Ely-Eye — local multimodal context operating system on Qwen3.5-9B

Real-models implementation: multimodal ingestion, Context Cartridge memory,
sparse + Qwen embedding retrieval, Cache Fabric, Qwen runtime adapters, verifier
contracts, Visual HashHop proofs, BF16 LoRA fine-tuning, and a local dashboard.
Verified by a 15-check PRD proof suite on the live model.

scripts/download-models.ps1

@@ -0,0 +1,7 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+
+.\.venv\Scripts\python.exe .\scripts\download_models.py --model Qwen/Qwen3.5-9B
+.\.venv\Scripts\python.exe .\scripts\download_models.py --model Qwen/Qwen3.5-0.8B
+.\.venv\Scripts\python.exe .\scripts\download_models.py --model Qwen/Qwen3-VL-Embedding-8B

scripts/download_models.py

@@ -0,0 +1,17 @@
+from __future__ import annotations
+
+import argparse
+
+from huggingface_hub import snapshot_download
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", required=True)
+    args = parser.parse_args()
+    path = snapshot_download(repo_id=args.model, resume_download=True)
+    print(path)
+
+
+if __name__ == "__main__":
+    main()

scripts/install-cuda-torch.ps1

@@ -0,0 +1,5 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+.\.venv\Scripts\python.exe -m pip install --upgrade --no-cache-dir --index-url https://download.pytorch.org/whl/cu130 torch==2.12.0+cu130 torchvision==0.27.0+cu130
+.\.venv\Scripts\python.exe -c "import torch; print(torch.__version__); print(torch.cuda.is_available()); print(torch.version.cuda)"

scripts/install.ps1

@@ -0,0 +1,21 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+
+if (-not (Test-Path ".venv")) {
+  python -m venv .venv
+}
+
+.\.venv\Scripts\python.exe -m pip install --upgrade pip setuptools wheel
+.\.venv\Scripts\python.exe -m pip install -e .\backend
+.\.venv\Scripts\python.exe -m pip install --upgrade --no-cache-dir unsloth unsloth_zoo
+.\.venv\Scripts\python.exe -m pip install --upgrade --no-cache-dir --index-url https://download.pytorch.org/whl/cu130 torch==2.12.0+cu130 torchvision==0.27.0+cu130
+.\.venv\Scripts\python.exe -m pip install --upgrade --no-cache-dir transformers==5.9.0 datasets==4.8.5 dill==0.4.1 multiprocess==0.70.19 fsspec==2026.2.0 safetensors==0.8.0rc0
+
+Push-Location .\frontend
+npm install
+Pop-Location
+
+Copy-Item .env.example .env -ErrorAction SilentlyContinue
+
+.\.venv\Scripts\ely-eye.exe status

scripts/launch-ktransformers-extreme.ps1

@@ -0,0 +1,23 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+
+if (-not (Get-Command wsl.exe -ErrorAction SilentlyContinue)) {
+  throw "WSL is required for the KTransformers Linux runtime."
+}
+
+$wslRoot = (wsl.exe -d Ubuntu wslpath -a (Get-Location).Path).Trim()
+$command = @"
+set -euo pipefail
+cd "$wslRoot"
+source .venv-linux/bin/activate
+export SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1
+python -m sglang.launch_server \
+  --model-path Qwen/Qwen3.5-9B \
+  --port 8002 \
+  --context-length 1010000 \
+  --kt-cpuinfer 16 \
+  --cpu-offload-gb 48
+"@
+
+wsl.exe -d Ubuntu -- bash -lc $command

scripts/launch-llamacpp-offline.ps1

@@ -0,0 +1,35 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+
+if (-not $env:ELY_EYE_GGUF_MODEL) {
+  throw "Set ELY_EYE_GGUF_MODEL to the absolute GGUF model path."
+}
+
+$server = $env:LLAMA_CPP_SERVER
+if (-not $server) {
+  $server = "llama-server"
+}
+
+if (-not (Get-Command $server -ErrorAction SilentlyContinue)) {
+  $wingetPackageRoot = Join-Path $env:LOCALAPPDATA "Microsoft\WinGet\Packages"
+  $installed = Get-ChildItem $wingetPackageRoot -Recurse -Filter llama-server.exe -ErrorAction SilentlyContinue |
+    Where-Object { $_.FullName -like "*ggml.llamacpp*" } |
+    Select-Object -First 1
+  if ($installed) {
+    $server = $installed.FullName
+  }
+}
+
+if (-not (Get-Command $server -ErrorAction SilentlyContinue)) {
+  throw "llama-server executable is missing. Install ggml.llamacpp or set LLAMA_CPP_SERVER."
+}
+
+& $server `
+  --model $env:ELY_EYE_GGUF_MODEL `
+  --host 127.0.0.1 `
+  --port 8003 `
+  --ctx-size 262144 `
+  --flash-attn `
+  --cache-type-k q4_0 `
+  --cache-type-v q4_0

