GitLab

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>

**Description**
Add FP4 precision support for cublaslt_gemm benchmark.

**Major Revision**
- Add new type `fp4e2m1` and `__nv_fp4_e2m1`.
- For FP4 matmul, precision of MatrixC (add) should be FP16, precision
of MatricD (output) should be FP4, otherwise, it will not work.
- Add macro `CUDA_VERSION` to resolve the compatibility issue of
different CUDA versions.

---------
Co-authored-by: Ubuntu <aiperf@aiperf000000.hp5z1gqeinfufbj2u3jcty5fme.cdmx.internal.cloudapp.net>
Co-authored-by: AVA <39534996+avazr@users.noreply.github.com>
Co-authored-by: Guoshuai Zhao <microsoft@microsoft.com>

**Description**
Enable autotuning as an opt-in mode when benchmarking cublasLt via
`cublaslt_gemm`

The implementation is based on
https://github.com/NVIDIA/CUDALibrarySamples/blob/master/cuBLASLt/LtSgemmSimpleAutoTuning/sample_cublasLt_LtSgemmSimpleAutoTuning.cu

The behavior of original benchmark command remains unchanged, e.g.:
- `cublaslt_gemm -m 2048 -n 12288 -k 1536 -w10000 -i 1000 -t fp8e4m3`

The new opt-in options are `-a` (for autotune) and `-I` (for autotune
iterations, default is 50, same as the default for `-i`) and `-W` (for
autotune warmups, default=20, same as the default for `-w`), e.g.:
- `cublaslt_gemm -m 2048 -n 12288 -k 1536 -w 10000 -i 1000 -t fp8e4m3
-a`
- `cublaslt_gemm -m 2048 -n 12288 -k 1536 -w 10000 -i 1000 -t fp8e4m3 -a
-I 10 -W 10`

**Note:** This PR also changes the default `gemm_compute_type` for BF16
and FP16 to `CUBLAS_COMPUTE_32F`.

**Further observations:** 
1. The support matrix of the `cublaslt_gemm` could be further extended
in the future to support non-FP16 output as well for FP8 inputs.
2. Currently, the input matrices are initialized with values of 1.0 and
2.0 which makes them less demanding in terms of power. Another future
extension could be to enable another fill mode for, say, uniform random
numbers between -1 and 1.
3. cuBLAS workspace recommendations are listed under
https://docs.nvidia.com/cuda/cublas/#cublassetworkspace



Update (June 10, 2025): verified using higher level test driver with
these commands:

1. inline:
```
python3 -c "                                                                            
from superbench.benchmarks import BenchmarkRegistry, Platform
from superbench.common.utils import logger

parameters = (
    '--num_warmup 10 --num_steps 50 '
    '--shapes 512,512,512 1024,1024,1024 --in_types fp16 fp32 '
    '--enable_autotune --num_warmup_autotune 20 --num_steps_autotune 50'
)
context = BenchmarkRegistry.create_benchmark_context(
    'cublaslt-gemm', platform=Platform.CUDA, parameters=parameters
)
benchmark = BenchmarkRegistry.launch_benchmark(context)
logger.info('Result: {}'.format(benchmark.result))
"
```

2. newly added script: 
`python3 examples/benchmarks/cublaslt_function.py`

---------
Co-authored-by: Babak Hejazi <babakh@nvidia.com>

**Description**
Added support for Grace CPU neo2 architecture in CPU Stream. Now CPU
Stream supports dual socket benchmarking.

