feat(server): flash neoX (#133)

.github/workflows/build.yaml

@@ -20,6 +20,10 @@ on:
     branches:
       - 'main'
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   build-and-push-image:
     runs-on: ubuntu-latest

.github/workflows/tests.yaml

@@ -11,6 +11,10 @@ on:
       - "Cargo.lock"
       - "rust-toolchain.toml"
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   run_tests:
     runs-on: ubuntu-20.04

Dockerfile

@@ -43,7 +43,7 @@ ENV LANG=C.UTF-8 \
     CONDA_DEFAULT_ENV=text-generation \
     PATH=$PATH:/opt/miniconda/envs/text-generation/bin:/opt/miniconda/bin:/usr/local/cuda/bin
 
-RUN apt-get update && apt-get install -y unzip curl libssl-dev && rm -rf /var/lib/apt/lists/*
+RUN apt-get update && apt-get install -y git curl libssl-dev && rm -rf /var/lib/apt/lists/*
 
 RUN cd ~ && \
     curl -L -O https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh && \
@@ -53,10 +53,13 @@ RUN cd ~ && \
 
 WORKDIR /usr/src
 
+# Install torch
+RUN pip install torch --extra-index-url https://download.pytorch.org/whl/cu118 --no-cache-dir
+
 COPY server/Makefile server/Makefile
 
-# Install specific version of torch
+# Install specific version of flash attention
-RUN cd server && make install-torch
+RUN cd server && make install-flash-attention
 
 # Install specific version of transformers
 RUN cd server && BUILD_EXTENSIONS="True" make install-transformers

server/Makefile

@@ -1,4 +1,5 @@
 transformers_commit := 2b57aa18da658e7d2f42ef6bd5b56751af582fef
+flash_att_commit := 4d87e4d875077ad9efd25030efa4ab0ba92c19e1
 
 gen-server:
 	# Compile protos
@@ -12,13 +13,19 @@ install-transformers:
 	# Install specific version of transformers with custom cuda kernels
 	pip uninstall transformers -y || true
 	rm -rf transformers || true
-	rm -rf transformers-$(transformers_commit) || true
+	git clone https://github.com/OlivierDehaene/transformers.git
-	curl -L -O https://github.com/OlivierDehaene/transformers/archive/$(transformers_commit).zip
+	cd transformers && git checkout $(transformers_commit)
-	unzip $(transformers_commit).zip
-	rm $(transformers_commit).zip
-	mv transformers-$(transformers_commit) transformers
 	cd transformers && python setup.py install
 
+install-flash-attention:
+	# Install specific version of flash attention
+	pip install packaging
+	pip uninstall flash_attn rotary_emb dropout_layer_norm -y || true
+	rm -rf flash-attention || true
+	git clone https://github.com/HazyResearch/flash-attention.git
+	cd flash-attention && git checkout $(flash_att_commit)
+	cd flash-attention && python setup.py install && cd csrc/layer_norm && python setup.py install && cd ../rotary && python setup.py install
+
 install-torch:
 	# Install specific version of torch
 	pip install torch --extra-index-url https://download.pytorch.org/whl/cu118 --no-cache-dir

server/text_generation_server/models/__init__.py

@@ -1,5 +1,7 @@
+import os
 import torch
 
+from loguru import logger
 from transformers import AutoConfig
 from typing import Optional
 
@@ -12,6 +14,14 @@ from text_generation_server.models.santacoder import SantaCoder
 from text_generation_server.models.gpt_neox import GPTNeoxSharded
 from text_generation_server.models.t5 import T5Sharded
 
+try:
+    from text_generation_server.models.flash_neox import FlashNeoX, FlashNeoXSharded
+    FLASH_NEOX = torch.cuda.is_available() and int(os.environ.get("FLASH_NEOX", 0)) == 1
+except ImportError:
+    if int(os.environ.get("FLASH_NEOX", 0)) == 1:
+        logger.exception("Could not import FlashNeoX")
+    FLASH_NEOX = False
+
 __all__ = [
     "Model",
     "BLOOM",
@@ -26,6 +36,10 @@ __all__ = [
     "get_model",
 ]
 
+if FLASH_NEOX:
+    __all__.append(FlashNeoX)
+    __all__.append(FlashNeoXSharded)
+
 # The flag below controls whether to allow TF32 on matmul. This flag defaults to False
 # in PyTorch 1.12 and later.
 torch.backends.cuda.matmul.allow_tf32 = True
@@ -59,9 +73,11 @@ def get_model(
 
     if config.model_type == "gpt_neox":
         if sharded:
-            return GPTNeoxSharded(model_id, revision, quantize=quantize)
+            neox_cls = FlashNeoXSharded if FLASH_NEOX else GPTNeoxSharded
+            return neox_cls(model_id, revision, quantize=quantize)
         else:
-            return CausalLM(model_id, revision, quantize=quantize)
+            neox_cls = FlashNeoX if FLASH_NEOX else CausalLM
+            return neox_cls(model_id, revision, quantize=quantize)
 
     if config.model_type == "t5":
         if sharded:

server/text_generation_server/models/causal_lm.py

@@ -64,7 +64,6 @@ class CausalLMBatch(Batch):
         inputs = []
         next_token_choosers = []
         stopping_criterias = []
-        input_lengths = []
 
