| """TransformerEngine DotProductAttention (cuDNN fused, thd, head_dim 256, bf16) fwd/bwd at the |
| training sequence lengths vs an fp32 SDPA reference. DPA returns [T, H*D].""" |
| import torch |
|
|
| print("torch", torch.__version__, "| device", torch.cuda.get_device_name(0), flush=True) |
| import transformer_engine.pytorch as te |
| from transformer_engine.pytorch.attention import DotProductAttention |
| print("TE", getattr(te, "__version__", "?"), flush=True) |
| dev = "cuda" |
| torch.manual_seed(0) |
|
|
|
|
| def bad(t): |
| return bool(torch.isnan(t).any() or torch.isinf(t).any()) |
|
|
|
|
| def rel(a, b): |
| return ((a.float() - b.float()).abs().max() / b.float().abs().max().clamp_min(1e-6)).item() |
|
|
|
|
| H, Hkv, D = 6, 1, 256 |
| for Tl in ([400, 300], [4000, 3500], [8700, 9100, 400], [12288], [9000, 3200, 80, 8]): |
| try: |
| T = sum(Tl) |
| cu = torch.tensor([0] + list(torch.cumsum(torch.tensor(Tl), 0)), dtype=torch.int32, device=dev) |
| dpa = DotProductAttention(num_attention_heads=H, kv_channels=D, num_gqa_groups=Hkv, attention_dropout=0.0, |
| qkv_format="thd", attn_mask_type="padding_causal").to(dev) |
| q = torch.randn(T, H, D, device=dev, dtype=torch.bfloat16, requires_grad=True) |
| k = torch.randn(T, Hkv, D, device=dev, dtype=torch.bfloat16, requires_grad=True) |
| v = torch.randn(T, Hkv, D, device=dev, dtype=torch.bfloat16, requires_grad=True) |
| out = dpa(q, k, v, cu_seqlens_q=cu, cu_seqlens_kv=cu, max_seqlen_q=max(Tl), max_seqlen_kv=max(Tl)) |
| out = out.view(T, H, D) |
| do = torch.randn_like(out) |
| (out * do).sum().backward() |
| fused = {n: t.grad.detach().clone() for n, t in (("q", q), ("k", k), ("v", v))} |
| print(f"TE fused thd hd256 T={T} seqs={Tl}: out nan/inf {bad(out)} | grad nan/inf", {n: bad(t) for n, t in fused.items()}, flush=True) |
| for t in (q, k, v): |
| t.grad = None |
| ref_out = [] |
| for i in range(len(Tl)): |
| s, e = int(cu[i]), int(cu[i + 1]) |
| qi = q[s:e].float().transpose(0, 1) |
| ki = k[s:e].float().transpose(0, 1).expand(H, -1, -1) |
| vi = v[s:e].float().transpose(0, 1).expand(H, -1, -1) |
| oi = torch.nn.functional.scaled_dot_product_attention(qi, ki, vi, is_causal=True) |
| ref_out.append(oi.transpose(0, 1)) |
| ref = torch.cat(ref_out, dim=0) |
| (ref * do.float()).sum().backward() |
| print(f" fwd rel err {rel(out, ref):.3e} | grad rel err " + ", ".join(f"{n}={rel(fused[n], t.grad):.3e}" for n, t in (("q", q), ("k", k), ("v", v))), flush=True) |
| del q, k, v, out, do, fused, ref, dpa |
| torch.cuda.empty_cache() |
| except Exception as exc: |
| print(f" case {Tl} failed: {type(exc).__name__}: {str(exc)[:300]}", flush=True) |
| print("DONE", flush=True) |
|
|
