Continum-Computing/ssm-bench

Fix the mojibake and stale platform comment in .gitignore 22 days ago

good first issue packaging

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Mark each README metric as measured or planned 22 days ago

documentation good first issue

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Benchmark the remaining GPU targets: H100, L40S, L4 22 days ago

help wanted hardware

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Implement a hybrid (Jamba-style) model 22 days ago

help wanted model-coverage

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Make the Makefile work on Windows or state that it is POSIX-only 22 days ago

good first issue tooling

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Add tests for profiling/, including a T4-versus-2080Ti peak regression test 22 days ago

help wanted testing

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Add CI running make test and make lint 22 days ago

help wanted ci

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Pin the git ref in modal_benchmark.py 22 days ago

good first issue reproducibility

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Correct the Tesla T4 peak values in the results-schema example 22 days ago

documentation good first issue

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Resolve the empty configs/ directory 22 days ago

documentation good first issue

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Implement energy per token measurement via pynvml 22 days ago

help wanted gate baseline

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Add CUDA graph capture and decode-step timing 22 days ago

help wanted gate methodology

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Integrate official Mamba CUDA kernels (mamba-ssm) for the SSM path 22 days ago

help wanted gate baseline

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Integrate FlashAttention-2 as the attention baseline 22 days ago

help wanted gate baseline

Continum-Computing/ssm-bench

0

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Python

#arithmetic-intensity#attention-mechanism#cuda#deep-learning#flash-attention#gpu-benchmarking#hardware-algorithm-co-design#hybrid-models#jamba#mamba#mamba-2#mamba-3#memory-bandwidth#ml-inference#nvidia-gpu#pytorch#roofline-analysis#ssm#state-space-models#throughput-per-watt

Open Issues Need Help