chore(release/1.5.0): cherry-pick #13640, #14317, #14525, #12856, #13951

[ishandhanani]

Summary

Cherry-picks the following merged PRs into release/1.5.0 as a single release branch PR:

• feat(llm): expose training metadata for RL rollouts #13640: feat(llm): expose training metadata for RL rollouts
• fix(sglang): preserve stop tokens and incremental streams #14317: fix(sglang): preserve stop tokens and incremental streams
• feat(frontend): expose detailed finish reason #14525: feat(frontend): expose detailed finish reason
• feat(rl): add Slime external rollout example #12856: feat(rl): add Slime external rollout example
• fix(sglang): sync discovery from native pause state #13951: fix(sglang): sync discovery from native pause state

PR #14525 depends on the prompt_token_ids support introduced in #13640, so #13640 is included as a dependency.

Validation

• cargo fmt --all
• Targeted cargo test -p dynamo-llm --no-default-features --lib filters passed:
  • delta (63 tests)
  • extensions (47 tests)
  • timing (32 tests)
  • preprocessor (188 tests)
  • sglang (19 tests)
• Full cargo test -p dynamo-llm run shows unrelated network/timing failures in the local macOS environment.

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-01a0b071-be4e-7559-adc0-e6ff593e3699.docs.buildwithfern.com/dynamo/dev

[dmitry-tokarev-nv]

Backport review: 5 cherry-picks onto release/1.5.0

I compared each cherry-pick against the commit that merged on main, not against the
branch head, because all five landed as squash merges. 46 of the 47 file changes are
byte-identical. The one file that did not come across was already on the branch. The Rust
and the Python code both work on the older base.

Approved. One P3 below.

[P3] The description does not say that one file from #13640 was not carried.
lib/llm/src/protocols/openai/chat_completions/delta.rs. The omission is correct,
because the branch already holds that change at line 125. Please add one line to the
description so the next reader does not have to prove it again.

Per-pull-request comparison: source merge commit against cherry-pick

I split each diff per file, then took a SHA-256 of the added lines and of the removed
lines on each side. Equal byte counts prove nothing, so the checksums are the test.

PR	source merge on main	cherry-pick	files	result
#13640	50c6fea10d	931bae38b8	11 vs 10	10 identical, 1 not carried
#14317	8dcfd50cde	a086dffa25	12 vs 12	all identical
#14525	c3e05f0244	73c0a329c3	4 vs 4	all identical
#12856	09497580bb	06a4db49c3	12 vs 12	all identical
#13951	8c5a737231	f27386479c	8 vs 8	all identical

c3e05f0244 has 8dcfd50cde as its parent, so #14525 sat directly on top of #14317 on
main. The cherry-picks keep that order.

The one omission, and why it is correct

lib/llm/src/protocols/openai/chat_completions/delta.rs is the only file from a source
pull request that is not in its cherry-pick. #13640 changed it by +51/-23 on main.

The branch already had that change. I took all 48 added lines of the #13640 hunk and
looked for each one in the file at the merge base 3507149353. 47 match word for word,
including the test test_chat_logprobs_without_top_logprobs_include_sampled_token. The
48th line differs because the branch wraps that call in a guard the main version at
that time did not have:

let converted = if self.suppress_top_logprobs {
    Vec::new()
} else {
    convert_backend_top_logprobs(top_lps, t, *tid, lp, return_as_ids)
};

That guard is at lib/llm/src/protocols/openai/chat_completions/delta.rs:125. It
survives all five cherry-picks: suppress_top_logprobs appears 4 times in the file at
the merge base and 4 times at the head.

Does each change still work on the older base

The branch differs from main in the files these patches touch, so a clean apply is not
proof. The largest gaps at the point each patch applied:

file	branch against main
lib/llm/src/preprocessor.rs	+142/-1026
components/src/dynamo/frontend/sglang_prepost.py	+55/-17
components/src/dynamo/sglang/request_handlers/handler_base.py	+45/-12
lib/llm/src/backend.rs	+3/-4

What the ported code calls, and whether the branch has it:

• #13640 calls RequestTracker::set_prompt_token_ids and prompt_token_ids(). The same
  set adds both in lib/llm/src/protocols/common/timing.rs.
• #13951 calls resolve_configured_engine_routes from
  components/src/dynamo/sglang/engine_routes.py. That file is byte-identical on both
  branches, so the "pause_generation:tm" descriptor form works.
• #14317 renames _strip_trailing_eos_token_ids to _strip_matched_stop_token_ids. No
  caller of the old name is left on the branch.
• #14317 gives flush_pending a new keyword-only argument stop_terminated with no
  default. Both call sites, at components/src/dynamo/frontend/sglang_processor.py:878
  and :915, pass it.
• #13951 deletes components/src/dynamo/sglang/pause.py. Nothing on the branch still
  imports it.
• The Slime example uses apiVersion: nvidia.com/v1beta1. That version is served and is
  the storage version on this branch, and 75 other examples already use it.

On the x86_64 Linux box with the pinned 1.96.1 toolchain:

cargo test -p dynamo-llm --no-default-features --lib
test result: ok. 2388 passed; 0 failed; 3 ignored

All 19 changed Python files compile.

Release-branch rules, push shape, and what the branch gained

• Base branch is release/1.5.0, which is what the title says.
• No version file, pin, or lock file is in the 41-file diff.
• CODEOWNERS is generated. This set changes it together with
  .github/codeowners/areas.yaml, one line each, and the codeowners check passed.
• The issue timeline has no head_ref_force_pushed event. The 5 commits sit in a line
  on the merge base 3507149353.
• Since that merge base, release/1.5.0 gained #15017 and #15027. Neither touches any of
  the 41 files, so no contract moved under this set.
• The lineage carries no v1.5.0 release tag, only -dev tags, so removing the two
  control/* engine routes lands before the release, not after it.

Test lane, and mutation results with controls

The lane is sglang-runtime / Test cuda13.0, amd64, which runs
pytest -m "pre_merge and sglang and gpu_0" from .github/workflows/pr.yaml. Run
against the changed test files on this branch:

683/763 tests collected (80 deselected)

The new test tests/test_sglang_gms_runtime_routes.py::test_sglang_pause_resume_use_native_engine_routes
is in that set. The 4 tests in components/src/dynamo/thunderagent_router/tests/test_args.py
are not, because the file carries no sglang mark. They pass on their own.

I saw 14 failures in components/src/dynamo/sglang/tests/test_sglang_unit.py. The merge
base gives the same 14 failures with the same names, so they belong to the local
environment and not to this set.

Mutation, both directions, with a SHA-256 after each write:

step	file checksum	result
baseline	de5bc2b05b0f92cd	1 passed
rename the pause route in tests/gpu_memory_service/common/runtime.py	e8f27858a534bf3a	1 failed
restore	de5bc2b05b0f92cd	1 passed
baseline	56f8d071f062b6ee	1 passed
return early in _strip_matched_stop_token_ids	192eaf11dbebe002	1 failed
restore	56f8d071f062b6ee	1 passed

Controls: components/src/dynamo/common/backend/tests/test_logprobs.py stayed at
39 passed, 1 skipped through every step of the first mutation. For the second, the
enclosing class went from 30 passed to 2 failed and 28 passed, so 28 tests did not move.

Claims in the description, checked

Every number in the Validation section is exact.

cargo fmt --all -- --check   -> clean
delta        -> 63 passed
extensions   -> 47 passed
timing       -> 32 passed
preprocessor -> 188 passed
sglang       -> 19 passed