repo_name
stringlengths
1
62
dataset
stringclasses
1 value
lang
stringclasses
11 values
pr_id
int64
1
20.1k
owner
stringlengths
2
34
reviewer
stringlengths
2
39
diff_hunk
stringlengths
15
262k
code_review_comment
stringlengths
1
99.6k
xFasterTransformer
github_2023
cpp
375
intel
Duyi-Wang
@@ -0,0 +1,348 @@ +// Copyright (c) 2023 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless require...
I prefer to prepare KVCache at a high level api. We can assume all seqs' KVCache is ready.
xFasterTransformer
github_2023
cpp
348
intel
abenmao
@@ -596,4 +648,135 @@ void crossAttnShardedHead(T1 *output, const T1 *query, const float *attnMask, in } // end for b } +/** + * @brief Cross attention with head granularity (including copy key/value to KV Cache) + * @note if causal = True, attnMask is not used + * @tparam T Data type + * @tparam KVCacheT KV ca...
"const float *" is ok both for alibiSlopes and attnMask
xFasterTransformer
github_2023
cpp
348
intel
abenmao
@@ -596,4 +648,135 @@ void crossAttnShardedHead(T1 *output, const T1 *query, const float *attnMask, in } // end for b } +/** + * @brief Cross attention with head granularity (including copy key/value to KV Cache) + * @note if causal = True, attnMask is not used + * @tparam T Data type + * @tparam KVCacheT KV ca...
During the decoding phase, the softmax computing is the same whether causal is False๏ผˆchatglm๏ผ‰ or True, so โ€œcausalโ€ can be omitted in the โ€œif conditionโ€
xFasterTransformer
github_2023
cpp
348
intel
abenmao
@@ -596,4 +648,135 @@ void crossAttnShardedHead(T1 *output, const T1 *query, const float *attnMask, in } // end for b } +/** + * @brief Cross attention with head granularity (including copy key/value to KV Cache) + * @note if causal = True, attnMask is not used + * @tparam T Data type + * @tparam KVCacheT KV ca...
For the Baichuan model, alibiSlopes[i] need to be used in the softmax calculation; I will add a calculation function later.
xFasterTransformer
github_2023
cpp
371
intel
Duyi-Wang
@@ -0,0 +1,216 @@ +// Copyright (c) 2024 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless require...
default value for prefixID is better to be -1? since the ID starts from 0.
xFasterTransformer
github_2023
cpp
371
intel
Duyi-Wang
@@ -0,0 +1,216 @@ +// Copyright (c) 2024 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless require...
=๏ผŸ
xFasterTransformer
github_2023
cpp
369
intel
Duyi-Wang
@@ -530,13 +530,67 @@ class CommonDecoder : public AbstractDecoder { TimeLine t("Decoder.forward"); TimeLine t1("Decoder.embedding"); + // Prepare input + int totInputSeqLen = 0; + std::vector<int> allInputIds; + for (auto seq : seqs) { + // TODO: maybe not cor...
Assume input is prompt+past_generated_tokens. getInputSeqLen() gets the size of prompt, getLatestToken() return the past_generated_tokens.
xFasterTransformer
github_2023
cpp
369
intel
changqi1
@@ -67,63 +67,71 @@ class SequenceIDManager { // The SequenceMeta is one sequence of batch inputs and includes the generated tokens. class SequenceMeta { public: - SequenceMeta(std::vector<int32_t> &_inputTokens) + SequenceMeta(std::vector<int32_t> &_promptTokens) : sequenceID(SequenceIDManager::getIn...
ไผ˜ๅŒ–่ฟ”ๅ›žๅ€ผไผš่ฟ›่กŒcopy๏ผŒๅฝฑๅ“ๆ€ง่ƒฝ
xFasterTransformer
github_2023
cpp
369
intel
changqi1
@@ -67,63 +67,71 @@ class SequenceIDManager { // The SequenceMeta is one sequence of batch inputs and includes the generated tokens. class SequenceMeta { public: - SequenceMeta(std::vector<int32_t> &_inputTokens) + SequenceMeta(std::vector<int32_t> &_promptTokens) : sequenceID(SequenceIDManager::getIn...
ไผ˜ๅŒ–่ฟ”ๅ›žๅ€ผไผš่ฟ›่กŒcopy๏ผŒๅฝฑๅ“ๆ€ง่ƒฝ
xFasterTransformer
github_2023
cpp
370
intel
Duyi-Wang
@@ -47,4 +47,4 @@ class OptDecoder : public CommonDecoder<Attention<WeiT, QKPO_Dummy, LayerNorm>, LayerNorm finalLN; }; -REGISTER_MODEL(OptDecoder, gpt) +REGISTER_MODEL(OptDecoder, opt)
It should be 'gpt' since the model type in the converted config.ini is gpt.
xFasterTransformer
github_2023
cpp
370
intel
Duyi-Wang
@@ -122,3 +122,5 @@ template <typename WeiT, typename KVCacheT> void OptDecoder<WeiT, KVCacheT>::lastLayerNormForward(float *input, float *output, int rows) { finalLN.forward(input, output, rows); } + +IMPLEMENT_MODEL(OptDecoder, opt)
Should be gpt.
xFasterTransformer
github_2023
cpp
366
intel
abenmao
@@ -36,9 +36,17 @@ class Model { void config(SearcherConfig &config_, const std::vector<std::vector<int>> &stopWordsList_ = {}); + void set_input(std::vector<int32_t> &inputIds_, int batchSize_, int maxLen_ = -1, int numBeams_ = 1, + int numBeamHypsToKeep_ = 1, float lenPenalty_ = 1.0, bool doEar...
default to false
xFasterTransformer
github_2023
cpp
332
intel
Duyi-Wang
@@ -0,0 +1,210 @@ +// Copyright (c) 2024 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless require...
