Skip to content

fix(gateway): DeepSeek cache-hit tokens billing + disjoint input (rant 2026-08-23T08:20:38) #128

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

Merged
argszero merged 1 commit into from
Aug 23, 2026
Merged

fix(gateway): DeepSeek cache-hit tokens billing + disjoint input (rant 2026-08-23T08:20:38) #128

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
114 changes: 98 additions & 16 deletions src/sse.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,6 +162,10 @@ struct StreamUsage {
cache_read_input_tokens: Option<u32>,
#[serde(default)]
cache_creation_input_tokens: Option<u32>,
// DeepSeek 原生顶层拼写（rant 2026-08-23T08:20:38 P0 缓存计费）：
// prompt_tokens = prompt_cache_hit_tokens + prompt_cache_miss_tokens
#[serde(default)]
prompt_cache_hit_tokens: u32,
}

#[derive(Debug, Deserialize)]
Expand Down Expand Up @@ -254,11 +258,13 @@ pub fn openai_sse_to_anthropic<E: std::error::Error + Send + 'static>(
*pending_usage = Some(usage_json.clone());
}
// 计量：openai 原样 usage → (prompt, cached, completion)
// input 做 disjoint：prompt_tokens 含缓存命中部分，扣除后避免重复计费（rant 2026-08-23T08:20:38）
if let Some(u) = &chunk.usage {
let cached = extract_cache_read_tokens(u).unwrap_or(0);
record_usage(
&usage,
u.prompt_tokens as f64,
extract_cache_read_tokens(u).unwrap_or(0) as f64,
u.prompt_tokens.saturating_sub(cached) as f64,
cached as f64,
u.completion_tokens as f64,
);
}
Expand Down Expand Up @@ -575,14 +581,19 @@ fn build_anthropic_usage_json(usage: &StreamUsage) -> Value {
}

fn extract_cache_read_tokens(usage: &StreamUsage) -> Option<u32> {
if let Some(v) = usage.cache_read_input_tokens {
return Some(v);
// 优先级（rant 2026-08-23T08:20:38）：DeepSeek 原生顶层 → OpenAI → Anthropic
if usage.prompt_cache_hit_tokens > 0 {
return Some(usage.prompt_cache_hit_tokens);
}
usage
if let Some(v) = usage
.prompt_tokens_details
.as_ref()
.map(|d| d.cached_tokens)
.filter(|&v| v > 0)
{
return Some(v);
}
usage.cache_read_input_tokens
}

fn map_stop_reason(finish_reason: Option<&str>) -> Option<String> {
Expand Down Expand Up @@ -651,14 +662,17 @@ pub fn openai_sse_to_openai_responses<E: std::error::Error + Send + 'static>(
Err(_) => continue,
};
// 计量：openai usage 字段 → (prompt, cached, completion)
// 统一走 StreamUsage 提取（三拼写兼容）+ input disjoint（rant 2026-08-23T08:20:38）
if let Some(u) = v.get("usage") {
let input = u.get("prompt_tokens").and_then(|x| x.as_f64()).unwrap_or(0.0);
let cached = u
.pointer("/prompt_tokens_details/cached_tokens")
.and_then(|x| x.as_f64())
.unwrap_or(0.0);
let output = u.get("completion_tokens").and_then(|x| x.as_f64()).unwrap_or(0.0);
record_usage(&usage, input, cached, output);
if let Ok(su) = serde_json::from_value::<StreamUsage>(u.clone()) {
let cached = extract_cache_read_tokens(&su).unwrap_or(0);
record_usage(
&usage,
su.prompt_tokens.saturating_sub(cached) as f64,
cached as f64,
su.completion_tokens as f64,
);
}
}

let id = v.get("id").and_then(|i| i.as_str()).unwrap_or("").to_string();
Expand Down Expand Up @@ -956,7 +970,7 @@ pub fn anthropic_sse_to_openai<E: std::error::Error + Send + 'static>(
output_tokens = u.get("output_tokens").and_then(|x| x.as_f64()).unwrap_or(0.0);
cached_tokens = u.get("cache_read_input_tokens").and_then(|x| x.as_f64()).unwrap_or(cached_tokens);
}
record_usage(&usage, input_tokens, cached_tokens, output_tokens);
record_usage(&usage, (input_tokens - cached_tokens).max(0.0), cached_tokens, output_tokens);
let stop_reason = v.get("delta").and_then(|d| d.get("stop_reason")).and_then(|s| s.as_str());
let finish = match stop_reason {
Some("max_tokens") => "length",
Expand All @@ -980,7 +994,7 @@ pub fn anthropic_sse_to_openai<E: std::error::Error + Send + 'static>(
}
"message_stop" => {
if !finished {
record_usage(&usage, input_tokens, cached_tokens, output_tokens);
record_usage(&usage, (input_tokens - cached_tokens).max(0.0), cached_tokens, output_tokens);
let data = json!({
"id": format!("chatcmpl-{msg_id}"),
"object": "chat.completion.chunk",
Expand Down Expand Up @@ -1016,7 +1030,7 @@ pub fn anthropic_sse_to_openai<E: std::error::Error + Send + 'static>(
}
// 上游无 message_stop 直接断流 → 补 finish + [DONE]
if !finished {
record_usage(&usage, input_tokens, cached_tokens, output_tokens);
record_usage(&usage, (input_tokens - cached_tokens).max(0.0), cached_tokens, output_tokens);
let data = json!({
"id": format!("chatcmpl-{msg_id}"),
"object": "chat.completion.chunk",
Expand Down Expand Up @@ -1184,7 +1198,8 @@ pub fn responses_sse_to_openai_chat<E: std::error::Error + Send + 'static>(
.and_then(|x| x.as_f64())
.unwrap_or(0.0);
let output = u.get("output_tokens").and_then(|x| x.as_f64()).unwrap_or(0.0);
record_usage(&usage, input, cached, output);
// input disjoint：input_tokens 含缓存命中部分（rant 2026-08-23T08:20:38）
record_usage(&usage, (input - cached).max(0.0), cached, output);
}
if !finished {
let data = json!({
Expand Down Expand Up @@ -1435,4 +1450,71 @@ mod tests {
};
assert_eq!((i, o), (3.0, 0.0), "usage recorded");
}

