diff --git a/benchmark_configs/xdit/flux.yaml b/benchmark_configs/xdit/flux.yaml index f97d583..ffe6635 100644 --- a/benchmark_configs/xdit/flux.yaml +++ b/benchmark_configs/xdit/flux.yaml @@ -228,10 +228,10 @@ warmup_calls: 2 num_iterations: 5 guidance_scale: 0.0 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 use_torch_compile: true use_fp8_gemms: true - + cache_method: dbcache - name: flux.usp_1k.2gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.1-dev @@ -246,14 +246,12 @@ warmup_calls: 2 num_iterations: 5 guidance_scale: 0.0 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 ulysses_degree: 2 - ring_degree: 1 use_torch_compile: true use_fp8_gemms: true - enable_tiling: true use_parallel_vae: true - + cache_method: dbcache - name: flux.usp_1k.4gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.1-dev @@ -268,13 +266,12 @@ warmup_calls: 2 num_iterations: 5 guidance_scale: 0.0 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 + use_torch_compile: true use_fp8_gemms: true - enable_tiling: true use_parallel_vae: true - + cache_method: dbcache - name: flux.2k.1gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.1-dev @@ -289,11 +286,11 @@ warmup_calls: 2 num_iterations: 5 guidance_scale: 0.0 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - enable_tiling: true use_fp8_gemms: true - + enable_tiling: true + cache_method: dbcache - name: flux.usp_2k.2gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.1-dev @@ -308,14 +305,13 @@ warmup_calls: 2 num_iterations: 5 guidance_scale: 0.0 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 ulysses_degree: 2 - ring_degree: 1 use_torch_compile: true - enable_tiling: true - use_parallel_vae: true use_fp8_gemms: true - + use_parallel_vae: true + enable_tiling: true + cache_method: dbcache - name: flux.usp_2k.4gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.1-dev @@ -330,10 +326,10 @@ warmup_calls: 2 num_iterations: 5 guidance_scale: 0.0 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 use_torch_compile: true use_fp8_gemms: true - enable_tiling: true use_parallel_vae: true + enable_tiling: true + cache_method: dbcache diff --git a/benchmark_configs/xdit/flux2.yaml b/benchmark_configs/xdit/flux2.yaml index c631297..a929285 100644 --- a/benchmark_configs/xdit/flux2.yaml +++ b/benchmark_configs/xdit/flux2.yaml @@ -623,18 +623,17 @@ prompt: Add a cool hat to the cat height: 1024 width: 1024 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 - input_images: /app/data/flux_cat.png guidance_scale: 4.0 - attention_backend: aiter_flydsl - ulysses_degree: 1 - ring_degree: 1 + input_images: /app/data/flux_cat.png + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - enable_sequential_cpu_offload: true - + use_fp8_gemms: true + enable_group_cpu_offload: true + cache_method: dbcache - name: flux2.i2i_1k.2gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -644,21 +643,19 @@ prompt: Add a cool hat to the cat height: 1024 width: 1024 - num_inference_steps: 4 - max_sequence_length: 256 + num_inference_steps: 50 + max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 - input_images: /app/data/flux_cat.png guidance_scale: 4.0 - attention_backend: aiter_flydsl + input_images: /app/data/flux_cat.png + attention_backend: aiter_flydsl_fp8 ulysses_degree: 2 - ring_degree: 1 fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true use_fp8_gemms: true - + cache_method: dbcache - name: flux2.i2i_1k.4gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -668,22 +665,19 @@ prompt: Add a cool hat to the cat height: 1024 width: 1024 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 - input_images: /app/data/flux_cat.png guidance_scale: 4.0 - attention_backend: aiter_flydsl + input_images: /app/data/flux_cat.png + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true - enable_slicing: true use_fp8_gemms: true - + cache_method: dbcache - name: flux2.t2i_1k.1gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -693,17 +687,16 @@ prompt: Cat astronaut raising a cat flag on the surface of moon height: 1024 width: 1024 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 guidance_scale: 