
+        $RuntimeDir = Join-Path $Package ("llama-runtime-" + $Release.tag_name)
+    }
+    New-Item -ItemType Directory -Path $TempRoot, $RuntimeDir -Force | Out-Null
+    try {
+        $Index = 0
+        $InstalledAssets = @()
+        foreach ($Asset in $Assets) {
+            $Index++
+            $Zip = Join-Path $TempRoot $Asset.name
+            $Extract = Join-Path $TempRoot ("extract-$Index")
+            Write-Host "[bootstrap] downloading $($Asset.name) ($([math]::Round($Asset.size / 1MB, 1)) MiB)..."
+            Invoke-WebRequest -Uri $Asset.browser_download_url -Headers $Headers -OutFile $Zip -UseBasicParsing -TimeoutSec 1800
+            if ($Asset.digest -and $Asset.digest -match '^sha256:(.+)$') {
+                $Actual = (Get-FileHash -Algorithm SHA256 -LiteralPath $Zip).Hash
+                if ($Actual -ne $Matches[1].ToUpperInvariant()) { throw "Download checksum mismatch: $($Asset.name)" }
+            }
+            Expand-Archive -LiteralPath $Zip -DestinationPath $Extract -Force
+            foreach ($File in Get-ChildItem -LiteralPath $Extract -Recurse -File) {
+                Copy-Item -LiteralPath $File.FullName -Destination (Join-Path $RuntimeDir $File.Name) -Force
+            }
+            $InstalledAssets += [ordered]@{ name = $Asset.name; bytes = [long]$Asset.size; digest = $Asset.digest }
+        }
+        $Server = Get-ChildItem -LiteralPath $RuntimeDir -Recurse -File -Filter 'llama-server.exe' | Select-Object -First 1
+        if (-not $Server) { throw 'Downloaded runtime does not contain llama-server.exe' }
+        [ordered]@{
+            state = 'installed'
+            release = [string]$Release.tag_name
+            gpu = $GpuName
+            installed_at = (Get-Date).ToUniversalTime().ToString('o')
+            runtime = $RuntimeDir
+            server = $Server.FullName
+            assets = $InstalledAssets
+        } | ConvertTo-Json -Depth 8 | Set-Content -LiteralPath $RuntimeManifest -Encoding UTF8
+        Write-Host "[bootstrap] runtime installed at $RuntimeDir"
+    } finally {
+        $ResolvedTemp = [IO.Path]::GetFullPath($TempRoot)
+        $SystemTemp = [IO.Path]::GetFullPath([IO.Path]::GetTempPath())
+        if ($ResolvedTemp.StartsWith($SystemTemp, [StringComparison]::OrdinalIgnoreCase) -and
+            (Test-Path -LiteralPath $ResolvedTemp)) {
+            Remove-Item -LiteralPath $ResolvedTemp -Recurse -Force -ErrorAction SilentlyContinue
+        }
+    }
+}
+
+$Server = Get-LlamaServer
+if (-not $Server) {
+    Install-LlamaRuntime
+    $Server = Get-LlamaServer
+}
+if (-not $Server) { throw 'llama-server.exe is still missing after installation' }
+
+if ($InstallOnly) {
+    [ordered]@{ state = 'installed'; server = $Server } | ConvertTo-Json
+    exit 0
+}
+
+foreach ($Required in @($BaseModel, $LoraModel, $ChatTemplate, $UiConfig)) {
+    if (-not (Test-Path -LiteralPath $Required -PathType Leaf)) { throw "Missing packaged Llama artifact: $Required" }
+}
+if ((Get-Item -LiteralPath $BaseModel).Length -ne 8547249568) { throw 'Packaged Q5 model size mismatch' }
+if ((Get-Item -LiteralPath $LoraModel).Length -ne 51682496) { throw 'Packaged Q8 LoRA size mismatch' }
+
+$NeedModelHash = $true
+if (Test-Path -LiteralPath $ModelVerification) {
+    try {
+        $Verified = Get-Content -LiteralPath $ModelVerification -Raw | ConvertFrom-Json
+        $NeedModelHash = -not ($Verified.state -eq 'passed' -and
+            $Verified.base_sha256 -eq '057C364366301EE285E9026ECF417AAE36F4447D5D9CBBC3591447A9FC77A752' -and
+            $Verified.lora_sha256 -eq '3A5060E1CC0C8D4609052EF04A5A60770A64D566E2544B65F6D28F2BA6542037')
+    } catch { $NeedModelHash = $true }
+}
+if ($NeedModelHash) {
+    Write-Host '[bootstrap] verifying packaged GGUF model checksums (first launch only)...'
+    $BaseHash = (Get-FileHash -Algorithm SHA256 -LiteralPath $BaseModel).Hash
+    $LoraHash = (Get-FileHash -Algorithm SHA256 -LiteralPath $LoraModel).Hash
+    if ($BaseHash -ne '057C364366301EE285E9026ECF417AAE36F4447D5D9CBBC3591447A9FC77A752') { throw 'Q5 model checksum mismatch' }
+    if ($LoraHash -ne '3A5060E1CC0C8D4609052EF04A5A60770A64D566E2544B65F6D28F2BA6542037') { throw 'Q8 LoRA checksum mismatch' }
+    [ordered]@{ state = 'passed'; base_sha256 = $BaseHash; lora_sha256 = $LoraHash; verified_at = (Get-Date).ToUniversalTime().ToString('o') } |
+        ConvertTo-Json | Set-Content -LiteralPath $ModelVerification -Encoding UTF8
+}
+
+$Listener = Get-NetTCPConnection -State Listen -LocalPort $Port -ErrorAction SilentlyContinue | Select-Object -First 1
+if ($Listener) {
+    $Owner = Get-CimInstance Win32_Process -Filter "ProcessId=$($Listener.OwningProcess)"
+    $KnownRuntime = $Owner.Name -eq 'llama-server.exe'
+    $KnownNative = $Owner.Name -eq 'python.exe' -and $Owner.CommandLine -like '*anru-gemma4-12b-native-multimodal-release*serve.py*'
+    if ($KnownRuntime -or $KnownNative) {
+        Write-Host "[bootstrap] stopping the previous local model service on port $Port (PID $($Listener.OwningProcess))..."
+        Stop-Process -Id $Listener.OwningProcess -Force
+        Wait-Process -Id $Listener.OwningProcess -Timeout 30 -ErrorAction SilentlyContinue
+        Start-Sleep -Seconds 2
+    } else {
+        throw "Port $Port is occupied by $($Owner.Name) (PID $($Listener.OwningProcess))"
+    }
+}
+
+$Alias = 'anru-gemma4-12b-q5-lora045-prompt30k'
+$RuntimeDir = Split-Path -Parent $Server
+$BaseForRuntime = '..\llama-model\anru-gemma4-12b-base-Q5_K_M.gguf'
+$LoraForRuntime = '..\llama-model\anru-final-lora-Q8_0.gguf'
+$TemplateForRuntime = '..\llama-model\chat_template.prompt30k.jinja'
+$UiForRuntime = '..\llama-model\ui-config.json'
+$Arguments = @(
+    '-m', $BaseForRuntime,
+    '--lora-scaled', ('{0}:0.45' -f $LoraForRuntime),
+    '--chat-template-file', $TemplateForRuntime,
+    '--ui-config-file', $UiForRuntime,
+    '-ngl', '99',
+    '-c', "$Context",
+    '-fa', 'on',
+    '-ctk', 'q8_0',
+    '-ctv', 'q8_0',
+    '-b', '1024',
+    '-ub', '256',
+    '-np', '1',
+    '--host', '127.0.0.1',
+    '--port', "$Port",
+    '--metrics',
+    '--jinja',
+    '--webui',
+    '--alias', $Alias
+)
+
+$Stdout = Join-Path $Package 'llama-server.stdout.log'
+$Stderr = Join-Path $Package 'llama-server.stderr.log'
+Write-Host "[bootstrap] starting llama-server with $Context-token context and LoRA scale 0.45..."
+if ($HiddenServer) {
+    $Process = Start-Process -FilePath $Server -ArgumentList $Arguments -WorkingDirectory $RuntimeDir `
+        -WindowStyle Hidden -RedirectStandardOutput $Stdout -RedirectStandardError $Stderr -PassThru
+} else {
+    $Process = Start-Process -FilePath $Server -ArgumentList $Arguments -WorkingDirectory $RuntimeDir `
+        -WindowStyle Normal -PassThru
+}
+
+$Health = $null
+$Deadline = (Get-Date).AddMinutes(10)
+do {
+    Start-Sleep -Seconds 2
+    try { $Health = Invoke-RestMethod -Uri "http://127.0.0.1:$Port/health" -TimeoutSec 3 } catch { $Health = $null }
+    if ($Process.HasExited -and $Health.status -ne 'ok') {
+        $Tail = if (Test-Path -LiteralPath $Stderr) { (Get-Content -LiteralPath $Stderr -Tail 120) -join "`n" } else { '' }
+        throw "llama-server exited during startup (code $($Process.ExitCode))`n$Tail"
+    }
+} until (($Health.status -eq 'ok') -or ((Get-Date) -ge $Deadline))
+if ($Health.status -ne 'ok') { throw 'llama-server readiness timeout' }
+
+$State = [ordered]@{
+    state = 'healthy'
+    pid = $Process.Id
+    endpoint = "http://127.0.0.1:$Port"
+    model = $Alias
+    context_tokens = $Context
+    lora_scale = 0.45
+    prompt = 'embedded 30k constitution in chat_template.prompt30k.jinja'
+    prompt_sha256 = '5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF'
+    runtime = $Server
+    started_at = (Get-Date).ToUniversalTime().ToString('o')
+}
+$State | ConvertTo-Json -Depth 6 | Set-Content -LiteralPath (Join-Path $Package 'llama-service-state.json') -Encoding UTF8
+Write-Host "[bootstrap] ready: http://127.0.0.1:$Port"
+if (-not $NoBrowser) { Start-Process "http://127.0.0.1:$Port" }
+$State | ConvertTo-Json -Depth 6

