minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[AMD] Fall back to CPU tensor for decode retraction on ROCm (#36343)

Browse Source
This commit is contained in:
Bingxu Chen
2026-08-26 18:32:51 -07:00
committed by GitHub
parent 20621aa14b
commit 862f9ed3d8
1 changed files with 4 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/mem_cache/kv_cache_builder.py
+4
View File
@@ -47,6 +47,7 @@ from sglang.srt.runtime_context import (
get_parallel, get_parallel,
get_schedule, get_schedule,
) )
from sglang.srt.utils import is_hip
if TYPE_CHECKING: if TYPE_CHECKING:
@@ -143,6 +144,9 @@ def resolve_decode_retraction_backup(*, tp_worker: BaseTpWorker) -> str:
backend = ( backend = (
"host_pool" "host_pool"
if disagg.disaggregation_mode == "decode" if disagg.disaggregation_mode == "decode"
# Large ROCm retraction restores can fault the GPU process. Keep
# host_pool opt-in on HIP until the retraction path is safe at scale.
and not is_hip()
and not get_parallel().dcp_enabled and not get_parallel().dcp_enabled
and not disagg.disaggregation_decode_enable_radix_cache and not disagg.disaggregation_decode_enable_radix_cache
# KV offload already owns a host pool; a second one double-books host memory. # KV offload already owns a host pool; a second one double-books host memory.