We've delved into all details and internals of all stages. We've covered all steps of the complete end-to-end flow of our project, except one, I saved it for last: The LlamaTransformer.Forward(...) method.
If you are familiar with PyTorch framework, you should be aware of what the forward function means. In this project, this naming style was kept, it means, this function is the real place for performing calculations.
In this chapter, following 15. MAKING PREDICTION with LLAMA MODEL - 2, and 16. MAKING PREDICTION with LLAMA MODEL - 3 chapters, we walk through the LlamaTransformer.Forward(...) method and its components.
Important note: The given example tensor shapes and values are for the first iteration that the prompt tokens is processed at, for this input prompt string:
promptString = "<|begin_of_text|><|start_header_id|>system<|end_header_id|>
You are Einstein<|eot_id|><|start_header_id|>user<|end_header_id|>
Describe your theory.<|eot_id|><|start_header_id|>assistant<|end_header_id|>
"
// the shape of original inputTokens is {200},
// prompt tokens length is 22,
// shape of inputTokens argument is {22}, because it is the part of inputTokens containing prompt tokens.from src/model/llamatransformer.go
func (lt *LlamaTransformer) Forward(infContext *InferenceContext, inputTokens *ml.Tensor, startPos int) (*ml.Tensor, error) {
if inputTokens.Size[0] == 0 {
return nil, fmt.Errorf("empty token array")
}
common.GLogger.DebugPrintf("LlamaTransformer.Forward started for inputTokens: shape(%v), startPos: %d -> tensor inputTensor", inputTokens.Size, startPos)
...
}We can see output lines in the "debug.log" file if debugging is enabled, as follows:
[DEBUG] ... Created input tokens tensor: shape([200]) ...
[DEBUG] ... =======================================
[DEBUG] ... Running Transformer.Forward for curPos: 22, prevPos: 0, inputTokensSlice: shape([22])...Our LlamaTransformer.Forward(...) method calls the LlamaTransformer.prepare(...) method, to make it to create:
inputTensor: The tensor contains embedding vectors of the token ids ininputTokenstensor, with shape of{22, 4096}(remember theinputTokensSlicehas shape of{22}in our first iteration example case and4096is one dimension size of Llama model's embedding layer),freqsCis: The tensor contains a slice ofLlamaTransformer.PrecomputedFreqsCistaken by between indices 0 and 21. TheLlamaTransformer.PrecomputedFreqsCistensor was precomputed before and we have covered this subject in previous chapters. The shape of thefreqsCistensor is{22, 64}in our first iteration example case,mask: A triangular upper tensor consists of "negative infinities" and "ones (1s)", with shape of{22, 22}in our first iteration example case,err: If the function finishes with an error, this represents it, otherwisenil.
from src/model/llamatransformer.go
func (lt *LlamaTransformer) Forward(infContext *InferenceContext, inputTokens *ml.Tensor, startPos int) (*ml.Tensor, error) {
...
inputTensor, freqsCis, mask, err := lt.prepare(inputTokens, startPos)
...
}from src/model/llamatransformer.go
func (lt *LlamaTransformer) prepare(inputTokens *ml.Tensor, startPos int) (inputTensor *ml.Tensor, freqsCis *ml.Tensor, mask *ml.Tensor, err error) {
sequenceLength := inputTokens.Size[0]
common.GLogger.DebugPrintf("LlamaTransformer.prepare started for inputTokens: shape(%v), startPos: %d. sequenceLength: %d", inputTokens.Size, startPos, sequenceLength)
inputTensor, err = ml.Fwd_Get_Rows(lt.tok_embd, inputTokens)
...
}
Diagram: Creating inputTensor. For the complete diagram, click here.
The inputTensor is created via ml.Fwd_Get_Rows function.
This step aims to retrieve embedding vectors of the token ids in inputTokens tensor (in our case it contains 32 token id integers), the result will be a tensor with shape of {22, 4096}.
We have the embedding layer tensor at lt.tok_embd already loaded to LlamaTransformer struct. In our case, shape of our embedding layer is {128256, 4096}, which has 128256 rows corresponding to supported tokens in our vocabulary, 4096 columns corresponding to dimensions of one token embedding vector.
This function simply takes token embedding vectors with {1, 4096} shape by using our token ids as item indices/row indices in a list. It finds start and end indices at the embedding.RawData byte array of corresponding row, for each token independently, then collects all of them into dst tensor.
from src/ml/operations_impl.go
func Fwd_Get_Rows(embedding *Tensor, tokens *Tensor) (*Tensor, error) {
if err := processErrors(
checkIsMatrix(embedding),
checkIsVector(tokens),
); err != nil {
return nil, err
}
if tokens.DataType != DT_INT32 {
return nil, fmt.Errorf("tensor is not in data type %s: \"%s\" is %s", DT_INT32, tokens.Name, tokens.DataType)
}
sequenceLength := tokens.Size[0]
embeddingDim := embedding.Size[1]
dst := NewEmptyTensor([]int{sequenceLength, embeddingDim}, embedding.DataType)
rowCount := tokens.GetElementCount()
for rowIdx := 0; rowIdx < rowCount; rowIdx++ {
rowVal, err := tokens.GetItem([]int{rowIdx})
if err != nil {
return nil, err
}
row := int(rowVal.(int32))
readOffsetStart := embedding.calculateByteOffset([]int{row, 0})
readOffsetEnd := embedding.calculateByteOffset([]int{row + 1, 0})
rowBytes := embedding.RawData[readOffsetStart:readOffsetEnd]
writeOffsetStart := dst.calculateByteOffset([]int{rowIdx, 0})
copy(dst.RawData[writeOffsetStart:], rowBytes)
}
return dst, nil
}We can see output lines in the "debug.log" file if debugging is enabled, as follows:
[DEBUG] ... LlamaTransformer.Forward started for inputTokens: shape([22]), startPos: 0 -> tensor inputTensor...
[DEBUG] ... LlamaTransformer.prepare started for inputTokens: shape([22]), startPos: 0. sequenceLength: 22...from src/model/llamatransformer.go
func (lt *LlamaTransformer) prepare(inputTokens *ml.Tensor, startPos int) (inputTensor *ml.Tensor, freqsCis *ml.Tensor, mask *ml.Tensor, err error) {
...
freqsCis, err = lt.PrecomputedFreqsCis.Slice([]int{startPos}, []int{startPos + sequenceLength})
...
}We've precomputed the lt.PrecomputedFreqsCis in previous steps. It has the shape of {4096, 64}, which has 4096 rows corresponding to 2 times of 2048 max sequence length, 64 columns corresponding to half of dimensions of dim / N_Heads (head count) = 4096 / 32 = 128. It is 64, because we take 128 numbers as pairs to create one complex number, one as real part and other one as imaginary part. For more information, refer to 10. ROPE ROTARY POSITIONAL EMBEDDINGS and 10.BONUS. PRECOMPUTING FREQUENCY TENSOR.
Here, we take a slice of the lt.PrecomputedFreqsCis into freqsCis with shape of {22, 64} in our case.
from src/model/llamatransformer.go
func (lt *LlamaTransformer) prepare(inputTokens *ml.Tensor, startPos int) (inputTensor *ml.Tensor, freqsCis *ml.Tensor, mask *ml.Tensor, err error) {
...
if sequenceLength > 1 {
negativeInfinity := dtype.BFloat16fromFloat32(float32(math.Inf(-1)))
if mask, err = ml.Full([]int{sequenceLength, sequenceLength}, ml.DT_BF16, negativeInfinity); err != nil {
return
}
if mask, err = ml.TriangularUpper(mask, 1); err != nil {
return
}
}
var maskSize []int
if mask != nil {
maskSize = mask.Size
}
common.GLogger.DebugPrintf("LlamaTransformer.prepare finished inputTensor: shape(%v), freqsCis: shape(%v), mask: shape(%v)", inputTensor.Size, freqsCis.Size, maskSize)
return
}In Wikipedia, it writes:
Llama (Large Language Model Meta AI) is a family of autoregressive large language models (LLMs), released by Meta AI starting in February 2023.Llama is an auto-regressive model, which means, remember this is a time-series model, it always handles the information behind the current position. So, if we add -Inf (negative infinity) value to the values we want to ignore and add 0 (zero) to the values we want to pay attention to, our goal could be achieved.
