From e4d09ca6256f64d04663beff2f5cbeb678bafd12 Mon Sep 17 00:00:00 2001 From: Karol Damaszke Date: Wed, 2 Jul 2025 16:15:39 +0300 Subject: [PATCH] Enable block_softmax_adjustment on Gaudi2 --- vllm_hpu_extension/features.py | 1 - 1 file changed, 1 deletion(-) diff --git a/vllm_hpu_extension/features.py b/vllm_hpu_extension/features.py index c6234eedc..c71e724d5 100644 --- a/vllm_hpu_extension/features.py +++ b/vllm_hpu_extension/features.py @@ -53,7 +53,6 @@ def get_features(): Value('fp32_alibi_biases', True, env_var='VLLM_ALIBI_USE_FLOAT32_BIASES'), Value('fp32_softmax', ModelType('qwen2')), Value('fused_block_softmax_adjustment', All(VersionRange(">=1.22.0.494"), - Hardware('gaudi3'), Kernel(block_softmax_adjustment), Not(ModelType('qwen2')))), Value('flex_impl', False, env_var='VLLM_PROMPT_USE_FLEX_ATTENTION'),