้˜ˆๅ€ผ`4`ไปŽ็Žฏๅขƒไธญpiplineๆ•ฐ้‡ๆˆ–่€…่‡ชๅฎšไน‰่ฎพ็ฝฎ๏ผŸ
xFasterTransformer
github_2023
cpp
332
intel
Duyi-Wang
@@ -0,0 +1,210 @@ +// Copyright (c) 2024 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless require...
set()? return bool. If not find return false else true.
xFasterTransformer
github_2023
cpp
332
intel
Duyi-Wang
@@ -0,0 +1,210 @@ +// Copyright (c) 2024 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless require...
bool? if key is existed, return false?
xFasterTransformer
github_2023
cpp
332
intel
Duyi-Wang
@@ -0,0 +1,210 @@ +// Copyright (c) 2024 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless require...
beamsearch้œ€่ฆๅ†ๅŒ…ไธ€ๅฑ‚๏ผŒๆฏ”ๅฆ‚็”จSampleGroupMetaใ€‚ https://github.com/vllm-project/vllm/blob/main/vllm/sequence.py#L547
xFasterTransformer
github_2023
cpp
332
intel
Duyi-Wang
@@ -0,0 +1,210 @@ +// Copyright (c) 2024 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless require...
Not used.
xFasterTransformer
github_2023
cpp
332
intel
pujiang2018
@@ -0,0 +1,275 @@ +// Copyright (c) 2024 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless require...
Is it better to design function like below to merge addNextToken and setStep: void stepForward(int id) (and in future, the id could be a vector) inside that function, we can manage step and other changed members.
xFasterTransformer
github_2023
cpp
332
intel
pujiang2018
@@ -64,6 +64,7 @@ struct DecoderContext { int inputSeqLen; // For custom usage int reserved1; + int sequenceID;
why do we need it here?
xFasterTransformer
github_2023
cpp
332
intel
Duyi-Wang
@@ -0,0 +1,286 @@ +// Copyright (c) 2024 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless require...
Make SequenceGroup as the base unit of Sequencepool.
xFasterTransformer
github_2023
cpp
321
intel
pujiang2018
@@ -621,3 +626,43 @@ class DecoderUtil { return false; } }; + +class DecoderFactory { +public: + using CreateFunc = std::function<AbstractDecoder *(const std::string &)>; + + static void Register(const std::string &key, CreateFunc createFunc) { GetRegistry()[key] = createFunc; } + + static A...
it's better to move REGISTER_DECODER and related code to CommonDecoder or a separate place. for util folder, it should be completely decoupled from the structure related to xFT. In future, we need to make 'util' a real utility folder.
xFasterTransformer
github_2023
cpp
321
intel
pujiang2018
@@ -40,3 +40,19 @@ class Baichuan : public CommonDecoder<BaichuanAttention<WeiT, LlamaRotaryEmbeddi TokenEmbedding<float16_t> *embedding; RmsNorm finalLN; }; + +REGISTER_DECODER(Baichuan, baichuan, float)
Move to cpp file is better, suppose 2 cpp files include baichuan.h, then the model will be registered twice.
xFasterTransformer
github_2023
others
337
intel
marvin-Yu
@@ -105,169 +118,183 @@ rls_test_case=$( # | qwen-14b | โˆš | โˆš | โˆš | โˆš | โˆš | โˆš | 32 | 32 | # llama-2-7b with short prompt & full data type: -bash run_benchmark.sh -m llama-2-7b -d fp16 -i 1 -w 0 -in 32 -out 32 -s 1 -bash run_benchmark.sh -m llama-2-7b -d bf16 -i 1 -w 0 -in 32 -out 32 -s 1 -...
typing error
xFasterTransformer
github_2023
others
337
intel
marvin-Yu
@@ -41,48 +41,61 @@ _test_case=$( # | gemma-7b | โˆš | โˆš | ร— | ร— | ร— | ร— | 32 | 32 |
Should we add the "-kvd" parameter to this table for better understanding of our test case coverage?
xFasterTransformer
github_2023
cpp
334
intel
marvin-Yu
@@ -0,0 +1,163 @@ +#include "attention_kernels.h" +#include "gemm_kernel_ext.h" +#include "gtest/gtest.h" + +#include <cstdlib> + +// Define a fixture for the unit tests +class AttentionKernelsTest : public ::testing::Test { +protected: + void SetUp() override { + // Set up any necessary data or resources for...
didn't see what the specific differences are between these two sets of test cases. Can we use a more specific naming convention for the test cases? Also, since we already support the int8 kvcache format, should it be reflected in this unit test?
xFasterTransformer
github_2023
cpp
259
intel
pujiang2018
@@ -218,6 +218,7 @@ class Attention { bool useSelfAttn, bool doLnBefore, int *positionIds = nullptr) { auto hiddenSize = ctx->hiddenSize; + auto attSize = ctx->attHeadNum * ctx->attHeadSize;
if 'attSize' not used, let's remove it.