Example config for this arch support:
```yaml
    cpu-stream:numa0:
      timeout: *default_timeout
      modes:
      - name: local
        parallel: no
      parameters:
        cpu_arch: neo2
        numa_mem_nodes: 0
        cores: 0 1 2 3 4 5 6 7 8
    cpu-stream:numa1:
      timeout: *default_timeout
      modes:
      - name: local
        parallel: no
      parameters:
        cpu_arch: neo2
        numa_mem_nodes: 1
        cores: 64 65 66 67 68 69 70 71 72
    cpu-stream:numa-spread:
      timeout: *default_timeout
      modes:
      - name: local
        parallel: no
      parameters:
        cpu_arch: neo2
        numa_mem_nodes: 0 1
        cores: 0 1 2 3 4 5 6 7 8 64 65 66 67 68 69 70 71 72
```

---------
Co-authored-by: dpower4 <dilipreddi@gmail.com>

Added GPU Stream benchmark - measures the GPU memory bandwidth and
efficiency for double datatype through various memory operations
including copy, scale, add, and triad.
- added documentation for `gpu-stream` detailing its introduction,
metrics, and descriptions.
- added unit tests for `gpu-stream`. Example output is in
`superbenchmark/tests/data/gpu_stream.log`.

In the current implementation, the CPU‑stream benchmark code renames the
binary before the microbench base class can verify its existence,
causing the default‐binary check to fail.

This PR adds a “default” binary—built with the standard compile
parameters—so that the base class can always find and validate it. Once
the default binary is in place, the CPU‑stream code will rename it as
needed and re‑check its presence before running the benchmark.

The PR also enable CPU stream in the default settings.

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

adding gb200 cuda arch flag for cublaslt compilation

**Description**
Updated docker for 12.8
Use cutlass latest relase 3.8 with ARCH 100(blackwell) support
add latest nccl-test release with ARCH 100(blackwell) 
Updated msccl to support build for sm_100
No breaking changes, so backward compatible tested with  cuda 12.4

---------
Co-authored-by: Hongtao Zhang <garyworkzht@gmail.com>

Improves logging info for diagnosis rule op baseline errors. This allows
developers to easily detect errors in their rule files as well as
baseline files, improving end-user experience.

Co-authored-by: Yifan Xiong <yifan.xiong@microsoft.com>
Co-authored-by: Hongtao Zhang <garyworkzht@gmail.com>

Root Cause:

1. '_get_all_test_cases()' was called in '_parser' while '_parser' was
defined in the base class.
2.  in '_get_all_test_cases()', cmd path was not included.

Fix:

1. Remove '_get_all_test_cases()' from '_parser'.
2. Construct path for cmd.

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

**Description**

1. Fixed the bug that nvbandwidth benchmark need to handle 'N/A' values
in nvbandwidth cmd output.
2. Replaced the input format of test cases with a list.
3. Add nvbandwidth configuration example in default config files.

---------
Co-authored-by: hongtaozhang <hongtaozhang@microsoft.com>
Co-authored-by: Yifan Xiong <yifan.xiong@microsoft.com>

**Description**
Today I was running a benchmark on my machine. And encountered a fancy
issue with tensorrt-inference.
I got code 33, which according to the source code is:
```
MICROBENCHMARK_RESULT_PARSING_FAILURE = 33
```
I dived deep into the code and found out the following problem. The
parser stumbled upon getting to the following line:
```
[11/28/2024-17:03:11] [I] Latency: min = 7.2793 ms, max = 10.1606 ms, mean = 7.41642 ms, median = 7.39551 ms, percentile(99%) = 8 ms
```
I ran it separately on the code and found out that the regular
expression was not suitable for the cases like this, when you encounter
an INT as a result in milliseconds.
That's why this pull request is created.
I came up with the closest possible regular expression to fix this issue
and not to introduce any other bug.

**Major Revision**
- 0.11.0

**Description**
Introduce architecture support for version 10.0 in gemm-flops.

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>

Fix ordering of args in err messages.

Upgrade dependency versions in Azure pipeline:

* Remove Python 3.6 and add Python 3.10 for cpu-unit-test
* Upgrade CUDA from 11.1 to 12.4 for cuda-unit-test
* Update labels accordingly

---------
Co-authored-by: Dilip Patlolla <dilipreddi@gmail.com>

**Description**

Add nvbandwidth benchmark.

