2014 lines
60 KiB
JSON
2014 lines
60 KiB
JSON
{
|
|
"mode": "block-gate",
|
|
"environment": {
|
|
"platform": "Linux-6.17.0-1026-nvidia-aarch64-with-glibc2.39",
|
|
"python": "3.12.3 (main, Mar 23 2026, 19:04:32) [GCC 13.3.0]",
|
|
"torch": "2.9.1+cu130",
|
|
"torch_cuda": "13.0",
|
|
"device": "NVIDIA GB10",
|
|
"device_capability": [
|
|
12,
|
|
1
|
|
],
|
|
"driver": null,
|
|
"git_commit": null,
|
|
"checkpoint_path": "/models/minimax_h3_fl2va_pruned_nvfp4.safetensors",
|
|
"checkpoint_sha256": null,
|
|
"checkpoint_hash_note": "not calculated",
|
|
"environment_switches": {
|
|
"CUDA_DEVICE_MAX_CONNECTIONS": "1",
|
|
"CUDA_DEVICE_MAX_COPY_CONNECTIONS": "4",
|
|
"CUDA_HOME": "/usr/local/cuda",
|
|
"CUDA_INC_PATH": "/usr/local/cuda/include",
|
|
"CUDA_MANAGED_FORCE_DEVICE_ALLOC": "1",
|
|
"CUDA_MODULE_LOADING": "EAGER",
|
|
"CUDA_VERSION": "13.0.2",
|
|
"H3_FUSED_ELEMENTWISE": "1",
|
|
"H3_MODEL_PATH": "/models/minimax_h3_ref2va_pruned_nvfp4.safetensors",
|
|
"H3_NVFP4_MODULATE_FUSION": "1",
|
|
"H3_NVFP4_SCALE_BACKEND": "vortex",
|
|
"H3_NVFP4_SCALE_VERSION": "1",
|
|
"H3_NVFP4_SWIGLU_FUSION": "1",
|
|
"H3_SAGE_QKV_LAYOUT": "strided_nhd",
|
|
"TORCH_COMPILE_DISABLE": "0",
|
|
"TORCH_CUDA_ARCH_LIST": "12.1a",
|
|
"TORCH_EXTENSIONS_DIR": "/opt/h3-blackwell-runtime/.torch_extensions"
|
|
},
|
|
"extension": {
|
|
"cuda_version": 13000,
|
|
"cublas_version": 130100,
|
|
"stream_k_public_control": false,
|
|
"stream_k_note": "CUDA 13 cuBLASLt exposes no documented MatmulAlgoConfig attribute that directly selects Stream-K. Negative SPLITK_NUM values returned by heuristics are preserved as undocumented library sentinels, not claimed as public Stream-K control."
|
|
},
|
|
"comfy_kitchen": "0.2.31 package without __version__"
|
|
},
|
|
"workload": {
|
|
"resolution": [
|
|
1344,
|
|
768
|
|
],
|
|
"frames": 124,
|
|
"steps": 12,
|
|
"sampler_step": 1,
|
|
"seed": 440420,
|
|
"text_tokens": 100,
|
|
"tokens": 37810,
|
|
"hidden_shape": [
|
|
37810,
|
|
5376
|
|
],
|
|
"segments": [
|
|
[
|
|
0,
|
|
100,
|
|
1
|
|
],
|
|
[
|
|
100,
|
|
514,
|
|
2
|
|
],
|
|
[
|
|
514,
|
|
37810,
|
|
0
|
|
]
|
|
]
|
|
},
|
|
"retained_blocks": [
|
|
0,
|
|
24,
|
|
49
|
|
],
|
|
"immutable_cloned_block_inputs": {
|
|
"0": [
|
|
37810,
|
|
5376
|
|
],
|
|
"24": [
|
|
37810,
|
|
5376
|
|
],
|
|
"49": [
|
|
37810,
|
|
5376
|
|
]
|
|
},
|
|
"fc2_boundary": {
|
|
"block": 24,
|
|
"gate_up_shape": [
|
|
37810,
|
|
28672
|
|
],
|
|
"activation_qdata_shape": [
|
|
37824,
|
|
7168
|
|
],
|
|
"weight_qdata_shape": [
|
|
5376,
|
|
7168
|
|
],
|
|
"logical_mnk": [
|
|
37810,
|
|
5376,
|
|
14336
|
|
],
|
|
"descriptor_mnk_after_padding": [
|
|
37824,
|
|
5376,
|
|
14336
|
|
],
|
|
"producer": "vortex_native_quantize_swiglu_nvfp4",
|
|
"no_bias": true
|
|
},
|
|
"baseline_kernel_metadata": {
|
|
"path": "fc2.forward_swiglu -> accepted producer -> Comfy Kitchen 0.2.31 scaled_mm_nvfp4",
|
|
"descriptors": {
|
|
"packed_input_output": "row-major [M,K] @ [N,K].T -> BF16 [M,N]",
|
|
"block_scale_mode": "VEC16_UE4M3",
|
|
"compute_and_scale": "FP32",
|
|
"scalar_pointer_mode": "device",
|
|
"bias": null,
|
|
"beta": 0.0,
|
|
"comfy_kitchen_version": "0.2.31"
|
|
},
|
|
"profiler_cuda_events_available": false,
|
|
"profiler_note": "Torch profiler returned no CUDA kernel events on this build; use --mode profile with NCU for kernel metadata.",
|
|
"top_cuda_events": [],
|
|
"output_sha256": "ad26c744afbb0249bff9afea503984e51217f85a59434e44fde347ae7d9b852f"
|
|
},
|
|
"heuristics": [
|
|
{
|
|
"max_workspace_bytes": 0,
|
|
"requested_count": 32,
|
|
"returned_count": 5,
|
|
"algorithms": [
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": -2,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"max_workspace_bytes": 4194304,
|
|
"requested_count": 32,
|
|
"returned_count": 7,
|
|
"algorithms": [
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": -2,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"max_workspace_bytes": 8388608,
|
|
"requested_count": 32,
|
|
"returned_count": 7,
|
|
"algorithms": [
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": -2,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"max_workspace_bytes": 16777216,
|
|
"requested_count": 32,
|
|
"returned_count": 6,
|
|
