[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
@@ -0,0 +1,29 @@
model_name: "Qwen/Qwen3-4B-FP8"
num_questions: 1319
num_fewshot: 5
max_tokens: 256
use_chat_completions: false
temperature: 1.0
seed: 42
# Reference accuracy from the non-watermarked DSpark regression test.
accuracy_threshold: 0.801
tolerance: 0.02
# Measured TP1 acceptance length: 4.115; allow a 0.2 regression.
min_acceptance_length: 3.915
env:
VLLM_USE_V2_MODEL_RUNNER: "1"
server_args: >-
--tensor-parallel-size 2
--max-model-len 4096
--gpu-memory-utilization 0.92
--no-enable-prefix-caching
--speculative-config.method dspark
--speculative-config.model deepseek-ai/dspark_qwen3_4b_block7
--speculative-config.num_speculative_tokens 7
--speculative-config.draft_sample_method probabilistic
--watermark-config.algorithm dual_key_gumbel
--watermark-config.key 42
--watermark-config.context_width 4
--watermark-config.deduplicate_contexts none
@@ -0,0 +1 @@
Qwen3-4B-FP8-DSpark-watermark-TP2.yaml