mirror of
https://github.com/BerriAI/litellm.git
synced 2026-09-14 23:21:35 +00:00
692226e334
The first version of this PR carried three files. Comparing it against #34752 (SCX.ai), the last JSON-configured provider merged here, showed it was missing most of what a provider needs: litellm/constants.py base URL + provider name litellm/types/utils.py LlmProviders.HUBRIS model_prices_and_context_window.json 14 models litellm/model_prices_and_context_window_backup.json litellm/provider_endpoints_support_backup.json litellm/proxy/public_endpoints/provider_create_fields.json ui/litellm-dashboard/public/assets/logos/hubris.svg Model entries deliberately carry no cost fields. Hubris bills in rubles at a rate that moves with the cost of buying dollars, so a hardcoded USD-per-token figure would be wrong within weeks. 118 chat entries already in this file are priced the same way, dashscope's whole catalogue among them. Context windows, output limits and capability flags come from the gateway's own catalogue. The endpoint support entry is corrected against the live API. Six of its twelve flags were wrong: embeddings, image_generations, audio_transcriptions and audio_speech were declared unsupported when the routes exist, and a2a and interactions were declared supported when they return 404. Every flag was re-checked by hitting the route: 401 means it exists and wants a key, 404 means it does not exist. The two overstated flags were the bad ones — LiteLLM would have promised users endpoints that are not there. Verified: check_provider_folders_documented.py passes (29 openai_like providers, 178 provider entries, ours among them), all touched JSON parses, LlmProviders.HUBRIS resolves. |
||
|---|---|---|
| .. | ||
| litellm-dashboard | ||
| Dockerfile | ||
| nginx.conf | ||