Add support for transformers-neuronx continuous batching

Makefile

@@ -40,7 +40,7 @@ PACKAGE_FILES = $(PACKAGE_PYTHON_FILES)  \
 $(PACKAGE_DIST) $(PACKAGE_WHEEL): $(PACKAGE_FILES)
 	python -m build
 
-TGI_VERSION ?= 1.4.0
+TGI_VERSION ?= 1.4.1
 
 neuronx-tgi: $(PACKAGE_DIST)
 	docker build --rm -f text-generation-inference/Dockerfile \

optimum/exporters/neuron/base.py

@@ -379,7 +379,7 @@ class NeuronDecoderConfig(NeuronConfig):
         be passed to export the model,
     - NEURONX_CLASS (`str`) -- the name of the transformers-neuronx class to instantiate for the model.
     It is a full class name defined relatively to the transformers-neuronx module, e.g. `gpt2.model.GPT2ForSampling`
-    [`~optimum.utils.DummyInputGenerator`] specifying how to create dummy inputs.
+    - CONTINUOUS_BATCHING (`bool`, , defaults to `False`) -- Whether the model supports continuous batching or not.
 
     The NEURONX_CLASS must always be defined in each model configuration.
 
@@ -389,6 +389,7 @@ class NeuronDecoderConfig(NeuronConfig):
 
     INPUT_ARGS = ("batch_size", "sequence_length")
     NEURONX_CLASS = None
+    CONTINUOUS_BATCHING = False
 
     def __init__(self, task: str):
         if not is_transformers_neuronx_available():
@@ -404,3 +405,7 @@ def __init__(self, task: str):
     @property
     def neuronx_class(self):
         return self._neuronx_class
+
+    @property
+    def continuous_batching(self):
+        return self.CONTINUOUS_BATCHING

optimum/exporters/neuron/model_configs.py

@@ -431,6 +431,7 @@ class GPT2NeuronConfig(TextNeuronDecoderConfig):
 @register_in_tasks_manager("llama", "text-generation")
 class LLamaNeuronConfig(TextNeuronDecoderConfig):
     NEURONX_CLASS = "llama.model.LlamaForSampling"
+    CONTINUOUS_BATCHING = True
 
 
 @register_in_tasks_manager("t5-encoder", "text2text-generation")
@@ -533,3 +534,4 @@ def generate_io_aliases(self, model):
 @register_in_tasks_manager("mistral", "text-generation")
 class MistralNeuronConfig(TextNeuronDecoderConfig):
     NEURONX_CLASS = "mistral.model.MistralForSampling"
+    CONTINUOUS_BATCHING = True

optimum/neuron/modeling.py

@@ -16,7 +16,7 @@
 
 import copy
 import logging
-from typing import TYPE_CHECKING, Dict, Optional, Union
+from typing import TYPE_CHECKING, Dict, List, Optional, Union
 
 import torch
 from transformers import (
@@ -656,14 +656,14 @@ def __init__(
         generation_config: Optional["GenerationConfig"] = None,
     ):
         super().__init__(config, checkpoint_dir, compiled_dir=compiled_dir, generation_config=generation_config)
-        self.cur_len = 0
-        self.batch_size = self.model.config.batch_size
-        self.max_length = self.model.config.n_positions
+        self.batch_size = self.config.neuron["batch_size"]
+        self.max_length = self.config.neuron["sequence_length"]
+        self.continuous_batching = self.model.neuron_config and self.model.neuron_config.continuous_batching
         # The generate method from GenerationMixin expects the device attribute to be set
         self.device = torch.device("cpu")
 
     def reset_generation(self):
-        self.cur_len = 0
+        pass
 
     @add_start_docstrings_to_model_forward(
         NEURON_CAUSALLM_MODEL_FORWARD_DOCSTRING
@@ -688,32 +688,78 @@ def forward(
             return ModelOutput([("logits", out_logits)])
         return (out_logits,)
 