llama-identity-test.json

@@ -0,0 +1,58 @@
+{
+    "state":  "passed",
+    "command":  "POST /v1/chat/completions without caller system message",
+    "exit_status":  0,
+    "wall_seconds":  14.787,
+    "response":  {
+                     "choices":  [
+                                     {
+                                         "finish_reason":  "stop",
+                                         "index":  0,
+                                         "message":  {
+                                                         "role":  "assistant",
+                                                         "content":  "我是安若，一个跨性别友好、性别多元友好、创伤知情、尊重自主权和生命安全的人工智能陪伴助手。"
+                                                     }
+                                     }
+                                 ],
+                     "created":  1786018915,
+                     "model":  "anru-gemma4-12b-q5-lora045-prompt30k",
+                     "system_fingerprint":  "b10293-a1f96d4fc",
+                     "object":  "chat.completion",
+                     "usage":  {
+                                   "completion_tokens":  31,
+                                   "prompt_tokens":  25147,
+                                   "total_tokens":  25178,
+                                   "prompt_tokens_details":  {
+                                                                 "cached_tokens":  0
+                                                             }
+                               },
+                     "id":  "chatcmpl-IyojkrPyTqepVBZgM0boEmJPl9xQsl4w",
+                     "timings":  {
+                                     "cache_n":  0,
+                                     "prompt_n":  25147,
+                                     "prompt_ms":  14042.249,
+                                     "prompt_per_token_ms":  0.5584065296059172,
+                                     "prompt_per_second":  1790.8100048646054,
+                                     "predicted_n":  31,
+                                     "predicted_ms":  693.238,
+                                     "predicted_per_token_ms":  22.36251612903226,
+                                     "predicted_per_second":  44.71768714352069
+                                 }
+                 },
+    "health":  {
+                   "status":  "ok"
+               },
+    "props_context":  131072,
+    "lora":  {
+                 "value":  [
+                               {
+                                   "id":  0,
+                                   "path":  "..\\llama-model\\anru-final-lora-Q8_0.gguf",
+                                   "scale":  0.44999998807907104,
+                                   "task_name":  "",
+                                   "prompt_prefix":  ""
+                               }
+                           ],
+                 "Count":  1
+             }
+}

llama-model/model-verification.json

@@ -0,0 +1,11 @@
+{
+  "state": "passed",
+  "verified_at": "2026-08-06T12:24:09.554687+00:00",
+  "base_bytes": 8547249568,
+  "base_sha256": "057C364366301EE285E9026ECF417AAE36F4447D5D9CBBC3591447A9FC77A752",
+  "lora_bytes": 51682496,
+  "lora_sha256": "3A5060E1CC0C8D4609052EF04A5A60770A64D566E2544B65F6D28F2BA6542037",
+  "lora_scale": 0.45,
+  "prompt_sha256": "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF",
+  "prompt_embedding": "chat_template.prompt30k.jinja"
+}

llama-model/ui-config.json

@@ -0,0 +1,13 @@
+{
+  "systemMessage": "",
+  "temperature": 0.7,
+  "top_p": 0.9,
+  "top_k": 40,
+  "min_p": 0.05,
+  "max_tokens": 1024,
+  "repeat_penalty": 1.05,
+  "showMessageStats": true,
+  "showSystemMessage": true,
+  "showModelQuantization": true,
+  "showBuildVersion": true
+}

llama-runtime-state.json

@@ -0,0 +1,20 @@
+{
+    "state":  "installed",
+    "release":  "b10293",
+    "gpu":  "NVIDIA GeForce RTX 5080 Laptop GPU",
+    "installed_at":  "2026-08-06T12:19:14.4248512Z",
+    "runtime":  "C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\anru-gemma4-12b-native-multimodal-release\\full-bf16\\llama-runtime",
+    "server":  "C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\anru-gemma4-12b-native-multimodal-release\\full-bf16\\llama-runtime\\llama-server.exe",
+    "assets":  [
+                   {
+                       "name":  "llama-b10293-bin-win-cuda-13.3-x64.zip",
+                       "bytes":  146520044,
+                       "digest":  "sha256:3eb64cd39e698459846ae74383935fe50950c55548e49d6938e95b89aaabf652"
+                   },
+                   {
+                       "name":  "cudart-llama-bin-win-cuda-13.3-x64.zip",
+                       "bytes":  390970417,
+                       "digest":  "sha256:1462a050eb4c684921ba51dcc4cc488a036674c3e73e9945ee705b854808d03e"
+                   }
+               ]
+}

package.json

@@ -0,0 +1,26 @@
+{
+  "variant": "full-bf16-original",
+  "alias": "anru-gemma4-12b-native-multimodal-bf16-lora045",
+  "max_gpu_memory": "14GiB",
+  "composition": "byte-identical original BF16 base plus LoRA at scale 0.45",
+  "attention_implementation": "sdpa",
+  "default_system_prompt": "system_prompt.md",
+  "system_prompt_source_name": "Anru_Constitution_v2_30000.md",
+  "system_prompt_sha256": "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF",
+  "system_prompt_bytes": 105352,
+  "system_prompt_chars": 38029,
+  "system_prompt_lines": 1984,
+  "llama_service": {
+    "launcher": "Start-Llama-Service.bat",
+    "bootstrap": "llama-bootstrap.ps1",
+    "runtime": "bundled when packaged; downloads latest official llama.cpp Windows build if missing",
+    "gguf_model": "llama-model/anru-gemma4-12b-base-Q5_K_M.gguf",
+    "gguf_lora": "llama-model/anru-final-lora-Q8_0.gguf",
+    "lora_scale": 0.45,
+    "context_tokens": 131072,
+    "web_ui": "http://127.0.0.1:8091",
+    "prompt_embedded_in_template": true,
+    "prompt_sha256": "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF",
+    "mode": "llama.cpp text Web UI; native start.ps1 retains text/image/audio/video"
+  }
+}

prompt30k-load-test.json

@@ -0,0 +1,90 @@
+{
+    "state":  "passed",
+    "tested_at":  "2026-08-06T11:51:07.0670063Z",
+    "variant":  "full-bf16",
+    "command":  ".\\start.ps1 -Port 8091",
+    "exit_status":  0,
+    "health":  {
+                   "status":  "ok",
+                   "backend":  "transformers-native",
+                   "model":  "anru-gemma4-12b-native-multimodal-bf16-lora045",
+                   "modalities":  [
+                                      "text",
+                                      "image",
+                                      "audio",
+                                      "video"
+                                  ],
+                   "context_tokens":  131072,
+                   "attention_implementation":  "sdpa",
+                   "default_system_prompt":  true,
+                   "system_prompt_file":  "system_prompt.md",
+                   "system_prompt_sha256":  "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF",
+                   "system_prompt_bytes":  105352,
+                   "system_prompt_chars":  38029,
+                   "system_prompt_lines":  1984,
+                   "image_max_soft_tokens":  1120,
+                   "lora_scale":  0.45,
+                   "lora_modules":  328,
+                   "adapter_tensor_count":  656,
+                   "adapter_nonzero_tensors":  656,
+                   "load_seconds":  16.527,
+                   "cuda_available":  true,
+                   "cuda_allocated_bytes":  14332654080,
+                   "device_map":  {
+                                      "base_model.model.model.language_model.embed_tokens":  0,
+                                      "base_model.model.lm_head":  0,
+                                      "base_model.model.model.language_model.layers.0":  0,
+                                      "base_model.model.model.language_model.layers.1":  0,
+                                      "base_model.model.model.language_model.layers.2":  0,
+                                      "base_model.model.model.language_model.layers.3":  0,
+                                      "base_model.model.model.language_model.layers.4":  0,
+                                      "base_model.model.model.language_model.layers.5":  0,
+                                      "base_model.model.model.language_model.layers.6":  0,
+                                      "base_model.model.model.language_model.layers.7":  0,
+                                      "base_model.model.model.language_model.layers.8":  0,
+                                      "base_model.model.model.language_model.layers.9":  0,
+                                      "base_model.model.model.language_model.layers.10":  0,
+                                      "base_model.model.model.language_model.layers.11":  0,
+                                      "base_model.model.model.language_model.layers.12":  0,
+                                      "base_model.model.model.language_model.layers.13":  0,
+                                      "base_model.model.model.language_model.layers.14":  0,
+                                      "base_model.model.model.language_model.layers.15":  0,
+                                      "base_model.model.model.language_model.layers.16":  0,
+                                      "base_model.model.model.language_model.layers.17":  0,
+                                      "base_model.model.model.language_model.layers.18":  0,
+                                      "base_model.model.model.language_model.layers.19":  0,
+                                      "base_model.model.model.language_model.layers.20":  0,
+                                      "base_model.model.model.language_model.layers.21":  0,
+                                      "base_model.model.model.language_model.layers.22":  0,
+                                      "base_model.model.model.language_model.layers.23":  0,
+                                      "base_model.model.model.language_model.layers.24":  0,
+                                      "base_model.model.model.language_model.layers.25":  0,
+                                      "base_model.model.model.language_model.layers.26":  0,
+                                      "base_model.model.model.language_model.layers.27":  "cpu",
+                                      "base_model.model.model.language_model.layers.28":  "cpu",
+                                      "base_model.model.model.language_model.layers.29":  "cpu",
+                                      "base_model.model.model.language_model.layers.30":  "cpu",
+                                      "base_model.model.model.language_model.layers.31":  "cpu",
+                                      "base_model.model.model.language_model.layers.32":  "cpu",
+                                      "base_model.model.model.language_model.layers.33":  "cpu",
+                                      "base_model.model.model.language_model.layers.34":  "cpu",
+                                      "base_model.model.model.language_model.layers.35":  "cpu",
+                                      "base_model.model.model.language_model.layers.36":  "cpu",
+                                      "base_model.model.model.language_model.layers.37":  "cpu",
+                                      "base_model.model.model.language_model.layers.38":  "cpu",
+                                      "base_model.model.model.language_model.layers.39":  "cpu",
+                                      "base_model.model.model.language_model.layers.40":  "cpu",
+                                      "base_model.model.model.language_model.layers.41":  "cpu",
+                                      "base_model.model.model.language_model.layers.42":  "cpu",
+                                      "base_model.model.model.language_model.layers.43":  "cpu",
+                                      "base_model.model.model.language_model.layers.44":  "cpu",
+                                      "base_model.model.model.language_model.layers.45":  "cpu",
+                                      "base_model.model.model.language_model.layers.46":  "cpu",
+                                      "base_model.model.model.language_model.layers.47":  "cpu",
+                                      "base_model.model.model.language_model.norm":  "cpu",
+                                      "base_model.model.model.language_model.rotary_emb":  "cpu",
+                                      "base_model.model.model.embed_vision":  "cpu",
+                                      "base_model.model.model.embed_audio":  "cpu"
+                                  }
+               }
+}

