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 11, 08:45 UTC
[Mamba] Removed disable cascade attn in MambaModelConfig (#30712)
Signed-off-by: asafg <39553475+Josephasafg@users.noreply.github.com>
Asaf Joseph Gardin committed
Dec 17, 2025 at 10:48 UTC
a9e15c21efbbc5b4a7a1e69e40378fdfe1acdcb7
1 file changed
-6
vllm/model_executor/models/config.py
-6
@@ -308,12 +308,6 @@ class MambaModelConfig(VerifyAndUpdateConfig):
308
if cache_config.mamba_block_size is None:
309
cache_config.mamba_block_size = model_config.max_model_len
310
311
- # TODO(tdoublep): remove once cascade attention is supported
312
- logger.info(
313
- "Disabling cascade attention since it is not supported for hybrid models."
314
- )
315
- model_config.disable_cascade_attn = True
316
-
311
312
class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
313
@classmethod