-    def prepare_inputs_for_generation(
-        self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs
-    ) -> Dict[str, torch.Tensor]:
-        # convert attention_mask to start_ids
+    def get_start_ids(
+        self,
+        input_ids: torch.Tensor,
+        attention_mask: torch.Tensor,
+        seq_ids: Optional[torch.Tensor] = None,
+    ):
+        # The start_ids parameter has different meanings:
+        # - for continuous (unpadded) batching it corresponds to the sequence id,
+        # - for static batching it corresponds to the start of the padded sequence.
+        if self.continuous_batching:
+            if seq_ids is None:
+                seq_ids = torch.arange(input_ids.shape[0])
+            else:
+                assert seq_ids.shape[0] == input_ids.shape[0]
+            return seq_ids
         start_ids = None
         if attention_mask is not None:
             _, start_ids = attention_mask.max(axis=1)
-
-        if self.cur_len > 0:
-            # Only pass the last tokens of each sample
-            input_ids = input_ids[:, -1:]
-            # Specify the single index at which the new keys and values need to be stored
-            cache_ids = torch.as_tensor([self.cur_len], dtype=torch.int32)
-        else:
-            # cache_ids will be set directly by the parallel context encoding code
-            cache_ids = None
-
-        # Increment the current cache index
-        self.cur_len += input_ids.shape[-1]
-        model_inputs = {
+        return start_ids
+
+    def get_cache_ids(self, attention_mask: torch.tensor, prefill: bool):
+        cache_n, cache_len = attention_mask.shape
+        if self.continuous_batching:
+            # Evaluate the inputs that are not masked for each sequence
+            input_length = attention_mask.sum(axis=1)
+            if not prefill:
+                # When decoding, cache_ids contains a single value per sequence
+                return (input_length - 1).unsqueeze(1)
+            # When prefilling, cache_ids is an increasing range
+            cache_ids = torch.zeros_like(attention_mask)
+            for i in range(cache_n):
+                cur_length = input_length[i]
+                cache_ids[i, :cur_length] = torch.arange(cur_length)
+            return cache_ids
+        # Static batching
+        return None if prefill else torch.tensor([cache_len - 1], dtype=torch.int32)
+
+    def prepare_inputs_for_prefill(
+        self, input_ids: torch.Tensor, attention_mask: torch.Tensor, seq_ids: Optional[List[int]] = None
+    ) -> Dict[str, torch.Tensor]:
+        start_ids = self.get_start_ids(input_ids, attention_mask, seq_ids=seq_ids)
+        cache_ids = self.get_cache_ids(attention_mask, prefill=True)
+        if self.continuous_batching and torch.any(attention_mask[:, 0] == 0):
+            # Inputs are left padded: we need to invert padding as continuous batching requires right-padding
+            batch_size, seq_len = input_ids.shape
+            input_length = attention_mask.sum(axis=1)
+            new_input_ids = torch.zeros_like(input_ids)
+            for i in range(batch_size):
+                cur_length = input_length[i]
+                new_input_ids[i, :cur_length] = input_ids[i, seq_len - cur_length :]
+            input_ids = new_input_ids
+        return {
             "input_ids": input_ids,
             "cache_ids": cache_ids,
             "start_ids": start_ids,
         }
 
-        return model_inputs
+    def prepare_inputs_for_decode(
+        self,
+        input_ids: torch.Tensor,
+        attention_mask: torch.Tensor,
+        seq_ids: Optional[List[int]] = None,
+    ) -> Dict[str, torch.Tensor]:
+        start_ids = self.get_start_ids(input_ids, attention_mask, seq_ids=seq_ids)
+        cache_ids = self.get_cache_ids(attention_mask, prefill=False)
+        # Only pass the last tokens of each sample
+        input_ids = input_ids[:, -1:]
+        return {
+            "input_ids": input_ids,
+            "cache_ids": cache_ids,
+            "start_ids": start_ids,
+        }
 
     def can_generate(self) -> bool:
         """Returns True to validate the check made in `GenerationMixin.generate()`."""
@@ -775,7 +821,7 @@ def generate(
                 f"The input sequence length ({sequence_length}) exceeds the model static sequence length ({self.max_length})"
             )
         padded_input_ids = input_ids
-        padded_attention_mask = attention_mask
+        padded_attention_mask = torch.ones_like(input_ids) if attention_mask is None else attention_mask
         if batch_size > self.batch_size:
             raise ValueError(
                 f"The specified batch_size ({batch_size}) exceeds the model static batch size ({self.batch_size})"
@@ -784,18 +830,15 @@ def generate(
             logger.warning("Inputs will be padded to match the model static batch size. This will increase latency.")
             padding_shape = [self.batch_size - batch_size, sequence_length]
             padding = torch.full(padding_shape, fill_value=self.config.eos_token_id, dtype=torch.int64)
-            padded_input_ids = torch.cat([input_ids, padding])
-            if attention_mask is not None:
-                padding = torch.zeros(padding_shape, dtype=torch.int64)
-                padded_attention_mask = torch.cat([attention_mask, padding])
-        # Drop the current generation context and clear the Key/Value cache
-        self.reset_generation()
+            padded_input_ids = torch.cat([padded_input_ids, padding])
+            padding = torch.zeros(padding_shape, dtype=torch.int64)
+            padded_attention_mask = torch.cat([padded_attention_mask, padding])
 
         output_ids = self.generate_tokens(
             padded_input_ids,
             selector,
             batch_size,
-            attention_mask=padded_attention_mask,
+            padded_attention_mask,
             **model_kwargs,
         )
         return output_ids[:batch_size, :]
@@ -805,7 +848,7 @@ def generate_tokens(
         input_ids: torch.LongTensor,
         selector: TokenSelector,
         batch_size: int,
-        attention_mask: Optional[torch.Tensor] = None,
+        attention_mask: torch.Tensor,
         **model_kwargs,
     ) -> torch.LongTensor:
         r"""
@@ -831,17 +874,15 @@ def generate_tokens(
         unfinished_sequences = torch.zeros(input_ids.shape[0], dtype=torch.long, device=input_ids.device)
         unfinished_sequences[:batch_size] = 1
 
+        # Prefill and obtain the first token
+        model_inputs = self.prepare_inputs_for_prefill(input_ids, attention_mask)
+        outputs = self(
+            **model_inputs,
+            return_dict=True,
+        )
+
         # auto-regressive generation
         while True:
-            # prepare model inputs
-            model_inputs = self.prepare_inputs_for_generation(input_ids, attention_mask, **model_kwargs)
-
-            # forward pass to get next token
-            outputs = self(
-                **model_inputs,
-                return_dict=True,
-            )
-
             next_token_logits = outputs.logits[:, -1, :]
 
             next_tokens = selector.select(input_ids, next_token_logits)
@@ -851,10 +892,7 @@ def generate_tokens(
 
             # update inputs for the next step
             input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
-            if attention_mask is not None:
-                attention_mask = torch.cat(
-                    [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
-                )
+            attention_mask = torch.cat([attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1)
 
             # if eos_token was found in one sentence, set sentence to finished
             unfinished_sequences = unfinished_sequences * next_tokens.ne(selector.eos_token_id)
@@ -867,4 +905,11 @@ def generate_tokens(
             if selector.stopping_criteria(input_ids, None):
                 break
 
+            # forward pass to get next token
+            model_inputs = self.prepare_inputs_for_decode(input_ids, attention_mask)
+            outputs = self(
+                **model_inputs,
+                return_dict=True,
+            )
+
         return input_ids

optimum/neuron/modeling_decoder.py

@@ -35,6 +35,7 @@
 
 
 if is_transformers_neuronx_available():
+    from transformers_neuronx.config import ContinuousBatchingConfig, NeuronConfig
     from transformers_neuronx.module import save_split
 
 
@@ -131,16 +132,26 @@ def __init__(
 
         exporter = get_exporter(config, task)
 
-        # transformers-neuronx uses f32/f16 instead of fp32/fp16
-        auto_cast_type = auto_cast_type.replace("p", "")
+        tnx_kwargs = {
+            "batch_size": batch_size,
+            "tp_degree": num_cores,
+            # transformers-neuronx uses f32/f16 instead of fp32/fp16
+            "amp": auto_cast_type.replace("p", ""),
+        }
+        if batch_size > 1 and exporter.continuous_batching:
+            # Continuous batching is always enabled for models that support it because static batching
+            # is broken for these models:  see https://github.com/aws-neuron/transformers-neuronx/issues/79
+            tnx_kwargs["neuron_config"] = NeuronConfig(
+                continuous_batching=ContinuousBatchingConfig(batch_size_for_shared_caches=batch_size)
+            )
+            tnx_kwargs["n_positions"] = [sequence_length]
+            tnx_kwargs["context_length_estimate"] = [sequence_length]
+        else:
+            tnx_kwargs["n_positions"] = sequence_length
+
+        # Instantiate neuronx model
         checkpoint_path = checkpoint_dir.name if isinstance(checkpoint_dir, TemporaryDirectory) else checkpoint_dir
-        neuronx_model = exporter.neuronx_class.from_pretrained(
-            checkpoint_path,
-            batch_size=batch_size,
-            n_positions=sequence_length,
-            tp_degree=num_cores,
-            amp=auto_cast_type,
-        )
+        neuronx_model = exporter.neuronx_class.from_pretrained(checkpoint_path, **tnx_kwargs)
 
