add intel xpu support for TGI (huggingface#1475)

# What does this PR do?

<!--
Congratulations! You've made it this far! You're not quite done yet
though.

Once merged, your PR is going to appear in the release notes with the
title you set, so make sure it's a great title that fully reflects the
extent of your awesome contribution.

Then, please replace this with a description of the change and which
issue is fixed (if applicable). Please also include relevant motivation
and context. List any dependencies (if any) that are required for this
change.

Once you're done, someone will review your PR shortly (see the section
"Who can review?" below to tag some potential reviewers). They may
suggest changes to make the code even better. If no one reviewed your PR
after a week has passed, don't hesitate to post a new comment
@-mentioning the same persons---sometimes notifications get lost.
-->

<!-- Remove if not applicable -->

Fixes # (issue)


## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
      Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
      to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
- [ ] Did you write any new necessary tests?


## Who can review?

Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.

<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @


@OlivierDehaene OR @Narsil

 -->

---------

Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
Co-authored-by: Morgan Funtowicz <funtowiczmo@gmail.com>
Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com>

Author: 3 people

.github/workflows/build.yaml

@@ -274,12 +274,105 @@ jobs:
           cache-from: type=registry,ref=registry.internal.huggingface.tech/api-inference/community/text-generation-inference:cache-rocm,mode=min
           cache-to: type=registry,ref=registry.internal.huggingface.tech/api-inference/community/text-generation-inference:cache-rocm,mode=min
 
+  build-and-push-image-intel:
+    concurrency:
+      group: ${{ github.workflow }}-build-and-push-image-intel-${{ github.head_ref || github.run_id }}
+      cancel-in-progress: true
+    needs:
+      - start-runner
+      - build-and-push-image # Wait for the main docker image to be built
+      - integration-tests # Wait for the main integration-tests
+    runs-on: ${{ needs.start-runner.outputs.label }} # run the job on the newly created runner
+    permissions:
+      contents: write
+      packages: write
+      # This is used to complete the identity challenge
+      # with sigstore/fulcio when running outside of PRs.
+      id-token: write
+      security-events: write
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v3
+      - name: Initialize Docker Buildx
+        uses: docker/setup-buildx-action@v2.0.0
+        with:
+          install: true
+      - name: Inject slug/short variables
+        uses: rlespinasse/github-slug-action@v4.4.1
+      - name: Tailscale
+        uses: tailscale/github-action@7bd8039bf25c23c4ab1b8d6e2cc2da2280601966
+        with:
+          authkey: ${{ secrets.TAILSCALE_AUTHKEY }}
+      - name: Login to GitHub Container Registry
+        if: github.event_name != 'pull_request'
+        uses: docker/login-action@v2
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+      - name: Login to internal Container Registry
+        uses: docker/login-action@v2.1.0
+        with:
+          username: ${{ secrets.TAILSCALE_DOCKER_USERNAME }}
+          password: ${{ secrets.TAILSCALE_DOCKER_PASSWORD }}
+          registry: registry.internal.huggingface.tech
+      - name: Login to Azure Container Registry
+        if: github.event_name != 'pull_request'
+        uses: docker/login-action@v2.1.0
+        with:
+          username: ${{ secrets.AZURE_DOCKER_USERNAME }}
+          password: ${{ secrets.AZURE_DOCKER_PASSWORD }}
+          registry: db4c2190dd824d1f950f5d1555fbadf0.azurecr.io
+      # If pull request
+      - name: Extract metadata (tags, labels) for Docker
+        if: ${{ github.event_name == 'pull_request' }}
+        id: meta-pr
+        uses: docker/metadata-action@v4.3.0
+        with:
+          images: |
+            registry.internal.huggingface.tech/api-inference/community/text-generation-inference
+          tags: |
+            type=raw,value=sha-${{ env.GITHUB_SHA_SHORT }}-intel
+      # If main, release or tag
+      - name: Extract metadata (tags, labels) for Docker
+        if: ${{ github.event_name != 'pull_request' }}
+        id: meta
+        uses: docker/metadata-action@v4.3.0
+        with:
+          flavor: |
+            latest=false
+          images: |
+            registry.internal.huggingface.tech/api-inference/community/text-generation-inference
+            ghcr.io/huggingface/text-generation-inference
+            db4c2190dd824d1f950f5d1555fbadf0.azurecr.io/text-generation-inference
+          tags: |
+            type=semver,pattern={{version}}-intel
+            type=semver,pattern={{major}}.{{minor}}-intel
+            type=raw,value=latest-intel,enable=${{ github.ref == format('refs/heads/{0}', github.event.repository.default_branch) }}
+            type=raw,value=sha-${{ env.GITHUB_SHA_SHORT }}-intel
+      - name: Build and push Docker image
+        id: build-and-push
+        uses: docker/build-push-action@v4
+        with:
+          context: .
+          file: Dockerfile_intel
+          push: true
+          platforms: 'linux/amd64'
+          build-args: |
+            GIT_SHA=${{ env.GITHUB_SHA }}
+            DOCKER_LABEL=sha-${{ env.GITHUB_SHA_SHORT }}-intel
+          tags: ${{ steps.meta.outputs.tags || steps.meta-pr.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels || steps.meta-pr.outputs.labels }}
+          cache-from: type=registry,ref=registry.internal.huggingface.tech/api-inference/community/text-generation-inference:cache-intel,mode=min
+          cache-to: type=registry,ref=registry.internal.huggingface.tech/api-inference/community/text-generation-inference:cache-intel,mode=min
+
   stop-runner:
     name: Stop self-hosted EC2 runner
     needs:
       - start-runner
       - build-and-push-image
       - build-and-push-image-rocm
+      - build-and-push-image-intel
       - integration-tests
     runs-on: ubuntu-latest
     env:

