Cleanup for PR

Signed-off-by: Dom Brown <[email protected]>

Author: DomBrown

cpp/CMakeLists.txt

@@ -358,7 +358,7 @@ endif()
 
 setup_sanitizers()
 
-set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --expt-extended-lambda -lineinfo")
+set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --expt-extended-lambda")
 set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --expt-relaxed-constexpr")
 if(FAST_MATH)
   set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --use_fast_math")

cpp/tensorrt_llm/common/attentionOp.cpp

@@ -30,7 +30,6 @@
 #include "tensorrt_llm/runtime/utils/mpiUtils.h"
 #include <algorithm>
 #include <cstdint>
-#include <cstdlib>
 #include <type_traits>
 
 using namespace tensorrt_llm::kernels;
@@ -286,12 +285,8 @@ bool AttentionOp::convertMMHAParamsToXQAParams(tensorrt_llm::kernels::XQAParams&
     xqaParams.fp4_out_sf_scale = generationsParams.attention_output_sf_scale;
     xqaParams.start_token_idx_sf = generationsParams.start_token_idx_sf;
 
-    xqaParams.num_tokens = generationsParams.num_tokens;
     // Cross attention parameters.
     xqaParams.encoder_input_lengths = generationsParams.encoder_input_lengths;
-    // xqaParams.cross_kv = generationsParams.cross_kv;
-    // xqaParams.cross_kv_length = generationsParams.cross_kv_length;
-    // xqaParams.num_encoder_tokens = generationsParams.num_encoder_tokens;
 
     return true;
 }
@@ -2239,7 +2234,7 @@ int AttentionOp::enqueueGeneration(EnqueueGenerationParams<T> const& params, cud
         }
         else
         {
-            TLLM_LOG_DEBUG("XQA kernels are not selected in the generation phase. mEnableXQA: %d", mEnableXQA);
+            TLLM_LOG_DEBUG("XQA kernels are not selected in the generation phase.");
         }
     }
 

cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/xqaParams.h

@@ -106,12 +106,8 @@ struct XQAParams
 
     void* quant_q_buffer_ptr = nullptr;
 
-    int num_tokens;
     // for cross attention
     int32_t const* encoder_input_lengths = nullptr;
-    // void const* cross_kv = nullptr;
-    // int32_t cross_kv_length = 0;
-    // int32_t num_encoder_tokens = 0;
 
     cudaStream_t stream = 0;
 
@@ -182,11 +178,7 @@ struct XQAParams
            << "total_num_input_tokens :" << total_num_input_tokens << std ::endl
            << "is_fp8_output :" << (is_fp8_output ? "true" : "false") << std ::endl
            << "fp8_out_scale :" << fp8_out_scale << std ::endl
-           << "encoder_input_lengths: " << encoder_input_lengths
-           << std::endl
-           //<< "cross_kv: " << cross_kv << std::endl
-           //<< "cross_kv_length: " << cross_kv_length << std::endl
-           //<< "num_encoder_tokens: " << num_encoder_tokens << std::endl
+           << "encoder_input_lengths: " << encoder_input_lengths << std::endl
            << "stream :" << stream;
 
         return ss.str();

cpp/tensorrt_llm/plugins/gptAttentionPlugin/gptAttentionPlugin.cpp

@@ -1041,7 +1041,6 @@ int GPTAttentionPlugin::enqueueSome(int32_t seqIdxBeg, int32_t localNbSeq, int32
 
         common_enqueue_params.input_seq_length = max_context_q_len;
         common_enqueue_params.max_past_kv_length = max_context_kv_len;
-
         EnqueueContextParams<T> enqueue_params{common_enqueue_params};
         enqueue_params.attention_packed_mask = attention_packed_mask;
         enqueue_params.host_block_offsets = host_block_offsets;