         # Parse batch
         padding_right_offset = 0

server/text_generation_server/models/flash_neox.py

@@ -0,0 +1,601 @@
+import torch
+import torch.distributed
+
+from accelerate import init_empty_weights
+from dataclasses import dataclass
+from opentelemetry import trace
+from safetensors import safe_open
+from transformers import AutoTokenizer, PreTrainedTokenizerBase, AutoConfig
+from typing import Optional, Tuple, List, Type, Union
+
+from text_generation_server.models import Model
+from text_generation_server.models.flash_neox_modeling import (
+    FlashGPTNeoXForCausalLM,
+    TensorParallelEmbedding,
+    TensorParallelRowLinear,
+    TensorParallelColumnLinear,
+)
+from text_generation_server.models.types import (
+    Batch,
+    PrefillTokens,
+    Generation,
+    GeneratedText,
+)
+from text_generation_server.pb import generate_pb2
+from text_generation_server.utils import (
+    NextTokenChooser,
+    StoppingCriteria,
+    Sampling,
+    initialize_torch_distributed,
+    weight_files,
+)
+
+tracer = trace.get_tracer(__name__)
+
+
+@dataclass
+class FlashNeoXBatch(Batch):
+    batch_id: int
+    requests: List[generate_pb2.Request]
+
+    # Decoder values
+    input_ids: torch.Tensor
+    position_ids: torch.Tensor
+    # cumulative sequence lengths
+    cu_seqlens: torch.Tensor
+    max_seqlen: int
+    past_key_values: Optional[torch.Tensor]
+
+    # All tokens
+    all_input_ids: List[List[int]]
+
+    # Lengths of all generations present in the batch
+    input_lengths: List[int]
+
+    # Generation helpers
+    next_token_choosers: List[NextTokenChooser]
+    stopping_criterias: List[StoppingCriteria]
+
+    def to_pb(self) -> generate_pb2.Batch:
+        return generate_pb2.Batch(
+            id=self.batch_id, requests=self.requests, size=len(self)
+        )
+
+    @classmethod
+    def from_pb(
+        cls,
+        pb: generate_pb2.Batch,
+        tokenizer: PreTrainedTokenizerBase,
+        device: torch.device,
+    ) -> "CausalLMBatch":
+        input_ids = []
+        position_ids = []
+        cu_seqlens = [0]
+        max_seqlen = 0
+
+        input_lengths = []
+        all_input_ids = []
+
+        next_token_choosers = []
+        stopping_criterias = []
+
+        # Cumulative length
+        cumulative_length = 0
+
+        # Parse batch
+        for r in pb.requests:
+            tokenized_input = tokenizer(r.inputs, return_tensors="pt")[
+                "input_ids"
+            ].squeeze(0)
+            input_ids.append(tokenized_input)
+            all_input_ids.append(tokenized_input.tolist())
+
+            input_length = len(tokenized_input)
+            max_seqlen = max(max_seqlen, input_length)
+            input_lengths.append(input_length)
+
+            # Position ids
+            position_ids.append(torch.arange(0, input_length, dtype=torch.int32))
+
+            # Add cumulative lengths of all previous inputs
+            cu_seqlens.append(cumulative_length + input_length)
+
+            next_token_choosers.append(NextTokenChooser.from_pb(r.parameters, device))
+            stopping_criterias.append(
+                StoppingCriteria.from_pb(r.stopping_parameters, tokenizer)
+            )
+
+            # Update
+            cumulative_length += input_length
+
+        input_ids = torch.concat(input_ids).unsqueeze(1)
+        position_ids = torch.concat(position_ids)
+        cu_seqlens = torch.tensor(cu_seqlens, dtype=torch.int32)
+
+        return cls(
+            batch_id=pb.id,
+            requests=pb.requests,
+            input_ids=input_ids,
+            position_ids=position_ids,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+            past_key_values=None,
+            input_lengths=input_lengths,
+            all_input_ids=all_input_ids,
+            next_token_choosers=next_token_choosers,
+            stopping_criterias=stopping_criterias,
+        )
+
+    @classmethod
+    @tracer.start_as_current_span("concatenate")
+    def concatenate(cls, batches: List["CausalLMBatch"]) -> "CausalLMBatch":
+        # Batch attributes
+        requests = []
+        input_lengths = []
+        all_input_ids = []
+        next_token_choosers = []
+        stopping_criterias = []
+
+        # Batch tensors
+        input_ids = []
+        position_ids = []
+        cu_seqlens = [torch.tensor([0], dtype=torch.int32)]
+        max_seqlen = 0
+        past_key_values = []
+
+        # Cumulative length
+        cumulative_length = torch.tensor(0)
+
+        for i, batch in enumerate(batches):
+            requests.extend(batch.requests)
+            input_lengths.extend(batch.input_lengths)
+            all_input_ids.extend(batch.all_input_ids)
+            next_token_choosers.extend(batch.next_token_choosers)
+            stopping_criterias.extend(batch.stopping_criterias)
+
+            # Add cumulative lengths of all previous inputs
+            cu_seqlens.append(batch.cu_seqlens[1:] + cumulative_length)
+
+            input_ids.append(batch.input_ids)
+            position_ids.append(batch.position_ids)
+            past_key_values.append(batch.past_key_values)
+
+            max_seqlen = max(max_seqlen, batch.max_seqlen)
+
+            # Update
+            cumulative_length += batch.cu_seqlens[-1]
+
+        input_ids = torch.concat(input_ids)
+        position_ids = torch.concat(position_ids)
+        # Concat on dim=1 as first dim represents the model layers
+        past_key_values = torch.concat(past_key_values, dim=1)
+        cu_seqlens = torch.concat(cu_seqlens)
+
+        return FlashNeoXBatch(
+            batch_id=batches[0].batch_id,
+            requests=requests,
+            input_ids=input_ids,
+            position_ids=position_ids,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+            past_key_values=past_key_values,
+            input_lengths=input_lengths,
+            all_input_ids=all_input_ids,
+            next_token_choosers=next_token_choosers,
+            stopping_criterias=stopping_criterias,
+        )
+
+    def __len__(self):
+        return len(self.requests)
+
+
+class FlashNeoX(Model):
+    def __init__(self, model_id: str, revision: Optional[str] = None, quantize=False):
+        if torch.cuda.is_available():
+            device = torch.device("cuda")
+            dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+        else:
+            raise NotImplementedError("FlashNeoX is only available on GPU")
+
+        if quantize:
+            raise NotImplementedError("FlashNeoX does not support quantization")
+
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_id, revision=revision, padding_side="left"
+        )
+        self.model = (
+            FlashGPTNeoXForCausalLM.from_pretrained(
+                model_id,
+                revision=revision,
+                torch_dtype=dtype,
+            )
+            .eval()
+            .cuda()
+        )
+        tokenizer.pad_token_id = (
+            self.model.config.pad_token_id
+            if self.model.config.pad_token_id is not None
+            else self.model.config.eos_token_id
+        )
+
+        super(FlashNeoX, self).__init__(
+            tokenizer=tokenizer,
+            device=device,
+        )
+
+    @property
+    def batch_type(self) -> Type[FlashNeoXBatch]:
+        return FlashNeoXBatch
+
+    def decode(self, generated_ids: Union[torch.Tensor, List[int]]) -> str:
+        return self.tokenizer.decode(
+            generated_ids, skip_special_tokens=True, cleanup_tokenization_spaces=False
+        )
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        max_s: int,
+        past_key_values: Optional = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Model Forward
+        return self.model.forward(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            cu_seqlens=cu_seqlens,
+            max_s=max_s,
+            past_key_values=past_key_values,
+        )
+
+    @tracer.start_as_current_span("generate_token")
+    def generate_token(
+        self, batch: FlashNeoXBatch
+    ) -> Tuple[List[Generation], Optional[FlashNeoXBatch]]:
+        # Better to send to device here to avoid device issues in concatenate
+        position_ids = batch.position_ids.to(self.device, non_blocking=True)
+        cu_seqlens = batch.cu_seqlens.to(self.device, non_blocking=True)
+        input_ids = batch.input_ids.squeeze(1).to(self.device)
+
+        out, present = self.forward(
+            input_ids,
+            position_ids,
+            cu_seqlens,
+            batch.max_seqlen,
+            batch.past_key_values,
+        )
+
+        # List of indices to cache
+        next_batch_keep_indices = []
+
+        # New values for next forward
+        next_batch_input_ids = []
+        next_batch_position_ids = []
+        next_batch_cu_seqlens = [0]
+        next_batch_max_seqlen = 0
+        next_batch_past_key_values = []
+        next_batch_input_lengths = []
+        next_batch_all_input_ids = []
+
+        # Cumulative length
+        cumulative_length = 0
+
+        # Results
+        generations: List[Generation] = []
+
+        # Zipped iterator
+        iterator = zip(
+            batch.requests,
+            batch.input_lengths,
+            batch.next_token_choosers,
+            batch.stopping_criterias,
+            batch.all_input_ids,
+        )
+
+        # For each member of the batch
+        for i, (
+            request,
+            input_length,
+            next_token_chooser,
+            stopping_criteria,
+            all_input_ids,
+        ) in enumerate(iterator):
+            # Indexing metadata
+            start_index = cumulative_length
+            end_index = cumulative_length + input_length
+
+            if batch.past_key_values is None:
+                # Prefill mode
+                # out is of shape [cumulative_sequence_lengths, vocab_size]
+                logits = out[start_index:end_index]
+            else:
+                # Decode mode
+                # out is of shape [batch_size, vocab_size]
+                logits = out[i].unsqueeze(0)
+
+            # Select next token
+            next_token_id, logprobs = next_token_chooser(all_input_ids, logits)
+            # Copy to cpu to avoid other copies when indexing and calling .item()
+            next_token_id = next_token_id.to("cpu", non_blocking=True)
+            logprobs = logprobs.to("cpu")
+
+            next_token_id_squeezed = next_token_id.squeeze()
+            next_token_id_item = next_token_id_squeezed.item()
+
+            # Append next token to all tokens
+            all_input_ids.append(next_token_id_item)
+            new_input_length = input_length + 1
+
+            # Generated token
+            next_token_logprob = logprobs[-1, next_token_id]
+            next_token_text = self.decode_token(
+                next_token_id_item,
+            )
+
+            # Evaluate stopping criteria
+            stop, reason = stopping_criteria(
+                next_token_id_item,
+                next_token_text,
+            )
+
+            if stop:
+                # Decode generated tokens
+                output_text = self.decode(
+                    all_input_ids[-stopping_criteria.current_tokens :]
+                )
+                # Get seed
+                if isinstance(next_token_chooser.choice, Sampling):
+                    seed = next_token_chooser.choice.seed
+                else:
+                    seed = None
+
+                generated_text = GeneratedText(
+                    output_text, stopping_criteria.current_tokens, reason, seed
+                )
+            else:
+                # Keep request in the batch
+                next_batch_keep_indices.append(i)
+                generated_text = None
+
+                # Get sequence present
+                seq_present = present[:, start_index:end_index]
+                # Pad it for next iter attention
+                past = torch.nn.functional.pad(seq_present, (0, 0, 0, 0, 0, 0, 0, 1))
+                next_batch_past_key_values.append(past)
+
+                next_batch_input_ids.append(next_token_id)
+                next_batch_position_ids.append(input_length)
+                # Cumulative sum
+                next_batch_cu_seqlens.append(
+                    next_batch_cu_seqlens[-1] + new_input_length
+                )
+                next_batch_input_lengths.append(new_input_length)
+                next_batch_all_input_ids.append(all_input_ids)
+                next_batch_max_seqlen = max(next_batch_max_seqlen, new_input_length)
+
+            # Prefill
+            if stopping_criteria.current_tokens == 1:
+                # Remove generated token to only have prefill and add nan for first prompt token
+                prefill_logprobs = [float("nan")] + logprobs.gather(
+                    1, torch.tensor(all_input_ids[1:]).unsqueeze(1)
+                ).squeeze(1)[:-1].tolist()
+                prefill_token_ids = all_input_ids[:-1]
+                prefill_texts = self.tokenizer.batch_decode(
+                    prefill_token_ids,
+                    clean_up_tokenization_spaces=False,
+                    skip_special_tokens=False,
+                )
+                prefill_tokens = PrefillTokens(
+                    prefill_token_ids, prefill_logprobs, prefill_texts
+                )
+            else:
+                prefill_tokens = None
+
+            generation = Generation(
+                request.id,
+                prefill_tokens,
+                next_token_id_item,
+                next_token_logprob,
+                next_token_text,
+                next_token_id_item in self.all_special_ids,
+                generated_text,
+            )
+
+            generations.append(generation)
+            cumulative_length += input_length
+
+        # We finished all generations in the batch; there is no next batch
+        if not next_batch_keep_indices:
+            return generations, None
+
+        # If we finished at least one generation, we need to evict the indices of the generations that finished
+        # from the values of the next batch
+        if len(next_batch_keep_indices) != len(batch):
+            # Apply indices to requests, token_choosers and stopping_criterias that need to be cached
+            next_batch_requests = [batch.requests[i] for i in next_batch_keep_indices]
+            next_batch_next_token_choosers = [
+                batch.next_token_choosers[i] for i in next_batch_keep_indices
+            ]
+            next_batch_stopping_criterias = [
+                batch.stopping_criterias[i] for i in next_batch_keep_indices
+            ]
+        else:
+            next_batch_requests = batch.requests
