From cee1caaf476f989329232f3217a620a53cae315e Mon Sep 17 00:00:00 2001 From: Thomas Wang Date: Mon, 22 Jun 2026 22:35:45 +0800 Subject: [PATCH] [AMD] Fix nightly-8-gpu-mi35x-deepseek-v4-flash-rocm720 OOM issue (#28941) --- .../srt/mem_cache/deepseek_v4_memory_pool.py | 44 +++++++++---------- 1 file changed, 22 insertions(+), 22 deletions(-) diff --git a/python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py b/python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py index d521ac425..c9219bd4d 100644 --- a/python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py +++ b/python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py @@ -578,29 +578,29 @@ class DeepSeekV4TokenToKVPool(BaseSWAKVPool): global_page_size=swa_page_size, ) - c4_kv_pool_type = DeepSeekV4SingleKVPool - if enable_hisparse: - c4_kv_pool_type = HiSparseC4DevicePool - self.c4_kv_pool = self._make_kv_pool( - size=c4_size, - page_size=c4_page_size, - dtype=dtype, - layer_num=c4_layer_num, - device=device, - enable_memory_saver=enable_memory_saver, - global_page_size=page_size, - cls=c4_kv_pool_type, - ) + c4_kv_pool_type = DeepSeekV4SingleKVPool + if enable_hisparse: + c4_kv_pool_type = HiSparseC4DevicePool + self.c4_kv_pool = self._make_kv_pool( + size=c4_size, + page_size=c4_page_size, + dtype=dtype, + layer_num=c4_layer_num, + device=device, + enable_memory_saver=enable_memory_saver, + global_page_size=page_size, + cls=c4_kv_pool_type, + ) - self.c128_kv_pool = self._make_kv_pool( - size=c128_size, - page_size=c128_page_size, - dtype=dtype, - layer_num=c128_layer_num, - device=device, - enable_memory_saver=enable_memory_saver, - global_page_size=page_size, - ) + self.c128_kv_pool = self._make_kv_pool( + size=c128_size, + page_size=c128_page_size, + dtype=dtype, + layer_num=c128_layer_num, + device=device, + enable_memory_saver=enable_memory_saver, + global_page_size=page_size, + ) indexer_size = ( self.c4_logical_size