opensource/mateclaw
2
0
Fork 0
mirror of https://gitee.com/mateos/mateclaw.git synced 2026-09-13 03:13:41 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

fix(llm): pre-filter non-chat DashScope modalities + quieter probe logging

Browse Source
This commit is contained in:
matevip 2026-04-16 18:15:48 +08:00
parent adfe23cd2c
commit c4883e108c
1 changed files with 62 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

74
mateclaw-server/src/main/java/vip/mate/llm/service/ModelDiscoveryService.java
View File

@ -36,8 +36,9 @@ public class ModelDiscoveryService {
// Virtual-thread executor for parallel model probing (lightweight, short-lived)
private static final ExecutorService PROBE_EXECUTOR = Executors.newVirtualThreadPerTaskExecutor();
// Probe concurrency cap — avoid flooding the provider with concurrent ping requests
private static final int MAX_PROBE_CONCURRENCY = 5;
// Probe concurrency cap — reduced from 5 to 3 because higher parallelism
// triggered 429 Throttling.RateQuota on DashScope during bulk refresh
private static final int MAX_PROBE_CONCURRENCY = 3;
// Per-model probe timeout (short; we only need to know "yes/no usable")
private static final long PROBE_TIMEOUT_SECONDS = 12;
@ -60,19 +61,57 @@ public class ModelDiscoveryService {
private static final java.util.regex.Pattern DASHSCOPE_NATIVE_UNSUPPORTED_PATTERN =
java.util.regex.Pattern.compile("^qwen\\d+\\.\\d+.*", java.util.regex.Pattern.CASE_INSENSITIVE);
/**
* Parameter-size suffixes used by DashScope open-source base models (e.g.
* {@code qwen3-0.6b}, {@code qwen3-8b}, {@code qwen3-32b}, {@code qwen3-30b-a3b}).
* These are catalog entries, not DashScope-hosted chat endpoints, and the native
* protocol returns {@code InvalidParameter: parameter.enable_thinking ...} for
* any attempt to invoke them. Pre-filter them out of discovery.
*/
private static final java.util.regex.Pattern DASHSCOPE_OPEN_SOURCE_SIZE_PATTERN =
java.util.regex.Pattern.compile("^qwen\\d+-\\d+(?:\\.\\d+)?b(?:-.*)?$", java.util.regex.Pattern.CASE_INSENSITIVE);
/**
* Prefixes that identify non-chat modalities (image generation, vision understanding,
* TTS, ASR, omni/multimodal bases, realtime speech, live translation, OCR, speech-to-
* speech, voice-clone). They are catalog-visible but have different endpoints than
* the native chat-generation API, so probing them via chat always fails with
* "url error". Blocking them up-front cuts discovery time and log noise dramatically.
*/
private static final Set<String> DASHSCOPE_NON_CHAT_PREFIXES = Set.of(
// Vision understanding / OCR
"qwen-vl-",
"qwen3-vl-",
// Image generation / edit
"qwen-image-",
"qwen3-image-",
// TTS / ASR / speech-to-speech / voice
"qwen-tts-",
"qwen3-tts-",
"qwen-asr-",
"qwen3-asr-",
"qwen-s2s-",
"qwen3-s2s-",
// Omni multimodal bases
"qwen-omni-",
"qwen3-omni-",
// Audio understanding
"qwen-audio-",
// Live translation
"qwen-livetranslate-",
"qwen3-livetranslate-"
);
/**
* Allow-list prefixes for DashScope models that are known to work on the native
* protocol. An empty set means "no prefix filter" (we still apply DENY).
* Extend conservatively as we verify additional families.
* chat protocol. An empty set means "no prefix filter" (we still apply DENY).
* Extend conservatively as new families are verified.
*/
private static final Set<String> DASHSCOPE_NATIVE_ALLOW_PREFIXES = Set.of(
"qwen-", // qwen-max / qwen-plus / qwen-turbo / qwen-coder-* / qwen-long
"qwen2-", // qwen2 series
"qwen3-", // qwen3-max / qwen3-plus / qwen3-coder / qwen3-235b-*
"qwen-vl-", // vision-language
"qwen-audio-",
"qwen-omni-",
"deepseek-", // deepseek-v3.x
"deepseek-", // deepseek-v3.x / deepseek-r1*
"baichuan",
"yi-",
"llama"
@ -139,15 +178,21 @@ public class ModelDiscoveryService {
}
/**
* Return true if a DashScope model id is allowed on the native protocol:
* not in the explicit DENY set, doesn't match the dot-version unsupported
* pattern, and matches at least one ALLOW prefix (or the allow list is empty).
* Return true if a DashScope model id is allowed on the native chat protocol.
* Rejection rules (in order):
* 1. Explicit DENY set (e.g. qwen3.5-max)
* 2. Dot-version family pattern (qwen3.5-*, qwen3.6-*, ...)
* 3. Non-chat modality prefix (vl, image, tts, asr, omni, audio, s2s, ocr, livetranslate)
* 4. Open-source parameter-size suffix (qwen3-8b, qwen3-32b, qwen3-30b-a3b, qwen3-0.6b ...)
* 5. Must start with a known ALLOW prefix (qwen-, qwen2-, qwen3-, deepseek-, ...)
*/
private static boolean isDashScopeModelIdAcceptable(String modelId) {
if (modelId == null || modelId.isBlank()) return false;
String lower = modelId.toLowerCase();
if (DASHSCOPE_NATIVE_DENY.contains(lower)) return false;
if (DASHSCOPE_NATIVE_UNSUPPORTED_PATTERN.matcher(lower).matches()) return false;
if (DASHSCOPE_NON_CHAT_PREFIXES.stream().anyMatch(lower::startsWith)) return false;
if (DASHSCOPE_OPEN_SOURCE_SIZE_PATTERN.matcher(lower).matches()) return false;
if (DASHSCOPE_NATIVE_ALLOW_PREFIXES.isEmpty()) return true;
return DASHSCOPE_NATIVE_ALLOW_PREFIXES.stream().anyMatch(lower::startsWith);
}
@ -196,7 +241,12 @@ public class ModelDiscoveryService {
} catch (Exception e) {
dto.setProbeOk(false);
dto.setProbeError(shortError(e));
log.info("[ModelDiscovery] Probe failed for model={}: {}", dto.getId(), dto.getProbeError());
// DEBUG level — per-model probe failures are an expected part of bulk
// discovery (DashScope lists many deprecated/restricted models). The
// aggregate "Probe results: X passed, Y failed" summary below is
// sufficient for normal operations. Enable DEBUG for ModelDiscovery
// if you need to inspect individual reasons.
log.debug("[ModelDiscovery] Probe failed for model={}: {}", dto.getId(), dto.getProbeError());
} finally {
sem.release();
}