+            next_batch_next_token_choosers = batch.next_token_choosers
+            next_batch_stopping_criterias = batch.stopping_criterias
+
+        # Create final next batch tensors
+        next_batch_position_ids = torch.tensor(
+            next_batch_position_ids, dtype=torch.int32
+        )
+        next_batch_cu_seqlens = torch.tensor(next_batch_cu_seqlens, dtype=torch.int32)
+        if len(next_batch_keep_indices) > 1:
+            next_batch_input_ids = torch.concat(next_batch_input_ids)
+            next_batch_past_key_values = torch.concat(next_batch_past_key_values, dim=1)
+        else:
+            next_batch_input_ids = next_batch_input_ids[0]
+            next_batch_past_key_values = next_batch_past_key_values[0]
+
+        next_batch = FlashNeoXBatch(
+            batch_id=batch.batch_id,
+            requests=next_batch_requests,
+            input_ids=next_batch_input_ids,
+            position_ids=next_batch_position_ids,
+            cu_seqlens=next_batch_cu_seqlens,
+            max_seqlen=next_batch_max_seqlen,
+            past_key_values=next_batch_past_key_values,
+            input_lengths=next_batch_input_lengths,
+            all_input_ids=next_batch_all_input_ids,
+            next_token_choosers=next_batch_next_token_choosers,
+            stopping_criterias=next_batch_stopping_criterias,
+        )
+        return generations, next_batch
+
+
+class FlashNeoXSharded(FlashNeoX):
+    def __init__(
+        self, model_id: str, revision: Optional[str] = None, quantize: bool = False
+    ):
+        self.process_group, self.rank, self.world_size = initialize_torch_distributed()
+        self.master = self.rank == 0
+        if torch.cuda.is_available():
+            device = torch.device(f"cuda:{self.rank}")
+            dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+        else:
+            raise NotImplementedError("FlashNeoX is only available on GPU")
+
+        if quantize:
+            raise NotImplementedError("FlashNeoX does not support quantization")
+
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_id, revision=revision, padding_side="left"
+        )
+
+        config = AutoConfig.from_pretrained(
+            model_id, revision=revision, tp_parallel=True
+        )
+
+        torch.distributed.barrier(group=self.process_group)
+        filenames = weight_files(model_id, revision=revision, extension=".safetensors")
+
+        with init_empty_weights():
+            model = FlashGPTNeoXForCausalLM(config)
+
+        torch.distributed.barrier(group=self.process_group)
+        self.load_weights(
+            model,
+            filenames,
+            quantize=quantize,
+            device=device,
+            rank=self.rank,
+            world_size=self.world_size,
+        )
+        self.model = model.eval().to(dtype)
+        torch.distributed.barrier(group=self.process_group)
+        super(FlashNeoX, self).__init__(
+            tokenizer=tokenizer,
+            device=device,
+        )
+
+    @staticmethod
+    def load_weights(
+        model,
+        filenames: List[str],
+        quantize: bool,
+        device: torch.device,
+        rank: int,
+        world_size: int,
+    ):
+        parameters = dict(model.named_parameters())
+        for file in filenames:
+            with safe_open(
+                file, framework="pt", device=str(device) if not quantize else "cpu"
+            ) as f:
+                for name in f.keys():
+                    module_name, param_name = name.rsplit(".", 1)
+                    module = model.get_submodule(module_name)
+
+                    current_parameter_tensor = parameters.get(name, None)
+
+                    slice_ = f.get_slice(name)
+
+                    if isinstance(module, TensorParallelColumnLinear):
+                        size = slice_.get_shape()[0]
+                        block_size = size // world_size
+                        start = rank * block_size
+                        stop = (rank + 1) * block_size
+                        tensor = slice_[start:stop]
+                    elif isinstance(module, TensorParallelRowLinear):
+                        if param_name == "weight":
+                            size = slice_.get_shape()[1]
+                            block_size = size // world_size
+                            start = rank * block_size
+                            stop = (rank + 1) * block_size
+                            tensor = slice_[:, start:stop]
+                        else:
+                            tensor = slice_[:]
+                            # XXX: Hack for Rowlinear to add the bias only once.
+                            if rank != 0:
+                                tensor = torch.zeros_like(tensor)
+                    elif isinstance(module, TensorParallelEmbedding):
+                        size = slice_.get_shape()[0]
+                        block_size = size // world_size
+                        start = rank * block_size
+                        stop = (rank + 1) * block_size
+                        tensor = slice_[start:stop]
+                    elif name == "embed_out.weight" and model.gpt_neox.tp_embeddings:
+                        size = slice_.get_shape()[0]
+                        block_size = size // world_size
+                        start = rank * block_size
+                        stop = (rank + 1) * block_size
+                        tensor = slice_[start:stop]
+                    else:
+                        try:
+                            tensor = slice_[:]
+                        except:
+                            tensor = f.get_tensor(name)
+
+                    if (
+                        current_parameter_tensor is not None
+                        and current_parameter_tensor.shape != tensor.shape
+                    ):
+                        raise ValueError(
+                            f"Name {name} -- Current {current_parameter_tensor.shape} and got {tensor.shape}"
+                        )
+
+                    tensor = tensor.contiguous()
+
+                    if current_parameter_tensor is not None:
+                        module._parameters[param_name] = tensor
+                    else:
+                        module._buffers[param_name] = tensor
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        max_s: int,
+        past_key_values: Optional = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if self.model.gpt_neox.tp_embeddings:
+            logits, present = self.model.forward(
+                input_ids=input_ids,
+                position_ids=position_ids,
+                cu_seqlens=cu_seqlens,
+                max_s=max_s,
+                past_key_values=past_key_values,
+            )
+
+            # Logits are sharded, so we need to gather them
+            world_logits = [torch.empty_like(logits) for _ in range(self.world_size)]
+            torch.distributed.all_gather(world_logits, logits, group=self.process_group)
+            world_logits = torch.cat(world_logits, dim=1)
+
+            return world_logits, present
+        # While the model itself is sharded, the embeddings might not as they might not be dividable by num-shard
+        else:
+            return super(FlashNeoXSharded, self).forward(
+                input_ids, position_ids, cu_seqlens, max_s, past_key_values
+            )

