refactor: unify attention parameter structures and update function signatures for clarity.

Author: XuZhang99

xllm/core/kernels/param.h

@@ -34,10 +34,8 @@ struct RotaryParams {
   torch::Tensor cu_query_lens;
   bool interleaved;
   bool discrete;
-  bool dynamic_ntk;
+  bool dynamic_ntk = false;
   int max_query_len;
-
-  RotaryParams() : position_ids(std::nullopt), dynamic_ntk(false) {}
 };
 
 // Activation parameters
@@ -48,14 +46,8 @@ struct ActivationParams {
   std::optional<torch::Tensor> cusum_token_count;
   std::string act_mode;
   bool is_gated;
-  int start_expert_id;
-  int expert_size;
-
-  ActivationParams()
-      : bias(std::nullopt),
-        cusum_token_count(std::nullopt),
-        start_expert_id(0),
-        expert_size(0) {}
+  int start_expert_id = 0;
+  int expert_size = 0;
 };
 
 // Reshape paged cache parameters
@@ -65,82 +57,46 @@ struct ReshapePagedCacheParams {
   torch::Tensor k_cache;
   torch::Tensor v_cache;
   torch::Tensor slot_mapping;
-  bool direction;
-
-  ReshapePagedCacheParams() : direction(false) {}
+  bool direction = false;
 };
 
-// Prefill attention parameters
-struct PrefillAttentionParams {
+// Attention parameters
+struct AttentionParams {
+  // common parameters
   torch::Tensor query;
-  torch::Tensor key;
-  torch::Tensor value;
   torch::Tensor output;
   torch::Tensor output_lse;
-  torch::Tensor query_start_loc;
-  torch::Tensor seq_start_loc;
   std::optional<torch::Tensor> alibi_slope;
-  std::optional<torch::Tensor> attn_bias;
   std::optional<torch::Tensor> q_quant_scale;
-  std::optional<torch::Tensor> k_quant_scale;
-  std::optional<torch::Tensor> v_quant_scale;
   std::optional<torch::Tensor> out_quant_scale;
-  std::optional<torch::Tensor> block_tables;
-  int max_query_len;
+  std::optional<torch::Tensor> block_table;
+  std::string compute_dtype;
   int max_seq_len;
-  float scale;
-  bool is_causal;
   int window_size_left;
-  int window_size_right;
-  std::string compute_dtype;
-  bool return_lse;
-
-  FlashAttentionParams()
-      : alibi_slope(std::nullopt),
-        attn_bias(std::nullopt),
-        q_quant_scale(std::nullopt),
-        k_quant_scale(std::nullopt),
-        v_quant_scale(std::nullopt),
-        out_quant_scale(std::nullopt),
-        block_tables(std::nullopt),
-        is_causal(true),
-        window_size_right(-1),
-        return_lse(false) {}
-};
+  int window_size_right = -1;
+  float scale;
+  bool return_lse = false;
 
