[BugFix] Stop dummy runs from writing mamba state through stale block-table rows (#49757)
Signed-off-by: Nick Hill <nickhill123@gmail.com> Signed-off-by: Jeff Ma <jeffjma@umich.edu> Co-authored-by: Jeff Ma <jeffjma@umich.edu>
This commit is contained in:
@@ -6064,7 +6064,13 @@ class GPUModelRunner(
|
||||
num_reqs=num_reqs_padded,
|
||||
max_query_len=max_query_len,
|
||||
ubatch_slices=(ubatch_slices_padded if pad_attn else ubatch_slices),
|
||||
for_cudagraph_capture=is_graph_capturing,
|
||||
# FULL replay reads capture-time metadata buffers. Re-stage them
|
||||
# from the zeroed dummy block tables instead of retaining state
|
||||
# indices from the previous real batch.
|
||||
for_cudagraph_capture=(
|
||||
is_graph_capturing
|
||||
or cudagraph_runtime_mode == CUDAGraphMode.FULL
|
||||
),
|
||||
slot_mappings=slot_mappings_by_group,
|
||||
use_spec_decode=self.speculative_config is not None,
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user