docker(xpu): drop redundant setvars.sh from torch_memory_saver RUN

[arathi-hlab]

Summary

Follow-up to #38617. That PR turned a silent exit 3 into a visible one, which surfaced the real bug: setvars.sh should not have been in this RUN at all. Removing it fixes the nightly and gets rid of the fragility for good.

Failing nightly run (post-#38617): https://github.com/sgl-project/sglang/actions/runs/34330707892

Root cause

The base image intel/deep-learning-essentials:2026.0.0-devel-ubuntu24.04 already runs setvars.sh in its own build layers and freezes the resulting env into image ENV. Every subsequent RUN starts with the SYCL toolchain fully wired up:

$ docker run --rm intel/deep-learning-essentials:2026.0.0-devel-ubuntu24.04 \
    env | grep -E '^(PATH|LIBRARY_PATH|CMPLR_ROOT|SETVARS_COMPLETED)'
PATH=/opt/intel/oneapi/mpi/…/bin:…/compiler/2026.0/bin:…
LIBRARY_PATH=/opt/intel/oneapi/tcm/1.5/lib:…/compiler/2026.0/lib:…
CMPLR_ROOT=/opt/intel/oneapi/compiler/2026.0
SETVARS_COMPLETED=1

Re-sourcing setvars.sh in the RUN was redundant. It was also actively breaking the build: setvars.sh sees SETVARS_COMPLETED=1, refuses to re-run, prints its usage, and exits 3. --force gets dropped by dash's . builtin (Docker's default sh -c on Ubuntu 24.04), so setvars.sh never actually sees it. The exit-in-a-sourced-script kills the parent shell before pip runs, defeating even the || echo … guard from #38617.

Fix

Drop the setvars.sh invocation from the RUN and let the base image's env do its job:

RUN TMS_PLATFORM=xpu pip install --no-cache-dir --no-build-isolation \
    git+https://github.com/fzyzcjy/torch_memory_saver.git@${TORCH_MEMORY_SAVER_REF}

Also simplify CMD for the same reason — the container's ENV already has everything set, so source setvars.sh --force && exec bash reduces to just bash.

Why CI didn't catch this

.github/workflows/release-docker-intel-xpu-nightly.yml only fires on:

on:
  workflow_dispatch:
  schedule:
    - cron: '0 12 * * *'

There is no pull_request trigger. Any PR that touches docker/xpu.Dockerfile or python/pyproject_xpu.toml can merge without ever building the affected image — the first time the change actually runs is the next nightly, hours after merge. That's exactly the pathway #29935 → the original exit 3, and #38617's diagnostic fix → this exit 3.

Recommended follow-up (kept out of this PR because updating workflow files requires the workflow OAuth scope, which the author's token doesn't hold — a maintainer or a second PR can land it):

on:
  workflow_dispatch:
  schedule:
    - cron: '0 12 * * *'
  pull_request:
    paths:
      - 'docker/xpu.Dockerfile'
      - 'python/pyproject_xpu.toml'
      - '.github/workflows/release-docker-intel-xpu-nightly.yml'

With that in place, any PR that could break the XPU image gets a docker build in CI and can't merge with a broken Dockerfile — the same protection the CUDA image already has.

Test plan

• Trigger Release Docker Images Nightly (Intel XPU) on this branch → build reaches Push intel/sglang-dev and succeeds.
• docker run --rm <built-image> python -c "import torch_memory_saver; print(torch_memory_saver.__file__)" succeeds.

---

CI States

Latest PR Test (Base): ✅ Run #34337535233
Latest PR Test (Extra): ✅ Run #34428340387
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.