A runner and the objects it builds freeze the placement they describe (#40341)

This commit is contained in:
Cheng Wan
2026-09-21 12:24:17 -07:00
committed by GitHub
parent 2d0e94e3a3
commit 65be3fa71a
13 changed files with 289 additions and 110 deletions
@@ -91,7 +91,8 @@ def _make_prefill_aware_swa_runner(
kv_cache_dtype=torch.float16,
kv_cache_dtype_str="auto",
page_size=1,
ps=SimpleNamespace(attn_cp_size=1, tp_size=1),
attn_cp_size=1,
tp_size=1,
is_draft_worker=False,
server_args=server_args,
attention_chunk_size=None,