Skip to content

feat(client): pass through sampling-mask logprobs and by-handle payload segments #170

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
faresobeid wants to merge 2 commits into
base: perf/packed-completion-logprobs
Choose a base branch
from
Open

feat(client): pass through sampling-mask logprobs and by-handle payload segments #170

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ecfeef7
feat(client): pass through sampling-mask logprobs
faresobeid Oct 6, 2026
d3a1af0
feat(client): pass through by-handle payload segments
faresobeid Oct 3, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 11 additions & 4 deletions renderers/client.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -261,9 +261,9 @@ async def generate(
Returns a dict with: request_id, prompt_ids, renderer_prompt_ids,
mm_placeholders, completion_ids, completion_logprobs, content,
reasoning_content, tool_calls, finish_reason, routed_experts,
multi_modal_data, prompt_attribution. ``renderer_prompt_ids`` is the
unexpanded logical prompt when ``process_multimodal=False`` and ``None``
otherwise.
sampling_mask, sampling_mask_logprobs, payload, multi_modal_data,
prompt_attribution. ``renderer_prompt_ids`` is the unexpanded logical
prompt when ``process_multimodal=False`` and ``None`` otherwise.

``prompt_attribution`` is the renderer's :class:`RenderedTokens` for
the prompt — either the one this call computed via
Expand Down Expand Up @@ -396,8 +396,11 @@ def _prepare():
routed_experts = choice.get("routed_experts")
# vLLM's native kept-set sampling masks (``--return-sampling-mask``):
# one list of surviving vocab ids per completion token, or (prime-rl's
# server) packed CSR ``{"ids", "counts"}`` base64 int32 arrays.
# server) packed CSR ``{"ids", "counts"}`` base64 int32 arrays. With score
# centering on, prime-rl's server also returns the sampler's renormalized
# logprob of each of those ids (packed float32, parallel to the ids).
sampling_mask = choice.get("sampling_mask")
sampling_mask_logprobs = choice.get("sampling_mask_logprobs")

# /inference/v1/generate returns finish_reason in {"stop","length",...} —
# never "tool_calls" (a chat-completions concept). Promote stop→tool_calls
Expand Down Expand Up @@ -429,6 +432,10 @@ def _prepare():
"reasoning_complete": parsed.reasoning_complete,
"routed_experts": routed_experts,
"sampling_mask": sampling_mask,
"sampling_mask_logprobs": sampling_mask_logprobs,
# By-handle side arrays: segments of a shared payload file the inference
# server wrote instead of returning routed_experts / sampling_mask inline.
"payload": choice.get("payload"),
# The mm sidecar consumed on the request side, surfaced back so
# callers can persist it on the trajectory step for downstream
# multi-turn bridging and training-sample construction.
Expand Down
Loading