[TPU] Add TPU specific var VLLM_TPU_MOST_MODEL_LEN (#19919)

Signed-off-by: Chenyaaang <chenyangli@google.com>
This commit is contained in:
Chenyaaang
2025-06-25 15:51:02 -07:00
committed by GitHub
parent 55c65ab495
commit 2d7620c3eb
5 changed files with 184 additions and 76 deletions
-10
View File
@@ -122,16 +122,6 @@ class TpuPlatform(Platform):
PallasAttentionBackend)
cache_config.block_size = PallasAttentionBackend.get_page_size(
vllm_config) # type: ignore[assignment]
min_page_size = PallasAttentionBackend.get_min_page_size(
vllm_config)
if min_page_size > cache_config.block_size:
logger.warning(
"Increase the page size from %s to %s to make sure there's"
"no SMEM OOM",
cache_config.block_size,
min_page_size,
)
cache_config.block_size = min_page_size # type: ignore[assignment]
parallel_config = vllm_config.parallel_config
scheduler_config = vllm_config.scheduler_config