Dockerfile_intel

@@ -0,0 +1,105 @@
+FROM lukemathwalker/cargo-chef:latest-rust-1.75 AS chef
+WORKDIR /usr/src
+
+ARG CARGO_REGISTRIES_CRATES_IO_PROTOCOL=sparse
+
+FROM chef as planner
+COPY Cargo.toml Cargo.toml
+COPY rust-toolchain.toml rust-toolchain.toml
+COPY proto proto
+COPY benchmark benchmark
+COPY router router
+COPY launcher launcher
+RUN cargo chef prepare --recipe-path recipe.json
+
+FROM chef AS builder
+
+ARG GIT_SHA
+ARG DOCKER_LABEL
+
+RUN PROTOC_ZIP=protoc-21.12-linux-x86_64.zip && \
+    curl -OL https://github.com/protocolbuffers/protobuf/releases/download/v21.12/$PROTOC_ZIP && \
+    unzip -o $PROTOC_ZIP -d /usr/local bin/protoc && \
+    unzip -o $PROTOC_ZIP -d /usr/local 'include/*' && \
+    rm -f $PROTOC_ZIP
+
+COPY --from=planner /usr/src/recipe.json recipe.json
+RUN cargo chef cook --release --recipe-path recipe.json
+
+COPY Cargo.toml Cargo.toml
+COPY rust-toolchain.toml rust-toolchain.toml
+COPY proto proto
+COPY benchmark benchmark
+COPY router router
+COPY launcher launcher
+RUN cargo build --release
+
+
+# Text Generation Inference base image for Intel
+FROM intel/intel-extension-for-pytorch:2.1.10-xpu as base
+
+USER root
+# libssl.so.1.1 is not installed on Ubuntu 22.04 by default, install it
+RUN wget http://nz2.archive.ubuntu.com/ubuntu/pool/main/o/openssl/libssl1.1_1.1.1f-1ubuntu2_amd64.deb && \
+    dpkg -i ./libssl1.1_1.1.1f-1ubuntu2_amd64.deb
+
+
+RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB \
+| gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list
+
+RUN apt-get update && apt install -y intel-basekit xpu-smi cmake python3-dev ninja-build
+
+# Text Generation Inference base env
+ENV HUGGINGFACE_HUB_CACHE=/data \
+    HF_HUB_ENABLE_HF_TRANSFER=1 \
+    PORT=80
+
+
+WORKDIR /usr/src
+# Build pytorch and ipex
+RUN git clone https://github.com/intel/intel-extension-for-pytorch && cd intel-extension-for-pytorch && git checkout -b xpu_main origin/xpu-main
+RUN git clone https://github.com/pytorch/pytorch.git && cd pytorch && git checkout 209f2fa8ff86652f67d75c2f19bf9cb9942fd018 && git apply /usr/src/intel-extension-for-pytorch/torch_patches/00*.patch
+
+# Install server
+COPY proto proto
+COPY server server
+COPY server/Makefile server/Makefile
+RUN cd server && \
+    make gen-server && \
+    pip install -r requirements_cuda.txt && \
+    pip install ".[accelerate, peft, outlines]" --no-cache-dir
+
+ENV CCL_ROOT=/opt/intel/oneapi/ccl/latest
+ENV I_MPI_ROOT=/opt/intel/oneapi/mpi/latest
+ENV FI_PROVIDER_PATH=/opt/intel/oneapi/mpi/latest/opt/mpi/libfabric/lib/prov:/usr/lib/x86_64-linux-gnu/libfabric
+ENV DIAGUTIL_PATH=/opt/intel/oneapi/compiler/latest/etc/compiler/sys_check/sys_check.sh
+ENV CCL_CONFIGURATION=cpu_gpu_dpcpp
