Skip to content

fix(scheduler): clamp admission to the actual KV pool so oversized prompts fail loudly #118

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ascorb12 wants to merge 1 commit into
base: main
Choose a base branch
from
+10 −3
Open

fix(scheduler): clamp admission to the actual KV pool so oversized prompts fail loudly #118

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 10 additions & 3 deletions python/freetoken/scheduler/scheduler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -487,10 +487,17 @@ def _process_one_msg(self, msg: BaseBackendMsg) -> None:
)
return
input_len, max_seq_len = len(msg.input_ids), self.engine.max_seq_len
max_output_len = max_seq_len - input_len
# max_seq_len is the model's advertised context, which can far exceed the
# KV pool actually allocated (see issue #111): a prompt that passes this
# check but can never be granted enough pages is queued forever with no
# error and no log line. Clamp admission to the real pool so oversized
# prompts fail loudly with the same error clients already understand.
pool_tokens = self.engine.num_pages * self.config.page_size
effective_max = min(max_seq_len, pool_tokens)
max_output_len = effective_max - input_len
if max_output_len <= 0:
logger.warning_rank0(
f"Input sequence length {input_len} exceeds {max_seq_len}, "
f"Input sequence length {input_len} exceeds {effective_max}, "
f"request {msg.uid} is dropped."
)
# Tell the client instead of dropping silently — otherwise its wait_for_ack
Expand All @@ -502,7 +509,7 @@ def _process_one_msg(self, msg: BaseBackendMsg) -> None:
# "prompt is too long: N tokens > M" is the phrasing Claude Code and
# OpenClaw match on; the Anthropic wire has no error code to read.
error=(
f"prompt is too long: {input_len} tokens > {max_seq_len} maximum "
f"prompt is too long: {input_len} tokens > {effective_max} maximum "
f"(prompt + generation); shorten the prompt or increase the KV "
f"cache budget"
),
Expand Down