fix(engines/base): default use_cuda_graph=False for paged attention continuous batching

[Qubitium]

Summary

transformers continuous batching defaults use_cuda_graph=True for flash-attention paths. When attn_implementation contains the paged marker (e.g. "paged|flash_attention_2"), ModelRunner captures a CUDA graph around the model forward, and the paged decode path calls flash_paged._paged_decode_forward → flash_attn_with_kvcache → flash_attn_gpu.fwd_kvcache, which is not capture-safe. It fails with CUDA error: operation not permitted when stream is capturing.

This change adds a __post_init__ to BaseEnginePagedBatchingConfig so that when use_cuda_graph is not explicitly set and attn_implementation contains a paged token, it defaults to (False, False). Paged Flash Attention still runs, but CUDA graph capture is disabled.

def __post_init__(self) -> None:
    if self.use_cuda_graph is not None:
        return
    attn = getattr(self, "attn_implementation", None)
    if isinstance(attn, str) and "paged" in [part.strip() for part in attn.split("|")]:
        self.use_cuda_graph = (False, False)

This keeps the fix in one place (Evalution) so GPTQModel/Transformers engine configs always get the safe default, and downstream callers don't need to set it themselves.

Version bump

• pyproject.toml: 0.0.9 → 0.0.10
• evalution/version.py fallback: 0.0.9 → 0.0.10

Verification

• ruff check evalution/engines/base.py passes.
• PYTHON_GIL=0 pytest -q tests/test_transformer.py passes (62/62).
• PYTHON_GIL=0 pytest -q tests/test_gptqmodel_engine.py -k "not generate_and_score_on_cuda" passes; the skipped integration test fails because gptqmodel is not importable under PYTHON_GIL=0 in this environment, which is unrelated to this change.
• End-to-end gsm8k_platinum_cot (64 samples) on a pre-quantized Llama 3.2 1B GPTQ checkpoint with attn_implementation=paged|flash_attention_2 runs backend=continuous_batching paged_attention=True and completes with no CUDA graph-capture error.

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring