Capability-Aware Fast Paths

When optimizing for speed, make the fast path both correct for the actual hardware and lightweight in implementation.

1) Keep kernel support documentation accurate and capability-aware

• Maintain an explicit GPU architecture support matrix for optimized kernels.
• Only claim “Hopper-only” (or similar) if that’s true; otherwise document the actual compiled architectures.
• Ensure the code selects an appropriate fallback (e.g., SDPA) for unsupported architectures.

Example (pattern):

# Supported FA3 archs reflect what kernels are compiled for
# sm90 (Hopper), sm89 (Ada), sm80/sm86 (Ampere)
# sm100 (Blackwell): fallback to SDPA until FA3 is recompiled

def _load_flash_attention_3():
    if not torch.cuda.is_available():
        return None
    # ... select FA3 vs fallback based on device capability ...

2) Avoid unnecessary overhead in parsing/utility code

• If checkpoint filenames are already guaranteed to match model_<step>.pt, prefer simple split/string ops over regex.
• Remove extra path manipulation (basename) when you already operate on plain filenames.

Example (pattern):

# checkpoint_files contains only filenames like model_123456.pt
last_step = int(max(f.split("_")[-1].split(".")[0] for f in checkpoint_files))

Result: faster execution with fewer CPU-side bottlenecks, and fewer correctness surprises caused by stale/incorrect hardware assumptions.