GitLab

Adds opt-in deterministic training mode to SuperBench's PyTorch model
benchmarks. When enabled --enable-determinism. PyTorch deterministic
algorithms are enforced, and per-step numerical fingerprints (loss,
activation means) are recorded as metrics. These can be compared across
runs using the existing sb result diagnosis pipeline to verify bit-exact
reproducibility — useful for hardware validation and platform
comparison.
 
Flags added - 

--enable-determinism
--check-frequency: Number of steps after which you want the metrics to
be recorded
--deterministic-seed

Changes - 

Updated pytorch_base.py to handle deterministic settings, logging.
Added a new example script: pytorch_deterministic_example.py
Added a test file: test_pytorch_determinism_all.py to verify everything
works as expected.

Usage - 

Step 1: Run 1 - Run with --enable-determinism and the necessary metrics
will be recorded in the results-summary.jsonl file
Step 2: Generate the baseline file from the Run 1 results using - sb
result generate-baseline
Step 3: Run 2 - Run with --enable-determinism and the necessary metrics
will be recorded in the results-summary.jsonl file on a different
machine (or the same machine)
Step 4: Run diagnosis on the results generated from the 2 runs using the
- sb result diagnosis command

Note - 
1. Make sure all the parameters are constant between the 2 runs 
2. Running the diagnosis command requires the rules.yaml file

---------
Co-authored-by: Aishwarya Tonpe <aishwarya.tonpe25@gmail.com>
Co-authored-by: Ubuntu <rdadmin@HPCPLTNODE0.n3kgq4m0lhoednrx3hxtad2nha.cdmx.internal.cloudapp.net>

To improve benchmark debugging, the following debug methods were added:

pytorch profiler in model benchmark

- SB_ENABLE_PYTORCH_PROFILER: switch to enable/disable
- SB_TORCH_PROFILER_TRACE_DIR: log path
These 2 runtime variables need to be configured in SB config file.

nsys in SB runner

- SB_ENABLE_NSYS: switch to enable/disable 
- SB_NSYS_TRACE_DIR: log path
These 2 runtime variables need to be configured in runner's ENV

---------
Co-authored-by: Hongtao Zhang <hongtaozhang@microsoft.com>

**Description**
add option to exclude data copy time in model benchmarks.

**Major Revision**
- add an option --no_copy
- move start time after data copy finish

**Description**

Cherry-pick bug fixes from v0.12.0 to main.

**Major Revisions**

* #725
* #727
* #728
Co-authored-by: Hongtao Zhang <hongtaozhang@microsoft.com>
Co-authored-by: Yifan Xiong <yixio@microsoft.com>
Co-authored-by: Guoshuai Zhao <guzhao@microsoft.com>

---------
Co-authored-by: Hongtao Zhang <hongtaozhang@microsoft.com>

Added MoE model using MixtralConfig. 

1. Added 8x7b and 8x22b variants 
2. Requires high VRAM as all experts are loaded in memory. Thus,
disabled training due to memory constraint on test worker.

---------
Co-authored-by: Hongtao Zhang <garyworkzht@gmail.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Hongtao Zhang <hongtaozhang@microsoft.com>

**Description**
Add deepseek megatron-lm benchmark.

---------
Co-authored-by: yukirora <yuting.jiang@microsoft.com>
Co-authored-by: Hongtao Zhang <garyworkzht@gmail.com>
Co-authored-by: Hongtao Zhang <hongtaozhang@microsoft.com>

**Description**
Add cuda 12.9 dockerfile and build in pipeline.

---------
Co-authored-by: Guoshuai Zhao <microsoft@microsoft.com>
Co-authored-by: Hongtao Zhang <hongtaozhang@microsoft.com>
Co-authored-by: Hongtao Zhang <garyworkzht@gmail.com>

Added llama benchmark - training and inference in accordance with the
existing pytorch models implementation like gpt2, lstm etc.

- added llama fp8 unit test for better code coverage, to reduce memory
required
- updated transformers version >= 4.28.0 for LLamaConfig
- set tokenizers version <= 0.20.3 to avoid 0.20.4 version
[issues](https://github.com/huggingface/tokenizers/issues/1691

) with
py3.8
- added llama2 to tensorrt
- llama2 tests not added to test_tensorrt_inference_performance.py due
to large memory requirement for worker gpu. tests validated separately
on gh200

---------
Co-authored-by: dpatlolla <dpatlolla@microsoft.com>

**Description**

Cherry-pick bug fixes from v0.10.0 to main.

