fix: Fixes for bugs encountered during generation & evaluation

[peri044]

This PR has some fixes for a few bugs encountered while running the kernel generation & evaluation using a local hosted model.
Please review

cc: @simonguozirui

[peri044]

@simonguozirui @anneouyang Can you take a look at this ? Also, I have been trying to generate kernels using https://github.com/ScalingIntelligence/KernelBench?tab=readme-ov-file#run-on-all-problems but when I evaluate them, I get poor pass@k results with strong models (eg: GPT-5.2). The commands are

# For generation
uv run python scripts/generate_samples.py run_name=test_hf_level_1_gpt5.2_fp16 dataset_src=huggingface level=1 num_workers=50  model_name="azure/openai/gpt-5.2" server_type=openai server_address=localhost server_port=9000  log_prompt=True verbose=True num_samples=50 max_tokens=32768 precision="fp32" reasoning_effort="high" is_reasoning_model=True subset="(1, 10)"

# For evaluation
uv run python scripts/eval_from_generations.py run_name=test_hf_level_1_gpt5.2_fp32_subset10_check_kernel_false dataset_src=local level=1 num_gpu_devices=1 timeout=300 subset="(0, 10)" num_samples_per_problem=50 pass_at_k_values="[1, 3]"

I get syntax errors, lockfile errors and missing headerfile errors for level 1 problem solutions. I believe this is not expected since level - 1 problems are almost solved ? Can you share some insights and a reproducer of results ? Thanks

[peri044]

@AffectionateCurry @simonguozirui @anneouyang @nataliakokoromyti - would appreciate any insights or help on this thread.