feat(server): support OPT models (#55)

OPT models do not all have a `tokenizer.json` file on the hub at the
moment. Can't merge for now.

README.md

@@ -54,11 +54,12 @@ to power LLMs api-inference widgets.
 ## Optimized architectures
 
 - [BLOOM](https://huggingface.co/bigscience/bloom)
-- [Galactica](https://huggingface.co/facebook/galactica-120b)
-- [SantaCoder](https://huggingface.co/bigcode/santacoder)
-- [GPT-Neox](https://huggingface.co/EleutherAI/gpt-neox-20b)
 - [FLAN-T5](https://huggingface.co/google/flan-t5-xxl)
+- [Galactica](https://huggingface.co/facebook/galactica-120b)
+- [GPT-Neox](https://huggingface.co/EleutherAI/gpt-neox-20b)
 - [Llama](https://github.com/facebookresearch/llama)
+- [OPT](https://huggingface.co/facebook/opt-66b)
+- [SantaCoder](https://huggingface.co/bigcode/santacoder)
 
 Other architectures are supported on a best effort basis using:
 

launcher/src/main.rs

@@ -349,8 +349,8 @@ fn main() -> ExitCode {
             Err(TryRecvError::Empty) => {
                 sleep(Duration::from_millis(100));
             }
-            Ok(ShardStatus::Failed((rank, err))) => {
-                tracing::error!("Shard {} failed to start:\n{}", rank, err);
+            Ok(ShardStatus::Failed(rank)) => {
+                tracing::error!("Shard {} failed to start.", rank);
                 shutdown_shards(shutdown, &shutdown_receiver);
                 return ExitCode::FAILURE;
             }
@@ -457,8 +457,8 @@ fn main() -> ExitCode {
     let mut exit_code = ExitCode::SUCCESS;
 
     while running.load(Ordering::SeqCst) {
-        if let Ok(ShardStatus::Failed((rank, err))) = status_receiver.try_recv() {
-            tracing::error!("Shard {rank} failed:\n{err}");
+        if let Ok(ShardStatus::Failed(rank)) = status_receiver.try_recv() {
+            tracing::error!("Shard {rank} failed.");
             exit_code = ExitCode::FAILURE;
             break;
         };
@@ -488,7 +488,7 @@ fn main() -> ExitCode {
 #[derive(Debug)]
 enum ShardStatus {
     Ready,
-    Failed((usize, String)),
+    Failed(usize),
 }
 
 #[allow(clippy::too_many_arguments)]
@@ -627,9 +627,7 @@ fn shard_manager(
                     tracing::error!("Please install it with `make install-server`")
                 }
             }
-            status_sender
-                .send(ShardStatus::Failed((rank, err.to_string())))
-                .unwrap();
+            status_sender.send(ShardStatus::Failed(rank)).unwrap();
             return;
         }
     };
@@ -658,11 +656,7 @@ fn shard_manager(
     loop {
         // Process exited
         if p.poll().is_some() {
-            let mut err = String::new();
-            p.stderr.take().unwrap().read_to_string(&mut err).unwrap();
-            status_sender
-                .send(ShardStatus::Failed((rank, err)))
-                .unwrap();
+            status_sender.send(ShardStatus::Failed(rank)).unwrap();
             return;
         }
 

server/text_generation_server/models/__init__.py

@@ -1,4 +1,3 @@
-import os
 import torch
 
 from loguru import logger
@@ -11,6 +10,7 @@ from text_generation_server.models.causal_lm import CausalLM
 from text_generation_server.models.flash_causal_lm import FlashCausalLM
 from text_generation_server.models.bloom import BLOOM, BLOOMSharded
 from text_generation_server.models.seq2seq_lm import Seq2SeqLM
+from text_generation_server.models.opt import OPT, OPTSharded
 from text_generation_server.models.galactica import Galactica, GalacticaSharded
 from text_generation_server.models.santacoder import SantaCoder
 from text_generation_server.models.gpt_neox import GPTNeoxSharded
@@ -36,7 +36,11 @@ __all__ = [
     "GalacticaSharded",
     "GPTNeoxSharded",
     "Seq2SeqLM",
+    "Galactica",
+    "GalacticaSharded",
     "SantaCoder",
+    "OPT",
+    "OPTSharded",
     "T5Sharded",
     "get_model",
 ]
@@ -48,9 +52,11 @@ if FLASH_ATTENTION:
     __all__.append(FlashLlama)
     __all__.append(FlashLlamaSharded)
 
-FLASH_ATT_ERROR_MESSAGE = "{} requires Flash Attention CUDA kernels to be installed.\n" \
-                          "Use the official Docker image (ghcr.io/huggingface/text-generation-inference:latest) " \
+FLASH_ATT_ERROR_MESSAGE = (
+    "{} requires Flash Attention CUDA kernels to be installed.\n"
+    "Use the official Docker image (ghcr.io/huggingface/text-generation-inference:latest) "
     "or install flash attention with `cd server && make install install-flash-attention`"
+)
 
