Skip to content

feat: parse vLLM's native sampling_mask field #144

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

Merged
hallerite merged 3 commits into from
Sep 1, 2026
Merged

feat: parse vLLM's native sampling_mask field #144

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
04d4704
feat: parse vLLM's native sampling_mask field
mikasenghaas Aug 29, 2026
f062242
Merge remote-tracking branch 'origin/main' into feat/native-sampling-…
mikasenghaas Aug 31, 2026
a9dccfb
Merge remote-tracking branch 'origin/main' into feat/native-sampling-…
mikasenghaas Sep 1, 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
2 changes: 1 addition & 1 deletion renderers/base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2095,7 +2095,7 @@ def build_trajectory_step(
"completion_mask": [True] * len(completion_ids),
"completion_logprobs": [0.0] * len(completion_ids),
"routed_experts": None,
"kept_tokens": None,
"sampling_mask": None,
}
if (
full_rendered.multi_modal_data is not None
Expand Down
10 changes: 4 additions & 6 deletions renderers/client.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,6 @@

_request_logger = logging.getLogger("renderers.client")
ROUTED_EXPERTS_DATA_PREFIX = b'"routed_experts":{"data":"'
KEPT_TOKENS_IDS_PREFIX = b'"kept_tokens":{"ids":"'
# vLLM uses this value both when sampled-token evidence is missing and as a
# lower-bound clamp, so receiving it cannot prove the real logprob was returned.
VLLM_LOGPROB_SENTINEL = -9999.0
Expand Down Expand Up @@ -129,12 +128,9 @@ def _strip_base64_field(raw: bytes, prefix: bytes) -> tuple[bytes, memoryview |

def parse_generate_response(raw: bytes) -> dict[str, Any]:
stripped, routed_data = _strip_base64_field(raw, ROUTED_EXPERTS_DATA_PREFIX)
stripped, kept_ids_data = _strip_base64_field(stripped, KEPT_TOKENS_IDS_PREFIX)
payload: dict[str, Any] = json.loads(stripped)
if routed_data is not None:
payload["choices"][0]["routed_experts"]["data"] = routed_data
if kept_ids_data is not None:
payload["choices"][0]["kept_tokens"]["ids"] = kept_ids_data
return payload


Expand Down Expand Up @@ -374,7 +370,9 @@ def _prepare():
parsed = renderer.parse_response(completion_ids, tools=tools)

routed_experts = choice.get("routed_experts")
kept_tokens = choice.get("kept_tokens")
# vLLM's native kept-set sampling masks (``--return-sampling-mask``):
# one list of surviving vocab ids per completion token.
sampling_mask = choice.get("sampling_mask")

# /inference/v1/generate returns finish_reason in {"stop","length",...} —
# never "tool_calls" (a chat-completions concept). Promote stop→tool_calls
Expand Down Expand Up @@ -403,7 +401,7 @@ def _prepare():
"tool_calls": parsed.tool_calls,
"finish_reason": finish_reason,
"routed_experts": routed_experts,
"kept_tokens": kept_tokens,
"sampling_mask": sampling_mask,
# 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