Profiling in PyTorch (Part 3): Attention is all you profile

The series "Profiling in PyTorch" is meant to make you comfortable reading profiler traces and tables. In Part 1 we profiled basic math operations like addition and multiplication. We saw how the profiler table uncovers hotspots, and how the profiler trace shows the order in which an algorithm runs over time.

In Part 2 we wrapped that addition and multiplication into a torch linear layer. We then stacked several linear layers on top of each other (a multilayer perceptron) and profiled that. Along the way we also profiled fused and hand-tuned kernels.

From the perspective of the Transformer architecture, the next logical step for us to profile is yet another fundamental algorithm, attention. While being infamous for its quadratic-time complexity, many clever tricks exist to mitigate that issue and make it fast. Our goal here is not to cover every trick in detail. Instead, we want to see how each one looks different under the profiler.

The scripts for this blog post live here: 04_a_naive_attention.py, 04_b_inplace_ops_attention.py, 04_c_sdpa_attention.py, and 04_d_kernels_attention.py. Like before, it helps to open them in a separate tab and walk through the code as you read. We use an NVIDIA A100-SXM4-80GB GPU to run the scripts. It is really easy to set up a GPU on the Hugging Face infrastructure and experiment with the scripts using Dev Mode with Spaces. One could also run the scripts with the Hugging Face Jobs pipeline.

Naive attention

Attention works with Queries (q), Keys (k), and Values (v). The interaction between them can be written as a short sequence of steps:

1. Build the attention scores scores: matmul(q, k.T)
2. Scale the scores: scores * scale
3. Apply a causal mask to the scores: scores.masked_fill(mask, "-inf")
4. Normalize the scores with softmax to get the attention weights attn: softmax(scores)
5. Reweight the values with those weights: matmul(attn, v)

So attention is really a collection of primitive operations. Some of them we already know (the matmuls), and the rest are easy to spot. Let's write a naive attention module in PyTorch and profile it.

class NaiveCausalAttention(nn.Module):
    def __init__(self, head_dim):
        super().__init__()
        self.scale = 1.0 / math.sqrt(head_dim)

    def forward(self, q, k, v, mask):
        scores = torch.matmul(q, k.transpose(-2, -1))
        scores = scores * self.scale
        scores = scores.masked_fill(mask, float("-inf"))
        attn = torch.softmax(scores, dim=-1)
        out = torch.matmul(attn, v)
        return out

Before opening the trace, let's do our usual exercise and guess what we should see. Tracing the forward of this module, we expect:

• a matmul kernel (q . k.T)
• a mul kernel (the scaling)
• an operation for the masking
• a softmax kernel
• a matmul kernel (atten . v)

uv run 04_a_naive_attention.py
uvx trace-util -f traces/ -b <hf_uname>/traces

Figure 1 shows the CPU lane of the profile (the GPU lane is folded so it does not overwhelm us). Inside attn_fwd (our annotated forward call) we can see exactly the operations we guessed. The matmul is an old friend by now, and the new operations are easy to spot:

• mul: the scaling
• masked_fill: the causal masking
• softmax: the softmax kernel

Now let's unfold the GPU lane and see which kernels were actually launched.

Figure 2 shows the GPU lane next to the CPU lane. Let's zoom into a single attn_fwd block on the GPU lane to look at the kernels one by one.

Figure 3 lets us read off the individual kernels for one profiler step:

1. matmul (query and key)
2. mul (scaling)
3. memory copy 🤔
4. causal masking
5. softmax (produces the attention weights)
6. matmul (attention weights and values)

Five of these are expected. The memory copy is the odd one out, so where does this come from? The clue is that PyTorch has in-place operations. When you operate on a tensor the ordinary (out-of-place) way, PyTorch often makes a copy, applies the requested operation to it, and returns the copy. Following the sequence of operations, the culprit here is our masked_fill.

What if we replaced this with an in-place operation?

Naive attention with inplace causal masking

All we change is masked_fill to masked_fill_ (note the trailing underscore, PyTorch's convention for in-place operations), and we run the same script.

def forward(self, q, k, v, mask):
    # q, k, v: [batch, heads, seq, head_dim]
    scores = torch.matmul(q, k.transpose(-2, -1))  # [batch, heads, seq, seq]
    scores = torch.mul(scores, self.scale)
-    scores = scores.masked_fill(mask, float("-inf"))
+    scores.masked_fill_(mask, float("-inf"))
    attn = torch.softmax(scores, dim=-1)
    out = torch.matmul(attn, v)  # [batch, heads, seq, head_dim]
    return out

Let's look at the trace and see if something changed.

uv run 04_b_inplace_ops_attention.py
uvx trace-util -f traces/ -b <hf_uname>/traces

The in-place version (Figure 5) wraps far fewer CPU ops inside the masking step than the out-of-place version (Figure 4). This is an encouraging signal. Let's unfold the GPU lane to confirm what happened there.