-// Decode attention parameters
-struct DecodeAttentionParams {
-  torch::Tensor query;
-  torch::Tensor k_cache;
-  torch::Tensor output;
-  torch::Tensor block_table;
-  torch::Tensor seq_lens;
-  torch::Tensor v_cache;
-  torch::Tensor output_lse;
-  std::optional<torch::Tensor> q_quant_scale;
+  // prefill parameters
+  std::optional<torch::Tensor> key;
+  std::optional<torch::Tensor> value;
+  std::optional<torch::Tensor> query_start_loc;
+  std::optional<torch::Tensor> seq_start_loc;
+  std::optional<torch::Tensor> attn_bias;
+  std::optional<torch::Tensor> k_quant_scale;
+  std::optional<torch::Tensor> v_quant_scale;
+  int max_query_len;
+  bool is_causal = true;
+
+  // decode parameters
+  std::optional<torch::Tensor> k_cache;
+  std::optional<torch::Tensor> v_cache;
+  std::optional<torch::Tensor> block_table;
+  std::optional<torch::Tensor> seq_lens;
   std::optional<torch::Tensor> k_cache_quant_scale;
   std::optional<torch::Tensor> v_cache_quant_scale;
-  std::optional<torch::Tensor> out_quant_scale;
-  std::optional<torch::Tensor> alibi_slope;
   std::optional<torch::Tensor> mask;
-  std::string compute_dtype;
-  int max_seq_len;
-  int window_size_left;
-  int window_size_right;
-  float scale;
-  bool return_lse;
-  int kv_cache_quant_bit_size;
-
-  DecodeAttentionParams()
-      : q_quant_scale(std::nullopt),
-        k_cache_quant_scale(std::nullopt),
-        v_cache_quant_scale(std::nullopt),
-        out_quant_scale(std::nullopt),
-        alibi_slope(std::nullopt),
-        mask(std::nullopt),
-        window_size_right(-1),
-        return_lse(false),
-        kv_cache_quant_bit_size(-1) {}
+  int kv_cache_quant_bit_size = -1;
 };
 
 // Fused layer norm parameters
@@ -157,21 +113,9 @@ struct FusedLayerNormParams {
   std::optional<torch::Tensor> normed_out;
   std::string mode;
   double eps;
-  bool store_output_before_norm;
-  bool store_output_after_norm;
-  bool dynamic_quant;
-
-  FusedLayerNormParams()
-      : residual(std::nullopt),
-        beta(std::nullopt),
-        bias(std::nullopt),
-        quant_scale(std::nullopt),
-        residual_out(std::nullopt),
-        smooth_quant_scale(std::nullopt),
-        normed_out(std::nullopt),
-        store_output_before_norm(false),
-        store_output_after_norm(false),
-        dynamic_quant(false) {}
+  bool store_output_before_norm = false;
+  bool store_output_after_norm = false;
+  bool dynamic_quant = false;
 };
 
 // Matmul parameters
@@ -182,8 +126,6 @@ struct MatmulParams {
   std::optional<torch::Tensor> c;
   double alpha;
   double beta;
-
-  MatmulParams() : bias(std::nullopt), c(std::nullopt) {}
 };
 
 // Fused MoE parameters
@@ -203,39 +145,17 @@ struct FusedMoEParams {
   bool renormalize;
   bool gated;
   std::string act_mode;
-  std::string scoring_func;
-  int start_expert_id;
-  int block_n;
-  bool avg_moe;
+  std::string scoring_func = "softmax";
+  int start_expert_id = 0;
+  int block_n = 0;
+  bool avg_moe = false;
   std::optional<torch::Tensor> class_reduce_weight;
   std::optional<torch::Tensor> class_expert_id;
   std::optional<std::vector<bool>> w1_quant_flag;
   std::optional<std::vector<bool>> w2_quant_flag;
-  int world_size;
-  int shared_expert_num;
-  std::string parallel_mode;
-
-  FusedMoEParams()
-      : bias1(std::nullopt),
-        bias2(std::nullopt),
-        residual(std::nullopt),
-        input_smooth(std::nullopt),
-        act_smooth(std::nullopt),
-        w1_scale(std::nullopt),
-        w2_scale(std::nullopt),
-        w1_quant_flag(std::nullopt),
-        w2_quant_flag(std::nullopt),
-        scoring_func("softmax"),
-        start_expert_id(0),
-        block_n(0),
-        avg_moe(false),
-        class_reduce_weight(std::nullopt),
-        class_expert_id(std::nullopt),
-        w1_quant_flag(std::nullopt),
-        w2_quant_flag(std::nullopt),
-        world_size(0),
-        shared_expert_num(0),
-        parallel_mode("ep") {}
+  int world_size = 0;
+  int shared_expert_num = 0;
+  std::string parallel_mode = "ep";
 };
 }  // namespace kernel
 }  // namespace xllm