+ENV MANPATH=/opt/intel/oneapi/mpi/latest/share/man:/opt/intel/oneapi/mpi/latest/share/man:/opt/intel/oneapi/compiler/latest/share/man
+ENV CMAKE_PREFIX_PATH=/opt/intel/oneapi/mkl/latest/lib/cmake:/opt/intel/oneapi/compiler/latest
+ENV CMPLR_ROOT=/opt/intel/oneapi/compiler/latest
+ENV LIBRARY_PATH=/opt/intel/oneapi/mpi/latest/lib:/opt/intel/oneapi/ccl/latest/lib/:/opt/intel/oneapi/mkl/latest/lib/:/opt/intel/oneapi/compiler/latest/lib
+ENV OCL_ICD_FILENAMES=libintelocl_emu.so:libalteracl.so:/opt/intel/oneapi/compiler/latest/lib/libintelocl.so
+ENV CLASSPATH=/opt/intel/oneapi/mpi/latest/share/java/mpi.jar:/opt/intel/oneapi/mpi/latest/share/java/mpi.jar
+ENV LD_LIBRARY_PATH=/opt/intel/oneapi/ccl/latest/lib/:/opt/intel/oneapi/mpi/latest/opt/mpi/libfabric/lib:/opt/intel/oneapi/mpi/latest/lib:/opt/intel/oneapi/mkl/latest/lib:/opt/intel/oneapi/compiler/latest/opt/compiler/lib:/opt/intel/oneapi/compiler/latest/lib:/opt/intel/oneapi/lib:/opt/intel/oneapi/lib/intel64:
+ENV MKLROOT=/opt/intel/oneapi/mkl/latest
+ENV NLSPATH=/opt/intel/oneapi/mkl/latest/share/locale/%l_%t/%N:/opt/intel/oneapi/compiler/latest/lib/locale/%l_%t/%N
+ENV PATH=/opt/intel/oneapi/mpi/latest/opt/mpi/libfabric/bin:/opt/intel/oneapi/mpi/latest/bin:/opt/intel/oneapi/mpi/latest/opt/mpi/libfabric/bin:/opt/intel/oneapi/mkl/latest/bin/:/opt/intel/oneapi/compiler/latest/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
+ENV CPATH=/opt/intel/oneapi/mpi/latest/include:/opt/intel/oneapi/ccl/latest/include:/opt/intel/oneapi/mkl/latest/include
+ENV CCL_ZE_IPC_EXCHANGE=sockets
+
+
+RUN pip uninstall -y torch && cd pytorch && git submodule update --init --recursive && python setup.py install
+RUN pip uninstall -y intel-extension-for-pytorch && cd intel-extension-for-pytorch && git submodule update --init --recursive && USE_AOT_DEVLIST='pvc' BUILD_SEPARATE_OPS=ON BUILD_WITH_CPU=ON USE_XETLA=ON python setup.py install
+
+# Install benchmarker
+COPY --from=builder /usr/src/target/release/text-generation-benchmark /usr/local/bin/text-generation-benchmark
+# Install router
+COPY --from=builder /usr/src/target/release/text-generation-router /usr/local/bin/text-generation-router
+# Install launcher
+COPY --from=builder /usr/src/target/release/text-generation-launcher /usr/local/bin/text-generation-launcher
+
+# Final image
+FROM base
+
+ENTRYPOINT ["text-generation-launcher"]
+CMD ["--json-output"]

launcher/src/env_runtime.rs

@@ -7,14 +7,17 @@ pub(crate) struct Env {
     git_sha: &'static str,
     docker_label: &'static str,
     nvidia_env: String,
+    xpu_env: String,
 }
 
 impl Env {
     pub fn new() -> Self {
         let nvidia_env = nvidia_smi();
+        let xpu_env = xpu_smi();
 
