Commit 4d09df0
committed
fix: HttpRangeReader stall detection, CDN re-resolve, HF API resolve
Root cause: 21 GB tensor reads stall silently when HuggingFace CDN
drops the TCP connection. curl sits forever with no data.
Fixes:
- --speed-limit 100000 --speed-time 30: abort if < 100 KB/s for 30s
- Re-resolve URL on 403 (CDN token expiry after ~1 hour)
- Segment-aligned fetches (no overlapping reads on sequential access)
- 6 retries (was 4) with capped exponential backoff (max 32s)
- from_hf() constructor: resolves via huggingface_hub Python API first,
falls back to curl HEAD, then HF REST API. Stores repo/filename for
automatic re-resolution on token expiry.
- resolve_hf_url() tries 3 methods: Python HF API → curl HEAD → REST API
https://claude.ai/code/session_01HmdXNPit7QsTCfhJFef3Ee1 parent b00df16 commit 4d09df0
1 file changed
Lines changed: 227 additions & 71 deletions
0 commit comments