chore(release): backport #33853 and #33864 onto patch-1.93.0rc2 to complete the 1.93.0 stable cut

[yuneng-berri]

Relevant issues

Completes the 1.93.0 stable cut on top of #33847. That PR was merged from its 11-pick state, so two staging changes intended for the cut still needed to land on patch-1.93.0rc2: #33853, which restores fresh-database Docker deployments (prisma CLI and engines baked at a fixed path, the fix for the fresh-deploy migration failures tracked since v1.90.0), and #33864, the hardening follow-up to #33721 that was disclosed on #33847 (a deployment configured with a non-numeric token limit made the whole GET /v1/models listing 500; it now degrades to a response without limits, restoring the listing behavior the line had before the cost-map switch)

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

What is included

• fix(docker): bake prisma CLI and engines at a fixed path so fresh-DB migrations work for any uid offline #33853 fix(docker): bake prisma CLI and engines at a fixed path so fresh-DB migrations work for any uid offline (verbatim pick of staging squash 567ebcb3e9)
• fix(router): treat malformed configured token limits as absent on /v1/models #33864 fix(router): treat malformed configured token limits as absent on /v1/models (pick of staging squash ef7007c3dd; every added and removed line is byte-identical to the source, the patch-id differs only through surrounding context the line does not carry)

Both commits carry (cherry picked from commit ...) footers whose SHAs are reachable from litellm_internal_staging; no merge commits, no _experimental/out/ artifacts, no version bump (the branch still rides the pending 1.93.0). Neither pick touches pyproject.toml or uv.lock, so the green image-scan and osv-scan dispatch runs from #33847 remain representative for the dependency set; #33853 changes Dockerfile/docker/Dockerfile.database, which the image-scan workflow does not build, and is a patch-identical pick of the change staging validated the same day

Known noise on this line

tests/test_litellm/proxy/test_proxy_utils.py::test_get_custom_url fails on the untouched base (environment-dependent URL expectation) exactly as documented on #33847; it is the same single failure before and after these picks

Screenshots / Proof of Fix

Targeted test files (tests/test_litellm/proxy/test_proxy_utils.py + tests/test_litellm/test_router.py) on this branch: 172 passed, 1 failed (the known-noise test above). That includes #33864's three regression tests, which fail on the base without the guard with the exact defect (ValueError: invalid literal for int() with base 10: '128,000')

Live proxy on this branch with a deployment configured as model_info: {max_input_tokens: "128,000"}, the request that returned 500 on the pre-guard picks (demonstrated on #33847):

$ curl -s -o /dev/null -w 'HTTP %{http_code}\n' http://localhost:4001/v1/models -H 'Authorization: Bearer sk-1234'
HTTP 200
# from the same response:
gpt-4o          in: 128000  out: 16384   (cost map)
custom-good     in: 32000   out: 8000    (configured model_info)
bad-limit-model in: None    out: None    (malformed value treated as absent)

Type

🐛 Bug Fix

Changes

Two cherry-picks only; no hand-written code. With these, patch-1.93.0rc2 carries the full intended 1.93.0 stable set and release tooling can cut v1.93.0 from the branch

[greptile-apps]

Greptile Summary

This PR backports two cherry-picks onto patch-1.93.0rc2 to complete the 1.93.0 stable cut: a Dockerfile fix that bakes Prisma CLI and engines at a fixed /opt/prisma path (restoring fresh-database migration support for non-root UIDs), and a router.py guard that degrades malformed configured token limits to None instead of surfacing a ValueError as a 500 on /v1/models.

• Dockerfile / docker/Dockerfile.database: prisma generate now runs with HOME=/opt/prisma and PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries, and the runtime image pins PRISMA_CLI_PATH, PRISMA_OFFLINE_MODE=true, and related env vars; a build-time test -x / test -f guard verifies the baked artifacts before the image is finalised.
• litellm/router.py: get_configured_token_limits now wraps int() conversion in a _as_int helper that returns None on any TypeError / ValueError, and explicitly excludes bool values from being coerced to 0/1.
• Tests: Three additive-only test functions cover the full malformed-value matrix, valid numeric-string coercion, and the end-to-end /v1/models listing path — all new, none modify existing assertions.

Confidence Score: 5/5

Both fixes are narrowly scoped, additive-only, and come with dedicated regression tests that fail on the base without the guard.

The router change is a one-function defensive wrapper around an existing int() cast, the Docker change moves Prisma artifacts to a fixed world-readable path with a build-time binary check, and no existing tests were modified — only new ones added.

No files require special attention; all five changed files are straightforward and well-tested.

Important Files Changed

Filename	Overview
Dockerfile	Bakes Prisma CLI and engines at /opt/prisma with fixed env vars; adds runtime sanity checks confirming the CLI binary and index.js are in place before the image is finalized.
docker/Dockerfile.database	Identical Prisma path-pinning change as main Dockerfile; both are consistent, and the database-specific Dockerfile now also gains the PRISMA_OFFLINE_MODE and CLI_PATH env vars.
litellm/router.py	Introduces _as_int helper in get_configured_token_limits that silently degrades malformed configured token limits (e.g. "128,000") to None instead of propagating a ValueError that caused a 500 on /v1/models.
tests/test_litellm/test_router.py	Adds two new regression tests: one covering the full malformed-value matrix (empty string, "unlimited", comma-formatted, list, dict, bool) and one confirming valid numeric strings still coerce correctly.
tests/test_litellm/proxy/test_proxy_utils.py	Adds end-to-end regression test for the /v1/models listing path: verifies that a model configured with max_input_tokens="128,000" produces a 200 response with the token-limit fields absent rather than raising a 500.

_{Reviews (1): Last reviewed commit: "fix(router): treat malformed configured ..." | Re-trigger Greptile}