         Self {
             nvidia_env: nvidia_env.unwrap_or("N/A".to_string()),
+            xpu_env: xpu_env.unwrap_or("N/A".to_string()),
             cargo_target: env!("VERGEN_CARGO_TARGET_TRIPLE"),
             cargo_version: env!("VERGEN_RUSTC_SEMVER"),
             git_sha: option_env!("VERGEN_GIT_SHA").unwrap_or("N/A"),
@@ -31,7 +34,8 @@ impl fmt::Display for Env {
         writeln!(f, "Cargo version: {}", self.cargo_version)?;
         writeln!(f, "Commit sha: {}", self.git_sha)?;
         writeln!(f, "Docker label: {}", self.docker_label)?;
-        write!(f, "nvidia-smi:\n{}", self.nvidia_env)?;
+        writeln!(f, "nvidia-smi:\n{}", self.nvidia_env)?;
+        write!(f, "xpu-smi:\n{}", self.xpu_env)?;
 
         Ok(())
     }
@@ -43,3 +47,10 @@ fn nvidia_smi() -> Option<String> {
     let output = nvidia_smi.replace('\n', "\n   ");
     Some(output.trim().to_string())
 }
+
+fn xpu_smi() -> Option<String> {
+    let output = Command::new("xpu-smi").arg("discovery").output().ok()?;
+    let xpu_smi = String::from_utf8(output.stdout).ok()?;
+    let output = xpu_smi.replace('\n', "\n   ");
+    Some(output.trim().to_string())
+}

server/text_generation_server/models/cache_manager.py

@@ -2,6 +2,7 @@
 import torch
 
 from typing import Optional, List, Tuple
+from text_generation_server.utils.import_utils import IS_XPU_SYSTEM
 
 BLOCK_SIZE: int = 16
 # Will be set in warmup
@@ -24,7 +25,10 @@ def __init__(
         self.repeat_slots = repeat_slots
 
         element_size = torch.tensor([], dtype=dtype).element_size()
-        x = self.block_size // element_size
+        if IS_XPU_SYSTEM:
+            x = 1
+        else:
+            x = self.block_size // element_size
 
         self.kv_cache = [
             (

server/text_generation_server/models/custom_modeling/flash_dbrx_modeling.py

@@ -21,8 +21,10 @@
 from transformers.configuration_utils import PretrainedConfig
 from typing import Optional, List, Tuple, Any
 from loguru import logger
+from text_generation_server.utils.import_utils import IS_XPU_SYSTEM
 
-from vllm.model_executor.layers.fused_moe import fused_moe
+if not IS_XPU_SYSTEM:
+    from vllm.model_executor.layers.fused_moe import fused_moe
 from text_generation_server.utils import paged_attention, flash_attn
 from text_generation_server.utils.layers import (
     FastLinear,

server/text_generation_server/models/custom_modeling/flash_mixtral_modeling.py

@@ -24,7 +24,10 @@
 import numpy as np
 
 from torch import nn
-from vllm.model_executor.layers.fused_moe import fused_moe
+from text_generation_server.utils.import_utils import IS_XPU_SYSTEM
+
+if not IS_XPU_SYSTEM:
+    from vllm.model_executor.layers.fused_moe import fused_moe
 from transformers.activations import ACT2FN
 from transformers.configuration_utils import PretrainedConfig
 from typing import Optional, List, Tuple

server/text_generation_server/models/flash_causal_lm.py

@@ -33,7 +33,7 @@
 from text_generation_server.utils.dist import MEMORY_FRACTION
 
 tracer = trace.get_tracer(__name__)
-
+from text_generation_server.utils.import_utils import IS_CUDA_SYSTEM, IS_ROCM_SYSTEM, IS_XPU_SYSTEM
 
