[Bugfix] Fix mamba+dflash for MRV2 (#47698)

This commit is contained in:
Benjamin Chislett
2026-07-07 15:59:13 -04:00
committed by GitHub
parent 2f3f441f84
commit 7bd154375d
@@ -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,