biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Simplify (and fix) passing of guided decoding backend options (#17008)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-04-29 20:02:23 +01:00
committed by GitHub
parent 2fa2a50bf9
commit a6977dbd15
17 changed files with 309 additions and 217 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/v1/engine/processor.py
View File

@@ -144,7 +144,7 @@ class Processor:
if not params.guided_decoding or not self.decoding_config:
return
engine_level_backend = self.decoding_config.guided_decoding_backend
engine_level_backend = self.decoding_config.backend
if params.guided_decoding.backend:
# Request-level backend selection is not supported in V1.
# The values may differ if `params` is reused and was set
@@ -152,8 +152,8 @@ class Processor:
# request. We remember that it was set as a result of `auto`
# using the `_auto` option set on the backend in the params.
if (params.guided_decoding.backend != engine_level_backend
and not (engine_level_backend == "auto" and "_auto"
in params.guided_decoding.backend_options())):
and not (engine_level_backend == "auto"
and params.guided_decoding.backend_was_auto)):
raise ValueError(
"Request-level structured output backend selection is no "
"longer supported. The request specified "
@@ -189,7 +189,7 @@ class Processor:
# are not supported in xgrammar. Fall back to guidance.
params.guided_decoding.backend = "guidance"
# Remember that this backend was set automatically
params.guided_decoding.add_option("_auto")
params.guided_decoding.backend_was_auto = True
def process_inputs(
self,