 # The flag below controls whether to allow TF32 on matmul. This flag defaults to False
 # in PyTorch 1.12 and later.
@@ -100,13 +106,17 @@ def get_model(
         if sharded:
             if FLASH_ATTENTION:
                 return FlashLlamaSharded(model_id, revision, quantize=quantize)
-            raise NotImplementedError(
-                FLASH_ATT_ERROR_MESSAGE.format(f"Sharded Llama")
-            )
+            raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format(f"Sharded Llama"))
         else:
             llama_cls = FlashLlama if FLASH_ATTENTION else CausalLM
             return llama_cls(model_id, revision, quantize=quantize)
 
+    if config.model_type == "opt":
+        if sharded:
+            return OPTSharded(model_id, revision, quantize=quantize)
+        else:
+            return OPT(model_id, revision, quantize=quantize)
+
     if model_type == "t5":
         if sharded:
             return T5Sharded(model_id, revision, quantize=quantize)

server/text_generation_server/models/bloom.py

@@ -62,7 +62,7 @@ class BLOOMSharded(BLOOM):
         self.master = self.rank == 0
         if torch.cuda.is_available():
             device = torch.device(f"cuda:{self.rank}")
-            dtype = torch.bfloat16
+            dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float32
         else:
             device = torch.device("cpu")
             dtype = torch.float32
@@ -122,13 +122,6 @@ class BLOOMSharded(BLOOM):
                     slice_ = f.get_slice(name)
 
                     if isinstance(module, TensorParallelColumnLinear):
-                        if param_name == "weight":
-                            size = slice_.get_shape()[0]
-                            block_size = size // world_size
-                            start = rank * block_size
-                            stop = (rank + 1) * block_size
-                            tensor = slice_[start:stop]
-                        else:
                         size = slice_.get_shape()[0]
                         block_size = size // world_size
                         start = rank * block_size

server/text_generation_server/models/galactica.py

@@ -19,8 +19,9 @@ from transformers.models.opt.parallel_layers import (
 )
 
 from text_generation_server.models import CausalLM
-from text_generation_server.pb import generate_pb2
 from text_generation_server.models.causal_lm import CausalLMBatch
+from text_generation_server.pb import generate_pb2
+from text_generation_server.models.opt import OPT
 from text_generation_server.utils import (
     NextTokenChooser,
     StoppingCriteria,
@@ -158,7 +159,7 @@ class GalacticaCausalLMBatch(CausalLMBatch):
         )
 
 
-class Galactica(CausalLM):
+class Galactica(OPT):
     @property
     def batch_type(self) -> Type[CausalLMBatch]:
         return GalacticaCausalLMBatch
@@ -192,7 +193,7 @@ class GalacticaSharded(Galactica):
         self.master = self.rank == 0
         if torch.cuda.is_available():
             device = torch.device(f"cuda:{self.rank}")
-            dtype = torch.bfloat16
+            dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float32
         else:
             device = torch.device("cpu")
             dtype = torch.float32
@@ -253,13 +254,6 @@ class GalacticaSharded(Galactica):
                     slice_ = f.get_slice(name)
 
                     if isinstance(module, TensorParallelColumnLinear):
-                        if param_name == "weight":
-                            size = slice_.get_shape()[0]
-                            block_size = size // world_size
-                            start = rank * block_size
-                            stop = (rank + 1) * block_size
-                            tensor = slice_[start:stop]
-                        else:
                         size = slice_.get_shape()[0]
                         block_size = size // world_size
                         start = rank * block_size

server/text_generation_server/models/gpt_neox.py

@@ -38,7 +38,7 @@ class GPTNeoxSharded(CausalLM):
         self.master = self.rank == 0
         if torch.cuda.is_available():
             device = torch.device(f"cuda:{self.rank}")
-            dtype = torch.bfloat16
+            dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float32
         else:
             device = torch.device("cpu")
             dtype = torch.float32

