Reproduction link or steps

Workload-dependent bloat on a large monorepo. I do not have a minimal stackblitz repro yet — happy to bisect/profile further if the maintainers point at suspect areas.

Public repo used for the comparison: https://github.com/lobehub/lobehub
Branch / pinned package state: canary (currently vite ^8.0.9, @vitejs/plugin-react ^5.1.4).

Steps (both runs in the same checkout, only swapping vite + plugin-react versions; backend Next.js was NOT started, only the SPA dev server):

1. pnpm install
2. pnpm run dev:spa (which runs vite --port 9876 against the repo's vite.config.ts)
3. Wait until VITE … ready
4. Headless Chrome via Playwright page.goto(/) + waitForTimeout(25_000) — this is important; vite triggers ✨ new dependencies optimized → reloading ~20–28 s after ready, so a shorter window catches the page mid-reload, not its settled first paint. Both versions render the same SPA home (sidebar Search / Home / Tasks / Pages / Agents / Generation / Community / Resources + chat input + model picker), with pageErrors empty and the only console errors being the expected 500/502 on TRPC because no backend is running.
5. Sample the entire process tree with vmmap -summary <pid> plus children.

For the v7 control I had to:

• Pin vite ^7.3.3, @vitejs/plugin-react ^4.7.0.
• Drop experimental.bundledDev: false (v8-only).
• Drop build.rolldownOptions (v8-only).
• Replace resolve: { tsconfigPaths: true } (v8-only) with the vite-tsconfig-paths plugin.

The same v7-compatible vite.config.ts was used for the v8 run as well, so the only variables across the two runs are the vite + plugin-react versions themselves.

What is expected?

Vite 8 dev memory should be in the same order of magnitude as vite 7 for the same workload — ideally lower, since rolldown is intended to be a more efficient bundler than esbuild + rollup. At minimum the integrated single-process design should not regress total physical footprint vs v7's two-process model.

What is actually happening?

vite 8's main node process plateaus at ~7-9 GB physical memory while vite 7's process tree (main + esbuild service) settles under 1 GB on the same workload — roughly +7× on tree physical footprint, +4–5× on peak.

Notes on methodology:

• macOS RSS for the esbuild Go process is mmap-inflated; vmmap "Physical footprint" is the more reliable cross-process metric.
• experimental.bundledDev is at its default false, so rolldown's dev engine itself is not driving the workload — the increase appears in the regular dep-optimize / module-graph dev path that now routes through rolldown instead of esbuild.
• I posted earlier numbers in this issue claiming +50%; those were wrong because (a) the chrome warmup window I used was shorter than the time it takes for the optimized dependencies changed → reloading cycle to settle, and (b) I was sampling v7's esbuild during its transient post-bundle peak, before the service idled back down to ~174 MB. The numbers above are the corrected, fully-settled measurement. See the comment on this issue for the full methodology trace.

System Info

System:
  OS: macOS 26.4.1
  CPU: (10) arm64 Apple M5
  Memory: 9.12 GB free / 32.00 GB
Binaries:
  Node: 24.15.0
  pnpm: 10.33.0
  npm:  11.6.2
Browsers:
  Chrome: 147.0.7727.138
npmPackages (vite 8 run):
  vite:                 8.0.11
  @vitejs/plugin-react: 5.2.0
  rolldown (bundled):   1.0.0-rc.18
npmPackages (vite 7 control):
  vite:                 7.3.3
  @vitejs/plugin-react: 4.7.0
  vite-tsconfig-paths:  6.1.1

Any additional comments?

• This is workload-dependent: the lobe-chat monorepo has a heavy server.warmup.clientFiles glob pre-bundling many packages/**/src/**/*.ts files. I assume that interacts with rolldown's dep handling more aggressively than v7's esbuild.
• I have not bisected @vitejs/plugin-react 4 → 5 separately, so a non-zero share of the regression could live there rather than in rolldown — but plugin-react@5 is required by vite 8, so users cannot avoid it on v8.
• Filing here per the discussion that vite 8 dev now goes through rolldown end-to-end. Happy to refile at vitejs/vite if this is more appropriately routed there.
• I also bisected across all 31 publicly released vite 8 versions (19 betas + 12 stables); see the bisect comment for the table. Short version: the regression is present from 8.0.0-beta.0 (rolldown 1.0.0-beta.53) onward — there is no within-8.x patch where the footprint jumps. The slow drift across stables (8.0.0 ~6.2 GB → 8.0.11 ~6.9 GB) is dwarfed by the v7→v8 step.
• Happy to capture node --inspect heap snapshots, run with RUSTFLAGS=-C debug-assertions=on, or share a profile of the napi binding allocations if any of that helps localize where the ~6 GB of resident memory in the rolldown native module lives.