server/text_generation_server/models/flash_neox_modeling.py

@@ -0,0 +1,637 @@
+import torch
+import torch.distributed
+
+import torch.nn.functional as F
+
+from torch import nn
+from transformers.activations import ACT2FN
+from transformers.modeling_utils import PreTrainedModel
+from transformers.models.gpt_neox import GPTNeoXConfig
+
+# Flash attention imports
+import rotary_emb
+import flash_attn_cuda
+import dropout_layer_norm
+
+from flash_attn.layers.rotary import RotaryEmbedding
+
+
+class TensorParallelColumnLinear(nn.Linear):
+    def __init__(
+        self,
+        in_features,
+        out_features,
+        process_group: torch.distributed.ProcessGroup,
+        bias=True,
+        device=None,
+        dtype=None,
+    ):
+        self.process_group = process_group
+        self.tp_world_size = process_group.size()
+        assert out_features % self.tp_world_size == 0
+        out_features = out_features // self.tp_world_size
+
+        super().__init__(
+            in_features=in_features,
+            out_features=out_features,
+            bias=bias,
+            device=device,
+            dtype=dtype,
+        )
+
+    @staticmethod
+    def linear(input, weight, bias):
+        return F.linear(input, weight, bias)
+
+    def forward(self, input):
+        return self.linear(input, self.weight, self.bias)
+
+
+class TensorParallelRowLinear(nn.Linear):
+    def __init__(
+        self,
+        in_features,
+        out_features,
+        process_group: torch.distributed.ProcessGroup,
+        bias=True,
+        device=None,
+        dtype=None,
+    ):
+        self.process_group = process_group
+        self.tp_world_size = process_group.size()
+        assert in_features % self.tp_world_size == 0
+        in_features = in_features // self.tp_world_size
+
+        super().__init__(
+            in_features=in_features,
+            out_features=out_features,
+            bias=bias,
+            device=device,
+            dtype=dtype,
+        )
+
+    @staticmethod
+    def linear(input, weight, bias):
+        return F.linear(input, weight, bias)
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        out = self.linear(input, self.weight, self.bias)
+        torch.distributed.all_reduce(out, group=self.process_group)
+
+        return out
+
+
+class TensorParallelEmbedding(nn.Embedding):
+    def __init__(
+        self,
+        num_embeddings,
+        embedding_dim,
+        process_group: torch.distributed.ProcessGroup,
+        padding_idx=None,
+        max_norm=None,
+        norm_type=2.0,
+        scale_grad_by_freq=False,
+        sparse=False,
+        _weight=None,
+        device=None,
+        dtype=None,
+    ):
+        self.process_group = process_group
+        self.tp_rank = process_group.rank()
+        self.tp_world_size = process_group.size()
+
+        self.original_num_embeddings = num_embeddings
+
+        assert num_embeddings % self.tp_world_size == 0
+        block_size = num_embeddings // self.tp_world_size
+        # inputs in `[min_id, max_id[` are handled by `self` to get embeddings
+        self.min_id = self.tp_rank * block_size
+        self.max_id = (self.tp_rank + 1) * block_size
+
+        super().__init__(
+            block_size,
+            embedding_dim,
+            padding_idx=padding_idx,
+            max_norm=max_norm,
+            norm_type=norm_type,
+            scale_grad_by_freq=scale_grad_by_freq,
+            sparse=sparse,
+            _weight=_weight,
+            device=device,
+            dtype=dtype,
+        )
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        # Sanity check
+        if torch.any(
+            torch.logical_or(0 > input, input >= self.original_num_embeddings)
+        ):
+            raise IndexError(
+                f"Input is required to be in [0, {self.original_num_embeddings}[, got min: {torch.min(input)} and max: {torch.max(input)}"
+            )
+
+        # `0` if input is in the correct interval, else `1`
+        input_mask = torch.logical_or(self.min_id > input, input >= self.max_id)
+        # translate for [0, self.max_id - self.min_id[
+        input = input - self.min_id
+        # default all out of bounds values to `0`
+        input[input_mask] = 0
+        out = super().forward(input)
+        out[input_mask] = 0.0
+        torch.distributed.all_reduce(out, group=self.process_group)
+        return out
+
+
+class PositionRotaryEmbedding(RotaryEmbedding):
+    def _update_cos_sin_cache(self, dtype, device, seqlen):
+        # Reset the tables if the sequence length has changed,
+        # or if we're on a new device (possibly due to tracing for instance)
+        if (
+            seqlen > self._seq_len_cached
+            or self._cos_cached.device != device
+            or self._cos_cached.dtype != dtype
+        ):
+            self._seq_len_cached = seqlen
+            t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
+            # Don't do einsum, it converts fp32 to fp16
+            # freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            freqs = torch.outer(t, self.inv_freq.to(device=t.device))
+            self._cos_cached = torch.cos(freqs).to(dtype)
+            self._sin_cached = torch.sin(freqs).to(dtype)
+
+    def get_cos_sin(self, position_ids: torch.Tensor, max_s: int, dtype: torch.dtype):
+        """
+        Return cos and sin for the asked position ids
+        """
+
+        self._update_cos_sin_cache(dtype, position_ids.device, max_s)
+
+        cos = torch.index_select(self._cos_cached, 0, position_ids)
+        sin = torch.index_select(self._sin_cached, 0, position_ids)
+        return cos.unsqueeze(1), sin.unsqueeze(1)
+
+    def forward(self, qkv: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor):
+        rotary_dim = cos.shape[-1]
+        q1 = qkv[:, 0, :, :rotary_dim]
+        q2 = qkv[:, 0, :, rotary_dim : 2 * rotary_dim]
+        k1 = qkv[:, 1, :, :rotary_dim]
+        k2 = qkv[:, 1, :, rotary_dim : 2 * rotary_dim]
+
+        rotary_emb.apply_rotary(q1, q2, cos, sin, q1, q2, False)
+        rotary_emb.apply_rotary(k1, k2, cos, sin, k1, k2, False)
+        return qkv
+
+
+class FlashNeoxAttention(torch.nn.Module):
+    def __init__(
+        self, num_heads, hidden_size, rotary_pct, rotary_emb_base, process_group=None
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.head_size = hidden_size // num_heads
+
+        rotary_ndims = int(self.head_size * rotary_pct)
+        self.rotary_emb = PositionRotaryEmbedding(rotary_ndims, base=rotary_emb_base)
+        self.softmax_scale = self.head_size ** (-0.5)
+
+        if process_group is None:
+            self.query_key_value = nn.Linear(hidden_size, 3 * hidden_size)
