Fused attention never returns when the bool mask is a run-time tensor, at 2176 to 2304 query rows

Offen
#4,932 0 Kommentare 0 Reaktionen 0 zugewiesene Personen Auf GitHub ansehen

Maintainer antworten meist innerhalb von 6 Tagen

Dieses Issue hat noch niemand übernommen.

Bewertung

Schwierigkeit
5/5
Geschätzter Aufwand
Über eine Woche
Anfängerfreundlichkeit
8/100
Issue-Typ
Bug
Klarheit
Größtenteils klar
Aktivitätsstatus
Aktiv
Tech-Stack
python, pytorch
Bereich
ai-infra-agents

Rechercherichtung

The hang is in the _gemm_mha_v2 fused attention kernel for a bool mask that is a runtime input, at query lengths 2176 to 2304; the repro in the issue is a single script repro.py that takes the length as an argument. Start by running it at 2176 and at 2048 to confirm the size dependence, then compare with --const. A fix needs a kernel-level change in the closed-source TensorRT attention path, so there is no open source file to edit here.

Vom Indexierungsmodell aus dem Issue-Text verfasst.

Beschreibung

Module:Runtime

What is wrong

An IAttention layer (bf16, softmax, decomposable=True) with a bool mask that is a network input hangs on its
first execute. One _gemm_mha_v2 kernel (grid 132 x 384 on H100) stays busy forever, and execute_async_v3 never
finishes. CUDA_LAUNCH_BLOCKING=1 shows the kernel hangs on its own. It does not depend on the input values.

It depends on the size and on the mask pattern. A square causal mask hangs for N = 2176, 2240, 2303 and 2304. It
runs for 1024, 2048, 2305, 2368, 2432, 3456, 3584 and 4096. 2184 and 2240 queries against 3467 keys also hang. The
same mask as a build-time constant runs in 0.47 ms with correct results. An all-true run-time mask also runs.

Anyone who builds attention masks inside the network can hit this at these sizes. That is the default when a framework keeps mask
computation out of constant folding, which Torch-TensorRT does.

What this blocks

No policy in pytorch/executorch#24353 is blocked by this today.
Any model that builds its mask inside the network, which is the Torch-TensorRT default, never returns from its first execute at these sizes, with no error.
A fix lets such models keep run-time masks at every size.

Repro

Save as repro.py. Needs tensorrt and torch (torch only for buffers and the reference). It prints HANG after 60 s.

import argparse, os, threading, time
import numpy as np
import tensorrt as trt
import torch

p = argparse.ArgumentParser()
p.add_argument("n", type=int, help="query length (and key length unless --keys)")
p.add_argument("--keys", type=int, default=None)
p.add_argument("--hq", type=int, default=24)
p.add_argument("--hkv", type=int, default=8)
p.add_argument("--const", action="store_true", help="mask as a build-time constant")
p.add_argument("--ones", action="store_true", help="all-true mask instead of causal")
p.add_argument("--timeout", type=float, default=60)
a = p.parse_args()
sq, skv, d = a.n, a.keys or a.n, 128
mask = torch.ones(sq, skv, dtype=torch.bool) if a.ones else torch.arange(sq)[:, None] + (skv - sq) >= torch.arange(skv)[None, :]
q = torch.randn(1, a.hq, sq, d, device="cuda", dtype=torch.bfloat16)
k = torch.randn(1, a.hkv, skv, d, device="cuda", dtype=torch.bfloat16)
v = torch.randn(1, a.hkv, skv, d, device="cuda", dtype=torch.bfloat16)

logger = trt.Logger(trt.Logger.WARNING)
builder = trt.Builder(logger)
net = builder.create_network(0)
tq = net.add_input("q", trt.bfloat16, tuple(q.shape))
tk = net.add_input("k", trt.bfloat16, tuple(k.shape))
tv = net.add_input("v", trt.bfloat16, tuple(v.shape))
if a.const:
    m = mask.reshape(1, 1, sq, skv).numpy()
    tm = net.add_constant(m.shape, trt.Weights(m)).get_output(0)
else:
    tm = net.add_input("mask", trt.bool, (1, 1, sq, skv))
s = np.full((1, 1, 1, 1), 1 / np.sqrt(d), dtype=np.float32)
sc = net.add_cast(net.add_constant(s.shape, trt.Weights(s)).get_output(0), trt.bfloat16).get_output(0)
sq_ = net.add_elementwise(tq, sc, trt.ElementWiseOperation.PROD).get_output(0)
att = net.add_attention_v2(sq_, tk, tv, trt.AttentionNormalizationOp.SOFTMAX, trt.CausalMaskKind.NONE)
att.mask = tm
att.decomposable = True
net.mark_output(att.get_output(0))
plan = builder.build_serialized_network(net, builder.create_builder_config())
eng = trt.Runtime(logger).deserialize_cuda_engine(plan)
ctx = eng.create_execution_context()
out = torch.empty(1, a.hq, sq, d, device="cuda", dtype=torch.bfloat16)
bufs = {"q": q, "k": k, "v": v}
if not a.const:
    bufs["mask"] = mask.reshape(1, 1, sq, skv).cuda().contiguous()
for name, t in bufs.items():
    ctx.set_tensor_address(name, t.data_ptr())
ctx.set_tensor_address(eng.get_tensor_name(eng.num_io_tensors - 1), out.data_ptr())
threading.Timer(a.timeout, lambda: (print(f"HANG: execute did not finish in {a.timeout:.0f} s", flush=True), os._exit(2))).start()
stream = torch.cuda.Stream()
t = time.time()
ctx.execute_async_v3(stream.cuda_stream)
stream.synchronize()
ms = (time.time() - t) * 1e3
ref = torch.nn.functional.scaled_dot_product_attention(q.float(), k.float(), v.float(), attn_mask=mask.cuda(), enable_gqa=True)
err = ((out.float() - ref).abs().max() / ref.abs().max()).item()
print(f"ok: {ms:.2f} ms (first call), max abs error / max abs = {err:.2e}", flush=True)
os._exit(0)
python repro.py 2304            # HANG: execute did not finish in 60 s
python repro.py 2176            # HANG
python repro.py 2048            # ok, about 0.24 ms
python repro.py 2305            # ok, about 0.35 ms
python repro.py 2304 --const    # ok: same mask as a constant
python repro.py 2304 --ones     # ok: all-true mask

Versions: H100 80GB HBM3 (sm_90), driver 580.126.09, TensorRT 11.4.0.106 (pip wheel), bf16, 24 query heads and
8 key/value heads (24/24 hangs too), head size 128.

Expected

The layer finishes, as it does with a constant mask or at nearby sizes.

What we tried

  • Builder optimization levels 0 to 5: all hang.
  • Padding keys to a multiple of 64 or 128: still hangs.
  • decomposable=False: the build fails ("not supported by a dedicated kernel").
  • A bool mask turned into an additive mask inside the network: still hangs. An additive bf16 mask given as an
    input runs.
  • Workaround: make the mask a build-time constant when it does not change between calls.
Vorherrschende Sprache
C++
Sterne
13.5k
Forks
2.4k
Ø Merge
6 T. 21 Std.
Gemergte PRs (30 T.)
3

Entwicklungsumgebung

Erste Schritte

  1. Lesen Sie das ganze Issue und danach den Beitragsleitfaden des Projekts.
  2. Schreiben Sie ins Issue, dass Sie es übernehmen — das erspart doppelte Arbeit.
  3. Forken Sie das Repository und arbeiten Sie in einem Branch.
  4. Öffnen Sie einen Pull Request, der die Issue-Nummer nennt.

Mehr aus NVIDIA/TensorRT

Alle Issues in NVIDIA/TensorRT

Ähnliche Issues

Weitere Issues zu C++

Neue Issues direkt in Ihr Postfach

Eine kurze Übersicht über anfängerfreundliche GitHub-Issues.