         if compiled_dir is not None:
             # Specify the path where compiled artifacts are stored before conversion

tests/generation/conftest.py

@@ -72,7 +72,7 @@ def export_seq2seq_model_class(request):
 @requires_neuronx
 def neuron_decoder_path(export_decoder_id):
     model = NeuronModelForCausalLM.from_pretrained(
-        export_decoder_id, export=True, batch_size=1, sequence_length=100, num_cores=2
+        export_decoder_id, export=True, batch_size=2, sequence_length=100, num_cores=2
     )
     model_dir = TemporaryDirectory()
     model_path = model_dir.name

tests/generation/test_generate.py

@@ -18,20 +18,12 @@
 import pytest
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
-from transformers.generation import StoppingCriteria
 
-from optimum.neuron import NeuronModelForCausalLM, NeuronModelForSeq2SeqLM
+from optimum.neuron import NeuronModelForSeq2SeqLM
 from optimum.neuron.utils.testing_utils import is_inferentia_test, is_trainium_test, requires_neuronx
 from optimum.neuron.utils.training_utils import patch_generation_mixin_to_general_neuron_generation_mixin
 
 
-def _test_model_generation(model, tokenizer, batch_size, input_length, **gen_kwargs):
-    input_ids = torch.ones((batch_size, input_length), dtype=torch.int64)
-    with torch.inference_mode():
-        sample_output = model.generate(input_ids, **gen_kwargs)
-        assert sample_output.shape[0] == batch_size
-
-
 def _test_model_generation_trn(model, tokenizer, batch_size, input_length, **gen_kwargs):
     import torch_xla.core.xla_model as xm
 
@@ -43,58 +35,6 @@ def _test_model_generation_trn(model, tokenizer, batch_size, input_length, **gen
     assert sample_output.shape[0] == batch_size
 
 
-@pytest.mark.parametrize(
-    "gen_kwargs",
-    [
-        {"do_sample": True},
-        {"do_sample": True, "temperature": 0.7},
-        {"do_sample": False},
-        {"do_sample": False, "repetition_penalty": 1.2},
-    ],
-    ids=["sample", "sample-with-temp", "greedy", "greedy_no-repeat"],
-)
-@is_inferentia_test
-@requires_neuronx
-def test_decoder_generation(neuron_decoder_path, gen_kwargs):
-    model = NeuronModelForCausalLM.from_pretrained(neuron_decoder_path)
-    tokenizer = AutoTokenizer.from_pretrained(neuron_decoder_path)
-    _test_model_generation(model, tokenizer, model.batch_size, 10, **gen_kwargs)
-
-
-@is_inferentia_test
-@requires_neuronx
-def test_model_generation_input_dimensions(neuron_decoder_path):
-    model = NeuronModelForCausalLM.from_pretrained(neuron_decoder_path)
-    tokenizer = AutoTokenizer.from_pretrained(neuron_decoder_path)
-    # Using valid input dimensions
-    _test_model_generation(model, tokenizer, model.batch_size, model.max_length // 2)
-    # Using an incompatible batch_size
-    with pytest.raises(ValueError, match="The specified batch_size"):
-        _test_model_generation(model, tokenizer, model.batch_size + 1, model.max_length)
-    # Using an incompatible input length
-    with pytest.raises(ValueError, match="The input sequence length"):
-        _test_model_generation(model, tokenizer, model.batch_size, input_length=model.max_length * 2)
-
-
-@is_inferentia_test
-@requires_neuronx
-def test_decoder_generation_custom_stopping_criteria():
-    model_id = "hf-internal-testing/tiny-random-gpt2"
-    model = NeuronModelForCausalLM.from_pretrained(model_id, export=True, batch_size=1)
-
-    class CustomStoppingCriteria(StoppingCriteria):
-        def __init__(self):
-            self.called = False
-
-        def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
-            self.called = True
-            return True
-
-    criteria = CustomStoppingCriteria()
-    model.generate(input_ids=torch.ones([1, 10], dtype=torch.int64), stopping_criteria=[criteria])
-    assert criteria.called, "Custom StoppingCriteria should have been called"
-
-
 @is_inferentia_test
 @requires_neuronx
 def test_seq2seq_generation_beam(neuron_seq2seq_beam_path):