16GB mem unable to build candle anymore

HI, try build candle on my 16GB mem, it keep throw out error like this:

```
killed out of memory
```
cargo build can not succeed.

```
export NVCC_THREADS=1
1348* export MAX_JOBS=1
1349* NVCC_PREPEND_FLAGS="-t 1" cargo build --release --features "cuda flash-attn" --jobs 1
```

tried constraint threads still not work.

Why it was so strict to GPU mem when build candle.