用xFormers构建内存高效Transformer教程

In this tutorial, we implement xFormers: a practical toolkit for building fast, memory-efficient Transformer models on GPUs. We begin by validating memory-efficient attention against a standard attention implementation, then compare their speed and memory consumption across different sequence lengths. We then examine causal masking, packed variable-length sequences, grouped-query attention, and custom ALiBi positional biases. Finally, we combine these techniques into a trainable GPT-style model that uses xFormers attention, SwiGLU feed-forward layers, and automatic mixed-precision training. Setting Up xFormers and Validating Memory-Efficient Attention Copy CodeCopiedUse a different Browserimport subprocess, sys def _pip(*a): subprocess.run([sys.executable, "-m", "pip", "install", *a], check=False) try: import xformers except Exception: _pip("-q", "-U", "xformers") import math, time import torch, torch.nn as nn, torch.nn.functional as F import xformers, xformers.ops as xops from xformers.ops import fmha ab = fmha.attn_bias assert torch.cuda.is_available(), ( "No GPU detected. In Colab: Runtime → Change runtime type → GPU, then re-run.") device = "cuda" torch.manual_seed(0) print("torch :", torch.__version__) print("xformers :", xformers.__version__) print("GPU :", torch.cuda.get_device_name(0)) print("\n--- xformers.info (which kernels are built/available) ---") try: subprocess.run([sys.executable, "-m", "xformers.info"], check=False) except Exception as e: print("xformers.info unavailable:", e) def cuda_time(fn, iters=20, warmup=5): for _ in range(warmup): fn() torch.cuda.synchronize() s, e = (torch.cuda.Event(enable_timing=True) for _ in range(2)) s.record() for _ in range(iters): fn() e.record(); torch.cuda.synchronize() return s.elapsed_time(e) / iters def peak_mem_mb(fn): torch.cuda.empty_cache(); torch.cuda.reset_peak_memory_stats() fn(); torch.cuda.synchronize() return torch.cuda.max_memory_allocated() / 1e6 def vanilla_attention(q, k, v, causal=False): """Reference attention that MATERIALIZES the [B,H,M,M] score matrix. Inputs are xformers-layout [B, M, H, K].""" q, k, v = (t.transpose(1, 2).float() for t in (q, k, v)) scores = (q @ k.transpose(-2, -1)) / math.sqrt(q.shape[-1]) if causal: M = scores.shape[-1] m = torch.triu(torch.ones(M, M, device=q.device, dtype=torch.bool), 1) scores = scores.masked_fill(m, float("-inf")) out = scores.softmax(-1) @ v return out.transpose(1, 2) print("\n" + "="*70 + "\n1. memory_efficient_attention basics + correctness\n" + "="*70) B, M, H, K = 2, 512, 8, 64 q, k, v = (torch.randn(B, M, H, K, device=device, dtype=torch.float16) for _ in range(3)) out_xf = xops.memory_efficient_attention(q, k, v) out_ref = vanilla_attention(q, k, v).half() print("output shape :", tuple(out_xf.shape), "(layout B, M, H, K)") print("max abs diff vs ref : {:.2e}".format((out_xf - out_ref).abs().max().item())) print("-> it's EXACT attention (fp16 rounding only), just computed without") print(" ever storing the full MxM score matrix.") We install and import xFormers, verify GPU availability, and inspect the attention kernels supported by the environment. We define helper functions for measuring CUDA execution time and peak memory consumption. We then validate memory-efficient attention against standard attention to confirm that both produce results that closely match each other. Benchmarking Memory and Speed Against Naive Causal Attention Copy CodeCopiedUse a different Browserprint("\n" + "="*70 + "\n2. Memory & speed vs naive attention (fwd+bwd)\n" + "="*70) print(f"{'seqlen':>8} | {'naive MB':>10} | {'xformers MB':>12} | {'naive ms':>9} | {'xf ms':>7}") print("-"*60) for M in [512, 1024, 2048, 4096]: q, k, v = (torch.randn(2, M, 8, 64, device=device, dtype=torch.float16, requires_grad=True) for _ in range(3)) def run_xf(): o = xops.memory_efficient_attention(q, k, v); o.sum().backward() def run_naive(): o = vanilla_attention(q, k, v); o.sum().backward() try: nm = peak_mem_mb(run_naive); nt = cuda_time(run_naive, 