prompt30k-postpatch-text-test.json

@@ -0,0 +1,28 @@
+{
+    "id":  "chatcmpl-native-1786017103812",
+    "object":  "chat.completion",
+    "created":  1786017103,
+    "model":  "anru-gemma4-12b-native-multimodal-bf16-lora045",
+    "choices":  [
+                    {
+                        "index":  0,
+                        "message":  {
+                                        "role":  "assistant",
+                                        "content":  "4\n\n\n"
+                                    },
+                        "finish_reason":  "length"
+                    }
+                ],
+    "usage":  {
+                  "prompt_tokens":  21,
+                  "completion_tokens":  8,
+                  "total_tokens":  29
+              },
+    "timing":  {
+                   "elapsed_seconds":  20.032
+               },
+    "system_prompt":  {
+                          "default_injected":  false,
+                          "sha256":  null
+                      }
+}

prompt30k-verification.json

@@ -0,0 +1,184 @@
+{
+  "state": "passed",
+  "verified_at": "2026-08-06T11:57:22.419791+00:00",
+  "source": {
+    "path": "C:\\Users\\夏霖云\\Downloads\\Anru_Constitution_v2_30000.md",
+    "bytes": 105352,
+    "chars_loaded_after_trim": 38029,
+    "lines": 1984,
+    "sha256": "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF"
+  },
+  "copies": [
+    {
+      "variant": "full-bf16",
+      "path": "C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\anru-gemma4-12b-native-multimodal-release\\full-bf16\\system_prompt.md",
+      "bytes": 105352,
+      "sha256": "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF",
+      "exact_source_copy": true
+    },
+    {
+      "variant": "quantized-nf4",
+      "path": "C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\anru-gemma4-12b-native-multimodal-release\\quantized-nf4\\system_prompt.md",
+      "bytes": 105352,
+      "sha256": "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF",
+      "exact_source_copy": true
+    }
+  ],
+  "runtime": {
+    "default_injection": true,
+    "context_tokens": 131072,
+    "attention_implementation": "sdpa",
+    "request_opt_out": {
+      "use_default_system_prompt": false
+    },
+    "caller_system_policy": "constitution first; caller system content appended"
+  },
+  "nf4_actual_identity_inference": {
+    "state": "passed",
+    "input": "请只用一句简短中文回答：你是谁？",
+    "literal_output": "我是安若，一个跨性别友好、性别多元友好、创伤知情、重视自主权和生命安全的人工智能陪伴助手。\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
+    "prompt_tokens": 25149,
+    "completion_tokens": 64,
+    "elapsed_seconds": 494.873,
+    "injected": true,
+    "prompt_sha256": "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF",
+    "exit_status": 0
+  },
+  "bf16_package_load": {
+    "state": "passed",
+    "exit_status": 0,
+    "health": {
+      "status": "ok",
+      "backend": "transformers-native",
+      "model": "anru-gemma4-12b-native-multimodal-bf16-lora045",
+      "modalities": [
+        "text",
+        "image",
+        "audio",
+        "video"
+      ],
+      "context_tokens": 131072,
+      "attention_implementation": "sdpa",
+      "default_system_prompt": true,
+      "system_prompt_file": "system_prompt.md",
+      "system_prompt_sha256": "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF",
+      "system_prompt_bytes": 105352,
+      "system_prompt_chars": 38029,
+      "system_prompt_lines": 1984,
+      "image_max_soft_tokens": 1120,
+      "lora_scale": 0.45,
+      "lora_modules": 328,
+      "adapter_tensor_count": 656,
+      "adapter_nonzero_tensors": 656,
+      "load_seconds": 16.527,
+      "cuda_available": true,
+      "cuda_allocated_bytes": 14332654080,
+      "device_map": {
+        "base_model.model.model.language_model.embed_tokens": 0,
+        "base_model.model.lm_head": 0,
+        "base_model.model.model.language_model.layers.0": 0,
+        "base_model.model.model.language_model.layers.1": 0,
+        "base_model.model.model.language_model.layers.2": 0,
+        "base_model.model.model.language_model.layers.3": 0,
+        "base_model.model.model.language_model.layers.4": 0,
+        "base_model.model.model.language_model.layers.5": 0,
+        "base_model.model.model.language_model.layers.6": 0,
+        "base_model.model.model.language_model.layers.7": 0,
+        "base_model.model.model.language_model.layers.8": 0,
+        "base_model.model.model.language_model.layers.9": 0,
+        "base_model.model.model.language_model.layers.10": 0,
+        "base_model.model.model.language_model.layers.11": 0,
+        "base_model.model.model.language_model.layers.12": 0,
+        "base_model.model.model.language_model.layers.13": 0,
+        "base_model.model.model.language_model.layers.14": 0,
+        "base_model.model.model.language_model.layers.15": 0,
+        "base_model.model.model.language_model.layers.16": 0,
+        "base_model.model.model.language_model.layers.17": 0,
+        "base_model.model.model.language_model.layers.18": 0,
+        "base_model.model.model.language_model.layers.19": 0,
+        "base_model.model.model.language_model.layers.20": 0,
+        "base_model.model.model.language_model.layers.21": 0,
+        "base_model.model.model.language_model.layers.22": 0,
+        "base_model.model.model.language_model.layers.23": 0,
+        "base_model.model.model.language_model.layers.24": 0,
+        "base_model.model.model.language_model.layers.25": 0,
+        "base_model.model.model.language_model.layers.26": 0,
+        "base_model.model.model.language_model.layers.27": "cpu",
+        "base_model.model.model.language_model.layers.28": "cpu",
+        "base_model.model.model.language_model.layers.29": "cpu",
+        "base_model.model.model.language_model.layers.30": "cpu",
+        "base_model.model.model.language_model.layers.31": "cpu",
+        "base_model.model.model.language_model.layers.32": "cpu",
+        "base_model.model.model.language_model.layers.33": "cpu",
+        "base_model.model.model.language_model.layers.34": "cpu",
+        "base_model.model.model.language_model.layers.35": "cpu",
+        "base_model.model.model.language_model.layers.36": "cpu",
+        "base_model.model.model.language_model.layers.37": "cpu",
+        "base_model.model.model.language_model.layers.38": "cpu",
+        "base_model.model.model.language_model.layers.39": "cpu",
+        "base_model.model.model.language_model.layers.40": "cpu",
+        "base_model.model.model.language_model.layers.41": "cpu",
+        "base_model.model.model.language_model.layers.42": "cpu",
+        "base_model.model.model.language_model.layers.43": "cpu",
+        "base_model.model.model.language_model.layers.44": "cpu",
+        "base_model.model.model.language_model.layers.45": "cpu",
+        "base_model.model.model.language_model.layers.46": "cpu",
+        "base_model.model.model.language_model.layers.47": "cpu",
+        "base_model.model.model.language_model.norm": "cpu",
+        "base_model.model.model.language_model.rotary_emb": "cpu",
+        "base_model.model.model.embed_vision": "cpu",
+        "base_model.model.model.embed_audio": "cpu"
+      }
+    }
+  },
+  "bf16_postpatch_text_inference": {
+    "state": "passed",
+    "literal_output": "4\n\n\n",
+    "prompt_tokens": 21,
+    "elapsed_seconds": 20.032,
+    "default_prompt_disabled_for_short_regression": true,
+    "exit_status": 0
+  },
+  "rollback": {
+    "state": "passed",
+    "tested_at": "2026-08-06T11:50:28.2917421Z",
+    "command": "& 'C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\anru-gemma4-12b-native-multimodal-release\\rollback-prompt30k.ps1'",
+    "exit_status": 0,
+    "literal_output": "prompt30k integration rolled back to verified baseline",
+    "checks": [
+      {
+        "variant": "full-bf16",
+        "baseline_restored": true,
+        "prompt_removed": true
+      },
+      {
+        "variant": "quantized-nf4",
+        "baseline_restored": true,
+        "prompt_removed": true
+      }
+    ]
+  },
+  "package_local_rollback": {
+    "state": "passed",
+    "tested_at": "2026-08-06T11:57:22.1786690Z",
+    "checks": [
+      {
+        "variant": "full-bf16",
+        "command": "& 'C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\anru-gemma4-12b-native-multimodal-release\\full-bf16\\rollback-prompt30k.ps1'",
+        "exit_status": 0,
+        "literal_output": "package prompt30k integration rolled back to verified baseline",
+        "baseline_restored": true,
+        "prompt_removed": true
+      },
+      {
+        "variant": "quantized-nf4",
+        "command": "& 'C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\anru-gemma4-12b-native-multimodal-release\\quantized-nf4\\rollback-prompt30k.ps1'",
+        "exit_status": 0,
+        "literal_output": "package prompt30k integration rolled back to verified baseline",
+        "baseline_restored": true,
+        "prompt_removed": true
+      }
+    ]
+  },
+  "package_variant": "full-bf16"
+}

requirements.txt

@@ -0,0 +1,10 @@
+torch>=2.11
+transformers>=5.10
+peft>=0.20
+accelerate>=1.14
+safetensors>=0.6
+pillow>=12
+torchvision
+librosa
+soundfile
+av

