Reuse the existing MoE workspace for DeepGEMM grouped-GEMM warmup activations, scales, and output, avoiding a second large scratch allocation after profiling. Ordinary linear warmup keeps temporary buffers so dense models do not acquire a persistent workspace.

B300 validation with two-layer dummy block-FP8 Qwen3 and Qwen3-MoE models (hidden size 1024, intermediate size 3072, 1024 batched tokens, eager execution):

• Both completed startup and generated the same four tokens as the baseline.
• Final workspace stayed at 0 MiB (dense) and 54 MiB (MoE).
• Additional MoE warmup peak fell from 39.25 MiB to 7.04 MiB.

PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 .venv/bin/python -m pytest \
  tests/model_executor/test_deep_gemm_warmup.py -q
# 6 passed

.venv/bin/pre-commit run --files \
  vllm/model_executor/warmup/deep_gemm_warmup.py
# Passed

Alternative to #57258, which moves warmup before profiling. Related #49469 removes DP scaling; this change preserves warmup shapes and ordering. Duplicate-work searches found no existing PR implementing this workspace reuse.

Codex assisted with implementation and validation. Reviewed by submitter.