 @dataclass
 class FlashCausalLMBatch(Batch):
@@ -752,7 +752,10 @@ def cuda_graph_warmup(self, bs: int, max_s: int, max_bt: int):
 
     def warmup(self, batch: FlashCausalLMBatch):
         # The warmup batch is the biggest batch we could ever receive
-        torch.cuda.empty_cache()
+        if IS_CUDA_SYSTEM or IS_ROCM_SYSTEM:
+            torch.cuda.empty_cache()
+        elif IS_XPU_SYSTEM:
+            torch.xpu.empty_cache()
         try:
             cache_manager = set_cache_manager(
                 batch.blocks,
@@ -772,20 +775,29 @@ def warmup(self, batch: FlashCausalLMBatch):
                 f"You need to decrease `--max-batch-prefill-tokens`"
             ) from e
 
-        torch.cuda.synchronize(self.device)
+        if IS_CUDA_SYSTEM or IS_ROCM_SYSTEM:
+            torch.cuda.synchronize(self.device)
+        elif IS_XPU_SYSTEM:
+            torch.xpu.synchronize(self.device)
 
         # Inspired by the original implementation in [vllm](https://github.com/vllm-project/vllm)
         # Calculate the number of blocks that can be allocated with the free memory
         dtype_size = torch.tensor([], dtype=self.dtype).element_size()
         cache_block_size = BLOCK_SIZE * self.num_kv_heads * self.head_size
         total_cache_size = self.num_layers * cache_block_size * 2 * dtype_size
 
-        total_free_memory, _ = torch.cuda.mem_get_info(self.device)
-        total_gpu_memory = torch.cuda.get_device_properties(self.device).total_memory
+        if IS_CUDA_SYSTEM or IS_ROCM_SYSTEM:
+            total_free_memory, _ = torch.cuda.mem_get_info(self.device)
+            total_gpu_memory = torch.cuda.get_device_properties(self.device).total_memory
 
-        free_memory = max(
-            0, total_free_memory - (1 - MEMORY_FRACTION) * total_gpu_memory
-        )
+            free_memory = max(
+                0, total_free_memory - (1 - MEMORY_FRACTION) * total_gpu_memory
+            )
+        elif IS_XPU_SYSTEM:
+            total_gpu_memory = torch.xpu.get_device_properties(self.device).total_memory
+            free_memory = int(total_gpu_memory *0.5)
+        else:
+            raise NotImplementedError("FlashModel is only available on GPU")
 
         num_blocks = (
             # Leave 5% for some wiggle room

server/text_generation_server/models/flash_llama.py

@@ -18,6 +18,7 @@
 
 tracer = trace.get_tracer(__name__)
 
+from text_generation_server.utils.import_utils import IS_XPU_SYSTEM
 
 class FlashLlama(FlashCausalLM):
     def __init__(
@@ -33,6 +34,9 @@ def __init__(
         if torch.cuda.is_available():
             device = torch.device(f"cuda:{rank}")
             dtype = torch.float16 if dtype is None else dtype
+        elif IS_XPU_SYSTEM:
+            device = torch.device(f"xpu:{rank}")
+            dtype = torch.float16 if dtype is None else dtype
         else:
             raise NotImplementedError("FlashLlama is only available on GPU")
 

server/text_generation_server/models/flash_mistral.py

@@ -33,8 +33,9 @@
 # Will be set in init
 SLIDING_WINDOW: Optional[int] = None
 SLIDING_WINDOW_BLOCKS: Optional[int] = None
+from text_generation_server.utils.import_utils import IS_XPU_SYSTEM
 
-MEM_POOL = torch.cuda.graph_pool_handle()
+MEM_POOL = torch.cuda.graph_pool_handle() if torch.cuda.is_available() else None
 
 
 def set_sliding_window(sliding_window: int, sliding_window_blocks: int):
@@ -316,6 +317,9 @@ def __init__(
         if torch.cuda.is_available():
             device = torch.device(f"cuda:{rank}")
             dtype = torch.float16 if dtype is None else dtype
+        elif IS_XPU_SYSTEM:
+            device = torch.device(f"xpu:{rank}")
+            dtype = torch.float16 if dtype is None else dtype
         else:
             raise NotImplementedError("FlashMistral is only available on GPU")