fix(sambanova): emit empty delta on usage-only final stream chunk

[ebarkhordar]

Description

SambaNovaCloud sets stream_options={"include_usage": True} by default (base.py:201-204), so a streaming response ends with a usage-only chunk where choices == []. In stream_chat, that final chunk takes the else branch (base.py:553-561), which builds additional_kwargs but never reassigns content_delta. The trailing yield then re-emits the previous chunk's token as its delta:

yield ChatResponse(
    message=ChatMessage(role=role, content=content, additional_kwargs=additional_kwargs),
    delta=content_delta,   # still the last text chunk's content on the usage chunk
    raw=partial_response,
)

So "".join(r.delta for r in client.stream_chat(...)) no longer equals the final message.content: any consumer that rebuilds the text from deltas (the usual streaming pattern) gets the last token duplicated. message.content itself stays correct, only the re-emitted delta is wrong.

The invariant a streaming generator should hold is that the concatenation of the deltas equals the final message.content, and a usage-only chunk carries no new text, so its delta must be "". This is what the OpenAI-compatible reference implementations do.

There is no tracking issue; this was found by reading the code and reproduced as described below.

Fix

Set content_delta = "" in the else (empty-choices) branch so the usage-only chunk yields an empty delta. One line plus a comment; no behavior changes for the text chunks or for message.content.

Version Bump?

• Yes (0.6.0 to 0.6.1)

Type of Change

• Bug fix (non-breaking change which fixes an issue)

How Has This Been Tested?

• I added a new unit test.

Added test_stream_chat_usage_chunk_emits_empty_delta. It stubs _handle_streaming_request to yield a text chunk followed by the usage-only chunk (choices == [], plus usage/model/system_fingerprint/created) that the provider sends when include_usage=True, then asserts "".join(deltas) == final content == "Hello" and that the last delta is "".

Verified in a clean Docker container (python:3.11) against current main:

• Before the fix the test fails with assert 'HelloHello' == 'Hello'.
• After the fix it passes, and the full tests/test_llms_sambanovasystems.py file is green (the live-API tests are skipped without a key).
• ruff check, ruff format --check, and codespell pass on the two changed files.

Not verified: I did not call the real SambaNova Cloud endpoint (no API key). The test reproduces the usage-chunk shape via a stub, using the exact fields the existing else branch already reads.

AI assistance: this change was prepared with AI assistance. I reproduced the bug on main, validated the fix with the Docker differential and repo lint above, and reviewed the diff line by line.

[ebarkhordar]

No rush on this. CI is green and the branch still applies cleanly on main, so it is ready whenever it fits your review queue. Happy to add a regression test for the usage-only final chunk if that would help.

[ebarkhordar]

Correcting a line in my note above: I said CI is green, and that was wrong. The only check reporting on this PR is check-new-pyproject. Unit Testing, Build Package, Core Typecheck, Linting and CodeQL are all held awaiting approval of the fork workflow run, so none of them has executed against this branch and I have no CI evidence to point at.

What I do have is local: the added regression test fails on main and passes with the fix. The branch does still apply cleanly on main.

[ebarkhordar]

Closing this to keep my open PR list at a size I can actually shepherd. Nothing here is
blocked on anyone and this is not a nudge.

The change still applies if it is useful later, and I am happy to reopen and rebase it
at any point. The branch stays up so reopening is a click.