fix params issue

Signed-off-by: yuhangh <[email protected]>

Author: heyuhhh

cpp/tensorrt_llm/common/attentionOp.h

@@ -494,11 +494,11 @@ class AttentionOp
             mPosShiftEnabled, mPagedContextFMHA, mFP8ContextFMHA, mFP8AttenOutput, mFP8ContextMLA, mFP8GenerationMLA,
             mChunkPrefillBufferBatchSize, mDenseContextFMHA, mHasFullAttentionMask, mIsSpecDecodingEnabled,
             mUseSpecDecoding, mIsSpecDecTree, mSpecDecodingIsGenerationLengthVariable, mSpecDecodingMaxGenerationLength,
-            mIsMLAEnabled, mIsGenerationMLA, mUseGenFlashMLA, mUseSparseAttention, mMLAParams.data(), mCpSize, mCpRank,
-            mCpGroup, mNumAttnHeads, mNumAttnKVHeads, mNumKVHeadsOrigin, mAttnTpSize, mAttnTpRank, mAttnCpSize,
-            mAttnCpRank, mUlyssesMQABroadcast, mEnableContextFMHA, mFMHAForceFP32Acc, mMultiBlockMode, mEnableXQA,
-            mUseKVCache, mSkipAttn, mFuseFp4Quant, mRuntimeSparseAttentionParams.data(), mNbMultiBlockSemaphores,
-            mAttentionChunkSize.value_or(-1));
+            mIsMLAEnabled, mIsGenerationMLA, mUseGenFlashMLA, mUseSparseAttention, mUseTllmGenSparseAttention,
+            mMLAParams.data(), mCpSize, mCpRank, mCpGroup, mNumAttnHeads, mNumAttnKVHeads, mNumKVHeadsOrigin,
+            mAttnTpSize, mAttnTpRank, mAttnCpSize, mAttnCpRank, mUlyssesMQABroadcast, mEnableContextFMHA,
+            mFMHAForceFP32Acc, mMultiBlockMode, mEnableXQA, mUseKVCache, mSkipAttn, mFuseFp4Quant,
+            mRuntimeSparseAttentionParams.data(), mNbMultiBlockSemaphores, mAttentionChunkSize.value_or(-1));
     };
 
 private:

cpp/tensorrt_llm/thop/attentionOp.cpp

@@ -684,6 +684,8 @@ void attention(torch::Tensor q, torch::optional<torch::Tensor> k, torch::optiona
     op->mUseSpecDecoding = spec_decoding_bool_params[1];       // use_spec_decoding
     op->mIsSpecDecTree = spec_decoding_bool_params[2];         // is_spec_dec_tree
 
+    op->mUseSparseAttention = false;
+    op->mUseTllmGenSparseAttention = false;
     if ((sparse_kv_indices.has_value() && sparse_kv_indices.value().numel() > 0)
         || (sparse_attn_indices.has_value() && sparse_attn_indices.value().numel() > 0))
     {