HF "deepseek-ai/deepseek-v3" TransformerLayer fails to run with thunder.jit and thunderfx

@kshitij12345 is already working on this.

Since Apr 29, 2025.

Repro
```python
import nvfuser
import pytest
import transformers
import torch
import torch.distributed as dist
from contextlib import contextmanager
from torch.distributed.tensor import DTensor
from torch.distributed.tensor.parallel import (
parallelize_module,
RowwiseParallel,
ColwiseParallel,
)

@contextmanager
def default_tensor_type(dtype=torch.float32, device="cpu"):
# Save
prev_dtype = torch.get_default_dtype()
prev_device = torch.get_default_device()

# Set
torch.set_default_dtype(dtype)
torch.set_default_device(device)

yield

# Restore
torch.set_default_dtype(prev_dtype)
torch.set_default_device(prev_device)

config = transformers.AutoConfig.from_pretrained(
"deepseek-ai/deepseek-v3", trust_remote_code=True
)

# Create only one layer which is sufficient for the test.
config.num_hidden_layers = 1
# Without this, the first and only layer will have a dense MLP instead of MoE.
config.first_k_dense_replace = 0
# Disable quantization so the test can run on A100 and is made easier for nvFuser.
delattr(config, "quantization_config")

torch.manual_seed(0)

with default_tensor_type(dtype=config.torch_dtype, device="cuda"):
model = transformers.AutoModel.from_config(config, trust_remote_code=True)
# Training is unavailable (cf. https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/main/modeling_deepseek.py#L439)
model.eval()

transformer_layer = model.layers[0]

batch_size = 1
seq_len = 2048
inp = torch.randn(batch_size, seq_len, config.hidden_size)
mask = transformers.modeling_attn_mask_utils._prepare_4d_causal_attention_mask(
None, [batch_size, seq_len], inp, past_key_values_length=0
)
import thunder
from thunder.dynamo import thunderfx

# AssertionError: A symbol [Symbol name=_set_grad_enabled_with_warning] was called while processing a primitive
# tfn = thunder.jit(transformer_layer)

# torch._dynamo.exc.TorchRuntimeError: Failed running call_function >(*(FakeTensor(..., size=(), dtype=torch.int64), 0), **{}): 'ndarray' object has no attribute 'add'
tfn = thunderfx(transformer_layer)

# Error: torch._dynamo.exc.TorchRuntimeError: Failed running call_function >(*(FakeTensor(..., size=(), dtype=torch.int64), 0), **{}): 'ndarray' object has no attribute 'add'
# tfn = torch.compile(transformer_layer)

# Eager Works
# tfn = transformer_layer
(out,) = tfn(inp, attention_mask=mask)

# thunder.last_traces(tfn)[-1].save_trace("deepseek_tfm_layer_fwd.py")
# thunder.last_backward_traces(tfn)[-1].save_trace("deepseek_tfm_layer_bwd.py")
# Finish all computation and communication. Otherwise,
# destroy_process_group may deadlock.
torch.cuda.synchronize()

assert out.size() == (batch_size, seq_len, config.hidden_size)
assert out.dtype == config.torch_dtype
assert out.is_cuda

```

In case of `thunderfx`, the error comes from torch-dynamo (torch.compile)

Assigning myself for investigation.