[Distributed] Implement universal batch_decode & decode_in_flight for llama2 & llama3, with deterministic or multinomial (topk) decoding (handle both sentencepiece (llama2) and tiktoken (llama3)) (#1234)

* working multi-prompt same lengths

* working multi-prompt multiple lengths

* tighten up results decoding and display

* improve batch_decode_next_tokens

* update _decode_in_flight

* move prompt outside of main, auto-update batch size based on prompt

* faster batch_decode_next_tokens, add topk/temperature option

* ruff format and check

* simplify decode step, remove old comments

* add explanatory comment on topk min check

Author: lessw2020

dist_run.py

@@ -16,6 +16,8 @@
 
 import torch
 import torch.distributed as dist
+from torch.distributed.pipelining import PipelineStage, ScheduleGPipe
+from torchchat.cli.builder import _initialize_tokenizer, TokenizerArgs
 
 from torchchat.distributed.logging_utils import SingletonLogger
 
@@ -33,8 +35,6 @@
     get_num_params,
     GPUMemoryMonitor,
 )
-from torch.distributed.pipelining import PipelineStage, ScheduleGPipe
-from torchchat.cli.builder import _initialize_tokenizer, TokenizerArgs
 from torchchat.model import ModelArgs, Transformer, TransformerArgs
 from torchchat.utils.build_utils import set_precision
 
@@ -189,23 +189,49 @@ def _create_padded_prompts(
 
 def _batch_decode_next_tokens(
     output: torch.Tensor,
-    pos: int,
+    pos: List[int],
+    step: int = -1,
+    temperature: float = 1.0,
+    topk: int = 10,
 ) -> torch.Tensor:
     """
-    Decode the next token for each prompt in the batch.
+    Decode the next token for each prompt in the batch. Adds temperature option for non-deterministic decoding.
+
     Args:
         output (torch.Tensor): The output tensor to decode.
-        pos: the position of the `output` to decode in the sequence length dimension.
+        pos (List[int]): The positions of the `output` to decode in the sequence length dimension.
+        step (int): Step indicator. If -1, use positions from `pos`. Otherwise, use the first token.
+        temperature (float): Sampling temperature for non-deterministic decoding.
 
     Returns:
-        Decoded token ids.
+        torch.Tensor: Decoded token ids.
     """
-    # Take the next token logits for each prompt
-    next_token_logits = output[:, pos, :]
-    # Argmax (deterministic) TODO: add temperature
-    next_token = torch.argmax(next_token_logits, dim=-1)
-    # Token ids in int tensor form
-    return next_token
+    batch_size, seq_len, vocab_size = output.shape
+
+    if step != -1:
+        next_token_logits = output[:, 0, :]
+    else:
+        # get the logits for each prompt at the specified positions
+        next_token_logits = output[torch.arange(batch_size), torch.tensor(pos) - 1]
+
+    if temperature != 1.0:
+        next_token_logits = next_token_logits / temperature
+
+    # Uses top-k sampling if temperature is not 1.0, otherwise use argmax
+    if temperature != 1.0:
+        top_k = min(topk, vocab_size)  # Ensure top-k is not greater than vocab size
+        top_k_logits, top_k_indices = torch.topk(next_token_logits, k=top_k, dim=-1)
+        probs = torch.softmax(top_k_logits, dim=-1)
+        next_token_indices = torch.multinomial(probs, num_samples=1).squeeze(-1)
+        next_tokens = top_k_indices.gather(
+            -1, next_token_indices.unsqueeze(-1)
+        ).squeeze(-1)
+    else:
+        # Argmax (deterministic)
+        next_tokens = torch.argmax(next_token_logits, dim=-1)
+
+    logger.info(f"{color.yellow}Next tokens: {color.blue}{next_tokens}{color.reset}")
+    return next_tokens
 
 
 def _update_padded_sequence(
@@ -218,11 +244,32 @@ def _update_padded_sequence(
         # logger.info(f"updated prompt {i} with new token {new_token[i, 0]}")
 
 
+# Decode token id into string and print it
+def _decode_in_flight(token, tokenizer, tp_rank):
+    """decode token ids for all prompts in the batch and log them"""
+    token_str = tokenizer.decode(token.tolist())
+    # print the token string on tp rank 0
+    if tp_rank == 0:
+        logger.info(
+            f"{color.green} responses ====>>>> "
+            f"{color.blue} {token_str} {color.reset}"
+        )
+
+
 def _cleanup():
     dist.barrier()
     dist.destroy_process_group()
 
 
+prompt = [
+    "What is Snow?",
+    "Who is Santa Claus?",
+    "Where does Santa live?",
+    # "Who is Abraham Lincoln?",
+    # "How are models trained?",
+]
+
+
 def main(args):
     model_name = args.model_name
     pp_degree = args.pp
@@ -293,7 +340,7 @@ def main(args):
     # Batch size. Since we push batches dynamically through the pipeline rather
     # than chunking them, this is effectively micro-batch size in pipeline
     # sense. Thus it is interchangeable with micro-batch size below.
-    batch_size = 4
+    batch_size = len(prompt)
     seqlen_prefill = 1024  # sequence length
     dim = 4096  # embedding dimension
 
