[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:
Nick Hill
2026-07-29 01:09:41 +00:00
committed by GitHub
co-authored by Jeff Ma
parent 56f31af62a
commit 6fbbcf2151
5 changed files with 86 additions and 2 deletions
+5
View File
@@ -144,6 +144,11 @@ class BlockTable:
block_table_np = self.block_table.np
block_table_np[tgt, :num_blocks] = block_table_np[src, :num_blocks]
self.num_blocks_per_row[tgt] = num_blocks
# Clear the vacated source row: dummy-run batches dereference stale
# rows as mamba state slots and write state in place there, possibly
# after the blocks have been freed and reallocated.
block_table_np[src, :num_blocks] = 0
self.num_blocks_per_row[src] = 0
def swap_row(self, src: int, tgt: int) -> None:
src_tgt, tgt_src = [src, tgt], [tgt, src]
+7 -1
View File
@@ -163,7 +163,13 @@ class BlockTables:
# Therefore, this method must return the persistent tensor
# with the same memory address as that used during the model's forward pass,
# rather than allocating a new tensor.
return tuple(block_table[:num_reqs] for block_table in self.input_block_tables)
#
# Zero the rows so dummy runs write mamba state to the reserved null
# block rather than through the previous real step's (stale) block
# ids, which may point at blocks since freed and reallocated.
return tuple(
block_table[:num_reqs].zero_() for block_table in self.input_block_tables
)
def compute_slot_mappings(
self,
+4
View File
@@ -1290,6 +1290,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
slot_mappings,
self.attn_groups,
self.kv_cache_config,
# 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_capture=dummy_run and batch_desc.cg_mode == CUDAGraphMode.FULL,
)
input_ids = input_batch.input_ids
+7 -1
View File
@@ -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,
)