76 lines
2.7 KiB
JSON
76 lines
2.7 KiB
JSON
|
{
|
||
|
|
"device": "NVIDIA GB10",
|
||
|
|
"compute_capability": "SM121",
|
||
|
|
"torch": "2.9.1+cu130",
|
||
|
|
"resolution": [
|
||
|
|
1344,
|
||
|
|
768
|
||
|
|
],
|
||
|
|
"frames": 124,
|
||
|
|
"packed_tokens": 37810,
|
||
|
|
"seed": 440420,
|
||
|
|
"attention": "sage2",
|
||
|
|
"representative_block_results": [
|
||
|
|
{
|
||
|
|
"block": 0,
|
||
|
|
"comfy_scale_p50_seconds": 0.5251280819647945,
|
||
|
|
"vortex_scale_p50_seconds": 0.5088961640140042,
|
||
|
|
"speedup": 1.031896326006387,
|
||
|
|
"latency_reduction": 0.0309103978786619,
|
||
|
|
"checksum": -8279139.0
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"block": 24,
|
||
|
|
"comfy_scale_p50_seconds": 0.5646577654988505,
|
||
|
|
"vortex_scale_p50_seconds": 0.5149167295021471,
|
||
|
|
"speedup": 1.0966001552227602,
|
||
|
|
"latency_reduction": 0.08809059050619694,
|
||
|
|
"checksum": 89071408.0
|
||
|
|
},
|
||
|
|
{
|
||
|
|
"block": 49,
|
||
|
|
"comfy_scale_p50_seconds": 0.524194426019676,
|
||
|
|
"vortex_scale_p50_seconds": 0.5067900969879702,
|
||
|
|
"speedup": 1.034342283196033,
|
||
|
|
"latency_reduction": 0.03320204902570367,
|
||
|
|
"checksum": -40617144320.0
|
||
|
|
}
|
||
|
|
],
|
||
|
|
"block_24_quantization_p50_milliseconds": {
|
||
|
|
"attn_qkv_proj": {
|
||
|
|
"comfy": 9.078,
|
||
|
|
"vortex": 3.908
|
||
|
|
},
|
||
|
|
"attn_out_proj": {
|
||
|
|
"comfy": 10.035,
|
||
|
|
"vortex": 5.151
|
||
|
|
},
|
||
|
|
"mlp_fc1": {
|
||
|
|
"comfy": 7.27,
|
||
|
|
"vortex": 5.182
|
||
|
|
},
|
||
|
|
"mlp_fc2": {
|
||
|
|
"comfy": 19.187,
|
||
|
|
"vortex": 10.163
|
||
|
|
}
|
||
|
|
},
|
||
|
|
"attention_findings": [
|
||
|
|
"Sage2 was substantially faster than PyTorch SDPA, forced cuDNN SDPA, and FlashAttention-4 at 37,810 tokens.",
|
||
|
|
"Sage2 NHD produced an exact output but did not materially improve median total path time and had worse tail latency.",
|
||
|
|
"Strict Sol-Attn reached about 157 ms kernel time versus about 260 ms for Sage2, but its output differed materially from Sage2 and is not a correctness-preserving replacement.",
|
||
|
|
"Explicit Sage2 kernel variants did not provide an exact speed improvement over the Sage2 dispatcher."
|
||
|
|
],
|
||
|
|
"validation": {
|
||
|
|
"linear_reference_diff": "zero for all four Vortex-scale plus Comfy-pack H3 linears",
|
||
|
|
"block_checksums": "identical for Comfy and Vortex paths at blocks 0, 24, and 49",
|
||
|
|
"remote_contracts": "28 passed",
|
||
|
|
"full_model_smoke": "not run: the deployed image predates the distributed runner; direct preview did not produce an artifact while resident services remained active"
|
||
|
|
},
|
||
|
|
"notes": [
|
||
|
|
"The optimization changes activation-scale discovery only; Comfy Kitchen still packs activations and executes scaled_mm_nvfp4.",
|
||
|
|
"The default H3_NVFP4_SCALE_BACKEND=torch path is unchanged.",
|
||
|
|
"The native Vortex packer is not enabled because the attention output projection failed parity in this profile.",
|
||
|
|
"Resident H3 and Qwen services remained running during profiling, so medians are preferred over means and tail latency requires an isolated rerun."
|
||
|
|
]
|
||
|
|
}
|