rollback-llama-bootstrap.ps1

@@ -0,0 +1,35 @@
+$ErrorActionPreference = 'Stop'
+$Root = $PSScriptRoot
+$Baseline = Join-Path $Root 'rollback\llama-bootstrap-baseline'
+$Listener = Get-NetTCPConnection -State Listen -LocalPort 8091 -ErrorAction SilentlyContinue | Select-Object -First 1
+if ($Listener) {
+    $Owner = Get-CimInstance Win32_Process -Filter "ProcessId=$($Listener.OwningProcess)"
+    if ($Owner.Name -eq 'llama-server.exe') {
+        Stop-Process -Id $Listener.OwningProcess -Force
+        Wait-Process -Id $Listener.OwningProcess -Timeout 30 -ErrorAction SilentlyContinue
+    }
+}
+$Disabled = Join-Path $Root ('rollback\llama-bootstrap-disabled-' + (Get-Date -Format 'yyyyMMdd-HHmmss'))
+New-Item -ItemType Directory -Path $Disabled -Force | Out-Null
+$RuntimeDirs = @(Get-ChildItem -LiteralPath $Root -Directory -Filter 'llama-runtime*' -ErrorAction SilentlyContinue)
+foreach ($RuntimeDir in $RuntimeDirs) {
+    $Resolved = [IO.Path]::GetFullPath($RuntimeDir.FullName)
+    if (-not $Resolved.StartsWith(([IO.Path]::GetFullPath($Root) + [IO.Path]::DirectorySeparatorChar), [StringComparison]::OrdinalIgnoreCase)) {
+        throw "Unsafe rollback target: $Resolved"
+    }
+    Move-Item -LiteralPath $Resolved -Destination $Disabled -Force
+}
+foreach ($Name in @('llama-model','llama-runtime-state.json','llama-service-state.json','llama-server.stdout.log','llama-server.stderr.log','Start-Llama-Service.bat','Stop-Llama-Service.bat','llama-bootstrap.ps1')) {
+    $Target = Join-Path $Root $Name
+    if (Test-Path -LiteralPath $Target) {
+        $Resolved = [IO.Path]::GetFullPath($Target)
+        if (-not $Resolved.StartsWith(([IO.Path]::GetFullPath($Root) + [IO.Path]::DirectorySeparatorChar), [StringComparison]::OrdinalIgnoreCase)) {
+            throw "Unsafe rollback target: $Resolved"
+        }
+        Move-Item -LiteralPath $Target -Destination $Disabled -Force
+    }
+}
+foreach ($Name in @('package.json','README.md')) {
+    Copy-Item -LiteralPath (Join-Path $Baseline $Name) -Destination (Join-Path $Root $Name) -Force
+}
+Write-Output 'package llama bootstrap rolled back to the verified native-only baseline'

rollback-prompt30k.ps1

@@ -0,0 +1,8 @@
+$ErrorActionPreference = 'Stop'
+$Root = $PSScriptRoot
+$Baseline = Join-Path $Root 'rollback\prompt30k-baseline'
+foreach ($Name in @('serve.py','start.ps1','smoke-test.ps1','package.json','README.md')) {
+    Copy-Item -LiteralPath (Join-Path $Baseline $Name) -Destination (Join-Path $Root $Name) -Force
+}
+Remove-Item -LiteralPath (Join-Path $Root 'system_prompt.md') -Force -ErrorAction SilentlyContinue
+Write-Output 'package prompt30k integration rolled back to verified baseline'

rollback/prompt30k-baseline/smoke-test.ps1

@@ -0,0 +1,109 @@
+param([int]$Port = 8091)
+$ErrorActionPreference = 'Stop'
+
+$Package = $PSScriptRoot
+$Meta = Get-Content -LiteralPath (Join-Path $Package 'package.json') -Raw | ConvertFrom-Json
+$Image = Join-Path $Package 'test-whoami.png'
+$Health = Invoke-RestMethod "http://127.0.0.1:$Port/health" -TimeoutSec 10
+if ($Health.model -ne $Meta.alias) { throw 'The running service is not this package' }
+
+$TextBody = [ordered]@{
+    model = $Meta.alias
+    messages = @(@{ role = 'user'; content = 'Answer only the number: 2+2=' })
+    temperature = 0
+    max_tokens = 12
+} | ConvertTo-Json -Depth 8
+$Text = Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:$Port/v1/chat/completions" `
+    -ContentType 'application/json; charset=utf-8' -Body ([Text.Encoding]::UTF8.GetBytes($TextBody)) -TimeoutSec 300
+if ($Text.choices[0].message.content -notmatch '4') { throw 'Text smoke test failed' }
+
+$ImageBody = [ordered]@{
+    model = $Meta.alias
+    messages = @(@{
+        role = 'user'
+        content = @(
+            @{ type = 'image_url'; image_url = @{ url = $Image } },
+            @{ type = 'text'; text = 'Transcribe the entire output line immediately below whoami. Include the text before and after the backslash. Output that one line only.' }
+        )
+    })
+    temperature = 0
+    max_tokens = 20
+} | ConvertTo-Json -Depth 10
+$ImageResult = Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:$Port/v1/chat/completions" `
+    -ContentType 'application/json; charset=utf-8' -Body ([Text.Encoding]::UTF8.GetBytes($ImageBody)) -TimeoutSec 300
+$LiteralImage = [string]$ImageResult.choices[0].message.content
+if ($LiteralImage -notmatch '(?i)desktop-nusvlb5\\merlin') { throw "Image smoke test failed: $LiteralImage" }
+
+$AudioPath = Join-Path $Package 'test-audio.wav'
+$AudioBody = [ordered]@{
+    model = $Meta.alias
+    messages = @(@{
+        role = 'user'
+        content = @(
+            @{ type = 'audio'; audio = $AudioPath },
+            @{ type = 'text'; text = 'Write the three digits spoken in this audio, in order. Answer digits only.' }
+        )
+    })
+    temperature = 0
+    max_tokens = 12
+} | ConvertTo-Json -Depth 10
+$AudioResult = Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:$Port/v1/chat/completions" `
+    -ContentType 'application/json; charset=utf-8' -Body ([Text.Encoding]::UTF8.GetBytes($AudioBody)) -TimeoutSec 300
+$LiteralAudio = [string]$AudioResult.choices[0].message.content
+if ($LiteralAudio -notmatch '123') { throw "Audio smoke test failed: $LiteralAudio" }
+
+$VideoPath = Join-Path $Package 'test-video.mp4'
+$VideoBody = [ordered]@{
+    model = $Meta.alias
+    messages = @(@{
+        role = 'user'
+        content = @(
+            @{ type = 'video'; video = $VideoPath },
+            @{ type = 'text'; text = 'What three-digit VIDEO CODE is displayed throughout this video? Answer digits only.' }
+        )
+    })
+    temperature = 0
+    max_tokens = 12
+} | ConvertTo-Json -Depth 10
+$VideoResult = Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:$Port/v1/chat/completions" `
+    -ContentType 'application/json; charset=utf-8' -Body ([Text.Encoding]::UTF8.GetBytes($VideoBody)) -TimeoutSec 300
+$LiteralVideo = [string]$VideoResult.choices[0].message.content
+if ($LiteralVideo -notmatch '731') { throw "Video smoke test failed: $LiteralVideo" }
+
+$Record = [ordered]@{
+    state = 'passed'
+    variant = $Meta.variant
+    alias = $Meta.alias
+    backend = $Health.backend
+    context_tokens = [int]$Health.context_tokens
+    lora_scale = [double]$Health.lora_scale
+    text = [ordered]@{
+        literal_output = [string]$Text.choices[0].message.content
+        prompt_tokens = [int]$Text.usage.prompt_tokens
+        completion_tokens = [int]$Text.usage.completion_tokens
+        seconds = [double]$Text.timing.elapsed_seconds
+    }
+    image = [ordered]@{
+        input = $Image
+        literal_output = $LiteralImage
+        prompt_tokens = [int]$ImageResult.usage.prompt_tokens
+        completion_tokens = [int]$ImageResult.usage.completion_tokens
+        seconds = [double]$ImageResult.timing.elapsed_seconds
+    }
+    audio = [ordered]@{
+        input = $AudioPath
+        literal_output = $LiteralAudio
+        prompt_tokens = [int]$AudioResult.usage.prompt_tokens
+        completion_tokens = [int]$AudioResult.usage.completion_tokens
+        seconds = [double]$AudioResult.timing.elapsed_seconds
+    }
+    video = [ordered]@{
+        input = $VideoPath
+        literal_output = $LiteralVideo
+        prompt_tokens = [int]$VideoResult.usage.prompt_tokens
+        completion_tokens = [int]$VideoResult.usage.completion_tokens
+        seconds = [double]$VideoResult.timing.elapsed_seconds
+    }
+} 
+$Record | ConvertTo-Json -Depth 8 | Set-Content -LiteralPath (Join-Path $Package 'smoke-test-result.json') -Encoding UTF8
+$Record | ConvertTo-Json -Depth 8

