Skip to content

[nvbugs/5333742] fix MTP illegal memory access in cuda graph warmup #5947

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
merged 2 commits into from
Jul 12, 2025
Merged

[nvbugs/5333742] fix MTP illegal memory access in cuda graph warmup #5947

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
0ed52e0
set previous offsets to zero for the special cases in warmup.
lfr-0531 Jul 11, 2025
a1249ab
add test.
lfr-0531 Jul 11, 2025
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
10 changes: 10 additions & 0 deletions tensorrt_llm/_torch/pyexecutor/model_engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1408,6 +1408,16 @@ def _prepare_tp_inputs(
self.previous_batch_indices_cuda[:previous_batchs]],
non_blocking=True)

if (not self._disable_overlap_scheduler
and next_draft_tokens_device is None
and len(extend_requests) > 0):
# During warmup, for those generation requests, we don't have previous tensors,
# so we need to set the previous_pos_id_offsets and previous_kv_lens_offsets to zeros
# to skip the value changes in _preprocess_inputs. Otherwise, there will be illegal memory access
# when writing key/values to the KV cache.
self.previous_pos_id_offsets_cuda *= 0
self.previous_kv_lens_offsets_cuda *= 0

position_ids = torch.tensor(position_ids,
dtype=torch.int,
pin_memory=True)
Expand Down
1 change: 1 addition & 0 deletions tests/integration/test_lists/test-db/l0_dgx_b200.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@ l0_dgx_b200:
- accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8_4gpus[pp4-fp8kv=True-attn_backend=TRTLLM-torch_compile=False]
- accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[tp4-mtp_nextn=0-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=True]
- accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[tp4-mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=True]
- accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[tp4-mtp_nextn=2-attention_dp=False-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
- accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[ep4-mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False]
- accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[ep4-mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=True]
- accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[tp2pp2-mtp_nextn=0-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=True]
Expand Down