Skip to content

[TRTLLM-15403][fix] VisualGen: warn when a requested attention backend silently falls back to VANILLA #17698

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
chang-l merged 2 commits into from
Aug 19, 2026
Merged

[TRTLLM-15403][fix] VisualGen: warn when a requested attention backend silently falls back to VANILLA #17698

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
258823d
[TRTLLM-15403][fix] VisualGen: warn when a requested attention backen…
chang-l Aug 14, 2026
2ea2c80
Merge branch 'main' into fix/vgoa-attn-backend-fallback-warning
chang-l Aug 18, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions tensorrt_llm/_torch/visual_gen/models/cosmos3/transformer_cosmos3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -513,6 +513,13 @@ def __init__(
if model_config.attention.backend == "TRTLLM":
# TRTLLM backend is not supported for Cosmos3CrossAttention
model_config.attention.backend = "VANILLA"
# Warn once per (module class, requested, resolved) triple so the
# fallback is visible without per-module-instance log spam.
logger.warning_once(
f"{type(self).__name__}: requested attention backend {original_backend} is not "
f"supported for Cosmos3 cross-attention; falling back to VANILLA.",
key=(type(self).__name__, original_backend, "VANILLA"),
)

super().__init__(
hidden_size=hidden_size,
Expand Down
9 changes: 9 additions & 0 deletions tensorrt_llm/_torch/visual_gen/modules/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@
import torch
import torch.nn as nn

from tensorrt_llm.logger import logger
from tensorrt_llm.visual_gen.sparse_attention import SkipSoftmaxAttentionConfig

from ...modules.linear import Linear, TensorParallelMode, WeightMode, WeightsLoadingConfig
Expand Down Expand Up @@ -105,6 +106,14 @@ def __init__(
# Cross-attention fallback: TRTLLM and CUTEDSL VSA are self-attn only.
if self.qkv_mode == QKVMode.SEPARATE_QKV and (base_backend == "TRTLLM" or _is_vsa):
backend_name = "VANILLA"
requested = f"{base_backend} (VSA)" if _is_vsa else base_backend
# Warn once per (module class, requested, resolved) triple so the
# fallback is visible without per-module-instance log spam.
logger.warning_once(
f"{type(self).__name__}: requested attention backend {requested} does not "
f"support qkv_mode=SEPARATE_QKV; falling back to {backend_name}.",
key=(type(self).__name__, requested, backend_name),
)
else:
backend_name = base_backend

Expand Down
Loading