[BugFix] MLA + V1, illegal memory access and accuracy issues (#14253)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
This commit is contained in:
@@ -456,8 +456,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
|
||||
# Some attention backends (namely MLA) may want to separate requests
|
||||
# based on if the attention computation will be compute-bound or
|
||||
# memory-bound. This gives them a hook to do that.
|
||||
self.attn_metadata_builder.reorder_batch(self.input_batch,
|
||||
scheduler_output)
|
||||
modified_batch = self.attn_metadata_builder.reorder_batch(
|
||||
self.input_batch, scheduler_output)
|
||||
if modified_batch:
|
||||
self.input_batch.refresh_sampling_metadata()
|
||||
|
||||
# OPTIMIZATION: Start copying the block table first.
|
||||
# This way, we can overlap the copy with the following CPU operations.
|
||||
|
||||
Reference in New Issue
Block a user