Fix max seq length bug

Differential Revision: D84562463

Pull Request resolved: #15084

(cherry picked from commit f1e2548)

Author: kimishpatel

extension/llm/custom_ops/custom_ops.py

@@ -207,8 +207,8 @@ def _validate_update_cache_params(
             1
         ), f"Start position {start_pos} must be less than sequence length {cache.size(1)}"
 
-        torch._check((start_pos + seq_len) < cache.size(1))
-        assert (start_pos + seq_len) < cache.size(
+        torch._check((start_pos + seq_len) <= cache.size(1))
+        assert (start_pos + seq_len) <= cache.size(
             1
         ), f"Start position  + length = {start_pos + seq_len} must be less than sequence length {cache.size(1)}"
 

extension/llm/export/builder.py

@@ -142,7 +142,8 @@ def __init__(
                     {1: torch.export.Dim("token_dim", max=self.max_seq_len - 1)},
                 )
             else:
-                # Two input arguments: tokens and input_pos but input_pos is static shape
+                # Two input arguments: tokens and input_pos but input_pos is static shape.
+
                 self.dynamic_shapes = (
                     {1: torch.export.Dim("token_dim", max=self.max_seq_len)},
                     {"input_pos": {0: 1}},