fix llama buck build (#13169)

Summary:

Some recent changes break the llama buck build

Reviewed By: rohansjoshi

Differential Revision: D79753385

Author: cccclai

examples/models/llama/evaluate/eager_eval.py

@@ -12,6 +12,7 @@
 from lm_eval.models.huggingface import HFLM as eval_wrapper
 from pytorch_tokenizers.llama2c import Llama2cTokenizer as SentencePieceTokenizer
 from pytorch_tokenizers.tiktoken import TiktokenTokenizer as Tiktoken
+from pytorch_tokenizers.hf_tokenizer import HuggingFaceTokenizer
 
 from torch import nn
 
@@ -24,7 +25,7 @@ class EagerEvalWrapper(eval_wrapper):
     def __init__(
         self,
         model: nn.Module,
-        tokenizer: Union[SentencePieceTokenizer, Tiktoken],
+        tokenizer: Union[SentencePieceTokenizer, Tiktoken, HuggingFaceTokenizer],
         max_seq_length: Optional[int] = None,
         use_kv_cache: bool = False,
     ):

examples/qualcomm/oss_scripts/llama/TARGETS

@@ -15,10 +15,30 @@ python_library(
     ],
 )
 
+python_library(
+    name = "decoder_utils",
+    srcs = [
+        "decoder_utils.py",
+    ],
+    deps = [
+        "//caffe2:torch",
+        "//executorch/examples/models/llama:eval_library",
+    ],
+)
+
+python_library(
+    name = "decoder_constants",
+    srcs = [
+        "decoder_constants.py",
+    ],
+)
+
 python_library(
     name = "llama_lib",
     srcs = ["llama.py"],
     deps = [
+        ":decoder_constants",
+        ":decoder_utils",
         "//executorch/examples/models/llama:source_transformation",
         "//caffe2:torch",
         "//executorch/backends/qualcomm/partition:partition",

examples/qualcomm/oss_scripts/llama/decoder_utils.py

@@ -44,14 +44,15 @@ def __init__(
         tokenizer: Union[
             SentencePieceTokenizer, TiktokenTokenizer, HuggingFaceTokenizer
         ],
-        max_seq_length: Optional[int],
+        max_seq_length: int,
         ar_len: int,
         use_kv_cache: bool,
         get_example_inputs: Callable,
         kv_updater: Callable,
         use_i64_token: bool,
     ):
         # n seq len = n-1 cache len, so we len(inps) = n-1 during _model_call
+        assert max_seq_length is not None, "max_seq_length must be provided"
         super().__init__(
             model=model, tokenizer=tokenizer, max_seq_length=max_seq_length - 1
         )
@@ -156,6 +157,7 @@ def __init__(
         )
         self.adb.push(inputs=[], input_list="", files=[self.runtime_tokenizer_path])
         # n seq len = n-1 cache len, so we len(inps) = n-1 during _model_call
+        # pyre-ignore
         super().__init__(None, tokenizer, max_seq_length - 1)
 
     def _model_call(self, inps):
@@ -278,6 +280,7 @@ def kv_inference(
         else:
             raise RuntimeError("Unknown tokenizer")
     else:
+        # pyre-ignore
         token_list = prompt.flatten().tolist()
     pos = len(token_list) if len(token_list) < ar_len else ar_len
     dtype = torch.int64 if use_i64_token else torch.int32
@@ -359,6 +362,7 @@ def prefill_inference(
         else:
             raise RuntimeError("Unknown tokenizer")
     else:
+        # pyre-ignore
         token_list = prompt.flatten().tolist()
 
     pos = len(token_list)
@@ -405,7 +409,7 @@ def graph_module_inference(
     max_seq_len=512,
     kv_updater=smart_mask_updater,
     use_i64_token=False,
-    event_name: str = None,
+    event_name: Optional[str] = None,
 ):
     if args.tasks is None:
         if use_kv_cache: