A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Misc] Correct an outdated comment. (#24765)
Signed-off-by: Russell Bryant <rbryant@redhat.com>
Russell Bryant committed
Sep 13, 2025 at 03:34 UTC
4dad72f0d903933429e96227894f98a2a6e30fa7
1 file changed
+4
-4
vllm/v1/engine/processor.py
+4
-4
@@ -260,10 +260,10 @@ class Processor:
260
else:
261
# NOTE: engine_level_backend must be "auto" here, because we have
262
# checked supported_backends above.
263
- # "auto" is an opt-in to opinionated behavior where we try to
264
- # choose a backend based on request contents. This is not the
265
- # default as it is less predictable and subject to change
266
- # between releases as feature support changes.
263
+ # In this mode, we set opinionated defaults based on what we think
264
+ # will satisfy the most use cases without having to worry about
265
+ # this setting. We include fallback behavior here, but not with any
266
+ # other setting where a specific backend was specified.
267
try:
268
validate_xgrammar_grammar(params)
269
params.guided_decoding.backend = "xgrammar"