@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 7888e1d0a3

Fork 0

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 15, 19:45 UTC

[V1] TPU - Enable prefix caching by default (#14773)

Alexander Matveev committed Mar 13, 2025 at 23:40 UTC 7888e1d0a3eb83becda81bbad5e9848dbb598453

1 file changed -6

vllm/platforms/tpu.py

-6

@@ -108,12 +108,6 @@ class TpuPlatform(Platform):

108 parallel_config.worker_cls = \

109 "vllm.worker.tpu_worker.TPUWorker"

110

111 - # Adjust scheduler config for V1

112 - # TODO: Add support for these

113 - if envs.VLLM_USE_V1 and vllm_config.cache_config.enable_prefix_caching:

114 - logger.warning("[V1][TPU] Disable prefix caching")

115 - vllm_config.cache_config.enable_prefix_caching = False

116 -

111 assert not vllm_config.speculative_config, (

112 "Speculative decoding is not yet supported for TPU backend")

113