[NA] [BE] fix: register fireworks_ai as a canonical provider so Fireworks model prices load

[Anuj7411]

Details

Fireworks AI's pricing entries are silently dropped by Opik today. model_prices_and_context_window.json carries roughly 260 non-zero-cost entries tagged litellm_provider: "fireworks_ai" (the accounts/fireworks/models/* catalog spanning the llama, qwen, deepseek, glm, kimi and minimax families), and none of them load because fireworks_ai is not registered in PROVIDERS_MAPPING; buildModelPrice returns null for any entry whose provider is not in that map, so every Fireworks span ends up with no cost. This follows the exact pattern of the earlier provider-registration fixes for azure (#7262), xai (#7433), deepseek (#7434) and perplexity (#7432).

• Register fireworks_ai -> fireworks_ai in PROVIDERS_MAPPING so the entries load
• Route fireworks_ai through textGenerationWithCacheCostOpenAI in PROVIDERS_CACHE_COST_CALCULATOR; Fireworks' LiteLLM cost calculator (litellm/llms/fireworks_ai/cost_calculator.py) delegates to generic_cost_per_token, so its usage payload uses the OpenAI shape (cached tokens under prompt_tokens_details.cached_tokens), and the eight cache-priced Fireworks models today (glm-4p7, glm-5p1, kimi-k2p5, minimax-m2p1) get the same cache-read discount OpenAI/Azure/xAI/DeepSeek receive
• PROVIDERS_CACHE_COST_CALCULATOR was at the ten-pair ceiling of Map.of, so it moves to Map.ofEntries to take the eleventh entry

Change checklist

• User facing
• Documentation update

Issues

No ticket; discovered by scanning litellm_provider values in the bundled price map for providers with priced models that are absent from PROVIDERS_MAPPING.

AI-WATERMARK

AI-WATERMARK: yes

• Tools: Claude Code
• Model(s): Claude Opus 4.8
• Scope: investigation of the drop path, the two-line mapping change, and the parameterized test
• Human verification: model resolution and both cost figures checked against the bundled model_prices_and_context_window.json before submission

Testing

Added a parameterized test (calculateCostHandlesFireworksModels) covering both branches:

• fireworks_ai/accounts/fireworks/models/llama-v3p1-70b-instruct (input 9e-7, output 9e-7, no cache) routes through textGenerationCost; 1000 prompt + 200 completion tokens = 0.00108
• fireworks_ai/accounts/fireworks/models/kimi-k2p5 (input 6e-7, output 3e-6, cache_read 1e-7) routes through textGenerationWithCacheCostOpenAI; with 300 of 1000 prompt tokens cached, 700 * 6e-7 + 200 * 3e-6 + 300 * 1e-7 = 0.00105

Both cases fail against the current code (the models do not resolve without the mapping entry, so cost comes back as zero) and pass with the fix.

Documentation

No documentation changes needed; provider pricing is driven entirely by the bundled price map and the provider registration in CostService.

Follow-up

The same drop affects several other high-volume OpenAI-compatible inference providers that could be registered the same way if the team wants them: openrouter (~92 priced models), novita (~85), deepinfra (~67), together_ai (~36), databricks (~28) and moonshot (~22). Happy to open follow-ups for any of those; keeping this PR scoped to Fireworks alone.

[andrescrz]

Hi @Anuj7411

Thanks for the contribution! I reviewed the PR and everything looks good from my side. I've approved it, and all CI checks have passed successfully.

We'll merge it shortly, and the change will be included in the next Opik release.

Thanks again for contributing to Opik!