85 lines
3 KiB
JSON
85 lines
3 KiB
JSON
{
|
|
"device": "NVIDIA GB10",
|
|
"compute_capability": "SM121",
|
|
"torch": "2.9.1+cu130",
|
|
"attention": "sage2",
|
|
"resolution": [1344, 768],
|
|
"frames": 124,
|
|
"packed_tokens": 37810,
|
|
"seed": 440420,
|
|
"feature_flag": "H3_SAGE_QKV_LAYOUT=strided_nhd",
|
|
"dataflow": {
|
|
"qkv_input": "projection-strided NHD views with no Q/K/V materialization",
|
|
"qk_transform": "in-place Comfy Kitchen RMSNorm and RoPE on strided views",
|
|
"attention_output": "contiguous Sage2 NHD reshaped as a token-major view for out_proj",
|
|
"removed_copies_per_block": 4
|
|
},
|
|
"isolated_attention_layout": {
|
|
"hnd_mean_seconds": 0.3037509555928409,
|
|
"strided_nhd_mean_seconds": 0.2602574324118905,
|
|
"speedup": 1.167117314490817,
|
|
"max_abs_difference": 0.0,
|
|
"mean_abs_difference": 0.0
|
|
},
|
|
"real_block_results": [
|
|
{
|
|
"block": 0,
|
|
"hnd_p50_seconds": 0.4870640740264207,
|
|
"strided_nhd_p50_seconds": 0.4714779225178063,
|
|
"speedup": 1.0330580728475696,
|
|
"max_abs_difference": 0.0
|
|
},
|
|
{
|
|
"block": 24,
|
|
"hnd_p50_seconds": 0.4879330809926614,
|
|
"strided_nhd_p50_seconds": 0.4716692080255598,
|
|
"speedup": 1.034481523683056,
|
|
"max_abs_difference": 0.0
|
|
},
|
|
{
|
|
"block": 49,
|
|
"hnd_p50_seconds": 0.48766444699140266,
|
|
"strided_nhd_p50_seconds": 0.47158038747147657,
|
|
"speedup": 1.0341067184879458,
|
|
"max_abs_difference": 0.0
|
|
}
|
|
],
|
|
"sampling_2_step": {
|
|
"hnd_seconds": 52.49690314201871,
|
|
"strided_nhd_seconds": 50.95522483601235,
|
|
"speedup": 1.0302555490819223,
|
|
"checksums": [-276716.375, 727.1075439453125],
|
|
"peak_allocated_bytes_saved": 1185838592,
|
|
"peak_reserved_bytes_saved": 2036334592
|
|
},
|
|
"sampling_12_step": {
|
|
"component1_hnd_seconds": 301.04568115097936,
|
|
"strided_nhd_seconds": 290.2269145210157,
|
|
"speedup": 1.0372769239814257,
|
|
"latency_reduction": 0.03593729226940101,
|
|
"checksums": [-91481.7890625, 678.401611328125],
|
|
"cumulative_reduction_vs_unfused_310_11_second_baseline": 0.06411616431876277
|
|
},
|
|
"nsight_block_recapture": {
|
|
"tool": "Nsight Systems 2025.3.2",
|
|
"hnd_kernel_launches": 53,
|
|
"strided_nhd_kernel_launches": 49,
|
|
"removed_kernel_launches": 4,
|
|
"removed_kernels": "three Q/K/V transpose-contiguous copies and one attention-output transpose-contiguous copy",
|
|
"large_bf16_layout_copy_instances_after": 0
|
|
},
|
|
"validation": {
|
|
"isolated_attention": "bit-exact",
|
|
"blocks_0_24_49": "bit-exact",
|
|
"full_50_block_two_step_trajectory": "identical video and audio checksums",
|
|
"canonical_12_step_trajectory": "identical video and audio checksums",
|
|
"deployed_tests": "34 passed"
|
|
},
|
|
"source_artifacts": [
|
|
"benchmarks/gb10-sage-strided-nhd-contract-1344x768-124f-seed440420.json",
|
|
"benchmarks/gb10-sage-strided-nhd-blocks-1344x768-124f-seed440420.json",
|
|
"benchmarks/gb10-component2-layout-hnd-2step.json",
|
|
"benchmarks/gb10-component2-layout-strided-nhd-2step.json",
|
|
"benchmarks/gb10-component2-layout-strided-nhd-12step.json"
|
|
]
|
|
}
|