"algorithms": [
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": -2,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": -2,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"max_workspace_bytes": 33554432,
|
|
"requested_count": 32,
|
|
"returned_count": 6,
|
|
"algorithms": [
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": -2,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": -2,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"max_workspace_bytes": 67108864,
|
|
"requested_count": 32,
|
|
"returned_count": 6,
|
|
"algorithms": [
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": -2,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": -2,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
}
|
|
}
|
|
]
|
|
}
|
|
],
|
|
"explicit_split_k_checks": [
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 1.0,
|
|
"state": 0,
|
|
"api_status": 0,
|
|
"valid": true,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
},
|
|
"requested_config": {
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 2,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 0.0,
|
|
"state": 15,
|
|
"api_status": 15,
|
|
"valid": false,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
},
|
|
"requested_config": {
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"split_k": 2,
|
|
"reduction_scheme": 2
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 2,
|
|
"reduction_scheme": 4,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 0.0,
|
|
"state": 15,
|
|
"api_status": 15,
|
|
"valid": false,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
},
|
|
"requested_config": {
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"split_k": 2,
|
|
"reduction_scheme": 4
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 4,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 0.0,
|
|
"state": 15,
|
|
"api_status": 15,
|
|
"valid": false,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
},
|
|
"requested_config": {
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"split_k": 4,
|
|
"reduction_scheme": 2
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 4,
|
|
"reduction_scheme": 4,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 0.0,
|
|
"state": 15,
|
|
"api_status": 15,
|
|
"valid": false,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
},
|
|
"requested_config": {
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"split_k": 4,
|
|
"reduction_scheme": 4
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 8,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 0.0,
|
|
"state": 15,
|
|
"api_status": 15,
|
|
"valid": false,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
},
|
|
"requested_config": {
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"split_k": 8,
|
|
"reduction_scheme": 2
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 8,
|
|
"reduction_scheme": 4,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 0.0,
|
|
"state": 15,
|
|
"api_status": 15,
|
|
"valid": false,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
},
|
|
"requested_config": {
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"split_k": 8,
|
|
"reduction_scheme": 4
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 16,
|
|
"reduction_scheme": 2,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 0.0,
|
|
"state": 15,
|
|
"api_status": 15,
|
|
"valid": false,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
},
|
|
"requested_config": {
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"split_k": 16,
|
|
"reduction_scheme": 2
|
|
}
|
|
},
|
|
{
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 16,
|
|
"reduction_scheme": 4,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"required_workspace_bytes": 0,
|
|
"waves": 0.0,
|
|
"state": 15,
|
|
"api_status": 15,
|
|
"valid": false,
|
|
"capabilities": {
|
|
"split_k_support": 1,
|
|
"reduction_scheme_mask": 6,
|
|
"cta_swizzle_support": 0,
|
|
"custom_option_max": 0,
|
|
"strided_batch_support": 1,
|
|
"out_of_place_result_support": 1,
|
|
"tile_ids": [
|
|
20
|
|
],
|
|
"stages_ids": [
|
|
37
|
|
],
|
|
"inner_cluster_shape_capability_note": "This CUDA 13 cublasLt.h exposes config IDs but no public capability attributes that enumerate inner/cluster shape IDs."
