Houmeres/LLeMbas
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 3 Wiki Activity

Ask the endpoint what a streamed reply cost

Browse Source 
A streamed completion carries no token counts unless you ask for them, and
`stream_options: {include_usage: true}` is how. Not every server implements
it, and an unknown key is a 400 from some -- the same hazard as sending a
tools array to an endpoint without support. So it is asked for once per base
URL per process, and an endpoint that refuses is remembered and retried
without it. The retry is safe because the status is checked before a single
line is read: nothing has been yielded, so there is nothing to duplicate.

chunk_usage() reads the resulting chunk. It needed no change to the loop
above it: a usage chunk carries `choices: []`, which is exactly the shape
delta_text, delta_reasoning, delta_tool_calls and finish_reason have always
returned early on. All-zero counts are treated as absent, because some
servers attach zeros to every chunk and the real numbers only at the end.

services/tokens.py is the fallback for endpoints that never report: four
characters to a token, counting the tools array because thirteen schemas is
a meaningful slice of a short window, and counting nothing for an image
because its cost depends on tiling and an invented number would be worse
than the omission. Crude on purpose -- a real tokeniser means one per model
family, for a figure that is displayed beside a tilde.

Nothing uses any of this yet.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Jaroslav Beneš
2026-08-01 00:36:12 +02:00
parent 2fe736aa6a
commit 9e2caeac48
3 changed files with 290 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/lembas/services/llm/openai_client.py
+76
View File
@@ -190,6 +190,12 @@ def context_from(entry: dict[str, Any]) -> int:
return 0
# Endpoints that rejected `stream_options`, so it is asked for once per base URL
# per process and then never again. Not persisted: it is a property of whatever
# is running there now, and a restart is the right time to find out afresh.
_NO_STREAM_OPTIONS: set[str] = set()
async def stream_chat(
endpoint: Endpoint,
payload: dict[str, Any],
@@ -198,8 +204,41 @@ async def stream_chat(
Yields the raw upstream chunks; interpreting them is the caller's job. The
terminating "[DONE]" sentinel is consumed here and not yielded.
`stream_options` asks for the final usage chunk, which is the only way to
learn what a streamed reply actually cost. Not every server implements it,
and an unknown key is a 400 from some of them -- the same hazard as sending
a `tools` array to an endpoint without support. So it is asked for once,
and an endpoint that refuses is remembered and never asked again. Retrying
is safe because the status is checked before a single line is read: nothing
has been yielded, so there is nothing to duplicate.
"""
wants_usage = endpoint.base_url not in _NO_STREAM_OPTIONS
try:
async for chunk in _stream_once(endpoint, payload, usage=wants_usage):
yield chunk
except LLMError as exc:
if not wants_usage or exc.status_code not in (400, 422):
raise
_NO_STREAM_OPTIONS.add(endpoint.base_url)
log.info(
"%s rejected stream_options; token counts will be estimated there",
endpoint.base_url,
)
async for chunk in _stream_once(endpoint, payload, usage=False):
yield chunk
async def _stream_once(
endpoint: Endpoint,
payload: dict[str, Any],
*,
usage: bool,
) -> AsyncIterator[dict[str, Any]]:
body = {**payload, "stream": True}
if usage:
body["stream_options"] = {"include_usage": True}
try:
async with (
@@ -318,6 +357,43 @@ def finish_reason(chunk: dict[str, Any]) -> str:
return ""
def chunk_usage(chunk: dict[str, Any]) -> dict[str, int] | None:
"""Token counts from a usage chunk, or None if this is not one.
A usage chunk carries `choices: []`, which is exactly the shape delta_text,
delta_reasoning, delta_tool_calls and finish_reason all return early on --
they have always tolerated it, so nothing else needs to change to let one
through.
Fields are read defensively because "the endpoint returned something odd"
must never be the reason a reply fails; a bad shape simply means no counts.
"""
try:
raw = chunk.get("usage")
except AttributeError:
return None
if not isinstance(raw, dict):
return None
counts: dict[str, int] = {}
for key in ("prompt_tokens", "completion_tokens", "total_tokens"):
value = raw.get(key)
if isinstance(value, bool) or not isinstance(value, (int, float)):
continue
if value >= 0:
counts[key] = int(value)
# Some servers send a usage object of zeros on every chunk and the real
# numbers only at the end. All-zero is indistinguishable from that, and
# treating it as an answer would freeze the count at nothing.
if not counts or not any(counts.values()):
return None
counts.setdefault(
"total_tokens", counts.get("prompt_tokens", 0) + counts.get("completion_tokens", 0)
)
return counts
def delta_text(chunk: dict[str, Any]) -> str:
"""Pull the text out of one streamed chunk, tolerating provider variation."""
try:
src/lembas/services/tokens.py
+75
View File
@@ -0,0 +1,75 @@
"""A rough token count, for when the endpoint does not give a real one.
Deliberately crude. Counting tokens properly means the model's own tokeniser,
which means shipping one per model family and a dependency that has to be kept
in step with them -- for a number that is displayed beside a `~` and used to
decide when to summarise.
Four characters per token is the usual English approximation. It is optimistic
on code and badly wrong on CJK, which is why anything derived from it is marked
as an estimate everywhere it is shown.
"""
from __future__ import annotations
from typing import Any
CHARS_PER_TOKEN = 4
# What a message costs beyond its text: the role, the delimiters a chat template
# wraps each turn in. Small, but a hundred-turn conversation is a hundred of them.
PER_MESSAGE_OVERHEAD = 4
def estimate(text: str) -> int:
"""Roughly how many tokens a piece of text is."""
if not text:
return 0
return max(1, round(len(text) / CHARS_PER_TOKEN))
def estimate_content(content: Any) -> int:
"""A message's content, whether it is a plain string or typed parts.
An image part contributes nothing: its cost depends on the model's tiling,
and a number invented here would be worse than the omission.
"""
if isinstance(content, str):
return estimate(content)
if isinstance(content, list):
return sum(
estimate(part.get("text", ""))
for part in content
if isinstance(part, dict) and part.get("type") == "text"
)
return 0
def estimate_messages(messages: list[dict[str, Any]]) -> int:
total = 0
for message in messages:
if not isinstance(message, dict):
continue
total += PER_MESSAGE_OVERHEAD + estimate_content(message.get("content"))
# A tool result the model reads back is part of the window too.
for call in message.get("tool_calls") or []:
function = (call or {}).get("function") or {}
total += estimate(str(function.get("name", "")))
total += estimate(str(function.get("arguments", "")))
return total
def estimate_request(payload: dict[str, Any]) -> int:
"""What a whole request body costs, tools included.
The tools array is sent on every request when tools are offered and is not
small -- thirteen schemas is a meaningful slice of a short window, and
leaving it out would make the percentage read low exactly when it matters.
"""
total = estimate_messages(payload.get("messages") or [])
for tool in payload.get("tools") or []:
function = (tool or {}).get("function") or {}
total += estimate(str(function.get("name", "")))
total += estimate(str(function.get("description", "")))
total += estimate(str(function.get("parameters", "")))
return total