test_populate_grads for nanogpt - use smaller model for avoiding OOMs…

… in CIs

thunder/tests/test_grad.py

@@ -1395,7 +1395,7 @@ def test_populate_grads_nanogpt(executor, device, dtype):
     from thunder.benchmarks import NanoGPTBenchmark, NanoGPTConfig
 
     # NOTE Currently setting dropout to zero for reproducibility, other settings taken from gpt2 config
-    config = NanoGPTConfig(dropout=0, n_layer=12, n_head=12, n_embd=768)
+    config = NanoGPTConfig(dropout=0, n_layer=12, n_head=12, n_embd=384)
 
     bench = NanoGPTBenchmark(config=config, requires_grad=True, device=device, dtype=dtype)
     model = bench.fn()