zefancainlp's picture
download
raw
16.2 kB
{
"scope": "Saved-source graph audit and CPU failure-mechanism reproduction",
"api_calls": 0,
"gpu_evaluations": 0,
"training_updates": 0,
"source_graphs": {
"historical_step3.py": {
"source_sha256": "21a0f0f854f124c1af737fc620a8ab87fcc5843eaef0a4680c7d07a6c2bf5d9c",
"defined_functions": [
"batched_matmul_kernel",
"batched_matmul_triton",
"fused_proj_gate_out_kernel",
"fused_proj_gate_kernel",
"custom_kernel"
],
"reachable_functions": [
"batched_matmul_triton",
"custom_kernel",
"fused_proj_gate_kernel"
],
"reachable_jit_functions": [
"fused_proj_gate_kernel"
],
"dormant_jit_functions": [
"batched_matmul_kernel",
"fused_proj_gate_out_kernel"
],
"edges": {
"batched_matmul_triton": [],
"custom_kernel": [
"batched_matmul_triton",
"fused_proj_gate_kernel"
],
"fused_proj_gate_kernel": []
},
"live_torch_and_launch_call_sites": [
{
"owner": "batched_matmul_triton",
"line": 45,
"call": "torch.bmm"
},
{
"owner": "custom_kernel",
"line": 168,
"call": "F.layer_norm"
},
{
"owner": "custom_kernel",
"line": 178,
"call": "torch.empty"
},
{
"owner": "custom_kernel",
"line": 179,
"call": "torch.empty"
},
{
"owner": "custom_kernel",
"line": 180,
"call": "torch.empty"
},
{
"owner": "custom_kernel",
"line": 196,
"call": "fused_proj_gate_kernel[grid_m, grid_n]"
},
{
"owner": "custom_kernel",
"line": 213,
"call": "F.layer_norm"
},
{
"owner": "custom_kernel",
"line": 215,
"call": "F.linear"
}
],
"precision_prefix_lines": [
"x = F.layer_norm(input_tensor, normalized_shape=(dim,), weight=norm_weight, bias=norm_bias, eps=eps)",
"out = F.layer_norm(out, normalized_shape=(hidden_dim,), weight=to_out_norm_w, bias=to_out_norm_b, eps=eps)"
]
},
"linked_step3.py": {
"source_sha256": "40f295d342f0244445e620b1eb72c039965d2da2a0c5ad51de36a7bb89d4ab0c",
"defined_functions": [
"batched_matmul_kernel",
"batched_matmul_triton",
"custom_kernel"
],
"reachable_functions": [
"batched_matmul_triton",
"custom_kernel"
],
"reachable_jit_functions": [],
"dormant_jit_functions": [
"batched_matmul_kernel"
],
"edges": {
"batched_matmul_triton": [],
"custom_kernel": [
"batched_matmul_triton"
]
},
"live_torch_and_launch_call_sites": [
{
"owner": "batched_matmul_triton",
"line": 58,
"call": "torch.bmm"
},
{
"owner": "custom_kernel",
"line": 83,
"call": "F.layer_norm"
},
{
"owner": "custom_kernel",
"line": 87,
"call": "torch.cat([left_proj_w, right_proj_w], dim=0).to"
},
{
"owner": "custom_kernel",
"line": 90,
"call": "torch.cat([left_gate_w, right_gate_w], dim=0).to"
},
{
"owner": "custom_kernel",
"line": 92,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 94,
"call": "F.linear(x, gate_weight).sigmoid"
},
{
"owner": "custom_kernel",
"line": 96,
"call": "F.linear(x, out_gate_w).sigmoid"
},
{
"owner": "custom_kernel",
"line": 112,
"call": "F.layer_norm"
},
{
"owner": "custom_kernel",
"line": 115,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 87,
"call": "torch.cat"
},
{
"owner": "custom_kernel",
"line": 90,
"call": "torch.cat"
},
{
"owner": "custom_kernel",
"line": 94,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 96,
"call": "F.linear"
}
],
"precision_prefix_lines": [
"x_norm = F.layer_norm(input_tensor, normalized_shape=(dim,), weight=norm_weight, bias=norm_bias, eps=eps)",
"out_norm = F.layer_norm(out, normalized_shape=(hidden_dim,), weight=to_out_norm_w, bias=to_out_norm_b, eps=eps)"
]
},
"historical_later_best.py": {
"source_sha256": "3e8c298ab404ea167652b4bff3903c7fa921dd83aa4ffe2ac44908e5e10a4cc3",
"defined_functions": [
"batched_matmul_kernel",
"batched_matmul_triton",
"fused_proj_gate_out_kernel",
"fused_proj_gate_out_opt_kernel",
"postprocess_kernel",
"fused_proj_gate_out_direct_kernel",
"fused_norm_gate_kernel",
"custom_kernel"
],
"reachable_functions": [
"custom_kernel",
"fused_norm_gate_kernel",
"fused_proj_gate_out_direct_kernel"
],
"reachable_jit_functions": [
"fused_norm_gate_kernel",
"fused_proj_gate_out_direct_kernel"
],
"dormant_jit_functions": [
"batched_matmul_kernel",
"fused_proj_gate_out_kernel",
"fused_proj_gate_out_opt_kernel",
"postprocess_kernel"
],
"edges": {
"custom_kernel": [
"fused_norm_gate_kernel",
"fused_proj_gate_out_direct_kernel"
],
"fused_norm_gate_kernel": [],
"fused_proj_gate_out_direct_kernel": []
},
"live_torch_and_launch_call_sites": [
{
"owner": "custom_kernel",
"line": 349,
"call": "F.layer_norm"
},
{
"owner": "custom_kernel",
"line": 364,
"call": "torch.empty"
},
{
"owner": "custom_kernel",
"line": 365,
"call": "torch.empty_like"
},
{
"owner": "custom_kernel",
"line": 366,
"call": "torch.empty_like"
},
{
"owner": "custom_kernel",
"line": 372,
"call": "fused_proj_gate_out_direct_kernel[grid_m, grid_n]"
},
{
"owner": "custom_kernel",
"line": 373,
"call": "torch.bmm"
},
{
"owner": "custom_kernel",
"line": 375,
"call": "torch.empty"
},
{
"owner": "custom_kernel",
"line": 379,
"call": "fused_norm_gate_kernel[grid]"
},
{
"owner": "custom_kernel",
"line": 381,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 357,
"call": "torch.ones"
}
],
"precision_prefix_lines": [
"x = F.layer_norm(input_tensor, (dim,), weight=norm_w, bias=norm_b, eps=eps)"
]
},
"step0.py": {
"source_sha256": "d4fe30c2270c484dace78e1462a1d1d3a2e172fe8aefd0212f682e33bdc18174",
"defined_functions": [
"batched_matmul_kernel",
"batched_matmul_triton",
"custom_kernel"
],
"reachable_functions": [
"batched_matmul_kernel",
"batched_matmul_triton",
"custom_kernel"
],
"reachable_jit_functions": [
"batched_matmul_kernel"
],
"dormant_jit_functions": [],
"edges": {
"batched_matmul_kernel": [],
"batched_matmul_triton": [
"batched_matmul_kernel"
],
"custom_kernel": [
"batched_matmul_triton"
]
},
"live_torch_and_launch_call_sites": [
{
"owner": "batched_matmul_triton",
"line": 96,
"call": "torch.empty"
},
{
"owner": "batched_matmul_triton",
"line": 113,
"call": "batched_matmul_kernel[grid]"
},
{
"owner": "custom_kernel",
"line": 174,
"call": "F.layer_norm"
},
{
"owner": "custom_kernel",
"line": 187,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 188,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 202,
"call": "F.linear(x_norm_h, left_gate_w).sigmoid"
},
{
"owner": "custom_kernel",
"line": 203,
"call": "F.linear(x_norm_h, right_gate_w).sigmoid"
},
{
"owner": "custom_kernel",
"line": 204,
"call": "F.linear(x_norm_h, out_gate_w).sigmoid"
},
{
"owner": "custom_kernel",
"line": 231,
"call": "F.layer_norm"
},
{
"owner": "custom_kernel",
"line": 240,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 202,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 203,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 204,
"call": "F.linear"
}
],
"precision_prefix_lines": [
"x_norm = F.layer_norm(",
"out_norm = F.layer_norm("
]
},
"shape_failure.py": {
"source_sha256": "a3ace375a6a1ae8a5d72a0366721670956328ef1f0591fe24952b822544e7c69",
"defined_functions": [
"batched_matmul_kernel",
"batched_matmul_triton",
"_apply_mask_kernel",
"_apply_mask_triton",
"pfgv_k_8c88bc056b6c_batched_matmul_kernel",
"pfgv_k_8c88bc056b6c_batched_matmul_triton",
"fused_proj_gate_outgate_kernel",
"custom_kernel"
],
"reachable_functions": [
"custom_kernel",
"fused_proj_gate_outgate_kernel",
"pfgv_k_8c88bc056b6c_batched_matmul_kernel",
"pfgv_k_8c88bc056b6c_batched_matmul_triton"
],
"reachable_jit_functions": [
"fused_proj_gate_outgate_kernel",
"pfgv_k_8c88bc056b6c_batched_matmul_kernel"
],
"dormant_jit_functions": [
"_apply_mask_kernel",
"batched_matmul_kernel"
],
"edges": {
"custom_kernel": [
"fused_proj_gate_outgate_kernel",
"pfgv_k_8c88bc056b6c_batched_matmul_triton"
],
"fused_proj_gate_outgate_kernel": [],
"pfgv_k_8c88bc056b6c_batched_matmul_kernel": [],
"pfgv_k_8c88bc056b6c_batched_matmul_triton": [
"pfgv_k_8c88bc056b6c_batched_matmul_kernel"
]
},
"live_torch_and_launch_call_sites": [