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

**Description**
Add micro benchmark to measure general CPU bandwidth and latency without 'mlc'.

Test output:
```
{
"cpu-memory-bw-latency/return_code": 0,
"cpu-memory-bw-latency/mem_bandwidth_matrix_numa_0_1_bw": 5388.75021,
"cpu-memory-bw-latency/mem_bandwidth_matrix_numa_0_1_lat": 0.185571786,
"cpu-memory-bw-latency/mem_bandwidth_matrix_numa_1_0_bw": 4634.82028,
"cpu-memory-bw-latency/mem_bandwidth_matrix_numa_1_0_lat": 0.215758096,
}
```

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

Add support for arm64 build:

- Updated dockerfile for arm64 build
- extend cpu stream compilation for neoverse 
- handle onnxruntime-gpu installation
- third party builds filtering based on arch
- disable cuda decode perf build for non x86

The current GPU Copy BW Performance fails on Nvidia Grace systems. This
is due to the memory only numa node and thus the numa_run_on_node fails
for such nodes and halts completely.

This fix checks for the presence of assigned CPU cores for the numa
node, on checking if it has no cpu cores assigned, it skips that
specific node during the args creation and continues.

**Description**
Cherry pick bug fixes from v0.11.0 to main

**Major Revision**
* #645 
* #648 
* #646 
* #647 
* #651 
* #652 
* #650

---------
Co-authored-by: hongtaozhang <hongtaozhang@microsoft.com>
Co-authored-by: Yifan Xiong <yifan.xiong@microsoft.com>

**Description**
Fix executor for Benchmark Execution Without Explicit Framework Field

**Description**
Fix bug of failure test and warning of pandas in data diagnosis.

**Major Revision**
- fix warning of pandas in replace and fillna due to type downcast
- fix bug of failure check function only check one matched metric rather
than all matched metrics
- fix bug when converting regex into str of metrics when there're more
than one match group

Add 10-hpcx.sh to /etc/profile.d
Update the Docker exec command to ensure a persistent HPCX environment.

**Description**
Add support GPU ARCH 8.9 for NVIDIA L4/L40/L40s GPUs in gemm-flops.

Update `omegaconf` version to
[2.3.0](https://pypi.org/project/omegaconf/2.3.0/) as omegaconf 2.0.6
has a non-standard dependency specifier PyYAML>=5.1.*. pip 24.1 will
enforce this behaviour change.
Discussion can be found at https://github.com/pypa/pip/issues/12063.

**Description**
Adds hipblasLt tuning to dist-inference cpp implementation.

**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**
`add_compile_options` will not work for ROCm build, change it to setting
`CMAKE_CXX_FLAGS`.

**Description**
Add distributed inference benchmark cpp implementation.

**Description**
upgrade to rocm5.7 dockerfile.

---------
Co-authored-by: yukirora <yuting.jiang@microsoft.com>

Benchmarks: Micro benchmark - Add one-to-all, all-to-one, all-to-all support to gpu_copy_bw_performance (#588)

**Description**
Add one-to-all, all-to-one, all-to-all support to
gpu_copy_bw_performance, and fix performance bug in gpu_copy

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

**Description**
Revise NCCL/RCCL benchmarks to graph mode add latency metrics.

**Description**
Benchmarks: micro benchmark - Support cpu-gpu and gpu-cpu in
ib-validation

**Major Revision**
- Support cpu-gpu and gpu-cpu in ib-validation


**Minor Revision**
- support multi msg size, multi direction, multi ib commands in
ib-validation

**Description**
Add AMD support in monitor.

**Major Revision**
- Add library pyrsmi to collect metrics.
- Currently can get device_utilization, device_power, device_used_memory
and device_total_memory.

Upgrade Docker image to CUDA 12.2 for H100:
* upgrade base image to 23.10
* fix onnxruntime version in python3.10
* fix compilation errors