Add DG_MINIMIZE_NUM_SMS env var to control SM minimization

[yurekami]

Summary

This PR adds an environment variable DG_MINIMIZE_NUM_SMS to control whether DeepGEMM should minimize the number of SMs used for kernel execution, addressing #239.

Background

The current SM minimization logic in get_best_config() recomputes the minimal number of SMs required:

if (ArchSpec::should_minimize_num_sms()) {
    num_min_sms = ceil_div(ceil_div(m, best_block_m) * ceil_div(n, best_block_n) * num_groups, best_num_waves);
    num_min_sms = align(num_min_sms, best_multicast_config.num_multicast);
}

This can invoke kernel compilation multiple times with different configurations, causing unstable time-to-first-token (TTFT) during inference.

Changes

• Modified SM90ArchSpec::should_minimize_num_sms() to check env var
• Modified SM100ArchSpec::should_minimize_num_sms() to check env var

Usage

# Enable SM minimization (default, current behavior)
export DG_MINIMIZE_NUM_SMS=1

# Disable SM minimization for stable TTFT
export DG_MINIMIZE_NUM_SMS=0

Benefits

• DG_MINIMIZE_NUM_SMS=1 (default): Better L2 cache usage, reduced GPU frequency drops
• DG_MINIMIZE_NUM_SMS=0: More predictable compilation, stable TTFT

Test Plan

• Verify default behavior is unchanged (SM minimization enabled)
• Verify DG_MINIMIZE_NUM_SMS=0 disables SM minimization
• Compare TTFT stability with env var disabled

Fixes #239

🤖 Generated with Claude Code