Attention projections (QKV, O) disaggregation

.gitignore

@@ -193,3 +193,6 @@ lora_training_logs
 Untitled-1.ipynb
 Untitled-2.ipynb
 tests/inference/python_test_configs/*.json
+
+core.*
+fine_grained_alignment_config.json

examples/python/native/ops/inc_multihead_self_attention.py

@@ -11,8 +11,6 @@ def test_inc_multihead_self_attention(
         kdim: int = 0,
         vdim: int = 0,
         dropout: float = 0.0,
-        bias: bool = True,
-        add_bias_kv: bool = False,
         add_zero_attn: bool = False,
         data_type: DataType = DataType.DT_NONE,
         kernel_initializer=None,
@@ -34,8 +32,6 @@ def test_inc_multihead_self_attention(
         kdim=kdim,
         vdim=vdim,
         dropout=dropout,
-        bias=bias,
-        add_bias_kv=add_bias_kv,
         add_zero_attn=add_zero_attn,
         data_type=data_type,
         kernel_initializer=kernel_initializer,
@@ -85,8 +81,6 @@ def test_inc_multihead_self_attention(
         kdim=0,  # Example value for kdim
         vdim=0,  # Example value for vdim
         dropout=0.1,  # Example value for dropout
-        bias=True,
-        add_bias_kv=False,
         add_zero_attn=False,
         data_type=DataType.DT_FLOAT,
         kernel_initializer=None,  # Example value for kernel_initializer

examples/python/native/ops/inc_multihead_self_attention_verify.py

@@ -11,8 +11,6 @@ def test_inc_multihead_self_attention_verify(
         kdim: int = 0,
         vdim: int = 0,
         dropout: float = 0.0,
-        bias: bool = True,
-        add_bias_kv: bool = False,
         add_zero_attn: bool = False,
         data_type: DataType = DataType.DT_NONE,
         kernel_initializer=None,
@@ -34,8 +32,6 @@ def test_inc_multihead_self_attention_verify(
         kdim=kdim,
         vdim=vdim,
         dropout=dropout,
-        bias=bias,
-        add_bias_kv=add_bias_kv,
         add_zero_attn=add_zero_attn,
         data_type=data_type,
         kernel_initializer=kernel_initializer,
@@ -85,8 +81,6 @@ def test_inc_multihead_self_attention_verify(
         kdim=0,  # Example value for kdim
         vdim=0,  # Example value for vdim
         dropout=0.1,  # Example value for dropout
-        bias=True,
-        add_bias_kv=False,
         add_zero_attn=False,
         data_type=DataType.DT_FLOAT,
         kernel_initializer=None,  # Example value for kernel_initializer

examples/python/native/ops/inc_multiquery_self_attention.py

@@ -12,8 +12,6 @@ def test_inc_multiquery_self_attention(
         kdim: int = 0,
         vdim: int = 0,
         dropout: float = 0.0,
-        bias: bool = True,
-        add_bias_kv: bool = False,
         add_zero_attn: bool = False,
         data_type: DataType = DataType.DT_NONE,
         kernel_initializer=None,
@@ -36,8 +34,6 @@ def test_inc_multiquery_self_attention(
         kdim=kdim,
         vdim=vdim,
         dropout=dropout,
-        bias=bias,
-        add_bias_kv=add_bias_kv,
         add_zero_attn=add_zero_attn,
         data_type=data_type,
         kernel_initializer=kernel_initializer,
@@ -89,8 +85,6 @@ def test_inc_multiquery_self_attention(
         kdim=0,  # Example value for kdim
         vdim=0,  # Example value for vdim
         dropout=0.1,  # Example value for dropout
-        bias=True,
-        add_bias_kv=False,
         add_zero_attn=False,
         data_type=DataType.DT_FLOAT,
         kernel_initializer=None,  # Example value for kernel_initializer

examples/python/native/ops/inc_multiquery_self_attention_verify.py

@@ -12,8 +12,6 @@ def test_inc_multiquery_self_attention_verify(
         kdim: int = 0,
         vdim: int = 0,
         dropout: float = 0.0,
-        bias: bool = True,
-        add_bias_kv: bool = False,
         add_zero_attn: bool = False,
         data_type: DataType = DataType.DT_NONE,
         kernel_initializer=None,
@@ -36,8 +34,6 @@ def test_inc_multiquery_self_attention_verify(
         kdim=kdim,
         vdim=vdim,
         dropout=dropout,
-        bias=bias,
-        add_bias_kv=add_bias_kv,
         add_zero_attn=add_zero_attn,
         data_type=data_type,
         kernel_initializer=kernel_initializer,
@@ -89,8 +85,6 @@ def test_inc_multiquery_self_attention_verify(
         kdim=0,  # Example value for kdim
         vdim=0,  # Example value for vdim
         dropout=0.1,  # Example value for dropout
-        bias=True,
-        add_bias_kv=False,
         add_zero_attn=False,
         data_type=DataType.DT_FLOAT,
         kernel_initializer=None,  # Example value for kernel_initializer