+            self.dense = nn.Linear(hidden_size, hidden_size)
+        else:
+            self.num_heads = self.num_heads // process_group.size()
+            self.query_key_value = TensorParallelColumnLinear(
+                hidden_size,
+                3 * hidden_size,
+                process_group=process_group,
+            )
+            self.dense = TensorParallelRowLinear(
+                hidden_size,
+                hidden_size,
+                process_group=process_group,
+            )
+        self.swap_dims = True
+
+    # TODO: remove and swap dims when loading weights
+    def _swap_dims(self):
+        """Swap dims for the first inference to avoid an additional permute"""
+        self.query_key_value.weight = torch.nn.Parameter(
+            self.query_key_value.weight.view(
+                self.num_heads, 3, self.head_size, self.hidden_size
+            )
+            .permute(1, 0, 2, 3)
+            .reshape(-1, self.hidden_size)
+        )
+        self.query_key_value.bias = torch.nn.Parameter(
+            self.query_key_value.bias.view(self.num_heads, 3, self.head_size)
+            .permute(1, 0, 2)
+            .reshape(-1)
+        )
+        self.swap_dims = False
+
+    def forward(
+        self,
+        hidden_states,
+        cos,
+        sin,
+        cu_seqlens,
+        max_s,
+        layer_past,
+        layer_past_present_indices,
+        cu_seqlens_q,
+    ):
+        if self.swap_dims:
+            self._swap_dims()
+
+        qkv = self.query_key_value(hidden_states)
+        qkv = qkv.view(-1, 3, self.num_heads, self.head_size)
+        qkv_rot = self.rotary_emb(qkv, cos, sin)
+
+        # Prefill
+        if layer_past_present_indices is None:
+            # Copy to layer past
+            layer_past[...] = qkv_rot[:, 1:]
+
+            # output
+            attn_output = torch.empty_like(qkv[:, 0])
+            # flash attention
+            flash_attn_cuda.fwd(
+                qkv[:, 0],
+                qkv[:, 1],
+                qkv[:, 2],
+                attn_output,
+                cu_seqlens,
+                cu_seqlens,
+                max_s,
+                max_s,
+                0.0,
+                self.softmax_scale,
+                False,
+                True,
+                False,
+                0,
+                None,
+            )
+        # Decode
+        else:
+            query = qkv_rot[:, 0]
+            # Add present to the layer_past tensor at the correct indices
+            layer_past[layer_past_present_indices] = qkv_rot[:, 1:]
+
+            # output
+            attn_output = torch.empty_like(query)
+            # flash attention
+            flash_attn_cuda.fwd(
+                query,
+                layer_past[:, 0],
+                layer_past[:, 1],
+                attn_output,
+                cu_seqlens_q,
+                cu_seqlens,
+                1,
+                max_s,
+                0.0,
+                self.softmax_scale,
+                False,
+                False,
+                False,
+                0,
+                None,
+            )
+
+        return self.dense(attn_output.view(-1, self.num_heads * self.head_size))
+
+
+class FlashMLP(nn.Module):
+    def __init__(self, act, hidden_size, intermediate_size, process_group=None):
+        super().__init__()
+        self.act = (
+            ACT2FN[act]
+            if "gelu" not in act
+            else lambda x: torch.nn.functional.gelu(x, approximate="tanh")
+        )
+
+        if process_group is None:
+            self.dense_h_to_4h = nn.Linear(hidden_size, intermediate_size)
+            self.dense_4h_to_h = nn.Linear(intermediate_size, hidden_size)
+        else:
+            self.dense_h_to_4h = TensorParallelColumnLinear(
+                hidden_size,
+                intermediate_size,
+                process_group=process_group,
+            )
+            self.dense_4h_to_h = TensorParallelRowLinear(
+                intermediate_size,
+                hidden_size,
+                process_group=process_group,
+            )
+        self.heuristic = "auto"
+        self.process_group = process_group
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense_h_to_4h(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.dense_4h_to_h(hidden_states)
+        return hidden_states
+
+
+class FlashNeoXLayer(nn.Module):
+    def __init__(
+        self,
+        num_heads,
+        act,
+        hidden_size,
+        intermediate_size,
+        rotary_pct,
+        rotary_emb_base,
+        layer_norm_eps,
+        use_parallel_residual,
+        process_group=None,
+    ):
+        super().__init__()
+        self.use_parallel_residual = use_parallel_residual
+        self.input_layernorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)
+        self.post_attention_layernorm = nn.LayerNorm(hidden_size, eps=layer_norm_eps)
+        self.attention = FlashNeoxAttention(
+            num_heads, hidden_size, rotary_pct, rotary_emb_base, process_group
+        )
+        self.mlp = FlashMLP(act, hidden_size, intermediate_size, process_group)
+
+    def forward(
+        self,
+        hidden_states,
+        residual,
+        cos,
+        sin,
+        cu_seqlens,
+        max_s,
+        layer_past,
+        layer_past_present_indices,
+        cu_seqlens_q,
+    ):
+        if self.use_parallel_residual:
+            # faster input layer norm
+            ln1_hidden_states, *rest = dropout_layer_norm.dropout_add_ln_fwd(
+                hidden_states,
+                None,
+                self.input_layernorm.weight,
+                self.input_layernorm.bias,
+                None,
+                None,
+                None,
+                None,
+                0.0,
+                self.input_layernorm.eps,
+                1.0,
+                0,
+                None,
+                False,
+                False,
+            )
+
+            attn_output = self.attention(
+                ln1_hidden_states,
+                cos,
+                sin,
+                cu_seqlens,
+                max_s,
+                layer_past,
+                layer_past_present_indices,
+                cu_seqlens_q,
+            )
+
+            # faster post attention layer norm
+            ln2_hidden_states, *rest = dropout_layer_norm.dropout_add_ln_fwd(
+                hidden_states,
+                None,
+                self.post_attention_layernorm.weight,
+                self.post_attention_layernorm.bias,
+                None,
+                None,
+                None,
+                None,
+                0.0,
+                self.post_attention_layernorm.eps,
+                1.0,
+                0,
