[watermarking] Dual-key gumbel-max watermarking for speculative decoding support (#56122)

Signed-off-by: Raphael Rialland <raphael.rialland@mistral.ai>
Signed-off-by: Simon Veitner <sveitner@redhat.com>
Signed-off-by: Tomas Ruiz <tomas.ruiz.te@gmail.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Simon Veitner <sveitner@redhat.com>
Co-authored-by: Tomas Ruiz <tomas.ruiz.te@gmail.com>
This commit is contained in:
Raphaël Rialland
2026-09-12 18:42:01 +00:00
committed by GitHub
co-authored by OpenAI Codex Simon Veitner Tomas Ruiz
parent 1ee4be4dbc
commit 7ee8a6dd01
28 changed files with 3460 additions and 245 deletions
@@ -137,7 +137,7 @@ class DSparkSpeculator(DFlashSpeculator):
# sample_pos is the predicted token's position P. Sampling keys a draw
# by the position before the sampled token, P-1.
return gumbel_sample(
sampled = gumbel_sample(
logits,
idx_map,
self.temperature,
@@ -149,6 +149,14 @@ class DSparkSpeculator(DFlashSpeculator):
logits_cache_col=self._step_cols[step],
use_fp64=self.use_fp64_gumbel,
)
if self.draft_watermarker is not None:
sampled = self.draft_watermarker.sample(
logits,
sampled,
idx_map,
self.temperature,
)
return sampled
def _sample_sequential(self, num_reqs: int, head_hidden: torch.Tensor) -> None:
if self._draft_topk is not None: