[jhen0409]

Overview

Ref #27357 (gzip only, no brotli).

This adds a LLAMA_ZLIB CMake option (default ON, gracefully disabled when zlib is not found) that builds the vendored cpp-httplib with CPPHTTPLIB_ZLIB_SUPPORT. cpp-httplib already implements both directions:

• Requests with Content-Encoding: gzip are decompressed automatically (decompressed size capped by CPPHTTPLIB_PAYLOAD_MAX_LENGTH, 100MB).
  This helps us save a lot of traffic in long, multi-turn conversations or even multimodal message conversations on compression API requests.
• JSON responses are compressed per Accept-Encoding negotiation. SSE streaming is unaffected (text/event-stream is excluded by httplib).
• SSE streaming is unaffected (text/event-stream is excluded by httplib)

The fix in cpp-httplib

Enabling zlib exposed a double-encoding bug in cpp-httplib: the web UI serves build-time-gzipped assets with Content-Encoding: gzip set by the handler, and httplib compresses them a second time (Safari renders raw gzip bytes). The patch skips compression when the handler already set Content-Encoding, deciding the coding once so headers and body agree.

I'm submitted this patch upstream (PR: yhirose/cpp-httplib#2575, merged) and will keep this PR as a draft until the upstream author agrees with the approach the cpp-httplib released new version.

Additional information

Validation:

• gzip request body on /v1/chat/completions (text + image): same 200 and identical prompt_tokens as the uncompressed request; the gzip request even hits the prompt cache populated by the plain one, confirming the decompressed bytes are byte-identical.
• Web UI loads correctly on Safari and Chrome (single Content-Encoding header, single gzip layer); Accept-Encoding: gzip on /props returns a compressed response; SSE streaming unaffected.

Possible follow-ups: compressing request bodies from the web UI, brotli.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, helped by Claude Code.

[jhen0409]

Rebased master, now it's ready for review.

Also, I found it cannot set quality param for brotli, so it's significantly increases response time by default, I think the benefits of brotli support to be minimal at the moment.