// ── 缓存计费三拼写 + disjoint（rant 2026-08-23T08:20:38 P0）──

#[test]
fn deepseek_cache_hit_tokens_spelling() {
// DeepSeek 原生顶层 prompt_cache_hit_tokens：prompt_tokens=100 含 90 命中
// → input disjoint 为 10、cached=90、output=50；下游 input_tokens 同样不含命中
let slot = usage_slot();
let chunks = sse_chunks(vec![
"data: {\"id\":\"c1\",\"model\":\"m1\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"x\"},\"finish_reason\":null}]}\n\n".to_string(),
"data: {\"id\":\"c1\",\"model\":\"m1\",\"choices\":[{\"index\":0,\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":100,\"prompt_cache_hit_tokens\":90,\"completion_tokens\":50}}\n\n".to_string(),
"data: [DONE]\n\n".to_string(),
]);
let out = collect(openai_sse_to_anthropic(chunks, slot.clone()));
let Some((i, c, o)) = *slot.lock().unwrap() else {
panic!("usage missing");
};
assert_eq!((i, c, o), (10.0, 90.0, 50.0), "deepseek disjoint usage");
assert!(
out.contains("\"input_tokens\":10"),
"downstream input_tokens disjoint: {out}"
);
assert!(
out.contains("\"cache_read_input_tokens\":90"),
"downstream cache_read forwarded: {out}"
);
}

#[test]
fn openai_cached_tokens_spelling_disjoint() {
// OpenAI 拼写 prompt_tokens_details.cached_tokens
let slot = usage_slot();
let chunks = sse_chunks(vec![
"data: {\"id\":\"c1\",\"model\":\"m1\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"x\"},\"finish_reason\":null}]}\n\n".to_string(),
"data: {\"id\":\"c1\",\"model\":\"m1\",\"choices\":[{\"index\":0,\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":100,\"prompt_tokens_details\":{\"cached_tokens\":90},\"completion_tokens\":50}}\n\n".to_string(),
"data: [DONE]\n\n".to_string(),
]);
let out = collect(openai_sse_to_anthropic(chunks, slot.clone()));
let Some((i, c, o)) = *slot.lock().unwrap() else {
panic!("usage missing");
};
assert_eq!((i, c, o), (10.0, 90.0, 50.0), "openai disjoint usage");
assert!(
out.contains("\"cache_read_input_tokens\":90"),
"downstream cache_read forwarded: {out}"
);
}

#[test]
fn anthropic_cache_read_spelling_disjoint() {
// Anthropic 拼写 cache_read_input_tokens（message_start 计 input，message_delta 计 output）
let slot = usage_slot();
let chunks = sse_chunks(vec![
"event: message_start\ndata: {\"type\":\"message_start\",\"message\":{\"id\":\"m1\",\"model\":\"claude-x\",\"usage\":{\"input_tokens\":100,\"cache_read_input_tokens\":90,\"output_tokens\":0}}}\n\n".to_string(),
"event: content_block_start\ndata: {\"type\":\"content_block_start\",\"index\":0,\"content_block\":{\"type\":\"text\",\"text\":\"\"}}\n\n".to_string(),
"event: content_block_delta\ndata: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"text_delta\",\"text\":\"Hi\"}}\n\n".to_string(),
"event: content_block_stop\ndata: {\"type\":\"content_block_stop\",\"index\":0}\n\n".to_string(),
"event: message_delta\ndata: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":\"end_turn\"},\"usage\":{\"output_tokens\":50}}\n\n".to_string(),
"event: message_stop\ndata: {\"type\":\"message_stop\"}\n\n".to_string(),
]);
let out = collect(anthropic_sse_to_openai(chunks, slot.clone()));
assert!(out.contains("data: [DONE]"), "done: {out}");
let Some((i, c, o)) = *slot.lock().unwrap() else {
panic!("usage missing");
};
assert_eq!((i, c, o), (10.0, 90.0, 50.0), "anthropic disjoint usage");
}
}
Loading