fix: replace mutable default input_kwargs={} with None in HuggingfaceEngine

[kuishou68]

Closes #10476

Problem

Four static methods in HuggingfaceEngine use input_kwargs: Optional[dict[str, Any]] = {} as a default parameter. This is a Python anti-pattern: the same dict object is shared across all invocations that omit the argument. Since each method calls .pop() on it, the shared dict is drained after the first call — subsequent callers relying on the default receive an already-mutated (empty) dict.

Affected methods:

• _process_args
• _chat
• _stream_chat
• _get_scores

Fix

Change the default to None and initialise input_kwargs = {} at the top of each function body, following the standard Python idiom.

[gemini-code-assist]

Code Review

This pull request replaces mutable default arguments for input_kwargs with None and adds support for the minicpmv4_6 model type. The review feedback highlights that input_kwargs should be copied to prevent side effects from .pop() calls mutating the caller's dictionary. Additionally, the reviewer suggests making downsample_mode configurable per request via input_kwargs rather than relying on a global environment variable.

[gemini-code-assist]

There are two issues with this new block:

1. Using os.getenv for a generation parameter like downsample_mode makes it globally fixed for the process, preventing per-request configuration. It would be better to allow this to be passed via input_kwargs with the environment variable as a fallback.
2. If minicpmv4_6 is a multimodal model architecture similar to minicpmv, it likely requires the same input_ids and tokenizer keys in gen_kwargs that are set in the preceding block (lines 202-208). By keeping this check separate, those necessary configurations are skipped, which may lead to runtime errors or incorrect model behavior.

Suggested change

	if getattr(model.config, "model_type", None) == "minicpmv4_6":
	gen_kwargs["downsample_mode"] = os.getenv("DOWNSAMPLE_MODE", "16x")
	if getattr(model.config, "model_type", None) == "minicpmv4_6":
	gen_kwargs["downsample_mode"] = input_kwargs.pop("downsample_mode", os.getenv("DOWNSAMPLE_MODE", "16x"))

[gemini-code-assist]

While this change correctly addresses the shared mutable default argument issue, it still allows the input_kwargs dictionary to be mutated if one is provided by the caller. Since this method calls .pop() on the dictionary later (e.g., lines 123-133), it will have the side effect of draining the caller's dictionary. It is safer to create a local copy to avoid these side effects.

        input_kwargs = input_kwargs.copy() if input_kwargs is not None else {}

[gemini-code-assist]

Similar to the issue in _process_args, this method should copy input_kwargs to prevent the .pop() calls in the subsequent _process_args call from mutating the dictionary passed by the caller.

        input_kwargs = input_kwargs.copy() if input_kwargs is not None else {}

[gemini-code-assist]

Similar to the issue in _process_args, this method should copy input_kwargs to prevent the .pop() calls in the subsequent _process_args call from mutating the dictionary passed by the caller.

        input_kwargs = input_kwargs.copy() if input_kwargs is not None else {}

[gemini-code-assist]

This method calls input_kwargs.pop("max_length", None) on line 331. To avoid mutating the dictionary provided by the caller, it is recommended to work on a copy.

        input_kwargs = input_kwargs.copy() if input_kwargs is not None else {}