|
|
},
|
|
"requested_config": {
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0,
|
|
"inner_shape": null,
|
|
"cluster_shape": null,
|
|
"split_k": 16,
|
|
"reduction_scheme": 4
|
|
}
|
|
}
|
|
],
|
|
"selected": {
|
|
"algorithm_id": 70,
|
|
"tile_id": 20,
|
|
"stages_id": 37,
|
|
"split_k": 1,
|
|
"reduction_scheme": 0,
|
|
"custom_option": 0,
|
|
"cta_swizzle": 0
|
|
},
|
|
"block_gate": [
|
|
{
|
|
"block": 0,
|
|
"only_monkeypatched_method": "block.mlp.fc2.forward_swiglu",
|
|
"accepted_gate_and_residual_path_preserved": true,
|
|
"candidate_vs_baseline": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "6bdb03e2bdf7733fde92b2cf5c78a373e9818cb5ec9d213d873a632fad0af147",
|
|
"expected_sha256": "6bdb03e2bdf7733fde92b2cf5c78a373e9818cb5ec9d213d873a632fad0af147"
|
|
},
|
|
"baseline_vs_traversal": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "6bdb03e2bdf7733fde92b2cf5c78a373e9818cb5ec9d213d873a632fad0af147",
|
|
"expected_sha256": "6bdb03e2bdf7733fde92b2cf5c78a373e9818cb5ec9d213d873a632fad0af147"
|
|
},
|
|
"candidate_vs_traversal": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "6bdb03e2bdf7733fde92b2cf5c78a373e9818cb5ec9d213d873a632fad0af147",
|
|
"expected_sha256": "6bdb03e2bdf7733fde92b2cf5c78a373e9818cb5ec9d213d873a632fad0af147"
|
|
},
|
|
"timing": {
|
|
"order": "AB/BA alternates by round",
|
|
"baseline": {
|
|
"samples_ms": [
|
|
457.58404541015625,
|
|
455.8675537109375,
|
|
455.042724609375,
|
|
456.5163879394531,
|
|
456.7074279785156,
|
|
459.7113037109375,
|
|
456.3467102050781,
|
|
459.0717468261719,
|
|
457.6736755371094,
|
|
458.6617736816406,
|
|
458.61083984375,
|
|
456.4378662109375,
|
|
457.4807434082031,
|
|
458.178466796875,
|
|
457.4196472167969,
|
|
458.90869140625,
|
|
458.4776306152344,
|
|
458.2261962890625,
|
|
458.10919189453125,
|
|
460.418701171875
|
|
],
|
|
"p50_ms": 457.8914337158203,
|
|
"p95_ms": 459.7466735839844,
|
|
"mean_ms": 457.7725662231445
|
|
},
|
|
"candidate": {
|
|
"samples_ms": [
|
|
428.5652160644531,
|
|
425.1132507324219,
|
|
420.88983154296875,
|
|
420.71435546875,
|
|
419.1617736816406,
|
|
421.9552307128906,
|
|
419.37896728515625,
|
|
419.3441467285156,
|
|
419.79351806640625,
|
|
419.0124206542969,
|
|
420.5487976074219,
|
|
420.9779052734375,
|
|
420.5229187011719,
|
|
420.88104248046875,
|
|
420.3728942871094,
|
|
420.11834716796875,
|
|
420.3209533691406,
|
|
423.73919677734375,
|
|
419.7261657714844,
|
|
423.4720458984375
|
|
],
|
|
"p50_ms": 420.5358581542969,
|
|
"p95_ms": 425.2858489990234,
|
|
"mean_ms": 421.2304489135742
|
|
},
|
|
"parity": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "6bdb03e2bdf7733fde92b2cf5c78a373e9818cb5ec9d213d873a632fad0af147",
|
|
