fix: deterministic prefix matching and correct Llama 3.x context windows
Two fixes in openaiContextWindows.ts: 1. Sort lookup keys by length descending in lookupByModel() so the most specific prefix always wins. Without this, 'gpt-4-turbo-preview' could match 'gpt-4' (8k) instead of 'gpt-4-turbo' (128k) depending on V8's object key iteration order. 2. Update Llama 3.1/3.2/3.3 context windows from 8,192 to 128,000. These models support 128k context natively (Meta official specs). The previous 8k value was Ollama's default num_ctx, not the model's actual capability, causing premature auto-compact warnings.
This commit is contained in:
@@ -50,9 +50,11 @@ const OPENAI_CONTEXT_WINDOWS: Record<string, number> = {
|
||||
'gemini-2.5-flash': 1_048_576,
|
||||
|
||||
// Ollama local models
|
||||
'llama3.3:70b': 8_192,
|
||||
'llama3.1:8b': 8_192,
|
||||
'llama3.2:3b': 8_192,
|
||||
// Llama 3.1+ models support 128k context natively (Meta official specs).
|
||||
// Ollama defaults to num_ctx=8192 but users can configure higher values.
|
||||
'llama3.3:70b': 128_000,
|
||||
'llama3.1:8b': 128_000,
|
||||
'llama3.2:3b': 128_000,
|
||||
'qwen2.5-coder:32b': 32_768,
|
||||
'qwen2.5-coder:7b': 32_768,
|
||||
'deepseek-coder-v2:16b': 163_840,
|
||||
@@ -122,7 +124,11 @@ const OPENAI_MAX_OUTPUT_TOKENS: Record<string, number> = {
|
||||
|
||||
function lookupByModel<T>(table: Record<string, T>, model: string): T | undefined {
|
||||
if (table[model] !== undefined) return table[model]
|
||||
for (const key of Object.keys(table)) {
|
||||
// Sort keys by length descending so the most specific prefix wins.
|
||||
// Without this, 'gpt-4-turbo-preview' could match 'gpt-4' (8k) instead
|
||||
// of 'gpt-4-turbo' (128k) depending on V8's key iteration order.
|
||||
const sortedKeys = Object.keys(table).sort((a, b) => b.length - a.length)
|
||||
for (const key of sortedKeys) {
|
||||
if (model.startsWith(key)) return table[key]
|
||||
}
|
||||
return undefined
|
||||
|
||||
Reference in New Issue
Block a user