# SPEC — vram-detect-cross-platform

Parent: `runnable-test-suite` (see PARENT_SPEC.md).

## Scope

Make `scripts/vram_detect.py` work on macOS, Windows, and Linux without behavior
change on Cachyos/Linux. The existing version was developed on Cachyos and only
fully works on Linux. Pure detection logic — no CLI/JSON-shape changes.

## Files this sub-task touches (and ONLY this)

- `scripts/vram_detect.py` — surgical edits to detection functions only.

## MUST NOT touch

- `tests/test_vram_detect.py` (subtask-3 owns all vram_detect tests)
- Any other test file
- Any documentation, prompt, or install script (subtask-1 owns those)
- CLI args, JSON output shape, main() flow — only detection internals

## Functions to refactor (by current line in `scripts/vram_detect.py`)

### `run_command()` (vram_detect.py:54-70)
- On Windows, route PowerShell cmdlets via `powershell -NoProfile -NoLogo -Command "..."` wrapper.
- Keep `shutil.which()` gating so missing tools return None on all OSes.
- Do not break Linux path.

### `detect_gpu_vram()` (vram_detect.py:73-108)
Branch on `platform.system()`:
- **Linux**: keep `nvidia-smi` → `lspci -vnn` path exactly as-is. Regression guard.
- **Darwin** (macOS): add `system_profiler SPDisplaysDataType` parser.
  Parse `VRAM (Total):` line for Intel Macs, and for Apple Silicon unified memory,
  detect `Chipset Model: Apple M*` and treat total RAM as shared VRAM (call
  `sysctl -n hw.memsize` once and use that number, since Apple Silicon has no
  dedicated VRAM). Print clearly which kind was detected.
  Fallback (if `system_profiler` missing): `ioreg -c IOPlatformDevice -r -d 1`.
- **Windows**: add `wmic path win32_VideoController get AdapterRAM,Name /format:list`
  (deprecated but ubiquitous; works on Win10/11). Sum `AdapterRAM=` values across
  GPUs. PowerShell fallback:
  `powershell -NoProfile -Command "Get-CimInstance Win32_VideoController | Select-Object AdapterRAM"`

### `detect_ram()` (vram_detect.py:137-162)
Branch on `platform.system()`:
- **Linux**: keep `/proc/meminfo` path. Parse `MemTotal` + `MemAvailable`. Regression guard.
- **Darwin**: keep `sysctl -n hw.memsize` path (currently the fallback; promote to
  the macOS branch as primary). Returns (total_kb, total_kb) because macOS doesn't
  expose "available RAM" via sysctl directly — leave available == total. Print clear
  "available RAM detection not supported on macOS, reporting total" message once.
- **Windows**: add `wmic ComputerSystem get TotalPhysicalMemory /format:list`.
  Returns bytes — divide by 1024 for KB. PowerShell fallback:
  `powershell -NoProfile -Command "(Get-CimInstance Win32_ComputerSystem).TotalPhysicalMemory"`

### `MODEL_CONTEXT_WINDOWS` dict (vram_detect.py:24-47)
Additive only. Add these local-LLM entries with context sizes from public model cards:
- `llama-3.1-8b`: 128_000
- `llama-3.3-70b`: 128_000
- `qwen2.5-7b`: 128_000 (Qwen2.5 supports up to 128k per model card)
- `qwen2.5-72b`: 128_000
- `mistral-7b`: 32_000
- `mistral-large`: 128_000
- `deepseek-r1`: 64_000 (DeepSeek-R1)
- `deepseek-v3`: 64_000
- `glm-4`: 128_000
- `glm-4.5`: 128_000
- `gemma-2`: 8_000
- `gemma-2-27b`: 8_000
- `phi-3`: 128_000
- `phi-4`: 16_000

Add a docstring comment above each: `# Source: <model card URL or repo>` — never
fabricate. If a number is uncertain, use the smaller conservative value and
leave a comment noting the uncertainty.

### `detect_model_context()` (vram_detect.py:175-228)
Add an `ollama list` probe when no config file names a model AND the system has
`ollama` on PATH. Steps:
1. `ollama list` → parse first non-header row's NAME column (strip `:latest` tag).
2. Look up the cleaned name in `MODEL_CONTEXT_WINDOWS` via existing `_lookup_model_context`.
3. If matched, return that context. If not matched, fall through to fail-open `0`.

Do NOT modify any other code path in `detect_model_context`.

## MUST NOT regress

- The existing Linux output of `python3 vram_detect.py` must produce byte-identical
  stdout (after the equivalent hardware probe) on the original Cachyos box. Subtask-3
  will write a Linux-fixture test to lock this in.
- Do not remove or alter any existing OpenAI/Anthropic entry in `MODEL_CONTEXT_WINDOWS`.

## Acceptance criteria

1. `python3 scripts/vram_detect.py` on Darwin prints `gpu_vram_gb > 0` (currently prints 0).
2. `python3 scripts/vram_detect.py` JSON contains the same keys, same order, same types.
3. `python3 -m py_compile scripts/vram_detect.py` exits 0.
4. On a Linux fixture (simulated by subtask-3 tests with patched `platform.system`),
   stdout matches the pre-refactor output line-by-line for GPU/RAM sections.
5. `python3 scripts/vram_detect.py` does not crash on Windows stub (subtask-3 sets
   `monkeypatch.setattr(platform, "system", lambda: "Windows")` and mocks subprocess).
6. No new third-party imports.

## Anti-spin rails

- Unknown model name → fail open with `0`. NEVER guess a context window.
- If `platform.system()` returns an unexpected string (e.g. "AIX"), fall through to
  Linux path or print "Unsupported OS: X" and return 0s. Do not crash.
- If `system_profiler` output format on the local M-series Mac is different from what
  you parsed, STOP and report. Don't patch a half-working parser.

## Hardware context (this box)

- Darwin arm64, Python 3.9.6 (stock CommandLineTools).
- `system_profiler SPDisplaysDataType` is the canonical probe.
- You can iterate locally by running `python3 scripts/vram_detect.py` after each edit.

## Recommended approach

1. Add the local-LLM entries to `MODEL_CONTEXT_WINDOWS` first (mechanical).
2. Refactor `detect_ram()` with a `platform.system()` dispatch — easiest, lowest risk.
3. Refactor `detect_gpu_vram()` — hardest, leave for after RAM is green.
4. Add the `ollama list` probe.
5. Wrapping `run_command()` for Windows PowerShell — defer until last.
6. After each function, run `python3 scripts/vram_detect.py` and confirm no crash + correct output.
7. Do NOT touch `tests/test_vram_detect.py`. Subtask-3 will write tests against your function signatures.