scripts/launch-sglang-extreme.ps1

@@ -0,0 +1,25 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+
+if (-not (Get-Command wsl.exe -ErrorAction SilentlyContinue)) {
+  throw "WSL is required for the SGLang Linux runtime."
+}
+
+$override = '{"text_config":{"rope_parameters":{"mrope_interleaved":true,"mrope_section":[11,11,10],"rope_type":"yarn","rope_theta":10000000,"partial_rotary_factor":0.25,"factor":4.0,"original_max_position_embeddings":262144}}}'
+$wslRoot = (wsl.exe -d Ubuntu wslpath -a (Get-Location).Path).Trim()
+$command = @"
+set -euo pipefail
+cd "$wslRoot"
+source .venv-linux/bin/activate
+export SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1
+python -m sglang.launch_server \
+  --model-path Qwen/Qwen3.5-9B \
+  --port 8000 \
+  --tp-size 1 \
+  --mem-fraction-static 0.88 \
+  --context-length 1010000 \
+  --json-model-override-args '$override'
+"@
+
+wsl.exe -d Ubuntu -- bash -lc $command

scripts/launch-sglang-live.ps1

@@ -0,0 +1,26 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+
+if (-not (Get-Command wsl.exe -ErrorAction SilentlyContinue)) {
+  throw "WSL is required for the SGLang Linux runtime."
+}
+
+$wslRoot = (wsl.exe -d Ubuntu wslpath -a (Get-Location).Path).Trim()
+$command = @"
+set -euo pipefail
+cd "$wslRoot"
+source .venv-linux/bin/activate
+python -m sglang.launch_server \
+  --model-path Qwen/Qwen3.5-9B \
+  --port 8000 \
+  --tp-size 1 \
+  --mem-fraction-static 0.82 \
+  --context-length 262144 \
+  --reasoning-parser qwen3 \
+  --speculative-algo NEXTN \
+  --speculative-num-steps 3 \
+  --speculative-num-draft-tokens 4
+"@
+
+wsl.exe -d Ubuntu -- bash -lc $command

scripts/launch-vllm-extreme.ps1

@@ -0,0 +1,26 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+
+if (-not (Get-Command wsl.exe -ErrorAction SilentlyContinue)) {
+  throw "WSL is required for the vLLM Linux runtime."
+}
+
+$wslRoot = (wsl.exe -d Ubuntu wslpath -a (Get-Location).Path).Trim()
+$ropeScaling = '{"rope_type":"yarn","factor":4.0,"original_max_position_embeddings":262144}'
+$command = @"
+set -euo pipefail
+cd "$wslRoot"
+source .venv-linux-vllm/bin/activate
+python -m vllm.entrypoints.openai.api_server \
+  --model Qwen/Qwen3.5-9B \
+  --host 127.0.0.1 \
+  --port 8001 \
+  --max-model-len 1010000 \
+  --enable-prefix-caching \
+  --gpu-memory-utilization 0.88 \
+  --trust-remote-code \
+  --rope-scaling '$ropeScaling'
+"@
+
+wsl.exe -d Ubuntu -- bash -lc $command

scripts/launch-vllm-live.ps1

@@ -0,0 +1,24 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+
+if (-not (Get-Command wsl.exe -ErrorAction SilentlyContinue)) {
+  throw "WSL is required for the vLLM Linux runtime."
+}
+
+$wslRoot = (wsl.exe -d Ubuntu wslpath -a (Get-Location).Path).Trim()
+$command = @"
+set -euo pipefail
+cd "$wslRoot"
+source .venv-linux-vllm/bin/activate
+python -m vllm.entrypoints.openai.api_server \
+  --model Qwen/Qwen3.5-9B \
+  --host 127.0.0.1 \
+  --port 8001 \
+  --max-model-len 262144 \
+  --enable-prefix-caching \
+  --gpu-memory-utilization 0.82 \
+  --trust-remote-code
+"@
+
+wsl.exe -d Ubuntu -- bash -lc $command

