[Bugfix] Fix mamba+dflash for MRV2 (#47698)
This commit is contained in:
@@ -135,9 +135,6 @@ class DFlashSpeculator(DraftModelSpeculator):
|
||||
]
|
||||
assert self.draft_kv_cache_group_ids, "No draft attention groups found."
|
||||
self.draft_kv_cache_group_id = self.draft_kv_cache_group_ids[0]
|
||||
self.draft_block_size = self.block_tables.block_sizes[
|
||||
self.draft_kv_cache_group_id
|
||||
]
|
||||
|
||||
# Per-group context slot buffers for the precompute (one row per group).
|
||||
self._context_slot_mappings = torch.zeros(
|
||||
@@ -337,7 +334,7 @@ class DFlashSpeculator(DraftModelSpeculator):
|
||||
last_sampled,
|
||||
next_prefill_tokens,
|
||||
self.block_tables.input_block_tables[gid],
|
||||
self.block_tables.block_sizes[gid],
|
||||
self.block_tables.kernel_block_sizes[gid],
|
||||
self.parallel_drafting_token_id,
|
||||
self.num_query_per_req,
|
||||
self.num_speculative_steps,
|
||||
|
||||
Reference in New Issue
Block a user