Mila
Deep Neural Network Library
Loading...
Searching...
No Matches
Dnn.Components.LlamaTransformer Module Reference

Exported Modules

module  Serialization.PretrainedReader
module  Dnn.Components.TokenEmbedding
module  Dnn.ActivationType
module  Dnn.Components.RmsNorm
module  Compute.DeviceTypeTraits
module  Serialization.Tensor
module  Serialization.ModelArchive
module  Dnn.Components.Linear
module  Dnn.Quantization.Weight.Policies
module  Dnn.TensorDataTypeTraits
module  Compute.GqaState
module  Dnn.TensorDataType
module  Dnn.ComponentType
module  Compute.ExecutionContextFactory
module  Dnn.Components.Rope
module  Compute.DeviceId
module  Dnn.ModelType
module  Compute.Device
module  Dnn.Tensor
module  Compute.ExecutionContext
module  Dnn.LanguageNetwork
module  Compute.DeviceType
module  Dnn.ITensor
module  Dnn.TensorTypes
module  Compute.CpuMemoryResource
module  Dnn.Component
module  Dnn.Quantization.KvCache.Policy

Classes

class  Mila::Dnn::LlamaBlock< TDeviceType, TPrecision, TWeightQuant, TKvPolicy >
class  Mila::Dnn::LlamaConfig
 Network-level configuration for LLaMA-style transformer networks. More...
class  Mila::Dnn::LlamaTransformer< TDeviceType, TPrecision, TWeightQuantization, TKvCachePolicy >
 LLaMA-style transformer (decoder-only) for autoregressive token prediction. More...

Functions

int64_t Mila::Dnn::computePrefillChunkSize (int64_t batch, int64_t num_heads, int64_t head_dim, int64_t context_length, int64_t precision_bytes)
LlamaConfig Mila::Dnn::Llama2_13B ()
 Llama 2 13B.
LlamaConfig Mila::Dnn::Llama2_70B ()
 Llama 2 70B.
LlamaConfig Mila::Dnn::Llama2_7B ()
 Llama 2 7B.
LlamaConfig Mila::Dnn::Llama3_1_405B ()
 Llama 3.1 405B.
LlamaConfig Mila::Dnn::Llama3_1_70B ()
 Llama 3.1 70B.
LlamaConfig Mila::Dnn::Llama3_1_8B ()
 Llama 3.1 8B.
LlamaConfig Mila::Dnn::Llama3_2_1B ()
 Usage Examples:
LlamaConfig Mila::Dnn::Llama3_2_3B ()
 Llama 3.2 3B.
LlamaConfig Mila::Dnn::Llama3_70B ()
 Llama 3 70B (Original release).
LlamaConfig Mila::Dnn::Llama3_8B ()
 Llama 3 8B (Original release).

Variables

constexpr int64_t Mila::Dnn::kPrefillScratchByteCap = int64_t{ 1536 } * 1024 * 1024

Files

file  Mila/Src/Dnn/Components/Transformers/LlaMa/Llama.ixx
 LLaMA-style decoder-only transformer network.
file  Mila/Src/Dnn/Components/Transformers/LlaMa/Llama.Block.ixx
 LLaMA transformer block – module partition of LlamaTransformer.
file  Mila/Src/Dnn/Components/Transformers/LlaMa/Llama.Config.ixx
 LLaMA network-level configuration.
file  Mila/Src/Dnn/Components/Transformers/LlaMa/Llama.Presets.ixx