[python_pytorch_profiler] Add new check - #152
Draft
KowalskiThomas wants to merge 1 commit into
Draft
Conversation
This comment has been minimized.
This comment has been minimized.
KowalskiThomas
force-pushed
the
kowalski/python_pytorch_profiler-add-new-check
branch
from
June 4, 2026 09:08
2ab5eed to
1767942
Compare
gh-worker-dd-mergequeue-cf854d Bot
pushed a commit
to DataDog/dd-trace-py
that referenced
this pull request
Jun 12, 2026
## Description This PR makes our PyTorch integration create profiles that actually look like profiles. Previously, we would get events from the PyTorch Profiler, sample them, and report them without looking at their context. As a result, we would get a "flat flame graph" where every frame would be at the same level. ### Results **Before my changes** flame graphs are "flat" <img width="1878" height="123" alt="image" src="https://github.com/user-attachments/assets/000a32c6-fc49-47f1-8f14-82f1e4153987" /> **After my changes** flame graphs are flame graphs (note: different code being profiled) <img width="1743" height="460" alt="image" src="https://github.com/user-attachments/assets/0bd15ca7-794b-414f-a86c-bc2569ea6ce2" /> [Live example profile](https://ddstaging.datadoghq.com/profiling/explorer?query=service%3Akowalski-python-torch-5&agg_m=count&agg_m_source=base&agg_t=count&color_by=total&fromUser=true&group_by=line&my_code=disabled&refresh_mode=paused&viz=flame_graph&from_ts=1780646782169&to_ts=1780650382169&live=false) (while ~stocks~ retention lasts!) <details> The previous flame graph was obtained by profiling the following script. ```py import os from time import time import torch from torch import nn from torch.profiler import ProfilerActivity class FeedForward(nn.Module): def __init__(self, d_model: int, d_ff: int) -> None: super().__init__() self.w1 = nn.Linear(d_model, d_ff) self.act = nn.GELU() self.w2 = nn.Linear(d_ff, d_model) self.norm = nn.LayerNorm(d_model) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.norm(x + self.w2(self.act(self.w1(x)))) class TransformerBlock(nn.Module): def __init__(self, d_model: int, n_heads: int, d_ff: int) -> None: super().__init__() self.attn = nn.MultiheadAttention(d_model, n_heads, batch_first=True) self.attn_norm = nn.LayerNorm(d_model) self.ff = FeedForward(d_model, d_ff) def forward(self, x: torch.Tensor) -> torch.Tensor: attn_out, _ = self.attn(x, x, x) x = self.attn_norm(x + attn_out) return self.ff(x) class DeepModel(nn.Module): def __init__(self, d_model: int = 256, n_heads: int = 4, d_ff: int = 512, n_layers: int = 6) -> None: super().__init__() self.embed = nn.Linear(64, d_model) self.blocks = nn.ModuleList([TransformerBlock(d_model, n_heads, d_ff) for _ in range(n_layers)]) self.head = nn.Sequential( nn.Linear(d_model, d_ff), nn.GELU(), nn.Linear(d_ff, d_model), nn.LayerNorm(d_model), nn.Linear(d_model, 16), ) def forward(self, x: torch.Tensor) -> torch.Tensor: x = self.embed(x) for block in self.blocks: x = block(x) return self.head(x) def main() -> None: torch.set_num_threads(1) model = DeepModel() x = torch.randn(8, 32, 64) # (batch, seq_len, input_dim) execution_time_sec = float(os.getenv("EXECUTION_TIME_SEC", "30")) end = time() + execution_time_sec with torch.no_grad(): while time() < end: with torch.profiler.profile(activities=[ProfilerActivity.CPU]): for _ in range(10): model(x) main() ``` </details> ### Performance Note this theoretically comes at a performance cost, as we need to reconstruct the event parent tree. However, looking at a Python profile for the Profiler (convenient that our Profiler is written in Python!) The top lines are 1. `events = prof.events()` -- this line hasn't changed and used to be called as much as it is today. 2. `self_cpu_time: int = e.self_cpu_time_total` -- this is new but is already more than 10x less CPU-intensive than 1. 3. `handle.push_frame(parent.name, _FILE_PLACEHOLDER, 0, 0)` -- this is not new and not called more often than it used to be 4. `handle.flush_sample()` -- this is not new and not called more often than it used to be 5. `handle.push_frame(f"PYTORCH_{device_type_str}", _DEVICE_FRAME_FILE_NAME, 0, 0)` -- this is not new and not called more often than it used to be In short, the most resource-hungry functions on the PyTorch Profiler at the moment are functions that already were there before, and that already were CPU-hungry before. The added logic (to reconstruct the tree) is practically irrelevant as far as performance goes. ### Related work New prof-correctness check: DataDog/prof-correctness#152 Co-authored-by: thomas.kowalski <thomas.kowalski@datadoghq.com>
KowalskiThomas
force-pushed
the
kowalski/python_pytorch_profiler-add-new-check
branch
from
July 7, 2026 08:21
1767942 to
94feef2
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What is this PR?
This PR is a correctness check specifically associated to this PR: DataDog/dd-trace-py#18457.