+                None,
+                False,
+                False,
+            )
+
+            mlp_output = self.mlp(ln2_hidden_states)
+            return mlp_output + attn_output + hidden_states, None
+        else:
+            # faster input layer norm
+            hidden_states, residual, *rest = dropout_layer_norm.dropout_add_ln_fwd(
+                hidden_states,
+                residual,
+                self.input_layernorm.weight,
+                self.input_layernorm.bias,
+                None,
+                None,
+                None,
+                None,
+                0.0,
+                self.input_layernorm.eps,
+                1.0,
+                0,
+                None,
+                False,
+                False,
+            )
+
+            hidden_states = self.attention(
+                hidden_states,
+                cos,
+                sin,
+                cu_seqlens,
+                max_s,
+                layer_past,
+                layer_past_present_indices,
+                cu_seqlens_q,
+            )
+
+            # faster post attention layer norm
+            hidden_states, residual, *rest = dropout_layer_norm.dropout_add_ln_fwd(
+                hidden_states,
+                residual,
+                self.post_attention_layernorm.weight,
+                self.post_attention_layernorm.bias,
+                None,
+                None,
+                None,
+                None,
+                0.0,
+                self.post_attention_layernorm.eps,
+                1.0,
+                0,
+                None,
+                False,
+                False,
+            )
+
+            mlp_output = self.mlp(hidden_states)
+
+            return mlp_output, residual
+
+
+class FlashGPTNeoXPreTrainedModel(PreTrainedModel):
+    config_class = GPTNeoXConfig
+    base_model_prefix = "gpt_neox"
+    supports_gradient_checkpointing = False
+    _no_split_modules = None
+
+
+class FlashGPTNeoXModel(FlashGPTNeoXPreTrainedModel):
+    def __init__(self, config, process_group=None):
+        super().__init__(config)
+        self.config = config
+
+        self.tp_embeddings = False
+        if process_group is not None:
+            self.tp_rank = process_group.rank()
+            self.tp_world_size = process_group.size()
+            if config.vocab_size % self.tp_world_size == 0:
+                self.tp_embeddings = True
+
+        if self.tp_embeddings:
+            self.embed_in = TensorParallelEmbedding(
+                config.vocab_size, config.hidden_size, process_group=process_group
+            )
+        else:
+            self.embed_in = nn.Embedding(config.vocab_size, config.hidden_size)
+
+        self.layers = nn.ModuleList(
+            [
+                FlashNeoXLayer(
+                    config.num_attention_heads,
+                    config.hidden_act,
+                    config.hidden_size,
+                    config.intermediate_size,
+                    config.rotary_pct,
+                    config.rotary_emb_base,
+                    config.layer_norm_eps,
+                    config.use_parallel_residual,
+                    process_group,
+                )
+                for _ in range(config.num_hidden_layers)
+            ]
+        )
+        self.final_layer_norm = nn.LayerNorm(
+            config.hidden_size, eps=config.layer_norm_eps
+        )
+
+        self.gradient_checkpointing = False
+
+        self.head_size = self.layers[0].attention.head_size
+        self.num_heads = self.layers[0].attention.num_heads
+
+    def forward(
+        self,
+        input_ids,
+        position_ids,
+        cu_seqlens,
+        max_s,
+        past_key_values=None,
+    ):
+        hidden_states = self.embed_in(input_ids)
+
+        # Prefill
+        if past_key_values is None:
+            # Create past tensor
+            past_key_values = hidden_states.new_empty(
+                (
+                    len(self.layers),
+                    len(hidden_states),
+                    2,
+                    self.num_heads,
+                    self.head_size,
+                )
+            )
+            layer_past_present_indices = None
+            cu_seqlens_q = None
+        # Decode
+        else:
+            # Create indices from cumulative sequence lengths
+            layer_past_present_indices = cu_seqlens[1:] - 1
+            cu_seqlens_q = torch.arange(
+                len(cu_seqlens), dtype=torch.int32, device=hidden_states.device
+            )
+
+        # Get rotary cos and sin for this forward
+        # Avoid to index in each layer
+        cos, sin = self.layers[0].attention.rotary_emb.get_cos_sin(
+            position_ids, max_s, hidden_states.dtype
+        )
+
+        residual = None
+        for i, layer in enumerate(self.layers):
+            hidden_states, residual = layer(
+                hidden_states,
+                residual,
+                cos,
+                sin,
+                cu_seqlens,
+                max_s,
+                past_key_values[i],
+                layer_past_present_indices,
+                cu_seqlens_q,
+            )
+
+        # Faster final layer norm
+        hidden_states, *rest = dropout_layer_norm.dropout_add_ln_fwd(
+            hidden_states,
+            residual,
+            self.final_layer_norm.weight,
+            self.final_layer_norm.bias,
+            None,
+            None,
+            None,
+            None,
+            0.0,
+            self.final_layer_norm.eps,
+            1.0,
+            0,
+            None,
+            False,
+            False,
+        )
+
+        return hidden_states, past_key_values
+
+
+class FlashGPTNeoXForCausalLM(FlashGPTNeoXPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        if config.tp_parallel:
+            process_group = torch.distributed.distributed_c10d._get_default_group()
+        else:
+            process_group = None
+
+        self.gpt_neox = FlashGPTNeoXModel(config, process_group)
+
+        if self.gpt_neox.tp_embeddings:
+            self.embed_out = nn.Linear(
+                config.hidden_size,
+                config.vocab_size // process_group.size(),
+                bias=False,
+            )
+        else:
+            self.embed_out = nn.Linear(
+                config.hidden_size, config.vocab_size, bias=False
+            )
+
+    def forward(
+        self,
+        input_ids,
+        position_ids,
+        cu_seqlens,
+        max_s,
+        past_key_values=None,
+    ):
+        hidden_states, present = self.gpt_neox(
+            input_ids, position_ids, cu_seqlens, max_s, past_key_values
+        )
+        return self.embed_out(hidden_states), present