scripts/run-all-prd-finetunes.ps1

@@ -0,0 +1,100 @@
+param(
+  [string]$CartridgeId = "ely_eye_prd_7e5cd27b8284",
+  [int]$MaxLength = 8192
+)
+
+$ErrorActionPreference = "Stop"
+$root = Split-Path -Parent $PSScriptRoot
+Set-Location $root
+
+$proofImage = Join-Path $root ".ely_eye\proofs\ely-eye-dashboard.png"
+$coreDataset = Join-Path $root ".ely_eye\training\hme_prd_sft.jsonl"
+$visualDataset = Join-Path $root ".ely_eye\training\hme_visual_sft.jsonl"
+$routerDataset = Join-Path $root ".ely_eye\training\hme_router_sft.jsonl"
+$retrievalDataset = Join-Path $root ".ely_eye\training\hme_retrieval_pairs.jsonl"
+
+.\.venv\Scripts\ely-eye.exe build-prd-training-data `
+  --prd-path .\PRD.md `
+  --output-path $coreDataset `
+  --max-sections 48
+
+.\.venv\Scripts\ely-eye.exe build-visual-training-data `
+  --prd-path .\PRD.md `
+  --output-path $visualDataset `
+  --cartridge-id $CartridgeId `
+  --image-path $proofImage
+
+.\.venv\Scripts\ely-eye.exe build-router-training-data `
+  --prd-path .\PRD.md `
+  --output-path $routerDataset `
+  --cartridge-id $CartridgeId
+
+.\.venv\Scripts\ely-eye.exe build-retrieval-training-data `
+  --prd-path .\PRD.md `
+  --output-path $retrievalDataset `
+  --cartridge-id $CartridgeId `
+  --max-sections 32
+
+.\.venv\Scripts\ely-eye.exe train-lora `
+  $coreDataset `
+  .\.ely_eye\adapters\hme_core_lora_qwen35_9b `
+  --max-steps 37 `
+  --kind hme_core_lora `
+  --rank 16 `
+  --alpha 32 `
+  --max-length $MaxLength `
+  --gradient-accumulation-steps 4 `
+  --cartridge-id $CartridgeId
+
+.\.venv\Scripts\ely-eye.exe train-lora `
+  $visualDataset `
+  .\.ely_eye\adapters\hme_vision_lora_qwen35_9b `
+  --max-steps 20 `
+  --kind hme_vision_lora `
+  --rank 4 `
+  --alpha 8 `
+  --max-length $MaxLength `
+  --gradient-accumulation-steps 4 `
+  --cartridge-id $CartridgeId
+
+.\.venv\Scripts\ely-eye.exe train-lora `
+  $visualDataset `
+  .\.ely_eye\adapters\hme_ttt_vl_qwen35_9b `
+  --max-steps 20 `
+  --kind hme_ttt_vl `
+  --rank 4 `
+  --alpha 8 `
+  --max-length $MaxLength `
+  --gradient-accumulation-steps 4 `
+  --cartridge-id $CartridgeId
+
+.\.venv\Scripts\ely-eye.exe train-lora `
+  $visualDataset `
+  .\.ely_eye\adapters\hme_visual_mtp_qwen35_9b `
+  --max-steps 20 `
+  --kind hme_visual_mtp `
+  --rank 4 `
+  --alpha 8 `
+  --max-length $MaxLength `
+  --gradient-accumulation-steps 4 `
+  --cartridge-id $CartridgeId
+
+.\.venv\Scripts\ely-eye.exe train-lora `
+  $routerDataset `
+  .\.ely_eye\adapters\hme_router_qwen35_08b `
+  --max-steps 16 `
+  --kind hme_router `
+  --rank 8 `
+  --alpha 16 `
+  --max-length $MaxLength `
+  --gradient-accumulation-steps 4 `
+  --cartridge-id $CartridgeId
+
+.\.venv\Scripts\ely-eye.exe train-retrieval-lora `
+  $retrievalDataset `
+  .\.ely_eye\adapters\hme_retrieval_qwen3_vl_embedding_8b `
+  --max-steps 32 `
+  --rank 8 `
+  --alpha 16 `
+  --max-length $MaxLength `
+  --cartridge-id $CartridgeId