8, 3) except RuntimeError: nm, nt = float("nan"), float("nan"); torch.cuda.empty_cache() xm = peak_mem_mb(run_xf); xt = cuda_time(run_xf, 8, 3) print(f"{M:>8} | {nm:>10.0f} | {xm:>12.0f} | {nt:>9.2f} | {xt:>7.2f}") print("-> naive memory grows ~4x per doubling of M (it stores BxHxMxM);") print(" xformers grows ~linearly and stays fast.") print("\n" + "="*70 + "\n3. Causal attention via LowerTriangularMask\n" + "="*70) B, M, H, K = 2, 256, 8, 64 q, k, v = (torch.randn(B, M, H, K, device=device, dtype=torch.float16) for _ in range(3)) out_causal = xops.memory_efficient_attention(q, k, v, attn_bias=ab.LowerTriangularMask()) ref_causal = vanilla_attention(q, k, v, causal=True).half() print("causal max abs diff : {:.2e}".format((out_causal - ref_causal).abs().max().item())) print("-> the mask is implicit; no MxM boolean tensor is allocated.") We benchmark naive attention and xFormers attention across progressively longer sequences using forward and backward passes. We compare their execution times and peak GPU memory usage to observe how xFormers avoids quadratic memory growth. We also apply an implicit lower-triangular mask and verify causal attention against the reference implementation. Packing Variable-Length Sequences and Running Grouped-Query Attention Copy CodeCopiedUse a different Browserprint("\n" + "="*70 + "\n4. Variable-length packed batch — no padding waste\n" + "="*70) seqlens = [37, 120, 8, 200] total = sum(seqlens) H, K = 8, 64 q = torch.randn(1, total, H, K, device=device, dtype=torch.float16) k = torch.randn(1, total, H, K, device=device, dtype=torch.float16) v = torch.randn(1, total, H, K, device=device, dtype=torch.float16) try: bias = ab.BlockDiagonalMask.from_seqlens(seqlens) out_packed = xops.memory_efficient_attention(q, k, v, attn_bias=bias) s0 = seqlens[0] ref0 = vanilla_attention(q[:, :s0], k[:, :s0], v[:, :s0]).half() print("packed shape :", tuple(out_packed.shape), "(all", total, "tokens, no pad)") print("segment-0 max diff : {:.2e}".format((out_packed[:, :s0] - ref0).abs().max().item())) cbias = ab.BlockDiagonalCausalMask.from_seqlens(seqlens) _ = xops.memory_efficient_attention(q, k, v, attn_bias=cbias) print("-> also did a packed CAUSAL pass. This is how vLLM-style engines") print(" batch requests of different lengths with zero padding overhead.") splits = bias.split(out_packed) print("recovered segments :", [tuple(t.shape) for t in splits]) except Exception as e: print("BlockDiagonalMask path skipped on this version/backend:", repr(e)) print("\n" + "="*70 + "\n5. Grouped-query attention (5-D BMGHK layout)\n" + "="*70) B, M, K = 2, 256, 64 n_q_heads, n_kv_heads = 8, 2 G, Hq = n_kv_heads, n_q_heads // n_kv_heads try: qg = torch.randn(B, M, G, Hq, K, device=device, dtype=torch.float16) kg = torch.randn(B, M, G, 1, K, device=device, dtype=torch.float16) vg = torch.randn(B, M, G, 1, K, device=device, dtype=torch.float16) out_gqa = xops.memory_efficient_attention(qg, kg, vg) print("GQA output shape :", tuple(out_gqa.shape), "= [B, M, G, Hq, K]") print(f"-> {n_q_heads} query heads, only {n_kv_heads} KV heads: smaller KV-cache,") print(" which is exactly what Llama-/Mistral-class models use at inference.") except Exception as e: print("GQA 5-D path skipped on this version/backend:", repr(e)) We concatenate variable-length sequences and use BlockDiagonalMask to prevent attention from crossing sequence boundaries without padding. We recover the individual outputs and also perform packed causal attention for decoder-style workloads. We then demonstrate grouped-query attention, where multiple query heads share fewer key-value heads to reduce KV-cache requirements. Adding a Custom ALiBi Additive Positional Bias Copy CodeCopiedUse a different Browserprint("\n" + "="*70 + "\n6. Custom ALiBi additive bias\n" + "="*70) B, M, H, K = 1, 128, 8, 64 q, k, v = (torch.randn(B, M, H, K, device=device, dtype=torch.float16) for _ in range(3)) try: slopes = (2.0 ** (-8.0 / H)) ** torch.arange(1, H + 1, device=device) po

原文超出正文长度上限，此处截断——上游还有内容，完整版见上方「原文 ↗」。