server/text_generation_server/utils/tokens.py

@@ -24,7 +24,7 @@ class Sampling:
         self.seed = seed
 
     def __call__(self, logits):
-        probs = torch.nn.functional.softmax(logits)
+        probs = torch.nn.functional.softmax(logits, -1)
         next_tokens = torch.multinomial(probs, num_samples=1, generator=self.generator)
         return next_tokens
 

server/text_generation_server/utils/watermark.py

@@ -17,6 +17,7 @@ import os
 
 import torch
 from transformers import LogitsProcessor
+from typing import List, Union
 
 GAMMA = os.getenv("WATERMARK_GAMMA", 0.5)
 DELTA = os.getenv("WATERMARK_DELTA", 2.0)
@@ -36,7 +37,15 @@ class WatermarkLogitsProcessor(LogitsProcessor):
         self.rng = torch.Generator(device=device)
         self.hash_key = hash_key
 
-    def _seed_rng(self, input_ids: torch.LongTensor) -> None:
+    def _seed_rng(self, input_ids: Union[List[int], torch.LongTensor]):
+        if isinstance(input_ids, list):
+            assert (
+                len(input_ids) >= 1
+            ), "requires at least a 1 token prefix sequence to seed rng"
+            prev_token = input_ids[-1]
+        else:
+            input_ids = input_ids[0]
+            assert len(input_ids) == 1
             assert (
                 input_ids.shape[-1] >= 1
             ), "requires at least a 1 token prefix sequence to seed rng"
@@ -44,15 +53,16 @@ class WatermarkLogitsProcessor(LogitsProcessor):
         self.rng.manual_seed(self.hash_key * prev_token)
 
     def _get_greenlist_ids(
-        self, input_ids: torch.LongTensor, max_value: int
+        self,
-    ) -> list[int]:
+        input_ids: Union[List[int], torch.LongTensor],
+        max_value: int,
+        device: torch.device,
+    ) -> List[int]:
         # seed the rng using the previous tokens/prefix
         self._seed_rng(input_ids)
 
         greenlist_size = int(max_value * self.gamma)
-        vocab_permutation = torch.randperm(
+        vocab_permutation = torch.randperm(max_value, device=device, generator=self.rng)
-            max_value, device=input_ids.device, generator=self.rng
-        )
         greenlist_ids = vocab_permutation[:greenlist_size]
         return greenlist_ids
 
@@ -73,10 +83,11 @@ class WatermarkLogitsProcessor(LogitsProcessor):
         return scores
 
     def __call__(
-        self, input_ids: torch.LongTensor, scores: torch.FloatTensor
+        self, input_ids: Union[List[int], torch.LongTensor], scores: torch.FloatTensor
     ) -> torch.FloatTensor:
-        assert len(input_ids) == 1
+        greenlist_ids = self._get_greenlist_ids(
-        greenlist_ids = self._get_greenlist_ids(input_ids[0], scores.shape[-1])
+            input_ids, scores.shape[-1], scores.device
+        )
         green_tokens_mask = self._calc_greenlist_mask(
             scores=scores, greenlist_token_ids=greenlist_ids
         )