scripts/run-prd-finetune.ps1

@@ -0,0 +1,41 @@
+param(
+  [int]$MaxSteps = 37,
+  [int]$Rank = 16,
+  [int]$Alpha = 32,
+  [int]$MaxLength = 8192,
+  [int]$GradientAccumulationSteps = 4,
+  [string]$CartridgeId = ""
+)
+
+$ErrorActionPreference = "Stop"
+$root = Split-Path -Parent $PSScriptRoot
+Set-Location $root
+
+$dataset = Join-Path $root ".ely_eye\training\hme_prd_sft.jsonl"
+$adapter = Join-Path $root ".ely_eye\adapters\hme_core_lora_qwen35_9b"
+
+.\.venv\Scripts\ely-eye.exe build-prd-training-data `
+  --prd-path .\PRD.md `
+  --output-path $dataset `
+  --max-sections 48
+
+$args = @(
+  "train-lora",
+  $dataset,
+  $adapter,
+  "--max-steps", "$MaxSteps",
+  "--kind", "hme_core_lora",
+  "--rank", "$Rank",
+  "--alpha", "$Alpha",
+  "--max-length", "$MaxLength",
+  "--gradient-accumulation-steps", "$GradientAccumulationSteps"
+)
+
+if ($CartridgeId.Length -gt 0) {
+  $args += @("--cartridge-id", $CartridgeId)
+}
+
+.\.venv\Scripts\ely-eye.exe @args
+
+Write-Host "Adapter ready: $adapter"
+Write-Host "Use with: `$env:ELY_EYE_ADAPTER_DIR='$adapter'"

scripts/serve-backend.ps1

@@ -0,0 +1,4 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+.\.venv\Scripts\ely-eye.exe serve --host 127.0.0.1 --port 8765

scripts/serve-frontend.ps1

@@ -0,0 +1,6 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+Push-Location .\frontend
+npm run dev
+Pop-Location

scripts/setup-linux-runtimes.ps1

@@ -0,0 +1,23 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+
+if (-not (Get-Command wsl.exe -ErrorAction SilentlyContinue)) {
+  throw "WSL is required for SGLang, vLLM, and KTransformers runtime profiles."
+}
+
+wsl.exe -d Ubuntu -- bash -lc "command -v nvidia-smi >/dev/null && nvidia-smi -L >/dev/null"
+wsl.exe -d Ubuntu -- bash -lc "command -v curl >/dev/null || (apt-get update && apt-get install -y curl)"
+wsl.exe -d Ubuntu -- bash -lc "command -v /root/.local/bin/uv >/dev/null || curl -LsSf https://astral.sh/uv/install.sh | sh"
+$wslRoot = (wsl.exe -d Ubuntu wslpath -a (Get-Location).Path).Trim()
+wsl.exe -d Ubuntu -- bash -lc "cd '$wslRoot' && /root/.local/bin/uv venv .venv-linux --python 3.12"
+wsl.exe -d Ubuntu -- bash -lc "cd '$wslRoot' && /root/.local/bin/uv venv .venv-linux-vllm --python 3.12"
+wsl.exe -d Ubuntu -- bash -lc "cd '$wslRoot' && source .venv-linux/bin/activate && /root/.local/bin/uv pip install --prerelease=allow -r backend/requirements-sglang-linux.txt"
+wsl.exe -d Ubuntu -- bash -lc "cd '$wslRoot' && source .venv-linux-vllm/bin/activate && /root/.local/bin/uv pip install --prerelease=allow -r backend/requirements-vllm-linux.txt"
+wsl.exe -d Ubuntu -- bash -lc "cd '$wslRoot' && source .venv-linux/bin/activate && /root/.local/bin/uv pip install --prerelease=allow -r backend/requirements-ktransformers-linux.txt"
+
+if (-not (Get-Command llama-server -ErrorAction SilentlyContinue)) {
+  winget install --id ggml.llamacpp -e --accept-package-agreements --accept-source-agreements
+}
+
+.\.venv\Scripts\ely-eye.exe runtime-profiles

scripts/verify.ps1

@@ -0,0 +1,10 @@
+$ErrorActionPreference = "Stop"
+
+Set-Location (Split-Path -Parent $PSScriptRoot)
+
+.\.venv\Scripts\python.exe -m compileall .\backend\src
+.\.venv\Scripts\ely-eye.exe status
+
+Push-Location .\frontend
+npm run build
+Pop-Location