Optimize FP8 Triton kernels for better performance

[yurekami]

Summary

This PR addresses Issue #1052 by optimizing the Triton FP8 kernels for improved performance and correctness.

Changes

1. act_quant_kernel improvements:

• Added boundary masking for partial blocks to prevent out-of-bounds memory access
• Added n_elements parameter for proper boundary handling
• Extracted FP8_E4M3_MAX constant (448.0) for code clarity

2. fp8_gemm_kernel optimizations:

• Extended autotuning configurations with larger block sizes (up to 128x256)
• Added dynamic num_warps calculation based on block dimensions for optimal GPU occupancy
• Added M dimension to autotune key for better configuration selection
• Introduced explicit stride parameters for flexible memory layouts
• Improved code documentation

3. Autotuning configuration improvements:

• Expanded block_m options: [16, 32, 64] → [16, 32, 64, 128]
• Expanded block_n options: [32, 64, 128] → [32, 64, 128, 256]
• Reduced num_stages options: [3, 4, 5, 6] → [3, 4, 5] for faster tuning
• Added tile size limit (16384 elements) to avoid excessive register pressure

Expected Benefits

• Better memory access patterns: Explicit strides allow the compiler to generate more efficient code
• Improved GPU utilization: Dynamic num_warps adapts to block size for optimal occupancy
• Faster autotuning: Reduced configuration space while covering important cases
• Correctness: Boundary masking prevents out-of-bounds access in edge cases

Test plan

• Python syntax validation passes (python3 -m py_compile inference/kernel.py)
• Run inference with sample input to verify output correctness
• Benchmark on representative workloads to measure performance improvement

Related Issues

Closes #1052

🤖 Generated with Claude Code

[esball1]

yeah, just, you know... did you even optimize something?