examples/python/native/ops/spec_inc_multihead_self_attention.py

@@ -11,8 +11,6 @@ def test_spec_inc_multihead_self_attention(
         kdim: int = 0,
         vdim: int = 0,
         dropout: float = 0.0,
-        bias: bool = True,
-        add_bias_kv: bool = False,
         add_zero_attn: bool = False,
         data_type: DataType = DataType.DT_NONE,
         kernel_initializer=None,
@@ -34,8 +32,6 @@ def test_spec_inc_multihead_self_attention(
         kdim=kdim,
         vdim=vdim,
         dropout=dropout,
-        bias=bias,
-        add_bias_kv=add_bias_kv,
         add_zero_attn=add_zero_attn,
         data_type=data_type,
         kernel_initializer=kernel_initializer,
@@ -85,8 +81,6 @@ def test_spec_inc_multihead_self_attention(
         kdim=0,  # Example value for kdim
         vdim=0,  # Example value for vdim
         dropout=0.1,  # Example value for dropout
-        bias=True,
-        add_bias_kv=False,
         add_zero_attn=False,
         data_type=DataType.DT_FLOAT,
         kernel_initializer=None,  # Example value for kernel_initializer

examples/python/native/ops/spec_inc_multiquery_self_attention.py

@@ -12,8 +12,6 @@ def test_spec_inc_multiquery_self_attention(
         kdim: int = 0,
         vdim: int = 0,
         dropout: float = 0.0,
-        bias: bool = True,
-        add_bias_kv: bool = False,
         add_zero_attn: bool = False,
         data_type: DataType = DataType.DT_NONE,
         kernel_initializer=None,
@@ -36,8 +34,6 @@ def test_spec_inc_multiquery_self_attention(
         kdim=kdim,
         vdim=vdim,
         dropout=dropout,
-        bias=bias,
-        add_bias_kv=add_bias_kv,
         add_zero_attn=add_zero_attn,
         data_type=data_type,
         kernel_initializer=kernel_initializer,
@@ -89,8 +85,6 @@ def test_spec_inc_multiquery_self_attention(
         kdim=0,  # Example value for kdim
         vdim=0,  # Example value for vdim
         dropout=0.1,  # Example value for dropout
-        bias=True,
-        add_bias_kv=False,
         add_zero_attn=False,
         data_type=DataType.DT_FLOAT,
         kernel_initializer=None,  # Example value for kernel_initializer

include/flexflow/flexflow_c.h

@@ -445,12 +445,16 @@ flexflow_tensor_t flexflow_model_add_inc_multihead_self_attention(
     int kdim,
     int vdim,
     float dropout,
-    bool bias,
-    bool add_bias_kv,
     bool add_zero_attn,
     enum DataType data_type,
     flexflow_initializer_t kernel_initializer_,
     bool apply_rotary_embedding,
+    float rope_theta,
+    char const *rope_type,
+    float rope_factor,
+    float low_freq_factor,
+    float high_freq_factor,
+    int original_max_position_embeddings,
     bool scaling_query,
     float scaling_factor,
     bool qk_prod_scaling,
@@ -465,12 +469,16 @@ flexflow_tensor_t flexflow_model_add_spec_inc_multihead_self_attention(
     int kdim,
     int vdim,
     float dropout,
-    bool bias,
-    bool add_bias_kv,
     bool add_zero_attn,
     enum DataType data_type,
     flexflow_initializer_t kernel_initializer_,
     bool apply_rotary_embedding,
+    float rope_theta,
+    char const *rope_type,
+    float rope_factor,
+    float low_freq_factor,
+    float high_freq_factor,
+    int original_max_position_embeddings,
     bool scaling_query,
     float scaling_factor,
     bool qk_prod_scaling,
@@ -485,12 +493,16 @@ flexflow_tensor_t flexflow_model_add_inc_multihead_self_attention_verify(
     int kdim,
     int vdim,
     float dropout,
-    bool bias,
-    bool add_bias_kv,
     bool add_zero_attn,
     enum DataType data_type,
     flexflow_initializer_t kernel_initializer_,
     bool apply_rotary_embedding,
+    float rope_theta,
+    char const *rope_type,
+    float rope_factor,
+    float low_freq_factor,
+    float high_freq_factor,
+    int original_max_position_embeddings,
     bool scaling_query,
     float scaling_factor,
     bool qk_prod_scaling,
@@ -506,12 +518,16 @@ flexflow_tensor_t flexflow_model_add_inc_multiquery_self_attention(
     int kdim,
     int vdim,
     float dropout,
-    bool bias,
-    bool add_bias_kv,
     bool add_zero_attn,
     enum DataType data_type,
     flexflow_initializer_t kernel_initializer_,
     bool apply_rotary_embedding,
+    float rope_theta,
+    char const *rope_type,
+    float rope_factor,
+    float low_freq_factor,
+    float high_freq_factor,
+    int original_max_position_embeddings,
     bool scaling_query,
     float scaling_factor,
     bool qk_prod_scaling,
@@ -527,12 +543,16 @@ flexflow_tensor_t flexflow_model_add_spec_inc_multiquery_self_attention(
     int kdim,
     int vdim,
     float dropout,
-    bool bias,
-    bool add_bias_kv,
     bool add_zero_attn,
     enum DataType data_type,
     flexflow_initializer_t kernel_initializer_,
     bool apply_rotary_embedding,
+    float rope_theta,
+    char const *rope_type,
+    float rope_factor,
+    float low_freq_factor,
+    float high_freq_factor,
+    int original_max_position_embeddings,
     bool scaling_query,
     float scaling_factor,
     bool qk_prod_scaling,
@@ -548,12 +568,16 @@ flexflow_tensor_t flexflow_model_add_inc_multiquery_self_attention_verify(
     int kdim,
     int vdim,
     float dropout,
-    bool bias,
-    bool add_bias_kv,
     bool add_zero_attn,
     enum DataType data_type,
     flexflow_initializer_t kernel_initializer_,
     bool apply_rotary_embedding,
+    float rope_theta,
+    char const *rope_type,
+    float rope_factor,
+    float low_freq_factor,
+    float high_freq_factor,
+    int original_max_position_embeddings,
     bool scaling_query,
     float scaling_factor,
     bool qk_prod_scaling,

