[cloudgpuapp]

Adds CloudGPU as a model provider, so a graph config can use "model": "cloudgpu/deepseek-v4-flash" with a CLOUDGPU_API_KEY.

CloudGPU (cloudgpu.app) has an OpenAI-compatible model API at https://cloudgpu.app/v1: a model API for DeepSeek, Kimi, GLM, MiniMax, Qwen and more.

Changes

• scrapegraphai/models/cloudgpu.py: CloudGPU(ChatOpenAI) wrapper, same shape as MiniMax: maps api_key, defaults the base URL to https://cloudgpu.app/v1 unless base_url / openai_api_base is given.

• scrapegraphai/models/__init__.py: re-export CloudGPU, matching the other model wrappers.

• scrapegraphai/graphs/abstract_graph.py: cloudgpu added to the known providers and to the custom-class branch, returning CloudGPU(**llm_params).

• scrapegraphai/helpers/models_tokens.py: cloudgpu context lengths for seven chat models, taken from context_length in GET https://cloudgpu.app/v1/models.

• examples/smart_scraper_graph/cloudgpu/smart_scraper_cloudgpu.py: SmartScraper example.

• tests/test_cloudgpu_model.py: base URL / key mapping, custom base URL kept, token table entries, and cloudgpu/<model> resolution through SmartScraperGraph (graph creation patched, no network).

Testing

• ruff check, black --check, and isort --check-only on the touched files: clean.

• pytest tests/test_cloudgpu_model.py: 4 passed. These tests do not call the network.

• Context lengths were checked against the public GET https://cloudgpu.app/v1/models (no key). Every model id in this PR is in that list, with the context_length values used here:

deepseek-v4-flash 131072

• deepseek-v4-pro 131072

• glm-5.3 131072

• kimi-k2.6 262144

• minimax-m3 1000000

• qwen3-coder-480b 262144

• qwen3.5-397b 262144

• Live graph execution was not run.

• make lint is not clean on pre/beta without this change (pre-existing ruff, black, and isort findings in unrelated files). mypy reports a missing annotation on CloudGPU.__init__, the same pattern as MiniMax and AtlasCloud, plus existing errors in abstract_graph.py.

• A unit run excluding integration, e2e, and requires_api_key was 272 passed, 98 failed, 1 error. The failures are pre-existing (missing Playwright browser, network graph tests, and token-table assertions). The same four failures in tests/test_models_tokens.py and tests/graphs/abstract_graph_test.py::test_llm_missing_tokens fail on pre/beta without this change. The new CloudGPU tests are not among the failures.

Checklist

• Follows the existing provider pattern (MiniMax, DeepSeek)

• Example added under examples/smart_scraper_graph/

• Unit tests added (no network)

• Target branch pre/beta per CONTRIBUTING

Disclosure: I work on CloudGPU.

Open in Web Open in Cursor