xFasterTransformer
github_2023
cpp
259
intel
pujiang2018
@@ -199,7 +211,16 @@ class LlamaMLP : public SingletonBase<LlamaMLP<WeiT>> { const float *sumB = gateWeightSum.Data(); ImT *C = output.Data(); - ctx->mmHelper->compute_silu(false, M, N, K, 1.0f, A, lda, B, scaleB, zeroB, sumB, 0.0f, C, ldc); + if (ctx->actType == DecoderContext::SILU) ...
let's use the original path.
xFasterTransformer
github_2023
cpp
259
intel
changqi1
@@ -453,6 +453,36 @@ class DecoderUtil { } } + // compute gelu on the left half and then add it with the right half + template <typename T1, typename T2> + static void geluSum(hpj::Matrix<T1> &src, hpj::Matrix<T2> &dst) { + __m512 c1 = _mm512_set1_ps(0.044715f);
const
xFasterTransformer
github_2023
cpp
330
intel
pujiang2018
@@ -0,0 +1,66 @@ +// Copyright (c) 2023-2024 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless req...
What's the error when put such macro to cpp file?
xFasterTransformer
github_2023
others
330
intel
Duyi-Wang
@@ -1,5 +1,5 @@ #!/bin/bash -# set -x
unset
xFasterTransformer
github_2023
others
330
intel
changqi1
@@ -0,0 +1,27 @@ +{ + "architectures": [ + "Qwen2ForCausalLM" + ], + "attention_dropout": 0.0, + "bos_token_id": 151643, + "eos_token_id": 151645, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 2816, + "max_position_embeddings": 32768, + "max_window_laye...
what is the meaning for max_window_layers?
xFasterTransformer
github_2023
cpp
320
intel
abenmao
@@ -463,8 +463,9 @@ class Attention { auto srcV = value.Row(b * ctx->inputSeqLen + seq) + h * headSize; auto dstV = presentValue.getSequence(pastSeqLen + seq, b, h); - xft::copy(dstK, srcK, headSize); - xft::copy(dstV, srcV, headSize); + ...
Since there are type conversions in storeKVCache. The old comments can be removed.
xFasterTransformer
github_2023
cpp
320
intel
changqi1
@@ -478,19 +478,26 @@ class Attention { for (int seq = 0; seq < ctx->inputSeqLen; ++seq) { auto src = kv.Row(bdx * ctx->inputSeqLen + seq) + hdx * ctx->attHeadSize; auto dst = presentKV.getSequence(pastSeqLen + seq, bdx, hdx); - xft::copy(dst, src, ctx->attHeadSize); + ...
gemm1 -> gemm_query?
xFasterTransformer
github_2023
cpp
320
intel
changqi1
@@ -504,12 +511,18 @@ class Attention { // score: M * K(keyLen), value: K * headSize, output: M * headSize template <typename T1, typename T2, typename T3> - void gemm2(T1 *score, T2 *value, T3 *output, int M, int headSize, int K, int lds, int ldv, int ldo) { + void gemm2(T1 *score, const std::tuple<T...
gemm2 -> gemm_score?
xFasterTransformer
github_2023
others
323
intel
pujiang2018
@@ -183,6 +183,9 @@ add_definitions(-DAVX512_FP32_WEIGHT_ONLY_NF4=true) # add_definitions(-DSTEP_BY_STEP_ATTN=true) add_definitions(-DUSE_SHM=true) option(XFT_BUILD_TESTS "Build xfastertransformer unit tests" OFF) +if(XFT_BUILD_TESTS) +add_definitions(-DUNDEBUG=true)
why do we need UNDEBUG?
xFasterTransformer
github_2023
cpp
319
intel
pujiang2018
@@ -70,6 +76,7 @@ class Model { std::vector<int32_t> inputIds; int batchSize; int seqLen; + int vocabSize_;
Why do we need vocabSize in Model class? if needed, could you pls make it in same naming style?
xFasterTransformer
github_2023
cpp
319
intel
pujiang2018
@@ -129,6 +129,25 @@ struct TorchAutoModel : torch::CustomClassHolder { doSample, temperature, topK, topP, repetitionPenalty, stopWordsList_int32); } + torch::Tensor forward(torch::Tensor &inputIds) { + int batchSize = inputIds.size(0); + int seqLen = inputIds.size(1); + ...
For the original code, why 'input' method is not called 'setInput', input is more like a keyword, :(
xFasterTransformer
github_2023
cpp
319
intel
Duyi-Wang
@@ -129,6 +129,25 @@ struct TorchAutoModel : torch::CustomClassHolder { doSample, temperature, topK, topP, repetitionPenalty, stopWordsList_int32); } + torch::Tensor forward(torch::Tensor &inputIds) { + int batchSize = inputIds.size(0); + int seqLen = inputIds.size(1); + ...
Shape is correct? The decoder just returns the last token's logits.
xFasterTransformer
github_2023
cpp
319
intel
Duyi-Wang
@@ -129,6 +129,25 @@ struct TorchAutoModel : torch::CustomClassHolder { doSample, temperature, topK, topP, repetitionPenalty, stopWordsList_int32); } + torch::Tensor forward(torch::Tensor &inputIds) { + int batchSize = inputIds.size(0); + int seqLen = inputIds.size(1); + ...
Sync in multi-ranks?
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -87,65 +87,112 @@ void ChatGLM2RotaryEmbedding::glm2CalEmb() { // x_out2 = x_out2.flatten(3) // return torch.cat((x_out2, x_pass), dim=-1) -void ChatGLM2RotaryEmbedding::forward(float *buf, int bufStride, int batch_size, int seq_len, int qk_size, - int hidden_size_per_attention_head, const int *po...
