Builder keeps one host copy of a weight for every layer that reads it transposed

Ouverte
#4,934 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub

Les mainteneurs répondent en général sous 6 jours

Personne n'a encore pris cette issue.

Évaluation

Difficulté
4/5
Temps estimé
3-5 jours
Accessibilité débutants
25/100
Type d'issue
Bug
Clarté
Clairement spécifiée
Activité
Active
Stack technique
cpp, python

Piste de recherche

Start by running python trt_transpose_readers.py one_constant 10 and python trt_transpose_readers.py pretransposed 10 to reproduce the gap in host peak memory. The payload names no source files, so the builder code that handles transposed matrix-multiply weights has to be located. Done means one transposed constant read by N layers keeps host peak close to the pretransposed numbers, about 6.9 GiB for N=10.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

Module:Engine Build

What is wrong

When several matrix multiplies read the same weight with MatrixOperation.TRANSPOSE, the builder seems to keep a separate host copy of that weight for each reader. In the script below, one 1 GiB float32 weight read by 10 layers needs 16.1 GiB of host memory to build. Transposed once up front and then read by the same 10 layers, it needs 6.9 GiB. Each extra transposed reader adds about 1 GiB, one more copy of the weight.

Who hits it: any model that runs one block many times in one graph. Examples are a flow-matching or diffusion policy with its denoising loop unrolled, or classifier-free guidance that runs the same layers twice. Linear layers read their weight transposed, so this is the common case. Through Torch-TensorRT, a 3.5 GiB float32 policy that runs its action transformer 10 times needed 20.5 GiB of host memory to compile, and 11.0 GiB once the converter shared one transposed copy itself (pytorch/TensorRT#4965). Related: pytorch/TensorRT#4867.

What this blocks

xvla and eo1 (pytorch/executorch#24353) read the same weights in many layers, so their compile keeps one extra host copy per reader.
Torch-TensorRT 2.15 avoids it only with the opt-in deduplicate_weights (pytorch/TensorRT#4965). Anyone who builds with the TensorRT API, or without that option, still pays it.
A fix in the builder keeps that memory flat however many layers read a weight, which matters most on Jetsons with little memory.

Repro

python trt_transpose_readers.py MODE [N] [D], with MODE one of:

  • per_reader: N constant layers that point at the same host memory, each read transposed.
  • one_constant: one constant layer, read transposed by N layers.
  • pretransposed: one constant, transposed once with a shuffle layer, read by N layers without a transpose flag.

Default N = 10, D = 16384 (1 GiB float32). TF32 is off.

mode N = 1 N = 5 N = 10
per_reader 16061 MiB (15.7x)
one_constant 6918 MiB (6.8x) 10939 MiB (10.7x) 16060 MiB (15.7x)
pretransposed 6931 MiB (6.8x) 6931 MiB (6.8x) 6930 MiB (6.8x)

Host peak during the build, weight 1024 MiB, host memory before the build 1603 MiB. The plan is 1024 MiB in every case.

Expected

Host memory that does not grow with the number of readers, as in pretransposed: at most one builder copy of the weight, however many layers read it transposed.

Versions

TensorRT 11.4.0.106, numpy 2.5.4, H100, driver 580.126.09, Linux x86_64.

trt_transpose_readers.py
import sys, threading, time
import numpy as np
import tensorrt as trt

MODE = sys.argv[1]
N = int(sys.argv[2]) if len(sys.argv) > 2 else 10
D = int(sys.argv[3]) if len(sys.argv) > 3 else 16384
peak = [0]


def rss_mib():
    for line in open("/proc/self/status"):
        if line.startswith("VmRSS"):
            return int(line.split()[1]) // 1024


def sampler():
    while True:
        peak[0] = max(peak[0], rss_mib())
        time.sleep(0.05)


w = (np.random.default_rng(0).standard_normal((D, D)) / np.sqrt(D)).astype(np.float32)
w_mib = w.nbytes >> 20
builder = trt.Builder(trt.Logger(trt.Logger.WARNING))
net = builder.create_network(0)
x = net.add_input("x", trt.float32, (8, D))


def constant():
    return net.add_constant((D, D), trt.Weights(w)).get_output(0)


shared = constant() if MODE in ("one_constant", "pretransposed") else None
if MODE == "pretransposed":
    t = net.add_shuffle(shared)
    t.first_transpose = (1, 0)
    shared = t.get_output(0)
op = trt.MatrixOperation.NONE if MODE == "pretransposed" else trt.MatrixOperation.TRANSPOSE
for _ in range(N):
    weight = shared if shared is not None else constant()
    y = net.add_matrix_multiply(x, trt.MatrixOperation.NONE, weight, op).get_output(0)
    x = net.add_activation(y, trt.ActivationType.TANH).get_output(0)
net.mark_output(x)
cfg = builder.create_builder_config()
cfg.clear_flag(trt.BuilderFlag.TF32)

start = rss_mib()
threading.Thread(target=sampler, daemon=True).start()
plan = builder.build_serialized_network(net, cfg)
time.sleep(0.2)
print(f"{MODE}: N={N}, weight {w_mib} MiB, host before build {start} MiB, "
      f"build peak {peak[0]} MiB ({peak[0] / w_mib:.1f}x the weight), plan {plan.nbytes >> 20} MiB")

Actual

per_reader: N=10, weight 1024 MiB, host before build 1603 MiB, build peak 16061 MiB (15.7x the weight), plan 1024 MiB
one_constant: N=10, weight 1024 MiB, host before build 1603 MiB, build peak 16060 MiB (15.7x the weight), plan 1024 MiB
one_constant: N=5, weight 1024 MiB, host before build 1603 MiB, build peak 10939 MiB (10.7x the weight), plan 1024 MiB
one_constant: N=1, weight 1024 MiB, host before build 1603 MiB, build peak 6918 MiB (6.8x the weight), plan 1024 MiB
pretransposed: N=10, weight 1024 MiB, host before build 1603 MiB, build peak 6930 MiB (6.8x the weight), plan 1024 MiB
pretransposed: N=5, weight 1024 MiB, host before build 1603 MiB, build peak 6931 MiB (6.8x the weight), plan 1024 MiB
pretransposed: N=1, weight 1024 MiB, host before build 1603 MiB, build peak 6931 MiB (6.8x the weight), plan 1024 MiB
Langage dominant
C++
Étoiles
13.5k
Forks
2.4k
Merge moyen
6 j 21 h
PR mergées (30 j)
3

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de NVIDIA/TensorRT

Toutes les issues de NVIDIA/TensorRT

Issues similaires

Plus d'issues C++

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.