model_dump() in _generate_content_stream serializes full image contents per response (~32 MB wasted)

[feynnon]

Describe the bug

In _generate_content_stream() (and _generate_content(), and all async variants), every streaming response creates a full parameter_model.model_dump() to pass as kwargs to GenerateContentResponse._from_response():

# models.py, line 4324 (and ~35 other identical call sites)
return_value = types.GenerateContentResponse._from_response(
    response=response_dict, kwargs=parameter_model.model_dump()
)

_from_response() (types.py:6943) only reads kwargs['config']['response_schema'] and kwargs['config']['response_json_schema'] from that dict. It never accesses kwargs['contents'].

When contents includes inline image data (e.g. reference images for image generation), each model_dump() call serializes the entire image bytes into a Python dict — then discards them.

Measured impact

We instrumented the SDK (monkey-patched _GenerateContentParameters.model_dump to measure json.dumps(result) size) and ran isolated benchmarks:

Scenario	model_dump() size per call
No reference images	~1 KB
3 reference images (~3 MB each)	31.7 MB
5 reference images (~3 MB each)	52.8 MB

With 24 concurrent image generation threads (common for batch image generation), this creates ~760 MB of transient allocations that serve no purpose, contributing to OOM kills in memory-constrained environments.

To reproduce

1. Call generate_content_stream() with inline image data in contents (e.g. reference images via Part.from_bytes)
2. Monitor RSS memory — each streaming response allocates a dict equal in size to the full serialized request including all image bytes
3. With multiple concurrent requests, memory spikes rapidly

Proposed fix

Pass only the config to _from_response() instead of dumping the full parameter model:

# Instead of:
kwargs=parameter_model.model_dump()

# Pass just what _from_response needs:
kwargs={"config": parameter_model.config.model_dump() if parameter_model.config else {}}

This affects ~35 call sites in models.py (all _from_response calls follow the same pattern).

Environment

• SDK version: google-genai 1.63.0 (also confirmed present in 1.68.0)
• Python: 3.12
• OS: Linux (Docker) / macOS

Related

• slow memory leak #1258 reports a slow memory leak during streaming. Our instrumentation suggests model_dump() is a contributing factor — the large dicts hold references to image bytes that delay GC, especially under concurrent load.