From 7bd154375dc505046a6e59e6d8c884a9c6b8fc0f Mon Sep 17 00:00:00 2001 From: Benjamin Chislett Date: Wed, 8 Jul 2026 04:59:13 +0900 Subject: [PATCH] [Bugfix] Fix mamba+dflash for MRV2 (#47698) --- vllm/v1/worker/gpu/spec_decode/dflash/speculator.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/vllm/v1/worker/gpu/spec_decode/dflash/speculator.py b/vllm/v1/worker/gpu/spec_decode/dflash/speculator.py index d5d68a0146..b2762244a4 100644 --- a/vllm/v1/worker/gpu/spec_decode/dflash/speculator.py +++ b/vllm/v1/worker/gpu/spec_decode/dflash/speculator.py @@ -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,