Skip to content

Latest commit

 

History

History
252 lines (197 loc) · 11.9 KB

CHANGELOG.md

File metadata and controls

252 lines (197 loc) · 11.9 KB

CHANGELOG

v1.8.2

Performance

  • Enable flash attention by default for W8A8 dtype to accelerate the performance of the 1st token.

Benchmark

  • When the number of ranks is 1, run in single mode to avoid the dependency on mpirun.
  • Support SNC-3 platform.

v1.8.1

Functionality

  • Expose the interface of embedding lookup.

Performance

  • Optimized the performance of grouped query attention (GQA).
  • Enhanced the performance of creating keys for the oneDNN primitive cache.
  • Set the [bs][nh][seq][hs] layout as the default for KV Cache, resulting in better performance.
  • Improved the task split imbalance issue in self-attention.

v1.8.0 Continuous Batching on Single ARC GPU and AMX_FP16 Support.

Highlight

  • Continuous Batching on Single ARC GPU is supported and can be integrated by vllm-xft.
  • Introduce Intel AMX instructions support for float16 data type.

Models

  • Support ChatGLM4 series models.
  • Introduce BF16/FP16 full path support for Qwen series models.

BUG fix

  • Fixed memory leak of oneDNN primitive cache.
  • Fixed SPR-HBM flat QUAD mode detect issue in benchmark scripts.
  • Fixed heads Split error for distributed Grouped-query attention(GQA).
  • Fixed an issue with the invokeAttentionLLaMA API.

v1.7.3

BUG fix

  • Fixed SHM reduceAdd & rope error when batch size is large.
  • Fixed the issue of abnormal usage of oneDNN primitive cache.

v1.7.2 - Continuous batching feature supports Qwen 1.0 & hybrid data types.

Functionality

  • Add continuous batching support of Qwen 1.0 models.
  • Enable hybrid data types for continuous batching feature, including BF16_FP16, BF16_INT8, BF16_W8A8, BF16_INT4, BF16_NF4, W8A8_INT8, W8A8_int4, W8A8_NF4.

BUG fix

  • Fixed the convert fault in Baichuan1 models.

v1.7.1 - Continuous batching feature supports ChatGLM2/3.

Functionality

  • Add continuous batching support of ChatGLM2/3 models.
  • Qwen2Convert supports quantized Qwen2 models by GPTQ, such as GPTQ-Int8 and GPTQ-Int4, by param from_quantized_model="gptq".

BUG fix

  • Fixed the segament fault error when running with more than 2 ranks in vllm-xft serving.

v1.7.0 - Continuous batching feature supported.

Functionality

  • Refactor framework to support continuous batching feature. vllm-xft, a fork of vllm, integrates the xFasterTransformer backend and maintains compatibility with most of the official vLLM's features.
  • Remove FP32 data type option of KV Cache.
  • Add get_env() python API to get recommended LD_PRELOAD set.
  • Add GPU build option for Intel Arc GPU series.
  • Exposed the interface of the LLaMA model, including Attention and decoder.

Performance

  • Update xDNN to release v1.5.1
  • Baichuan series models supports full FP16 pipline to improve performance.
  • More FP16 data type kernel added, including MHA, MLP, YARN rotary_embedding, rmsnorm and rope.
  • Kernel implementation of crossAttnByHead.

Dependency

  • Bump torch to 2.3.0.

BUG fix

  • Fixed the segament fault error when running with more than 4 ranks.
  • Fixed the bugs of core dump && hang when running croos nodes.

v1.6.0 - Llama3 and Qwen2 series models supported.

Functionality

  • Support Llama3 and Qwen2 series models.
  • Add INT8 KV cache datatype, using kv_cache_dtype params to specify, including int8, fp16(default) and fp32.
  • More models enable full BF16 pipline, includes Chatglm2/3 and yarn-llama.
  • Add invokeMLPLLaMA FP16 API.
  • Support logits output using forward() api.

Dependency

  • Bump transformers to 4.40.0 to support Llama3 models.

Performance

  • Update xDNN to release v1.4.6

BUG fix

  • Fix numeric overflow when calculate softmax in sampling.
  • fix assert bug when concat gate&up.

v1.5.0 - Gemma series models supported.

Functionality

  • Support Gemma series medels, including Gemma and CodeGemma, and DeepSeek model.
  • Llama Converter support convert quantized huggingface model by params from_quantized_model='gptq into xFt format INT8/INT4 model files.
  • Support loading INT4 data weights directly from local files.
  • Optimize memory usage during QWen model conversion, particularly for QWen 72B.

Dependency

  • Bump transformers to 4.38.1 to support Gemma models.
  • Add protobuf to support new behavier in tokenzier.

Performance

  • Update xDNN to release v1.4.5
  • Add GPU kernel library gpuDNN v0.1 to support Intel Arc GPU series.
  • Optimize ROPE perfermance by reducing repeated sin and cos embedding table data.
  • Accelerate KVCache copy by increasing parallelism in self attention.
  • Accelerate addreduce operation in long sequence case by transposing KVCache and tuned comm.

BUG fix

  • Fix a incorrect computing which should be in float, but was in integer.
  • Fix timeline is disordered.
  • Fix runtime issue of Qwen when seq_length is bigger than 32768.

v1.4.0 - Fully BF16 support in Llama for better performance and serving framework support.