```c++ #include <immintrin.h> #include <stdio.h> void interleave_merge(__m512 a, __m512 b, __m512 *result) { // ็”ŸๆˆๆŽฉ็  //__m512i mask = _mm512_set_epi32(0x17, 0x07, 0x16, 0x06, 0x15, 0x05, 0x14, 0x04, 0x13, 0x03, 0x12, 0x02, 0x11, 0x01, 0x10, 0x00); __m512i mask = _mm512_set_epi32(0x1f, 0x0f, 0x1e, 0x...
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -87,65 +87,112 @@ void ChatGLM2RotaryEmbedding::glm2CalEmb() { // x_out2 = x_out2.flatten(3) // return torch.cat((x_out2, x_pass), dim=-1) -void ChatGLM2RotaryEmbedding::forward(float *buf, int bufStride, int batch_size, int seq_len, int qk_size, - int hidden_size_per_attention_head, const int *po...
Delete those
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -35,10 +36,12 @@ class ChatGLM2RotaryEmbedding { ~ChatGLM2RotaryEmbedding() {} - void forward(float *buf, int bufStride, int batch_size, int seq_len, int qk_size, - int hidden_size_per_attention_head, const int *position_ids); + // void forward(float *buf, int bufStride, int batch_size, int...
delete those.
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -21,8 +21,13 @@ template <typename WeiT> ChatGLM2<WeiT>::ChatGLM2(const std::string &modelPath, const std::string &modelType) - : CommonDecoder<Attention<WeiT, ChatGLM2RotaryEmbedding, RmsNorm, float, float, float, true>, - ChatGLM2MLP<WeiT, float, float, float, RmsNorm, true>>(modelPath, modelType...
delete those.
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -23,15 +23,23 @@ #include "token_embedding.h" template <typename WeiT> -class ChatGLM2 : public CommonDecoder<Attention<WeiT, ChatGLM2RotaryEmbedding, RmsNorm, float, float, float, true>, - ChatGLM2MLP<WeiT, float, float, float, RmsNorm, true>> { +class ChatGLM2 + : public CommonDecode...
delete those.
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -138,17 +108,52 @@ void ChatGLM2RotaryEmbedding::forward( float *psin = emb_sin + pos * dim; #pragma omp simd - for (int i = 0; i < dim; i += 2) { + for (int i = 0; i < half; i += 2) { auto t1 = p1[i]; - p1[i] = p1[i] * pcos[i ...
const __m512i
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -138,17 +108,52 @@ void ChatGLM2RotaryEmbedding::forward( float *psin = emb_sin + pos * dim; #pragma omp simd - for (int i = 0; i < dim; i += 2) { + for (int i = 0; i < half; i += 2) { auto t1 = p1[i]; - p1[i] = p1[i] * pcos[i ...
const __m512i
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -138,17 +108,52 @@ void ChatGLM2RotaryEmbedding::forward( float *psin = emb_sin + pos * dim; #pragma omp simd - for (int i = 0; i < dim; i += 2) { + for (int i = 0; i < half; i += 2) { auto t1 = p1[i]; - p1[i] = p1[i] * pcos[i ...
const __m512i
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -138,17 +108,52 @@ void ChatGLM2RotaryEmbedding::forward( float *psin = emb_sin + pos * dim; #pragma omp simd - for (int i = 0; i < dim; i += 2) { + for (int i = 0; i < half; i += 2) { auto t1 = p1[i]; - p1[i] = p1[i] * pcos[i ...
const __m512i
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -138,17 +108,52 @@ void ChatGLM2RotaryEmbedding::forward( float *psin = emb_sin + pos * dim; #pragma omp simd - for (int i = 0; i < dim; i += 2) { + for (int i = 0; i < half; i += 2) { auto t1 = p1[i]; - p1[i] = p1[i] * pcos[i ...
*result0 = _mm512_permutex2var_ps(a, mask0, b); Why not?
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -138,17 +108,52 @@ void ChatGLM2RotaryEmbedding::forward( float *psin = emb_sin + pos * dim; #pragma omp simd - for (int i = 0; i < dim; i += 2) { + for (int i = 0; i < half; i += 2) { auto t1 = p1[i]; - p1[i] = p1[i] * pcos[i ...
const __m512i
xFasterTransformer
github_2023
cpp
301
intel
changqi1
@@ -111,41 +113,87 @@ void ChatGLM2RotaryEmbedding::forward(float *buf, int bufStride, int batch_size, p1[i] = p1[i] * pcos[i] - p1[i + 1] * psin[i]; p1[i + 1] = p1[i + 1] * pcos[i] + t1 * psin[i]; } - off += bufStride; + off += qS...
Why not to use it to parallel for?
xFasterTransformer
github_2023
cpp
302
intel
abenmao
@@ -54,6 +54,38 @@ void invokeMLPLLaMA(DataType dt, int numTokens, int hiddenSize, int intermediate llama_mlp = it_created->second; } + ctx->resize(1, numTokens, 0); + llama_mlp->forward(ctx, (float *)const_cast<void *>(input), (float *)output, inputStride, outputStride, false); + ...
Can these duplicated or similar code lines for fp16 be reused for the bf16 section?
xFasterTransformer
github_2023
cpp
294
intel
changqi1
@@ -0,0 +1,231 @@ +#include <algorithm> +#include "attention_kernels.h" +#include "matmul_helper.h" +#include "gtest/gtest.h" + +// Reference implementation of matrix multiplication +static void mmRef( + bfloat16_t *A, bfloat16_t *B, bfloat16_t *C, int M, int N, int K, int lda, int ldb, int ldc, bool transB) { +...