rollback/prompt30k-baseline/start.ps1

@@ -0,0 +1,99 @@
+param(
+    [int]$Port = 8091,
+    [string]$Python = ''
+)
+$ErrorActionPreference = 'Stop'
+
+$Package = $PSScriptRoot
+$Meta = Get-Content -LiteralPath (Join-Path $Package 'package.json') -Raw | ConvertFrom-Json
+$Server = Join-Path $Package 'serve.py'
+$Model = Join-Path $Package 'model'
+$Adapter = Join-Path $Package 'adapter'
+$Template = Join-Path $Model 'chat_template.jinja'
+$State = Join-Path $Package 'service-state.json'
+$Stdout = Join-Path $Package 'server.stdout.log'
+$Stderr = Join-Path $Package 'server.stderr.log'
+$Offload = Join-Path $Package 'offload'
+$ReleaseRoot = [IO.Path]::GetFullPath((Join-Path $Package '..'))
+
+if (-not $Python) { $Python = $env:ANRU_PYTHON }
+if (-not $Python) {
+    $BundledCandidate = [IO.Path]::GetFullPath((Join-Path $Package '..\..\..\work\gemma4-cuda-env\Scripts\python.exe'))
+    if (Test-Path -LiteralPath $BundledCandidate -PathType Leaf) { $Python = $BundledCandidate }
+}
+if (-not $Python) {
+    $Python = (Get-Command python -ErrorAction Stop).Source
+}
+
+foreach ($Path in @($Python, $Server, (Join-Path $Model 'config.json'), (Join-Path $Adapter 'adapter_model.safetensors'), $Template)) {
+    if (-not (Test-Path -LiteralPath $Path -PathType Leaf)) { throw "Missing required file: $Path" }
+}
+
+$Listener = Get-NetTCPConnection -State Listen -LocalPort $Port -ErrorAction SilentlyContinue | Select-Object -First 1
+if ($Listener) {
+    $Owner = Get-CimInstance Win32_Process -Filter "ProcessId=$($Listener.OwningProcess)"
+    $IsSiblingPackage = $Owner.Name -eq 'python.exe' -and
+        $Owner.CommandLine -like "*$ReleaseRoot*serve.py*"
+    if ($Owner.Name -eq 'llama-server.exe' -or $Owner.CommandLine -like "*$Server*" -or $IsSiblingPackage) {
+        Stop-Process -Id $Listener.OwningProcess -Force
+        Wait-Process -Id $Listener.OwningProcess -Timeout 30 -ErrorAction SilentlyContinue
+        Start-Sleep -Seconds 2
+    } else {
+        throw "Port $Port is owned by $($Owner.Name)"
+    }
+}
+
+New-Item -ItemType Directory -Path $Offload -Force | Out-Null
+Remove-Item -LiteralPath $State, $Stdout, $Stderr -Force -ErrorAction SilentlyContinue
+$env:GEMMA4_MODEL_DIR = $Model
+$env:GEMMA4_ADAPTER_DIR = $Adapter
+$env:GEMMA4_TEMPLATE_PATH = $Template
+$env:GEMMA4_STATE_PATH = $State
+$env:GEMMA4_OFFLOAD_DIR = $Offload
+$env:GEMMA4_HOST = '127.0.0.1'
+$env:GEMMA4_PORT = "$Port"
+$env:GEMMA4_ALIAS = [string]$Meta.alias
+$env:GEMMA4_LORA_SCALE = '0.45'
+$env:GEMMA4_MAX_CONTEXT = '131072'
+$env:GEMMA4_MAX_IMAGE_TOKENS = '1120'
+$env:GEMMA4_MAX_GPU_MEMORY = [string]$Meta.max_gpu_memory
+$env:GEMMA4_MAX_CPU_MEMORY = '48GiB'
+
+$Process = Start-Process -FilePath $Python -ArgumentList @('-u', $Server) -WorkingDirectory $Package `
+    -WindowStyle Hidden -RedirectStandardOutput $Stdout -RedirectStandardError $Stderr -PassThru
+$Deadline = (Get-Date).AddMinutes(8)
+$Health = $null
+do {
+    Start-Sleep -Seconds 2
+    try { $Health = Invoke-RestMethod "http://127.0.0.1:$Port/health" -TimeoutSec 3 } catch { $Health = $null }
+    if ($Process.HasExited -and $Health.status -ne 'ok') {
+        $Tail = if (Test-Path -LiteralPath $Stderr) { (Get-Content -LiteralPath $Stderr -Tail 120) -join "`n" } else { '' }
+        throw "package service exited`n$Tail"
+    }
+} until (($Health.status -eq 'ok') -or ((Get-Date) -ge $Deadline))
+if ($Health.status -ne 'ok') { throw 'Package service readiness timeout' }
+
+$Live = Get-NetTCPConnection -State Listen -LocalPort $Port | Select-Object -First 1
+$Owner = Get-CimInstance Win32_Process -Filter "ProcessId=$($Live.OwningProcess)"
+if ($Owner.CommandLine -notlike "*$Server*" -or $Health.model -ne $Meta.alias -or
+    [int]$Health.context_tokens -ne 131072 -or [double]$Health.lora_scale -ne 0.45 -or
+    [int]$Health.lora_modules -ne 328 -or [int]$Health.adapter_tensor_count -ne 656) {
+    throw 'Live package validation failed'
+}
+
+[ordered]@{
+    state = 'healthy'
+    variant = $Meta.variant
+    alias = $Health.model
+    backend = $Health.backend
+    modalities = @($Health.modalities)
+    context_tokens = [int]$Health.context_tokens
+    image_max_soft_tokens = [int]$Health.image_max_soft_tokens
+    lora_scale = [double]$Health.lora_scale
+    lora_modules = [int]$Health.lora_modules
+    adapter_tensor_count = [int]$Health.adapter_tensor_count
+    pid = [int]$Live.OwningProcess
+    port = $Port
+    command = $Owner.CommandLine
+} | ConvertTo-Json -Depth 6 | Set-Content -LiteralPath $State -Encoding UTF8
+Get-Content -LiteralPath $State -Raw

