fix: update payload preparation to merge parameters into the output dictionary

[mishig25]

Fix feature extraction parameter structure for Inference Endpoints compatibility

Problem

The feature_extraction method's truncate parameter (and other parameters like normalize,
prompt_name, truncation_direction) is not working correctly when using custom Inference
Endpoints via base_url.

Root Cause

When the provider helper system was introduced in #2777, the parameter structure for HF
Inference API requests was inadvertently changed from root-level parameters to nested
parameters:

Before (working):

{
  "inputs": "text",
  "truncate": true,
  "normalize": true,
  "prompt_name": "query"
}

After #2777 (broken):

  {
    "inputs": "text",
    "parameters": {
      "truncate": true,
      "normalize": true,
      "prompt_name": "query"
    }
  }

Impact

• ❌ Python InferenceClient.feature_extraction(truncate=True) fails with custom endpoints
• ✅ JavaScript @huggingface/inference client works correctly (uses root-level parameters)
• ✅ HF Serverless Inference API works (likely handles both formats)

This inconsistency breaks existing Inference Endpoints that were deployed expecting the
original parameter format.

Solution

Restore the original parameter structure by modifying
HFInferenceTask._prepare_payload_as_dict() to place parameters at the root level instead of
nesting them under a "parameters" key.

Fixed payload:

  {
    "inputs": "text",
    "truncate": true,
    "normalize": true,
    "prompt_name": "query"
  }

Testing

Without this PR, the snippet below fails:

from huggingface_hub import InferenceClient

client = InferenceClient(base_url="MY_INFERENCE_ENDPOINT_URL", token="hf_TOKEN")

embeddings = client.feature_extraction("This is a test sentence. "*100, truncate=True)

print(embeddings)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Wauplin]

Hey @mishig25 , thanks for reporting + opened a PR. I just pushed 889a4ee so that this fix applies only to "feature-extraction" task and not to all tasks. I also added a quick test.

I'm approving right now but if that's fine with you, I'd prefer to wait until @hanouticelina 's back to have her opinion as well.

Note: I went back to our input.json specs which clearly states that parameters must be at root level.

[Wauplin]

no need for this change 50dba39

Suggested change

	__version__ = "0.33.0"
	__version__ = "0.33.0.dev0"

[mishig25]

I need it in the meantime. Otherwise, getting error when using with diffusers: version has to be xyz

[Wauplin]

ah yes ok so let's not forget about it before merging. Converting back to draft in the meantime

[hanouticelina]

Looks good to me! let's move this PR to ready for review and revert the package version change before merging. Once it's merged, we can go ahead and release a patch.

Side note: This also got me thinking that we should probably consider switch now to OpenAI-compatible specs for embeddings/feature-extraction for HF Inference API and update the SDKs accordingly (in another PRs of course!). cc @tomaarsen for viz

[mishig25]

actually, when I try after @Wauplin changes, I'm getting the error of parameters being ignored

[hanouticelina]

@mishig25 what is the exact error you're getting? with this PR, the following snippet works as expected and truncate is sent at the root level:

from huggingface_hub import InferenceClient

client = InferenceClient(base_url="MY_INFERENCE_ENDPOINT_URL", token="hf_TOKEN")

embeddings = client.feature_extraction("This is a test sentence. "*100, truncate=True)

print(embeddings)

[mishig25]

@hanouticelina it is working now, will merge it