Job instance 271223

Stderr output

<core_client_version>8.0.2</core_client_version>
<![CDATA[
<stderr_txt>
14:24:08 (13472): wrapper (7.7.26016): starting
14:24:08 (13472): wrapper: running ../../projects/axiom.heliex.net/axiom_worker_2.17_windows_x86_64.exe (wu.json result.json)
[CPU] Using NumPy
[Worker] Saved 5 gradient files to contribute folder (gradient sharing)
INFO: Downloading model: https://axiom.heliex.net/pyhelix/model/expert_190_v2.0.bin
INFO: Downloaded and cached model: expert_190_v2.0.bin (42580994 params)
[Worker] Running 30 batches for expert 190...
[Worker] FATAL ERROR: MemoryError: Unable to allocate 12.0 MiB for an array with shape (64, 64, 768) and data type float32
Traceback (most recent call last):
  File "worker_boinc_app.py", line 1699, in main
  File "worker_boinc_app.py", line 1322, in step
  File "simple_ml.py", line 1176, in backward
    return self.expert.backward_softmax_cross_entropy(x, y_idx, logits)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "simple_ml.py", line 649, in backward_softmax_cross_entropy
    self._backward_chunk(x, y_idx, logits, sc['ctx'], grads, B_total)
  File "simple_ml.py", line 736, in _backward_chunk
    d_h_pre_ln2, d_lg2, d_lb2 = _layernorm_backward(
                                ^^^^^^^^^^^^^^^^^^^^
  File "simple_ml.py", line 336, in _layernorm_backward
    d_x = inv_std * (dx_hat - np.mean(dx_hat, axis=-1, keepdims=True)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
numpy._core._exceptions._ArrayMemoryError: Unable to allocate 12.0 MiB for an array with shape (64, 64, 768) and data type float32

14:24:51 (13472): axiom_worker_2.17_windows_x86_64.exe exited; CPU time 50.546875
14:24:51 (13472): called boinc_finish(0)

</stderr_txt>
]]>