server/text_generation_server/models/opt.py

@@ -0,0 +1,224 @@
+import torch
+import torch.distributed
+
+from typing import List, Optional, Tuple
+
+from accelerate import init_empty_weights
+from safetensors import safe_open
+from transformers import (
+    AutoTokenizer,
+    AutoModelForCausalLM,
+    AutoConfig,
+)
+from transformers.models.opt.parallel_layers import (
+    TensorParallelColumnLinear,
+    TensorParallelEmbedding,
+    TensorParallelRowLinear,
+)
+
+from text_generation_server.models import CausalLM
+from text_generation_server.utils import (
+    initialize_torch_distributed,
+    weight_files,
+)
+
+HAS_BITS_AND_BYTES = True
+try:
+    import bitsandbytes as bnb
+    from bitsandbytes.nn import Int8Params
+except Exception as e:
+    HAS_BITS_AND_BYTES = False
+
+
+class OPT(CausalLM):
+    def forward(
+        self, input_ids, attention_mask, position_ids, past_key_values: Optional = None
+    ) -> Tuple[torch.Tensor, List[Tuple[torch.Tensor, torch.Tensor]]]:
+        """Overwrite forward to ignore position_ids"""
+
+        # Model Forward
+        outputs = self.model.forward(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=True,
+        )
+        return outputs.logits, outputs.past_key_values
+
+
+class OPTSharded(OPT):
+    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.float32
+        else:
+            device = torch.device("cpu")
+            dtype = torch.float32
+
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_id, revision=revision, padding_side="left", truncation_side="left"
+        )
+
+        config = AutoConfig.from_pretrained(
+            model_id, revision=revision, tp_parallel=True
+        )
+        tokenizer.pad_token_id = config.pad_token_id
+
+        torch.distributed.barrier(group=self.process_group)
+        filenames = weight_files(model_id, revision=revision, extension=".safetensors")
+
+        with init_empty_weights():
+            model = AutoModelForCausalLM.from_config(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(CausalLM, 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():
+                    if name == "lm_head.weight":
+                        continue
+
+                    module_name, param_name = name.rsplit(".", 1)
+                    module = model.get_submodule(module_name)
+                    current_tensor = parameters[name]
+
+                    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]
+                    else:
+                        tensor = slice_[:]
+
+                    if current_tensor.shape != tensor.shape:
+                        raise ValueError(
+                            f"Name {name} -- Current {current_tensor.shape} and got {tensor.shape}"
+                        )
+
+                    tensor = tensor.contiguous()
+
+                    if quantize:
+                        if not HAS_BITS_AND_BYTES:
+                            raise ImportError(
+                                "bitsandbytes is not available on your machine either because it is not installed "
+                                "or you don't have a GPU.\n"
+                                "You can install it with `pip install bitsandbytes`."
+                            )
+
+                        if (
+                            type(module)
+                            in [TensorParallelRowLinear, TensorParallelColumnLinear]
+                            and param_name == "weight"
+                        ):
+                            tensor = Int8Params(
+                                tensor,
+                                has_fp16_weights=False,
+                                requires_grad=False,
+                            ).to(device)
+                            state = bnb.MatmulLtState()
+                            state.threshold = 6.0
+                            state.has_fp16_weights = False
+                            state.memory_efficient_backward = False
+                            state.use_pool = True
+                            state.CB = tensor.CB
+                            state.SCB = tensor.SCB
+                            tensor.CB = None
+                            tensor.SCB = None
+
+                            def replace_linear(state):
+                                def linear(input, weight, bias):
+                                    out = bnb.matmul(
+                                        input,
+                                        weight,
+                                        state=state,
+                                        threshold=state.threshold,
+                                        bias=bias,
+                                    )
+
+                                    if state.CB is not None:
+                                        # we converted 8-bit row major to turing/ampere format
+                                        # in the first inference pass
+                                        # we no longer need the row-major weight
+                                        del state.CB
+                                        weight.data = state.CxB
+
+                                    return out
+
+                                return linear
+
+                            module.linear = replace_linear(state)
+
+                        else:
+                            tensor = tensor.to(device)
+
+                    module._parameters[param_name] = tensor
+                    if name == "model.decoder.embed_tokens.weight":
+                        model.lm_head._parameters["weight"] = tensor
+
+    def forward(
+        self, input_ids, attention_mask, position_ids, past_key_values: Optional = None
+    ):
+        outputs = self.model.forward(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=True,
+        )
+
+        # Logits are sharded, so we need to gather them
+        logits = [torch.empty_like(outputs.logits) for _ in range(self.world_size)]
+        torch.distributed.all_gather(logits, outputs.logits, group=self.process_group)
+        logits = torch.cat(logits, dim=2)
+
+        return logits, outputs.past_key_values

server/text_generation_server/models/t5.py

@@ -38,7 +38,7 @@ class T5Sharded(Seq2SeqLM):
         self.master = self.rank == 0
         if torch.cuda.is_available():
             device = torch.device(f"cuda:{self.rank}")
-            dtype = torch.bfloat16
+            dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float32
         else:
             device = torch.device("cpu")
             dtype = torch.float32

server/text_generation_server/utils/hub.py

@@ -50,7 +50,6 @@ def try_to_load_from_cache(
 
     refs_dir = repo_cache / "refs"
     snapshots_dir = repo_cache / "snapshots"
-    no_exist_dir = repo_cache / ".no_exist"
 
     # Resolve refs (for instance to convert main to the associated commit sha)
     if refs_dir.is_dir():
@@ -59,10 +58,6 @@ def try_to_load_from_cache(
             with revision_file.open() as f:
                 revision = f.read()
 
-    # Check if file is cached as "no_exist"
-    if (no_exist_dir / revision / filename).is_file():
-        return None
-
     # Check if revision folder exists
     if not snapshots_dir.exists():
         return None