Update code/train_sft.py with chunked cross-entropy and H200 turbo speed optimizations

code/train_sft.py

@@ -771,7 +771,7 @@ def main():
                         c_logit = raw_m.head(h_i[:, ci:ce, :])
                         c_lbl = s_lbl[ci:ce].unsqueeze(0)
                         sample_loss_sum += F.cross_entropy(
-                            c_logit.view(-1, V), c_lbl.view(-1),
+                            c_logit.reshape(-1, V), c_lbl.reshape(-1),
                             ignore_index=-100, reduction="sum"
                         ).item()