On the GPU lane the Memcpy kernel is gone for good (Figures 6 and 7). With a one line change we shaved a whole kernel off each forward pass. This may not look like much on its own, but remember this is a single attention operation. In the context of a transformer based large model (LLMs, Diffusion models, etc.), it repeats once per layer, and there are many layers, so the saving adds up quickly (and if it earns you a raise, sharing at least 10% with us feels only fair).

Out-of-place is PyTorch's default for a reason. To compute gradients, autograd has to remember the tensor values it saw on the forward pass, because many backward formulas reuse them. An in-place operation overwrites those values in memory, so the backward pass would read the wrong numbers. Due to the fact that we run forward under torch.no_grad, in-place is safe for us, with no backward pass and nothing to corrupt. It is also noteworthy that in-place operations do not only save time (like we see in our case) but also memory (due to no extra copy) which is great for large tensors like logits!

Scaled Dot Product Attention

We just built attention from primitives, and even shaved off a Memcpy. The good news is that the PyTorch team has done all of this for us, and packaged the whole pipeline into a single function:

from torch.nn import functional as F

F.scaled_dot_product_attention(q, k, v, is_causal=True)

This one line replaces our hand written module, and is_causal=True even saves us from building the mask by hand. It is worth pausing to appreciate how much this one call hides. And it hides more than just code lines. Scaled Dot Product Attention (SDPA) does not have a single implementation. Under the hood it dispatches to one of the several backends and picks the fastest one that supports our inputs (dtype, head dimension, mask, hardware, etc.).

The official SDPA tutorial walks us through this selection, and the backends themselves are listed in the torch.nn.attention.SDPBackend enum:

from torch.nn.attention import SDPBackend

BACKENDS = {
    "math": SDPBackend.MATH,
    "flash": SDPBackend.FLASH_ATTENTION,
    "efficient": SDPBackend.EFFICIENT_ATTENTION,
    "cudnn": SDPBackend.CUDNN_ATTENTION,
}

Normally SDPA chooses for us, but we can pin a specific backend with the torch.nn.attention.sdpa_kernel context manager. This is what we do in our scripts. This lets us profile each backend on its own and read how differently they show up in the trace. Let's go one at a time.

Math backend

uv run 04_c_sdpa_attention.py --backend math
uvx trace-util -f traces/ -b <hf_uname>/traces

Before we open anything, let's guess. We have replaced hand written attention (matmul, mul, mask, softmax, matmul) with a single one liner, so we should expect the trace to get simpler and faster. Fewer kernels, less CPU dispatch, maybe even a fused kernel. Let's check the profiler table first.

This is our first surprise, the one liner is 3.7x slower.

Opening the trace (Figure 9) shows why the alarm bells ring, the math backend launches 20 GPU kernels per forward instead of the 5 launched with our naive attention implementation (Figure 8). This is the opposite of what we guessed. Let's figure out why this happens.

Tensor cores left vacant

In Part 2 we learned to read a kernel name like a fingerprint. Let's use that habit here:

The A100s we used to capture these traces ship with Tensor Cores, specialised hardware for accelerated matmuls that is known to be far faster than the ordinary CUDA cores. To see why that matters here, it helps to know what lives inside a GPU. A Streaming Multiprocessor (SM) is the compute unit of a GPU, and each SM has two kinds of arithmetic units, the CUDA cores and the Tensor Cores. CUDA cores are general purpose and process a handful of elements at a time, while Tensor Cores multiply and accumulate a whole small matrix tile in a single instruction. So the question is simple, "Is each backend actually using the fast path?"

The kernel names answer it. The s16816 in the naive kernel (Figure 10) is the signature of a bfloat16 Tensor Core matmul (the 16x8x16 Tensor Core instruction), so the naive version is on the fast path. sgemm (Figure 11) is the classic single precision (FP32) matmul that runs on the ordinary CUDA cores. In other words, the math backend never touches the Tensor Cores at all: to trade speed for numerical accuracy it upcasts tensors to FP32 (doubling the data moved, even when the inputs are in bf16) and falls back to the slower CUDA cores.

Causal masks built

In the naive version we built the causal mask once and reused it. Here we passed is_causal=True and the math backend materialized one for us, on every single call. You can watch it happen on the CPU lane:

Here is what we see in Figure 12

aten::ones -> aten::tril            build a [seq, seq] lower-triangular matrix
aten::scalar_tensor -> aten::fill_  make the -inf fill value
aten::where                         turn it into an additive bias (0 or -inf)

On the GPU this shows up as a triu_tril_kernel, several where kernels, and an add_. The convenience flag that let us stop thinking about the mask did not remove the work, it just moved it one layer down, where the mask is rebuilt from scratch every forward.

The safe softmax

Our hand written version called plain aten::softmax. The math backend calls aten::_safe_softmax, and the difference is again visible as extra kernels (Figure 13):

A row that is fully masked (every entry -inf) would make an ordinary softmax compute exp(-inf)/sum(exp(-inf)) = 0/0 = NaN. _safe_softmax guards against exactly that. Our naive kernel never bothered, and would have quietly produced NaNs in that corner case.

Read the full original article:

HuggingFace Blog