Merge branch 'main' into load_var

Author: lessw2020

dist_run.py

@@ -10,6 +10,7 @@
 
 import argparse
 import os
+from enum import auto, Enum
 from pathlib import Path
 from types import SimpleNamespace
 from typing import Any, Dict, List, Optional, Tuple
@@ -49,6 +50,7 @@
 
 
 logger = SingletonLogger.get_logger()
+_tokenizer_type = None  # global variable to store the tokenizer type
 
 # Using model name to identify the model to load, for example "llama2-7b-chat".
 # You can change it to other values listed below.
@@ -59,6 +61,11 @@
 }
 
 
+class TokenizerType(Enum):
+    Tiktoken = auto()
+    SentencePiece = auto()
+
+
 def _init_distributed():
     dist.init_process_group("nccl")
     rank = dist.get_rank()
@@ -80,7 +87,10 @@ def _build_chat_tokenizer(
     model_name: str,
     model_base_name: Optional[str] = None,
 ) -> SentencePieceProcessor | TiktokenTokenizer:
-    """Builds a tokenizer for the given model name."""
+    """Builds a tokenizer for the given model name, and sets the global tokenizer type variable"""
+
+    global _tokenizer_type
+
     # Try to infer the model base name from the model name:
     # e.g. "llama2-7b-chat" -> "llama2"
     if model_base_name is None:
@@ -107,6 +117,15 @@ def _build_chat_tokenizer(
     logger.info(
         f"using tokenizer = {tokenizer.__class__.__module__}.{tokenizer.__class__.__name__}"
     )
+    # set global variable _tokenizer_type
+    if isinstance(tokenizer, TiktokenTokenizer):
+        _tokenizer_type = TokenizerType.Tiktoken
+    elif isinstance(tokenizer, SentencePieceProcessor):
+        _tokenizer_type = TokenizerType.SentencePiece
+    else:
+        raise ValueError(f"Unknown tokenizer type: {tokenizer.__class__}")
+
+    logger.info(f"tokenizer type = {_tokenizer_type}")
     return tokenizer
 
 
@@ -276,6 +295,7 @@ def _cleanup():
 
 prompt = [
     "What is Snow?",
+    # "Can you explain what is the purpose of back propagation in neural networks?",
     "Who is Santa Claus?",
     "Where does Santa live?",
     # "Who is Abraham Lincoln?",
@@ -494,7 +514,7 @@ def get_example_ins_outs(seqlen: int) -> Tuple[torch.Tensor, torch.Tensor]:
         group=pp_group,
     )
     # create schedule
-    decorder = ScheduleGPipe(decode_stage, 1)
+    decoder = ScheduleGPipe(decode_stage, 1)
 
     # Decoding
     with torch.no_grad(), CUDATrackTime() as timer:
@@ -517,11 +537,11 @@ def get_example_ins_outs(seqlen: int) -> Tuple[torch.Tensor, torch.Tensor]:
 
             # Run data through pipeline
             if pp_rank == first_pp_rank:
-                output = decorder.step(new_token, **kwargs)
+                output = decoder.step(new_token, **kwargs)
             elif pp_rank == last_pp_rank:
-                output = decorder.step(**kwargs)
+                output = decoder.step(**kwargs)
             else:  # middle pp ranks
-                decorder.step(**kwargs)
+                decoder.step(**kwargs)
 
             # Decode the output
             if pp_rank == last_pp_rank:
@@ -546,13 +566,16 @@ def get_example_ins_outs(seqlen: int) -> Tuple[torch.Tensor, torch.Tensor]:
         # token ids. Thus cat'ing along dim 1.
         res = torch.cat(res, dim=1)
         res_list = res.tolist()
-        if isinstance(tokenizer, TiktokenTokenizer):
+        if _tokenizer_type == TokenizerType.Tiktoken:
             # For TiktokenTokenizer, we need to decode prompt by prompt.
             # TODO: is there a better way to do this?
             responses = [tokenizer.decode(sequence) for sequence in res_list]
-        else:  # SentencePieceProcessor
+        elif _tokenizer_type == TokenizerType.SentencePiece:  # SentencePieceProcessor
             # For SentencePieceProcessor, we can decode the entire 2D list at once.
             responses = tokenizer.decode(res_list)
+        else:
+            raise ValueError(f"Unknown tokenizer type {_tokenizer_type}")
+
         # Show prompts and responses
         for prompt_text, response_text in zip(prompt, responses):
             logger.info(f"Prompt: {color.green}{prompt_text} {color.reset}")

