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 6, 21:45 UTC
Disable cascade attention by default (#36318)
Michael Goin committed
Mar 11, 2026 at 11:12 UTC
9c34e9d24fcd72834daf8b54f52667e3fa009d5f
1 file changed
+5
-4
vllm/config/model.py
+5
-4
@@ -217,12 +217,13 @@ class ModelConfig:
217
"""Whether to disable sliding window. If True, we will disable the sliding
218
window functionality of the model, capping to sliding window size. If the
219
model does not support sliding window, this argument is ignored."""
220
- disable_cascade_attn: bool = False
220
+ disable_cascade_attn: bool = True
221
"""Disable cascade attention for V1. While cascade attention does not
222
change the mathematical correctness, disabling it could be useful for
223
- preventing potential numerical issues. Note that even if this is set to
224
- False, cascade attention will be only used when the heuristic tells that
225
- it's beneficial."""
223
+ preventing potential numerical issues. This defaults to True, so users
224
+ must opt in to cascade attention by setting this to False. Even when this
225
+ is set to False, cascade attention will only be used when the heuristic
226
+ tells that it's beneficial."""
227
skip_tokenizer_init: bool = False
228
"""Skip initialization of tokenizer and detokenizer. Expects valid
229
`prompt_token_ids` and `None` for prompt from the input. The generated