**Major Revisions**

* Benchmarks: Microbenchmark - Support different hipblasLt data types in dist_inference #590
* Benchmarks: Microbenchmark - Support in-place for NCCL/RCCL benchmark #591
* Bug Fix - Fix NUMA Domains Swap Issue in NDv4 Topology File #592
* Benchmarks: Microbenchmark - Add data type option for NCCL and RCCL tests #595
* Benchmarks: Bug Fix - Make metrics of dist-inference-cpp aligned with PyTorch version #596
* CI/CD - Add ndv5 topo file #597
* Benchmarks: Microbenchmark - Improve AMD GPU P2P performance with fine-grained GPU memory #593
* Benchmarks: Build Pipeline - fix nccl and nccl test version to 2.18.3 to resolve hang issue in cuda12.2 docker #599
* Dockerfile - Bug fix for rocm docker build and deploy #598
* Benchmarks: Microbenchmark - Adapt to hipblasLt data type changes #603
* Benchmarks: Micro benchmarks - Update hipblaslt metric unit to tflops #604
* Monitor - U...

**Description**
Megatron-LM/Megatron-Deepspeed GPT pretrain benchmark

**Description**
Update 404 outdate reference links.

**Description**
Model benchmarks can stop due to `num_steps` or `duration` config which
will take effect when the value is set greater than 0.
If both are set greater than 0, the earliest condition reached will
work.

**Description**

Cherry-pick bug fixes from v0.8.0 to main.

**Major Revisions**

* Monitor - Fix the cgroup version checking logic (#502)
* Benchmark - Fix matrix size overflow issue in cuBLASLt GEMM (#503)
* Fix wrong torch usage in communication wrapper for Distributed
Inference Benchmark (#505)
* Analyzer: Fix bug in python3.8 due to pandas api change (#504)
* Bug - Fix bug to get metric from cmd when error happens (#506)
* Monitor - Collect realtime GPU power when benchmarking (#507)
* Add num_workers argument in model benchmark (#511)
* Remove unreachable condition when write host list (#512)
* Update cuda11.8 image to cuda12.1 based on nvcr23.03 (#513)
* Doc - Fix wrong unit of cpu-memory-bw-latency in doc (#515)
* Docs - Upgrade version and release note (#508)
Co-authored-by: guoshzhao <guzhao@microsoft.com>
Co-authored-by: Ziyue Yang <ziyyang@microsoft.com>
Co-authored-by: Yuting Jiang <yutingjiang@microsoft.com>

__Description__

Update TE FP8 model conversion.

__Major Revisions__
* Add 16-byte alignment comment.
* Fix TE layer parameters type.

Support Transformer Engine FP8 in existing PyTorch BERT/GPT2 models by
converting linear/layernorm to TE layers.

Fix potential barrier timeout in init_process_group due to race
condition of using the same port. Change to different ports when running
multiple models sequentially in one process.
For example, when running vgg11/13/16/19, will use port 29501~29504
respectively.

**Description**

Cherry-pick bug fixes from v0.7.0 to main.

**Major Revisions**

* Benchmarks - Fix missing include in FP8 benchmark (#460)
* Fix bug in TE BERT model (#461)
* Doc - Update benchmark doc (#465)
* Bug: Fix bug for incorrect datatype judgement in cublas-function
source code (#464)
* Support `sb deploy` without pulling image (#466)
* Docs - Upgrade version and release note (#467)
Co-authored-by: Russell J. Hewett <russell.j.hewett@gmail.com>
Co-authored-by: Yuting Jiang <yutingjiang@microsoft.com>

Support FP8 in PyTorch BERT models:

* add fp8 hybrid/e4m3/e5m2 in precision arguments
* build BERT encoders with `te.TransformerLayer` to repalce
`transformers.BertModel`
* wrap forward steps with fp8 autocast

**Description**
Add stdout logging util module and enable real-time logging flushing in executor

**Major Revision**
- Add stdout logging util module to redirect stdout into file log
- enable stdout logging in executor to write benchmark output into both stdout and file `sb-bench.log`
- enable real-time log flushing in run_command of microbenchmarks through config `log_flushing`

**Minor Revision**
- add log_n_step args to enable regular step time log in model benchmarks 
- udpate related docs

**Description**

Cherry-pick  bug fixes from v0.5.0 to main.