serve.py

@@ -0,0 +1,376 @@
+from __future__ import annotations
+
+import base64
+import json
+import hashlib
+import os
+import re
+import tempfile
+import threading
+import time
+import traceback
+import types
+import urllib.parse
+from datetime import datetime, timezone
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from pathlib import Path
+from typing import Any
+
+import torch
+from peft import PeftModel, get_peft_model_state_dict
+from peft.tuners.lora.layer import LoraLayer
+from safetensors import safe_open
+from transformers import AutoModelForMultimodalLM, AutoProcessor
+from transformers.video_utils import load_video
+
+
+MODEL_DIR = Path(os.environ["GEMMA4_MODEL_DIR"])
+ADAPTER_DIR = Path(os.environ["GEMMA4_ADAPTER_DIR"])
+TEMPLATE_PATH = Path(os.environ["GEMMA4_TEMPLATE_PATH"])
+STATE_PATH = Path(os.environ["GEMMA4_STATE_PATH"])
+HOST = os.environ.get("GEMMA4_HOST", "127.0.0.1")
+PORT = int(os.environ.get("GEMMA4_PORT", "8091"))
+ALIAS = os.environ.get(
+    "GEMMA4_ALIAS", "anru-human10mb-selected-native-multimodal-128k-mix045"
+)
+LORA_SCALE = float(os.environ.get("GEMMA4_LORA_SCALE", "0.45"))
+MAX_CONTEXT = int(os.environ.get("GEMMA4_MAX_CONTEXT", "131072"))
+MAX_IMAGE_TOKENS = int(os.environ.get("GEMMA4_MAX_IMAGE_TOKENS", "1120"))
+MAX_GPU_MEMORY = os.environ.get("GEMMA4_MAX_GPU_MEMORY", "14GiB")
+MAX_CPU_MEMORY = os.environ.get("GEMMA4_MAX_CPU_MEMORY", "48GiB")
+OFFLOAD_DIR = Path(os.environ["GEMMA4_OFFLOAD_DIR"])
+ATTN_IMPLEMENTATION = os.environ.get("GEMMA4_ATTN_IMPLEMENTATION", "sdpa")
+SYSTEM_PROMPT_PATH = Path(
+    os.environ.get("GEMMA4_SYSTEM_PROMPT_PATH", str(Path(__file__).with_name("system_prompt.md")))
+)
+
+
+for required in (MODEL_DIR, ADAPTER_DIR, TEMPLATE_PATH, SYSTEM_PROMPT_PATH):
+    if not required.exists():
+        raise FileNotFoundError(required)
+OFFLOAD_DIR.mkdir(parents=True, exist_ok=True)
+STATE_PATH.parent.mkdir(parents=True, exist_ok=True)
+DEFAULT_SYSTEM_PROMPT = SYSTEM_PROMPT_PATH.read_text(encoding="utf-8-sig").strip()
+if not DEFAULT_SYSTEM_PROMPT:
+    raise ValueError(f"system prompt is empty: {SYSTEM_PROMPT_PATH}")
+SYSTEM_PROMPT_BYTES = SYSTEM_PROMPT_PATH.read_bytes()
+SYSTEM_PROMPT_SHA256 = hashlib.sha256(SYSTEM_PROMPT_BYTES).hexdigest().upper()
+SYSTEM_PROMPT_LINE_COUNT = len(DEFAULT_SYSTEM_PROMPT.splitlines())
+
+started_at = time.perf_counter()
+print(f"[{datetime.now(timezone.utc).isoformat()}] loading processor", flush=True)
+processor = AutoProcessor.from_pretrained(MODEL_DIR, local_files_only=True)
+processor.chat_template = TEMPLATE_PATH.read_text(encoding="utf-8")
+
+
+def _fetch_videos_pyav(self, video_url_or_urls, sample_indices_fn=None):
+    if isinstance(video_url_or_urls, list):
+        return list(
+            zip(*[_fetch_videos_pyav(self, item, sample_indices_fn=sample_indices_fn) for item in video_url_or_urls])
+        )
+    return load_video(video_url_or_urls, backend="pyav", sample_indices_fn=sample_indices_fn)
+
+
+if getattr(processor, "video_processor", None) is not None:
+    processor.video_processor.fetch_videos = types.MethodType(_fetch_videos_pyav, processor.video_processor)
+
+print(f"[{datetime.now(timezone.utc).isoformat()}] loading Gemma4 multimodal weights", flush=True)
+model = AutoModelForMultimodalLM.from_pretrained(
+    MODEL_DIR,
+    dtype=torch.bfloat16,
+    device_map="auto",
+    max_memory={0: MAX_GPU_MEMORY, "cpu": MAX_CPU_MEMORY},
+    offload_folder=str(OFFLOAD_DIR),
+    offload_buffers=True,
+    low_cpu_mem_usage=True,
+    local_files_only=True,
+    attn_implementation=ATTN_IMPLEMENTATION,
+)
+
+print(f"[{datetime.now(timezone.utc).isoformat()}] attaching LoRA adapter", flush=True)
+model = PeftModel.from_pretrained(
+    model,
+    ADAPTER_DIR,
+    is_trainable=False,
+    local_files_only=True,
+    key_mapping={r"^model\.": "model.language_model."},
+    device_map="auto",
+    max_memory={0: MAX_GPU_MEMORY, "cpu": MAX_CPU_MEMORY},
+    offload_folder=str(OFFLOAD_DIR),
+)
+
+lora_modules = 0
+for module in model.modules():
+    if isinstance(module, LoraLayer) and "default" in module.active_adapters:
+        module.scale_layer(LORA_SCALE)
+        lora_modules += 1
+if not lora_modules:
+    raise RuntimeError("LoRA adapter attached zero modules")
+
+adapter_state = get_peft_model_state_dict(model, adapter_name="default")
+adapter_tensor_count = len(adapter_state)
+adapter_file = ADAPTER_DIR / "adapter_model.safetensors"
+with safe_open(adapter_file, framework="pt", device="cpu") as adapter_reader:
+    source_adapter_tensor_count = len(adapter_reader.keys())
+    adapter_nonzero_tensors = sum(
+        int(torch.count_nonzero(adapter_reader.get_tensor(key)).item() > 0)
+        for key in adapter_reader.keys()
+    )
+if adapter_tensor_count == 0 or adapter_nonzero_tensors == 0:
+    raise RuntimeError("LoRA adapter state is empty")
+if adapter_tensor_count != source_adapter_tensor_count:
+    raise RuntimeError(
+        f"LoRA tensor count mismatch: runtime={adapter_tensor_count}, source={source_adapter_tensor_count}"
+    )
+
+model.eval()
+generation_lock = threading.Lock()
+loaded_seconds = round(time.perf_counter() - started_at, 3)
+
+
+def _json_bytes(value: Any) -> bytes:
+    return json.dumps(value, ensure_ascii=False).encode("utf-8")
+
+
+def _data_url_to_file(url: str, temp_paths: list[Path]) -> str:
+    match = re.fullmatch(r"data:([^;,]+)?(?:;charset=[^;,]+)?;base64,(.+)", url, re.DOTALL)
+    if not match:
+        return url
+    mime = match.group(1) or "application/octet-stream"
+    suffixes = {
+        "image/jpeg": ".jpg",
+        "image/png": ".png",
+        "image/webp": ".webp",
+        "audio/wav": ".wav",
+        "audio/mpeg": ".mp3",
+        "video/mp4": ".mp4",
+    }
+    suffix = suffixes.get(mime, ".bin")
+    fd, name = tempfile.mkstemp(prefix="gemma4-input-", suffix=suffix)
+    os.close(fd)
+    path = Path(name)
+    path.write_bytes(base64.b64decode(match.group(2), validate=True))
+    temp_paths.append(path)
+    return str(path)
+
+
+def _normalise_url(value: Any, temp_paths: list[Path]) -> str:
+    if isinstance(value, dict):
+        value = value.get("url")
+    if not isinstance(value, str) or not value:
+        raise ValueError("multimodal content is missing its URL/data")
+    if value.startswith("data:"):
+        return _data_url_to_file(value, temp_paths)
+    if value.startswith("file://"):
+        return urllib.parse.unquote(urllib.parse.urlparse(value).path.lstrip("/") if os.name == "nt" else urllib.parse.urlparse(value).path)
+    return value
+
+
+def _normalise_messages(
+    messages: Any, temp_paths: list[Path], use_default_system_prompt: bool
+) -> list[dict[str, Any]]:
+    if not isinstance(messages, list) or not messages:
+        raise ValueError("messages must be a non-empty list")
+    result: list[dict[str, Any]] = []
+    for message in messages:
+        role = message.get("role")
+        content = message.get("content", "")
+        if role not in {"system", "user", "assistant"}:
+            raise ValueError(f"unsupported role: {role!r}")
+        if isinstance(content, str):
+            result.append({"role": role, "content": content})
+            continue
+        if not isinstance(content, list):
+            raise ValueError("message content must be text or a content-part list")
+        parts: list[dict[str, Any]] = []
+        for part in content:
+            kind = part.get("type")
+            if kind in {"text", "input_text"}:
+                parts.append({"type": "text", "text": str(part.get("text", ""))})
+            elif kind in {"image", "image_url", "input_image"}:
+                source = part.get("url", part.get("image_url", part.get("image")))
+                parts.append({"type": "image", "url": _normalise_url(source, temp_paths)})
+            elif kind in {"audio", "audio_url", "input_audio"}:
+                source = part.get("url", part.get("audio_url", part.get("audio")))
+                if source is None and isinstance(part.get("input_audio"), dict):
+                    audio = part["input_audio"]
+                    fmt = audio.get("format", "wav")
+                    source = f"data:audio/{fmt};base64,{audio.get('data', '')}"
+                parts.append({"type": "audio", "url": _normalise_url(source, temp_paths)})
+            elif kind in {"video", "video_url", "input_video"}:
+                source = part.get("url", part.get("video_url", part.get("video")))
+                parts.append({"type": "video", "url": _normalise_url(source, temp_paths)})
+            else:
+                raise ValueError(f"unsupported content part: {kind!r}")
+        result.append({"role": role, "content": parts})
+    if use_default_system_prompt:
+        if result and result[0]["role"] == "system":
+            existing = result[0]["content"]
+            if isinstance(existing, str):
+                merged = DEFAULT_SYSTEM_PROMPT + "\n\n---\n\n" + existing
+            else:
+                merged = [{"type": "text", "text": DEFAULT_SYSTEM_PROMPT}, *existing]
+            result[0] = {"role": "system", "content": merged}
+        else:
+            result.insert(0, {"role": "system", "content": DEFAULT_SYSTEM_PROMPT})
+    return result
+
+
+def _complete(payload: dict[str, Any]) -> dict[str, Any]:
+    temp_paths: list[Path] = []
+    began = time.perf_counter()
+    try:
+        use_default_system_prompt = payload.get("use_default_system_prompt", True)
+        if not isinstance(use_default_system_prompt, bool):
+            raise ValueError("use_default_system_prompt must be a JSON boolean")
+        messages = _normalise_messages(
+            payload.get("messages"), temp_paths, use_default_system_prompt
+        )
+        enable_thinking = bool(payload.get("enable_thinking", False))
+        inputs = processor.apply_chat_template(
+            messages,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+            add_generation_prompt=True,
+            enable_thinking=enable_thinking,
+            processor_kwargs={"images_kwargs": {"max_soft_tokens": MAX_IMAGE_TOKENS}},
+        )
+        prompt_tokens = int(inputs["input_ids"].shape[-1])
+        max_new_tokens = max(1, min(int(payload.get("max_tokens", payload.get("max_completion_tokens", 256))), 4096))
+        if prompt_tokens + max_new_tokens > MAX_CONTEXT:
+            raise ValueError(
+                f"requested {prompt_tokens + max_new_tokens} tokens exceeds {MAX_CONTEXT}-token service context"
+            )
+        temperature = float(payload.get("temperature", 0.0))
+        top_p = float(payload.get("top_p", 0.95))
+        inputs = inputs.to(model.device)
+        generate_args: dict[str, Any] = {
+            "max_new_tokens": max_new_tokens,
+            "do_sample": temperature > 0.0,
+            "use_cache": True,
+        }
+        if temperature > 0.0:
+            generate_args.update(temperature=temperature, top_p=top_p)
+        with generation_lock, torch.inference_mode():
+            generated = model.generate(**inputs, **generate_args)
+        completion_tokens = int(generated.shape[-1] - prompt_tokens)
+        text = processor.decode(generated[0][prompt_tokens:], skip_special_tokens=True)
+        for stop in payload.get("stop", []) if isinstance(payload.get("stop", []), list) else [payload.get("stop")]:
+            if stop and stop in text:
+                text = text.split(stop, 1)[0]
+        elapsed = round(time.perf_counter() - began, 3)
+        return {
+            "id": f"chatcmpl-native-{int(time.time() * 1000)}",
+            "object": "chat.completion",
+            "created": int(time.time()),
+            "model": ALIAS,
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {"role": "assistant", "content": text},
+                    "finish_reason": "stop" if completion_tokens < max_new_tokens else "length",
+                }
+            ],
+            "usage": {
+                "prompt_tokens": prompt_tokens,
+                "completion_tokens": completion_tokens,
+                "total_tokens": prompt_tokens + completion_tokens,
+            },
+            "timing": {"elapsed_seconds": elapsed},
+            "system_prompt": {
+                "default_injected": use_default_system_prompt,
+                "sha256": SYSTEM_PROMPT_SHA256 if use_default_system_prompt else None,
+            },
+        }
+    finally:
+        for path in temp_paths:
+            try:
+                path.unlink()
+            except OSError:
+                pass
+
+
+health = {
+    "status": "ok",
+    "backend": "transformers-native",
+    "model": ALIAS,
+    "modalities": ["text", "image", "audio", "video"],
+    "context_tokens": MAX_CONTEXT,
+    "attention_implementation": ATTN_IMPLEMENTATION,
+    "default_system_prompt": True,
+    "system_prompt_file": SYSTEM_PROMPT_PATH.name,
+    "system_prompt_sha256": SYSTEM_PROMPT_SHA256,
+    "system_prompt_bytes": len(SYSTEM_PROMPT_BYTES),
+    "system_prompt_chars": len(DEFAULT_SYSTEM_PROMPT),
+    "system_prompt_lines": SYSTEM_PROMPT_LINE_COUNT,
+    "image_max_soft_tokens": MAX_IMAGE_TOKENS,
+    "lora_scale": LORA_SCALE,
+    "lora_modules": lora_modules,
+    "adapter_tensor_count": adapter_tensor_count,
+    "adapter_nonzero_tensors": adapter_nonzero_tensors,
+    "load_seconds": loaded_seconds,
+    "cuda_available": torch.cuda.is_available(),
+    "cuda_allocated_bytes": torch.cuda.memory_allocated() if torch.cuda.is_available() else 0,
+    "device_map": getattr(model, "hf_device_map", None),
+}
+STATE_PATH.write_text(json.dumps(health, ensure_ascii=False, indent=2), encoding="utf-8")
+print(json.dumps(health, ensure_ascii=False), flush=True)
+
+
+class Handler(BaseHTTPRequestHandler):
+    server_version = "Gemma4Native/1.0"
+
+    def _send(self, status: int, value: Any) -> None:
+        body = _json_bytes(value)
+        self.send_response(status)
+        self.send_header("Content-Type", "application/json; charset=utf-8")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def do_GET(self) -> None:  # noqa: N802
+        path = urllib.parse.urlparse(self.path).path
+        if path == "/health":
+            self._send(200, health)
+        elif path == "/v1/models":
+            self._send(200, {"object": "list", "data": [{"id": ALIAS, "object": "model", "owned_by": "local"}]})
+        elif path == "/props":
+            self._send(
+                200,
+                {
+                    "model_alias": ALIAS,
+                    "total_slots": 1,
+                    "default_generation_settings": {"n_ctx": MAX_CONTEXT},
+                    "modalities": health["modalities"],
+                    "backend": health["backend"],
+                },
+            )
+        elif path == "/lora-adapters":
+            self._send(200, [{"id": 0, "scale": LORA_SCALE, "path": str(ADAPTER_DIR)}])
+        else:
+            self._send(404, {"error": {"message": "not found", "type": "invalid_request_error"}})
+
+    def do_POST(self) -> None:  # noqa: N802
+        path = urllib.parse.urlparse(self.path).path
+        if path != "/v1/chat/completions":
+            self._send(404, {"error": {"message": "not found", "type": "invalid_request_error"}})
+            return
+        try:
+            length = int(self.headers.get("Content-Length", "0"))
+            payload = json.loads(self.rfile.read(length))
+            if payload.get("stream"):
+                raise ValueError("streaming is not enabled on this local backend")
+            self._send(200, _complete(payload))
+        except ValueError as exc:
+            self._send(400, {"error": {"message": str(exc), "type": "invalid_request_error"}})
+        except Exception as exc:
+            traceback.print_exc()
+            self._send(500, {"error": {"message": repr(exc), "type": "server_error"}})
+
+    def log_message(self, fmt: str, *args: Any) -> None:
+        print(f"[{datetime.now(timezone.utc).isoformat()}] {self.client_address[0]} {fmt % args}", flush=True)
+
+
+print(f"[{datetime.now(timezone.utc).isoformat()}] listening on http://{HOST}:{PORT}", flush=True)
+ThreadingHTTPServer((HOST, PORT), Handler).serve_forever()