4.0 - attention_backend: aiter_flydsl - ulysses_degree: 1 - ring_degree: 1 + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - enable_sequential_cpu_offload: true - + use_fp8_gemms: true + enable_group_cpu_offload: true + cache_method: dbcache - name: flux2.t2i_1k.2gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -713,20 +706,18 @@ prompt: Cat astronaut raising a cat flag on the surface of moon height: 1024 width: 1024 - num_inference_steps: 4 - max_sequence_length: 256 + num_inference_steps: 50 + max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 guidance_scale: 4.0 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 ulysses_degree: 2 - ring_degree: 1 fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true use_fp8_gemms: true - + cache_method: dbcache - name: flux2.t2i_1k.4gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -736,20 +727,18 @@ prompt: Cat astronaut raising a cat flag on the surface of moon height: 1024 width: 1024 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 guidance_scale: 4.0 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true use_fp8_gemms: true - + cache_method: dbcache - name: flux2.multi_1k.1gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -759,20 +748,19 @@ prompt: the cat from image 1 has the same expression as cat on image 2 height: 1024 width: 1024 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 - input_images: - - /app/data/flux_cat.png - - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg guidance_scale: 4.0 - attention_backend: aiter_flydsl - ulysses_degree: 1 - ring_degree: 1 + input_images: + - /app/data/flux_cat.png + - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - enable_sequential_cpu_offload: true - + use_fp8_gemms: true + enable_group_cpu_offload: true + cache_method: dbcache - name: flux2.multi_1k.2gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -782,23 +770,21 @@ prompt: the cat from image 1 has the same expression as cat on image 2 height: 1024 width: 1024 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 - input_images: - - /app/data/flux_cat.png - - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg guidance_scale: 4.0 - attention_backend: aiter_flydsl + input_images: + - /app/data/flux_cat.png + - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg + attention_backend: aiter_flydsl_fp8 ulysses_degree: 2 - ring_degree: 1 fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true use_fp8_gemms: true - + cache_method: dbcache - name: flux2.multi_1k.4gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -808,23 +794,21 @@ prompt: the cat from image 1 has the same expression as cat on image 2 height: 1024 width: 1024 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 - input_images: - - /app/data/flux_cat.png - - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg guidance_scale: 4.0 - attention_backend: aiter_flydsl + input_images: + - /app/data/flux_cat.png + - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 fully_shard_degree: 4 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true use_fp8_gemms: true - + cache_method: dbcache - name: flux2.i2i_2k.1gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -834,19 +818,18 @@ prompt: Add a cool hat to the cat height: 2048 width: 2048 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 - input_images: /app/data/flux_cat.png guidance_scale: 4.0 - attention_backend: aiter_flydsl - ulysses_degree: 1 - ring_degree: 1 + input_images: /app/data/flux_cat.png + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - enable_sequential_cpu_offload: true + use_fp8_gemms: true enable_tiling: true - + enable_group_cpu_offload: true + cache_method: dbcache - name: flux2.i2i_2k.2gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -856,21 +839,21 @@ prompt: Add a cool hat to the cat height: 2048 width: 2048 - num_inference_steps: 4 - max_sequence_length: 512 + num_inference_steps: 50 + max_sequence_length: 256 warmup_calls: 2 num_iterations: 5 - input_images: /app/data/flux_cat.png guidance_scale: 4.0 - attention_backend: aiter_flydsl + input_images: /app/data/flux_cat.png + attention_backend: aiter_flydsl_fp8 ulysses_degree: 2 - ring_degree: 1 fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true use_fp8_gemms: true - + enable_tiling: true + vae_tile_latent_size: 64 + vae_tile_sample_size: 512 - name: flux2.i2i_2k.4gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -880,21 +863,20 @@ prompt: Add a cool hat to the cat height: 2048 width: 2048 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 - input_images: /app/data/flux_cat.png guidance_scale: 4.0 - attention_backend: aiter_flydsl + input_images: /app/data/flux_cat.png + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 - fully_shard_degree: 4 + fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true use_fp8_gemms: true - + enable_tiling: true + cache_method: dbcache - name: flux2.t2i_2k.1gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -904,19 +886,17 @@ prompt: Cat astronaut raising a cat flag on the surface of moon height: 2048 width: 2048 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 guidance_scale: 4.0 - attention_backend: aiter_flydsl - ulysses_degree: 1 - ring_degree: 1 + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - enable_sequential_cpu_offload: true + use_fp8_gemms: true enable_tiling: true - -# NOTE: requires PYTORCH_HIP_ALLOC_CONF=expandable_segments:True to run successfully + enable_group_cpu_offload: true + cache_method: dbcache - name: flux2.t2i_2k.2gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -926,20 +906,19 @@ prompt: Cat astronaut raising a cat flag on the surface of moon height: 2048 width: 2048 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 guidance_scale: 4.0 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 ulysses_degree: 2 - ring_degree: 1 fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true use_fp8_gemms: true - + enable_tiling: true + cache_method: dbcache - name: flux2.t2i_2k.4gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -949,20 +928,19 @@ prompt: Cat astronaut raising a cat flag on the surface of moon height: 2048 width: 2048 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 + guidance_scale: 4.0 + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 fully_shard_degree: 4 memory_efficient_sharding: true - guidance_scale: 4.0 - attention_backend: aiter_flydsl use_torch_compile: true - enable_tiling: true use_fp8_gemms: true - + enable_tiling: true + cache_method: dbcache - name: flux2.multi_2k.1gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -972,21 +950,20 @@ prompt: the cat from image 1 has the same expression as cat on image 2 height: 2048 width: 2048 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 - input_images: - - /app/data/flux_cat.png - - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg guidance_scale: 4.0 - attention_backend: aiter_flydsl - ulysses_degree: 1 - ring_degree: 1 + input_images: + - /app/data/flux_cat.png + - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - enable_sequential_cpu_offload: true + use_fp8_gemms: true enable_tiling: true - + enable_group_cpu_offload: true + cache_method: dbcache - name: flux2.multi_2k.2gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -996,23 +973,26 @@ prompt: the cat from image 1 has the same expression as cat on image 2 height: 2048 width: 2048 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 - input_images: - - /app/data/flux_cat.png - - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg guidance_scale: 4.0 + input_images: + - /app/data/flux_cat.png + - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg attention_backend: aiter_flydsl ulysses_degree: 2 - ring_degree: 1 fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true use_fp8_gemms: true - + enable_tiling: true + vae_tile_latent_size: 64 + vae_tile_sample_size: 512 + cache_method: dbcache + cache_config: + Fn_compute_blocks: 1 - name: flux2.multi_2k.4gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-dev @@ -1022,19 +1002,19 @@ prompt: the cat from image 1 has the same expression as cat on image 2 height: 2048 width: 2048 - num_inference_steps: 4 + num_inference_steps: 50 max_sequence_length: 512 warmup_calls: 2 num_iterations: 5 + guidance_scale: 4.0 + input_images: + - /app/data/flux_cat.png + - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 fully_shard_degree: 4 memory_efficient_sharding: true - input_images: - - /app/data/flux_cat.png - - https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/grumpy.jpg - guidance_scale: 4.0 - attention_backend: aiter_flydsl use_torch_compile: true - enable_tiling: true use_fp8_gemms: true + enable_tiling: true + cache_method: dbcache diff --git a/benchmark_configs/xdit/flux2_klein.yaml b/benchmark_configs/xdit/flux2_klein.yaml index 59759ca..2a01571 100644 --- a/benchmark_configs/xdit/flux2_klein.yaml +++ b/benchmark_configs/xdit/flux2_klein.yaml @@ -59,16 +59,13 @@ height: 2048 width: 2048 num_inference_steps: 4 - warmup_calls: 5 - num_iterations: 25 + warmup_calls: 2 + num_iterations: 5 guidance_scale: 1.0 - attention_backend: aiter_flydsl - ulysses_degree: 1 - ring_degree: 1 + attention_backend: aiter_flydsl_fp8 use_torch_compile: true enable_tiling: true use_fp8_gemms: true - - name: flux2_klein.2gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-klein-9B @@ -79,17 +76,15 @@ height: 2048 width: 2048 num_inference_steps: 4 - warmup_calls: 5 - num_iterations: 25 + warmup_calls: 2 + num_iterations: 5 guidance_scale: 1.0 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 ulysses_degree: 2 - ring_degree: 1 use_torch_compile: true + use_fp8_gemms: true enable_tiling: true use_parallel_vae: true - use_fp8_gemms: true - - name: flux2_klein.4gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.2-klein-9B @@ -100,13 +95,70 @@ height: 2048 width: 2048 num_inference_steps: 4 - warmup_calls: 5 - num_iterations: 25 + warmup_calls: 2 + num_iterations: 5 guidance_scale: 1.0 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 use_torch_compile: true use_fp8_gemms: true enable_tiling: true use_parallel_vae: true + +- name: flux2_klein_4b.1gpu.rdna4 + tags: [rdna4, gfx1201] + model: black-forest-labs/FLUX.2-klein-4B + runner: xdit + args: + seed: 42 + prompt: Cat astronaut raising a cat flag on the surface of the moon + height: 2048 + width: 2048 + num_inference_steps: 4 + warmup_calls: 2 + num_iterations: 5 + guidance_scale: 1.0 + attention_backend: aiter_flydsl_fp8 + use_torch_compile: true + use_fp8_gemms: true + enable_tiling: true + +- name: flux2_klein_4b.2gpu.rdna4 + tags: [rdna4, gfx1201] + model: black-forest-labs/FLUX.2-klein-4B + runner: xdit + args: + seed: 42 + prompt: Cat astronaut raising a cat flag on the surface of the moon + height: 2048 + width: 2048 + num_inference_steps: 4 + warmup_calls: 2 + num_iterations: 5 + guidance_scale: 1.0 + attention_backend: aiter_flydsl_fp8 + ulysses_degree: 2 + use_torch_compile: true + use_fp8_gemms: true + use_parallel_vae: true + enable_tiling: true + +- name: flux2_klein_4b.4gpu.rdna4 + tags: [rdna4, gfx1201] + model: black-forest-labs/FLUX.2-klein-4B + runner: xdit + args: + seed: 42 + prompt: Cat astronaut raising a cat flag on the surface of the moon + height: 2048 + width: 2048 + num_inference_steps: 4 + warmup_calls: 2 + num_iterations: 5 + guidance_scale: 1.0 + attention_backend: aiter_flydsl_fp8 + ulysses_degree: 4 + use_torch_compile: true + use_fp8_gemms: true + use_parallel_vae: true + enable_tiling: true diff --git a/benchmark_configs/xdit/flux_kontext.yaml b/benchmark_configs/xdit/flux_kontext.yaml index 751ffd9..b7ce619 100644 --- a/benchmark_configs/xdit/flux_kontext.yaml +++ b/benchmark_configs/xdit/flux_kontext.yaml @@ -64,19 +64,17 @@ prompt: A small cat sitting in a garden height: 1024 width: 1024 - num_inference_steps: 25 + num_inference_steps: 30 max_sequence_length: 256 warmup_calls: 2 num_iterations: 5 guidance_scale: 2.5 - attention_backend: aiter_flydsl - ulysses_degree: 1 - ring_degree: 1 input_images: /app/data/flux_cat.png resize_input_images: true + attention_backend: aiter_flydsl_fp8 use_torch_compile: true use_fp8_gemms: true - + cache_method: dbcache - name: flux_kontext.2gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.1-Kontext-dev @@ -86,21 +84,19 @@ prompt: A small cat sitting in a garden height: 1024 width: 1024 - num_inference_steps: 