This commit is contained in:
@@ -5376,6 +5376,12 @@ class GPUModelRunner(
|
||||
self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
|
||||
self.query_start_loc.copy_to_gpu()
|
||||
|
||||
# Sync block table CPU->GPU so cleared rows from
|
||||
# remove_request() are visible to the attention metadata
|
||||
# builder. Without this, stale block IDs from finished
|
||||
# requests can corrupt Mamba state.
|
||||
self.input_batch.block_table.commit_block_table(num_reqs_padded)
|
||||
|
||||
pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
|
||||
attn_metadata, _ = self._build_attention_metadata(
|
||||
num_tokens=num_tokens_unpadded,
|
||||
|
||||
Reference in New Issue
Block a user