[Fix] deepseek-v4.1: Reject encoder replay on a chunk too small for a hit or on other models (#45051)

This commit is contained in:
Liangsheng Yin
2026-10-18 17:55:35 -07:00
committed by GitHub
parent 3c560089e1
commit 59fec011dd
4 changed files with 26 additions and 2 deletions
@@ -238,15 +238,19 @@ def validate_deepseek_v41_encoder_replay_budget(server_args: ServerArgs) -> None
cfg = resolving_view(server_args)
if not cfg.enable_encoder_swa_bounded_replay:
return
# A hit's first chunk also runs up to 128 replay rows and floors its extend to
# whole pages, so a smaller chunk leaves the hit no page and it never admits.
min_chunk = 128 + (cfg.page_size or 1)
if (
cfg.max_running_requests is None
or cfg.max_running_requests <= 0
or not cfg.chunked_prefill_size
or cfg.chunked_prefill_size < 128
or cfg.chunked_prefill_size < min_chunk
):
raise ValueError(
"encoder SWA replay requires --max-running-requests > 0 and at least "
"128 chunked prefill tokens per attention DP rank"
f"{min_chunk} chunked prefill tokens per attention DP rank (128 replay "
"tokens plus one page)"
)
@@ -1026,6 +1026,8 @@ def handle_model_capability_adjustments(server_args: Any):
)
validate_deepseek_v41_encoder_replay_budget(server_args)
elif cfg.enable_encoder_swa_bounded_replay:
raise ValueError("--enable-encoder-swa-bounded-replay requires DeepSeek-V4.1")
def handle_mamba_radix_cache(server_args: Any, hf_config: Any):