[BUG][cuda.compute]: histogram cache key over-specializes on runtime-only values

[NaderAlAwar]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Performance

Component

cuda.compute (Python)

Describe the bug

In python/cuda_cccl/cuda/compute/algorithms/_histogram.py, _make_histogram_even_impl is cached on:

• num_output_levels_val
• lower_level_val
• upper_level_val
• level_dtype
• num_samples
• uses_privatized_smem

This means calls that differ only in lower/upper scalar values compile/cache separate histogram build artifacts, even when dtype, sample/output layouts, output level regime, and offset type are unchanged.

How to Reproduce

Run the histogram benchmark with configurations that vary only runtime histogram parameters:

pip install 'cuda-bench[cu13]'

python python/cuda_cccl/benchmarks/compute/histogram/even.py
--profile
--axis 'SampleT{ct}=I32'
--axis 'Elements{io}[pow2]=16'
--axis 'Bins=[32,128,2048,2097152]'
--axis 'Entropy=[0.201,1.000]'

Then inspect/cache-instrument make_histogram_even: current main creates separate cached build objects for exact num_output_levels_val, lower_level_val, and upper_level_val, even when the generated-code regime is unchanged.

A minimal object-level repro is to call make_histogram_even twice with the same sample/output layouts, same level_dtype, and same num_bins <= 256 regime, but different lower/upper bounds. Expected: cache reuse. Actual: cache miss because exact scalar values are part of the key.

Expected behavior

We should be able to reuse the cached histogram object even when lower_level_val and upper_level_val change

Reproduction link

No response

Operating System

No response

nvidia-smi output

No response

NVCC version

No response