"""Compare fused causal convolution outputs/gradients with PyTorch's reference.""" import json from pathlib import Path import torch import torch.nn.functional as F from causal_conv1d import causal_conv1d_fn torch.backends.cudnn.allow_tf32 = False torch.backends.cuda.matmul.allow_tf32 = False results = [] for dtype, channel_last, with_bias, length in ( (dtype, layout, bias, length) for dtype in [torch.float32, torch.bfloat16] for layout in [False, True] for bias in [False, True] for length in [511, 512, 513, 1025, 1537, 1730] ): torch.manual_seed(3407) if channel_last: x = torch.randn(2, length, 128, device="cuda", dtype=dtype).transpose(1, 2).requires_grad_() else: x = torch.randn(2, 128, length, device="cuda", dtype=dtype, requires_grad=True) w = torch.randn(128, 4, device="cuda", dtype=dtype, requires_grad=True) b = torch.randn(128, device="cuda", dtype=dtype, requires_grad=True) if with_bias else None inputs = (x, w, b) if with_bias else (x, w) grad = torch.randn_like(x) ref = F.silu(F.conv1d(x, w.unsqueeze(1), b, padding=3, groups=128)[..., :x.shape[-1]]) g_ref = torch.autograd.grad(ref, inputs, grad) # Match production's padding to a multiple of 8. The reference stays # unpadded, so this also verifies that padding does not change active tokens. padded = F.pad(x, (0, (-x.shape[-1]) % 8)) if channel_last: padded = padded.transpose(1, 2).contiguous().transpose(1, 2) out = causal_conv1d_fn(padded, w, b, activation="silu")[..., :x.shape[-1]] g_out = torch.autograd.grad(out, inputs, grad) errors = [] for name, actual, expected in zip(["output", "input_gradient", "weight_gradient", "bias_gradient"], [out, *g_out], [ref, *g_ref]): relative = ((actual.float() - expected.float()).norm() / expected.float().norm()).item() errors.append(relative) assert relative < (0.012 if dtype == torch.bfloat16 else 1e-4), (dtype, channel_last, with_bias, length, name, relative) results.append({"dtype": str(dtype), "channel_last": channel_last, "bias": with_bias, "sequence_length": length, "relative_l2_output_and_gradients": errors}) Path("kernel_correctness.json").write_text(json.dumps(results, indent=2)) summary = {dtype: max(max(r["relative_l2_output_and_gradients"]) for r in results if r["dtype"] == dtype) for dtype in ["torch.float32", "torch.bfloat16"]} print(json.dumps({"passed_cases": len(results), "maximum_relative_l2_error": summary}, indent=2), flush=True)