#pragma omp parallel for
xFasterTransformer
github_2023
cpp
272
intel
Duyi-Wang
@@ -0,0 +1,54 @@ +// Copyright (c) 2024 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required...
Add this ENV in environment.h?
xFasterTransformer
github_2023
python
276
intel
changqi1
@@ -210,3 +210,189 @@ def split_and_convert(self, input_dir, output_dir, dtype, processes): pool.starmap_async(self.split_and_convert_process, starmap_args) pool.close() pool.join() + + def split_and_convert_quantized_model(self, input_dir, output_dir, dtype, processes, quantizatio...
uint4x2
xFasterTransformer
github_2023
python
276
intel
changqi1
@@ -210,3 +210,189 @@ def split_and_convert(self, input_dir, output_dir, dtype, processes): pool.starmap_async(self.split_and_convert_process, starmap_args) pool.close() pool.join() + + def split_and_convert_quantized_model(self, input_dir, output_dir, dtype, processes, quantizatio...
quant_decoder_weights -> quant_weight_data_type int8,uint8,int4,uint4,fp8,fp4
xFasterTransformer
github_2023
python
276
intel
changqi1
@@ -53,16 +53,22 @@ def get_weight_data_type(self, dtype: str): else: raise Exception(f"{self.__class__.__name__} don't support convert weight to {dtype}.") - def convert(self, input_dir, output_dir=None, dtype: str = "fp16", processes=8): + def convert(self, input_dir, output_dir=None, dt...
\# from_quantized_model is from HF int8 model to xFT int8 model. def convert(self, input_dir, output_dir=None, dtype: str = "fp16", processes=8, from_quantized_model=None):
xFasterTransformer
github_2023
python
276
intel
changqi1
@@ -210,3 +210,189 @@ def split_and_convert(self, input_dir, output_dir, dtype, processes): pool.starmap_async(self.split_and_convert_process, starmap_args) pool.close() pool.join() + + def split_and_convert_quantized_model(self, input_dir, output_dir, dtype, processes, quantizatio...
[ERROR] Please input model must be {quantization} quantized model.
xFasterTransformer
github_2023
python
276
intel
changqi1
@@ -210,3 +210,189 @@ def split_and_convert(self, input_dir, output_dir, dtype, processes): pool.starmap_async(self.split_and_convert_process, starmap_args) pool.close() pool.join() + + def split_and_convert_quantized_model(self, input_dir, output_dir, dtype, processes, quantizatio...
\# load the quantized model, do not do GPTQ quantization.
xFasterTransformer
github_2023
python
276
intel
changqi1
@@ -210,3 +210,189 @@ def split_and_convert(self, input_dir, output_dir, dtype, processes): pool.starmap_async(self.split_and_convert_process, starmap_args) pool.close() pool.join() + + def split_and_convert_quantized_model(self, input_dir, output_dir, dtype, processes, quantizatio...
zeros is uint8/uint4 from QPTQ quantization, and need to convert to fp32 for xFT.
xFasterTransformer
github_2023
python
276
intel
changqi1
@@ -210,3 +210,189 @@ def split_and_convert(self, input_dir, output_dir, dtype, processes): pool.starmap_async(self.split_and_convert_process, starmap_args) pool.close() pool.join() + + def split_and_convert_quantized_model(self, input_dir, output_dir, dtype, processes, quantizatio...
scales is fp16/fp32/bf16 from QPTQ quantization, and need to convert to fp32 for xFT.
xFasterTransformer
github_2023
python
276
intel
changqi1
@@ -210,3 +210,189 @@ def split_and_convert(self, input_dir, output_dir, dtype, processes): pool.starmap_async(self.split_and_convert_process, starmap_args) pool.close() pool.join() + + def split_and_convert_quantized_model(self, input_dir, output_dir, dtype, processes, quantizatio...
convert uint8 qweight from QPTQ quantization, and need to convert to int8 for xFT. convert uint4 qweight from QPTQ quantization, and need to convert to uint4x2 for xFT.
xFasterTransformer
github_2023
cpp
279
intel
pujiang2018
@@ -21,6 +21,8 @@ #include "allocator.h" +extern bool kvTrans(); + /** * Tensor specially designed for KV Cache * Naturaly, it could be represented in the shape of [seq_length][batch_size][head_num][head_size]
Please also modify the comments here?
xFasterTransformer
github_2023
cpp
279
intel
pujiang2018
@@ -18,14 +18,41 @@ bool enableCATMLP() { static int catMlp = -1; - if (catMlp == -1) - catMlp = (getenv("ENABLE_CAT_MLP") ? atoi(getenv("ENABLE_CAT_MLP")) : 1); + if (catMlp == -1) catMlp = (getenv("ENABLE_CAT_MLP") ? atoi(getenv("ENABLE_CAT_MLP")) : 1); return catMlp == 1; } +bool tunedCo...
please remove it if not used.
xFasterTransformer
github_2023
cpp
279
intel
pujiang2018
@@ -18,14 +18,41 @@ bool enableCATMLP() { static int catMlp = -1; - if (catMlp == -1) - catMlp = (getenv("ENABLE_CAT_MLP") ? atoi(getenv("ENABLE_CAT_MLP")) : 1); + if (catMlp == -1) catMlp = (getenv("ENABLE_CAT_MLP") ? atoi(getenv("ENABLE_CAT_MLP")) : 1); return catMlp == 1; } +bool tunedCo...
it is not so easy to understand "Tuned communication". add some comment?
xFasterTransformer
github_2023
python
274
intel
changqi1
@@ -123,7 +123,14 @@ def split_and_convert(self, input_dir, output_dir, dtype, processes): config["llama"]["num_layer"] = str(hf_config["num_hidden_layers"]) config["llama"]["layernorm_eps"] = str(hf_config.get("rms_norm_eps", 1e-6)) config["llama"]["layernorm_type"] = "pre_layern...
