Masked bf16 attention rounds its scores to bf16 on Jetson Thor (sm_110), but not on A100

オープン
#4,938 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 6 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
8/100
issue の種類
バグ
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
cpp, python

調査の方向性

The repro is a single Python script that compiles one scaled_dot_product_attention with Torch-TensorRT and prints channel 0 for no mask, a bool mask and a float mask. Run it on a Jetson AGX Thor and on an A100 to confirm the masked outputs differ. Done means the masked Thor output matches the float32-score result (+0.4219) and the random-input error falls back to about the PyTorch level, which needs a fix inside TensorRT's attention kernels.

索引モデルが issue の本文から書いたものです。

説明

Module:Accuracy Module:Embedded

What is wrong

On Jetson AGX Thor, a bf16 attention layer with a mask (bool or additive float) computes its softmax scores in
bf16. Without a mask it keeps them in float32. On A100 (sm_80), with the same
TensorRT, the masked layer keeps float32 scores, like PyTorch.

When 2 scores are close, bf16 rounding makes them equal, so the softmax weights change. On random inputs the error
grows about 13 times.

What this blocks

On Jetson Thor, evo1 misses its accuracy check and fails (pytorch/executorch#24353).
lingbot_vla_v2 passes there only because LeRobot keeps its attention out of TensorRT on Thor. That splits it into hundreds of pieces and makes it slower than eager PyTorch.
A fix lets both use fused attention on Thor, as they do on the other GPUs.

Repro

The script compiles one scaled_dot_product_attention with Torch-TensorRT, which builds one IAttention layer.
A query and 2 keys are chosen so that their scaled scores (200.45 and 199.55) round to the same bf16 value. With
float32 scores, output channel 0 is +0.42. With bf16 scores it is 0.

import torch, torch.nn.functional as F, torch_tensorrt, tensorrt

B, H, L, S, D = 8, 8, 50, 805, 64
dev = "cuda"

class Attention(torch.nn.Module):
    def forward(self, q, k, v, mask=None):
        return F.scaled_dot_product_attention(q, k, v, attn_mask=mask)

def near_tie():
    q = torch.zeros(B, H, L, D, device=dev); q[..., 0], q[..., 1] = 16.0, 2.0
    k = torch.zeros(B, H, S, D, device=dev)
    k[..., 0, 0], k[..., 0, 1] = 100.0, 1.796875
    k[..., 1, 0], k[..., 1, 1] = 100.0, -1.796875
    v = torch.zeros(B, H, S, D, device=dev); v[..., 0, 0], v[..., 1, 0] = 1.0, -1.0
    return q.bfloat16(), k.bfloat16(), v.bfloat16()

def random_qkv():
    g = torch.Generator(device=dev).manual_seed(0)
    return tuple((torch.randn(B, H, n, D, device=dev, generator=g) * s).bfloat16()
                 for n, s in ((L, 2.0), (S, 3.0), (S, 1.0)))

def reference(q, k, v, mask):
    s = q.double() @ k.double().transpose(-2, -1) / D**0.5
    if mask is not None:
        s = s.masked_fill(~mask, float("-inf")) if mask.dtype == torch.bool else s + mask.double()
    return s.softmax(-1) @ v.double()

keep = torch.ones(B, 1, 1, S, dtype=torch.bool, device=dev)
keep[1::2, ..., S - S // 8:] = False  # key padding; keys 0 and 1 stay visible
additive = torch.zeros(B, 1, 1, S, device=dev).masked_fill(~keep, float("-inf")).bfloat16().expand(B, H, 1, S).contiguous()

print(torch.cuda.get_device_name(), torch.cuda.get_device_capability(), "TensorRT", tensorrt.__version__)
model = Attention().eval()
for name, mask in (("none", None), ("bool", keep), ("float", additive)):
    args = near_tie() + (() if mask is None else (mask,))
    trt = torch_tensorrt.dynamo.compile(torch.export.export(model, args), arg_inputs=list(args), min_block_size=1)
    with torch.no_grad():
        tie = trt(*args)[..., 0].float().mean().item()
        rargs = random_qkv() + (() if mask is None else (mask,))
        ref = reference(*rargs[:3], mask)
        ok = torch.isfinite(ref)
        err_trt = (trt(*rargs).double() - ref)[ok].abs().max().item()
        err_pt = (model(*rargs).double() - ref)[ok].abs().max().item()
    print(f"mask={name:5} channel 0: {tie:+.4f} (float32 scores give +0.4219); "
          f"random max error vs float64: TensorRT {err_trt:.4f}, PyTorch {err_pt:.4f}")

Results

no mask bool mask float mask random error, masked
Thor (11,0) +0.4219 +0.0000 +0.0000 0.1206 (PyTorch 0.0090)
A100 (8,0) +0.4219 +0.4219 +0.4219 0.0106 (PyTorch 0.0106)

Versions: torch 2.15.0 (CUDA 13.2), torch_tensorrt 2.15.0, TensorRT 11.4.0.106 (all from pip install "torch_tensorrt[executorch]"). Thor: JetPack 7.2.1, driver 595.78, CUDA 13.2.
A100: driver 580.126.09.

Actual

Jetson AGX Thor, driver 595.78:

NVIDIA Thor (11, 0) TensorRT 11.4.0.106
mask=none  channel 0: +0.4219 (float32 scores give +0.4219); random max error vs float64: TensorRT 0.0090, PyTorch 0.0084
mask=bool  channel 0: +0.0000 (float32 scores give +0.4219); random max error vs float64: TensorRT 0.1206, PyTorch 0.0090
mask=float channel 0: +0.0000 (float32 scores give +0.4219); random max error vs float64: TensorRT 0.1206, PyTorch 0.0090

A100 80 GB, driver 580.126.09:

NVIDIA PG509-210 (8, 0) TensorRT 11.4.0.106
mask=none  channel 0: +0.4219 (float32 scores give +0.4219); random max error vs float64: TensorRT 0.0106, PyTorch 0.0106
mask=bool  channel 0: +0.4219 (float32 scores give +0.4219); random max error vs float64: TensorRT 0.0106, PyTorch 0.0106
mask=float channel 0: +0.4219 (float32 scores give +0.4219); random max error vs float64: TensorRT 0.0106, PyTorch 0.0106

Expected

Float32 scores for masked attention on Thor, as on A100.

主要言語
C++
スター
13.5k
フォーク
2.4k
平均マージ
6日 21時間
マージ済み PR(30日)
3

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

NVIDIA/TensorRT のほかの issue

NVIDIA/TensorRT の issue をすべて見る

似ている issue

C++ の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。