ONNX gelu float16 标量常量导致 pass_level3 崩溃 (std::map::at)

[dawnwtc]

error log

error_log.md

context | 编译/运行环境 |

• PNNX 版本：20260526（pip install pnnx）
• 系统：Ubuntu 22.04
• PyTorch：2.13.0

how to reproduce | 复现步骤

单个 GELU 算子导出 ONNX，float16 精度即可复现：

import torch, torch.nn as nn
m = nn.GELU().to(torch.float16).eval()
x = torch.randn(1, 16, 16, 16, dtype=torch.float16)
torch.onnx.export(m, x, 'gelu_f16.onnx',
    input_names=['a'], output_names=['b'],
    opset_version=18, dynamo=True)

pnnx gelu_f16.onnx optlevel=2

而 float32 的同样模型正常：

# 把上面脚本里 float16 改成 float32，重新导出
pnnx gelu_f32.onnx optlevel=2  # ✅ 正常通过

根因定位

第 1 步：加载 ONNX 标量常量时，pass_onnx.cpp 约第 1010–1026 行：

// INT32、INT64、FLOAT、BOOL 都有对应的分支，会设置 params["value"]
else if (tensor.data_type() == onnx::TensorProto::BOOL)
{
    op_const->params["value"] = ...;  // ✅
}
else
{
    // FLOAT16 落到了这里
    fprintf(stderr, "unknown constant scalar type %d\n", ...);
    // 只打了个 warning，params["value"] 没有被设置！
}

此时 prim::Constant 节点已经创建，但缺少 "value" 参数。GELU 公式中的 0.5、1.0、√2 等标量全部变成了这种「空壳」Constant。

第 2 步：pass_level3/eliminate_noop_math.cpp 第 15 行：

static bool constant_is_all_constant(const Operator* op_constant, ...)
{
    const Parameter& param = op_constant->params.at("value");  // ← 崩溃！
    // "value" 这个 key 不存在 → std::map::at 抛 std::out_of_range

[dosubot]

Hi @dawnwtc! I'm Dosu and I'm helping the 腾讯开源's Org team.

Your root cause analysis is spot-on. The scalar constant loading code in pass_onnx.cpp handles INT32, INT64, FLOAT, and BOOL but completely lacks a FLOAT16 branch [1]. When a float16 scalar falls through to the else case, params["value"] is never set, and then eliminate_noop_math.cpp crashes on params.at("value") [2].

Interestingly, PNNX already has float16 conversion infrastructure — float16_to_float32() and float32_to_float16() helpers were added in 2023 [3] — but they were never wired into the ONNX scalar constant loading path.

A fix would add a FLOAT16 case to pass_onnx.cpp following the same pattern as the existing FLOAT branch, using float16_to_float32() to decode the raw half-precision bytes and store the result as a float parameter. Something like:

else if (tensor.data_type() == onnx::TensorProto::FLOAT16)
{
    uint16_t raw;
    if (tensor.raw_data().size() == sizeof(uint16_t))
        memcpy(&raw, tensor.raw_data().data(), sizeof(uint16_t));
    else if (tensor.int32_data_size())
        raw = (uint16_t)tensor.int32_data(0);
    op_const->params["value"] = float16_to_float32(raw);
}

No existing fix or workaround is currently documented for this issue.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.