Segmentation fault on nn.Linear

[brazhenko]

Hi, trying to use torch2trt on a simple example. But i receive a strange sefault on nn.Linear. Here are details.

Libs:

tensorrt-cu12==10.0.1
tensorrt-cu12-bindings==10.0.1
tensorrt-cu12-libs==10.0.1
torch==2.0.0
torch2trt==0.5.0

Code:

class Example(nn.Module):
    def __init__(
            self,
            c_in: int,
            d_in: int = 2,
            out_channels: int = 9,
            n_layer: int = 3,
            n_head: int = 4,
            n_embd: int = 256,
            n_tokens: int = 32,
            bias: bool = False,  # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster
    ):
        super().__init__()

        self.n_embd = n_embd
        self.n_layer = n_layer

        self.pe = nn.Embedding(n_tokens, n_embd)

        self.encoder_c = nn.Linear(c_in, n_embd)

    def forward(self, features_c, d_emb, K_caches, V_caches):
        features_c = features_c.unsqueeze(0)  # 1, X, 1, c_in
        d_emb = d_emb.unsqueeze(0)  # 1, X, H

        T = K_caches.shape[3] + 1

        print(features_c.device, d_emb.device, K_caches.device, V_caches.device, self.encoder_c.weight.device)

        return features_c # return here works fine

        features_emb = self.encoder_c(features_c)

        return features_emb # return here fails with segfault

...

    n_x = 180
    n_tokens = 32
    c_in = 339
    d_in = 4

    model = Example(c_in=c_in, d_in=d_in)

    model = model.to("cuda:0").eval()

    model_trt = torch2trt(mlp_last,
                          [x_last.to("cuda:0"), d_emb.to("cuda:0"), k_cache.to("cuda:0"), v_cache.to("cuda:0")],
                          log_level=trt.Logger.VERBOSE)

[jaybdub]

Hi @brazhenko ,

Thanks for reaching out!

Does this issue occur if you call torch2trt(..., use_onnx=True)?

John