From 61bd5fdcdaf3349e2a452445c34638df65f519e1 Mon Sep 17 00:00:00 2001 From: Iman Tabrizian <10105175+Tabrizian@users.noreply.github.com> Date: Tue, 19 May 2026 12:24:45 -0700 Subject: [PATCH] [None][fix] DSv4: drop duplicate indexer_k_dtype kwarg on V4 sparse attn indexer_k_dtype is already populated inside indexer_config by update_sparse_attention_indexer_config (which sources it from sparse_attention_config.indexer_k_dtype, falling back to the "fp8" default). Passing it again as an explicit kwarg to DeepSeekV4SparseAttentionConfig while also expanding **indexer_config triggers a duplicate keyword argument TypeError during executor init. Drop the explicit kwarg and let **indexer_config provide the value. Signed-off-by: Iman Tabrizian <10105175+Tabrizian@users.noreply.github.com> --- tensorrt_llm/_torch/model_config.py | 2 -- 1 file changed, 2 deletions(-) diff --git a/tensorrt_llm/_torch/model_config.py b/tensorrt_llm/_torch/model_config.py index 2ceab998b39e..308f079ea1e0 100644 --- a/tensorrt_llm/_torch/model_config.py +++ b/tensorrt_llm/_torch/model_config.py @@ -794,7 +794,6 @@ def update_sparse_attention_indexer_config(pretrained_config, kwargs): if checkpoint_window_size is None: checkpoint_window_size = getattr( pretrained_config, 'sliding_window', None) - indexer_k_dtype = sparse_attention_config.indexer_k_dtype if sparse_attention_config else 'fp8' if sparse_attention_config: compress_ratios = sparse_attention_config.compress_ratios window_size = sparse_attention_config.window_size @@ -838,7 +837,6 @@ def update_sparse_attention_indexer_config(pretrained_config, kwargs): 'sparse_attention_config'] = DeepSeekV4SparseAttentionConfig( compress_ratios=compress_ratios, window_size=window_size, - indexer_k_dtype=indexer_k_dtype, **indexer_config) else: raise ValueError(