torchchat/generate.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the license found in the
 # LICENSE file in the root directory of this source tree.
 import argparse
+import base64
 import itertools
 import logging
 import os
@@ -12,6 +13,7 @@
 
 from abc import ABC, abstractmethod
 from dataclasses import dataclass
+from io import BytesIO
 from os import PathLike
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Sequence, Tuple, Union
@@ -600,9 +602,8 @@ def generate(
 
         if len(prompt.shape) > 1:
             prompt = prompt.squeeze(0)
-        T = prompt.size(0)
-        max_new_tokens = min(max_new_tokens, max_seq_length - start_pos - T)
-        T_new = T + max_new_tokens
+        prompt_length = prompt.size(0)
+        max_new_tokens = min(max_new_tokens, max_seq_length - start_pos - prompt_length)
         # set up caches only if first inference
         if start_pos == 0:
             model = model.to(device=device)
@@ -616,7 +617,7 @@ def generate(
                         batch_size=1,
                         dtype=self.dtype,
                         encoder_max_seq_len=6404,
-                        decoder_max_seq_len=T_new,
+                        decoder_max_seq_len=max_seq_length,
                     )
                 else:
                     model.setup_caches(max_batch_size=1, max_seq_length=max_seq_length)
@@ -629,7 +630,7 @@ def generate(
                 model.reset_caches()
 
         input_pos = torch.arange(
-            start_pos, T + start_pos, device=device, dtype=torch.int
+            start_pos, prompt_length + start_pos, device=device, dtype=torch.int
         )
 
         prefill_t0 = time.perf_counter()
@@ -655,7 +656,9 @@ def generate(
         # max_new_tokens <= 2 means we are effectively not calling decode_n_tokens().
         callback(next_token.clone().view(-1), done_generating=max_new_tokens <= 2)
 
-        input_pos = torch.tensor([start_pos + T], device=device, dtype=torch.int)
+        input_pos = torch.tensor(
+            [start_pos + prompt_length], device=device, dtype=torch.int
+        )
         accept_counts = [0] * (
             speculate_k + 1
         )  # creates array of [0, 0, 0, ...] that is speculate_k + 1 long
@@ -678,7 +681,7 @@ def generate(
                 )
 
                 accept_counts[len(next_tokens) - 1] += 1
-                num_added = min(T_new - input_pos - 1, len(next_tokens))
+                num_added = min(max_new_tokens - input_pos - 1, len(next_tokens))
                 for token in next_tokens[:num_added,]:
                     callback(token)
                     yield token, None
@@ -741,6 +744,7 @@ def _gen_model_input(
         prompt: Union[str | List[Any]],
         image_prompts: Optional[List[str | Image.Image]] = None,
         max_new_tokens: Optional[int] = None,
+        max_seq_len: Optional[int] = 2048,
     ) -> Tuple[torch.Tensor, Optional[Dict[str, Any]]]:
         """
         Convert prompt and image prompts into consumable model input args.
@@ -757,7 +761,7 @@ def _gen_model_input(
             Tuple[torch.Tensor, Optional[Dict[str, Any]]]: Encoded prompt and batch config for multimodal models.
         """
 
-        # Not Llama 3.2 11B
+        # Text-Only model
         if self.model.config.model_type != ModelType.Flamingo:
             # Single String prompt
             if isinstance(prompt, str):
@@ -778,32 +782,69 @@ def _gen_model_input(
         assert (
             image_prompts is None or len(image_prompts) == 1
         ), "At most one image is supported at the moment"
+
         if image_prompts and isinstance(image_prompts[0], str):
             images = [Image.open(image_prompts[0])]
         else:
-            images = image_prompts
+            images = None
 
         assert (
             max_new_tokens is not None
         ), "max_new_tokens must be specified for Flamingo models"
-        assert isinstance(
-            prompt, str
-        ), "(Currently) prompt must be a str for Flamingo models"
 
