v1.0.0 (#727)

Cargo.toml

@@ -8,7 +8,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.9.4"
+version = "1.0.0"
 edition = "2021"
 authors = ["Olivier Dehaene"]
 homepage = "https://github.com/huggingface/text-generation-inference"

README.md

@@ -7,16 +7,14 @@
 <a href="https://github.com/huggingface/text-generation-inference">
   <img alt="GitHub Repo stars" src="https://img.shields.io/github/stars/huggingface/text-generation-inference?style=social">
 </a>
-<a href="https://github.com/huggingface/text-generation-inference/blob/main/LICENSE">
-  <img alt="License" src="https://img.shields.io/github/license/huggingface/text-generation-inference">
-</a>
 <a href="https://huggingface.github.io/text-generation-inference">
   <img alt="Swagger API documentation" src="https://img.shields.io/badge/API-Swagger-informational">
 </a>
-</div>
 
 A Rust, Python and gRPC server for text generation inference. Used in production at [HuggingFace](https://huggingface.co)
-to power LLMs api-inference widgets.
+to power Hugging Chat, the Inference API and Inference Endpoint.
+
+</div>
 
 ## Table of contents
 
@@ -85,7 +83,7 @@ The easiest way of getting started is using the official Docker container:
 model=tiiuae/falcon-7b-instruct
 volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run
 
-docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:0.9.4 --model-id $model
+docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:1.0.0 --model-id $model
 ```
 **Note:** To use GPUs, you need to install the [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html). We also recommend using NVIDIA drivers with CUDA version 11.8 or higher.
 
@@ -152,7 +150,7 @@ model=meta-llama/Llama-2-7b-chat-hf
 volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run
 token=<your cli READ token>
 
-docker run --gpus all --shm-size 1g -e HUGGING_FACE_HUB_TOKEN=$token -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:0.9.3 --model-id $model
+docker run --gpus all --shm-size 1g -e HUGGING_FACE_HUB_TOKEN=$token -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:1.0.0 --model-id $model
 ```
 
 ### A note on Shared Memory (shm)

docs/openapi.json

@@ -10,7 +10,7 @@
       "name": "Apache 2.0",
       "url": "https://www.apache.org/licenses/LICENSE-2.0"
     },
-    "version": "0.9.4"
+    "version": "1.0.0"
   },
   "paths": {
     "/": {

server/Makefile-vllm

@@ -1,4 +1,4 @@
-vllm_commit := 084ca75d4271f8f67be731bc58e0d41d8e0afd3a
+vllm_commit := d284b831c17f42a8ea63369a06138325f73c4cf9
 
 vllm:
     # Clone vllm

server/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "text-generation-server"
-version = "0.9.4"
+version = "1.0.0"
 description = "Text Generation Inference Python gRPC Server"
 authors = ["Olivier Dehaene <[email protected]>"]
 

server/text_generation_server/utils/layers.py

@@ -219,31 +219,36 @@ def load(config, prefix: str, weights):
         )
 
     def forward(self, input: torch.Tensor) -> torch.Tensor:
+        if not self.should_gather:
+            return super().forward(input)
+
         world_size = self.process_group.size()
-        # Fast branch for single requests
-        if (
-            self.should_gather
-            and len(input.shape) == 2
-            and isinstance(self.linear, FastLinear)
-            and input.shape[0] == 1
-        ):
+        if len(input.shape) == 2 and isinstance(self.linear, FastLinear):
             out_dim = self.linear.weight.shape[0]
 
-            world_out = input.new_empty(1, out_dim * world_size)
-            local_out = input.new_empty(1, out_dim)
+            if input.shape[0] == 1:
+                world_out = input.new_empty(1, out_dim * world_size)
+                local_out = input.new_empty(1, out_dim)
+                gather_input = local_out
+            else:
+                world_out = input.new_empty(out_dim * world_size, input.shape[0])
+                gather_input = input.new_empty(out_dim, input.shape[0])
+                local_out = gather_input.T
 
             torch.mm(input, self.linear.weight.T, out=local_out)
 
             torch.distributed.all_gather_into_tensor(
-                world_out, local_out, group=self.process_group
+                world_out, gather_input, group=self.process_group
             )
-            return world_out
 
-        output = super().forward(input)
-        if not self.should_gather:
-            return output
+            if input.shape[0] == 1:
+                return world_out
+            return world_out.T
 
-        world_output = [torch.empty_like(output) for _ in range(world_size)]
+        output = super().forward(input)
+        world_output = [
+            torch.empty_like(output) for _ in range(self.process_group.size())
+        ]
         torch.distributed.all_gather(world_output, output, group=self.process_group)
         world_output = torch.cat(world_output, dim=-1)
         return world_output