Build a Multi-LoRA Adapter Memory and Batch Sizing Calculator — implementation

## Build
Profile VRAM usage for serving concurrent LoRA adapters on open-weight base models using vLLM. Synthesize maximum concurrency limits and memory overheads using hosted Claude to generate deployment configuration.

## User
ML Deployment Specialist

## Workflow
1. Parse evaluation logs measuring VRAM footprint scaling per active LoRA adapter on local hardware.
2. Calculate maximum adapter concurrency before triggering out-of-memory errors or significant KV-cache eviction.
3. Prompt hosted Claude API to synthesize a deployment density report justifying adapter count limits.
4. Export a deployment configuration specifying max_loras and max_lora_rank for the vLLM engine.

## Interface
Web dashboard showing projected VRAM usage per adapter and configuration export options.

## Data
JSON arrays containing simplified synthetic VRAM profile measurements for LoRA adapters across various ranks.
Starter examples below are synthetic. They are not downloaded records.

## Source
[vLLM open-model serving](https://raw.githubusercontent.com/vllm-project/vllm/main/README.md)
References vLLM multi-LoRA serving features and configuration flags.

## Tools
anthropic-claude, fastapi, pydantic.
Synthesizes multi-LoRA deployment strategies based on memory profiles.
Hosted Claude API for technical synthesis; benchmarks focus on open-weight models.

## Deliverables
Submit runnable source, pinned dependencies, setup commands, a working interface, example outputs, and a recorded end-to-end demo. Include tests and an action log.

## Acceptance
- Correctly computes VRAM usage per adapter rank and base model overhead.
- Identifies hard limits on concurrent requests when serving heterogeneous LoRAs.
- Generates ready-to-run vLLM multi-LoRA launch arguments.
- Confines profiling strictly to specified open-weight model architectures.

## Starter Examples
```json
[
  {
    "input": {
      "base_model": "open-qwen-7b",
      "base_vram_gb": 14.5,
      "adapter_rank": 64,
      "vram_per_adapter_gb": 0.45,
      "total_vram_gb": 24,
      "target_adapters": 10
    },
    "expected_output": {
      "total_projected_vram_gb": 19,
      "viable_deployment": true,
      "vllm_launch_flag": "--enable-lora --max-loras 10 --max-lora-rank 64",
      "claude_rationale": "Base model and 10 adapters fit well within 24GB VRAM, leaving 5GB for KV cache and context generation."
    }
  },
  {
    "input": {
      "base_model": "open-qwen-7b",
      "base_vram_gb": 14.5,
      "adapter_rank": 128,
      "vram_per_adapter_gb": 0.9,
      "total_vram_gb": 24,
      "target_adapters": 20
    },
    "expected_output": {
      "total_projected_vram_gb": 32.5,
      "viable_deployment": false,
      "vllm_launch_flag": null,
      "claude_rationale": "Projected VRAM (32.5GB) exceeds hardware limits. Reduce target adapters."
    }
  }
]
```

## Safety
Evaluation applies exclusively to open-weight models. No access or extrapolation to proprietary Claude models.
Use mock external writes. Require human review before any real deployment.