**Major Revisions**

* Bug - Force to fix ort version as '1.10.0' (#343)
* Bug - Support no matching rules and unify the output name in result_summary (#345)
* Analyzer - Support regex in annotations of benchmark naming for metrics in rules (#344)
* Bug - Fix bugs in sync results on root rank for e2e model benchmarks (#342)
* Bug - Fix bug of duration feature for model benchmarks in distributed mode (#347)
* Docs - Upgrade version and release note (#348)
Co-authored-by: Yuting Jiang <v-yutjiang@microsoft.com>

**Description**
Use config `log_raw_data` to control whether log the raw data into file or not. The default value is `no`. We can set it as `yes` for some particular benchmarks to save the raw data into file, such as NCCL/RCCL test.

**Description**
Remove fp16 samples type converting time for training cnn and lstm inference.

**Description**
Remove fp16  samples type converting time for cnn and lstm models.

**Description**
The BatchNorm operator is not numerically stable in fp16.  PyTorch documentation recommends to keep the BN op in fp32 for fp16 AMP models.  Refer to https://pytorch.org/docs/stable/amp.html#ops-that-can-autocast-to-float32.  Preserving BN in fp32 for superbench more accurately reflects real workloads.

**Description**
Add support for pytorch>=1.9.0 of init_process_group.

**Major Revision**
- Use PrefixStore(TCPStore) to init_process_group manully for each model run

**Description**
Please write a brief description and link the related issue if have.

**Major Revision**
- Sync (do allreduce max) the E2E training results among all workers.
- Avoid using ':0' in metric name if there has only one rank having output.

**Description**
Add 50th, 90th, 95th, 99th, 99.9th latency metrics for ORT and pytorch inference benchmarks.

Add transformers for TensorRT inference.

**Description**
Unify metric names of benchmarks.

**Description**
Fix typo when set force_fp32 option.

**Description**
Add option `force_fp32` to use fp32 instead of tf32, only takes effect on Ampere or newer GPUs.

**Description**

Cherry-pick  bug fixes from v0.3.0 to main.

**Major Revisions**
* Docs - Upgrade version and release note (#209)
* Benchmarks: Build Pipeline - Update rccl-test git submodule to dc1ad48 (#210)
* Benchmarks: Update - Update benchmarks in configuration file (#208)
* CI/CD - Update GitHub Action VM (#211)
* Benchmarks: Fix Bug - Fix wrong parameters for gpu-sm-copy-bw in configuration examples (#203)
* CI/CD - Fix bug in build image for push event (#205)
* Benchmark: Fix Bug - fix error message of communication-computation-overlap (#204)
* Tool: Fix bug - Fix function naming issue in system info  (#200)
* CI/CD - Push images in GitHub Action (#202)
* Bug - Fix torch.distributed command for single node (#201)
* CLI - Integrate system info for node (#199)
* Benchmarks: Code Revision - Revise CMake files for microbenchmarks. (#196)
* CI/CD - Add ROCm image build in GitHub Actions (#194)
* Bug: Fix bug - fix bug of hipBusBandwidth build (#193)
* Benchmarks: Build Pipeline - Restore rocblas build logic (#197)
* Bug: Fix Bug - Add barrier before 'destroy_process_group' in model benchmarks (#198)
* Bug - Revise 'docker run' in sb deploy (#195)
* Bug - Fix Bug : fix bug of error param operations to operation in rccl-bw of hpe config (#190)
Co-authored-by: Yuting Jiang <v-yujiang@microsoft.com>
Co-authored-by: Guoshuai Zhao <guzhao@microsoft.com>
Co-authored-by: Ziyue Yang <ziyyang@microsoft.com>

**Description**
Set reduce type for current benchmarks' metrics, including model benchmarks and ShardingMatmul.

**Description**
Replace `sum(results) / len(results)` with `statistics.mean(results)`

__Description__
Cherry-pick bug fixes from v0.2.1 to main.

__Major Revisions__
* Fix bug of VGG models failed on A100 GPU with batch_size=128.
* Fix Ansible connection issue when running in localhost.
* Update version in packages and docs.

* add config file for ndv4.

* replace torch.optim.AdamW with transformers.AdamW.

Benchmarks: Add Feature - Add DistributedImpl and DistributedBackend arguments for micro benchmark. (#100)

Fix bugs and refine log in single GPU benchmarks:

* Fix none framework issue
* Fix empty parameter bug
* Remove missed mobilenet_v3 models
* Change benchmark registration log to debug level
* Add pid in logging
* Add missing benchmarks in default config
* Fix deprecated logging warn

* Clean up the cache.