Files
LeonandCodex 39a1a971c6 [Bugfix] Preserve streaming tool calls, errors and logprob metadata (#1317)
Backport related API streaming correctness fixes from Leonccaa/ReVolta.
Preserve every required call, engine termination reasons, suppressed-delta
logprobs and pre-token errors; omit empty call arrays and handle null top_logprobs.

ReVolta sources:
- d525a049dcf8d0b550bae2660d827a76d9e39656: [Bugfix] Stream every required tool call in each delta
  Adapted from vllm-project/vllm PR #60354, commit 7bbf3208a4620f176ebf6470578f0c7dc1e1aff6.
- 79b722bc1712c19029c8e97d05e3217011a107e9: [Bugfix] Preserve truncation reasons for tool calls
  Adapted from vllm-project/vllm PR #46303, commit 6f154ada6d8cdb3ba4224167bec287f01df6beeb.
- f90b9408f4cb3e9fcff0e9de710860a7ec1ca46b: [Bugfix] Retain logprobs when parsers suppress deltas
  Adapted from vllm-project/vllm PR #58583, commit 0caa4a5dd7ca9039a76b674660264d71020df7cc.
- 2f8ebb34418f033f3b237a752fbc56d0cfb2c9f4: [Bugfix] Report pre-token streaming errors before filtering
  Adapted from vllm-project/vllm PR #40986, commit 3a22e208ca3abea8d0ff477f2cc0909b731e1edb.
- 172c1ebd68f4e952da65eb391e76ce34f004a8e2: [Bugfix] Omit empty tool call arrays from API messages
  Adapted from vllm-project/vllm PR #44105, commit 29080b9b727bebdc9bd0978b90a1381c94700fce.
- 4ba2a9af4142da5f320e76251f7dc106109b68c5: [Bugfix] Treat null top_logprobs as sampled-token logprobs
  Adapted from vllm-project/vllm PR #47838, commit 2f8297631469ee33a168699b3d128d3448da9cbe.

Signed-off-by: Leonccaa <166551845+Leonccaa@users.noreply.github.com>
Co-authored-by: Codex <codex@openai.com>
2026-11-06 06:58:00 +08:00
..