Skip to content

[https://nvbugs/5580099][fix] Separate cuda graph workspace to prevent IMA #8685

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
MartinMarciniszyn merged 4 commits into from
Oct 29, 2025
Merged

[https://nvbugs/5580099][fix] Separate cuda graph workspace to prevent IMA #8685

Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
ad90d97
[https://nvbugs/5606268][fix] Separate cuda graph workspace to preven…
JunyiXu-nv Oct 27, 2025
682bd12
Minor fix
JunyiXu-nv Oct 27, 2025
f56299a
Fix test
JunyiXu-nv Oct 28, 2025
a7990ae
Fix test
JunyiXu-nv Oct 29, 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
14 changes: 12 additions & 2 deletions tensorrt_llm/_torch/attention_backend/trtllm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -536,6 +536,7 @@ def is_nvfp4_output_kernel_available(
@dataclass(kw_only=True)
class TrtllmAttentionMetadata(AttentionMetadata):
workspace: Optional[torch.Tensor] = None
cuda_graph_workspace: Optional[torch.Tensor] = None

# TrtllmAttention needs to know the beam width to access to the cache indirection buffer,
# when beam search is enabled.
Expand Down Expand Up @@ -693,6 +694,14 @@ def get_empty_like(like_tensor: torch.Tensor,
device='cuda',
dtype=torch.int8,
)

if self.is_cuda_graph and self.cuda_graph_workspace is None:
self.cuda_graph_workspace = torch.empty(
(0, ),
device='cuda',
dtype=torch.int8,
)

if self.kv_cache_manager is not None:
self.kv_cache_block_offsets = get_empty(
[
Expand Down Expand Up @@ -1276,8 +1285,9 @@ def forward(
host_kv_cache_pool_pointers=metadata.host_kv_cache_pool_pointers,
host_kv_cache_pool_mapping=metadata.host_kv_cache_pool_mapping,
block_ids_per_seq=metadata.block_ids_per_seq,
workspace=metadata.
workspace, # re-enable it, if pass None to it, fp8 mla will encounter invalid cuda free issue.
# re-enable it, if pass None to it, fp8 mla will encounter invalid cuda free issue.
workspace=metadata.workspace
if not metadata.is_cuda_graph else metadata.cuda_graph_workspace,
cache_indirection=metadata.cache_indirection,
kv_scale_orig_quant=self.kv_scale_orig_quant,
kv_scale_quant_orig=self.kv_scale_quant_orig,
Expand Down