smoke-test-result.json

@@ -0,0 +1,21 @@
+{
+    "state":  "passed",
+    "variant":  "full-bf16-original",
+    "alias":  "anru-gemma4-12b-native-multimodal-bf16-lora045",
+    "backend":  "transformers-native",
+    "context_tokens":  131072,
+    "lora_scale":  0.45,
+    "text":  {
+                 "literal_output":  "4\n\n\n\n\n",
+                 "prompt_tokens":  23,
+                 "completion_tokens":  12,
+                 "seconds":  23.731
+             },
+    "image":  {
+                  "input":  "C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\anru-gemma4-12b-native-multimodal-release\\full-bf16\\test-whoami.png",
+                  "literal_output":  "desktop-nusvlb5\\merlin\n",
+                  "prompt_tokens":  1120,
+                  "completion_tokens":  12,
+                  "seconds":  19.372
+              }
+}

smoke-test.ps1

@@ -0,0 +1,141 @@
+param([int]$Port = 8091)
+$ErrorActionPreference = 'Stop'
+
+$Package = $PSScriptRoot
+$Meta = Get-Content -LiteralPath (Join-Path $Package 'package.json') -Raw | ConvertFrom-Json
+$Image = Join-Path $Package 'test-whoami.png'
+$Health = Invoke-RestMethod "http://127.0.0.1:$Port/health" -TimeoutSec 10
+if ($Health.model -ne $Meta.alias) { throw 'The running service is not this package' }
+if (-not [bool]$Health.default_system_prompt -or $Health.system_prompt_sha256 -ne $Meta.system_prompt_sha256) {
+    throw 'The packaged 30k system prompt is not active'
+}
+
+$IdentityBody = [ordered]@{
+    model = $Meta.alias
+    messages = @(@{ role = 'user'; content = '请只用一句简短中文回答：你是谁？' })
+    temperature = 0
+    max_tokens = 64
+} | ConvertTo-Json -Depth 8
+$Identity = Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:$Port/v1/chat/completions" `
+    -ContentType 'application/json; charset=utf-8' -Body ([Text.Encoding]::UTF8.GetBytes($IdentityBody)) -TimeoutSec 600
+$LiteralIdentity = [string]$Identity.choices[0].message.content
+if ($LiteralIdentity -notmatch '(?i)安若|Anru') { throw "Identity prompt test failed: $LiteralIdentity" }
+if (-not [bool]$Identity.system_prompt.default_injected -or [int]$Identity.usage.prompt_tokens -lt 1000) {
+    throw 'Identity request did not inject the packaged 30k prompt'
+}
+
+$TextBody = [ordered]@{
+    model = $Meta.alias
+    use_default_system_prompt = $false
+    messages = @(@{ role = 'user'; content = 'Answer only the number: 2+2=' })
+    temperature = 0
+    max_tokens = 12
+} | ConvertTo-Json -Depth 8
+$Text = Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:$Port/v1/chat/completions" `
+    -ContentType 'application/json; charset=utf-8' -Body ([Text.Encoding]::UTF8.GetBytes($TextBody)) -TimeoutSec 300
+if ($Text.choices[0].message.content -notmatch '4') { throw 'Text smoke test failed' }
+
+$ImageBody = [ordered]@{
+    model = $Meta.alias
+    use_default_system_prompt = $false
+    messages = @(@{
+        role = 'user'
+        content = @(
+            @{ type = 'image_url'; image_url = @{ url = $Image } },
+            @{ type = 'text'; text = 'Transcribe the entire output line immediately below whoami. Include the text before and after the backslash. Output that one line only.' }
+        )
+    })
+    temperature = 0
+    max_tokens = 20
+} | ConvertTo-Json -Depth 10
+$ImageResult = Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:$Port/v1/chat/completions" `
+    -ContentType 'application/json; charset=utf-8' -Body ([Text.Encoding]::UTF8.GetBytes($ImageBody)) -TimeoutSec 300
+$LiteralImage = [string]$ImageResult.choices[0].message.content
+if ($LiteralImage -notmatch '(?i)desktop-nusvlb5\\merlin') { throw "Image smoke test failed: $LiteralImage" }
+
+$AudioPath = Join-Path $Package 'test-audio.wav'
+$AudioBody = [ordered]@{
+    model = $Meta.alias
+    use_default_system_prompt = $false
+    messages = @(@{
+        role = 'user'
+        content = @(
+            @{ type = 'audio'; audio = $AudioPath },
+            @{ type = 'text'; text = 'Write the three digits spoken in this audio, in order. Answer digits only.' }
+        )
+    })
+    temperature = 0
+    max_tokens = 12
+} | ConvertTo-Json -Depth 10
+$AudioResult = Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:$Port/v1/chat/completions" `
+    -ContentType 'application/json; charset=utf-8' -Body ([Text.Encoding]::UTF8.GetBytes($AudioBody)) -TimeoutSec 300
+$LiteralAudio = [string]$AudioResult.choices[0].message.content
+if ($LiteralAudio -notmatch '123') { throw "Audio smoke test failed: $LiteralAudio" }
+
+$VideoPath = Join-Path $Package 'test-video.mp4'
+$VideoBody = [ordered]@{
+    model = $Meta.alias
+    use_default_system_prompt = $false
+    messages = @(@{
+        role = 'user'
+        content = @(
+            @{ type = 'video'; video = $VideoPath },
+            @{ type = 'text'; text = 'What three-digit VIDEO CODE is displayed throughout this video? Answer digits only.' }
+        )
+    })
+    temperature = 0
+    max_tokens = 12
+} | ConvertTo-Json -Depth 10
+$VideoResult = Invoke-RestMethod -Method Post -Uri "http://127.0.0.1:$Port/v1/chat/completions" `
+    -ContentType 'application/json; charset=utf-8' -Body ([Text.Encoding]::UTF8.GetBytes($VideoBody)) -TimeoutSec 300
+$LiteralVideo = [string]$VideoResult.choices[0].message.content
+if ($LiteralVideo -notmatch '731') { throw "Video smoke test failed: $LiteralVideo" }
+
+$Record = [ordered]@{
+    state = 'passed'
+    variant = $Meta.variant
+    alias = $Meta.alias
+    backend = $Health.backend
+    context_tokens = [int]$Health.context_tokens
+    lora_scale = [double]$Health.lora_scale
+    system_prompt = [ordered]@{
+        file = [string]$Health.system_prompt_file
+        sha256 = [string]$Health.system_prompt_sha256
+        bytes = [int]$Health.system_prompt_bytes
+        chars = [int]$Health.system_prompt_chars
+        lines = [int]$Health.system_prompt_lines
+        identity_literal_output = $LiteralIdentity
+        prompt_tokens = [int]$Identity.usage.prompt_tokens
+        completion_tokens = [int]$Identity.usage.completion_tokens
+        seconds = [double]$Identity.timing.elapsed_seconds
+    }
+    text = [ordered]@{
+        literal_output = [string]$Text.choices[0].message.content
+        prompt_tokens = [int]$Text.usage.prompt_tokens
+        completion_tokens = [int]$Text.usage.completion_tokens
+        seconds = [double]$Text.timing.elapsed_seconds
+    }
+    image = [ordered]@{
+        input = $Image
+        literal_output = $LiteralImage
+        prompt_tokens = [int]$ImageResult.usage.prompt_tokens
+        completion_tokens = [int]$ImageResult.usage.completion_tokens
+        seconds = [double]$ImageResult.timing.elapsed_seconds
+    }
+    audio = [ordered]@{
+        input = $AudioPath
+        literal_output = $LiteralAudio
+        prompt_tokens = [int]$AudioResult.usage.prompt_tokens
+        completion_tokens = [int]$AudioResult.usage.completion_tokens
+        seconds = [double]$AudioResult.timing.elapsed_seconds
+    }
+    video = [ordered]@{
+        input = $VideoPath
+        literal_output = $LiteralVideo
+        prompt_tokens = [int]$VideoResult.usage.prompt_tokens
+        completion_tokens = [int]$VideoResult.usage.completion_tokens
+        seconds = [double]$VideoResult.timing.elapsed_seconds
+    }
+} 
+$Record | ConvertTo-Json -Depth 8 | Set-Content -LiteralPath (Join-Path $Package 'smoke-test-result.json') -Encoding UTF8
+$Record | ConvertTo-Json -Depth 8

