Skip to content

[https://nvbugs/6248764][fix] Normalize non-sliding KV windows to full attention in AutoDeploy #14906

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
eopXD merged 2 commits into from
Jun 5, 2026
Merged

[https://nvbugs/6248764][fix] Normalize non-sliding KV windows to full attention in AutoDeploy #14906

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
9b0dfd4
[https://nvbugs/6248764][fix] Normalize non-sliding KV windows to ful…
eopXD Jun 3, 2026
b3bf607
[https://nvbugs/6248764][fix] Unwaive gemma-3-1b-it AutoDeploy regist…
eopXD Jun 3, 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
9 changes: 9 additions & 0 deletions tensorrt_llm/_torch/auto_deploy/transform/library/kvcache.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -432,6 +432,15 @@ def _apply(
# None sentinel: pass literal None positionally, no resource allocated.
cache_in_nodes.append(None)
else:
# A window that can't slide within max_seq_len is functionally
# full attention; normalize it to 0 so the layer shares the
# full-attention pool instead of forking a redundant
# single-window pool.
if (
isinstance(resource_handler, KVPagedResourceHandler)
and resource_handler.sliding_window >= cm.info.max_seq_len
):
resource_handler.sliding_window = 0
resource_name = cm.add_resource(k, resource_handler)
node = self._process_cache_node(gm, resource_name)
cache_in_nodes.append(node)
Expand Down
1 change: 0 additions & 1 deletion tests/integration/test_lists/waives.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,6 @@ accuracy/test_llm_api.py::TestMistralNemo12B::test_fp8 SKIP (https://nvbugs/5413
accuracy/test_llm_api_autodeploy.py::TestGemma4MoE::test_bf16 SKIP (https://nvbugs/6158397)
accuracy/test_llm_api_autodeploy.py::TestGemmaE2B::test_gemma4_e2b_it SKIP (https://nvbugs/6194934)
accuracy/test_llm_api_autodeploy.py::TestMiniMaxM2::test_finegrained_fp8 SKIP (https://nvbugs/6158397)
accuracy/test_llm_api_autodeploy.py::TestModelRegistryAccuracy::test_autodeploy_from_registry[google_gemma-3-1b-it-False] SKIP (https://nvbugs/6248764)
accuracy/test_llm_api_autodeploy.py::TestModelRegistryAccuracy::test_autodeploy_from_registry[mistralai_Ministral-8B-Instruct-2410-False] SKIP (https://nvbugs/6248769)
accuracy/test_llm_api_autodeploy.py::TestModelRegistryAccuracy::test_autodeploy_from_registry[nvidia_Llama-3.1-8B-Instruct-NVFP4-True] SKIP (https://nvbugs/6245279)
accuracy/test_llm_api_autodeploy.py::TestNemotronNanoV3::test_accuracy[bf16-4-trtllm] SKIP (https://nvbugs/6185150)
Expand Down
Loading