[feature] Watermarked generation and detection (Gumbel-max algorithm) (#54053)

Signed-off-by: Raphael Rialland <raphael.rialland@mistral.ai>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Raphaël Rialland
2026-09-09 21:00:15 -07:00
committed by GitHub
co-authored by OpenAI Codex mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 8c34a37232
commit ea40bb9e90
32 changed files with 1693 additions and 15 deletions
+14
View File
@@ -106,6 +106,7 @@ from vllm.config.parallel import (
from vllm.config.scheduler import SchedulerPolicy
from vllm.config.utils import get_field
from vllm.config.vllm import OptimizationLevel, PerformanceMode
from vllm.config.watermarking import WatermarkConfig
from vllm.logger import init_logger, suppress_logging
from vllm.platforms import CpuArchEnum, current_platform
from vllm.plugins import load_general_plugins
@@ -656,6 +657,7 @@ class EngineArgs:
spec_method: str | None = None
spec_model: str | None = None
spec_tokens: int | None = None
watermark_config: dict[str, Any] | None = None
diffusion_config: dict[str, Any] | None = None
show_hidden_metrics_for_version: str | None = (
@@ -1680,6 +1682,8 @@ class EngineArgs:
vllm_group.add_argument(
"--spec-tokens", **speculative_kwargs["num_speculative_tokens"]
)
vllm_kwargs["watermark_config"]["type"] = optional_type(json.loads)
vllm_group.add_argument("--watermark-config", **vllm_kwargs["watermark_config"])
vllm_kwargs["diffusion_config"]["type"] = optional_type(json.loads)
vllm_group.add_argument(
"--diffusion-config", "-dc", **vllm_kwargs["diffusion_config"]
@@ -1986,6 +1990,14 @@ class EngineArgs:
cfg = json.loads(cfg)
return DiffusionConfig(**cfg)
def create_watermark_config(self) -> WatermarkConfig | None:
if self.watermark_config is None:
return None
cfg = self.watermark_config
if isinstance(cfg, str):
cfg = json.loads(cfg)
return WatermarkConfig(**cfg)
def create_observability_config(self) -> ObservabilityConfig:
return ObservabilityConfig(
show_hidden_metrics_for_version=self.show_hidden_metrics_for_version,
@@ -2366,6 +2378,7 @@ class EngineArgs:
target_parallel_config=parallel_config,
)
diffusion_config = self.create_diffusion_config()
watermark_config = self.create_watermark_config()
self._set_default_max_num_seqs_and_batched_tokens_args(
usage_context,
@@ -2602,6 +2615,7 @@ class EngineArgs:
kernel_config=kernel_config,
lora_config=lora_config,
speculative_config=speculative_config,
watermark_config=watermark_config,
diffusion_config=diffusion_config,
structured_outputs_config=self.structured_outputs_config,
observability_config=observability_config,