📌 Description

CuteDslMoEWrapper pre-allocates intermediate buffers sized for
max_num_tokens, but the autotuner can probe buckets larger than that
(e.g. 8192 tokens vs 2048 max). The GEMM kernels then write past
buffer bounds, silently corrupting model weights and eventually
triggering cudaErrorIllegalAddress.

Two fixes:

• Check num_tokens <= max_num_tokens before reusing pre-allocated
  buffers; fall back to dynamic allocation when exceeded.
• Move tuning_config to instance level so dummy expert IDs span all
  local experts (randint(0, num_experts)) instead of a hardcoded 8,
  which concentrated routing and inflated permutation buffer sizes.

🔍 Related Issues

feat: cuteDSL fp4 moe for better DSR1 performance.

🚀 Pull Request Checklist

Thank you for contributing to FlashInfer! Before we review your pull request, please make sure the following items are complete.

✅ Pre-commit Checks

• I have installed pre-commit by running pip install pre-commit (or used your preferred method).
• I have installed the hooks with pre-commit install.
• I have run the hooks manually with pre-commit run --all-files and fixed any reported issues.

If you are unsure about how to set up pre-commit, see the pre-commit documentation.

🧪 Tests

• Tests have been added or updated as needed.
• All tests are passing (unittest, etc.).

Reviewer Notes

Summary by CodeRabbit

• Performance Improvements

  • CUDA-graph preallocation now respects input-size limits to avoid over-allocation and improve memory behavior for larger token batches.
  • Memory gating refined to prevent oversized buffer reuse, improving inference stability under varying workloads.

• Stability & Tuning

  • Auto-tuner configuration is now per-run, improving tuning accuracy and consistency.
  • Runtime tuning initialization adapts to the configured expert count for more representative profiling.