launch_sglang_server.sh ignores CUDA_VISIBLE_DEVICES and always launches GPUs 0-7

[Oxygen56]

Problem

scripts/data/launch_sglang_server.sh currently hardcodes num_workers=8 and then launches workers by setting CUDA_VISIBLE_DEVICES=${gpu_id} for gpu_id=0..7.

This conflicts with the README guidance that users can reduce CUDA_VISIBLE_DEVICES for fewer GPUs. If a user only wants to run the data-generation servers on a subset such as GPUs 2,3, the launcher still starts eight workers and overrides each child process to use GPUs 0 through 7.

Reproduction

I reproduced this by running the launcher with a stub sglang command that records the environment passed to each worker:

CUDA_VISIBLE_DEVICES=2,3 bash scripts/data/launch_sglang_server.sh

Observed worker launches:

CUDA_VISIBLE_DEVICES=0 ... --port 30000
CUDA_VISIBLE_DEVICES=1 ... --port 30001
CUDA_VISIBLE_DEVICES=2 ... --port 30002
CUDA_VISIBLE_DEVICES=3 ... --port 30003
CUDA_VISIBLE_DEVICES=4 ... --port 30004
CUDA_VISIBLE_DEVICES=5 ... --port 30005
CUDA_VISIBLE_DEVICES=6 ... --port 30006
CUDA_VISIBLE_DEVICES=7 ... --port 30007

Expected behavior

When CUDA_VISIBLE_DEVICES=2,3 is set, the launcher should derive the worker GPU list from that value and start two workers, for example on ports 30000 and 30001.

The existing eight-GPU default should remain unchanged when CUDA_VISIBLE_DEVICES is not set.

Suggested fix

Parse CUDA_VISIBLE_DEVICES into a GPU list, derive the default num_workers from that list, keep the current 0,1,2,3,4,5,6,7 default when unset, and update the data-preparation docs so prepare_data.sh connects to the matching server ports.