prompt_embeds
Enable vLLM's --enable-prompt-embeds by default under Output Protection (OP).
Stained Glass Proxy sends Stained Glass Transformed prompt embeddings, so the OP Inference Server must
accept prompt_embeds. In vLLM this capability is gated behind enable_prompt_embeds, which is exposed only
as the --enable-prompt-embeds CLI flag.
This module flips the flag's argparse default to True.
Set SG_ENABLE_PROMPT_EMBEDS=0 to skip the patch and restore vLLM's default.
Warning
Under almost no circumstances should you need to import this module directly. If stainedglass_output_protection is installed, vLLM
loads it automatically via the vllm.general_plugins entry point.
Functions:
| Name | Description |
|---|---|
enable_prompt_embeds_by_default |
Default vLLM's |
Attributes:
| Name | Type | Description |
|---|---|---|
SG_ENABLE_PROMPT_EMBEDS_ENV_VAR |
Final[str]
|
Environment variable that disables our default flip when set to a falsy value. |
SG_ENABLE_PROMPT_EMBEDS_ENV_VAR
module-attribute
¶
Environment variable that disables our default flip when set to a falsy value.
enable_prompt_embeds_by_default
¶
Default vLLM's --enable-prompt-embeds flag to True.
Mutates the ModelConfig.enable_prompt_embeds dataclass field default.
Must run before the serve parser is built.
No-op when SG_ENABLE_PROMPT_EMBEDS is falsy.
Raises:
| Type | Description |
|---|---|
RuntimeError
|
If |
RuntimeError
|
If the default did not take effect, which would leave the server silently rejecting |