include/flexflow/inference.h

@@ -43,8 +43,43 @@ struct GenerationResult {
   std::vector<float> finetuning_losses;
 };
 
-#include <string>
-#include <vector>
+struct RotaryEmbeddingMeta {
+  bool apply_rotary_embedding = false;
+  float rope_theta = 10000.0f;
+  std::string rope_type = "default";
+  float factor = 8.0f;
+  float low_freq_factor = 1.0f;
+  float high_freq_factor = 4.0f;
+  int original_max_position_embeddings = 8192;
+
+  RotaryEmbeddingMeta(bool apply_rotary_embedding_ = false,
+                      float rope_theta_ = 10000.0f,
+                      std::string rope_type_ = "default",
+                      float factor_ = 8.0f,
+                      float low_freq_factor_ = 1.0f,
+                      float high_freq_factor_ = 4.0f,
+                      int original_max_position_embeddings_ = 8192)
+      : apply_rotary_embedding(apply_rotary_embedding_),
+        rope_theta(rope_theta_), rope_type(rope_type_), factor(factor_),
+        low_freq_factor(low_freq_factor_), high_freq_factor(high_freq_factor_),
+        original_max_position_embeddings(original_max_position_embeddings_) {}
+
+  friend std::ostream &operator<<(std::ostream &os,
+                                  RotaryEmbeddingMeta const &meta) {
+    os << std::boolalpha // To print bool as true/false instead of 1/0
+       << "RotaryEmbeddingMeta {\n"
+       << "  apply_rotary_embedding: " << meta.apply_rotary_embedding << ",\n"
+       << "  rope_theta: " << meta.rope_theta << ",\n"
+       << "  rope_type: \"" << meta.rope_type << "\",\n"
+       << "  factor: " << meta.factor << ",\n"
+       << "  low_freq_factor: " << meta.low_freq_factor << ",\n"
+       << "  high_freq_factor: " << meta.high_freq_factor << ",\n"
+       << "  original_max_position_embeddings: "
+       << meta.original_max_position_embeddings << "\n"
+       << "}";
+    return os;
+  }
+};
 
 std::string join_path(std::vector<std::string> const &paths);
 

include/flexflow/layer.h

@@ -32,11 +32,13 @@ class Layer {
   void add_float_property(std::string const &key, float value);
   void add_int_vector_property(std::string const &key,
                                std::vector<int> const &value);
+  void add_string_property(std::string const &key, std::string const &value);
   void add_initializer(std::string const &key, Initializer *initializer);
   bool get_int_property(std::string const &key, long long &value) const;
   bool get_float_property(std::string const &key, float &value) const;
   bool get_int_vector_property(std::string const &key,
                                std::vector<int> &value) const;
+  bool get_string_property(std::string const &key, std::string &value) const;
   bool get_initializer(std::string const &key, Initializer *&initializer) const;
   Tensor get_parameter(int index);
   void print();
@@ -59,6 +61,7 @@ class Layer {
   std::unordered_map<std::string, float> float_properties;
   std::unordered_map<std::string, Initializer *> initializers;
   std::unordered_map<std::string, std::vector<int>> int_vector_properties;
+  std::unordered_map<std::string, std::string> string_properties;
 };
 
 }; // namespace FlexFlow