Functionality

  • Introduce pure BF16 support to Llama series models, now can use fully BF16 data type to to utilize AMX more effectively when deploying Llama models.
  • Add MLServer serving framework support and demo in serving directory.
  • GCC for compiling release binary files has been updated from GCC 8.5 to GCC 12.
  • Introduce pipeline parallel feature for distributing deployment. Enabled by cmake .. -DWITH_PIPELINE_PARALLEL=ON in compilation and use XFT_PIPELINE_STAGE Marco to define pipeline parallel stages num.
  • Deprecate convert tool scripts in tools directory and it recommended to using Convert in xfastertransformer python wheel.
  • Support loading int8 data weights directly from local files.

Performance

  • Update xDNN to release v1.4.4.
  • Accelerate model weights loading by optimizing cast operation after loading and gain up to 50% speed up.
  • Optimize BF16 performance using AMX instruction when batchsize <= 8, and add XFT_USE_AMX_M to set threshold of M using AMX instead of AVX512, default 1.

Demo & Benchmark

  • Update dependency transformers requirement from 4.30.0 to 4.36.0 for high risk CVE Vulnerabilities.
  • Add distributed inference benchmark script which support deployment across platfrom.
  • Add single node platform support in benchmark script.
  • Add Yi model web demo.
  • Enhance the command-line chat mode in pytorch demo.py, using --chat true to enable.

BUG fix

  • Fix calculation issue in Qwen models and enhance LogN support for long token sequence.
  • Fix unsync issue in multi-rank model when do_sample is enabled.
  • Fix Baichuan models calculation and convert issue.
  • Fix repetition penalties not taking effect on other batches.

v1.3.1

BUG fix

  • Fix oneCCL environment is still needed when running in single-rank mode.

v1.3.0 - Qwen model support enhancement and added support for the SecLLM (YaRN-Llama) model.

Models

  • Introduce SecLLM(YaRN-Llama) model support.
  • Integrating the Qwen web demo, enhancing Qwen model support, and fix known issues in the Qwen convert tool.

Functionality

  • Introduce new generation configuration, repetition_penalty and stop_words_ids.
  • Rotary embedding supports BF16 data type now.
  • Introduce attention interfaces similar to page attention.
  • Add a whitelist to gather timeline events based on filtered events.

BUG fix

  • Fix libxft_comm_helper.so can't be found issue in multi-ranks mode.
  • Fix assert error in MLP when CAT_MLP opt is enabled.
  • Fix a w8a8 crash issue due to buffer size isn't big enough.
  • Correct GCC version for AVX512_BF16 instruction set.
  • Fix int32 overflow issue for larger size.

v1.2.0 - Qwen models and much more data types supported.

Models

  • Introduced Qwen models support and added the convert tool for Qwen models.
  • ChatGLM3 model is verfied and API supported.

Performance Optimizations

  • Update xDNN to version 1.4.2 to improve performance and support more data types.
  • Accelerate first token's generation with BF16-gemm Multi-Head Attention.

Functionality

  • Introduce more data types supports, including W8A8, INT4, and NF4. The hybrid data types between these new data types are supported.
  • Add accuracy evaluation script to assess the impact of different precisions on the text generation performance of the model.
  • Introduce XFT_VERBOSE macro to help profile model performance of each gemm. Set 1 to enable information ouput and default is 0.
  • Decouple oneCCL and MPI dependencies into a communication helper library. oneCCL environment is no longer needed when running in single-rank mode.

v1.1.0 - Baichuan models supported.

Models

  • Introduced Baichuan models support and added the convert tool for Baichuan models.

Performance Optimizations

  • Update xDNN to version 1.2.1 to improve performance of BF16 data type with AMX instruction on 4th generation Intel Xeon Scalable processors.
  • Improved performance of BF16 data type inference by adding matMul bf16bf16bf16 primitives and optimizing kernel selection strategy.
  • Improved performance of the model with unbalanced split allocation.

Functionality

  • Introduced prefix sharing feature.
  • Add sample strategy for token search, support temperature, top k, and top P parameter.
  • Introduce convert module to xfastertransformer python API.
  • Introduced grouped-query attention support for Llama2.
  • Auto-detect oneCCL environment and enter single-rank model if oneCCL does not exist.
  • Auto-detect oneCCL environment in compilation. If not detected, oneCCL will be built from source.
  • Add C++ exit function for multi-rank model.
  • Remove mklml 3rd party dependency.
  • Export normalization and position embedding C++ API, including alibi embedding and rotary embedding.
  • Introduced XFT_DEBUG_DIR environment value to specify the debug file directory.

BUG fix

  • Fix runtime issue of oneCCL shared memory model.
  • Fix path concat issue in convert tools.

This is the 1st official release of xFasterTransformer.🎇🎇🎇

Support models

  • ChatGLM-6B
  • ChatGLM2-6B
  • Llama 1, both 7B, 33B, and 65B
  • Llama 2, both 7B, 13B, and 70B
  • Opt larger than 1.3B

Features

  • Support Python and C++ API to integrate xFasterTransformer into the user's own solutions. Example codes are provided to demonstrate the usage.
  • Support hybrid data types such as BF16+FP16 and BF16+INT8 to accelerate the generation of the 1st token, in addition to supporting single data types like FP16, BF16, and INT8.
  • Support multiple instances to accelerate model inference, both locally and through the network.
  • Support Intel AMX instruction on 4th generation Intel Xeon Scalable processors.
  • Support 4th generation Intel Xeon Scalable processors with HBM which has a higher memory bandwidth and shows a much better performance on LLM.
  • Provide web demo scripts for users to show the performance of LLM models optimized by xFasterTransformer.
  • Support multiple distribution methods, both PyPI and docker images.