ๆŠŠLLaMaๅ’Œdeepseek็š„codeๅˆ†ๅผ€ๆฅ๏ผŒllama็š„codeๅฐฝ้‡ไธ่ฆๅŠจ๏ผŒ่ฎพ่ฎกๅˆฐdeepseek็š„ๅฏไปฅๅˆ›ๅปบๆ–ฐ็š„ๆ–‡ไปถ๏ผŒไฝ†้œ€่ฆ้›†ๆˆllama็š„codeใ€‚
xFasterTransformer
github_2023
cpp
274
intel
changqi1
@@ -46,5 +48,14 @@ class LlamaRotaryEmbedding { void llamaCalEmb(const float *inv_freq, const int max_position_embeddings); private: - static bool initialized; + bool initialized = false;
้œ€่ฆstatic็š„๏ผŒไฝ ่ฟ™sinๅ’Œcosไผšๆœ‰ๅคšไปฝ็›ธๅŒ็š„ๅฎžไพ‹๏ผŒไฝ†ไธ€ไธชmodelๅช้œ€่ฆไธ€ไปฝsinๅ’Œcos
xFasterTransformer
github_2023
cpp
274
intel
changqi1
@@ -17,41 +17,48 @@ #include "allocator.h" #include "compile_util.h" -static int inv_freq_size = -1; -static float *emb_cos = nullptr; -static float *emb_sin = nullptr; +LlamaRotaryEmbedding::LlamaRotaryEmbedding(DecoderContext *ctx) { + const std::string inv_freq_str = "inv_freq"; + const std::string emb_cos...
ๆŠŠLLaMaๅ’Œdeepseek็š„codeๅˆ†ๅผ€ๆฅ๏ผŒllama็š„codeๅฐฝ้‡ไธ่ฆๅŠจ๏ผŒๆถ‰ๅŠๅˆฐdeepseek็š„ๅฏไปฅๅˆ›ๅปบๆ–ฐ็š„ๆ–‡ไปถ๏ผŒไฝ†้œ€่ฆ้›†ๆˆllama็š„codeใ€‚่ฟ™ไธชไฝ ๅฏไปฅๆ–ฐๅปบไธ€ไธชdeepseek็š„ropeๆ–‡ไปถใ€‚
xFasterTransformer
github_2023
python
274
intel
changqi1
@@ -117,7 +117,7 @@ def build_inputs_chatglm(tokenizer, query: List[str], padding, history: List[Tup model_prompt = prompt_pool["chatglm2"] if "chatglm3" in args.model_name.lower(): model_prompt = prompt_pool["chatglm3"] - if "llama" in args.model_name.lower(): + if "llama" in args.model_na...
ๅ•็‹ฌ if deepseek
xFasterTransformer
github_2023
others
274
intel
changqi1
@@ -141,6 +142,7 @@ xFasterTransformer supports a different model format from Huggingface, but it's Supported model convert list: - LlamaConvert + - DeepseekConvert
put it down
xFasterTransformer
github_2023
python
274
intel
changqi1
@@ -34,6 +34,7 @@ def with_mpirun(): "automodel": ["AutoModel"], "tools": [ "LlamaConvert", + "DeepseekConvert",
put it down
xFasterTransformer
github_2023
cpp
278
intel
pujiang2018
@@ -0,0 +1,43 @@ +// Copyright (c) 2023 Intel Corporation
let's change to 2024
xFasterTransformer
github_2023
cpp
278
intel
pujiang2018
@@ -0,0 +1,48 @@ +// Copyright (c) 2023 Intel Corporation
let's change to 2024
xFasterTransformer
github_2023
cpp
278
intel
pujiang2018
@@ -0,0 +1,43 @@ +// Copyright (c) 2023 Intel Corporation +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required...
shall we combine batchSize and seqLen into one parameter? (think about continuous batching, we just need to pass the total id size)
xFasterTransformer
github_2023
others
262
intel
pujiang2018
@@ -1,5 +1,34 @@ # CHANGELOG +# [Version v1.4.0](https://github.com/intel/xFasterTransformer/releases/tag/v1.4.0) +v1.4.0 - Support fully BF16 inference deployment of Llama series and add serving supports. + +## Functionality +- Introduce pure BF16 support in Llama series models, now can use fully BF16 data type to ...