25 + num_inference_steps: 30 max_sequence_length: 256 warmup_calls: 2 num_iterations: 5 guidance_scale: 2.5 - attention_backend: aiter_flydsl - ulysses_degree: 2 - ring_degree: 1 input_images: /app/data/flux_cat.png resize_input_images: true + attention_backend: aiter_flydsl_fp8 + ulysses_degree: 2 use_torch_compile: true - enable_tiling: true - use_parallel_vae: true use_fp8_gemms: true - + use_parallel_vae: true + cache_method: dbcache - name: flux_kontext.4gpu.rdna4 tags: [rdna4, gfx1201] model: black-forest-labs/FLUX.1-Kontext-dev @@ -110,17 +106,16 @@ prompt: A small cat sitting in a garden height: 1024 width: 1024 - num_inference_steps: 25 + num_inference_steps: 30 max_sequence_length: 256 warmup_calls: 2 num_iterations: 5 guidance_scale: 2.5 - attention_backend: aiter_flydsl - ulysses_degree: 4 - ring_degree: 1 input_images: /app/data/flux_cat.png resize_input_images: true + attention_backend: aiter_flydsl_fp8 + ulysses_degree: 4 use_torch_compile: true use_fp8_gemms: true - enable_tiling: true use_parallel_vae: true + cache_method: dbcache diff --git a/benchmark_configs/xdit/qwen_image.yaml b/benchmark_configs/xdit/qwen_image.yaml index 30499f6..a3e2a56 100644 --- a/benchmark_configs/xdit/qwen_image.yaml +++ b/benchmark_configs/xdit/qwen_image.yaml @@ -51,65 +51,66 @@ model: Qwen/Qwen-Image-2512 runner: xdit args: + seed: 42 prompt: A cat holding a sign that says hello world height: 2048 width: 2048 - seed: 42 - num_inference_steps: 4 - guidance_scale: 0.0 + num_inference_steps: 50 warmup_calls: 2 num_iterations: 5 - attention_backend: aiter_flydsl - ulysses_degree: 1 - ring_degree: 1 + guidance_scale: 0.0 + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - enable_tiling: true - enable_slicing: true use_fp8_gemms: true enable_model_cpu_offload: true - + enable_tiling: true + cache_method: dbcache + cache_config: + residual_diff_threshold: 0.1 - name: qwen_image.2gpu.rdna4 tags: [rdna4, gfx1201] model: Qwen/Qwen-Image-2512 runner: xdit args: + seed: 42 prompt: A cat holding a sign that says hello world height: 2048 width: 2048 - seed: 42 - num_inference_steps: 4 - guidance_scale: 0.0 + num_inference_steps: 50 warmup_calls: 2 num_iterations: 5 - attention_backend: aiter_flydsl + guidance_scale: 0.0 + attention_backend: aiter_flydsl_fp8 ulysses_degree: 2 - ring_degree: 1 fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true - enable_slicing: true use_fp8_gemms: true - + use_parallel_vae: true + enable_tiling: true + cache_method: dbcache + cache_config: + residual_diff_threshold: 0.1 - name: qwen_image.4gpu.rdna4 tags: [rdna4, gfx1201] model: Qwen/Qwen-Image-2512 runner: xdit args: + seed: 42 prompt: A cat holding a sign that says hello world height: 2048 width: 2048 - seed: 42 - num_inference_steps: 4 - guidance_scale: 0.0 + num_inference_steps: 50 warmup_calls: 2 num_iterations: 5 - attention_backend: aiter_flydsl + guidance_scale: 0.0 + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 - fully_shard_degree: 4 + fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true - enable_slicing: true use_fp8_gemms: true + enable_tiling: true + cache_method: dbcache + cache_config: + residual_diff_threshold: 0.1 diff --git a/benchmark_configs/xdit/qwen_image_edit.yaml b/benchmark_configs/xdit/qwen_image_edit.yaml index e7a383a..73462bd 100644 --- a/benchmark_configs/xdit/qwen_image_edit.yaml +++ b/benchmark_configs/xdit/qwen_image_edit.yaml @@ -57,71 +57,65 @@ model: Qwen/Qwen-Image-Edit runner: xdit args: + seed: 42 prompt: Add a cool hat to the cat. - negative_prompt: " " + negative_prompt: ' ' height: 2048 width: 2048 - input_images: /app/data/flux_cat.png - seed: 42 - num_inference_steps: 4 - guidance_scale: 4.0 + num_inference_steps: 50 warmup_calls: 2 num_iterations: 5 - attention_backend: aiter_flydsl - ulysses_degree: 1 - ring_degree: 1 + guidance_scale: 4.0 + input_images: /app/data/flux_cat.png + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - enable_tiling: true - enable_slicing: true use_fp8_gemms: true enable_model_cpu_offload: true - + enable_tiling: true + cache_method: dbcache - name: qwen_image_edit.2gpu.rdna4 