Preserve OpenAI cached_tokens + Wan 2.7 R2V media types#10
Merged
Conversation
The Wan 2.7 R2V (`/v3/async/wan2.7-r2v`) endpoint requires each item
in the `media` array to carry a `type` value from the enum:
- `reference_image`
- `reference_video`
- `first_frame`
We were sending `image` and `video`, which Novita rejects with the
generic "failed to exec task" 500 — every R2V submission via the
playground / legacy `image_urls`+`video_urls` shape was failing
silently for that reason.
Two changes in `build_body`:
1. Repack each `image_urls[]` URL as `{type: "reference_image", url}`
and each `video_urls[]` URL as `{type: "reference_video", url}`.
No way to express `first_frame` or per-item `reference_voice`
from the legacy flat shape — callers who want those use the new
pass-through path below.
2. Pass `media` through the allowed-fields whitelist for the R2V
shape so direct API callers / a future media-editor UI can
submit the rich shape (`[{type, url, reference_voice?}, ...]`)
verbatim. The `!body.contains_key("media")` guard in the repack
block ensures the pass-through wins when both shapes are present.
Also cap the synthesised `media` array at 5 items to match Novita's
documented ceiling (combined images+videos ≤ 5), so users who upload
more get a deterministic truncate-from-front rather than a 422.
TensorZero normalized provider usage to {input_tokens, output_tokens}
and dropped OpenAI's prompt_tokens_details.cached_tokens. RouterBase
routes chat LLMs (incl. Claude/Gemini via Novita's OpenAI-compat
endpoint) through this gateway and needs the prompt-cache read count to
bill cache reads at the discounted rate and show users their savings.
- Add `cached_tokens: Option<u32>` to Usage (ts-bindings + skip-if-none).
- Parse prompt_tokens_details.cached_tokens in the OpenAI provider's
OpenAIUsage → Usage conversion.
- Thread through Usage::zero() and the streaming/cross-inference
aggregators (sum, treating None as 0). Anthropic/Bedrock/Vertex native
paths leave it None (out of scope; our Anthropic models use the
openai-compat path).
cargo check --package tensorzero-core (lib) clean; no new clippy warnings.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
I have read the Contributor License Agreement (CLA) and hereby sign the CLA. You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot. |
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.
RouterBase fork patches for the chat-LLM + media catalogue work upstream in
RouterBase/RouterBase.1. Preserve OpenAI
cached_tokensthroughUsage(6498b88)RouterBase routes chat LLMs (OpenAI / Anthropic / Gemini, all via Novita's OpenAI-compatible endpoint) through this gateway and needs the prompt-cache read count to bill cache reads at the discounted rate and show users their savings. Stock TZ normalizes provider usage to
{input_tokens, output_tokens}and dropsprompt_tokens_details.cached_tokens.Usagegainscached_tokens: Option<u32>(ts-bindings +skip_serializing_if/default).prompt_tokens_details.cached_tokensinOpenAIUsage → Usage.Usage::zero()and the streaming / cross-inference aggregators (sum,None-as-0).None(out of scope — our Anthropic models use the openai-compat path).cargo check --package tensorzero-core(lib) clean; no new clippy warnings.2. Wan 2.7 R2V media item types (54277f0)
Pre-existing fork commit (carried on this branch): fixes the Wan 2.7 reference-to-video request shape to match the upstream enum.
Notes
RouterBase/RouterBasePR #103 (submodule pointer bumped to 6498b88).prompt_cache_keyviaextra_body; this PR only handles surfacing the cache usage back to the caller.🤖 Generated with Claude Code