"expected_sha256": "6bdb03e2bdf7733fde92b2cf5c78a373e9818cb5ec9d213d873a632fad0af147"
|
|
}
|
|
},
|
|
"supplied_workspace_bytes": 0
|
|
},
|
|
{
|
|
"block": 24,
|
|
"only_monkeypatched_method": "block.mlp.fc2.forward_swiglu",
|
|
"accepted_gate_and_residual_path_preserved": true,
|
|
"candidate_vs_baseline": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "d8f5c6335e823dc4e3ed8cdd34d90dda0d69297ea47de9b328962dfae00c9f75",
|
|
"expected_sha256": "d8f5c6335e823dc4e3ed8cdd34d90dda0d69297ea47de9b328962dfae00c9f75"
|
|
},
|
|
"baseline_vs_traversal": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "d8f5c6335e823dc4e3ed8cdd34d90dda0d69297ea47de9b328962dfae00c9f75",
|
|
"expected_sha256": "d8f5c6335e823dc4e3ed8cdd34d90dda0d69297ea47de9b328962dfae00c9f75"
|
|
},
|
|
"candidate_vs_traversal": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "d8f5c6335e823dc4e3ed8cdd34d90dda0d69297ea47de9b328962dfae00c9f75",
|
|
"expected_sha256": "d8f5c6335e823dc4e3ed8cdd34d90dda0d69297ea47de9b328962dfae00c9f75"
|
|
},
|
|
"timing": {
|
|
"order": "AB/BA alternates by round",
|
|
"baseline": {
|
|
"samples_ms": [
|
|
458.4346008300781,
|
|
459.98040771484375,
|
|
457.2791748046875,
|
|
456.69989013671875,
|
|
458.1690368652344,
|
|
459.9689025878906,
|
|
459.0711975097656,
|
|
461.275390625,
|
|
458.7782897949219,
|
|
460.40740966796875,
|
|
459.8872985839844,
|
|
461.1615905761719,
|
|
459.39276123046875,
|
|
459.6435546875,
|
|
460.915283203125,
|
|
461.53076171875,
|
|
460.4534912109375,
|
|
460.2518615722656,
|
|
458.4383544921875,
|
|
461.34747314453125
|
|
],
|
|
"p50_ms": 459.9281005859375,
|
|
"p95_ms": 461.3566375732422,
|
|
"mean_ms": 459.6543365478516
|
|
},
|
|
"candidate": {
|
|
"samples_ms": [
|
|
418.2376708984375,
|
|
421.28668212890625,
|
|
417.4606018066406,
|
|
420.7315368652344,
|
|
421.40478515625,
|
|
419.4402770996094,
|
|
417.5429382324219,
|
|
419.78125,
|
|
419.68603515625,
|
|
421.91644287109375,
|
|
422.23394775390625,
|
|
420.081298828125,
|
|
419.1646728515625,
|
|
421.63360595703125,
|
|
422.85125732421875,
|
|
418.6084899902344,
|
|
420.3404235839844,
|
|
421.3265075683594,
|
|
422.9977111816406,
|
|
420.1449279785156
|
|
],
|
|
"p50_ms": 420.24267578125,
|
|
"p95_ms": 422.8585800170899,
|
|
"mean_ms": 420.3435531616211
|
|
},
|
|
"parity": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "d8f5c6335e823dc4e3ed8cdd34d90dda0d69297ea47de9b328962dfae00c9f75",
|
|
"expected_sha256": "d8f5c6335e823dc4e3ed8cdd34d90dda0d69297ea47de9b328962dfae00c9f75"
|
|
}
|
|
},
|
|
"supplied_workspace_bytes": 0
|
|
},
|
|
{
|
|
"block": 49,
|
|
"only_monkeypatched_method": "block.mlp.fc2.forward_swiglu",
|
|
"accepted_gate_and_residual_path_preserved": true,
|
|
