Fix prefix tuning finetune issue and update test (huggingface#975)

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

Author: sywangyi

optimum/habana/transformers/models/llama/modeling_llama.py

@@ -262,11 +262,10 @@ def update(self, prev, cur, dim, idx, inp_seq_len):
         if prev.shape == cur.shape:
             prev.copy_(cur)
             return orig_cur
-        if cur.shape[2] > 1 and cur.shape[2] <= prev.shape[2]:
+        if idx is not None and cur.shape[2] > 1 and cur.shape[2] <= prev.shape[2]:
             # Initialize
             prev[:, :, :inp_seq_len, :].copy_(cur)
             return orig_cur
-        assert cur.shape[2] == 1, f"Cannot update kv-cache. Unsupported shapes. prev:{prev.shape} cur:{cur.shape}"
         if idx is not None:
             prev.index_copy_(dim, idx - 1, cur)
             return prev

tests/baselines/llama_7b.json

@@ -230,16 +230,16 @@
                 "multi_card": {
                     "learning_rate": 5e-4,
                     "train_batch_size": 1,
-                    "train_runtime": 16.5,
-                    "train_samples_per_second": 63.161,
-                    "perplexity": 1.224,
+                    "train_runtime": 16.1,
+                    "train_samples_per_second": 63.249,
+                    "perplexity": 1.172,
                     "extra_arguments": [
                         "--num_virtual_tokens 8",
                         "--max_seq_length 64",
                         "--logging_steps 1",
                         "--report_to none",
                         "--max_steps 100",
-                        "--peft_type prompt_tuning",
+                        "--peft_type prefix_tuning",
                         "--max_seq_length 64",
                         "--lr_scheduler_type cosine",
                         "--warmup_steps 0",
@@ -256,16 +256,16 @@
                 "multi_card": {
                     "learning_rate": 5e-4,
                     "train_batch_size": 1,
-                    "train_runtime": 16.5,
+                    "train_runtime": 18.7,
                     "train_samples_per_second": 63.161,
-                    "perplexity": 1.224,
+                    "perplexity": 1.047,
                     "extra_arguments": [
                         "--num_virtual_tokens 8",
                         "--max_seq_length 64",
                         "--logging_steps 1",
                         "--report_to none",
                         "--max_steps 100",
-                        "--peft_type prompt_tuning",
+                        "--peft_type p_tuning",
                         "--max_seq_length 64",
                         "--lr_scheduler_type cosine",
                         "--warmup_steps 0",
@@ -276,4 +276,4 @@
             }
         }
     }
-}
+}