From 68ca6fd02c6cd705c4fa6872317611c83b26b4c6 Mon Sep 17 00:00:00 2001 From: Benjamin Chislett Date: Mon, 3 Aug 2026 10:21:50 -0400 Subject: [PATCH] [Bugfix] Remove bad startup assertion (#50869) Signed-off-by: Benjamin Chislett --- vllm/config/speculative.py | 25 ------------------------- 1 file changed, 25 deletions(-) diff --git a/vllm/config/speculative.py b/vllm/config/speculative.py index 07bdcac68b..aa3fc67ea1 100644 --- a/vllm/config/speculative.py +++ b/vllm/config/speculative.py @@ -1032,31 +1032,6 @@ class SpeculativeConfig: "Inkling MTP currently supports exactly one speculative token" ) - if self.method == "dspark": - # DSpark is a semi-autoregressive *block* drafter. A - # speculative length smaller than the checkpoint's block - # feeds the block / Markov-head machinery an unsupported - # layout and yields incorrect (garbled) output rather than - # merely lower acceptance. Require num_speculative_tokens to - # be at least the block size (e.g. 5 or 7 for DeepSeek-V4). - dspark_block_size = getattr( - self.draft_model_config.hf_config, - "dspark_block_size", - None, - ) - if ( - dspark_block_size is not None - and self.num_speculative_tokens < dspark_block_size - ): - raise ValueError( - "DSpark requires num_speculative_tokens >= " - f"dspark_block_size ({dspark_block_size}); got " - f"{self.num_speculative_tokens}. Smaller values " - "produce incorrect output. Use " - f"num_speculative_tokens={dspark_block_size} or " - "larger (e.g. 7)." - ) - self.draft_tensor_parallel_size = ( SpeculativeConfig._verify_and_get_draft_tp( self.target_parallel_config,