"candidate_vs_baseline": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "3723ebf82eb1e279cc004832b68d071422a45dcc74adac93010e5fde33b715d6",
|
|
"expected_sha256": "3723ebf82eb1e279cc004832b68d071422a45dcc74adac93010e5fde33b715d6"
|
|
},
|
|
"baseline_vs_traversal": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "3723ebf82eb1e279cc004832b68d071422a45dcc74adac93010e5fde33b715d6",
|
|
"expected_sha256": "3723ebf82eb1e279cc004832b68d071422a45dcc74adac93010e5fde33b715d6"
|
|
},
|
|
"candidate_vs_traversal": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "3723ebf82eb1e279cc004832b68d071422a45dcc74adac93010e5fde33b715d6",
|
|
"expected_sha256": "3723ebf82eb1e279cc004832b68d071422a45dcc74adac93010e5fde33b715d6"
|
|
},
|
|
"timing": {
|
|
"order": "AB/BA alternates by round",
|
|
"baseline": {
|
|
"samples_ms": [
|
|
455.2886962890625,
|
|
456.0704650878906,
|
|
455.62127685546875,
|
|
459.4961242675781,
|
|
456.5694580078125,
|
|
457.70477294921875,
|
|
458.9316101074219,
|
|
458.7506103515625,
|
|
457.3757629394531,
|
|
456.0811462402344,
|
|
458.1807861328125,
|
|
457.0354919433594,
|
|
459.8126220703125,
|
|
457.84747314453125,
|
|
458.5899963378906,
|
|
459.2362060546875,
|
|
460.5073547363281,
|
|
459.9343566894531,
|
|
457.9271240234375,
|
|
458.1864318847656
|
|
],
|
|
"p50_ms": 458.053955078125,
|
|
"p95_ms": 459.9630065917969,
|
|
"mean_ms": 457.95738830566404
|
|
},
|
|
"candidate": {
|
|
"samples_ms": [
|
|
414.79840087890625,
|
|
421.3484191894531,
|
|
416.3827209472656,
|
|
416.5849914550781,
|
|
419.1302490234375,
|
|
414.7456970214844,
|
|
419.312255859375,
|
|
415.2906494140625,
|
|
416.9346008300781,
|
|
416.1675720214844,
|
|
418.51458740234375,
|
|
417.76995849609375,
|
|
420.13006591796875,
|
|
418.7623596191406,
|
|
416.8761901855469,
|
|
418.8280029296875,
|
|
417.02777099609375,
|
|
417.6993408203125,
|
|
416.9612121582031,
|
|
419.1832580566406
|
|
],
|
|
"p50_ms": 417.3635559082031,
|
|
"p95_ms": 420.190983581543,
|
|
"mean_ms": 417.62241516113284
|
|
},
|
|
"parity": {
|
|
"bf16_exact": true,
|
|
"different_elements": 0,
|
|
"max_abs": 0.0,
|
|
"mean_abs": 0.0,
|
|
"actual_sha256": "3723ebf82eb1e279cc004832b68d071422a45dcc74adac93010e5fde33b715d6",
|
|
"expected_sha256": "3723ebf82eb1e279cc004832b68d071422a45dcc74adac93010e5fde33b715d6"
|
|
}
|
|
},
|
|
"supplied_workspace_bytes": 0
|
|
}
|
|
],
|
|
"errors_and_unsupported": [
|
|
{
|
|
"feature": "Stream-K",
|
|
"supported_public_control": false,
|
|
"reason": "CUDA 13 cuBLASLt exposes no documented MatmulAlgoConfig attribute that directly selects Stream-K. Negative SPLITK_NUM values returned by heuristics are preserved as undocumented library sentinels, not claimed as public Stream-K control."
|
|
}
|
|
]
|
|
}
|