Fix torchtitan eval logits

python/python/psyche/models/ttitan.py

@@ -436,7 +436,7 @@ def forward(
                     ),
                 )
                 if num_logits_to_keep:
-                    pred = pred[:, -num_logits_to_keep, :]
+                    pred = pred[:, -num_logits_to_keep:, :]
                 loss = None
                 if labels is not None:
                     if labels.shape != pred.shape[:2]:

shared/eval/src/harness.rs

@@ -493,29 +493,24 @@ impl PreparedTask {
 
                 // The request already contains [fewshot_tokens] + [question + choice_without_last_token]
                 let full_request = request;
-                let input_length = &full_request.len();
 
                 let request_tensor = Tensor::from_slice(&full_request)
                     .to(options.model.device())
                     .unsqueeze(0);
                 let (logits, _) = {
                     let _no_grad = tch::no_grad_guard();
-                    options
-                        .model
-                        .forward(&request_tensor, None, None, None, None, None)
+                    options.model.forward(
+                        &request_tensor,
+                        None,
+                        None,
+                        None,
+                        Some(choice.len() as i64),
+                        None,
+                    )
                 };
 
-                let logits = logits.unwrap().squeeze_dim(0).slice(0, 0, None, 1);
-
-                // Get tensor of shape `[choice.len(), vocab_size]` containing the
-                // model's logits for each token of the `choice` text.
-                // This should skip the fewshot tokens and get the tokens from the end.
-                let logits = logits.slice(
-                    0,
-                    *input_length as i64 - choice.len() as i64,
-                    *input_length as i64,
-                    1,
-                );
+                // Shape: [choice.len(), vocab_size]
+                let logits = logits.unwrap().squeeze_dim(0);
 
                 let greedy_tokens: Vec<i64> = logits.argmax(-1, false).try_into().unwrap();
                 let exact_match = greedy_tokens.eq(&choice);