tags: [rdna4, gfx1201] model: Qwen/Qwen-Image-Edit runner: xdit args: + seed: 42 prompt: Add a cool hat to the cat. - negative_prompt: " " + negative_prompt: ' ' height: 2048 width: 2048 - input_images: /app/data/flux_cat.png - seed: 42 - num_inference_steps: 4 - guidance_scale: 4.0 + num_inference_steps: 50 warmup_calls: 2 num_iterations: 5 - attention_backend: aiter_flydsl + guidance_scale: 4.0 + input_images: /app/data/flux_cat.png + attention_backend: aiter_flydsl_fp8 ulysses_degree: 2 - ring_degree: 1 fully_shard_degree: 2 memory_efficient_sharding: true use_torch_compile: true - enable_tiling: true - enable_slicing: true use_fp8_gemms: true - + enable_tiling: true + cache_method: dbcache - name: qwen_image_edit.4gpu.rdna4 tags: [rdna4, gfx1201] model: Qwen/Qwen-Image-Edit runner: xdit args: + seed: 42 prompt: Add a cool hat to the cat. - negative_prompt: " " + negative_prompt: ' ' height: 2048 width: 2048 - input_images: /app/data/flux_cat.png - seed: 42 - num_inference_steps: 4 - guidance_scale: 4.0 + num_inference_steps: 50 warmup_calls: 2 num_iterations: 5 - attention_backend: aiter_flydsl + guidance_scale: 4.0 + input_images: /app/data/flux_cat.png + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 - fully_shard_degree: 4 + fully_shard_degree: 2 memory_efficient_sharding: true - use_torch_compile: false - enable_tiling: true - enable_slicing: true + use_torch_compile: true use_fp8_gemms: true + enable_tiling: true + cache_method: dbcache diff --git a/benchmark_configs/xdit/stablediffusion_3_5.yaml b/benchmark_configs/xdit/stablediffusion_3_5.yaml index 192e21e..0dac75f 100644 --- a/benchmark_configs/xdit/stablediffusion_3_5.yaml +++ b/benchmark_configs/xdit/stablediffusion_3_5.yaml @@ -48,16 +48,15 @@ width: 1024 num_inference_steps: 28 warmup_calls: 2 - ulysses_degree: 1 - ring_degree: 1 - guidance_scale: 3.5 num_iterations: 5 - attention_backend: aiter_flydsl + guidance_scale: 3.5 + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - enable_model_cpu_offload: true enable_tiling: true enable_slicing: true - + vae_tile_latent_size: 64 + vae_tile_sample_size: 512 + cache_method: dbcache - name: stablediffusion_3_5.2gpu.rdna4 tags: [rdna4, gfx1201] model: stabilityai/stable-diffusion-3.5-large @@ -71,14 +70,14 @@ warmup_calls: 2 num_iterations: 5 guidance_scale: 3.5 - pipefusion_parallel_degree: 1 + attention_backend: aiter_flydsl_fp8 use_cfg_parallel: true - attention_backend: aiter_flydsl use_torch_compile: true - use_fp8_gemms: true + cache_method: dbcache enable_tiling: true enable_slicing: true - + vae_tile_latent_size: 64 + vae_tile_sample_size: 512 - name: stablediffusion_3_5.4gpu.rdna4 tags: [rdna4, gfx1201] model: stabilityai/stable-diffusion-3.5-large @@ -92,10 +91,12 @@ warmup_calls: 2 num_iterations: 5 guidance_scale: 3.5 - pipefusion_parallel_degree: 2 + attention_backend: aiter_flydsl_fp8 use_cfg_parallel: true - attention_backend: aiter_flydsl + ulysses_degree: 2 use_torch_compile: true - use_fp8_gemms: true + cache_method: dbcache enable_tiling: true - enable_slicing: true \ No newline at end of file + enable_slicing: true + vae_tile_latent_size: 64 + vae_tile_sample_size: 512 diff --git a/benchmark_configs/xdit/wan2_1.yaml b/benchmark_configs/xdit/wan2_1.yaml index 719acff..96fbc58 100644 --- a/benchmark_configs/xdit/wan2_1.yaml +++ b/benchmark_configs/xdit/wan2_1.yaml @@ -55,54 +55,3 @@ use_torch_compile: True attention_backend: cudnn -- name: wan2_1.2gpu.rdna4 - tags: [rdna4, gfx1201] - model: Wan-AI/Wan2.1-I2V-14B-720P-Diffusers - runner: xdit - args: - prompt: "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside." - height: 720 - width: 1280 - input_images: /app/data/wan_input.jpg - num_frames: 81 - ulysses_degree: 2 - ring_degree: 1 - fully_shard_degree: 2 - memory_efficient_sharding: true - use_parallel_vae: true - guidance_scale: 3.0 - seed: 123456 - warmup_calls: 1 - num_iterations: 1 - num_inference_steps: 40 - use_torch_compile: true - attention_backend: aiter_flydsl - enable_tiling: true - enable_slicing: true - use_fp8_gemms: true - -- name: wan2_1.4gpu.rdna4 - tags: [rdna4, gfx1201] - model: Wan-AI/Wan2.1-I2V-14B-720P-Diffusers - runner: xdit - args: - prompt: "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside." - height: 720 - width: 1280 - input_images: /app/data/wan_input.jpg - num_frames: 81 - ulysses_degree: 4 - ring_degree: 1 - fully_shard_degree: 4 - memory_efficient_sharding: true - use_parallel_vae: true - guidance_scale: 3.0 - seed: 123456 - warmup_calls: 1 - num_iterations: 1 - num_inference_steps: 40 - use_torch_compile: true - attention_backend: aiter_flydsl - enable_tiling: true - enable_slicing: true - use_fp8_gemms: true diff --git a/benchmark_configs/xdit/wan2_2.yaml b/benchmark_configs/xdit/wan2_2.yaml index 8074a2b..c16565f 100644 --- a/benchmark_configs/xdit/wan2_2.yaml +++ b/benchmark_configs/xdit/wan2_2.yaml @@ -424,52 +424,74 @@ spargeattn_cdfthreshold: 0.9 use_spargeattn_head_balance: True -- name: wan2_2.2gpu.rdna4 +- name: wan2_2_ti2v_5b.i2v.1gpu.rdna4 tags: [rdna4, gfx1201] - model: Wan-AI/Wan2.2-I2V-A14B-Diffusers + model: Wan-AI/Wan2.2-TI2V-5B-Diffusers runner: xdit args: - prompt: "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside." - height: 720 + seed: 42 + prompt: Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. + height: 736 width: 1280 + num_frames: 121 + num_inference_steps: 50 + warmup_calls: 2 + num_iterations: 5 + guidance_scale: 5.0 + flow_shift: 5 + task: i2v input_images: /app/data/wan_input.jpg - num_frames: 81 - ulysses_degree: 2 - ring_degree: 1 - fully_shard_degree: 2 - memory_efficient_sharding: true - use_parallel_vae: true - seed: 42 - warmup_calls: 1 - num_iterations: 1 - num_inference_steps: 40 + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - attention_backend: aiter_flydsl - enable_tiling: true - enable_slicing: true use_fp8_gemms: true + cache_method: dbcache + enable_tiling: true -- name: wan2_2.4gpu.rdna4 +- name: wan2_2_ti2v_5b.i2v.2gpu.rdna4 tags: [rdna4, gfx1201] - model: Wan-AI/Wan2.2-I2V-A14B-Diffusers + model: Wan-AI/Wan2.2-TI2V-5B-Diffusers runner: xdit args: - prompt: "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside." - height: 720 + seed: 42 + prompt: Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. + height: 736 width: 1280 + num_frames: 121 + num_inference_steps: 50 + warmup_calls: 2 + num_iterations: 5 + guidance_scale: 5.0 + flow_shift: 5 + task: i2v input_images: /app/data/wan_input.jpg - num_frames: 81 - ulysses_degree: 4 - ring_degree: 1 - fully_shard_degree: 4 - memory_efficient_sharding: true + attention_backend: aiter_flydsl_fp8 + ulysses_degree: 2 + use_torch_compile: true + use_fp8_gemms: true use_parallel_vae: true + cache_method: dbcache + enable_tiling: true + +- name: wan2_2_ti2v_5b.i2v.4gpu.rdna4 + tags: [rdna4, gfx1201] + model: Wan-AI/Wan2.2-TI2V-5B-Diffusers + runner: xdit + args: seed: 42 - warmup_calls: 1 - num_iterations: 1 - num_inference_steps: 40 + prompt: Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. + height: 736 + width: 1280 + num_frames: 121 + num_inference_steps: 50 + warmup_calls: 2 + num_iterations: 5 + guidance_scale: 5.0 + flow_shift: 5 + task: i2v + input_images: /app/data/wan_input.jpg + attention_backend: aiter_flydsl_fp8 + ulysses_degree: 4 use_torch_compile: true - attention_backend: aiter_flydsl - enable_tiling: true - enable_slicing: true - use_fp8_gemms: true \ No newline at end of file + use_fp8_gemms: true + use_parallel_vae: true + cache_method: dbcache diff --git a/benchmark_configs/xdit/z_image.yaml b/benchmark_configs/xdit/z_image.yaml index 91cec90..315592a 100644 --- a/benchmark_configs/xdit/z_image.yaml +++ b/benchmark_configs/xdit/z_image.yaml @@ -56,15 +56,15 @@ height: 1088 width: 1920 num_inference_steps: 50 - warmup_calls: 5 + warmup_calls: 2 + num_iterations: 5 guidance_scale: 4.0 - num_iterations: 25 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 use_torch_compile: true - enable_tiling: true - enable_slicing: true use_fp8_gemms: true - + cache_method: dbcache + cache_config: + residual_diff_threshold: 0.1 - name: z_image.2gpu.rdna4 tags: [rdna4, gfx1201] model: Tongyi-MAI/Z-Image @@ -75,17 +75,16 @@ height: 1088 width: 1920 num_inference_steps: 50 - warmup_calls: 5 + warmup_calls: 2 + num_iterations: 5 guidance_scale: 4.0 - num_iterations: 25 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 ulysses_degree: 2 - ring_degree: 1 use_torch_compile: true - enable_tiling: true - enable_slicing: true use_fp8_gemms: true - + cache_method: dbcache + cache_config: + residual_diff_threshold: 0.1 - name: z_image.4gpu.rdna4 tags: [rdna4, gfx1201] model: Tongyi-MAI/Z-Image @@ -96,13 +95,16 @@ height: 1088 width: 1920 num_inference_steps: 50 - warmup_calls: 5 + warmup_calls: 2 + num_iterations: 5 guidance_scale: 4.0 - num_iterations: 25 - attention_backend: aiter_flydsl + attention_backend: aiter_flydsl_fp8 ulysses_degree: 4 - ring_degree: 1 use_torch_compile: true - enable_tiling: true - enable_slicing: true use_fp8_gemms: true + use_parallel_vae: true + cache_method: dbcache + cache_config: + residual_diff_threshold: 0.08 + scm_policy: ultra + Fn_compute_blocks: 6 diff --git a/benchmark_configs/xdit/z_image_turbo.yaml b/benchmark_configs/xdit/z_image_turbo.yaml index cd9f8a2..d225940 100644 --- a/benchmark_configs/xdit/z_image_turbo.yaml +++ b/benchmark_configs/xdit/z_image_turbo.yaml @@ -8,14 +8,11 @@ height: 512 width: 320 num_inference_steps: 4 - warmup_calls: 3 - ulysses_degree: 1 - ring_degree: 1 + warmup_calls: 2 + num_iterations: 5 guidance_scale: 0.0 - num_iterations: 10 attention_backend: aiter_flydsl use_torch_compile: true - - name: z_image_turbo.2gpu.rdna4 tags: [rdna4, gfx1201] model: Tongyi-MAI/Z-Image-Turbo @@ -26,16 +23,12 @@ height: 512 width: 320 num_inference_steps: 4 - warmup_calls: 3 - ulysses_degree: 2 - ring_degree: 1 + warmup_calls: 2 + num_iterations: 5 guidance_scale: 0.0 - num_iterations: 10 attention_backend: aiter_flydsl + ulysses_degree: 2 use_torch_compile: true - enable_tiling: true - enable_slicing: true - - name: z_image_turbo.4gpu.rdna4 tags: [rdna4, gfx1201] model: Tongyi-MAI/Z-Image-Turbo @@ -46,12 +39,9 @@ height: 512 width: 320 num_inference_steps: 4 - warmup_calls: 3 - ulysses_degree: 4 - ring_degree: 1 + warmup_calls: 2 + num_iterations: 5 guidance_scale: 0.0 - num_iterations: 10 attention_backend: aiter_flydsl + ulysses_degree: 4 use_torch_compile: true - enable_tiling: true - enable_slicing: true diff --git a/docker/Dockerfile.rdna4 b/docker/Dockerfile.rdna4 index cc6f0d1..5b59fd1 100644 --- a/docker/Dockerfile.rdna4 +++ b/docker/Dockerfile.rdna4 @@ -6,17 +6,18 @@ # when running from repository root. # -ARG AITER_COMMIT="7d604afe5fa7efba63c0dce323b95d9daf2db112" +ARG AITER_REPO="https://github.com/pds-amd/aiter.git" +ARG AITER_COMMIT="e76910980066cc696c12f10745cde1bae94c933e" ARG DIFFUSERS_COMMIT="3759fab56d3170a04d747e918a13e55fda6681e2" ARG ARBITER_COMMIT="656b6689bdaea4f17afe7387f2d9f36e10867e32" ARG DISTVAE_COMMIT="6d8025c96b9975d45badae12d8ed9b6422e34e2c" -ARG XDIT_REPO="https://github.com/xdit-project/xDiT.git" -ARG XDIT_COMMIT="4433b81061d027e568d49c4812b4ed795782868f" +ARG XDIT_REPO="https://github.com/pds-amd/xDiT.git" +ARG XDIT_COMMIT="e42e9d822244bc4ffe7efd5b814a94c7b6946c51" ARG LONGCONTEXTATTENTION_COMMIT="631bdfd" -ARG FLYDSL_VERSION="0.2.0" +ARG FLYDSL_VERSION="0.2.2" # transformers is the one transitive dep we pin by hand (its majors break # diffusers/xfuser); everything else resolves from each package's own metadata. -ARG TRANSFORMERS_PIN="5.10.2" +ARG TRANSFORMERS_PIN="5.13.0" # ---- ROCm + PyTorch wheels (TheRock gfx120X) ------------------------------- # torch/vision/audio are pinned to EXACT ROCm version strings that exist only on @@ -254,6 +255,7 @@ ARG DISTVAE_COMMIT ARG XDIT_COMMIT ARG XDIT_REPO ARG LONGCONTEXTATTENTION_COMMIT +ARG AITER_REPO ARG FLYDSL_VERSION ARG TRANSFORMERS_PIN @@ -305,7 +307,7 @@ RUN git clone https://github.com/feifeibear/long-context-attention.git && \ # aiter - built from source ON PURPOSE: published aiter wheels are CDNA-only # (gfx942+gfx950), and the part we use on RDNA4 is the JIT FlyDSL aiter_flydsl # attention backend, which compiles at runtime. -RUN git clone https://github.com/ROCm/aiter.git && \ +RUN git clone "${AITER_REPO}" && \ cd aiter && \ git reset --hard ${AITER_COMMIT} && \ echo "aiter at $(git rev-parse HEAD)" && \