{
"owner": "custom_kernel",
"line": 174,
"call": "F.layer_norm"
},
{
"owner": "custom_kernel",
"line": 191,
"call": "torch.empty"
},
{
"owner": "custom_kernel",
"line": 192,
"call": "torch.empty"
},
{
"owner": "custom_kernel",
"line": 193,
"call": "torch.empty"
},
{
"owner": "custom_kernel",
"line": 208,
"call": "fused_proj_gate_outgate_kernel[grid]"
},
{
"owner": "custom_kernel",
"line": 218,
"call": "F.layer_norm"
},
{
"owner": "custom_kernel",
"line": 221,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 190,
"call": "torch.ones"
},
{
"owner": "pfgv_k_8c88bc056b6c_batched_matmul_triton",
"line": 96,
"call": "torch.empty"
},
{
"owner": "pfgv_k_8c88bc056b6c_batched_matmul_triton",
"line": 104,
"call": "pfgv_k_8c88bc056b6c_batched_matmul_kernel[grid]"
}
],
"precision_prefix_lines": [
"x_norm = F.layer_norm(input_tensor, normalized_shape=(dim,), weight=norm_weight, bias=norm_bias, eps=eps)",
"out_norm = F.layer_norm(out, normalized_shape=(hidden_dim,), weight=to_out_norm_w, bias=to_out_norm_b, eps=eps)"
]
},
"cauchy_failure_patch.py": {
"source_sha256": "2838eaf2872444d4c8306fbe9c71eda600a876db463dff27174cfdbf6d1447da",
"defined_functions": [
"custom_kernel"
],
"reachable_functions": [
"custom_kernel"
],
"reachable_jit_functions": [],
"dormant_jit_functions": [],
"edges": {
"custom_kernel": []
},
"live_torch_and_launch_call_sites": [
{
"owner": "custom_kernel",
"line": 42,
"call": "F.layer_norm"
},
{
"owner": "custom_kernel",
"line": 45,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 46,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 49,
"call": "torch.sigmoid"
},
{
"owner": "custom_kernel",
"line": 50,
"call": "torch.sigmoid"
},
{
"owner": "custom_kernel",
"line": 51,
"call": "torch.sigmoid"
},
{
"owner": "custom_kernel",
"line": 72,
"call": "torch.bmm"
},
{
"owner": "custom_kernel",
"line": 95,
"call": "torch.empty"
},
{
"owner": "custom_kernel",
"line": 105,
"call": "_fused_norm_gate_linear_kernel[grid_m, grid_n]"
},
{
"owner": "custom_kernel",
"line": 49,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 50,
"call": "F.linear"
},
{
"owner": "custom_kernel",
"line": 51,
"call": "F.linear"
}
],
"precision_prefix_lines": [
"x = input_tensor.to(torch.float16)",
"x = F.layer_norm(x, (dim,), weight=norm_w, bias=norm_b, eps=eps)"
]
}
},
"cpu_checks": {
"torch_version": "2.9.1+cu128",
"device": "cpu",
"saved_shape_failure": {
"helper": "pfgv_k_8c88bc056b6c_batched_matmul_triton",
"input_shape": [
2,
8,
8,
4
],
"exception": "ValueError",
"message": "too many values to unpack (expected 3)",
"original_gpu_error_family": "shape_stride_error",
"reproduced_before_any_gpu_launch": true
},
"saved_precision_failure": [
{
"distribution": "normal",
"seed": 20261105,
"shape": [
1,
128,
128,
128
],
"input_nonfinite": 0,
"max_abs_input": 5.05343770980835,
"nonfinite_after_premature_half_cast": 0,
"nonfinite_actual_candidate_layernorm": 0,
"nonfinite_preserve_fp32_layernorm": 0,
"scope": "CPU normalization-prefix mechanism check; no full candidate GPU re-score"
},
{
"distribution": "cauchy_gamma_2",
"seed": 20261105,
"shape": [
1,
128,
128,
128
],
"input_nonfinite": 0,
"max_abs_input": 2817710.25,
"nonfinite_after_premature_half_cast": 39,
"nonfinite_actual_candidate_layernorm": 4992,
"nonfinite_preserve_fp32_layernorm": 0,
"scope": "CPU normalization-prefix mechanism check; no full candidate GPU re-score"
}
]
},
"limitations": [
"No measurement of candidate latency or improvement.",
"CPU input samples are not the original saved GPU input tensors.",
"Call graph is static; call-site counts are not profiler timings.",
"Historical ordinary run used local training; linked run used API training."
]
}

Xet Storage Details

Size:
16.2 kB
·
Xet hash:
e9d174204b77f501ecdd517e2c274969b700be479956ec09d797b589d3cbc8e1

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.