[dockerfile]: swich to vllm==0.21.0

[princepride]

Purpose

Removed all SGLang-related Docker dependencies and replaced with vllm==0.21.0.

@gcanlin @aoshen02 I don't know how to test it because right now vime is still heavily dependent on SGLang.

[gemini-code-assist]

Code Review

This pull request migrates the Docker build process from an SGLang-based image to a vLLM-based image. Key changes include updating the base image, implementing a script to filter out SGLang dependencies from requirements.txt, and adding validation checks to ensure a clean vLLM environment. Review feedback suggests streamlining the VLLM_VERSION argument to avoid redundancy and removing an unreachable check for git conflict markers, as the shell is already configured to exit on error.

[gemini-code-assist]

The VLLM_VERSION is defined twice: once in the default value of BASE_IMAGE and once as a standalone ARG. This redundancy can lead to version mismatches if the base image is overridden but the version argument is not. Consider using VLLM_VERSION to construct the default BASE_IMAGE to ensure consistency.

ARG VLLM_VERSION=0.21.0
ARG BASE_IMAGE=vllm/vllm-openai:v${VLLM_VERSION}
FROM ${BASE_IMAGE}

# ======================================== Arguments =============================================

ARG VLLM_VERSION

[gemini-code-assist]

This check for conflict markers is redundant and unreachable. The SHELL instruction on line 12 enables pipefail and errexit (-e), so the build will fail immediately if git apply returns a non-zero exit code. Additionally, git apply --3way does not leave conflict markers in files upon failure; it simply fails the operation.

    git apply megatron.patch --3way && \
    rm megatron.patch && \

[gcanlin]

Purpose

Removed all SGLang-related Docker dependencies and replaced with vllm==0.21.0.

@gcanlin @aoshen02 I don't know how to test it because right now vime is still heavily dependent on SGLang.

I guess we need to wait #7 if we want to remove the dependencies from SGLang. But we can add vllm installation first. And after #7 done, we can remove SGLang totally in Dockerfile.

[aoshen02]

Is there any cuda/pytorch/transformers version conflict between original configuration and what vLLM 0.21.0 needs?

[aoshen02]

To let training not break, I think should still install sglang with --no-deps for now.

[aoshen02]

We need TE, it's a dependency of mcore

[princepride]

Is there any cuda/pytorch/transformers version conflict between original configuration and what vLLM 0.21.0 needs?

So we have better fallback to 0.20.2 ?

[aoshen02]

#12 I

Is there any cuda/pytorch/transformers version conflict between original configuration and what vLLM 0.21.0 needs?

So we have better fallback to 0.20.2 ?

Not sure.

[aoshen02]

#12 I have a follow up PR that fix some problems, please check.