Muse Glimmer launcher for DGX Spark

# Muse Glimmer 30B on DGX Spark

This launcher follows the verified vLLM DGX Spark recipe. It serves the
ModelOpt NVFP4 W4A4 checkpoint with Muse Glimmer's dedicated reasoning and tool
parsers, plus its DFlash speculative decoder.

## Quick start

Use the main launcher for start, stop, status, and logs:

```bash
cd ~/muse_glimmer

./muse_glimmer.sh start
./muse_glimmer.sh status
./muse_glimmer.sh logs
./muse_glimmer.sh stop
```

The existing `start.sh` and `stop.sh` wrappers remain available for backward
compatibility.

A normal start reuses the tested local runtime image. This prevents a mutable
upstream image tag from breaking an otherwise working restart. To deliberately
pull the current base image and rebuild the compatibility layer:

```bash
REBUILD_IMAGE=1 ./muse_glimmer.sh start
```

`stop` retains both downloaded checkpoints and the compiler caches. A cold
start can take several minutes because model loading, compilation, FP4 tuning,
and CUDA graph capture must finish before the API becomes ready.

Check readiness with:

```bash
./muse_glimmer.sh status
```

The server is ready only when the command reports:

```text
API status: ready
```

## API endpoint

The OpenAI-compatible API is served on port `8000`.

For requests from the DGX Spark itself:

```text
http://localhost:8000/v1
```

For requests from another computer on the current LAN:

```text
http://10.0.4.26:8000/v1
```

The served model name is:

```text
muse-glimmer-30b
```

The LAN address may change if the DGX Spark receives a different address from
DHCP. Update the examples below when necessary.

## API usage with `curl`

### 1. Set the API base URL

For local requests on the DGX Spark:

```bash
export MUSE_API_BASE="http://localhost:8000/v1"
```

For requests from another machine on the same LAN:

```bash
export MUSE_API_BASE="http://10.0.4.26:8000/v1"
```

Using an environment variable makes the remaining commands work with either
endpoint.

### 2. List served models

```bash
curl -sS "$MUSE_API_BASE/models"
```

Pretty-print the response with `jq`:

```bash
curl -sS "$MUSE_API_BASE/models" | jq
```

### 3. Check API readiness from a script

```bash
if curl -fsS "$MUSE_API_BASE/models" >/dev/null; then
  echo "Muse Glimmer API is ready"
else
  echo "Muse Glimmer API is not ready"
fi
```

### 4. Run a basic chat completion

Published request sampling is `temperature=1.0`, `top_p=0.95`, a