fix the corner case where T < 256

train_gpt2.cu

@@ -1733,7 +1733,7 @@ int main(int argc, char *argv[]) {
                 // on cuDNN 9.2.1 with cuDNN FrontEnd 1.5.2, T >= 256 seems bit-for-bit identical
                 // (but even if it wasn't fully identical that's probably not the end of the world)
                 // note this is still somewhat wasteful because we don't have a KV cache!
-                gpt2_forward(&model, gen_tokens, 1, CEIL_DIV(t, 256) * 256);
+                gpt2_forward(&model, gen_tokens, 1, CEIL_DIV(t, min(T,256)) * min(T,256));
                 // get the V-dimensional vector probs[0, t-1, :]
                 floatX* logits = model.acts.output + (t - 1) * model.config.padded_vocab_size;
                 // move probs back to CPU and sample (note we only move the first vocab_size logits, ignoring the padding)