[Bug] Duplicate Megatron-LM installation in build_conda.sh

[kaysonyu]

Description

The build_conda.sh script contains duplicate Megatron-LM cloning and installation steps, which causes unnecessary redundancy and incorrect directory structure during the build process.

Current Behavior

The script clones and installs Megatron-LM twice:

1. First installation (lines 49-51):

   git clone https://github.com/NVIDIA/Megatron-LM.git --recursive && \
       cd Megatron-LM && git checkout ${MEGATRON_COMMIT} && \
       pip install -e .

• Location: $BASE_DIR/sglang/Megatron-LM (incorrect, since the script is still in the sglang directory from line 29)
• Uses undefined variable ${MEGATRON_COMMIT} (not set anywhere in the script)

2. Second installation (lines 58-61):
   cd $BASE_DIR
   git clone https://github.com/NVIDIA/Megatron-LM.git --recursive &&
   cd Megatron-LM/ && git checkout core_v0.14.0 &&
   pip install -e .
   - Location: $BASE_DIR/Megatron-LM (correct)
   - Explicitly uses core_v0.14.0 version

The patch application at line 84 uses $BASE_DIR/Megatron-LM, which relies on the second installation.

Expected Behavior

Megatron-LM should be cloned and installed only once at the correct location with the proper version specified.

Comparison with Dockerfile

The docker/Dockerfile handles this correctly:

• Defines MEGATRON_COMMIT as an ARG (line 7)
• Clones Megatron-LM only once (lines 52-54)
• Has proper error handling for patch application

I'm uncertain about this. If this counts as a bug, I will fix it and submit a pull request.