To achieve this goal, we generate a mask tensor consisting of -Inf values at the upper triangle and 0 values at the lower triangle via ml.TriangularUpper function. This tensor will be used in our attention layer further.
[
[ 0, -Inf, -Inf, ..., -Inf, -Inf, -Inf],
[ 0, 0, -Inf, ..., -Inf, -Inf, -Inf],
[ 0, 0, 0, ..., -Inf, -Inf, -Inf],
...,
[ 0, 0, 0, ..., 0, -Inf, -Inf],
[ 0, 0, 0, ..., 0, 0, -Inf],
[ 0, 0, 0, ..., 0, 0, 0]
]
shape=[22 22], dtype=BF16Sources:
We can see output lines in the "debug.log" file if debugging is enabled, as follows:
[DEBUG] ... LlamaTransformer.prepare finished inputTensor: shape([22 4096]), freqsCis: shape([22 64]), mask: shape([22 22])...Recap:
The most important part of the transformer models that provide accurate outputs is the attention mechanism. Each "block" of Llama consists of a self-attention and a feed-forward neural network parts. The details will be explained further, but also we call these "block"s as "layer"s.A
LlamaTransformerBlockobject consists ofattn_norm(RMS normalization),attention(Attention mechanism),ffn_norm(RMS normalization), andfeedForward(Feed Forward Neural Network) modules. These modules operate respectively.The model configuration file "params.json" is parsed as ModelArgs object via JSON parser.
Diagram: Forward Pass Through Each Transformer Block. For the complete diagram, click here.
These "blocks" are also called as "layer"s. In our case, we use Llama 3.1 8B-Instruct model, in 8B model has 32 layers. Because we have model.ModelArgs.N_Layers = 32 as read from "params.json" configuration file. We've defined these "layers"/"transformer block layers" in previous steps as described in 09. IMPLEMENTING LLAMA MODEL ARCHITECTURE.
We initiate a for loop to iterate over 32 layers defined at lt.Layers and call LlamaTransformerBlock.Forward(...) method.
Important note: These 32 layers couldn't be run parallelly, because they take the output of the previous layer as input. This is a handicap for running parallel, so we need to try to parallelize the steps inside of them.
In this chapter, we won't delve into the details of a LlamaTransformerBlock.Forward(...), it is subject to another dedicated chapter further.
from src/model/llamatransformer.go
func (lt *LlamaTransformer) Forward(infContext *InferenceContext, inputTokens *ml.Tensor, startPos int) (*ml.Tensor, error) {
...
currentTensor := inputTensor
for layerIdx, layer := range lt.Layers {
startTime := time.Now()
common.GLogger.DebugPrintf("=======================================\n")
common.GLogger.DebugPrintf("Calling LlamaTransformerBlock.Forward for layer: %d / %d, startPos: %d -> tensor currentTensor", layerIdx+1, len(lt.Layers), startPos)
if currentTensor, err = layer.Forward(infContext, currentTensor, startPos, freqsCis, mask); err != nil {
return nil, err
}
infContext.Logf("Transformer block layer %d / %d was run, took %.4f sec(s)", layerIdx+1, len(lt.Layers), time.Since(startTime).Seconds())
}
...
}We can see output lines in the "debug.log" file if debugging is enabled, as follows:
[DEBUG] ... =======================================
[DEBUG] ... Calling LlamaTransformerBlock.Forward for layer: 1 / 32, startPos: 0 -> tensor currentTensor ...
...
[DEBUG] ... Calling LlamaTransformerBlock.Forward for layer: 2 / 32, startPos: 0 -> tensor currentTensor ...
...
[DEBUG] ... Calling LlamaTransformerBlock.Forward for layer: 32 / 32, startPos: 0 -> tensor currentTensor ...
[DEBUG] ... Returning tensor output: shape([22 4096]) ...