Since BF16 previously already supported (though not the fully support), suggest to modify it like: Introduce fully BF16 support in Llama series models to better use AMX
xFasterTransformer
github_2023
cpp
245
intel
changqi1
@@ -50,6 +50,8 @@ class MMHelper { std::cerr << "[Error] Wrong device type." << std::endl; std::exit(-1); } + + AMXThresholdM=Env::getAMXThresholdM();
็ฉบๆ ผ
xFasterTransformer
github_2023
cpp
252
intel
pujiang2018
@@ -0,0 +1,31 @@ +// Copyright (c) 2023 Intel Corporation
let's change to 2024
xFasterTransformer
github_2023
cpp
224
intel
pujiang2018
@@ -170,11 +202,33 @@ inline void float16_t::cvt_float16_to_float(const float16_t *src, float *dst, in } } +inline void float16_t::cvt_float16_to_float_MT(const float16_t *src, float *dst, int size) { + // Process 16 floats (AVX512 is a 512-bit SIMD register) + constexpr int kStep = 16; + int blockSize...
may have risk here. better not change.
xFasterTransformer
github_2023
python
33
intel
Duyi-Wang
@@ -152,7 +154,7 @@ def build_inputs_chatglm(tokenizer, query: List[str], padding, history: List[Tup Next_token_throughput = 1000 / latency_90 * args.batch_size print("\n") print("=" * 50 + args.model_name + " Final Performance" + "=" * 50) - print(f"First token Latency:\t{first_token_...
The input token length has been printed in Line 113.
xFasterTransformer
github_2023
others
221
intel
Duyi-Wang
@@ -14,7 +14,12 @@ # ============================================================================ cmake_minimum_required(VERSION 3.15.1) +find_package(MPI REQUIRED)
If oneCCL is not present in the user's environment๏ผŸ
xFasterTransformer
github_2023
cpp
221
intel
Duyi-Wang
@@ -134,7 +212,14 @@ class CommonDecoder : public AbstractDecoder { vocabSize, embeddingSize, maxPositions, maxPosEmbed, maxSeqLength, ropeParamsPtr); // Decoder - for (int i = 0; i < layers; ++i) { + if (layers % ctx->ppSize != 0) { + std::cerr << "Warning: layers c...
When layers is not divisible by ppSize, does it mean that a few layers (layers % ppSize) will not be processed? There is a warning but no termination if layers is not divisible by ppSize.
xFasterTransformer
github_2023
cpp
221
intel
Duyi-Wang
@@ -227,9 +312,19 @@ class CommonDecoder : public AbstractDecoder { int *positionIds = this->getPositionIds(ids, batchSize, inputSeqLen, step + this->prefixSharing); t1.release(); + // if current pipeline parallel stage rank isn't the first stage, should receive previous stage data + i...
This will reintroduce the MPI dependency into xft.so. It should be included in comm_helper.so and referenced through messager.
xFasterTransformer
github_2023
cpp
221
intel
Duyi-Wang
@@ -67,7 +67,28 @@ class Env { } } - static int getVerbose() { return verbose_value(); } + static int getVerbose() { return verboseValue(); } + +// Pipeline Parallel +private: + static int &pipelineValue() { + static int value = 1; + return value; + } + +public: + static voi...
Check if MPI_rank is divisible by ppStages?
xFasterTransformer
github_2023
cpp
221
intel
Duyi-Wang
@@ -46,7 +46,23 @@ std::vector<int> GreedySearch::getNextToken(int *ids, int batchSize, int seqLen) std::tuple<float *, int, int> result = decoder.forward(ids, dims, this->step++); - this->nextTokens = search(result); + DecoderContext *ctx = decoder.getContext(); + if (std::get<0>(result) == nullptr) ...
Use messager and comm_helper.so to decouple MPI dependency.
xFasterTransformer
github_2023
cpp
221
intel
pujiang2018
@@ -176,16 +201,21 @@ class Messenger { private: int size; int rank; + int color;
I don't know if color is a common concept. Is it easy to understand for others? add some comment?
xFasterTransformer
github_2023
cpp
225
intel
pujiang2018
@@ -392,7 +396,11 @@ void BeamSearch::searchTopK(std::tuple<float *, int, int> &result) { std::vector<long unsigned int> recvCount(msgerSize, static_cast<long unsigned int>(batchSize * numBeams)); messenger.allgatherv(maxVal, batchSize * numBeams, recvMax, recvCount); - float sumVal[batchSize...
Original float "sumVal[batchSize * numBeams] = {0};" should work.
xFasterTransformer
github_2023
cpp
225
intel
pujiang2018
@@ -373,7 +373,11 @@ void BeamSearch::searchTopK(std::tuple<float *, int, int> &result) { // 4. add beam socre. Initialize -1e9 to all beams except the first one if (msgerSize > 1) { // Get the maximum value of each beam through all instance - float maxVal[batchSize * numBeams] = {std::numeric...
I guess the author intended to set all values to lowest float. good fix.
xFasterTransformer
github_2023
cpp
215
intel
changqi1
@@ -149,33 +174,68 @@ void QwenRotaryEmbedding::forward( cur_emb_sin = std::get<1>(value); } - // for (size_t i = 0; i < emb_size; i++) { - // emb[i] = x[i] * emb_cos[position_ids[i % cached_size / dim]][i % dim]; - // int offset = (i % dim + this->inv_freq_size) % dim; - // floa...