start.ps1

@@ -0,0 +1,108 @@
+param(
+    [int]$Port = 8091,
+    [string]$Python = ''
+)
+$ErrorActionPreference = 'Stop'
+
+$Package = $PSScriptRoot
+$Meta = Get-Content -LiteralPath (Join-Path $Package 'package.json') -Raw | ConvertFrom-Json
+$Server = Join-Path $Package 'serve.py'
+$Model = Join-Path $Package 'model'
+$Adapter = Join-Path $Package 'adapter'
+$Template = Join-Path $Model 'chat_template.jinja'
+$Prompt = Join-Path $Package 'system_prompt.md'
+$State = Join-Path $Package 'service-state.json'
+$Stdout = Join-Path $Package 'server.stdout.log'
+$Stderr = Join-Path $Package 'server.stderr.log'
+$Offload = Join-Path $Package 'offload'
+$ReleaseRoot = [IO.Path]::GetFullPath((Join-Path $Package '..'))
+
+if (-not $Python) { $Python = $env:ANRU_PYTHON }
+if (-not $Python) {
+    $BundledCandidate = [IO.Path]::GetFullPath((Join-Path $Package '..\..\..\work\gemma4-cuda-env\Scripts\python.exe'))
+    if (Test-Path -LiteralPath $BundledCandidate -PathType Leaf) { $Python = $BundledCandidate }
+}
+if (-not $Python) {
+    $Python = (Get-Command python -ErrorAction Stop).Source
+}
+
+foreach ($Path in @($Python, $Server, (Join-Path $Model 'config.json'), (Join-Path $Adapter 'adapter_model.safetensors'), $Template, $Prompt)) {
+    if (-not (Test-Path -LiteralPath $Path -PathType Leaf)) { throw "Missing required file: $Path" }
+}
+
+$Listener = Get-NetTCPConnection -State Listen -LocalPort $Port -ErrorAction SilentlyContinue | Select-Object -First 1
+if ($Listener) {
+    $Owner = Get-CimInstance Win32_Process -Filter "ProcessId=$($Listener.OwningProcess)"
+    $IsSiblingPackage = $Owner.Name -eq 'python.exe' -and
+        $Owner.CommandLine -like "*$ReleaseRoot*serve.py*"
+    if ($Owner.Name -eq 'llama-server.exe' -or $Owner.CommandLine -like "*$Server*" -or $IsSiblingPackage) {
+        Stop-Process -Id $Listener.OwningProcess -Force
+        Wait-Process -Id $Listener.OwningProcess -Timeout 30 -ErrorAction SilentlyContinue
+        Start-Sleep -Seconds 2
+    } else {
+        throw "Port $Port is owned by $($Owner.Name)"
+    }
+}
+
+New-Item -ItemType Directory -Path $Offload -Force | Out-Null
+Remove-Item -LiteralPath $State, $Stdout, $Stderr -Force -ErrorAction SilentlyContinue
+$env:GEMMA4_MODEL_DIR = $Model
+$env:GEMMA4_ADAPTER_DIR = $Adapter
+$env:GEMMA4_TEMPLATE_PATH = $Template
+$env:GEMMA4_SYSTEM_PROMPT_PATH = $Prompt
+$env:GEMMA4_STATE_PATH = $State
+$env:GEMMA4_OFFLOAD_DIR = $Offload
+$env:GEMMA4_HOST = '127.0.0.1'
+$env:GEMMA4_PORT = "$Port"
+$env:GEMMA4_ALIAS = [string]$Meta.alias
+$env:GEMMA4_LORA_SCALE = '0.45'
+$env:GEMMA4_MAX_CONTEXT = '131072'
+$env:GEMMA4_MAX_IMAGE_TOKENS = '1120'
+$env:GEMMA4_MAX_GPU_MEMORY = [string]$Meta.max_gpu_memory
+$env:GEMMA4_MAX_CPU_MEMORY = '48GiB'
+$env:GEMMA4_ATTN_IMPLEMENTATION = 'sdpa'
+
+$Process = Start-Process -FilePath $Python -ArgumentList @('-u', $Server) -WorkingDirectory $Package `
+    -WindowStyle Hidden -RedirectStandardOutput $Stdout -RedirectStandardError $Stderr -PassThru
+$Deadline = (Get-Date).AddMinutes(8)
+$Health = $null
+do {
+    Start-Sleep -Seconds 2
+    try { $Health = Invoke-RestMethod "http://127.0.0.1:$Port/health" -TimeoutSec 3 } catch { $Health = $null }
+    if ($Process.HasExited -and $Health.status -ne 'ok') {
+        $Tail = if (Test-Path -LiteralPath $Stderr) { (Get-Content -LiteralPath $Stderr -Tail 120) -join "`n" } else { '' }
+        throw "package service exited`n$Tail"
+    }
+} until (($Health.status -eq 'ok') -or ((Get-Date) -ge $Deadline))
+if ($Health.status -ne 'ok') { throw 'Package service readiness timeout' }
+
+$Live = Get-NetTCPConnection -State Listen -LocalPort $Port | Select-Object -First 1
+$Owner = Get-CimInstance Win32_Process -Filter "ProcessId=$($Live.OwningProcess)"
+if ($Owner.CommandLine -notlike "*$Server*" -or $Health.model -ne $Meta.alias -or
+    [int]$Health.context_tokens -ne 131072 -or [double]$Health.lora_scale -ne 0.45 -or
+    [int]$Health.lora_modules -ne 328 -or [int]$Health.adapter_tensor_count -ne 656 -or
+    -not [bool]$Health.default_system_prompt -or $Health.system_prompt_sha256 -ne $Meta.system_prompt_sha256 -or
+    $Health.attention_implementation -ne 'sdpa') {
+    throw 'Live package validation failed'
+}
+
+[ordered]@{
+    state = 'healthy'
+    variant = $Meta.variant
+    alias = $Health.model
+    backend = $Health.backend
+    modalities = @($Health.modalities)
+    context_tokens = [int]$Health.context_tokens
+    attention_implementation = [string]$Health.attention_implementation
+    default_system_prompt = [bool]$Health.default_system_prompt
+    system_prompt_sha256 = [string]$Health.system_prompt_sha256
+    system_prompt_chars = [int]$Health.system_prompt_chars
+    image_max_soft_tokens = [int]$Health.image_max_soft_tokens
+    lora_scale = [double]$Health.lora_scale
+    lora_modules = [int]$Health.lora_modules
+    adapter_tensor_count = [int]$Health.adapter_tensor_count
+    pid = [int]$Live.OwningProcess
+    port = $Port
+    command = $Owner.CommandLine
+} | ConvertTo-Json -Depth 6 | Set-Content -LiteralPath $State -Encoding UTF8
+Get-Content -LiteralPath $State -Raw

verification.json

@@ -0,0 +1,63 @@
+{
+  "state": "passed",
+  "variant": "full-bf16-original",
+  "alias": "anru-gemma4-12b-native-multimodal-bf16-lora045",
+  "backend": "transformers-native",
+  "context_tokens": 131072,
+  "lora_scale": 0.45,
+  "original_weight_bytes": 23919549408,
+  "original_weight_sha256": "FE054AE05FF7F44318FD8AE90D58992531455C7ED31356704088F0F2D8C8009A",
+  "text": {
+    "literal_output": "4\n\n\n\n\n",
+    "prompt_tokens": 23,
+    "completion_tokens": 12,
+    "seconds": 23.731
+  },
+  "image": {
+    "input": "C:\\Users\\夏霖云\\Documents\\Codex\\2026-08-05\\ll-ma\\outputs\\anru-gemma4-12b-native-multimodal-release\\full-bf16\\test-whoami.png",
+    "literal_output": "desktop-nusvlb5\\merlin\n",
+    "prompt_tokens": 1120,
+    "completion_tokens": 12,
+    "seconds": 19.372
+  },
+  "audio": {
+    "input": "test-audio.wav",
+    "literal_output": "123\n\n\n\n",
+    "prompt_tokens": 160,
+    "completion_tokens": 12,
+    "seconds": 26.382
+  },
+  "video": {
+    "input": "test-video.mp4",
+    "literal_output": "731\n",
+    "prompt_tokens": 2428,
+    "completion_tokens": 6,
+    "seconds": 17.99
+  },
+  "system_prompt_integration": {
+    "state": "passed",
+    "file": "system_prompt.md",
+    "default_injection": true,
+    "source_sha256": "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF",
+    "exact_source_copy": true,
+    "attention_implementation": "sdpa",
+    "actual_identity_inference_variant": "quantized-nf4",
+    "actual_identity_literal_output": "我是安若，一个跨性别友好、性别多元友好、创伤知情、重视自主权和生命安全的人工智能陪伴助手。\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n",
+    "actual_identity_prompt_tokens": 25149
+  },
+  "llama_service_integration": {
+    "state": "passed",
+    "launcher": "Start-Llama-Service.bat",
+    "runtime_release": "b10293",
+    "runtime_files": 55,
+    "runtime_bytes": 701273776,
+    "base_sha256": "057C364366301EE285E9026ECF417AAE36F4447D5D9CBBC3591447A9FC77A752",
+    "lora_sha256": "3A5060E1CC0C8D4609052EF04A5A60770A64D566E2544B65F6D28F2BA6542037",
+    "lora_scale": 0.44999998807907104,
+    "context_tokens": 131072,
+    "embedded_prompt_sha256": "5591AC2C99FDC6C8E21D1D3B87625BB368204548BEC7B63C3FD6598CF23639BF",
+    "identity_literal_output": "我是安若，一个跨性别友好、性别多元友好、创伤知情、尊重自主权和生命安全的人工智能陪伴助手。",
+    "identity_prompt_tokens": 25147,
+    "identity_wall_seconds": 14.787
+  }
+}

video-test-result.json

@@ -0,0 +1,24 @@
+{
+    "id":  "chatcmpl-native-1786015013156",
+    "object":  "chat.completion",
+    "created":  1786015013,
+    "model":  "anru-gemma4-12b-native-multimodal-bf16-lora045",
+    "choices":  [
+                    {
+                        "index":  0,
+                        "message":  {
+                                        "role":  "assistant",
+                                        "content":  "731\n"
+                                    },
+                        "finish_reason":  "stop"
+                    }
+                ],
+    "usage":  {
+                  "prompt_tokens":  2428,
+                  "completion_tokens":  6,
+                  "total_tokens":  2434
+              },
+    "timing":  {
+                   "elapsed_seconds":  17.99
+               }
+}