Bench qwen selfhost#25
Merged
Merged
Conversation
The matrix kept failing whole runs on the first transient fault: Mistral's tokens-per-minute 429s (an agentic burst of large-context calls trips the rate limit even on a paid tier — it's a rate, not a quota, so adding credits doesn't help) and a self-hosted LM Studio endpoint closing the socket mid-response on a heavy request. Both are recoverable; _post_chat just gave up. _post_chat now retries 429/5xx and transport errors with exponential backoff (1,2,4,8,16s, capped at 60s), honoring a Retry-After header when present. After MAX_HTTP_RETRIES the last error still propagates, so a persistent outage is recorded as a real auth/infra failure — never masked as "found nothing". sleep is injectable for tests; non-retryable client errors (e.g. 400) are not retried. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…rix Halo A slow self-hosted endpoint (Qwen 3.6 27B on a Strix Halo APU) needs more than the 600s/call default: one big-context reasoning turn over freerdp's 1821-line planar.c exceeded it and failed the run as infra_error (a socket-read TimeoutError, not a crash — the APU's extra memory fixed the OOM the desktop GPUs hit). _http_timeout() now reads NELSON_HTTP_TIMEOUT (default 600); RawApiLoopRuntime threads it from the competitor's cost_model "http_timeout". Qwen's competitor is repointed from desktop LM Studio (qwen/qwen3.6-27b @ 192.168.122.1:1234) to the Strix Halo llama-server (unsloth/Qwen3.6-27B-GGUF:UD-Q8_K_XL @ 10.20.30.2:8080) with http_timeout 1800, its stable home after the desktop OOM-crashes. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Make Qwen work, add retry with backoff for models that fail for seeming infra reasons (rate limits, timeout, etc.).