Datasetpublic

Akshat/wan2.2-rocm-profiles

Wan2.2 Sequence Parallel ROCm Profiles (MI300X) This dataset contains PyTorch/Perfetto traces, offline execution logs, serving benchmarks, and comparative reports for Wan2.2-T2V-A14B sequence-parallel runs on AMD ROCm (gfx942, 8x MI300X node). Dataset Directory Structure reports/ / Root: wan22_rocm_sp_sweep_analysis.md: 3-way sequence parallel topology comparison report. wan22_profile_u4_r1_analysis.md: Detailed analysis of the Ulysses-4 topology.… See the full description on the dataset page: https://huggingface.co/datasets/Akshat/wan2.2-rocm-profiles.

sourceHugging Facecc-by-4.0updated 4mo agoView on Hugging Face

2likes88downloads

Dataset Card

Wan2.2 Sequence Parallel ROCm Profiles (MI300X)

This dataset contains PyTorch/Perfetto traces, offline execution logs, serving benchmarks, and comparative reports for Wan2.2-T2V-A14B sequence-parallel runs on AMD ROCm (gfx942, 8x MI300X node).

Dataset Directory Structure

—`reports/` / Root:
—wan22_rocm_sp_sweep_analysis.md: 3-way sequence parallel topology comparison report.
—wan22_profile_u4_r1_analysis.md: Detailed analysis of the Ulysses-4 topology.
—wan22_profile_u4_r2_analysis.md: Detailed analysis of the Ulysses-4 + Ring-2 hybrid topology.
—wan22_profile_u8_r1_analysis.md: Detailed analysis of the Ulysses-8 topology.
—wan22_repro_commands.md: Markdown file outlining the exact shell/python commands used to run the sweeps.
—wan22_rocm_env_flags.sh: Script containing the environment flags set during all runs.
—wan2_2_profileROCm_collect_env.txt: Complete PyTorch / ROCm collect-env profile of the docker system environment.
—`wan22_profile/` (Ulysses-8, 8 GPUs):
—torch_profile/: Raw PyTorch JSON Chrome traces per rank (trace_rank*.json.gz).
—run.log: Console log output.
—`wan22_profile_u4_r1/` (Ulysses-4, 4 GPUs):
—torch_profile/: Raw PyTorch JSON Chrome traces per rank.
—run.log: Console log output.
—`wan22_profile_u4_r2/` (Ulysses-4 + Ring-2, 8 GPUs):
—torch_profile/: Raw PyTorch JSON Chrome traces per rank.
—run.log: Console log output.
—`wan22_serving_bench/`:
—Online serving benchmarks, JSON performance records, and launching scripts.
—`profiles.tar.gz`:
—Full archive containing all the raw compressed chrome trace files (1.8 GB).

Environment Setup & Reproduction Flags

[!IMPORTANT] All optimization flags and environment configurations specified inside wan22_rocm_env_flags.sh (e.g. VLLM_ROCM_USE_AITER=1, etc.) were set to True/Active during all profiling, baseline, and serving benchmark sweeps.

During all profiling and serving sweeps, the system environment variables and optimizations were set using:

bash

source wan22_rocm_env_flags.sh set

These flags configure:

—AITER acceleration: VLLM_ROCM_USE_AITER=1 (validated for FlashAttention and GroupNorm kernels)
—Attention backend: DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
—ROCm specifics: Target architecture gfx942, multiproc method set to spawn, and disabling direct convolution debug modes to stabilize execution.

The exact software environment details (PyTorch, Triton, Python, ROCm SDK version, driver, and library versions) are documented in wan2_2_profileROCm_collect_env.txt.

Performance Summary (Offline Inference)

Workload: Wan2.2-T2V-A14B, 2 steps, 8 frames, 512x512

Topology / Config	Active GPUs	DiT E2E Time (s)	VAE Decode Time (s)	E2E Gen Time (s)
Ulysses-4 (`u4_r1`)	4	1.148 s	0.444 s	1.784 s
Ulysses-8 (`wan22_profile`)	8	9.976 s	0.184 s	10.100 s
Ulysses-4 + Ring-2 (`u4_r2`)	8	10.624 s	0.184 s	11.048 s

Performance Summary (Online Serving)

Workload: Wan2.2-T2V-A14B-Diffusers served via vLLM-Omni. Measure Mean and P99 Latency (seconds) at Concurrency = 1.

Dataset Config	Ulysses / TP	VAE Parallel	Mean Latency (s)	P99 Latency (s)	Gen Stage (s)	Peak Memory (GB)
Dataset A (480p)	Ulysses-4 / TP-1	VAE-1	12.64s	14.11s	11.25s	67.9 GB
Dataset A (480p)	Ulysses-4 / TP-1	VAE-4	10.45s	10.59s	7.73s	67.7 GB
Dataset A (480p)	Ulysses-8 / TP-1	VAE-8	8.02s	8.03s	6.90s	67.6 GB
Dataset B (720p)	Ulysses-4 / TP-1	VAE-1	47.24s	47.35s	43.67s	70.6 GB
Dataset B (720p)	Ulysses-4 / TP-1	VAE-4	38.26s	39.89s	36.31s	70.2 GB
Dataset C (Mix)	Ulysses-4 / TP-1	VAE-1	31.82s	47.43s	29.14s	71.1 GB
Dataset C (Mix)	Ulysses-4 / TP-1	VAE-4	25.34s	38.06s	23.31s	71.2 GB

Licensing

This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC-BY-4.0).