@@ -331,7 +378,9 @@ def main(args):
 
     # Helper function to get example inputs and outputs for the stages.
     def get_example_ins_outs(seqlen: int) -> Tuple[torch.Tensor, torch.Tensor]:
-        mb_ids = torch.randint(0, config.vocab_size, (batch_size, seqlen), device=device)
+        mb_ids = torch.randint(
+            0, config.vocab_size, (batch_size, seqlen), device=device
+        )
         activation = torch.rand(
             batch_size, seqlen, dim, device=device, dtype=model_dtype
         )
@@ -362,13 +411,6 @@ def get_example_ins_outs(seqlen: int) -> Tuple[torch.Tensor, torch.Tensor]:
     # pipelining effect.
     prefiller = ScheduleGPipe(prefill_stage, 1)
 
-    prompt = [
-        "What is a computer?",
-        "Where does Santa live?",
-        "Who is Abraham Lincoln?",
-        "How are models trained?",
-    ]
-
     start_pos = 0
 
     # Need these global ids due to the API definition of dist.send and recv
@@ -384,10 +426,6 @@ def get_example_ins_outs(seqlen: int) -> Tuple[torch.Tensor, torch.Tensor]:
     padded_sequence, prompt_lengths = _create_padded_prompts(
         input_ids, tokenizer, seqlen_prefill, start_pos, device
     )
-    # TODO: figure out how to set input_pos for each prompt in the batch then we
-    # can remove this limitation.
-    s = set(prompt_lengths)
-    assert len(s) == 1, f"prompt_lengths should be the same, got {s}"
 
     # Need these global ids due to the API definition of dist.send and recv
     first_pp_rank_global_id = dist.get_global_rank(pp_group, first_pp_rank)
@@ -396,6 +434,7 @@ def get_example_ins_outs(seqlen: int) -> Tuple[torch.Tensor, torch.Tensor]:
     # New token generated each iteration
     # need a row dimension for each prompt in the batch
     new_token = torch.zeros(batch_size, 1, device=device, dtype=torch.int64)
+    logger.info(f"{color.green}{new_token.shape=}, {new_token=}{color.reset}")
     # Store the generated tokens
     res = []
 
@@ -416,23 +455,13 @@ def get_example_ins_outs(seqlen: int) -> Tuple[torch.Tensor, torch.Tensor]:
         f"{color.green}Prefilling time: {timer.get_time()} {timer.unit} for rank {rank}{color.reset}"
     )
 
-    # Decode token id into string and print it
-    def decode_in_flight(token):
-        # Make a 2D tensor with ids on row dimension
-        unsqueezed = torch.unsqueeze(token, 1)
-        token_str = tokenizer.decode(unsqueezed.tolist())
-        if tp_rank == 0:
-            logger.info(
-                f"{color.green} responses ====>>>> "
-                f"{color.blue} {token_str} {color.reset}"
-            )
-
     # Decode the output -- first generated token
     if pp_rank == last_pp_rank:
-        new_token = _batch_decode_next_tokens(output, prompt_lengths[0] - 1)
+        logger.info(f"{color.green}Decoding...{prompt_lengths=}{color.reset}")
+        new_token = _batch_decode_next_tokens(output, prompt_lengths)
         res.append(new_token)
         if not args.disable_in_flight_decode:
-            decode_in_flight(new_token)
+            _decode_in_flight(new_token, tokenizer, tp_rank)
 
     # seqlen = 1 now
     seqlen_decode = 1
@@ -482,10 +511,11 @@ def decode_in_flight(token):
 
             # Decode the output
             if pp_rank == last_pp_rank:
-                new_token = _batch_decode_next_tokens(output, 0)
+                # logger.info(f"{color.red}Decoding...{output.shape=}{color.reset}")
+                new_token = _batch_decode_next_tokens(output, prompt_lengths, step)
                 res.append(new_token)
                 if not args.disable_in_flight_decode:
-                    decode_in_flight(new_token)
+                    _decode_in_flight(new_token, tokenizer, tp_rank)
 
             # Increment input position
             input_pos += 1
@@ -499,12 +529,17 @@ def decode_in_flight(token):
     # output formatted response via last pp group and tp rank 0
     if pp_rank == last_pp_rank and tp_rank == 0:
         # `res` is a list of tensors, each being a batch of generated token ids
-        res = torch.stack(res, dim=1)
-        res_list = res.tolist()
-        response = tokenizer.decode(res_list)
-        for i in range(len(response)):
-            logger.info(f"Prompt: {color.green}{prompt[i]} {color.reset}")
-            logger.info(f"Response: {color.red}{response[i]} {color.reset}")
+
+        res_stacked = torch.stack(res, dim=1)
+        res_list = res_stacked.tolist()
+
+        # Decode the output as comprehension instead of loop
+        responses = [tokenizer.decode(sequence) for sequence in res_list]
+
+        # Show prompts and responses
+        for prompt_text, response_text in zip(prompt, responses):
+            logger.info(f"Prompt: {color.green}{prompt_text} {color.reset}")
+            logger.info(f"Response: {color.red}{response_text} {color.reset}")
 
     # Cleanup
     _cleanup()