Rollback standard-layout cleanup

Restore every file from backup-pre-hf-cleanup-20260813 (2ca02f2cb4b845f95bd52f8f28972bd945d7c032).

.gitattributes

@@ -1,2 +1,18 @@
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-tokenizer.json filter=lfs diff=lfs merge=lfs -text
+*.gguf filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth 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
+*.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
+
+model/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+adapter/tokenizer.json 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 +1,60 @@
+{
+  "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,
+  "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,
+  "payload_bytes_excluding_manifest": 33416349568,
+  "hub_gitattributes_normalized": true,
+  "hub_gitattributes_sha256": "0045B59DBD28A94B5A1F1D5AAEB1FB87DECFA02F2B3095ADFCAB1BCCA03C9570"
+}

README.md

@@ -1,62 +1,152 @@
----
-library_name: transformers
-pipeline_tag: any-to-any
-license: apache-2.0
-license_link: https://ai.google.dev/gemma/docs/gemma_4_license
-base_model: google/gemma-4-12B-it
-model_name: Anrn-12B-R1
-tags:
-  - gemma4
-  - multimodal
-  - bf16
-  - text-generation
-  - image-text-to-text
-  - audio-text-to-text
-  - video-text-to-text
----
-
-# Anrn 12B R1
-
-This repository is the clean Hugging Face/Transformers checkpoint layout for the **original BF16 model weights** and the native multimodal processor used by Anrn 12B R1.
-
-## Contents
-
-- `model.safetensors` - original BF16 checkpoint
-- `config.json` - model architecture
-- `generation_config.json` - generation defaults
-- `processor_config.json` - native image, audio, and video processor configuration
-- `tokenizer.json` and `tokenizer_config.json` - tokenizer required by the unified processor
-- `chat_template.jinja` - chat formatting template
-
-The repository intentionally excludes LoRA adapters, GGUF quantizations, llama.cpp binaries, launch scripts, logs, test media, and internal verification artifacts.
-
-## Load with Transformers
-
-Install a Transformers release with Gemma 4 Unified support and the native
-vision dependencies:
-
-```bash
-pip install "transformers>=5.10.0" torch torchvision accelerate safetensors
+---
+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
 ```
 
-```python
-from transformers import AutoModelForMultimodalLM, AutoProcessor
-
-model_id = "chirs345678/Anrn-12B-R1"
-processor = AutoProcessor.from_pretrained(model_id)
-model = AutoModelForMultimodalLM.from_pretrained(
-    model_id,
-    dtype="auto",
-    device_map="auto",
-)
-```
-
-The checkpoint accepts text, image, audio, and video input and generates text.
-
-## Checkpoint note
-
-This clean revision contains the byte-identical original BF16 base checkpoint only. The previously bundled project LoRA is deliberately not part of this repository.
-
-## License
-
-Apache 2.0. See the linked Gemma 4 license information and upstream terms.
+## 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 @@
+0045B59DBD28A94B5A1F1D5AAEB1FB87DECFA02F2B3095ADFCAB1BCCA03C9570  .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/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6fda7ad849c2208c7a8d549b0f9588b5b2bf789b0faf2e44427b8d73171c44b6
+size 131224632

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/anru-final-lora-Q8_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a5060e1cc0c8d4609052ef04a5a60770a64d566e2544b65f6d28f2ba6542037
+size 51682496

llama-model/anru-gemma4-12b-base-Q5_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:057c364366301ee285e9026ecf417aae36f4447d5d9cbbc3591447a9fc77a752
+size 8547249568

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"
+                   }
+               ]
+}

llama-runtime/cublas64_13.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1d500d0cd892f5b8c6b6cdbffd82d0c55d5f5427215668e7ceb55aeeccc1b63
+size 51870320

llama-runtime/cublasLt64_13.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b592cd016d7673e9cb97716a22b27c4010ee635377a3ba28f37070a9bdb76a68
+size 460301424

llama-runtime/cudart64_13.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b00ca6f53699120da815bf3e06e2e4285fae2f201235b883dcbb50eec51e2a2a
+size 551024

llama-runtime/ggml-base.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4625dbacbfcbea34ed402a0dcc9225ff3df947dead0dd859dd9f81a983dc365
+size 773120

llama-runtime/ggml-cpu-alderlake.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2f3a7da3e2d1985faec025928f2da5badc4629b478c0ca8083b4418c90da7e1
+size 1180672

llama-runtime/ggml-cpu-cannonlake.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4193f2927723669ee7b47f107090f42a675b4fd84fb77403c7ce412632be3ccc
+size 1395200

llama-runtime/ggml-cpu-cascadelake.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7933b3891929de369149eb03bb607f8d75f2d0e81071adcc5fcbab6c027a1de9
+size 1381376

llama-runtime/ggml-cpu-cooperlake.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96f572bffc3ddbe7d48cbaf7c7f6dfb9d75e65dc8f43e62c55810223fd1dae36
+size 1382400

llama-runtime/ggml-cpu-haswell.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:258f2ad5fb79f498876a8d3f44b366bff4b8a7ddcfb18a61c72270b8da8886c4
+size 1184768

llama-runtime/ggml-cpu-icelake.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b88ac4c73d93b204fdcf82b813d52e8333252b011c871858e73ae524f78089d7
+size 1388032

llama-runtime/ggml-cpu-ivybridge.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b9018ea35e928b61445c501521ba279174b32d29cb32f7b87f11fbde1ccd590
+size 1073664

llama-runtime/ggml-cpu-piledriver.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1b6bf9d333d4ecb9f085e2ac3df7c5baf65b8f1dbaaaa0b75aa90dd8c8b3202
+size 1077248

llama-runtime/ggml-cpu-sandybridge.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:487df33cf344fe95424ab049fb3892d7d198c42434aea6e1b5e28d721ffc15e7
+size 1053696

llama-runtime/ggml-cpu-sapphirerapids.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a786981fe76c507d7ae79fab512118a680394df06c0e8bb207015b3a9441da48
+size 1659904

llama-runtime/ggml-cpu-skylakex.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fdd262c2df029c8ba7d417dce3473b0afb795ac6d9d41c60af0fba7d0f32997
+size 1389056

llama-runtime/ggml-cpu-sse42.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3199b17d4e6349c6c19b897ffa473b4f62b77298ea237b93fef891b074b337e
+size 876544

llama-runtime/ggml-cpu-x64.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d214739c43071fc37f2703d0014cdf1dff27f98faf23a9707071ba9d0ea1a6e1
+size 869376

llama-runtime/ggml-cpu-zen4.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:206d0301bf5aa90cba87b763346a5a50d0ac756cb0c4de1271cf91833fc39e08
+size 1389056

llama-runtime/ggml-cuda.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1f0ed3dc2d80fa7b4c33a3156c99f79da782152a05af87f4d4687a459ab6c41
+size 141608448

llama-runtime/ggml-rpc-server.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab7930b188e3610264c540cfda9df08df3595502f8298821191083c3d2483a33
+size 101376

llama-runtime/ggml-rpc.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b29961a29a1b2f2071f4b71bc3936efb54c221a5ec5f03c64b54e8ccf7b0798
+size 137728

llama-runtime/libomp140.x86_64.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a20c1e5c115c29771a12324513eb109badac72180f79481527ad79d996ffb33
+size 661856

llama-runtime/llama-bench-impl.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3951adfac8b9a334ebab53054bcc1fcde26a48f2b0bad43b8eba49b197d7bc0a
+size 394752