if ๅ’Œ else ไธญ็š„codeๆ˜ฏไธๆ˜ฏๅฏไปฅ็ปŸไธ€ๆˆไธ€ไปฝ๏ผŒๆˆ‘็œ‹็ปๅคงๅคšๆ•ฐcodeๆ˜ฏๅค็”จ็š„๏ผŒๅนถไธ” else ไธญ q_scale[seq] ไนŸๆ˜ฏ 1.0 ๏ผŸ
xFasterTransformer
github_2023
cpp
215
intel
changqi1
@@ -216,46 +277,103 @@ void QwenRotaryEmbedding::forward( cur_emb_sin = std::get<1>(value); } + if (seqLen + pastKeyLength > maxSeqLength && logn != nullptr) {
if ๅ’Œ else ไธญ็š„codeๆ˜ฏไธๆ˜ฏๅฏไปฅ็ปŸไธ€ๆˆไธ€ไปฝ๏ผŒๆˆ‘็œ‹็ปๅคงๅคšๆ•ฐcodeๆ˜ฏๅค็”จ็š„๏ผŒๅนถไธ” else ไธญ q_scale[seq] ไนŸๆ˜ฏ 1.0 ๏ผŸ
xFasterTransformer
github_2023
cpp
183
intel
pujiang2018
@@ -69,3 +69,26 @@ void repetitionPenaltyLogitsProcess(float penalty, float *logits, int sampleOffs } } } + +void stopWordsCheck(std::vector<int> &nextTokenIds, std::vector<std::vector<int>> &stopWordsList, + std::vector<std::vector<int>> &stopWordsIndex, std::vector<int> &doneBatch) { + for (i...
@Duyi-Wang Do you need openmp for large batch size?
xFasterTransformer
github_2023
cpp
73
intel
pujiang2018
@@ -78,7 +78,9 @@ class bfloat16_t { } static void cvt_float_to_bfloat16(const float *src, bfloat16_t *dst, int size); + static void batch_cvt_float_to_bfloat16(const float *src, bfloat16_t *dst, int size);
@changqi1 Do you accept adding it in bfloat16.h?
xFasterTransformer
github_2023
cpp
73
intel
pujiang2018
@@ -39,6 +42,12 @@ class Messenger { return; } + bf16_enable = (getenv("XFT_ONECCL_BF16") ? atoi(getenv("XFT_ONECCL_BF16")) : 0); + if (bf16_enable) { + printf("got 'XFT_ONECCL_BF16=%d', enable BF16 dtype comm.\n", bf16_enable); + buf_bf16 = new bfloat16_t[MAX...
suggest using aligned_alloc or malloc. as new will initialize all the data to 0, which is not needed here.
xFasterTransformer
github_2023
cpp
73
intel
pujiang2018
@@ -110,7 +123,20 @@ class Messenger { // From some example code of oneCCL, inplace reducing is supported template <typename T> void reduceAdd(T *sendBuf, T *recvBuf, size_t count) { + if constexpr (!std::is_same_v<T, float>) { + //todo(marvin): Consideration for additional optimization...
what if count > MAX_BF16_BUFFER?
xFasterTransformer
github_2023
cpp
73
intel
pujiang2018
@@ -110,7 +123,20 @@ class Messenger { // From some example code of oneCCL, inplace reducing is supported template <typename T> void reduceAdd(T *sendBuf, T *recvBuf, size_t count) { + if constexpr (!std::is_same_v<T, float>) { + //todo(marvin): Consideration for additional optimization...
Here the code looks like not formatted.
xFasterTransformer
github_2023
cpp
73
intel
pujiang2018
@@ -175,7 +201,8 @@ class Messenger { int size; int rank; bool local_ranks_flag; - + bfloat16_t* buf_bf16 = nullptr;
let's initialize it in the constructor to make sure the same style.
xFasterTransformer
github_2023
cpp
73
intel
changqi1
@@ -128,6 +130,18 @@ inline void bfloat16_t::cvt_float_to_bfloat16(const float *src, bfloat16_t *dst, } } +inline void bfloat16_t::batch_cvt_float_to_bfloat16(const float *src, bfloat16_t *dst, int count){ + constexpr int sizePerSplit = 1024;
Maybe use template to define blockSize or use totalSize/totalSystemThreads to define the blockSize.
xFasterTransformer
github_2023
cpp
157
intel
pujiang2018
@@ -85,6 +85,11 @@ class CommonDecoder : public AbstractDecoder { this->prefixSeqLen = 0; this->prefixSharing = false; + // Quantization config + const bool quant_decoder_weights = reader.GetBoolean(modelType, "quant_decoder_weights", false);
Shall we also use camelCase to name the params?
xFasterTransformer
github_2023
cpp
157
intel
pujiang2018
@@ -475,42 +481,152 @@ class CommonDecoder : public AbstractDecoder { int qkvSize = qSize + kvSize + kvSize; #define ALLOC(size, alignment) aligned_alloc((alignment), (size)) - float *qkvWeight = (float *)ALLOC(hiddenSize * qkvSize * sizeof(float), 64); + float *qkvWeight = nullptr; + i...
remove it?
xFasterTransformer
github_2023
cpp
157
intel
pujiang2018
@@ -475,42 +481,152 @@ class CommonDecoder : public AbstractDecoder { int qkvSize = qSize + kvSize + kvSize; #define ALLOC(size, alignment) aligned_alloc((alignment), (size)) - float *qkvWeight = (float *)ALLOC(hiddenSize * qkvSize * sizeof(float), 64); + float *qkvWeight = nullptr; + i...
maybe need to search all the printf to remove it or move it to debug code.
xFasterTransformer
github_2023
cpp
157
intel
pujiang2018
@@ -55,35 +55,69 @@ class Decoder { int getLayerId() { return layerIdx; } - void setWeights(DecoderContext *ctx, std::vector<float *> &params, bool trans = true) { - const float *queryWeight = params[0]; - const float *queryBias = params[1]; - const float *keyWeight = params[2]; - ...
I think it is better to change "int type" to "xft::DataType dt".