[v0.2.7] Multi-SM CUTLASS kernels with runtime dispatch

[m96-chan]

Summary

Implement SM-specific CUTLASS kernel variants with runtime dispatch for optimal performance across GPU architectures.

Motivation

• SM 80 (A100), SM 86 (RTX 30xx), SM 89 (RTX 40xx), SM 90 (H100) have different optimal configurations
• Single kernel compiled for lowest SM leaves performance on the table
• Runtime dispatch allows shipping one wheel that works optimally on all GPUs

Proposed Architecture

Wheel Structure

pygpukit/
 ├─ core/                  # Rust core
 ├─ native/
 │   ├─ ops/
 │   │   ├─ matmul_cutlass_sm80.cu   # Ampere (A100)
 │   │   ├─ matmul_cutlass_sm86.cu   # Ampere (RTX 30xx)
 │   │   ├─ matmul_cutlass_sm90.cu   # Hopper (H100)
 │   │   └─ matmul_fallback.cu       # SIMT fallback

Runtime Selector (Rust)

match device.sm {
    90.. => kernel_sm90,   // Hopper+
    86.. => kernel_sm86,   // RTX 30xx, RTX 40xx
    80.. => kernel_sm80,   // A100
    _    => fallback,      // Pre-Ampere (not officially supported)
}

Benefits

1. Optimal performance - Each SM gets tuned tile sizes, pipeline depth
2. Single wheel - No need for separate builds per architecture
3. Future-proof - Easy to add SM 100+ when available

Implementation Notes

• Each SM variant compiles with -arch=sm_XX
• Fallback uses SIMT (no TensorCore) for unsupported architectures
• Runtime detection via cudaGetDeviceProperties

Related

• [v1.0] CUTLASS FP32 SIMT kernel for full precision matmul #67 CUTLASS FP32 SIMT kernel (fallback implementation)
• Current: All kernels compile for SM 80 with -arch=sm_80

[m96-chan]

Completed in PR #76

Implementation

• Runtime SM detection via get_sm_tier() function in matmul_cutlass.cuh
• SM-specific kernel variants:
  • SM80 (A100): 4-stage pipeline, 48KB shared memory
  • SM86 (RTX 30xx): 5-stage pipeline, 100KB shared memory
  • SM89 (RTX 40xx): 6-stage pipeline
  • SM90+ (H100/Blackwell): CUTLASS 3.x with WGMMA/TMA

Files Changed

• native/ops/matmul_cutlass.cu - SM dispatch logic
• native/ops/matmul_cutlass.cuh - SM80/86/89 kernels
• native/ops/matmul_cutlass_sm90.cuh - SM90+ CUTLASS 3.x kernels
• .github/workflows/release.yml - Build SM80-90 (Linux), SM80-120 (Windows)

CI Verified

• ✓ Linux wheel (SM80/86/89/90)
• ✓ Windows wheel (SM80/86/89/90/100/120)