-        is_multimodal = images is not None
-        content = [{"type": "text", "content": prompt}]
+        image_found = False
+        messages = []
+        for message in prompt:
+            if isinstance(message["content"], str):
+                if not image_found and image_prompts:
+                    messages.append(
+                        Message(
+                            role=message["role"],
+                            content=[
+                                {"type": "image", "content": images[0]},
+                                {"type": "text", "content": message["content"]},
+                            ],
+                        )
+                    )
+                    image_found = True
+                else:
+                    messages.append(Message(**message))
+
+            elif isinstance(message["content"], list):
+                images = None
+                for content_dict in message["content"]:
+                    if content_dict["type"] == "text":
+                        prompt_arg = content_dict["text"]
+                    elif content_dict["type"] == "image_url":
+                        assert (
+                            images is None
+                        ), "At most one image is supported at the moment"
+
+                        base64_decoded = base64.b64decode(
+                            content_dict["image_url"].split(";base64,")[1]
+                        )
+                        images = [Image.open(BytesIO(base64_decoded))]
+                        image_found = True
+
+                is_multimodal = images is not None
+                content = [{"type": "text", "content": prompt_arg}]
+
+                if is_multimodal:
+                    content = [{"type": "image", "content": images[0]}] + content
 
-        if is_multimodal:
-            content = [{"type": "image", "content": images[0]}] + content
+                messages.append(
+                    Message(
+                        role=message["role"],
+                        content=content,
+                    )
+                )
 
-        messages = [
+        messages.append(
             Message(
-                role="user",
-                content=content,
-                eot=True,
-            ),
-            Message(role="assistant", content=""),
-        ]
+                role="assistant",
+                content="",
+            )
+        )
 
         transform = llama3_2_vision_transform(str(self.tokenizer_args.tokenizer_path))
 
@@ -812,7 +853,7 @@ def _gen_model_input(
         with device, set_default_dtype(self.dtype):
             data = transform({"messages": messages}, inference=True)
 
-            if is_multimodal:
+            if image_found:
                 batch = padded_collate_tiled_images_and_mask(
                     [data], pad_direction="left", pad_max_images=1
                 )
@@ -822,17 +863,27 @@ def _gen_model_input(
                 batch["encoder_input"]["images"] = batch["encoder_input"]["images"].to(
                     self.dtype
                 )
+
             else:
                 encoded = torch.tensor(data["tokens"], device=device).view(-1)
                 seq_len = encoded.size(0)
                 batch = {}
 
             total_response_length = seq_len + max_new_tokens
-            batch["causal_mask"] = torch.tril(
-                torch.ones(
-                    size=(total_response_length, total_response_length),
-                    dtype=torch.bool,
-                )
+            batch["causal_mask"] = torch.nn.functional.pad(
+                torch.tril(
+                    torch.ones(
+                        size=(total_response_length, total_response_length),
+                        dtype=torch.bool,
+                    )
+                ),
+                (
+                    0,
+                    max_seq_len - total_response_length,
+                    0,
+                    max_seq_len - total_response_length,
+                ),
+                value=0,
             )
 
         logging.debug(encoded)
@@ -845,12 +896,6 @@ def chat(
         if generator_args.chat_mode:
             print("Starting Interactive Chat")
 
-        encoded, batch = self._gen_model_input(
-            generator_args.prompt,
-            generator_args.image_prompts,
-            generator_args.max_new_tokens,
-        )
-
         model_size = sum(
             [
                 p.numel() * p.dtype.itemsize
@@ -896,6 +941,12 @@ def chat(
         max_seq_length = (
             text_transformer_args.max_seq_length if text_transformer_args else 2048
         )
+        encoded, batch = self._gen_model_input(
+            [{"role": "user", "content": generator_args.prompt}],
+            generator_args.image_prompts,
+            generator_args.max_new_tokens,
+            max_seq_length,
+        )
 
         if generator_args.chat_mode:
             print(
@@ -907,16 +958,16 @@ def chat(
             if get_system_prompt == "y" or get_system_prompt == "Y":
                 self.system_prompt = input("What is your system prompt? \n")
 
-        elif not generator_args.is_torchtune_model:
-            max_seq_length = min(
-                encoded.size(0) + generator_args.max_new_tokens,
-                (
-                    text_transformer_args.block_size
-                    if text_transformer_args is not None
-                    else 2048
-                ),
-                max_seq_length,
-            )
+        # elif not generator_args.is_torchtune_model:
+        #     max_seq_length = min(
+        #         encoded.size(0) + generator_args.max_new_tokens,
+        #         (
+        #             text_transformer_args.block_size
+        #             if text_transformer_args is not None
+        #             else 2048
+        #         ),
+        #         max_seq_length,
+        #     )
 
         max_seq_length = (
             max_seq_length + self.speculative_builder_args.speculate_k + 1